Datasets:

kenhktsui

/

test_arxiv_classification

like 0

arXiv:1707.02772v2 [cs.PL] 24 Mar 2018 Probabilistic Program Equivalence for NetKAT∗ STEFFEN SMOLKA, Cornell University, USA PRAVEEN KUMAR, Cornell University, USA NATE FOSTER, Cornell University, USA JUSTIN HSU, Cornell University, USA DAVID KAHN, Cornell University, USA DEXTER KOZEN, Cornell University, USA ALEXANDRA SILVA, University College London, UK We tackle the problem of deciding whether two probabilistic programs are equivalent in Probabilistic NetKAT, a formal language for specifying and reasoning about the behavior of packet-switched networks. We show that the problem is decidable for the history-free fragment of the language by developing an effective decision procedure based on stochastic matrices. The main challenge lies in reasoning about iteration, which we address by designing an encoding of the program semantics as a finite-state absorbing Markov chain, whose limiting distribution can be computed exactly. In an extended case study on a real-world data center network, we automatically verify various quantitative properties of interest, including resilience in the presence of failures, by analyzing the Markov chain semantics. 1 INTRODUCTION Program equivalence is one of the most fundamental problems in Computer Science: given a pair of programs, do they describe the same computation? The problem is undecidable in general, but it can often be solved for domain-specific languages based on restricted computational models. For example, a classical approach for deciding whether a pair of regular expressions denote the same language is to first convert the expressions to deterministic finite automata, which can then be checked for equivalence in almost linear time [32]. In addition to the theoretical motivation, there are also many practical benefits to studying program equivalence. Being able to decide equivalence enables more sophisticated applications, for instance in verified compilation and program synthesis. Less obviously—but arguably more importantly—deciding equivalence typically involves finding some sort of finite, explicit representation of the program semantics. This compact encoding can open the door to reasoning techniques and decision procedures for properties that extend far beyond straightforward program equivalence. With this motivation in mind, this paper tackles the problem of deciding equivalence in Probabilistic NetKAT (ProbNetKAT), a language for modeling and reasoning about the behavior of packet-switched networks. As its name suggests, ProbNetKAT is based on NetKAT [3, 9, 30], which is in turn based on Kleene algebra with tests (KAT), an algebraic system combining Boolean predicates and regular expressions. ProbNetKAT extends NetKAT with a random choice operator and a semantics based on Markov kernels [31]. The framework can be used to encode and reason about randomized protocols (e.g., a routing scheme that uses random forwarding paths to balance load [33]); describe uncertainty about traffic demands (e.g., the diurnal/nocturnal fluctuation in access patterns commonly seen in networks for large content providers [26]); and model failures (e.g., switches or links that are known to fail with some probability [10]). However, the semantics of ProbNetKAT is surprisingly subtle. Using the iteration operator (i.e., the Kleene star from regular expressions), it is possible to write programs that generate continuous distributions over an uncountable space of packet history sets [8, Theorem 3]. This makes reasoning about convergence non-trivial, and requires representing infinitary objects compactly ∗ This is a preliminary draft from March 28, 2018. 2 S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva in an implementation. To address these issues, prior work [31] developed a domain-theoretic characterization of ProbNetKAT that provides notions of approximation and continuity, which can be used to reason about programs using only discrete distributions with finite support. However, that work left the decidability of program equivalence as an open problem. In this paper, we settle this question positively for the history-free fragment of the language, where programs manipulate sets of packets instead of sets of packet histories (finite sequences of packets). Our decision procedure works by deriving a canonical, explicit representation of the program semantics, for which checking equivalence is straightforward. Specifically, we define a big-step semantics that interprets each program as a finite stochastic matrix—equivalently, a Markov chain that transitions from input to output in a single step. Equivalence is trivially decidable on this representation, but the challenge lies in computing the big-step matrix for iteration—intuitively, the finite matrix needs to somehow capture the result of an infinite stochastic process. We address this by embedding the system in a more refined Markov chain with a larger state space, modeling iteration in the style of a small-step semantics. With some care, this chain can be transformed to an absorbing Markov chain, from which we derive a closed form analytic solution representing the limit of the iteration by applying elementary matrix calculations. We prove the soundness of this approach formally. Although the history-free fragment of ProbNetKAT is a restriction of the general language, it captures the input-output behavior of a network—mapping initial packet states to final packet states—and is still expressive enough to handle a wide range of problems of interest. Many other contemporary network verification tools, including Anteater [22], Header Space Analysis [15], and Veriflow [17], are also based on a history-free model. To handle properties that involve paths (e.g., waypointing), these tools generate a series of smaller properties to check, one for each hop in the path. In the ProbNetKAT implementation, working with history-free programs can reduce the space requirements by an exponential factor—a significant benefit when analyzing complex randomized protocols in large networks. We have built a prototype implementation of our approach in OCaml. The workhorse of the decision procedure computes a finite stochastic matrix—representing a finite Markov chain—given an input program. It leverages the spare linear solver UMFPACK [5] as a back-end to compute limiting distributions, and incorporates a number of optimizations and symbolic techniques to compactly represent large but sparse matrices. Although building a scalable implementation would require much more engineering (and is not the primary focus of this paper), our prototype is already able to handle programs of moderate size. Leveraging the finite encoding of the semantics, we have carried out several case studies in the context of data center networks; our central case study models and verifies the resilience of various routing schemes in the presence of link failures. Contributions and outline. The main contribution of this paper is the development of a decision procedure for history-free ProbNetKAT. We develop a new, tractable semantics in terms of stochastic matrices in two steps, we establish the soundness of the semantics with respect to ProbNetKAT’s original denotational model, and we use the compact semantics as the basis for building a prototype implementation with which we carry out case studies. In Section 2 and Section 3 we introduce ProbNetKAT using a simple example and motivate the need for quantitative, probabilistic reasoning. In Section 4, we present a semantics based on finite stochastic matrices and show that it fully characterizes the behavior of ProbNetKAT programs on packets (Theorem 4.1). In this big-step semantics, the matrices encode Markov chains over the state space 2Pk . A single step of the chain models the entire execution of a program, going directly from the initial state corresponding to the set of input packets the final state corresponding to the set of output packets. Although this reduces Probabilistic Program Equivalence for NetKAT 3 program equivalence to equality of finite matrices, we still need to provide a way to explicitly compute them. In particular, the matrix that models iteration is given in terms of a limit. In Section 5 we derive a closed form for the big-step matrix associated with p ∗ , giving an explicit representation of the big-step semantics. It is important to note that this is not simply the calculation of the stationary distribution of a Markov chain, as the semantics of p ∗ is more subtle. Instead, we define a small-step semantics, a second Markov chain with a larger state space such that one transition models one iteration of p ∗ . We then show how to transform this finer Markov chain into an absorbing Markov chain, which admits a closed form solution for its limiting distribution. Together, the big- and small-step semantics enable us to analytically compute a finite representation of the program semantics. Directly checking these semantics for equality yields an effective decision procedure for program equivalence (Corollary 5.8). This is in contrast with the previous semantics [8], which merely provided an approximation theorem for the semantics of iteration p ∗ and was not suitable for deciding equivalence. In Section 6, we illustrate the practical applicability of our approach by exploiting the representation of ProbNetKAT programs as stochastic matrices to answer a number of questions of interest in real-world networks. For example, we can reduce loop termination to program equivalence: the fact that the while loop below terminates with probability 1 can be checked as follows: while ¬f =0 do (skip ⊕r f ←0) ≡ f ←0 We also present real-world case studies that use the stochastic matrix representation to answer questions about the resilience of data center networks in the presence of link failures. We discuss obstacles in extending our approach to the full ProbNetKAT language in Section 7, including a novel automata model encodable in ProbNetKAT for which equivalence seems challenging to decide. We survey related work in Section 8 and conclude in Section 9. 2 OVERVIEW This section introduces the syntax and semantics of ProbNetKAT using a simple example. We will also see how various properties, including program equivalence and also program ordering and quantitative computations over the output distribution, can be encoded in ProbNetKAT. Each of the analyses in this section can be automatically carried out in our prototype implementation. As our running example, consider the network shown in Figure 1. It connects Source and Destination hosts through a topology with three switches. Suppose we want to implement the following policy: forward packets from the Source to the Destination. We will start by building a straightforward implementation of this policy in ProbNetKAT and then verify that it correctly implements the specification embodied in the policy using program equivalence. Next, we will refine our implementation to improve its resilience to link failures and verify that the refinement is more resilient. Finally, we characterize the resilience of both implementations quantitatively. 2.1 Deterministic Programming and Reasoning We will start with a simple deterministic program that forwards packets from left to right through the topology. To a first approximation, a ProbNetKAT program can be thought of as a random function from input packets to output packets. We model packets as records, with fields for standard headers such as the source address (src) and destination address (dst) of a packet, as well as two fields switch (sw) and port (pt) identifying the current location of the packet. The precise field names and ranges turns out to be not so important for our purposes; what is crucial is that the number of fields and the size of their domains must be finite. NetKAT provides primitives f ←n and f =n to modify and test the field f of an incoming packet. A modification f ←n returns the input packet with the f field updated to n. A test f =n either 4 S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva Switch 3 1 2 3 1 3 2 1 Switch 1 Source 2 Switch 2 Destination Fig. 1. Example network. returns the input packet unmodified if the test succeeds, or returns the empty set if the test fails. There are also primitives skip and drop that behave like a test that always succeeds and fails, respectively. Programs p, q are assembled to larger programs by composing them in sequence (p ; q) or in parallel (p & q). NetKAT also provides the Kleene star operator p ∗ from regular expressions to iterate programs. ProbNetKAT extends NetKAT with an additional operator p ⊕r q that executes either p with probability r , or q with probability 1 − r . Forwarding. We now turn to the implementation of our forwarding policy. To route packets from Source to Destination, all switches can simply forward incoming packets out of port 2: p1 ≜ pt←2 p2 ≜ pt←2 p3 ≜ pt←2 This is achieved by modifying the port field (pt). Then, to encode the forwarding logic for all switches into a single program, we take the union of their individual programs, after guarding the policy for each switch with a test that matches packets at that switch: p ≜ (sw=1 ; p1 ) & (sw=2 ; p2 ) & (sw =3 ; p3 ) Note that we specify a policy for switch 3, even though it is unreachable. Now we would like to answer the following question: does our program p correctly forward packets from Source to Destination? Note however that we cannot answer the question by inspecting p alone, since the answer depends fundamentally on the network topology. Topology. Although the network topology is not programmable, we can still model its behavior as a program. A unidirectional link matches on packets located at the source location of the link, and updates their location to the destination of the link. In our example network (Figure 1), the link ℓi j from switch i to switch j , i is given by ℓi j ≜ sw=i ; pt=j ; sw←j ; pt←i We obtain a model for the entire topology by taking the union of all its links: t ≜ ℓ12 & ℓ13 & ℓ32 Although this example uses unidirectional links, bidirectional links can be modeled as well using a pair of unidirectional links. Network Model. A packet traversing the network is subject to an interleaving of processing steps by switches and links in the network. This is expressible in NetKAT using Kleene star as follows: M(p, t) ≜ (p ; t)∗ ; p However, the model M(p, t) captures the behavior of the network on arbitrary input packets, including packets that start at arbitrary locations in the interior of the network. Typically we are interested only in the behavior of the network for packets that originate at the ingress of the Probabilistic Program Equivalence for NetKAT 5 network and arrive at the egress of the network. To restrict the model to such packets, we can define predicates in and out and pre- and post-compose the model with them: in ; M(p, t) ; out For our example network, we are interested in packets originating at the Source and arriving at the Destination, so we define in ≜ sw=1 ; pt=1 out ≜ sw=2 ; pt=2 With a full network model in hand, we can verify that p correctly implements the desired network policy, i.e. forward packets from Source to Destination. Our informal policy can be expressed formally as a simple ProbNetKAT program: teleport ≜ sw←2 ; pt←2 We can then settle the correctness question by checking the equivalence in ; M(p, t) ; out ≡ in ; teleport Previous work [3, 9, 30] used NetKAT equivalence with similar encodings to reason about various essential network properties including waypointing, reachability, isolation, and loop freedom, as well as for the validation and verification of compiler transformations. Unfortunately, the NetKAT decision procedure [9] and other state of the art network verification tools [15, 17] are fundamentally limited to reasoning about deterministic network behaviors. 2.2 Probabilistic Programming and Reasoning Routing schemes used in practice often behave non-deterministically—e.g., they may distribute packets across multiple paths to avoid congestion, or they may switch to backup paths in reaction to failures. To see these sorts of behaviors in action, let’s refine our naive routing scheme p to make it resilient to random link failures. Link Failures. We will assume that switches have access to a boolean flag upi that is true if and only if the link connected to the switch at port i is transmitting packets correctly.1 To make the network resilient to a failure, we can modify the program for Switch 1 as follows: if the link ℓ12 is up, use the shortest path to Switch 2 as before; otherwise, take a detour via Switch 3, which still forwards all packets to Switch 2. pb1 ≜ (up2 =1 ; pt←2) & (up2 =0 ; pt←3) As before, we can then encode the forwarding logic for all switches into a single program: pb ≜ (sw=1 ; pb1 ) & (sw=2 ; p2 ) & (sw=3 ; p3 ) Next, we update our link and topology encodings. A link behaves as before when it is up, but drops all incoming packets otherwise: ℓbi j ≜ upj =1 ; ℓi j For the purposes of this example, we will consider failures of links connected to Switch 1 only: b t ≜ ℓb12 & ℓb13 & ℓ32 1 Modern switches use low-level protocols such as Bidirectional Forwarding Detection (BFD) to maintain healthiness information about the link connected to each port [4]. 6 S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva We also need to assume some failure model, i.e. a probabilistic model of when and how often links fail. We will consider three failure models: f 0 ≜ up2 ←1 ; up3 ←1  Ê 1 1 1 f1 ≜ f 0 @ , (up2 ←0) & (up3 ←1) @ , (up2 ←1) & (up3 ←0) @ 2 4 4 f 2 ≜ (up2 ←1 ⊕0.8 up2 ←0) ; (up2 ←1 ⊕0.8 up2 ←0) Intuitively, in model f 0 , links never fail; in f 1 , the links ℓ12 and ℓ13 can fail with probability 25% each, but at most one fails; in f 2 , the links can fail independently with probability 20% each. Finally, we can assemble the encodings of policy, topology, and failures into a refined model: b t, f ) ≜ var up2 ←1 in M(p, var up3 ←1 in M((f ; p), t) b wraps our previous model M with declarations of the two local variables up2 The refined model M and up3 , and it executes the failure model at each hop prior to switch and topology processing. As a quick sanity check, we can verify that the old model and the new model are equivalent in the absence of failures, i.e. under failure model f 0 : b b M(p, t) ≡ M(p, t , f0 ) Now let us analyze our resilient routing scheme pb. First, we can verify that it correctly routes packets to the Destination in the absence of failures by checking the following equivalence: b p, b in ; M(b t , f 0 ) ; out ≡ in ; teleport In fact, the scheme pb is 1-resilient: it delivers all packets as long as no more than 1 link fails. In particular, it behaves like teleport under failure model f 1 . In contrast, this is not true for our naive routing scheme p: b p, b b b in ; M(b t , f 1 ) ; out ≡ in ; teleport . in ; M(p, t , f 1 ) ; out Under failure model f 2 , neither of the routing schemes is fully resilient and equivalent to teleportation. However, it is reassuring to verify that the refined routing scheme pb performs strictly better than the naive scheme p, b b b p, b M(p, t , f 2 ) < M(b t , f2 ) where p < q means that q delivers packets with higher probability than p. Reasoning using program equivalences and inequivalences is helpful to establish qualitative properties such as reachability properties and program invariants. But we can also go a step further, and compute quantitative properties of the packet distribution generated by a ProbNetKAT program. For example, we may ask for the probability that the schemes deliver a packet originating at Source to Destination under failure model f 2 . The answer is 80% for the naive scheme, and 96% for the resilient scheme. Such a computation might be used by an Internet Service Provider (ISP) to check that it can meet its service-level agreements (SLA) with customers. In Section 6 we will analyze a more sophisticated resilient routing scheme and see more complex examples of qualitative and quantitative reasoning with ProbNetKAT drawn from real-world data center networks. But first, we turn to developing the theoretical foundations (Sections 3 to 5). Probabilistic Program Equivalence for NetKAT 3 7 BACKGROUND ON PROBABILISTIC NETKAT In this section, we review the syntax and semantics of ProbNetKAT [8, 31] and basic properties of the language, focusing on the history-free fragment. A synopsis appears in Figure 2. 3.1 Syntax A packet π is a record mapping a finite set of fields f1 , f2 , . . . , fk to bounded integers n. Fields include standard header fields such as source (src) and destination (dst) addresses, as well as two logical fields for the switch (sw) and port (pt) that record the current location of the packet in the network. The logical fields are not present in a physical network packet, but it is convenient to model them as if they were. We write π .f to denote the value of field f of π and π [f :=n] for the packet obtained from π by updating field f to n. We let Pk denote the (finite) set of all packets. ProbNetKAT expressions consist of predicates (t, u, . . .) and programs (p, q, . . .). Primitive predicates include tests (f =n) and the Boolean constants false (drop) and true (skip). Compound predicates are formed using the usual Boolean connectives of disjunction (t & u), conjunction (t ; u), and negation (¬t). Primitive programs include predicates (t) and assignments (f ←n). Compound programs are formed using the operators parallel composition (p & q), sequential composition (p ; q), iteration (p ∗ ), and probabilistic choice (p ⊕r q). The full version of the language also provides a dup primitives, which logs the current state of the packet, but we omit this operator from the history-free fragment of the language considered in this paper; we discuss technical challenges to handling full ProbNetKAT in Section 7. The probabilistic choice operator p ⊕r q executes p with probability r and q with probability 1 −r , where r is rational, 0 ≤ r ≤ 1. We often use an n-ary version and omit the r ’s as in p1 ⊕ · · · ⊕ pn , which is interpreted as executing one of the pi chosen with equal probability. This can be desugared into the binary version. Conjunction of predicates and sequential composition of programs use the same syntax (t ; u and p ; q, respectively), as their semantics coincide. The same is true for disjunction of predicates and parallel composition of programs (t & u and p & q, respectively). The negation operator (¬) may only be applied to predicates. The language as presented in Figure 2 only includes core primitives, but many other useful constructs can be derived. In particular, it is straightforward to encode conditionals and while loops: if t then p else q ≜ t ; p & ¬t ; q while t do p ≜ (t ; p)∗ ; ¬t These encodings are well known from KAT [19]. Mutable and immutable local variables can also be desugared into the core calculus (although our implementation supports them directly): var f ←n in p ≜ f ←n ; p ; f ←0 Here f is an otherwise unused field. The assignment f ←0 ensures that the final value of f is “erased” after the field goes out of scope. 3.2 Semantics In the full version of ProbNetKAT, the space 2H of sets of packet histories2 is uncountable, and programs can generate continuous distributions on this space. This requires measure theory and Lebesgue integration for a suitable semantic treatment. However, as programs in our history-free fragment can generate only finite discrete distributions, we are able to give a considerably simplified presentation (Figure 2). Nevertheless, the resulting semantics is a direct restriction of the general semantics originally presented in [8, 31]. 2A history is a non-empty finite sequence of packets modeling the trajectory of a single packet through the network. 8 S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva Semantics Syntax Naturals n ::= 0 | 1 | 2 | · · · Fields f ::= f1 | · · · | fk Packets Pk ∋ π ::= {f1 = n 1 , . . . , fk = nk } Probabilities r ∈ [0, 1] ∩ Q Predicates Programs JpK ∈ 2Pk → D(2Pk ) JdropK(a) ≜ δ (∅) JskipK(a) ≜ δ (a) Jf =nK(a) ≜ δ ({π ∈ a | π .f = n}) Jf ←nK(a) ≜ δ ({π [f :=n] | π ∈ a}) t, u ::= drop | skip | f =n | t &u | t ;u | ¬t False True Test Disjunction Conjunction Negation p, q ::= t | f ←n | p &q | p ;q | p ⊕r q | p∗ n ∈N Filter (0) ≜ skip, p (n+1) ≜ skip & p ; p (n) where p Assignment Union (Discrete) Probability Monad D Sequence Unit δ : X → D(X ) δ (x) ≜ δ x Choice Iteration Bind −† : (X → D(Y )) → D(X ) → D(Y ) Í f † (µ)(A) ≜ x ∈X f (x)(A) · µ(x) J¬tK(a) ≜ D(λb.a − b)(JtK(a)) Jp & qK(a) ≜ D(∪)(JpK(a) × JqK(a)) Jp ; qK(a) ≜ JqK† (JpK(a)) Jp ⊕r qK(a) ≜ rÄ · JpK(a) + (1 − r ) · JqK(a) Jp ∗ K(a) ≜ Jp (n) K(a) Fig. 2. ProbNetKAT core language: syntax and semantics. Proposition 3.1. Let L−M denote the semantics defined in [31]. Then for all dup-free programs p and inputs a ∈ 2Pk , we have JpK(a) = LpM(a), where we identify packets and histories of length one. Proof. The proof is given in Appendix A. □ For the purposes of this paper, we work in the discrete space 2Pk , i.e., the set of sets of packets. An outcome (denoted by lowercase variables a, b, c, . . . ) is a set of packets and an event (denoted by uppercase variables A, B, C, . . . ) is a set of outcomes. Given a discrete probability measure on this space, the probability of an event is the sum of the probabilities of its outcomes. Programs are interpreted as Markov kernels on the space 2Pk . A Markov kernel is a function 2Pk → D(2Pk ) in the probability (or Giry) monad D [11, 18]. Thus, a program p maps an input set of packets a ∈ 2Pk to a distribution JpK(a) ∈ D(2Pk ) over output sets of packets. The semantics uses the following probabilistic primitives:3 • For a discrete measurable space X , D(X ) denotes the set of probability measures over X ; that is, the set of countably additive functions µ : 2X → [0, 1] with µ(X ) = 1. • For a measurable function f : X → Y , D(f ) : D(X ) → D(Y ) denotes the pushforward along f ; that is, the function that maps a measure µ on X to D(f )(µ) ≜ µ ◦ f −1 = λA ∈ ΣY . µ({x ∈ X | f (x) ∈ A}) which is called the pushforward measure on Y . • The unit δ : X → D(X ) of the monad maps a point x ∈ X to the point mass (or Dirac measure) δ x ∈ D(X ). The Dirac measure is given by δ x (A) ≜ 1[x ∈ A] 3 The same primitives can be defined for uncountable spaces, as would be required to handle the full language. Probabilistic Program Equivalence for NetKAT 9 That is, the Dirac measure is 1 if x ∈ A and 0 otherwise. • The bind operation of the monad, −† : (X → D(Y )) → D(X ) → D(Y ) lifts a function f : X → D(Y ) with deterministic inputs to a function f † : D(X ) → D(Y ) that takes random inputs. Intuitively, this is achieved by averaging the output of f when the inputs are randomly distributed according to µ. Formally, Õ f † (µ)(A) ≜ f (x)(A) · µ(x). x ∈X • Given two measures µ ∈ D(X ) and ν ∈ D(Y ), µ × ν ∈ D(X × Y ) denotes their product measure. This is the unique measure satisfying: (µ × ν)(A × B) = µ(A) · ν (B) Intuitively, it models distributions over pairs of independent values. With these primitives at our disposal, we can now make our operational intuitions precise. Formal definitions are given in Figure 2. A predicate t maps (with probability 1) the set of input packets a ∈ 2Pk to the subset of packets b ⊆ a satisfying the predicate. In particular, the false primitive drop simply drops all packets (i.e., it returns the empty set with probability 1) and the true primitive skip simply keeps all packets (i.e., it returns the input set with probability 1). The test f =n returns the subset of input packets whose f -field contains n. Negation ¬t filters out the packets returned by t. Parallel composition p & q executes p and q independently on the input set, then returns the union of their results. Note that packet sets do not model nondeterminism, unlike the usual situation in Kleene algebras—rather, they model collections of packets traversing possibly different portions of the network simultaneously. Probabilistic choice p ⊕r q feeds the input to both p and q and returns a convex combination of the output distributions according to r . Sequential composition p ; q can be thought of as a two-stage probabilistic experiment: it first executes p on the input set to obtain a random intermediate result, then feeds that into q to obtain the final distribution over outputs. The outcome of q needs to be averaged over the distribution of intermediate results produced by p. It may be helpful to think about summing over the paths in a probabilistic tree diagram and multiplying the probabilities along each path. We say that two programs are equivalent, denoted p ≡ q, if they denote the same Markov kernel, i.e. if JpK = JqK. As usual, we expect Kleene star p ∗ to satisfy the characteristic fixed point equation p ∗ ≡ skip & p ; p ∗ , which allows it to be unrolled ad infinitum. Thus we define it as the supremum of its finite unrollings p (n) ; see Figure 2. This supremum is taken in a CPO (D(2Pk ), ⊑) of distributions that is described in more detail in § 3.3. The partial ordering ⊑ on packet set distributions gives rise to a partial ordering on programs: we write p ≤ q iff JpK(a) ⊑ JqK(a) for all inputs a ∈ 2Pk . Intuitively, p ≤ q iff p produces any particular output packet π with probability at most that of q for any fixed input. A fact that should be intuitively clear, although it is somewhat hidden in our presentation of the denotational semantics, is that the predicates form a Boolean algebra: Lemma 3.2. Every predicate t satisfies JtK(a) = δ a∩bt for a certain packet set bt ⊆ Pk, where • bdrop = ∅, • bskip = Pk, • bf =n = {π ∈ Pk | π .f = n}, 10 S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva • b¬t = Pk − bt , • bt &u = bt ∪ bu , and • bt ;u = bt ∩ bu . Proof. For drop, skip, and f =n, the claim holds trivially. For ¬t, t & u, and t ; u, the claim follows inductively, using that D(f )(δb ) = δ f (b) , δb × δc = δ (b,c) , and that f † (δb ) = f (b). The first and last equations hold because ⟨D, δ, −† ⟩ is a monad. □ 3.3 The CPO (D(2Pk ), ⊑) The space 2Pk with the subset order forms a CPO (2Pk , ⊆). Following Saheb-Djahromi [27], this CPO can be lifted to a CPO (D(2Pk ), ⊑) on distributions over 2Pk . Because 2Pk is a finite space, the resulting ordering ⊑ on distributions takes a particularly easy form: µ ⊑ν ⇐⇒ µ({a}↑) ≤ ν ({a}↑) for all a ⊆ Pk where {a}↑ ≜ {b | a ⊆ b} denotes upward closure. Intuitively, ν produces more outputs then µ. As was shown in [31], ProbNetKAT satisfies various monotonicity (and continuity) properties with respect to this ordering, including a ⊆ a ′ =⇒ JpK(a) ⊑ JpK(a ′) and n ≤ m =⇒ Jp (n) K(a) ⊑ Jp (m) K(a). As a result, the semantics of p ∗ as the supremum of its finite unrollings p (n) is well-defined. While the semantics of full ProbNetKAT requires domain theory to give a satisfactory characterization of Kleene star, a simpler characterization suffices for the history-free fragment: Lemma 3.3 (Pointwise Convergence). Let A ⊆ 2Pk . Then for all programs p and inputs a ∈ 2Pk , Jp ∗ K(a)(A) = lim Jp (n) K(a)(A). n→∞ Proof. See Appendix A □ This lemma crucially relies on our restrictions to dup-free programs and the space 2Pk . With this insight, we can now move to a concrete semantics based on Markov chains, enabling effective computation of program semantics. 4 BIG-STEP SEMANTICS The Scott-style denotational semantics of ProbNetKAT interprets programs as Markov kernels 2Pk → D(2Pk ). Iteration is characterized in terms of approximations in a CPO (D(2Pk ), ⊑) of distributions. In this section we relate this semantics to a Markov chain semantics on a state space consisting of finitely many packets. Since the set of packets Pk is finite, so is its powerset 2Pk . Thus any distribution over packet sets is discrete and can be characterized by a probability mass function, i.e. a function Õ f : 2Pk → [0, 1], f (b) = 1 b ⊆Pk It is convenient to view f as a stochastic vector, i.e. a vector of non-negative entries that sums to 1. The vector is indexed by packet sets b ⊆ Pk with b-th component f (b). A program, being a function that maps inputs a to distributions over outputs, can then be organized as a square matrix indexed by Pk in which the stochastic vector corresponding to input a appears as the a-th row. Pk Pk Thus we can interpret a program p as a matrix BJpK ∈ [0, 1]2 ×2 indexed by packet sets, where the matrix entry BJpKab denotes the probability that program p produces output b ∈ 2Pk on input a ∈ 2Pk . The rows of BJpK are stochastic vectors, each encoding the output distribution Probabilistic Program Equivalence for NetKAT 11 BJpK ∈ S(2Pk ) BJdropKab ≜ 1[b = ∅] BJp & qKab ≜ Õ 1[c ∪ d = b] · BJpKa,c · BJqKa,d c,d BJskipKab ≜ 1[a = b] BJf =nKab ≜ 1[b = {π ∈ a | π .f = n}] BJ¬tKab ≜ 1[b ⊆ a] · BJtKa,a−b BJf ←nKab ≜ 1[b = {π [f := n] | π ∈ a}] BJp ; qK ≜ BJpK · BJqK BJp ⊕r qK ≜ r · BJpK + (1 − r ) · BJqK BJp ∗ Kab ≜ lim BJp (n) Kab n→∞ Fig. 3. Big-Step Semantics: BJpKab denotes the probability that program p produces output b on input a. corresponding to a particular input set a. Such a matrix is called (right-)stochastic. We denote by S(2Pk ) the set of right-stochastic square matrices indexed by 2Pk . The interpretation of programs as stochastic matrices is largely straightforward and given formally in Figure 3. At a high level, deterministic program primitives map to simple (0, 1)-matrices, and program operators map to operations on matrices. For example, the program primitive drop is interpreted as the stochastic matrix ∅ b2 ... bn 1  ..  .. BJdropK = .  .  an  1  ∅ 0 · · · 0  .. . . ..  . .  .  0 · · · 0  a2 1 .. . an a1 = ∅ 1 (1) 1 that moves all probability mass to the ∅-column, and the primitive skip is the identity matrix. The formal definitions are given in Figure 3 using Iverson brackets: 1[φ] is defined to be 1 if φ is true, or 0 otherwise. As suggested by the picture in (1), a stochastic matrix B ∈ S(2Pk ) can be viewed as a Markov chain (MC), a probabilistic transition system with state space 2Pk that makes a random transition between states at each time step. The matrix entry Bab gives the probability that, whenever the system is in state a, it transitions to state b in the next time step. Under this interpretation, sequential composition becomes matrix product: a step from a to b in BJp ; qK decomposes into a step from a to some intermediate state c in BJpK and a step from c to the final state b in BJqK with probability Õ BJp ; qKab = BJpKac · BJqKcb = (BJpK · BJqK)ab . c 4.1 Soundness The main theoretical result of this section is that the finite matrix BJpK fully characterizes the behavior of a program p on packets. Theorem 4.1 (Soundness). For any program p and any sets a, b ∈ 2Pk , BJp ∗ K is well-defined, BJpK is a stochastic matrix, and BJpKab = JpK(a)({b}). Proof. It suffices to show the equality BJpKab = JpK(a)({b}); the remaining claims then follow by well-definedness of J−K. The equality is shown using Lemma 3.3 and a routine induction on p: For p = drop, skip, f =n, f ←n we have JpK(a)({b}) = δc ({b}) = 1[b = c] = BJpKab for c = ∅, a, {π ∈ a | π .f = n}, {π [f := n] | π ∈ a}, respectively. 12 S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva For ¬t we have, BJ¬tKab = 1[b ⊆ a] · BJtKa,a−b = 1[b ⊆ a] · JtK(a)({a − b}) = 1[b ⊆ a] · 1[a − b = a ∩ bt ] = 1[b ⊆ a] · 1[a − b = a − (H − bt )] = 1[b = a ∩ (H − bt )] = J¬tK(a)(b) (IH) (Lemma 3.2) (Lemma 3.2) For p & q, letting µ = JpK(a) and ν = JqK(a) we have Jp & qK(a)({b}) = (µ × ν )({(b1 , b2 ) | b1 ∪ b2 = b}) Í = b1,b2 1[b1 ∪ b2 = b] · (µ × ν )({(b1 , b2 )}) Í = Íb1,b2 1[b1 ∪ b2 = b] · µ({b1 }) · ν ({b2 }) = b1,b2 1[b1 ∪ b2 = b] · BJpKab1 · BJqKab2 = BJp & qKab (IH) where we use in the second step that b ⊆ Pk is finite, thus {(b1 , b2 ) | b1 ∪ b2 = b} is finite. For p ; q, let µ = JpK(a) and νc = JqK(c) and recall that µ is a discrete distribution on 2Pk . Thus Í Jp ; qK(a)({b}) = Íc ∈2Pk νc ({b}) · µ({c}) = c ∈2Pk BJqKc,b · BJpKa,c = BJp ; qKab . For p ⊕r q, the claim follows directly from the induction hypotheses. Finally, for p ∗ , we know that BJp (n) Kab = Jp (n) K(a)({b}) by induction hypothesis. The key to proving the claim is Lemma 3.3, which allows us to take the limit on both sides and deduce BJp ∗ Kab = lim BJp (n) Kab = lim Jp (n) K(a)({b}) = Jp ∗ K(a)({b}). n→∞ n→∞ □ Together, these results reduce the problem of checking program equivalence for p and q to checking equality of the matrices produced by the big-step semantics, BJpK and BJqK. Corollary 4.2. For programs p and q, JpK = JqK if and only if BJpK = BJqK. Proof. Follows directly from Theorem 4.1. □ Unfortunately, BJp ∗ K is defined in terms of a limit. Thus, it is not obvious how to compute the big-step matrix in general. The next section is concerned with finding a closed form for the limit, resulting in a representation that can be effectively computed, as well as a decision procedure. 5 SMALL-STEP SEMANTICS This section derives a closed form for BJp ∗ K, allowing to compute BJ−K explicitly. This yields an effective mechanism for checking program equivalence on packets. In the “big-step” semantics for ProbNetKAT, programs are interpreted as Markov chains over the state space 2Pk , such that a single step of the chain models the entire execution of a program, going directly from some initial state a (corresponding to the set of input packets) to the final state b (corresponding to the set of output packets). Here we will instead take a “small-step” approach and design a Markov chain such that one transition models one iteration of p ∗ . To a first approximation, the states (or configurations) of our probabilistic transition system are triples ⟨p, a, b⟩, consisting of the program p we mean to execute, the current set of (input) packets a, and an accumulator set b of packets output so far. The execution of p ∗ on input a ⊆ Pk starts from the initial state ⟨p ∗ , a, ∅⟩. It proceeds by unrolling p ∗ according to the characteristic equation Probabilistic Program Equivalence for NetKAT 1 ⟨p ∗ , a, b⟩ 13 ⟨skip & p ; p ∗ , a, b⟩ 1 ⟨p ; p ∗ , a, b ∪ a⟩ B Jp K a, a ′ BJpKa,a ′ ⟨p ∗ , a ′, b ∪ a⟩ Fig. 4. The small-step semantics is given by a Markov chain whose states are configurations of the form ⟨program, input set, output accumulator⟩. The three dashed arrows can be collapsed into the single solid arrow, rendering the program component superfluous. p ∗ ≡ skip & p ; p ∗ with probability 1: 1 ⟨p ∗ , a, ∅⟩ −−−−−−−−−→ ⟨skip & p ; p ∗ , a, ∅⟩ To execute a union of programs, we must execute both programs on the input set and take the union of their results. In the case of skip & p ; p ∗ , we can immediately execute skip by outputting the input set with probability 1, leaving the right hand side of the union: 1 ⟨skip & p ; p ∗ , a, ∅⟩ −−−−−−−−−→ ⟨p ; p ∗ , a, a⟩ To execute the sequence p ; p ∗ , we first execute p and then feed its (random) output into p ∗ : ∀a ′ : BJpKa, a ′ ⟨p ; p ∗ , a, a⟩ −−−−−−−−−→ ⟨p ∗ , a ′, a⟩ At this point the cycle closes and we are back to executing p ∗ , albeit with a different input set a ′ and some accumulated outputs. The structure of the resulting Markov chain is shown in Figure 4. At this point we notice that the first two steps of execution are deterministic, and so we can collapse all three steps into a single one, as illustrated in Figure 4. After this simplification, the program component of the states is rendered obsolete since it remains constant across transitions. Thus we can eliminate it, resulting in a Markov chain over the state space 2Pk × 2Pk . Formally, it can be defined concisely as SJpK ∈ S(2Pk × 2Pk ) SJpK(a,b),(a ′,b ′ ) ≜ 1[b ′ = b ∪ a] · BJpKa,a ′ As a first sanity check, we verify that the matrix SJpK defines indeed a Markov chain: Lemma 5.1. SJpK is stochastic. Proof. For arbitrary a, b ⊆ Pk, we have Õ Õ SJpK(a,b),(a ′,b ′ ) = 1[b ′ = a ∪ b] · BJpKa,a ′ a ′,b ′ a ′,b ′ = Õ Õ = Õ a′  1[b ′ = a ∪ b] · BJpKa,a ′ b′ BJpKa,a ′ = 1 a′ where, in the last step, we use that BJpK is stochastic (Theorem 4.1). □ Next, we show that steps in SJpK indeed model iterations of p ∗ . Formally, the (n + 1)-step of SJpK is equivalent to the big-step behavior of the n-th unrolling of p ∗ in the following sense: 14 S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva Proposition 5.2. BJp (n) Ka,b = Í a′ SJpKn+1 (a,∅),(a ′,b) Proof. Naive induction on the number of steps n ≥ 0 fails, because the hypothesis is too weak. We must first generalize it to apply to arbitrary start states in SJpK, not only those with empty accumulator. The appropriate generalization of the claim turns out to be: Lemma 5.3. Let p be program. Then for all n ∈ N and a, b, b ′ ⊆ Pk, Õ Õ 1[b ′ = a ′ ∪ b] · BJp (n) Ka,a ′ = SJpKn+1 (a,b),(a ′,b ′ ) a′ Proof. a′ By induction on n ≥ 0. For n = 0, we have Õ Õ 1[b ′ = a ′ ∪ b] · BJp (n) Ka,a ′ = 1[b ′ = a ′ ∪ b] · BJskipKa,a ′ a′ a′ = Õ 1[b ′ = a ′ ∪ b] · 1[a = a ′] a′ = 1[b ′ = a ∪ b] = 1[b ′ = a ∪ b] · Õ BJpKa,a ′ a′ = Õ SJpK(a,b),(a ′,b ′ ) a′ In the induction step (n > 0), Õ 1[b ′ = a ′ ∪ b] · BJp (n) Ka,a ′ a′ = Õ = Õ a′ 1[b ′ = a ′ ∪ b] · BJskip & p ; p (n−1) Ka,a ′ 1[b ′ = a ′ ∪ b] · Õ c a′ 1[a ′ = a ∪ c] · BJp ; p (n−1) Ka,c ! = Õ Õ = Õ c 1[b = a ∪ b] · 1[a = a ∪ c] · ′ ′ ′ Õ a′ k 1[b = a ∪ c ∪ b] · BJpKa,k · BJp ′ (n−1) c,k = Õ BJpKa,k · a′ k = Õ BJpKa,k · ÕÕ a′ = = Õ a′ Kk,c 1[b ′ = a ′ ∪ (a ∪ b)] · BJp (n−1) Kk,a ′ SJpKn(k,a∪b),(a ′,b ′ ) 1[k 2 = a ∪ b] · BJpKa,k1 · SJpKn(k1,k2 ),(a ′,b ′ ) k 1,k 2 ÕÕ a′ Õ a′ k = Õ BJpKa,k · BJp (n−1) Kk,c SJpK(a,b)(k1,k2 ) · SJpKn(k1,k2 ),(a ′,b ′ ) k 1,k 2 SJpKn+1 (a,b),(a ′,b ′ ) Proposition 5.2 then follows by instantiating Lemma 5.3 with b = ∅. □ □ Probabilistic Program Equivalence for NetKAT 5.1 15 Closed form Let (an , bn ) denote the random state of the Markov chain SJpK after taking n steps starting from (a, ∅). We are interested in the distribution of bn for n → ∞, since this is exactly the distribution of outputs generated by p ∗ on input a (by Proposition 5.2 and the definition of BJp ∗ K). Intuitively, the ∞-step behavior of SJpK is equivalent to the big-step behavior of p ∗ . The limiting behavior of finite state Markov chains has been well-studied in the literature (e.g., see [16]), and we can exploit these results to obtain a closed form by massaging SJpK into a so called absorbing Markov chain. A state s of a Markov chain T is called absorbing if it transitions to itself with probability 1: s (formally: Ts,s ′ = 1[s = s ′]) 1 A Markov chain T ∈ S(S) is called absorbing if each state can reach an absorbing state: n ∀s ∈ S. ∃s ′ ∈ S, n ≥ 0. Ts,s ′ > 0 and Ts ′,s ′ = 1 The non-absorbing states of an absorbing MC are called transient. Assume T is absorbing with nt transient states and na absorbing states. After reordering the states so that absorbing states appear before transient states, T has the form   I 0 T = R Q where I is the na × na identity matrix, R is an nt × na matrix giving the probabilities of transient states transitioning to absorbing states, and Q is an nt ×nt square matrix specifying the probabilities of transient states transitioning to transient states. Absorbing states never transition to transient states, thus the na × nt zero matrix in the upper right corner. No matter the start state, a finite state absorbing MC always ends up in an absorbing state eventually, i.e. the limit T ∞ ≜ limn→∞ T n exists and has the form   I 0 ∞ T = A 0 for an nt × na matrix A of so called absorption probabilities, which can be given in closed form: A = (I + Q + Q 2 + . . . ) R That is, to transition from a transient state to an absorbing state, the MC can first take an arbitrary number of steps between transient states, before taking a single and final step into an absorbing Í state. The infinite sum X ≜ n ≥0 Q n satisfies X = I + QX , and solving for X we get X = (I − Q)−1 and A = (I − Q)−1 R (2) (We refer the reader to [16] or Lemma A.2 in Appendix A for the proof that the inverse must exist.) Before we apply this theory to the small-step semantics SJ−K, it will be useful to introduce some T MC-specific notation. Let T be an MC. We write s − →n s ′ if s can reach s ′ in precisely n steps, i.e. T n > 0; and we write s − n > 0 for any if Ts,s → s ′ if s can reach s ′ in any number of steps, i.e. if Ts,s ′ ′ T T T n ≥ 0. Two states are said to communicate, denoted s ← → s ′, if s − → s ′ and s ′ − → s. The relation T ← → is an equivalence relation, and its equivalence classes are called communication classes. A communication class is called absorbing if it cannot reach any states outside the class. We sometimes T n . For the rest of the section, we fix a program p and write Pr[s − →n s ′] to denote the probability Ts,s ′ abbreviate BJpK as B and SJpK as S. Of central importance are what we will call the saturated states of S: 16 S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva Definition 5.4. A state (a, b) of S is called saturated if the accumulator b has reached its final S value, i.e. if (a, b) → − (a ′, b ′) implies b ′ = b. Once we have reached a saturated state, the output of p ∗ is determined. The probability of ending up in a saturated state with accumulator b, starting from an initial state (a, ∅), is Õ n lim S (a,∅),(a ′,b) n→∞ a′ that p ∗ and indeed this is the probability outputs b on input a by Proposition 5.2. Unfortunately, a saturated state is not necessarily absorbing. To see this, assume there exists only a single field f ranging over {0, 1} and consider the program p ∗ = (f ←0 ⊕1/2 f ←1)∗ . Then S has the form 0, 0 0, {0, 1} 1, 0 1, {0, 1} 0, ∅ where all edges are implicitly labeled with 12 , 0 denotes the packet with f set to 0 and 1 denotes the packet with f set to 1, and we omit states not reachable from (0, ∅). The two right most states are saturated; but they communicate and are thus not absorbing. We can fix this by defining the auxiliary matrix U ∈ S(2Pk × 2Pk ) as ( 1[a ′ = ∅] if (a, b) is saturated ′ U(a,b),(a ′,b ′ ) ≜ 1[b = b] · 1[a ′ = a] else It sends a saturated state (a, b) to the canonical saturated state (∅, b), which is always absorbing; and it acts as the identity on all other states. In our example, the modified chain SU looks as follows: 0, {0, 1} 0, 0 ∅, {0, 1} 0, ∅ 1, {0, 1} 1, 0 To show that SU is always an absorbing MC, we first observe: S Lemma 5.5. S, U , and SU are monotone in the following sense: (a, b) → − (a ′, b ′) implies b ⊆ b ′ (and similarly for U and SU ). Proof. For S and U the claim follows directly from their definitions. For SU the claim then follows compositionally. □ Now we can show: Proposition 5.6. Let n ≥ 1. (1) (SU )n = S n U (2) SU is an absorbing MC with absorbing states {(∅, b) | b ⊆ Pk}. Proof. (1) It suffices to show that U SU = SU . Suppose that U SU Pr[(a, b) −−−−→1 (a ′, b ′)] = p > 0. Probabilistic Program Equivalence for NetKAT 17 It suffices to show that this implies SU Pr[(a, b) −−→1 (a ′, b ′)] = p. If (a, b) is saturated, then we must have (a ′, b ′) = (∅, b) and U SU SU Pr[(a, b) −−−−→1 (∅, b)] = 1 = Pr[(a, b) −−→1 (∅, b)] U If (a, b) is not saturated, then (a, b) −→1 (a, b) with probability 1 and therefore U SU SU Pr[(a, b) −−−−→1 (a ′, b ′)] = Pr[(a, b) −−→1 (a ′, b ′)] (2) Since S and U are stochastic, clearly SU is a MC. Since SU is finite state, any state can reach an SU absorbing communication class. (To see this, note that the reachability relation −−→ induces a partial order on the communication classes of SU . Its maximal elements are necessarily absorbing, and they must exist because the state space is finite.) It thus suffices to show that a state set C ⊆ 2Pk × 2Pk in SU is an absorbing communication class iff C = {(∅, b)} for some b ⊆ Pk. B S “⇐”: Observe that ∅ − →1 a ′ iff a ′ = ∅. Thus (∅, b) → − 1 (a ′, b ′) iff a ′ = ∅ and b ′ = b, and likewise U (∅, b) −→1 (a ′, b ′) iff a ′ = ∅ and b ′ = b. Thus (∅, b) is an absorbing state in SU as required. SU “⇒”: First observe that by monotonicity of SU (Lemma 5.5), we have b = b ′ whenever (a, b) ←→ (a ′, b ′); thus there exists a fixed bC such that (a, b) ∈ C implies b = bC . SU Now pick an arbitrary state (a, bC ) ∈ C. It suffices to show that (a, bC ) −−→ (∅, bC ), because SU that implies (a, bC ) ←→ (∅, bC ), which in turn implies a = ∅. But the choice of (a, bC ) ∈ C was arbitrary, so that would mean C = {(∅, bC )} as claimed. SU To show that (a, bC ) −−→ (∅, bC ), pick arbitrary states such that S U (a, bC ) → − (a ′, b ′) −→1 (a ′′, b ′′) SU SU and recall that this implies (a, bC ) −−→ (a ′′, b ′′) by claim (1). Then (a ′′, b ′′) −−→ (a, bC ) because C is absorbing, and thus bC = b ′ = b ′′ by monotonicity of S, U , and SU . But (a ′, b ′) was chosen as an arbitrary state S-reachable from (a, bC ), so (a, b) and by transitivity (a ′, b ′) must be saturated. Thus a ′′ = ∅ by the definition of U . □ Arranging the states (a, b) in lexicographically ascending order according to ⊆ and letting n = |2Pk |, it then follows from Proposition 5.6.2 that SU has the form   In 0 SU = R Q where for a , ∅  (SU )(a,b),(a ′,b ′ ) = R Moreover, SU converges and its limit is given by  In ∞ (SU ) ≜ (I − Q)−1R Q  (a,b),(a ′,b ′ )  0 = lim (SU )n 0 n→∞ (3) We can use the modified Markov chain SU to compute the limit of S: Theorem 5.7 (Closed Form). Let a, b, b ′ ⊆ Pk. Then Õ n ∞ lim S (a,b),(a ′,b ′ ) = (SU )(a,b),(∅,b ′ ) n→∞ a′ (4) 18 S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva or, using matrix notation, lim Õ n→∞ a′ n S (−,−),(a ′,−)  Pk Pk Pk In = ∈ [0, 1](2 ×2 )×2 (I − Q)−1 R  (5) In particular, the limit in (4) exists and it can be effectively computed in closed-form. Proof. Using Proposition 5.6.1 in the second step and equation (3) in the last step, Õ Õ n lim S (a,b),(a (S n U )(a,b),(a ′,b ′ ) ′,b ′ ) = lim n→∞ n→∞ a′ = lim n→∞ = a′ Õ a′ (SU )n(a,b),(a ′,b ′ ) Õ ∞ (SU )∞ (a,b),(a ′,b ′ ) = (SU )(a,b),(∅,b ′ ) a′ (SU )∞ is computable because S and U are matrices over Q and hence so is (I − Q)−1 R. □ Corollary 5.8. For programs p and q, it is decidable whether p ≡ q. Proof. Recall from Corollary 4.2 that it suffices to compute the finite rational matrices BJpK and BJqK and check them for equality. But Theorem 5.7 together with Proposition 5.2 gives us an effective mechanism to compute BJ−K in the case of Kleene star, and BJ−K is straightforward to compute in all other cases. To summarize, we repeat the full chain of equalities we have deduced: Õ Jp ∗ K(a)({b}) = BJp ∗ Ka,b = lim BJp (n) Ka,b = lim SJpKn(a,∅),(a ′,b) = (SU )∞ (a,∅),(∅,b) n→∞ n→∞ a′ (From left to right: Theorem 4.1, Definition of BJ−K, Proposition 5.2, and Theorem 5.7.) 6 □ CASE STUDY: RESILIENT ROUTING We have build a prototype based on Theorem 5.7 and Corollary 5.8 in OCaml. It implements ProbNetKAT as an embedded DSL and compiles ProbNetKAT programs to transition matrices using symbolic techniques and a sparse linear algebra solver. A detailed description and performance evaluation of the implementation is beyond the scope of this paper. Here we focus on demonstrating the utility of such a tool by performing a case study with real-world datacenter topologies and resilient routing schemes. Recently proposed datacenter designs [1, 13, 14, 21, 24, 29] utilize a large number of inexpensive commodity switches, which improves scalability and reduces cost compared to other approaches. However, relying on many commodity devices also increases the probability of failures. A recent measurement study showed that network failures in datacenters [10] can have a major impact on application-level performance, leading to a new line of work exploring the design of fault-tolerant datacenter fabrics. Typically the topology and routing scheme are co-designed, to achieve good resilience while still providing good performance in terms of throughput and latency. 6.1 Topology and routing Datacenter topologies typically organize the fabric into multiple levels of switches. FatTree. A FatTree [1], which is a multi-level, multi-rooted tree, is perhaps the most common example of such a topology. Figure 5 shows a 3-level FatTree topology with 20 switches. The bottom level, edge, consists of top-of-rack (ToR) switches; each ToR switch connects all the hosts within a rack (not shown in the figure). These switches act as ingress and egress for intra-datacenter traffic. Probabilistic Program Equivalence for NetKAT 19 Core C ✗ Aggregation A ✗ A′ A′′ Edge s1 s2 s3 s4 s5 s6 s7 s8 Fig. 5. A FatTree topology with 20 switches. s1 s2 s3 s4 s5 s6 s7 s8 Fig. 6. An AB FatTree topology with 20 switches. The other two levels, aggregation and core, redundantly interconnect the switches from the edge layer. The redundant structure of a FatTree naturally lends itself to forwarding schemes that locally route around failures. To illustrate, consider routing from a source (s7) to a destination (s1) along shortest paths in the example topology. Packets are first forwarded upwards, until eventually there exists a downward path to s1. The green links in the figure depict one such path. On the way up, there are multiple paths at each switch that can be used to forward traffic. Thus, we can route around failures by simply choosing an alternate upward link. A common routing scheme is called equal-cost multi-path routing (ECMP) in the literature, because it chooses between several paths all having the same cost—e.g., path length. ECMP is especially attractive as is it can provide better resilience without increasing the lengths of forwarding paths. However, after reaching a core switch, there is a unique shortest path to the destination, so ECMP no longer provides any resilience if a switch fails in the aggregation layer (cf. the red cross in Figure 5). A more sophisticated scheme could take a longer (5-hop) detour going all the way to another edge switch, as shown by the red lines in the figure. Unfortunately, such detours inflate the path length and lead to increased latency and congestion. AB FatTree. FatTree’s unpleasantly long backup routes on the downward paths are caused by the symmetric wiring of aggregation and core switches. AB FatTrees [21] alleviate this flaw by skewing the symmetry of wiring. It defines two types of subtrees, differing in their wiring to higher levels. To illustrate, Figure 6 shows an example which rewires the FatTree from Figure 5 to make it an AB FatTree. It contains two types of subtrees: i) Type A: switches depicted in blue and wired to core using dashed lines, and ii) Type B: switches depicted in red and wired to core using solid lines. Type A subtrees are wired in a way similar to FatTree, but type B subtrees differ in their connections to core switches (see the original paper for full details [21]). This slight change in wiring enables shorter detours to route around failures in the downward direction. Consider again a flow involving the same source (s7) and destination (s1). As before, we have multiple options going upwards when following shortest paths (e.g., the one depicted in green), but we have a unique downward path once we reach the top. But unlike FatTree, if the aggregation switch on the downward path fails, we find that there is a short (3-hop) detour, as shown in blue. This backup path exists because the core switch, which needs to reroute traffic, is connected to aggregation switches of both types of subtrees. More generally, aggregation switches of the same type as the failed switch provide a 5-hop detour (as in a standard FatTrees); but aggregation switches of the opposite type can provide a more efficient 3-hop detour. 20 S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva // F10 with 3-hop & 5-hop rerouting // F10 without rerouting f10_3_5 := f10_0 := if at_ingress then (default <- 1); // ECMP, but don’t use inport if default = 1 then ( fwd_on_random_shortest_path f10_3; if at_down_port then (5hop_rr; default <- 0) // F10 with 3-hop rerouting ) else ( f10_3 := default <- 1; // back to default forwarding f10_0; fwd_downward_uniformly_at_random if at_down_port then 3hop_rr ) Fig. 7. ProbNetKAT implementation of F10 in three refinement steps. 6.2 ProbNetKAT implementation. Now we will see how to encode several routing schemes using ProbNetKAT and analyze their behavior in each topology under various failure models. Routing. F10 [21] provides a routing algorithm that combines the three routing and rerouting strategies we just discussed (ECMP, 3-hop rerouting, 5-hop rerouting) into a single scheme. We implemented it in three steps (see Figure 7). The first scheme, F100 , implements an approach similar to ECMP:4 it chooses a port uniformly at random from the set of ports connected to minimumlength paths to the destination. We exclude the port at which the packet arrived from this set; this eliminates the possibility of forwarding loops when routing around failures. Next, we improve the resilience of F100 by augmenting it with 3-hop rerouting if the next hop aggregation switch A along the downward shortest path from a core switch C fails. To illustrate, consider the blue path in Figure 6. We find a port on C that connects to an aggregation switch A′ of the opposite type than the failed aggregation switch, A, and forward the packet to A′. If there are multiple such ports that have not failed, we choose one uniformly at random. Normal routing continues at A′, and ECMP will know not to send the packet back to C. F103 implements this refinement. Note that if the packet is still parked at port whose adjacent link is down after executing F103 , it must be that all ports connecting to aggregation switches of the opposite type are down. In this case, we attempt 5-hop rerouting via an aggregation switch A′′ of the same type as A. To illustrate, consider the red path in Figure 6. We begin by sending the packet to A′′. To let A′′ know that it should not send the packet back up as normally, we set a flag default to false in the packet, telling A′′ to send the packet further down instead. From there, default routing continues. F103,5 implements this refinement. p Failure and Network model. We define a family of failure models fk in the style of Section 2. Let k ∈ N∪{∞} denote a bound on the maximum number of link failures that may occur simultaneously, and assume that links otherwise fail independently with probability 0 ≤ p < 1 each. We omit p when it is clear from context. For simplicity, to focus on the more complicated scenarios occurring on downward paths, we will model failures only for links connecting the aggregation and core layer. Our network model works much like the one from Section 2. However, we model a single destination, switch 1, and we elide the final hop to the appropriate host connected to this switch. M(p, t) ≜ in ; do (p ; t) while (¬sw=1) 4 ECMP implementations are usually based on hashing, which approximates random forwarding provided there is sufficient entropy in the header fields used to select an outgoing port. Probabilistic Program Equivalence for NetKAT k 0 1 2 3 4 ∞ b b b M(F10 0, t, f k ) M(F10 3, t, f k ) M(F10 3,5, t, f k ) ≡ teleport ≡ teleport ≡ teleport ✓ ✗ ✗ ✗ ✗ ✗ ✓ ✓ ✓ ✗ ✗ ✗ ✓ ✓ ✓ ✓ ✗ ✗ Table 1. Evaluating k-resilience of F10. 21 k 0 1 2 3 4 ∞ compare compare compare (F100, F103 ) (F103, F103,5 ) (F103,5, teleport) ≡ < < < < < ≡ ≡ ≡ < < < ≡ ≡ ≡ ≡ < < Table 2. Comparing schemes under k failures. The ingress predicate in is a disjunction of switch-and-port tests over all ingress locations. This first b t, f ) that integrates the failure model and declares all model is embedded into a refined model M(p, necessary local variables that track the healthiness of individual ports: b t, f ) ≜ var up1 ←1 in M(p, ... var upd ←1 in M((f ; p), t) Here d denotes the maximum degree of all nodes in the FatTree and AB FatTree topologies from Figures 5 and 6, which we encode as programs fattree and abfattree. much like in Section 2.2. 6.3 Checking invariants We can gain confidence in the correctness of our implementation of F10 by verifying that it maintains certain key invariants. As an example, recall our implementation of F103,5 : when we perform 5-hop rerouting, we use an extra bit (default) to notify the next hop aggregation switch to forward the packet downwards instead of performing default forwarding. The next hop follows this instruction and also sets default back to 1. By design, the packet can not be delivered to the destination with default set to 0. To verify this property, we check the following equivalence: b b ∀t, k : M(F10 3,5 , t, f k ) ≡ M(F10 3,5 , t, f k ) ; default=1 We executed the check using our implementation for k ∈ {0, 1, 2, 3, 4, ∞} and t ∈ {fattree, abfattree}. As discussed below, we actually failed to implement this feature correctly on our first attempt due to a subtle bug—we neglected to initialize the default flag to 1 at the ingress. 6.4 F10 routing with FatTree We previously saw that the structure of FatTree doesn’t allow 3-hop rerouting on failures because all subtrees are of the same type. This would mean that augmenting ECMP with 3-hop rerouting should have no effect, i.e. 3-hop rerouting should never kick in and act as a no-op. To verify this, we can check the following equivalence: b b ∀k : M(F10 0 , fattree, f k ) ≡ M(F10 3 , fattree, f k ) We have used our implementation to check that this equivalence indeed holds for k ∈ {0, 1, 2, 3, 4, ∞}. 22 S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva 1.00 Pr[delivery] 0.95 0.90 AB FatTree, F10 no rerouting AB FatTree, F10 3-hop rerouting AB FatTree, F10 3+5-hop rerouting FatTree, F10 3+5-hop rerouting 0.85 0.80 1/128 1/64 1/32 1/16 Link failure probability 1/8 1/4 Fig. 8. Probability of delivery vs. link-failure probability. (k = ∞). 6.5 Refinement Recall that we implemented F10 in three stages. We started with a basic routing scheme (F100 ) based on ECMP that provides resilience on the upward path, but no rerouting capabilities on the downward paths. We then augmented this scheme by adding 3-hop rerouting to obtain F103 , which can route around certain failures in the aggregation layer. Finally, we added 5-hop rerouting to address failure cases that 3-hop rerouting cannot handle, obtaining F103,5 . Hence, we would expect the probability of packet delivery to increase with each refinement of our routing scheme. Additionally, we expect all schemes to deliver packets and drop packets with some probability under the unbounded failure model. These observations are summarized by the following ordering: b b b drop < M(F10 0 , t, f ∞ ) < M(F10 3 , t, f ∞ ) < M(F10 3,5 , t, f ∞ ) < teleport where t = abfattree and teleport ≜ sw←1. To our surprise, we were not able to verify this property initially, as our implementation indicated that the ordering b b M(F10 3 , t, f ∞ ) < M(F10 3,5 , t, f ∞ ) was violated. We then added a capability to our implementation to obtain counterexamples, and found that F103 performed better than F103,5 for packets π with π .default = 0. We were missing the first line in our implementation of F103,5 (cf., Figure 7) that initializes the default bit to 1 at the ingress, causing packets to be dropped! After fixing the bug, we were able to confirm the expected ordering. 6.6 k-resilience We saw that there exists a strict ordering in terms of resilience for F100 , F103 and F103,5 when an unbounded number of failures can happen. Another interesting way of measuring resilience is to count the minimum number of failures at which a scheme fails to guarantee 100% delivery. Using ProbNetKAT, we can measure this resilience by setting k in fk to increasing values and checking equivalence with teleportation. Table 1 shows the results based on our decision procedure for the AB FatTree topology from Figure 6. The naive scheme, F100 , which does not perform any rerouting, drops packets when a failure occurs on the downward path. Thus, it is 0-resilient. In the example topology, 3-hop rerouting Probabilistic Program Equivalence for NetKAT 23 1.0 Pr[hop count ≤ x] 0.9 0.8 AB FatTree, F10 no rerouting AB FatTree, F10 3-hop rerouting AB FatTree, F10 3+5-hop rerouting FatTree, F10 3+5-hop rerouting 0.7 0.6 2 4 6 8 Hop count 10 12 14 Fig. 9. Increased latency due to resilience. (k = ∞, p = 41 ) has two possible ways to reroute for the given failure. Even if only one of the type B subtrees is reachable, F103 can still forward traffic. However, if both the type B subtrees are unreachable, then F103 will not be able to reroute traffic. Thus, F103 is 2-resilient. Similarly, F103,5 can route as long as any aggregation switch is reachable from the core switch. For F103,5 to fail the core switch would need to be disconnected from all four aggregation switches. Hence it is 3-resilient. In cases where schemes are not equivalent to teleport, we can characterize the relative robustness by computing the ordering, as shown in Table 2. 6.7 Resilience under increasing failure rate We can also do more quantitative analyses such as evaluating the effect of increase failure probability of links on the probability of packet delivery. Figure 8 shows this analysis in a failure model in which an unbounded number of failures can occur simultaneously. We find that F100 ’s delivery probability dips significantly as the failure probability increases because F100 is not resilient to failures. In contrast, both F103 and F103,5 continue to ensure high probability of delivery by rerouting around failures. 6.8 Cost of resilience By augmenting naive routing schemes with rerouting mechanisms, we are able to achieve a higher degree of resilience. But this benefit comes at a cost. The detours taken to reroute traffic increase the latency (hop count) for packets. ProbNetKAT enables quantifying this increase in latency by augmenting our model with a counter that gets increased at every hop. Figure 9 shows the CDF of latency as the fraction of traffic delivered within a given hop count. On AB FatTree, we find that F100 delivers as much traffic as it can (≈80%) within a hop count ≤ 4 because the maximum length of a shortest path from any edge switch to s1 is 4 and F100 does not use any longer paths. F103 and F103,5 deliver the same amount of traffic with hop count ≤ 4. But, with 2 additional hops, they are able to deliver significantly more traffic because they perform 3-hop rerouting to handle certain failures. With 4 additional hops, F103,5 ’s throughput increases as 5-hop rerouting helps. We find that F103 also delivers more traffic with 8 hops—these are the cases when F103 performs 3-hop rerouting twice for a single packet as it encountered failure twice. Similarly, we see small increases 24 S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva 4.8 E[hop count | delivered] AB FatTree, F10 no rerouting AB FatTree, F10 3-hop rerouting AB FatTree, F10 3+5-hop rerouting FatTree, F10 3+5-hop rerouting 4.6 4.4 4.2 4.0 3.8 3.6 1/128 1/64 1/32 1/16 Link failure probability 1/8 1/4 Fig. 10. Expected hop-count conditioned on delivery. (k = ∞). in throughput for higher hop counts. We find that F103,5 improves resilience for FatTree too, but the impact on latency is significantly higher as FatTree does not support 3-hop rerouting. 6.9 Expected latency Figure 10 shows the expected hop-count of paths taken by packets conditioned on their delivery. Both F103 and F103,5 deliver packets with high probability even at high failure probabilities, as we saw in Figure 8. However, a higher probability of link-failure implies that it becomes more likely for these schemes to invoke rerouting, which increases hop count. Hence, we see the increase in expected hop-count as failure probability increases. F103,5 uses 5-hop rerouting to achieve more resilience compared to F103 , which performs only 3-hop rerouting, and this leads to slightly higher expected hop-count for F103,5 . We see that the increase is more significant for FatTree in contrast to AB FatTree because FatTree only supports 5-hop rerouting. As the failure probability increases, the probability of delivery for packets that are routed via the core layer decreases significantly for F100 (recall Figure 8). Thus, the distribution of delivered packets shifts towards those with direct 2-hop path via an aggregation switch (such as packets from s2 to s1), and hence the expected hop-count decreases slightly. 6.10 Discussion As this case study of resilient routing in datacenters shows, the stochastic matrix representation of ProbNetKAT programs and accompanying decision procedure enable us to answer a wide variety of questions about probabilistic networks completely automatically. These new capabilities represent a signficant advance over current network verification tools, which are based on deterministic packet-forwarding models [9, 15, 17, 22]. 7 DECIDING FULL PROBNETKAT: OBSTACLES AND CHALLENGES As we have just seen, history-free ProbNetKAT can describe sophisticated network routing schemes under various failure models, and program equivalence for the language is decidable. However, it is less expressive than the original ProbNetKAT language, which includes an additional primitive dup. Intuitively, this command duplicates a packet π ∈ Pk and outputs the word π π ∈ H, where H = Pk∗ is the set of non-empty, finite sequences of packets. An element of H is called a packet Probabilistic Program Equivalence for NetKAT 25 history, representing a log of previous packet states. ProbNetKAT policies may only modify the first (head) packet of each history; dup fixes the current head packet into the log by copying it. In this way, ProbNetKAT policies can compute distributions over the paths used to forward packets, instead of just over the final output packets. However, with dup, the semantics of ProbNetKAT becomes significantly more complex. Policies p now transform sets of packet histories a ∈ 2H to distributions JpK(a) ∈ D(2H ). Since 2H is uncountable, these distributions are no longer guaranteed to be discrete, and formalizing the semantics requires full-blown measure theory (see prior work for details [31]). Deciding program equivalence also becomes more challenging. Without dup, policies operate on sets of packets 2Pk ; crucially, this is a finite set and we can represent each set with a single state in a finite Markov chain. With dup, policies operate on sets of packet histories 2H . Since this set is not finite—in fact, it is not even countable—encoding each packet history as a state would give a Markov chain with infinitely many states. Procedures for deciding equivalence are not known for such systems. While in principle there could be a more compact representation of general ProbNetKAT policies as finite Markov chains or other models where equivalence is decidable, (e.g., weighted or probabilistic automata [7] or quantitative variants of regular expressions [2]), we suspect that deciding equivalence in the presence of dup is intractable. As evidence in support of this conjecture, we show that ProbNetKAT policies can simulate the following kind of probabilistic automata. This model appears to be new, and may be of independent interest. Definition 7.1. Let A be a finite alphabet. A 2-generative probabilistic automata is defined by a tuple (S, s 0 , ρ, τ ) where S is a finite set of states; s 0 ∈ S is the initial state; ρ : S → (A ∪ {_})2 maps each state to a pair of letters (u, v), where either u or v may be a special blank character _; and the transition function τ : S → D(S) gives the probability of transitioning from one state to another. The semantics of an automaton can be defined as a probability measure on the space A∞ × A∞ , where A∞ is the set of finite and (countably) infinite words over the alphabet A. Roughly, these measures are fully determined by the probabilities of producing any two finite prefixes of words (w, w ′) ∈ A∗ × A∗ . Presenting the formal semantics would require more concepts from measure theory and take us far afield, but the basic idea is simple to describe. An infinite trace of a 2-generative automata over states s 0 , s 1 , s 2 , . . . gives a sequence of pairs of (possibly blank) letters: ρ(s 0 ), ρ(s 1 ), ρ(s 2 ) . . . By concatenating these pairs together and dropping all blank characters, a trace induces two (finite or infinite) words over the alphabet A. For example, the sequence, (a 0 , _), (a 1 , _), (_, a 2 ), . . . gives the words a 0a 1 . . . and a 2 . . . . Since the traces are generated by the probabilistic transition function τ , each automaton gives rise to a probability measure over pairs of words. While we have no formal proof of hardness, deciding equivalence between these automata appears highly challenging. In the special case where only one word is generated (say, when the second component produced is always blank), these automata are equivalent to standard automata with ε-transitions (e.g., see [23]). In the standard setting, non-productive steps can be eliminated and the automata can be modeled as a finite state Markov chain, where equivalence is decidable. In our setting, however, steps producing blank letters in one component may produce non-blank letters in the other. As a result, it is not entirely clear how to eliminate these steps and encode our automata as a Markov chain. 26 S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva Returning to ProbNetKAT, 2-generative automata can be encoded as policies with dup. We sketch the idea here, deferring further details to Appendix B. Suppose we are given an automaton (S, s 0 , ρ, τ ). We build a ProbNetKAT policy over packets with two fields, st and id. The first field st ranges over the states S and the alphabet A, while the second field id is either 1 or 2; we suppose the input set has exactly two packets labeled with id = 1 and id = 2. In a set of packet history, the two active packets have the same value for st ∈ S—this represents the current state in the automata. Past packets in the history have st ∈ A, representing the words produced so far; the first and second components of the output are tracked by the histories with id = 1 and id = 2. We can encode the transition function τ as a probabilistic choice in ProbNetKAT, updating the current state st of all packets, and recording non-blank letters produced by ρ in the two components by applying dup on packets with the corresponding value of id. Intuitively, a set of packet histories generated by the resulting ProbNetKAT term describes a pair of words generated by the original automaton. With a bit more bookkeeping (see Appendix B), we can show that two 2-generative automata are equivalent if and only if their encoded ProbNetKAT policies are equivalent. Thus, deciding equivalence for ProbNetKAT with dup is harder than deciding equivalence for 2-generative automata. Showing hardness for the full framework is a fascinating open question. At the same time, deciding equivalence between 2-generative automata appears to require substantially new ideas; these insights could shed light on how to decide equivalence for the full ProbNetKAT language. 8 RELATED WORK A key ingredient that underpins the results in this paper is the idea of representing the semantics of iteration using absorbing Markov chains, and exploiting their properties to directly compute limiting distributions on them. Markov chains have been used by several authors to represent and to analyze probabilistic programs. An early example of using Markov chains for modeling probabilistic programs is the seminal paper by Sharir, Pnueli, and Hart [28]. They present a general method for proving properties of probabilistic programs. In their work, a probabilistic program is modeled by a Markov chain and an assertion on the output distribution is extended to an invariant assertion on all intermediate distributions (providing a probabilistic generalization of Floyd’s inductive assertion method). Their approach can assign semantics to infinite Markov chains for infinite processes, using stationary distributions of absorbing Markov chains in a similar way to the one used in this paper. Note however that the state space used in this and other work is not like ProbNetKAT’s current and accumulator sets (2P k × 2P k ), but is instead is the Cartesian product of variable assignments and program location. In this sense, the absorbing states occur for program termination, rather than for accumulation as in ProbNetKAT. Although packet modification is clearly related to variable assignment, accumulation does not clearly relate to program location. Readers familiar with prior work on probabilistic automata might wonder if we could directly apply known results on (un)decidability of probabilistic rational languages. This is not the case— probabilistic automata accept distributions over words, while ProbNetKAT programs encode distributions over languages. Similarly, probabilistic programming languages, which have gained popularity in the last decade motivated by applications in machine learning, focus largely on Bayesian inference. They typically come equipped with a primitive for probabilistic conditioning and often have a semantics based on sampling. Working with ProbNetKAT has a substantially different style, in that the focus is on on specification and verification rather than inference. Di Pierro, Hankin, and Wiklicky have used probabilistic abstract interpretation (PAI) to statically analyze probabilistic λ-calculus [6]. They introduce a linear operator semantics (LOS) and demonstrate a strictness analysis, which can be used in deterministic settings to replace lazy with Probabilistic Program Equivalence for NetKAT 27 eager evaluation without loss. Their work was later extended to a language called pW hile, using a store plus program location state-space similar to [28]. The language pW hile is a basic imperative language comprising while-do and if-then-else constructs, but augmented with random choice between program blocks with a rational probability, and limited to a finite of number of finitely-ranged variables (in our case, packet fields). The authors explicitly limit integers to finite sets for analysis purposes to maintain finiteness, arguing that real programs will have fixed memory limitations. In contrast to our work, they do not deal with infinite limiting behavior beyond stepwise iteration, and do not guarantee convergence. Probabilistic abstract interpretation is a new but growing field of research [34]. Olejnik, Wicklicky, and Cheraghchi provided a probabilistic compiler pwc for a variation of pW hile [25], implemented in OCaml, together with a testing framework. The pwc compiler has optimizations involving, for instance, the Kronecker product to help control matrix size, and a Julia backend. Their optimizations based on the Kronecker product might also be applied in, for instance, the generation of SJpK from BJpK, but we have not pursued this direction as of yet. There is a plenty of prior work on finding explicit distributions of probabilistic programs. Gordon, Henzinger, Nori, and Rajamani surveyed the state of the art with regard to probabilistic inference [12]. They show how stationary distributions on Markov chains can be used for the semantics of infinite probabilistic processes, and how they converge under certain conditions. Similar to our approach, they use absorbing strongly-connected-components to represent termination. Markov chains are used in many probabilistic model checkers, of which PRISM [20] is a prime example. PRISM supports analysis of discrete-time Markov chains, continuous-time Markov chains, and Markov decision processes. The models are checked against specifications written in temporal logics like PCTL and CSL. PRISM is written in Java and C++ and provides three model checking engines: a symbolic one with (multi-terminal) binary decision diagrams ((MT)BDDs), a sparse matrix one, and a hybrid. The use of PRISM to analyse ProbNetKAT programs is an interesting research avenue and we intend to explore it in the future. 9 CONCLUSION This paper settles the decidability of program equivalence for history-free ProbNetKAT. The key technical challenge is overcome by modeling the iteration operator as an absorbing Markov chain, which makes it possible to compute a closed-form solution for its semantics. The resulting tool is useful for reasoning about a host of other program properties unrelated to equivalence. Natural directions for future work include investigating equivalence for full ProbNetKAT, developing an optimized implementation, and exploring new applications to networks and beyond. REFERENCES [1] Mohammad Al-Fares, Alexander Loukissas, and Amin Vahdat. 2008. A Scalable, Commodity Data Center Network Architecture. In ACM SIGCOMM Computer Communication Review, Vol. 38. ACM, 63–74. [2] Rajeev Alur, Dana Fisman, and Mukund Raghothaman. 2016. Regular programming for quantitative properties of data streams. In ESOP 2016. 15–40. [3] Carolyn Jane Anderson, Nate Foster, Arjun Guha, Jean-Baptiste Jeannin, Dexter Kozen, Cole Schlesinger, and David Walker. 2014. NetKAT: Semantic Foundations for Networks. In POPL. 113–126. [4] Manav Bhatia, Mach Chen, Sami Boutros, Marc Binderberger, and Jeffrey Haas. 2014. Bidirectional Forwarding Detection (BFD) on Link Aggregation Group (LAG) Interfaces. RFC 7130. (Feb. 2014). https://doi.org/10.17487/RFC7130 [5] Timothy A. Davis. 2004. Algorithm 832: UMFPACK V4.3—an Unsymmetric-pattern Multifrontal Method. ACM Trans. Math. Softw. 30, 2 (June 2004), 196–199. https://doi.org/10.1145/992200.992206 [6] Alessandra Di Pierro, Chris Hankin, and Herbert Wiklicky. 2005. Probabilistic λ-calculus and quantitative program analysis. Journal of Logic and Computation 15, 2 (2005), 159–179. https://doi.org/10.1093/logcom/exi008 [7] Manfred Droste, Werner Kuich, and Heiko Vogler. 2009. Handbook of Weighted Automata. Springer. 28 S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva [8] Nate Foster, Dexter Kozen, Konstantinos Mamouras, Mark Reitblatt, and Alexandra Silva. 2016. Probabilistic NetKAT. In ESOP. 282–309. https://doi.org/10.1007/978-3-662-49498-1_12 [9] Nate Foster, Dexter Kozen, Matthew Milano, Alexandra Silva, and Laure Thompson. 2015. A Coalgebraic Decision Procedure for NetKAT. In POPL. ACM, 343–355. [10] Phillipa Gill, Navendu Jain, and Nachiappan Nagappan. 2011. Understanding Network Failures in Data Centers: Measurement, Analysis, and Implications. In ACM SIGCOMM. 350–361. [11] Michele Giry. 1982. A categorical approach to probability theory. In Categorical aspects of topology and analysis. Springer, 68–85. https://doi.org/10.1007/BFb0092872 [12] Andrew D Gordon, Thomas A Henzinger, Aditya V Nori, and Sriram K Rajamani. 2014. Probabilistic programming. In Proceedings of the on Future of Software Engineering. ACM, 167–181. https://doi.org/10.1145/2593882.2593900 [13] Chuanxiong Guo, Guohan Lu, Dan Li, Haitao Wu, Xuan Zhang, Yunfeng Shi, Chen Tian, Yongguang Zhang, and Songwu Lu. 2009. BCube: A High Performance, Server-centric Network Architecture for Modular Data Centers. ACM SIGCOMM Computer Communication Review 39, 4 (2009), 63–74. [14] Chuanxiong Guo, Haitao Wu, Kun Tan, Lei Shi, Yongguang Zhang, and Songwu Lu. 2008. Dcell: A Scalable and Fault-Tolerant Network Structure for Data Centers. In ACM SIGCOMM Computer Communication Review, Vol. 38. ACM, 75–86. [15] Peyman Kazemian, George Varghese, and Nick McKeown. 2012. Header Space Analysis: Static Checking for Networks. In USENIX NSDI 2012. 113–126. https://www.usenix.org/conference/nsdi12/technical-sessions/presentation/kazemian [16] John G Kemeny, James Laurie Snell, et al. 1960. Finite markov chains. Vol. 356. van Nostrand Princeton, NJ. [17] Ahmed Khurshid, Wenxuan Zhou, Matthew Caesar, and Brighten Godfrey. 2012. Veriflow: Verifying Network-Wide Invariants in Real Time. In ACM SIGCOMM. 467–472. [18] Dexter Kozen. 1981. Semantics of probabilistic programs. J. Comput. Syst. Sci. 22, 3 (1981), 328–350. https://doi.org/10. 1016/0022-0000(81)90036-2 [19] Dexter Kozen. 1997. Kleene Algebra with Tests. ACM TOPLAS 19, 3 (May 1997), 427–443. https://doi.org/10.1145/ 256167.256195 [20] M. Kwiatkowska, G. Norman, and D. Parker. 2011. PRISM 4.0: Verification of Probabilistic Real-time Systems. In Proc. 23rd International Conference on Computer Aided Verification (CAV’11) (LNCS), G. Gopalakrishnan and S. Qadeer (Eds.), Vol. 6806. Springer, 585–591. https://doi.org/10.1007/978-3-642-22110-1_47 [21] Vincent Liu, Daniel Halperin, Arvind Krishnamurthy, and Thomas E Anderson. 2013. F10: A Fault-Tolerant Engineered Network. In USENIX NSDI. 399–412. [22] Haohui Mai, Ahmed Khurshid, Rachit Agarwal, Matthew Caesar, P. Brighten Godfrey, and Samuel Talmadge King. 2011. Debugging the Data Plane with Anteater. In ACM SIGCOMM. 290–301. [23] Mehryar Mohri. 2000. Generic ε -removal algorithm for weighted automata. In CIAA 2000. Springer, 230–242. [24] Radhika Niranjan Mysore, Andreas Pamboris, Nathan Farrington, Nelson Huang, Pardis Miri, Sivasankar Radhakrishnan, Vikram Subramanya, and Amin Vahdat. 2009. Portland: A Scalable Fault-Tolerant Layer 2 Data Center Network Fabric. In ACM SIGCOMM Computer Communication Review, Vol. 39. ACM, 39–50. [25] Maciej Olejnik, Herbert Wiklicky, and Mahdi Cheraghchi. 2016. Probabilistic Programming and Discrete Time Markov Chains. (2016). http://www.imperial.ac.uk/media/imperial-college/faculty-of-engineering/computing/public/ MaciejOlejnik.pdf [26] Arjun Roy, Hongyi Zeng, Jasmeet Bagga, George Porter, and Alex C. Snoeren. 2015. Inside the Social Network’s (Datacenter) Network. In ACM SIGCOMM. 123–137. [27] N. Saheb-Djahromi. 1980. CPOs of measures for nondeterminism. Theoretical Computer Science 12 (1980), 19–37. https://doi.org/10.1016/0304-3975(80)90003-1 [28] Micha Sharir, Amir Pnueli, and Sergiu Hart. 1984. Verification of probabilistic programs. SIAM J. Comput. 13, 2 (1984), 292–314. https://doi.org/10.1137/0213021 [29] Ankit Singla, Chi-Yao Hong, Lucian Popa, and P Brighten Godfrey. 2012. Jellyfish: Networking Data Centers Randomly. In USENIX NSDI. 225–238. [30] Steffen Smolka, Spiros Eliopoulos, Nate Foster, and Arjun Guha. 2015. A Fast Compiler for NetKAT. In ICFP 2015. https://doi.org/10.1145/2784731.2784761 [31] Steffen Smolka, Praveen Kumar, Nate Foster, Dexter Kozen, and Alexandra Silva. 2017. Cantor Meets Scott: Semantic Foundations for Probabilistic Networks. In POPL 2017. https://doi.org/10.1145/3009837.3009843 [32] Robert Endre Tarjan. 1975. Efficiency of a Good But Not Linear Set Union Algorithm. J. ACM 22, 2 (1975), 215–225. https://doi.org/10.1145/321879.321884 [33] L. Valiant. 1982. A Scheme for Fast Parallel Communication. SIAM J. Comput. 11, 2 (1982), 350–361. [34] Di Wang, Jan Hoffmann, and Thomas Reps. 2018. PMAF: An Algebraic Framework for Static Analysis of Probabilistic Programs. In POPL 2018. https://www.cs.cmu.edu/~janh/papers/WangHR17.pdf Probabilistic Program Equivalence for NetKAT A 29 OMITTED PROOFS Lemma A.1. Let A be a finite boolean combination of basic open sets, i.e. sets of the form Ba = {a} ↑ for a ∈ ℘ω (H), and let L−M denote the semantics from [31]. Then for all programs p and inputs a ∈ 2H , Lp ∗ M(a)(A) = lim Lp (n) M(a)(A) n→∞ Proof. Using topological arguments, the claim follows directly from previous results: A is a Cantor-clopen set by [31] (i.e., both A and A are Cantor-open), so its indicator function 1A is Cantor-continuous. But µ n ≜ Lp (n) M(a) converges weakly to µ ≜ Lp ∗ M(a) in the Cantor topology (Theorem 4 in [8]), so ∫ ∫ lim Lp (n) M(a)(A) = lim 1Adµ n = 1Adµ = Lp ∗ M(a)(A) n→∞ n→∞ (To see why A and A are open in the Cantor topology, note that they can be written in disjunctive normal form over atoms B {h } .) □ Proof of Proposition 3.1. We only need to show that for dup-free programs p and history-free inputs a ∈ 2Pk , LpM(a) is a distribution on packets (where we identify packets and singleton histories). We proceed by structural induction on p. All cases are straightforward except perhaps the case of p ∗ . For this case, by the induction hypothesis, all Jp (n) K(a) are discrete probability distributions on packet sets, therefore vanish outside 2Pk . By Lemma A.1, this is also true of the limit Jp ∗ K(a), as its value on 2Pk must be 1, therefore it is also a discrete distribution on packet sets. □ Proof of Lemma 3.3. This follows directly from Lemma A.1 and Proposition 3.1 by noticing that any set A ⊆ 2Pk is a finite boolean combination of basic open sets. □ Lemma A.2. The matrix X = I − Q in Equation (2) of §5.1 is invertible. Proof. Let S be a finite set of states, |S | = n, M an S × S substochastic matrix (Mst ≥ 0, M1 ≤ 1). Í i A state s is defective if (M1)s < 1. We say M is stochastic if M1 = 1, irreducible if ( n−1 i=0 M )st > 0 (that is, the support graph of M is strongly connected), and aperiodic if all entries of some power of M are strictly positive. We show that if M is substochastic such that every state can reach a defective state via a path in the support graph, then the spectral radius of M is strictly less than 1. Intuitively, all weight in the system eventually drains out at the defective states. Let es , s ∈ S, Í be the standard basis vectors. As a distribution, esT is the unit point mass on s. For A ⊆ S, let e A = s ∈A es . The L 1 -norm of a substochastic vector is its total weight as a distribution. Multiplying on the right by M never increases total weight, but will strictly decrease it if there is nonzero weight on a defective state. Since every state can reach a defective state, this must happen Í after n steps, thus ∥esT M n ∥1 < 1. Let c = maxs ∥esT M n ∥1 < 1. For any y = s as es , Õ ∥yT M n ∥1 = ∥( as es )T M n ∥1 s ≤ Õ s |as | · ∥esT M n ∥1 ≤ Õ |as | · c = c · ∥yT ∥1 . s Then M n is contractive in the L 1 norm, so |λ| < 1 for all eigenvalues λ. Thus I − M is invertible because 1 is not an eigenvalue of M. □ 30 B S. Smolka, P. Kumar, N. Foster, J. Hsu, D. Kahn, D. Kozen, and A. Silva ENCODING 2-GENERATIVE AUTOMATA IN FULL PROBNETKAT To keep notation light, we describe our encoding in the special case where the alphabet A = {x, y}, there are four states S = {s 1 , s 2 , s 3 , s 4 }, the initial state is s 1 , and the output function ρ is ρ(s 1 ) = (x, _) ρ(s 2 ) = (y, _) ρ(s 3 ) = (_, x) ρ(s 4 ) = (_, y). Encoding general automata is not much more complicated. Let τ : S → D(S) be a given transition function; we write pi, j for τ (si )(s j ). We will build a ProbNetKAT policy simulating this automaton. Packets have two fields, st and id, where st ranges over S ∪A ∪ {•} and id ranges over {1, 2}. Define: p ≜ st=s 1 ; loop∗ ; st←• The initialization keeps packets that start in the initial state, while the final command marks histories that have exited the loop by setting st to be special letter •. The main program loop first branches on the current state st:  st=s 1      st=s 2  loop ≜ case  st=s 3     st=s 4  : state1 : state2 : state3 : state4 Then, the policy simulates the behavior from each state. For instance:    (if id=1 then st←x ; dup else skip) ; st←s 1 @ p1,1 ,  Ê  (if id=1 then st←y ; dup else skip) ; st←s 2 @ p1,2 ,  state1 ≜  (if id=2 then st←x ; dup else skip) ; st←s 3 @ p1,3 ,     (if id=2 then st←y ; dup else skip) ; st←s @ p 4 1,4  The policies state2, state3, state4 are defined similarly. Now, suppose we are given two 2-generative automata W ,W ′ that differ only in their transition functions. For simplicity, we will further assume that both systems have strictly positive probability of generating a letter in either component in finitely many steps from any state. Suppose they generate distributions µ, µ ′ respectively over pairs of infinite words Aω × Aω . Now, consider the encoded ProbNetKAT policies p, p ′. We argue that JpK = JqK if and only if µ = µ ′.5 First, it can be shown that JpK = Jp ′K if and only if JpK(e) = Jp ′K(e), where e ≜ {π π | π ∈ Pk}. and ν ′ Let ν = JpK(e) = is the following equality: Jp ′K(e). The key connection between the automata and the encoded policies µ(Su,v ) = ν(Tu,v ) (6) for every pair of finite prefixes u, v ∈ A∗ . In the automata distribution on the left, Su,v ⊆ Aω × Aω consists of all pairs of infinite strings where u is a prefix of the first component and v is a prefix of the second component. In the ProbNetKAT distribution on the right, we first encode u and v as packet histories. For i ∈ {1, 2} representing the component and w ∈ A∗ a finite word, define the history hi (w) ∈ H ≜ (st = •, id = i), (st = w[|w |], id = i), . . . , (st = w[1], id = i), (st = s 1 , id = i). The letters of the word w are encoded in reverse order because by convention, the head/newest packet is written towards the left-most end of a packet history, while the oldest packet is written 5 We will not present the semantics of ProbNetKAT programs with dup here; instead, the reader should consult earlier papers [8, 31] for the full development. Probabilistic Program Equivalence for NetKAT 31 towards the right-most end. For instance, the final letter w[|w |] is the most recent (i.e., the latest) letter produced by the policy. Then, Tu,v is the set of all history sets including h1 (u) and h2 (v): Tu,v ≜ {a ∈ 2H | h1 (u) ∈ a, h2 (v) ∈ a}. Now JpK = Jp ′K implies µ = µ ′, since Equation (6) gives µ(Su,v ) = µ ′(Su,v ). The reverse implication is a bit more delicate. Again by Equation (6), we have ν (Tu,v ) = ν ′(Tu,v ). We need to extend this equality to all cones, defined by packet histories h: B h ≜ {a ∈ 2H | h ∈ a}. This follows by expressing B h as boolean combinations of Tu,v , and observing that the encoded policy produces only sets of encoded histories, i.e., where the most recent state st is set to • and the initial state st is set to s 1 . 

Motif-based Rule Discovery for Predicting Real-valued Time Series∗†‡ Yuanduo He, Xu Chu, Juguang Peng, Jingyue Gao, Yasha Wang arXiv:1709.04763v4 [cs.AI] 2 Dec 2017 Key Laboratory of High Confidence Software Technologies, Ministry of Education, Beijing 100871, China {ydhe, chu xu, pgj.pku12, gaojingyue1997, yasha.wang}@pku.edu.cn Abstract Time series prediction is of great significance in many applications and has attracted extensive attention from the data mining community. Existing work suggests that for many problems, the shape in the current time series may correlate an upcoming shape in the same or another series. Therefore, it is a promising strategy to associate two recurring patterns as a rule’s antecedent and consequent: the occurrence of the antecedent can foretell the occurrence of the consequent, and the learned shape of consequent will give accurate predictions. Earlier work employs symbolization methods, but the symbolized representation maintains too little information of the original series to mine valid rules. The state-of-the-art work, though directly manipulating the series, fails to segment the series precisely for seeking antecedents/consequents, resulting in inaccurate rules in common scenarios. In this paper, we propose a novel motifbased rule discovery method, which utilizes motif discovery to accurately extract frequently occurring consecutive subsequences, i.e. motifs, as antecedents/consequents. It then investigates the underlying relationships between motifs by matching motifs as rule candidates and ranking them based on the similarities. Experimental results on real open datasets show that the proposed approach outperforms the baseline method by 23.9%. Furthermore, it extends the applicability from single time series to multiple ones. Introduction The prediction of real-valued time series is a topic of great significance and colossal enthusiasm in the data mining community, and has been applied extensively in many research areas (Hamilton 1994; Esling and Agon 2012). Most of the work in literature aims at modeling the dependencies among variables, and forecasting the next few values of a series based on the current values (Makridakis, Wheelwright, and Hyndman 2008). However, other work suggests that for many problems it is the shape of the current pattern rather than actual values that makes the prediction (Das et al. 1998). For clarity, we call the latter one, forecasting by ∗ The work remains incomplete and will be further refined. Read it on your own risk :). † We are grateful to Leye Wang, Junyi Ma, Zhu Jin, and Siqi Yang for their invaluable help. ‡ We are also grateful for the reviewers in AAAI18. shape, rule-based prediction, which is the subject of this paper. Informally1 , a rule A ⇒τ B associates a pattern A as the antecedent with a pattern B as the consequent in an upperbounding time interval τ . The rule-based prediction works as follows: if a shape resembling A is observed in a series, then a shape resembling B is supposed to appear in the same or another series, within the time interval τ . Usually, a rule refers to an “authentic rule”, an underlying relationship between two events implied by the two shapes. In most work, firstly the subsequence clustering method is employed to symbolize/discretize the series, and then the symbolic series rule discovery method is applied to find rules from real-valued series (Das et al. 1998). However, such work has met limited success and ends up discovering spurious rules (e.g. “good” rules discovered from random walk data), because the symbol representation by common symbolization methods is independent with the raw real-valued series (Keogh and Lin 2005). In conclusion, the symbolized time series maintains little information about the original series to mine valid rules. The state-of-the-art work (Shokoohi-Yekta et al. 2015) directly manipulates the series under the assumption that a rule is contained in a subsequence. It first selects a subsequence and then splits it into a rule’s antecedent and consequent. However, usually there is an interval between a rule’s antecedent and consequent. The splitting method will append the extra interval series to the antecedent/consequent, which fails to segment precisely for antecedents/consequents and results in rules with a lower prediction performance. Furthermore, it cannot be applied to find rules from multiple series, i.e. a shape in one series predicting a shape in another series. Only when the observed object presents repeatability, can predictions be valid/reasonable. We believe that a serviceable rule should appear repeatedly in a real-valued time series and therefore its antecedent and consequent must be recurring patterns of the series. Recent studies (Vahdatpour, Amini, and Sarrafzadeh 2009; Brown et al. 2013; Mueen 2014) show that Motifs, frequently occurring patterns in time series (Patel et al. 2002), contain key information about the series. Therefore, we could utilize motif dis1 A formal definition is given in the Preliminaries section. ries. • We develop a novel heuristic matching algorithm in search of the best combinations of motif instances. To accommodate the heuristic matching result, we modify Yekta’s scoring method leveraging Minimum Description Length (MDL) principle to evaluate each rule (ShokoohiYekta et al. 2015). • We evaluate our work on real open datasets, and the experimental results show that our method outperforms stateof-the-art work by 23.9% on average2 . Moreover, rules in multiple series can also be discovered. Figure 1: The electrical penetration graphs (EPG) data monitoring of the probing behavior of leafhoppers. There are three motifs in the series, motif 1 and motif 2 are overlapped in the green curves. Figure 2: An underlying rule with respect to motif mA and mB . TA and TB are two series. m˜A (i) and m˜B (j) are motif instances of mA and mB in TA and TB , respectively. m˜A (1) with m˜B (2) or m˜B (4) can form different rule instances, i.e. the two solid lines. covery methods to find all distinct motifs which accurately represent antecedent and consequent candidates of rules. However, three challenges are confronted: (1) there could be quite a few (overlapping) motifs in a single series (see Figure 1), let alone the dataset containing multiple series. How to screen for effective motifs as antecedents and consequents? (2) there could be many instances for two given effective motifs and different combinations of them will lead to different instances of a rule (see Figure 2). How to identify the best combinations for the underlying rule? (3) How to rank different rules given the instances of each rule? In this paper, we present a rule-based prediction method for real-valued time series from the perspective of motifs. For each pair of motifs as a candidate rule, a heuristicmatching-based scoring algorithm is developed to investigate the connection between them. We summarize contributions of this paper as follows: • We propose a novel rule discovery approach from the perspective of motifs. It can find rules with higher prediction performance and can also be applied to multiple time se- Related Work Early work extracts association rules from frequent patterns in the transactional database (Han, Pei, and Kamber 2011). Das et al. are the first to study the problem of finding rules from real-valued time series in a symbolization-based perspective (Das et al. 1998), followed by a series of work (Sang Hyun, Wesley, and others 2001; Wu, Salzberg, and Zhang 2004). However, it has been widely accepted that few work dealing with real-valued series discovers valid rules for prediction. Because the symbolized series by common symbolization methods, including subsequence clustering and Piecewise Linear Approximation (PLA), cannot effectively represent the original series. (Struzik 2003; Keogh and Lin 2005; Shokoohi-Yekta et al. 2015). Eamonn Keogh et al. (2005) demonstrate that clustering of time series subsequences is meaningless, resulting from the independency between the clustering centers and the input. Shokoohi-Yekta et al. (2015) point out that two time series can be differ only by a small discrepancy, yet have completely different PLA representations. Therefore, the symbolized series is irrelevant with the original real-valued one and the failure of finding valid rules is doomed. The state-of-the-art work (Y15) directly manipulates the real-valued series (Shokoohi-Yekta et al. 2015). Their method is found on the assumption that a rule is contained in a subsequence, which splits a subsequence into the rule’s antecedent and consequent. Usually, there is an interval between a rule’s antecedent/consequent, and the splitting method will append the extra series to the antecedent/consequent. The complicated diversity of the intervals could result in rules with bad prediction performance. Besides, the splitting method cannot be applied to discover rules from two series either, since a motif is a subsequence in a single series. The difference between Y15 with our work can be seen from the perspective of the different usages of motifs. In Y15, it is noted that “In principle we could use a brute force search, testing all subsequences of T”, which means that it can work without motifs; but our method cannot work without motifs, since we are trying to relate pairs of them. Even if Y15 method uses motifs to “provide a tractable search”, it is 2 This experiment is performed on a single dataset, which is inadequate. We will refine it. Please see future work. still different from ours. As it mentioned, “a good rule candidate must be a time series motif in T”, Y15 treats motifs as rule candidates, while our method takes motifs as candidates of antecedents and consequents, as we noted that “we could utilize motif discovery methods to find all distinct motifs which accurately represent antecedent and consequent candidates of rules”. Algorithm 1: Find top-K rules. Input: TA and TB are two time series. Output: Res is K best rules. 1 2 3 4 Preliminaries In this paper, we consider the problem of rule discovery from real-valued time series for rule-based prediction. Without loss of generality, the rule discovery problem is exemplified using two time series TA and TB , since finding rules from over two time series can be treated pairwisely. It can also be applied to the situation of only one series by letting TA = TB . We begin with the definition of a rule and its instance. Definition 1 (Rule). A rule r is a 4-tuple (mA , mB , τ, θ). mA , mB are the subsequences of two real-valued time series TA , TB , respectively, and also the antecedent and consequent of the rule r. τ is a non-negative value3 indicating the max length of time interval between the mA and mB . θ is a non-negative value as the trigger threshold for firing the rule. Definition 2 (Instance). A rule r’s instance e is a 3-tuple (m˜A , m˜B , τ̃ ). m˜A is the mA -like subsequence observed in series TA subject to d(m˜A , mA ) < θ. m˜B is the mB -like subsequence observed later than m˜A in series TB . τ̃ is the time interval between m˜A and m˜B . By the definitions, given a rule r = (mA , mB , τ, θ), if a subsequence m˜A of TA is observed and d(m˜A , mA ) < θ, the rule r is fired and a subsequence of m˜B similar with mB is supposed to be observed from TB within the subsequent τ time interval4 . Notice that there are no constraints for τ̃ and d(m˜B , mB ) in Definition 2, since an instance with τ̃ > τ or d(m˜B , mB )  0 can be viewed as a “bad” instance for r. An example is shown in Figure 2. Intuitively, if a rule is good, it must have many supporting instances in the time series, and then m˜A and m˜B will be frequently occurring patterns. By the definition of the motif (Patel et al. 2002), they are actually the motifs of TA and TB , respectively. Inspired by that, we make the following assumption. Assumption 1. If r = (mA , mB , τ, θ) is a rule, then mA ∈ MA , mB ∈ MB , where MA and MB are motif sets of TA and TB . Based on the assumption, the rule discovery problem can be formulated as follows. Given two time series TA , TB with their motif sets MA and MB , find top-K rules r = (ma , mb , τ, θ), where ma ∈ MA , and mb ∈ MB . Efficient algorithms for motif discovery have already been developed, and we could choose the widely-applied MK algorithm (Mueen et al. 2009)5 . 3 In real applications, τ is a predefined value according to the series, which avoids nonsense rules with infinity max time interval. 4 In fact, m˜B is observed after τ̃ time. 5 For concise, the footnotes of all Ks are omitted throughout this paper. In fact, Ks can be different. 5 MA , MB ← motifs(TA ), motifs(TB ) MA , MB ← sortK(MA ), sortK(MB ) forall (mA , mB ) ∈ MA × MB do score(mA , mB , TA , TB ) Res ← topK score rules Directly, we present the top-level algorithm 1. It first finds the motif set for each series in line 1 by MK algorithm. Line 2 sorts and returns the top K motifs according to the domain knowledge of the series. Then it traverses every pair of motifs, and scores the corresponding rule in line 3 to 4. It finally returns the best K rules. Methodologies Given a pair of motif ma and mb , the score algorithm aims at evaluating a rule candidate r based on the instances in the training data. We will use the example in Figure 2 to illustrate. The scoring approach consists of three steps: 1. Find out all mA , mB -like patterns in TA , TB . In this step, we use the sliding window method to select similar patterns by setting a threshold. In Figure 2, three mA -like patterns m˜A (i) (i = 1, 2, 3) and four mB -like patterns m˜B (i) (i = 1, .., 4) are discovered; 2. Match mA , mB -like patterns into rule instances, and search for the matching result that can support the rule most. A brutal search is not only intractable but also lack of robustness. Instead, we propose a heuristic matching algorithm according to the belief that a rule is preferred when it has (1) many instances and (2) a short max-length time interval. The lines (both dotted and solid) in Figure 3 are all possible instances and the solid ones {e1,2 , e2,3 } are the best instances, because it has the most number of instances (i.e. 2), and the average length of time intervals is also the smallest. 3. Score each instance with respect to the rule r, and then integrate to the final score. In this step, instead of Euclidean distance, we follow Yekta’s scoring method and further consider the ratio of antecedents being matched, i.e. 2/3. In the example, the two best instances {e1,2 , e2,3 } are evaluated respectively based on MDL principle. The final score for r is the sum of {e1,2 , e2,3 }’s results multiplied by 2/3. Step 1. Motif-like Pattern Discovery MK algorithm (Mueen et al. 2009) returns pairs of subsequences which are very similar with each other as motifs. In this step, given a motif m (a pair of subsequences) of a time series T , we need to search for all m-like non-overlapped subsequences in T . A direct approach based on sliding window method is as follows: (1) search all subsequences of the same length with g Figure 3: The graph G for the example in Figure 2. M A = (1) (2) (2) (3) (4) g {m˜A , m˜A }, MB = {m˜B , m˜B , m˜B }, and E = {e1,2 , e1,3 , e1,4 , e2,3 , e2,4 }. The best subset S is {e1,2 , e2,3 }, i.e., the solid edges. The crossed edges e2,3 and e1,4 cannot be chosen for S at the same time, according to the parallel constraint. m and calculate the distance between them and m; (2) set an appropriate threshold θ0 to filter the subsequences with distance smaller than θ0 ; (3) sort the remaining subsequences by distance and remove the overlap ones with larger distances. After that, the motif-like patterns are chosen as the non-overlapped subsequences with small distance. The rule threshold θ is set as the threshold for selecting antecedent-like subsequences. The complexity of sliding window method is determined by the number of windows and the distance computation procedure. In this step, the complexity is O(|m| · |T |). Step 2. Heuristic Matching The patterns found in step 1 can be combined into pairs as instances of a rule. This step aims at finding the best instance set that support the rule, which should satisfy the following conditions: (1) its cardinality is the largest among all possible sets; (2) the average length of interjacent intervals of instances is the smallest. The two condition come from the belief about what a good rule is. Modeling. To formulate the problem concretely, we introduce the following notations and construct a weighted bipartite graph. (1) g g M ˜A (1) , ..., m˜A (p) }, M , ..., m˜B (q) } A = {m B = {m˜B are the sets containing all subsequences similar with the rule’s antecedent mA and consequent mB , respectively. E = {ei,j = (m˜A (i) , m˜B (j) )|1 ≤ i ≤ p, 1 ≤ j ≤ q, 0 < t(m˜B (j) ) − t(m˜A (i) ) < τ }, where function t(·) returns the occurrence time of the pattern. E is the set of all feasible instances, since the antecedent must appear before the consequent and the interval between them cannot be too large. It imposes a structure on the set E that given m˜A (i) , for ∀j such that τ > t(m˜B (j) ) − t(m˜A (i) ) > 0, then ei,j ∈ E. Besides, let wi,j = t(m˜B (j) ) − t(m˜A (i) ) measure the length of interjacent interval of the instance ei,j . g g M A , MB , E make up a weighted bipartite graph G = g g (M ∪ M A B , E). Figure 3 shows the graph G of the example in Figure 2. Optimization. Using the notation introduced above, we restate the heuristic matching process as : Given a nong g complete weighted graph G(M A ∪ MB , E), find the instance set S ⊂ E subject to (1) |S| is maximized, and (2) W (S), the total weight of S, is minimized. One cannot simply apply algorithms solving assignment problem due to a parallel constraint. Concretely speaking, for any two instances, if the antecedent-like pattern in one instance appears earlier than the other’s, then its corresponding consequent-like pattern must also come earlier than the other’s. In the graph, this constraint requires no crossed edges in S, as is illustrated in Figure 3. Suppose that the max |S| is known as s somehow, we can solve the following 0-1 integer programming problem: minimize X wi,j xi,j subject to X xi,j = s x X xi,j ≤ 1, i (1a) (1b) X xi,j ≤ 1 (1c) j xi,j + xk,l ≤ 1, ∀ i > k and l < j xi,j ∈ {0, 1} (1d) (1e) The optimization variables xi,j ’s are 0, 1 variables, constrained by (1e), each of which represents the selection of corresponding instance ei,j . (1b) restricts that |S| is maximized as s. (1c) requires that at most one edge can be chosen in the graph G with respect to the same vertex. (1d) refers to the parallel constraint. Now consider how to solve for s. It is not the classical problem of maximum unweighted bipartite matching due to the parallel constraint and therefore it cannot be easily solved by max/min flow algorithms. We formulate it as another optimization problem. maximize X subject to X x i xi,j xi,j ≤ 1, (2a) X xi,j ≤ 1 (2b) j xi,j + xk,l ≤ 1, ∀ i > k and l < j xi,j ∈ {0, 1} (2c) (2d) The optimization problems are both 0-1 integer programming, which are NP-hard generally. Existing solvers (e.g. Matlab Optimization Toolbox) based on cutting plane method can handle these problems within a tolerable time. Step 3. MDL-based Instance Scoring In this step, given the best instance set S, we first evaluate each instance by the similarity between it and the rule r made up by mA and mB , and then aggregate the results to the score of rule r for further comparison. We first introduce the MDL-based scoring method, which is initially proposed in Shokoohi-Yekta et al. 2015. Intuitively, the more similar the shape in the instance is with respect to the rule’s consequent, the better the instance can support the rule. The Euclidean distance is the most widely accepted measure for similarity. However, the length of consequent varies in different rules, where Euclidean metric cannot fairly judge the differences between subsequences of different length. Inspired by the Minimum Description Length principle that any regularity in a given set of data can be used to compress the data, i.e. to describe it using fewer symbols than needed to describe the data literally (Grünwald 2007), it is possible to take the rule’s consequent as the regularity and measure how many bits can be saved by compressing the shape in the instances according to the regularity using Huffman coding. A concrete example can be found in ShokoohiYekta et al. 2015. To use MDL principle, the series must be digitized first, and let dl(·) be the coding length. The digitization loses little information of the raw data according. The number of bits saved for instance e by encoding m˜B with respect to r’s consequent mB is as below: bit saved(e, r) = dl(e.m˜B ) − dl(e.m˜B |r.mB ) (3) The above is the original version of MDL-based scoring method developed by Yekta et al. We further take the ratio of antecedent-like pattern being matched into consideration. Intuitively, when the ratio is too small, indicating the number of matched instances is much less than the times that the antecedent is fired, the rule shouldn’t be considered a good rule. Therefore, the final score for a rule r is: score(r) = X
s bit saved(e, r) − dl(r.mB ) (4) g |M A| e∈S Experiment Evaluation We evaluate our method on real open datasets. Top rules discovered by our method and the baseline method from the same training data are compared and analyzed. In addition, we also validate the applicability of our method on multiple series. Experiment Setup The baseline method (Y15) is the state-of-the-art work by Yekta et al. The experiment is conducted on two open metering datasets. One is Almanac of Minutely Power dataset (AMPds), mainly containing electricity stream data at one minute intervals per meter for 2 years of monitoring in a single house (Makonin et al. 2013; 2016). The other is UK Domestic Appliance-Level Electricity (UK-DALE) dataset, which records both the whole-house mains power demand every six seconds as well as power demand from individual appliances every six seconds from five houses (Kelly and Knottenbelt 2015). Settings. Two groups of experiments are performed to (1) evaluate the prediction performance of our method and (2) validate the applicability of multiple series. • On Single Series. We utilize the aggregate power series of clothes washer and clothes dryer for 1 year from AMPds, which is also used by Y15 as the experiment dataset. The series of first month is used to discover rules, while the rest is used for testing. We select 5 top rules by each method and evaluate the prediction performance on the test data; • One Multiple Series. We attempt to discover rules from the separated power series of total 52 appliances from the house 1 of UK-DALE dataset, such as boiler, laptop, etc. We run our method on each pair of the series to search for valid rules. Metric. To measure the prediction performance of rules, we adopt the same metric Q proposed by Yekta et al. as: PN d(mB , ui ) Q(r) = Pi=1 , (5) N i=1 d(mB , vi ) where N is the total firing number of the rule r in the test data, and d(mB , x) is Euclidean distance between the consequent mB with the the shape beginning by position x. ui and vi are the i-th firing position and the i-th randomly chosen position, respectively. The denominator is used to normalize Q to a value between 0 and 1. The final Q is averaged after 1000 measurements. A smaller Q indicates a better prediction performance. The Q close to 1 suggests the rule r is no better than random guessing and Q close to 0 indicates that the rule r captures the structure in the data and predicts accurately. On Single Series To compare Y15 with our method, we select top 5 rules discovered by each method from the training data, and then evaluate them on the test data. Result. The top 5 ranked rules’ Qs are listed in Table 1. The top rules discovered by our methods are better than those by Y15. Specifically, our method outperforms Y15 by 23.9% on average. Comparison. To demonstrate the reason why the rules discovered by our method outperform those by Y15, we take a close look at the 5-th top rule, rY (by Y15) and rO (by this paper), and scrutinize their prediction results on the test data in Figure 4. As is shown in Figure 4a and 4b, the 5-th top rules rY and rO are quite different from each other though they are describing the same thing6 . rY comes from a splitting point at 10%-th of a 120-long subsequence, whose antecedent is 12 in length and consequent is 108 in length, whereas rO takes two motifs as its antecedent and consequent, whose lengths 6 Actually, they both imply the fact that the clothes dryer is often used after the clothes washer. Y15 MBP 1 2 3 4 5 Mean 0.389 0.340 0.436 0.299 0.398 0.337 0.481 0.310 0.424 0.341 0.426 0.324 Table 1: The prediction performance Q of top 5 rules on the test data. (a) The 5-th top rule rY by Y15 (b) The 5-th top rule rO by our method (c) A firing by rY (d) A firing by rO Figure 4: The two 5-th top rules with the prediction result around a same position. In (a) and (b), the red patterns are the antecedent of the rule, while the blue ones are the consequent. The rule’s threshold θ is set to 5 and max length time interval τ is set to 300 in both methods. (c) (d) depict the prediction results around the same position by both rules. The black curve is the real time series. The red curve shows the position where the rule is fired, while the blue curve is the best prediction during the max time interval. are 50 and 30 respectively. In contrast, rY has more reasonable antecedent and consequent. To illustrate, consider the case in Figure 4c and 4d. The antecedents of both rules present a good match and trigger both rules around the same position. However, rO gives an accurate prediction,while rY predicts a shape with a clear discrepancy to the real series. The interval before rY ’s consequent is so long that the consequent misses the best matching position. Intuitively, rY ’s consequent can be viewed as rO ’s consequent appended by a piece of noise series, which results in the mismatch of the consequent. The inaccurate prediction has its root in the splitting method of Y15, which inevitably adds some extra noises to the “real” antecedent/consequent, because the splitting method cannot position the boundaries of the antecedent/consequent. Our method, however, directly finds the key part of the series, i.e. motif, as the rule’s antecedent/consequent. Additionally, any rule discovered by Y15 is a split of a motif in this experiment. Since the split parts are also frequent patterns in the series, they can be discovered as motifs, i.e. the elements of MA and MB . Therefore, rules discovered by Y15 can also be found by our method. Discussion. In the electricity datasets, zero series7 is recognized as motif by MK algorithm because it is also a frequent pattern (though meaningless). To avoid such motifs, we sort the discovered motifs by the “roughness” of the shape and choose the top ones. Commonly, the sorting process, mentioned in the line 2 of Algorithm 1, is relevant with 7 The values in the series are almost all 0s, indicating no appliance is being used. Figure 5: A rule rE discovered in the series of hair dryer and straightener. The red curve is the antecedent, while the blue curve is the consequent. the characteristics of the time series. On Multiple Series We attempt to discover rules in multiple appliance series from the UK-DALE datasets, which include 52 kinds of appliances. The original data is sampled every six seconds, and we resample it per minute for efficiency consideration. Result. A serviceable rule rE discovered is from the power series of hair dryer and straighteners, the antecedent and consequent of which are shown in Figure 5. The rule rE describes the relationship between the usage of hair dryer and straightener. An interesting fact is that the rule’s antecedents and consequents are interchangeable, coinciding the common sense that the two appliances, hair dryer and straightener, are often used at the same time. To illustrate rE concretely, we list an overview of 4 power series in Figure 6, including hair dryer, straightener, breadmaker and laptop. The series of hair dryer and straightener are well matched, whereas the rest combinations are ranked References Figure 6: The overviews of four appliances’ power series. relatively lower. Conclusion and Discussion In this paper, we have introduced a novel rule-based prediction method for real-valued time series from the perspective of motifs. We preliminarily explore the possibility of relating two motifs as rule candidate. It first leverages motif discovery to segment the time series precisely for seeking recurring patterns as antecedents/consequents, and then investigates the underlying temporal relationships between motifs by combing motifs as rule candidates and ranking them based on the similarities. However, as is mentioned before, this work itself is incomplete and will be refined. We further consider the following two problems: First, current experiment mainly uses one kind of open dataset, i.e. household electricity usage. We will search for more open datasets to comprehensively evaluate the performance of our method. Second, in this work we evaluate each rule from the perspective of prediction. However, prediction is only a single aspect of a rule. We will try to develop more metrics that can reveal the inner connections within rules. Acknowlegements This work is supported by the program JS71-16-005. Brown, A. E.; Yemini, E. I.; Grundy, L. J.; Jucikas, T.; and Schafer, W. R. 2013. A dictionary of behavioral motifs reveals clusters of genes affecting caenorhabditis elegans locomotion. Proceedings of the National Academy of Sciences 110(2):791–796. Das, G.; Lin, K.-I.; Mannila, H.; Renganathan, G.; and Smyth, P. 1998. Rule discovery from time series. In KDD, volume 98, 16–22. Esling, P., and Agon, C. 2012. Time-series data mining. ACM Computing Surveys (CSUR) 45(1):12. Grünwald, P. D. 2007. The minimum description length principle. MIT press. Hamilton, J. D. 1994. Time series analysis, volume 2. Princeton university press Princeton. Han, J.; Pei, J.; and Kamber, M. 2011. Data mining: concepts and techniques. Elsevier. Kelly, J., and Knottenbelt, W. 2015. The UK-DALE dataset, domestic appliance-level electricity demand and whole-house demand from five UK homes. 2(150007). Keogh, E., and Lin, J. 2005. Clustering of time-series subsequences is meaningless: implications for previous and future research. Knowledge and information systems 8(2):154– 177. Makonin, S.; Popowich, F.; Bartram, L.; Gill, B.; and Bajic, I. V. 2013. Ampds: A public dataset for load disaggregation and eco-feedback research. In Electrical Power & Energy Conference (EPEC), 2013 IEEE, 1–6. IEEE. Makonin, S.; Ellert, B.; Bajic, I. V.; and Popowich, F. 2016. Electricity, water, and natural gas consumption of a residential house in Canada from 2012 to 2014. Scientific Data 3(160037):1–12. Makridakis, S.; Wheelwright, S. C.; and Hyndman, R. J. 2008. Forecasting methods and applications. John wiley & sons. Mueen, A.; Keogh, E.; Zhu, Q.; Cash, S.; and Westover, B. 2009. Exact discovery of time series motifs. In Proceedings of the 2009 SIAM International Conference on Data Mining, 473–484. SIAM. Mueen, A. 2014. Time series motif discovery: dimensions and applications. Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery 4(2):152–159. Patel, P.; Keogh, E.; Lin, J.; and Lonardi, S. 2002. Mining motifs in massive time series databases. In Data Mining, 2002. ICDM 2003. Proceedings. 2002 IEEE International Conference on, 370–377. IEEE. Sang Hyun, P.; Wesley, W.; et al. 2001. Discovering and matching elastic rules from sequence databases. Fundamenta Informaticae 47(1-2):75–90. Shokoohi-Yekta, M.; Chen, Y.; Campana, B.; Hu, B.; Zakaria, J.; and Keogh, E. 2015. Discovery of meaningful rules in time series. In Proceedings of the 21th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 1085–1094. ACM. Struzik, Z. R. 2003. Time series rule discovery: Tough, not meaningless. Lecture notes in computer science 32–39. Vahdatpour, A.; Amini, N.; and Sarrafzadeh, M. 2009. Toward unsupervised activity discovery using multidimensional motif detection in time series. In IJCAI, volume 9, 1261–1266. Wu, H.; Salzberg, B.; and Zhang, D. 2004. Online eventdriven subsequence matching over financial data streams. In Proceedings of the 2004 ACM SIGMOD international conference on Management of data, 23–34. ACM. 

arXiv:1402.2031v1 [cs.CV] 10 Feb 2014 Deeply Coupled Auto-encoder Networks for Cross-view Classification Wen Wang, Zhen Cui, Hong Chang, Shiguang Shan, Xilin Chen Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China {wen.wang, zhen.cui, hong.chang, shiguang.shan, xilin.chen}@vipl.ict.ac.cn November 2013 Abstract age retrieval, interests are taken in comparing two types of heterogeneous images, which may come from different views or even different sensors. Since the spanned feature spaces are quite different, it is very difficult to classify these images across views directly. To decrease the discrepancy across views, most of previous works endeavored to learn view-specific linear transforms and to project cross-view samples into a common latent space, and then employed these newly generated features for classification. The comparison of heterogeneous samples extensively exists in many applications, especially in the task of image classification. In this paper, we propose a simple but effective coupled neural network, called Deeply Coupled Autoencoder Networks (DCAN), which seeks to build two deep neural networks, coupled with each other in every corresponding layers. In DCAN, each deep structure is developed via stacking multiple discriminative coupled auto-encoders, a denoising auto-encoder trained with maximum margin criterion consisting of intra-class compactness and inter-class penalty. This single layer component makes our model simultaneously preserve the local consistency and enhance its discriminative capability. With increasing number of layers, the coupled networks can gradually narrow the gap between the two views. Extensive experiments on cross-view image classification tasks demonstrate the superiority of our method over state-of-the-art methods. 1 Though there are lots of approaches used to learn viewspecific projections, they can be divided roughly based on whether the supervised information is used. Unsupervised methods such as Canonical Correlation Analysis (CCA)[14] and Partial Least Square (PLS) [26] are employed to the task of cross-view recognition. Both of them attempt to use two linear mappings to project samples into a common space where the correlation is maximized, while PLS considers the variations rather than only the correlation in the target space. Besides, with use of the mutual information, a Coupled InformationTheoretic Encoding (CITE) method is developed to narrow the inter-view gap for the specific photo-sketch recognition task. And in [30], a semi-coupled dictionary is used to bridge two views. All the methods above consider to reduce the discrepancy between two views, however, the label information is not explicitly taken into account. With label information available, many methods were further developed to learn a discriminant common space For instance, Discriminative Canonical Correlation Analysis (DCCA) [16] is proposed as an extension of CCA. And Introduction Real-world objects often have different views, which might be endowed with the same semantic. For example, face images can be captured in different poses, which reveal the identity of the same object; images of one face can also be in different modalities, such as pictures under different lighting condition, pose, or even sketches from artists. In many computer vision applications, such as im1 In [22], with an additional local smoothness constraints, two linear projections are simultaneously learnt for Common Discriminant Feature Extraction (CDFE). There are also other such methods as the large margin approach [8] and the Coupled Spectral Regression (CSR) [20]. Recently, multi-view analysis [27, 15] is further developed to jointly learn multiple specific-view transforms when multiple views (usually more than 2 views) can be available. Although the above methods have been extensively applied in the cross-view problem, and have got encouraging performances, they all employed linear transforms to capture the shared features of samples from two views. However, these linear discriminant analysis methods usually depend on the assumption that the data of each class agrees with a Gaussian distribution, while data in real world usually has a much more complex distribution [33]. It indicates that linear transforms are insufficient to extract the common features of cross-view images. So it’s natural to consider about learning nonlinear features. A recent topic of interest in nonlinear learning is the research in deep learning. Deep learning attempts to learn nonlinear representations hierarchically via deep structures, and has been applied successfully in many computer vision problems. Classical deep learning methods often stack or compose multiple basic building blocks to yield a deeper structure. See [5] for a recent review of Deep Learning algorithms. Lots of such basic building blocks have been proposed, including sparse coding [19], restricted Boltzmann machine (RBM) [12], autoencoder [13, 6], etc. Specifically, the (stacked) autoencoder has shown its effectiveness in image denoising [32], domain adaptation [7], audio-visual speech classification [23], etc. As we all known, the kernel method, such as Kernel Canonical Correlation Analysis(Kernel CCA) [1], is also a widely used approach to learn nonlinear representations. Compared with the kernel method, deep learning is much more flexible and time-saving because the transform is learned rather than fixed and the time needed for training and inference process is beyond the limit of the size of training set. Inspired by the deep learning works above, we intend to solve the cross-view classification task via deep networks. It’s natural to build one single deep neural network with samples from both views, but this kind of network can’t handle complex data from totally different modalities and may suffer from inadequate representation capacity. Another way is to learn two different deep neural networks with samples of the different views. However, the two independent networks project samples from different views into different spaces, which makes comparison infeasible. Hence, building two neural networks coupled with each other seems to be a better solution. In this work, we propose a Deeply Coupled Autoencoder Networks(DCAN) method that learns the common representations to conduct cross-view classification by building two neural networks deeply coupled respectively, each for one view. We build the DCAN by stacking multiple discriminative coupled auto-encoders, a denoising auto-encoder with maximum margin criterion. The discriminative coupled auto-encoder has a similar input corrupted and reconstructive error minimized mechanism with the denoising auto-encoder proposed in [28], but is modified by adding a maximum margin criterion. This kind of criterion has been used in previous works, like [21, 29, 35], etc. Note that the counterparts from two views are added into the maximum margin criterion simultaneously since they both come from the same class, which naturally couples the corresponding layer in two deep networks. A schematic illustration can be seen in Fig.1. The proposed DCAN is related to Multimodal Autoencoders [23], Multimodal Restricted Boltzmann Machines and Deep Canonical Correlation Analysis [3]. The first two methods tend to learn a single network with one or more layers connected to both views and to predict one view from the other view, and the Deep Canonical Correlation Analysis build two deep networks, each for one view, and only representations of the highest layer are constrained to be correlated. Therefore, the key difference is that we learn two deep networks coupled with each other in representations in each layer, which is of great benefits because the DCAN not only learn two separate deep encodings but also makes better use of data from the both two views. What’s more, these differences allow for our model to handle the recognition task even when data is impure and insufficient. The rest of this paper is organized as follows. Section 2 details the formulation and solution to the proposed Deeply Coupled Auto-encoder Networks. Experimental results in Section 3 demonstrate the efficacy of the DCAN. In section 4 a conclusion is given. 2 2 Deeply Coupled Networks Auto-encoder min fx ,fy s.t. In this section, we first present the basic idea. The second part gives a detailed description of the discriminative coupled auto-encoder. Then, we describe how to stack multiple layers to build a deep network. Finally, we briefly describe the optimization of the model. 2.1 L(X, fx ) + L(Y, fy ) G1 (Hx , Hy ) − G2 (Hx , Hy ) ≤ ε, (1) (2) where X, Y denote inputs from the two views, and Hx , Hy denote hidden representations of the two views respectively. fx : X −→ Hx , fy : Y −→ Hy are the transforms we intend to learn, and we denote the reconstructive error as L(·), and maximum margin criterion as G1 (·) − G2 (·), which are described detailedly in the next subsection.ε is the threshold of the maximum margin criterion. Basic Idea As shown in Fig.1, the Deeply Coupled Auto-encoder Networks(DCAN) consists of two deep networks coupled with each other, and each one is for one view. The network structures of the two deep networks are just like the left-most and the right-most parts in Fig.1, where circles means the units in each layers (pixels in a input image for the input layer and hidden representation in higher layers), and arrows denote the full connections between adjacent layers. And the middle part of Fig.1 illustrates how the whole network projects samples in different views into a common space and gradually enhances the separability with increasing layers. 2.2 Discriminative coupled auto-encoder In the problem of cross-view, there are two types of heterogenous samples. Without loss of generality, we denote samples from one view as X = [x1 , · · · , xn ] , and those from the other view as Y = [y1 , · · · , yn ], in which n is the sample sizes. Noted that the corresponding labels are known, and Hx , Hy denote hidden representations of the two views we want to learn. The DCAN attempts to learn two nonlinear transforms fx : X −→ Hx and fy : Y −→ Hy that can project the samples from two views to one discriminant common space respectively, in which the local neighborhood relationship as well as class separability should be well preserved for each view. The auto-encoder like structure stands out in preserving the local consistency, and the denoising form enhances the robustness of learnt representations. However, the discrimination isn’t taken into consideration. Therefore, we modify the denoising autoencoder by adding a maximum margin criterion consisting of intra-class compactness and inter-class penalty. And the best nonlinear transformation is a trade-off between local consistency preserving and separability enhancing. Just like the one in denoising auto-encoder, the reconstructive error L(·) in Eq.(1) is formulated as follows: The two deep networks are both built through stacking multiple similar coupled single layer blocks because a single coupled layer might be insufficient, and the method of stacking multiple layers and training each layer greedily has be proved efficient in lots of previous works, such as those in [13, 6]. With the number of layers increased, the whole network can compactly represent a significantly larger set of transforms than shallow networks , and gradually narrow the gap with the discriminative capacity enhanced. We use a discriminative coupled auto-encoders trained with maximum margin criterion as a single layer component. Concretely, we incorporate the additional noises in the training process while maximizing the margin criterion, which makes the learnt mapping more stable as well as discriminant. Note that the maximum margin criterion also works in coupling two corresponding layers. Formally, the discriminative coupled auto-encoder can be written as follows: L(X, Θ) = X Ex̃∼P (x̃|x) kx̂ − xk (3) Eỹ∼P (ỹ|y) kŷ − yk (4) x∈X p L(Y, Θ) = X y∈Y p 3 Figure 1: An illustration of our proposed DCAN. The left-most and right-most schematic show the structure of the two coupled network respectively. And the schematic in the middle illustrates how the whole network gradually enhances the separability with increasing layers, where pictures with solid line border denote samples from view 1, those with dotted line border denote samples from view 2, and different colors imply different subjects. 4 where E calculates the expectation over corrupted ver- which can be formulated as follows, sions X̃, Ỹ of examples X, Y obtained from a corruption P 1 process P (x̃|x), P (ỹ|y). Θ = {Wx , Wy , bx , by , cx , cy } khi − hj k2 , (9) G2 (H) = 2N 2 Ii ,Ij ∈D specifies the two nonlinear transforms fx , fy , where Ij ∈KN N (Ii ) Wx , Wy is the weight matrix, and bx , by , cx , cy are the bias of encoder and decoder respectively, and X̂, Ŷ are where I belongs to the k nearest neighbors of I with j i calculated through the decoder process : different class labels, and N2 is the number of all pairs satisfying the condition. T And the function of G1 (H), G2 (H) is illustrated in the X̂ = s(Wx Hx + cx ) (5) middel part of Fig.1. In the projected common space deŶ = s(WyT Hy + cy ) noted by S, the compactness term G1 (·) shown by red ellipse works by pulling intra-class samples together while And hidden representations Hx , Hy are obtained from the penalty term G2 (·) shown by black ellipse tend to push the encoder that is a similar mapping with the decoder, adjacent inter-class samples away. Finally, by solving the optimization problem Eq.(1), we can learn a couple of nonlinear transforms fx , fy to transHx = s(Wx X̃ + bx ) (6) form the original samples from both views into a common Hy = s(Wy Ỹ + by ) space. where s is the nonlinear activation function, such as the point-wise hyperbolic tangent operation on linear projected features, i.e., s(x) = eax − e−ax eax + e−ax 2.3 Stacking coupled auto-encoder Through the training process above, we model the map between original sample space and a preliminary discriminant subspace with gap eliminated, and build a hidden representation H which is a trade-off between approximate preservation on local consistency and the distinction of the projected data. But since real-world data is highly complicated, using a single coupled layer to model the vast and complex real scenes might be insufficient. So we choose to stack multiple such coupled network layers described in subsection 2.2. With the number of layers increased, the whole network can compactly represent a significantly larger set of transforms than shallow networks, and gradually narrow the gap with the discriminative ability enhanced. Training a deep network with coupled nonlinear transforms can be achieved by the canonical greedy layer-wise approach [12, 6]. Or to be more precise, after training a single layer coupled network, one can compute a new feature H by the encoder in Eq.(6) and then feed it into the next layer network as the input feature. In practice, we find that stacking multiple such layers can gradually reduce the gap and improve the recognition performance (see Fig.1 and Section 3). (7) in which a is the gain parameter. Moreover, for the maximum margin criterion consisting of intra-class compactness and inter-class penalty, the constraint term G1 (·) − G2 (·) in Eq.(1) is used to realize coupling since samples of the same class are treated similarly no matter which view they are from. Assuming S is the set of sample pairs from the same class, and D is the set of sample pairs from different classes. Note that the counterparts from two views are naturally added into S, D since it’s the class rather than the view that are considered. Then, we characterize the compactness as follows, P 1 G1 (H) = 2N khi − hj k2 , (8) 1 Ii ,Ij ∈S where hi denotes theTcorresponding hidden representation of an input Ii ∈ X Y and is a sample from either view 1 or view 2, and N1 is the size of S. Meanwhile, the goal of the inter-class separability is to push the adjacent samples from different classes far away, 5 7 poses (−45◦ , −30◦ , −15◦ , 0◦ , 15◦ , 30◦ , 45◦ ), 3 expression (Neutral,Smile, Disgust), no flush illumination from We adopt the Lagrangian multiplier method to solve the 4 sessions are selected to validate our method. We ranobjective function Eq.(1) with the constraints Eq.(2) as domly choose 4 images for each pose of each subject, then follows: randomly partition the data into two parts: the training set with 231 subjects (i.e., 231 × 7 × 4 = 6468 images) and the testing set with the rest subjects. min λ(L(X, Θ) + L(Y, Θ)) + (G1 (H) − G2 (H))+ Θ CUHK Face Sketch FERET (CUFSF) dataset [34, 1 1 31] contains two types of face images: photo and sketch. 2 2 γ( kWx kF + kWy kF ) Total 1,194 images (one image per subject) were collected 2 2 (10) with lighting variations from FERET dataset [25]. For each subject, a sketch is drawn with shape exaggeration. where the first term is the the reconstruction error, the According to the configuration of [15], we use the first second term is the maximum margin criterion, and the last 700 subjects as the training data and the rest subjects as term is the shrinkage constraints called the Tikhonov reg- the testing data. ularizers in [11], which is utilized to decrease the magnitude of the weights and further to help prevent over-fitting. λ is the balance parameter between the local consistency 3.2 Settings and empirical separability. And γ is called the weight de- All images from Multi-PIE and CUFSF are cropped into cay parameter and is usually set to a small value, e.g., 64×80 pixels without any preprocess. We compare the 1.0e-4. proposed DCAN method with several baselines and stateTo optimize the objective function (10), we use back- of-the-art methods, including CCA [14], Kernel CCA [1], propagation to calculate the gradient and then employ the Deep CCA [3], FDA [4], CDFE [22], CSR [20], PLS limited-memory BFGS (L-BFGS) method [24, 17], which [26] and MvDA [15]. The first seven methods are pairis often used to solve nonlinear optimization problems wise methods for cross-view classification. MvDA jointly without any constraints. L-BFGS is particularly suitable learns all transforms when multiple views can be utilized, for problems with a large amount of variables under the and has achieved the state-of-the-art results in their remoderate memory requirement. To utilize L-BFGS, we ports [15]. need to calculate the gradients of the object function. ObThe Principal Component Analysis (PCA) [4] is used viously, the object function in (10) is differential to these for dimension reduction. In our experiments, we set the parameters Θ, and we use Back-propagation [18] method default dimensionality as 100 with preservation of most to derive the derivative of the overall cost function. In our energy except Deep CCA, PLS, CSR and CDFE, where setting, we find the objective function can achieve as fast the dimensionality are tuned in [50,1000] for the best perconvergence as described in [17]. formance. For all these methods, we report the best performance by tuning the related parameters according to their papers. Firstly, for Kernel CCA, we experiment with 3 Experiments Gaussian kernel and polynomial kernel and adjust the parameters to get the best performance. Then for Deep CCA In this section, the proposed DCAN is evaluated on two [3], we strictly follow their algorithms and tune all possidatasets, Multi-PIE [9] and CUHK Face Sketch FERET ble parameters, but the performance is inferior to CCA. (CUFSF) [34, 31]. One possible reason is that Deep CCA only considers the correlations on training data (as reported in their paper) so that the learnt mode overly fits the training data, which 3.1 Databases thus leads to the poor generality on the testing set. BeMulti-PIE dataset [9] is employed to evaluate face recog- sides, the parameter α and β are respectively traversed in nition across pose. Here a subset from the 337 subjects in [0.2,2] and [0.0001,1] for CDFE, the parameter λ and η 2.4 Optimization 6 Method Accuracy CCA[14] KernelCCA[10] DeepCCA[3] FDA[4] CDFE[22] CSR[20] PLS[26] MvDA[15] 0.698 0.840 0.599 0.814 0.773 0.580 0.574 0.867 DCAN-1 DCAN-2 DCAN-3 DCAN-4 0.830 0.877 0.884 0.879 Table 1: Evaluation on Multi-PIE database in terms of mean accuracy. DCAN-k means a stacked k-layer network. Figure 2: After learning common features by the crossview methods, we project the features into 2-D space by using the principal two components in PCA. The depicted samples are randomly chosen form Multi-PIE [9] dataset. are searched in [0.001,1] for CSR, and the reduced dimen- The “◦” and “+” points come from two views respecsionality is tuned for CCA, PLS, FDA and MvDA. tively. Different color points belong to different classes. As for our proposed DCAN, the performance on DCAN-k is our proposed method with a stacked k-layer CUFSF database of varied parameters, λ, k, is shown in neural network. Fig.3. In following experiments, we set λ = 0.2, γ = 1.0e − 4, k = 10 and a = 1. With increasing layers, the number of hidden neurons are gradually reduced by 10, on Multi-PIE data set. Since the images are acquired over i.e., 90, 80, 70, 60 if four layers. seven poses on Multi-PIE data set, in total 7 × 6 = 42 comparison experiments need to be conducted. The detailed results are shown in Table 2,where two poses are 3.3 Face Recognition across Pose used as the gallery and probe set to each other and the First, to explicitly illustrate the learnt mapping, we con- rank-1 recognition rate is reported. Further, the mean acduct an experiment on Multi-PIE dataset by projecting curacy of all pairwise results for each methods is also rethe learnt common features into a 2-D space with Princi- ported in Table 1. pal Component Analysis (PCA). As shown in Fig.2. The From Table 1, we can find the supervised methods exclassical method CCA can only roughly align the data cept CSR are significantly superior to CCA due to the in the principal directions and the state-of-the-art method use of the label information. And nonlinear methods MvDA [15] attempts to merge two types of data but seems except Deep CCA are significantly superior to the nonto fail. Thus, we argue that linear transforms are a little linear methods due to the use of nonlinear transforms. stiff to convert data from two views into an ideal com- Compared with FDA, the proposed DCAN with only one mon space. The three diagrams below shows that DCAN layer network can perform better with 1.6% improvement. can gradually separate samples from different classes with With increasing layers, the accuracy of DCAN reaches the increase of layers, which is just as we described in the a climax via stacking three layer networks. The reason above analysis. of the degradation in DCAN with four layers is mainly Next, we compare our methods with several state-of- the effect of reduced dimensionality, where 10 dimenthe-art methods for the cross-view face recognition task sions are cut out from the above layer network. Obvi7 −45◦ −30◦ −15◦ −45◦ −30◦ −15◦ 0◦ 15◦ 30◦ 45◦ 1.000 0.816 0.588 0.473 0.473 0.515 0.511 0.816 1.000 0.858 0.611 0.664 0.553 0.553 0.588 0.858 1.000 0.894 0.807 0.602 0.447 0◦ 15◦ 30◦ 0.473 0.611 0.894 1.000 0.909 0.604 0.484 0.473 0.664 0.807 0.909 1.000 0.874 0.602 0.515 0.553 0.602 0.604 0.874 1.000 0.768 45◦ −45◦ −30◦ −15◦ 0.511 −45◦ 0.553 −30◦ 0.447 −15◦ 0.484 0◦ 0.602 15◦ 0.768 30◦ 1.000 45◦ (a) CCA, Ave = 0.698 −45◦ −30◦ −15◦ −45◦ −30◦ −15◦ 0◦ 15◦ 30◦ 45◦ 1.000 0.854 0.598 0.425 0.473 0.522 0.523 0.854 1.000 0.844 0.578 0.676 0.576 0.566 0.598 0.844 1.000 0.806 0.807 0.602 0.424 0◦ 0.425 0.578 0.806 1.000 0.911 0.599 0.444 30◦ 0.522 0.576 0.602 0.599 0.866 1.000 0.756 45◦ −45◦ −30◦ −15◦ 0◦ 15◦ 30◦ 45◦ 1.000 0.854 0.714 0.595 0.557 0.633 0.608 0.854 1.000 0.867 0.746 0.688 0.697 0.606 0.714 0.867 1.000 0.887 0.808 0.704 0.579 0.523 0.566 −30◦ 0.424 −15◦ 0.444 0◦ 0.624 15◦ 0.756 30◦ 1.000 45◦ 1.000 0.847 0.754 0.686 0.573 0.610 0.664 −45◦ −30◦ −15◦ 0◦ 15◦ 30◦ 45◦ 1.000 0.856 0.807 0.757 0.688 0.708 0.719 0.872 1.000 0.874 0.854 0.777 0.735 0.715 0.819 0.881 1.000 0.896 0.854 0.788 0.697 30◦ 45◦ 0.756 0.858 0.911 1.000 0.938 0.759 0.759 0.706 0.808 0.880 0.938 1.000 0.922 0.845 0.726 0.801 0.861 0.759 0.922 1.000 0.912 0.737 0.757 0.765 0.759 0.845 0.912 1.000 0.847 1.000 0.911 0.847 0.807 0.766 0.635 0.754 0.911 1.000 0.925 0.896 0.821 0.602 0◦ 15◦ 30◦ 45◦ 0.686 0.847 0.925 1.000 0.964 0.872 0.684 0.573 0.807 0.896 0.964 1.000 0.929 0.768 0.610 0.766 0.821 0.872 0.929 1.000 0.878 0.664 0.635 0.602 0.684 0.768 0.878 1.000 (d) FDA, Ave = 0.814 0◦ 15◦ 30◦ 0.595 0.746 0.887 1.000 0.916 0.819 0.651 0.557 0.688 0.808 0.916 1.000 0.912 0.754 0.633 0.697 0.704 0.819 0.912 1.000 0.850 45◦ −45◦ −30◦ −15◦ 0.608 −45◦ 0.606 −30◦ 0.579 −15◦ 0.651 0◦ 0.754 15◦ 0.850 30◦ 1.000 45◦ 1.000 0.914 0.854 0.763 0.710 0.770 0.759 (e) CDFE, Ave = 0.773 −45◦ −30◦ −15◦ 0.810 0.892 1.000 0.911 0.880 0.861 0.765 −45◦ −30◦ −15◦ −45◦ (c) DeepCCA, Ave = 0.599 −45◦ −30◦ −15◦ 0.878 1.000 0.892 0.858 0.808 0.801 0.757 15◦ (b) KernelCCA, Ave = 0.840 15◦ 0.473 0.676 0.807 0.911 1.000 0.866 0.624 1.000 0.878 0.810 0.756 0.706 0.726 0.737 0◦ 0.914 1.000 0.947 0.858 0.812 0.861 0.766 0.854 0.947 1.000 0.923 0.880 0.894 0.775 0◦ 15◦ 30◦ 45◦ 0.763 0.858 0.923 1.000 0.938 0.900 0.750 0.710 0.812 0.880 0.938 1.000 0.923 0.807 0.770 0.861 0.894 0.900 0.923 1.000 0.934 0.759 0.766 0.775 0.750 0.807 0.934 1.000 (f) MvDA, Ave = 0.867 0◦ 15◦ 30◦ 0.730 0.825 0.869 1.000 0.916 0.834 0.752 0.655 0.754 0.865 0.938 1.000 0.918 0.832 0.708 0.737 0.781 0.858 0.900 1.000 0.909 45◦ −45◦ −30◦ −15◦ 0.686 −45◦ 0.650 −30◦ 0.681 −15◦ 0.790 0◦ 0.823 15◦ 0.916 30◦ 1.000 45◦ (g) DCAN-1, Ave = 0.830 1.000 0.927 0.867 0.832 0.765 0.779 0.794 0.905 1.000 0.929 0.876 0.865 0.832 0.777 0.876 0.954 1.000 0.925 0.907 0.870 0.785 0◦ 15◦ 30◦ 45◦ 0.783 0.896 0.905 1.000 0.951 0.916 0.812 0.714 0.850 0.905 0.958 1.000 0.945 0.876 0.779 0.825 0.867 0.896 0.929 1.000 0.938 0.796 0.730 0.757 0.808 0.874 0.949 1.000 (h) DCAN-3, Ave = 0.884 Table 2: Results of CCA, FDA [4], CDFE [22], MvDA [15] and DCAN on MultiPIE dataset in terms of rank-1 recognition rate. DCAN-k means a stacked k-layer network. Due to space limitation, the results of other methods cannot be reported here, but their mean accuracies are shown in Table 1. 8 Method Photo-Sketch Sketch-Photo CCA[14] KernelCCA[10] DeepCCA[3] CDFE[22] CSR[20] PLS[26] FDA[4] MvDA[15] 0.387 0.466 0.364 0.456 0.502 0.486 0.468 0.534 0.475 0.570 0.434 0.476 0.590 0.510 0.534 0.555 DCAN-1 DCAN-2 DCAN-3 0.535 0.603 0.601 0.555 0.613 0.652 67 70 60 65 64 63 62 61 55 0.01 0.2 0.4 λ (a) 0.6 0.8 1 60 2 4 6 8 10 12 k (b) Figure 3: The performance with varied parameter values for our proposed DCAN. The sketch and photo images in CUFSF [34, 31] are respectively used for the gallery and probe set. (a) Varied λ with fixed k = 10. (b) Varied k with fixed λ = 0.2. Table 3: Evluation on CUFSF database in terms of mean accuracy. DCAN-k means a stacked k-layer network. on a small sample size. The reasons lie in three folds: ously, compared with two-view based methods, the proposed DCAN with three layers improves the performance greatly (88.4% vs. 81.4%). Besides, MvDA also achieves a considerably good performance by using all samples from all poses. It is unfair to compare these two-view based methods (containing DCAN) with MvDA, because the latter implicitly uses additional five views information except current compared two views. But our method performs better than MvDA, 88.4% vs. 86.7%. As observed in Table 2, three-layer DCAN achieves a largely improvement compared with CCA,FDA,CDFE for all cross-view cases and MvDA for most of cross-view cases. The results are shown in Table 2 and Table 1. 3.4 Accuracy (%) Accuracy(%) 66 65 (1) The maximum margin criterion makes the learnt mapping more discriminative, which is a straightforward strategy in the supervised classification task. (2) Auto-encoder approximately preserves the local neighborhood structures. For this, Alain et al. [2] theoretically prove that the learnt representation by auto-encoder can recover local properties from the view of manifold. To further validate that, we employ the first 700 photo images from CUFSF database to perform the nonlinear self-reconstruction with auto-encoder. With the hidden presentations, we find the local neighbors with 1,2,3,4,5 neighbors can be preserved with the probability of 99.43%, 99.00%, 98.57%, 98.00% and 97.42% respectively. Thus, the use of auto-encoder intrinsically reduces the complexity of the discriminant model, which further makes the learnt model better generality on the testing set. Photo-Sketch Recognition Photo-Sketch recognition is conducted on CUFSF dataset. The samples come from only two views, photo and sketch. The comparison results are provided in Table 3. As shown in this table, since only two views can be utilized in this case, MvDA degrades to a comparable performance with those previous two-view based meth- (3) The deep structure generates a gradual model, which ods. Our proposed DCAN with three layer networks makes the learnt transform more robust. With only can achieve even better with more than 6% improvement, one layer, the model can’t represent the complex data which further indicates DCAN benefits from the nonlinvery well. But with layers goes deeper, the coupled ear and multi-layer structure. networks can learn transforms much more flexible and hence can be allowed to handle more complex Discussion and analysis: The above experiments data. demonstrate that our methods can work very well even 9 4 Conclusion In this paper, we propose a deep learning method, the Deeply Coupled Auto-encoder Networks(DCAN), which can gradually generate a coupled discriminant common representation for cross-view object classification. In each layer we take both local consistency and discrimination of projected data into consideration. By stacking multiple such coupled network layers, DCAN can gradually improve the learnt shared features in the common space. Moreover, experiments in the cross-view classification tasks demonstrate the superior of our method over other state-of-the-art methods. References [1] S. Akaho. A kernel method for canonical correlation analysis, 2006. 2, 6 [2] G. Alain and Y. Bengio. What regularized auto-encoders learn from the data generating distribution. arXiv preprint arXiv:1211.4246, 2012. 9 [3] G. Andrew, R. Arora, J. Bilmes, and K. Livescu. Deep canonical correlation analysis. 2, 6, 7, 9 [4] P. N. Belhumeur, J. P. Hespanha, and D. J. Kriegman. Eigenfaces vs. fisherfaces: Recognition using class specific linear projection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 19(7):711–720, 1997. 6, 7, 8, 9 [5] Y. Bengio, A. Courville, and P. Vincent. Representation learning: A review and new perspectives. 2013. 2 [6] Y. Bengio, P. Lamblin, D. Popovici, and H. Larochelle. Greedy layer-wise training of deep networks, 2007. 2, 3, 5 [7] M. Chen, Z. Xu, K. Weinberger, and F. Sha. Marginalized denoising autoencoders for domain adaptation, 2012. 2 [8] N. Chen, J. Zhu, and E. P. Xing. Predictive subspace learning for multi-view data: a large margin approach, 2010. 2 [9] R. Gross, I. Matthews, J. Cohn, T. Kanade, and S. Baker. The cmu multi-pose, illumination, and expression (multipie) face database, 2007. 6, 7 [10] D. R. Hardoon, S. Szedmak, and J. Shawe-Taylor. Canonical correlation analysis: An overview with application to learning methods. Neural Computation, 16(12):2639– 2664, 2004. 7, 9 [11] T. Hastie, R. Tibshirani, and J. J. H. Friedman. The elements of statistical learning, 2001. 6 [12] G. E. Hinton, S. Osindero, and Y.-W. Teh. A fast learning algorithm for deep belief nets. Neural computation, 18(7):1527–1554, 2006. 2, 5 10 [13] G. E. Hinton and R. R. Salakhutdinov. Reducing the dimensionality of data with neural networks. Science, 313(5786):504–507, 2006. 2, 3 [14] H. Hotelling. Relations between two sets of variates. Biometrika, 28(3/4):321–377, 1936. 1, 6, 7, 9 [15] M. Kan, S. Shan, H. Zhang, S. Lao, and X. Chen. Multiview discriminant analysis. pages 808–821, 2012. 2, 6, 7, 8, 9 [16] T.-K. Kim, J. Kittler, and R. Cipolla. Discriminative learning and recognition of image set classes using canonical correlations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(6):1005–1018, 2007. 1 [17] Q. V. Le, J. Ngiam, A. Coates, A. Lahiri, B. Prochnow, and A. Y. Ng. On optimization methods for deep learning, 2011. 6 [18] Y. LeCun, L. Bottou, G. B. Orr, and K.-R. Müller. Efficient backprop. In Neural networks: Tricks of the trade, pages 9–50. Springer, 1998. 6 [19] H. Lee, A. Battle, R. Raina, and A. Y. Ng. Efficient sparse coding algorithms, 2007. 2 [20] Z. Lei and S. Z. Li. Coupled spectral regression for matching heterogeneous faces, 2009. 2, 6, 7, 9 [21] H. Li, T. Jiang, and K. Zhang. Efficient and robust feature extraction by maximum margin criterion. Neural Networks, IEEE Transactions on, 17(1):157–165, 2006. 2 [22] D. Lin and X. Tang. Inter-modality face recognition. pages 13–26, 2006. 1, 6, 7, 8, 9 [23] J. Ngiam, A. Khosla, M. Kim, J. Nam, H. Lee, and A. Y. Ng. Multimodal deep learning, 2011. 2 [24] J. Nocedal and S. J. Wright. Numerical optimization, 2006. 6 [25] P. J. Phillips, H. Wechsler, J. Huang, and P. J. Rauss. The feret database and evaluation procedure for facerecognition algorithms. Image and vision computing, 16(5):295–306, 1998. 6 [26] A. Sharma and D. W. Jacobs. Bypassing synthesis: Pls for face recognition with pose, low-resolution and sketch, 2011. 1, 6, 7, 9 [27] A. Sharma, A. Kumar, H. Daume, and D. W. Jacobs. Generalized multiview analysis: A discriminative latent space, 2012. 2 [28] P. Vincent, H. Larochelle, Y. Bengio, and P.-A. Manzagol. Extracting and composing robust features with denoising autoencoders, 2008. 2 [29] F. Wang and C. Zhang. Feature extraction by maximizing the average neighborhood margin. In Computer Vision and Pattern Recognition, 2007. CVPR’07. IEEE Conference on, pages 1–8. IEEE, 2007. 2 [30] S. Wang, L. Zhang, Y. Liang, and Q. Pan. Semicoupled dictionary learning with applications to image super-resolution and photo-sketch synthesis, 2012. 1 [31] X. Wang and X. Tang. Face photo-sketch synthesis and recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 31(11):1955–1967, 2009. 6, 9 [32] J. Xie, L. Xu, and E. Chen. Image denoising and inpainting with deep neural networks, 2012. 2 [33] S. Yan, D. Xu, B. Zhang, H.-J. Zhang, Q. Yang, and S. Lin. Graph embedding and extensions: a general framework for dimensionality reduction. IEEE Transactions on Pattern Analysis and Machine Intelligence, 29(1):40–51, 2007. 2 [34] W. Zhang, X. Wang, and X. Tang. Coupled informationtheoretic encoding for face photo-sketch recognition, 2011. 6, 9 [35] B. Zhao, F. Wang, and C. Zhang. Maximum margin embedding. In Data Mining, 2008. ICDM’08. Eighth IEEE International Conference on, pages 1127–1132. IEEE, 2008. 2 11 

arXiv:1607.00062v1 [math.AG] 30 Jun 2016 LOCAL COHOMOLOGY AND BASE CHANGE KAREN E. SMITH f Abstract. Let X −→ S be a morphism of Noetherian schemes, with S reduced. For any closed subscheme Z of X finite over S, let j denote the open immersion X \ Z ֒→ X. Then for any coherent sheaf F on X \ Z and any index r ≥ 1, the sheaf f∗ (Rr j∗ F ) is generically free on S and commutes with base change. We prove this by proving a related statement about local cohomology: Let R be Noetherian algebra over a Noetherian domain A, and let I ⊂ R be an ideal such that R/I is finitely generated as an A-module. Let M be a finitely generated R-module. Then there exists a non-zero g ∈ A such that the local cohomology modules HIr (M )⊗A Ag are free over Ag and for any ring map A → L factoring through Ag , we have r HIr (M ) ⊗A L ∼ (M ⊗A L) for all r. = HI⊗ AL 1. Introduction In his work on maps between local Picard groups, Kollár was led to investigate the behavior of certain cohomological functors under base change [Kol]. The following theorem directly answers a question he had posed: f Theorem 1.1. Let X → S be a morphism of Noetherian schemes, with S reduced. Suppose that Z ⊂ X is closed subscheme finite over S, and let j denote the open embedding of its complement U. Then for any coherent sheaf F on U, the sheaves f∗ (Rr j∗ F ) are generically free and commute with base change for all r ≥ 1. Our purpose in this note is to prove this general statement. Kollár himself had proved a special case of this result in a more restricted setting [Kol, Thm 78]. We pause to say precisely what is meant by generically free and commutes with base change. Suppose H is a functor which, for every morphism of schemes X → S and every quasi-coherent sheaf F on X, produces a quasi-coherent sheaf H(F ) on S. We say H(F ) is generically free if there exists a dense open set S 0 of S over which Date: July 4, 2016. Thanks to János Kollár for encouraging me to write this paper, for his insistence that I establish the more general version of the theorem, and for sharing his insight as to why my arguments should go through more generally, especially the idea to reduce to the affine case in Section 2. I’d also like to thank Mel Hochster for listening to my arguments, Karl Schwede for reading a draft and suggesting the reference [DGI], and Barghav Bhatt for his interest. Financial support partially from NSF DMS-1501625. 1 2 KAREN E. SMITH p the OS -module H(F ) is free. If in addition, for every change of base T −→ S 0 , the natural map p∗ H(F ) → H(p∗X F ) of quasi-coherent sheaves on T is an isomorphism (where pX is the induced morphism X ×S T → X), then we say that H(F ) is generically free and commutes with base change. See [Kol, §72]. Remark 1.2. We do not claim the r = 0 case of Theorem 1.1; in fact, it is false. For a counterexample, consider the ring homomorphism splitting Z ֒→ Z × Q ։ Z. The corresponding morphism of Noetherian schemes Z = Spec(Z) ֒→ X = Spec(Z × Q) → S = Spec Z satisfies the hypothesis of Theorem 1.1. The open set U = X \ Z is the component Spec Q of X. The coherent sheaf determined by the module Q on U is not generically free over Z, since there is no open affine subset Spec Z[ n1 ] over which Q is a free module. [In this case, the map j is affine, so the higher direct image sheaves Rp j∗ F all vanish for p > 0.] On the other hand, if f is a map of finite type, then the r = 0 case of Theorem 1.1 can be deduced from Grothendiecks’s Lemma on Generic freeness; see Lemma 4.1. For the commutative algebraists, we record the following version of the main result, which is essentially just the statement in the affine case: Corollary 1.3. Let A be a reduced Noetherian ring. Let R be a Noetherian Aalgebra with ideal I ⊂ R such that the induced map A → R/I is finite. Then for any Noetherian R module M, the local cohomology modules HIi (M) are generically free and commute with base change over A for all i ≥ 0. Explicitly, this means that there exists element g not in any minimal prime of A such that the modules HIi (M) ⊗A Ag are free over Ag , and that for any algebra L over Ag , the natural map HIi (M) ⊗A L → HIi (M ⊗A L) is an isomorphism. Some version of Theorem 1.1 may follow from known statements in the literature, but looking through works of Grothendieck ([RD], [LC], [SGA2]) and [Conrad], I am not able to find it; nor presumably could Kollár. After this paper was written, I did find a related statement due to Hochster and Roberts [HR, Thm 3.4] in a special case, not quite strong enough to directly answer Kollár’s original question; furthermore, my proof is different and possibly of independent interest. In any case, there may be value in the self-contained proof here, which uses a relative form of Matlis duality proved here using only basic results about local cohomology well-known to most commutative algebraists. LOCAL COHOMOLOGY AND BASE CHANGE 3 2. Restatement of the Problem In this section, we show that Theorem 1.1 reduces to the following special statement: Theorem 2.1. Let A be a Noetherian domain. Let R = A[[x1 , . . . , xn ]] be a power series ring over A, and M a finitely generated R-module. Denote by I the ideal (x1 , . . . , xn ) ⊂ R. Then the local cohomology modules HIi (M) are generically free over A and commute with base change for all i. For basic definitions and properties of local cohomology modules, we refer to [LC]. For the remainder of this section, we show how Theorem 1.1 and Corollary 1.3 follow from Theorem 2.1. First, Theorem 1.1 is local on the base. Because the scheme S is reduced, it is the finite union of its irreducible components, each of which is reduced and irreducible, so it suffices to prove the result on each of them. Thus we can immediately reduce to the case where S = Spec A, for some Noetherian domain A. We now reduce to the case where X is affine as well. The coherent sheaf F on the open set U extends to a coherent sheaf on X, which we also denote by F . To simplify notation, let us denote the sheaf Rr j∗ F by H, which we observe vanishes outside the closed set Z. Each section is annihilated by a power of the ideal IZ of Z, so that although the sheaf of abelian groups H on Z is not an OZ -module, it does have the structure of a module over the sheaf of OS -algebras limt OI Xt , which we ←− Z bZ \ denote OX,Z ; put differently, H can be viewed as sheaf on the formal scheme X 0 i [ [ over S. Observe that i∗ O X,Z |X 0 = OX,Z , where X ֒→ X is the inclusion of any open set X 0 ⊂ X containing the generic points of all the components of Z. This means that H is generically free as an OS -module if and only if H|X 0 is, and there is no loss of generality in replacing X by any such open set X0 . In particular, we can choose such X 0 to be affine, thus reducing the proof of Theorem 1.1 to the case where both X and S are affine. We can now assume that X → S is the affine map of affine schemes corresponding to a ring homomorphism A → T . In this case the closed set Z is defined by some ideal I of T . Because Z is finite over S = Spec A, the composition A → T → T /I is a finite integral extension of A. The coherent sheaf F on U extends to a coherent sheaf F on X, which corresponds to a finitely generated T -module M. Since X = Spec T is affine, we have natural identifications for r ≥ 1 Rr j∗ F = H r (X \ Z, F ) = HIr+1 (M) 4 KAREN E. SMITH of modules over T [LC, Cor 1.9]. Thus we have reduced Theorem 1.1 to showing that if T is a Noetherian ring over a Noetherian domain A and I is any ideal such that T /I is finitely generated as an A-module, then for any finitely generated T -module M, the modules HIr+1(M) are generically free and commute with base change over A for r ≥ 1. In fact, we will be able to show this for all indices r ≥ −1. To get to the power series case, we first observe that for all i, every element of HIi (M) is annihilated by some power of I. This means that HIi (M) has the structure of a module over the I-adic completion T̂ I . There is no loss of generality in replacing T and M by their I-adic completions T̂ I and M̂ I —the module HIi (M) is canonically identified with HIi T̂ I (M̂ I ). So with out loss of generality, we assume that T is Iadically complete. Now, Lemma 2.2 below guarantees that T is a module-finite algebra over a power series ring A[[x1 , . . . , xn ]]. So the finitely generated T -module M is also a finitely generated module over A[[x1 , . . . , xn ]], and the computation of local cohomology is the same viewed over either ring. This means that to prove Theorem 1.1, it would suffice to prove Theorem 2.1. It only remains to prove Lemma 2.2. Lemma 2.2. Let A be a Noetherian ring. Let T be a Noetherian A-algebra containing an ideal I such that the composition of natural maps A → T ։ T /I is finite. Then there is a natural ring homomorphism from a power series ring A[[x1 , . . . , xn ]] → T̂ I := lim T /I t ←− t which is also finite. Proof of Lemma. Fix generators y1 , . . . , yn for the ideal I of T . Consider the Aalgebra homomorphism A[x1 , . . . , xn ] → T xi 7→ yi . We will show that this map induces a ring homomorphism A[[x1 , . . . , xn ]] → T̂ I which is finite. First note that for each natural number t, there is a naturally induced ring homomorphism (1) A[x1 , . . . , xn ] T → t t (x1 , . . . , xn ) I sending the class xi to the class yi . Claim: For every t, the map (1) is finite. Indeed, if t1 , . . . , td are elements of T whose classes modulo I are A-module generators for T /I, then the classes of t1 , . . . , td modulo I t are generators for T /I t as a module over A[x1 , . . . , xn ]/(x1 , . . . , xn )t . LOCAL COHOMOLOGY AND BASE CHANGE 5 The claim is straightforward to prove using induction on t and the exact sequence 0 → I t /I t+1 → T /I t+1 → T /I t → 0. We leave these details to the reader. Now to prove the lemma, we take the direct limit of the maps (1). Since at every stage, the target is generated over the source by the classes of t1 , . . . , td , also in the limit, T̂ I will be generated over A[[x1 , . . . , xn ]] by the images of t1 , . . . , td . So the induced ring homomorphism A[[x1 , . . . , xn ]] → T is finite.  Having reduced the proof of the main results discussed in the introduction to Theorem 2.1, the rest of the paper focuses on the local cohomology statement in the special case. Our proof of Theorem 2.1 uses an A-relative version of Matlis duality to convert the problem to an analagous one for finitely generated modules over a power series ring, where it will follow from the theorem on generic freeness. This relative version of Matlis duality might be of interest to commutative algebraists in other contexts, and holds in greater generality than what we develop here. To keep the paper as straightforward and readable as possible, we have chosen to present it only in the case we need to prove the main result. Some related duality is worked out in [DGI]. 3. A Relative Matlis Dual Functor 3.1. Matlis Duality. We first recall the classical Matlis duality in the complete local (Noetherian) case. Let (R, m) be a complete local ring, and let E be an injective hull of its residue field R/m. The Matlis dual functor HomR (−, E) is an exact contravariant functor on the category of R-modules. It takes each Noetherian R-module (i.e., one satisfying the ascending chain condition) to an Artinian R-module (i.e., one satisfying the descending chain condition) and vice-versa. Moreover, for any Artinian or Noetherian R-module H, we have a natural isomorphism H → HomR (HomR (H, E), E). That is, the Matlis dual functor defines an equivalence of categories between the category of Noetherian and the category of Artinian R-modules. See [LC], [BH, Thm 3.2.13] or [Hoch] for more on Matlis duality. 3.2. A relative version of Matlis duality. Let A be a domain. Let R be an Aalgebra, with ideal I ⊂ R such that R/I is finitely generated as an A-module. Define the relative Matlis dual functor to be the functor {R − modules} → {R − modules} M 7→ M ∨A := lim HomA (M/I t M, A). −→ t 6 KAREN E. SMITH We also denote M ∨A by Homcts A (M, A), since it is the submodule of HomA (M, A) consisting of maps continuous in the I-adic topology. That is, Homcts A (M, A) is the R-submodule of HomA (M, A) consisting of maps φ : M → A satisfying φ(I t M) = 0 for some t. Proposition 3.1. Let R be a Noetherian algebra over a Noetherian domain A, with ideal I ⊂ R such that R/I is finitely generated as an A-module. (1) The functor Homcts A (−, A) is naturally equivalent to the functor M 7→ HomR (M, Homcts A (R, A)). (2) The functor preserves exactness of sequences 0 → M1 → M2 → M3 → 0 of finitely generated R-modules, provided that the modules M3 /I n M3 are (locally) free A-modules for all n ≫ 0. Remark 3.2. If A = R/I is a field, then the relative Matlis dual specializes to the usual Matlis dual functor HomR (−, E), where E is the injective hull of the residue field of R at the maximal ideal I (denoted here now m). Indeed, one easily checks that Homcts A (R, A) is an injective hull of R/m. To wit, the R-module homomorphism ( R→A R/m → Homcts sending r mod m 7→ A (R, A) s 7→ rs mod m is a maximal essential extension of R/m. Proof of Proposition. Statement (1) follows from the adjointness of tensor and Hom, which is easily observed to restrict to the corresponding statement for modules of continuous maps. For (2), we need to show the sequence remains exact after applying the relative Matlis dual functor. The functor HomA (−, A) preserves left exactness: that is, (2) 0 → HomA (M3 , A) → HomA (M2 , A) → HomA (M1 , A) is exact. We want to show that, restricting to the submodules of continuous maps, we also have exactness at the right. That is, we need the exactness of (3) cts cts 0 → Homcts A (M3 , A) → HomA (M2 , A) → HomA (M1 , A) → 0. The exactness of (3) at all spots except the right is easy to verify using the description of a continuous map as one annihilated by a power of I. To check exactness of (3) at the right, we use the Artin Rees Lemma [AM, 10.10]. n Take φ ∈ Homcts A (M1 , A). By definition of continuous, we φ is annihilated by I for sufficiently large n. By the Artin-Rees Lemma, there exists t such that for all n ≥ t, we have I n+t M2 ∩ M1 ⊂ I n M1 . This means we have a surjection M1 /(I n+t M2 ∩ M1 ) ։ M1 /I n M1 . LOCAL COHOMOLOGY AND BASE CHANGE 7 Therefore the composition M1 /I n+t M2 ∩ M1 ։ M1 /I n M1 → A gives a lifting of φ to an element φ′ in HomA (M1 /I n+t M2 ∩ M1 , A). Now note that for n ≫ 0, we have exact sequences 0 → M1 /M1 ∩ I n+t M2 → M2 /I n+t M2 → M3 /I n+t M3 → 0, which are split over A by our assumption that M3 /I n+t M3 is projective. Thus (4) 0 → HomA (M3 /I n+t M3 , A) → HomA (M2 /I n+t M2 , A) → HomA (M1 /M1 ∩ I n+t M2 , A) → 0 is also split exact. This means we can pull φ′ ∈ HomA (M1 /I n+t M2 ∩ M1 , A) back to φ some element φ̃ in HomA (M2 /I n+t M2 , A). So our original continuous map M1 → A φ̃ is the restriction of some map M2 → A which satisfies φ̃(I n+t M2 ) = 0. This exactly says the continuous map φ on M1 extends to a continuous map φ̃ of M2 . That is, the sequence (3) is exact.  Remark 3.3. If M3 is a Noetherian module over a Noetherian algebra R over the Noetherian domain A, then the assumption that M3 /I n M3 is locally free for all n holds generically on A—that is, after inverting a single element of A. See Lemma 4.2. 4. Generic Freeness We briefly review Grothendieck’s idea of generic freeness, and use it to prove that the relative Matlis dual of a Noetherian R-module is generically free over A (under suitable Noetherian hypothesis on R and A). Let M be a module over a commutative domain A. We say that M is generically free over A if there exists a non-zero g ∈ A, such that M ⊗A Ag is a free Ag -module, where Ag denotes the localization of A at the element g. Likewise, a collection M of A-modules is simultaneously generically free over A if there exists a non-zero g ∈ A, such that M ⊗A Ag is a free for all modules M ∈ M. Note that any finite collection of generically free modules is always simultaneously generically free, since we can take g to be the product the gi that work for each of the Mi . Of course, finitely generated A-modules are always generically free. Grothendieck’s famous Lemma on Generic Freeness ensures that many other modules are as well: Lemma 4.1. [EGA, 6.9.2] Let A be a Noetherian domain. Let M be any finitely generated module over a finitely generated A-algebra T . Then M is generically free over A. We need a version of Generic Freeness for certain infinite families of modules over more general A-algebras: 8 KAREN E. SMITH Lemma 4.2. Let A be any domain. Let T be any Noetherian A-algebra, and I ⊂ T any ideal such that T /I is finite over A. Then for any Noetherian T -module M, the family of modules {M/I n M | n ≥ 1} is simultaneously generically free over A. That is, after inverting a single element of A, the modules M/I n M for all n ≥ 1 become free over A. Remark 4.3. We will make use of Lemma 4.2 in the case where T = A[[x1 , . . . , xn ]]. Proof. If M is finitely generated over T , then the associated graded module grI M = M/IM ⊕ IM/I 2 M ⊕ I 2 M/I 3 M ⊕ . . . is finitely generated over the associated graded ring grI T = T /I ⊕ I/I 2 ⊕ I 2 /I 3 ⊕ . . . , which is a homomorphic image of a polynomial ring over T /I. Hence grI T is a finitely generated A-algebra. Applying the Lemma of generic freeness to the grI T -module grI M, we see that after inverting a single non-zero element of g of A, the module grI M becomes A-free. Since grI M is graded over grI T and A acts in degree zero, clearly its graded pieces are also free after tensoring with Ag . We can thus replace A by Ag for suitable g, and assume that the I n M/I n+1 M are Ag -free for all n ≥ 0. Now consider the short exact sequences (5) 0 → I n M/I n+1 M → M/I n+1 M → M/I n M → 0, for each n ≥ 1. We already know that M/I 1 M and I n M/I n+1 M for all n ≥ 1 are free (over Ag ), so by induction, we conclude that the sequences (5) are all split over Ag for every n. In particular, the modules M/I n M are also free over Ag for all n ≥ 1. The lemma is proved.  Proposition 4.4. Let A be a Noetherian domain. Let R be any Noetherian Aalgebra with ideal I ⊂ R such that R/I is a finitely generated A-module. Then for any Noetherian R-module M, the relative Matlis dual Homcts A (M, A) is a generically free A-module. Furthermore, if g ∈ A is a non-zero element such that Ag ⊗A Homcts A (M, A) is free over Ag , then for any base change A → L factoring through Ag , the natural map cts Homcts A (M, A) ⊗A L → HomL (M ⊗A L, L) is an isomorphism of R ⊗A L-modules, functorial in M. Proof. We can invert one element of A so that each M/I t M is free over A; replace A by this localization. We now claim that the A-module   M cts HomA (M, A) = lim HomA ,A −→ I tM t LOCAL COHOMOLOGY AND BASE CHANGE 9 t is free. Indeed,
since each M/I M is free and has finite rank over A, its A-dual M HomA I t M , A is also free of finite rank. The direct limit is also A-free because the maps in the limit system are all split over A and injective. Indeed, if some finite A-linear combination of fi ∈ Homcts combination is A (M, A) is zero, then that same
, A of homomorphisms zero considered as elements of the free-submodule HomA IM tM in Homcts (M, A) killed by a large power of I. It follows that Homcts A A (M, A) is free over A, as desired. The result on base change follows as well, since tensor commutes with direct limits and with dualizing a finitely generated free module.  Remark 4.5. We can interpret Proposition 4.4 as saying that generically on A, the relative Matlis dual functor (applied to Noetherian R-modules) is exact and commutes with base change. 5. Relative Local Duality and the Proof of the Main Theorem The proof Theorem 2.1 and therefore of our main result answering Kollár’s question, follows from a relative version of Local Duality: Theorem 5.1. Let R be a power series ring A[[x1 , . . . , xn ]] over a Noetherian domain A, and let M be a finitely generated R-module. Then, after replacing A by its localization at one element, there is a functorial isomorphism for all i H i (M) ∼ = [Extn−i (M, R)]∨A . I R To prove Theorem 5.1, we need the following lemma. Lemma 5.2. Let R be a power series ring A[[x1 , . . . , xn ]] over a ring A. There is a ∼ n natural R-module isomorphism Homcts A (R, A) = HI (R), where I = (x1 , . . . , xn ). In particular, the relative Matlis dual functor can also be expressed M 7→ HomR (M, HIn (R)). Proof. We recall that HIi (R) is the i-th cohomology of the extended Čech complex K • on the elements x1 , . . . , xn . This is the complex M δ1 δn 0→R→ Rxi xj → · · · −→ Rx1 ⊕ Rx2 · · · ⊕ Rxn → Rx1 x2 ···xn → 0 i<j where the maps are (sums of) suitably signed localization maps. In particular, HIn (R) is the cokernel of δn , which can be checked to be the free A-module on (the classes of) the monomials xa11 . . . xann with all ai < 0. 1 L page 226 of [Hart], although I have included one extra map δ1 : R → Rxi sending f 7→ ( f1 , . . . , f1 ) in order to make the complex exact on the left, and my ring is a power series ring over A instead of a polynomial rings over a field. This is also discussed in [LC] page 22. 1See 10 KAREN E. SMITH Now define an explicit R-module isomorphism Φ from HIn (R) to Homcts A (R, A) by sending the (class of the) monomial xα to the map φα ∈ Homcts (R, A): A φα R −→ A ( xb11 +a1 +1 . . . xnbn +an +1 b1 bn x1 . . . xn 7→ 0 otherwise mod I if αi + βi ≥ −1 for all i Since {xβ | β ∈ Nn } is an A-basis for R, the map φα is a well-defined A-module map from R to A, and since it sends all but finitely many xβ to zero, it is I-adically continΦ uous. Thus the map HIn (R) −→ Homcts A (R, A) is is an A-module homomorphism; in fact, Φ is an A-module isomorphism, since it defines a bijection between the A-basis {xα | αi < 0} for HIn (R) and {πβ | βi ≥ 0} for Homcts A (R, A) (meaning the dual basis β on the free basis x for R over A) matching the indices αi ↔ βi = −αi − 1. It is easy to check that Φ is also R-linear, by thinking of it as “premultiplication by xα+1 ” on the cokernel of δn . Thus Φ identifies the R-modules HIn (R) and Homcts  A (R, A). Proof of Theorem 5.1. We proceed by proving that both modules are generically ison morphic to a third, namely TorR n−i (M, HI (R)). First, recall how to compute HIi (M). Let K • be the extended Čech complex on the elements x1 , . . . , xn M δ1 δn 0→R→ Rxi xj → · · · −→ Rx1 ⊕ Rx2 · · · ⊕ Rxn → Rx1 x2 ···xn → 0. i<j This is a complex of flat R-modules, all free over A, exact at every spot except the right end. Thus it is a flat R-module resolution of the local cohomology module HIn (R). The local cohomology module HIi (M) is the cohomology of this complex after tensoring over R with M, that is n HIi (M) = TorR n−i (M, HI (R)). n−i On the other hand, let us compute the relative Matlis dual of ExtR (M, R). Let P• • be a free resolution of M over R. The module ExtR (M, R) is the cohomology of the complex HomR (P• , R). We would like to say that the computation of the cohomology of this complex commutes with the relative Matlis dual functor, but the best we can say is that this is true generically on A. To see this, we will apply Lemma 4.2 to the following finite set of R-modules: • For i = 0, . . . , n, the image Di of the i-th map of the complex HomR (P• , R); n−i • For i = 0, . . . , n, the cohomology ExtR (M, R) of the same complex. Lemma 4.2 guarantees that the modules Di /I t Di and n−i n−i ExtR (M, R)/I t ExtR (M, R) LOCAL COHOMOLOGY AND BASE CHANGE 11 are all simultaneously generically free over A for all t ≥ 1. This allows us to break up the complex Ag ⊗A HomR (P• , R) into many short exact sequences, split over Ag , which satisfy the hypothesis of Proposition 3.1(2) (using Ag in place of A and Ag ⊗A R in place of R). It follows that the computation of cohomology of HomR (P• , R) commutes with the relative Matlis dual functor (generically on A). n−i Thus, after replacing A by a localization at one element, ExtR (M, R)]∨A is the cohomology of the complex HomR (HomR (P• , R), HIn (R)). In general, for any finitely generated projective module P and any module H (over any Noetherian ring R), the natural map P ⊗ H → Hom(Hom(P, R), H) sending a simple tensor x ⊗ h to the map which sends f ∈ Hom(P, R) to f (x)h, is an isomorphism, functorially in P and H. Thus we have a natural isomorphism of complexes P• ⊗ HIn (R) ∼ = HomR (HomR (P• , R), HIn (R)), and so [Extn−i (M, R)]∨A is identified with Torn−i (M, HIn (R)), which as we saw is identified with HIi (M). Since all identifications are functorial, we have proven the relative local duality ∼ = [Extn−i (M, R)]∨A , generically on A. HIi (M)  We can finally finish the proof of Theorem 1.1, and hence the main result: Proof of Theorem 2.1. Let R be a power series ring over a Noetherian domain A, and let M be any Noetherian R-module. We need to show that the local cohomology modules HIi (M) are generically free and commute with base change over A. In light of Corollary 4.4 , we can accomplish this by showing that HIi (M) is the relative Matlis dual of a Noetherian R-module, generically on A. But this is guaranteed by the relative local duality theorem Theorem 5.1, which guarantees that H i (M) ∼ = Extn−i (M, R)∨A I R generically on A.  Remark 5.3. One could obviously develop the theory of relative Matlis duality, especially Theorem 5.1, further; I wrote down only the simplest possible case and the simplest possible statements needed to answer Kollár’s question as directly as possible. 12 KAREN E. SMITH References [AM] M. F. Atiyah and I. G. MacDonald, Introduction to Commutative Algebra, Addison Wesley Series in Mathematics, Addison Wesley, London, (1969). [BH] Winfred Burns and Jürgen Herzog, Cohen-Macaulay Rings, Cambridge series in Advanced Mathematics, v 39. Cambridge University Press, (1993). [Conrad] Brian Conrad, Grothendieck Duality and Base Change, Lecture Notes in Mathematics 1750, Springer (2001). [DGI] W.G. Dwyer, J.P.C. Greenlees and S. Iyengar, Duality in algebra and topology, Advances in Mathematics Volume 200, Issue 2, 1 March 2006, Pages 357D402. [Eisen] David Eisenbud, Commutative Algebra with a view towards Algebraic Geometry, Graduate Texts in Mathematics 150, Springer (1995). [EGA] Alexander Grothendieck and Jean Dieudonné, Éléments de Geométrie Algébrique Chapter IV, part 2, Inst. Hautes Études Sci. Pub. Math. 24 (1965). [SGA2] Alexander Grothendieck and Michele Raynaud, Cohomologie locale des faisceaux cohérents et théorèmes de Lefschetz locaux et globaux (SGA 2), Documents Mathèmatiques (Paris) 4, Paris: Sociéité Mathèmatique de France, (2005) [1968], Laszlo, Yves, ed., arXiv:math/0511279, ISBN 978-2-85629-169-6, MR 2171939 [LC] Robin Hartshorne, Local cohomology. A seminar given by A. Grothendieck, Harvard University, Fall, 1961, Lecture notes in mathematics 41, Berlin, New York: Springer-Verlag, (1967). MR0224620 [RD] Robin Hartshorne, Residues and Duality: Lecture Notes of a Seminar on the Work of A. Grothendieck, Given at Harvard 1963 /64, Springer Verlag Lecture Notes in Mathematics, vol 20 (1966). [Hart] Robin Hartshorne, Algebraic Geometry Graduate Texts in Mathematics 52 Springer-Verlag, (2006). [Hoch] Mel Hochster, Lecture notes on Local Cohomology, unpublished, from his University of Michigan website http://www.math.lsa.umich.edu/ hochster/615W11/loc.pdf [HR] Mel Hochster and Joel Roberts, The Purity of Frobenius and Local Cohomology, Advances in Mathematics 21 117–172 (1976). [Kol] János Kollár, Maps between local Picard groups, arXiv:1407.5108v2 (preprint) 2014. 

Symmetric Tensor Completion from Multilinear Entries and Learning Product Mixtures over the Hypercube arXiv:1506.03137v3 [cs.DS] 23 Nov 2015 Tselil Schramm∗ Ben Weitz† November 25, 2015 Abstract We give an algorithm for completing an order-m symmetric low-rank tensor from its multilinear entries in time roughly proportional to the number of tensor entries. We apply our tensor completion algorithm to the problem of learning mixtures of product distributions over the hypercube, obtaining new algorithmic results. If the centers of the product distribution are linearly independent, then we recover distributions with as many as Ω(n) centers in polynomial time and sample complexity. In the general case, we recover distributions with as many as Ω̃(n) centers in quasi-polynomial time, answering an open problem of Feldman et al. (SIAM J. Comp.) for the special case of distributions with incoherent bias vectors. Our main algorithmic tool is the iterated application of a low-rank matrix completion algorithm for matrices with adversarially missing entries. ∗ UC Berkeley, tschramm@cs.berkeley.edu. Supported by an NSF Graduate Research Fellowship (NSF award no 1106400). † UC Berkeley, bsweitz@cs.berkeley.edu. Supported by an NSF Graduate Research Fellowship (NSF award no DGE 1106400). 1 Introduction Suppose we are given sample access to a distribution over the hypercube {±1}n , where each sample x is generated in the following manner: there are k product distributions D1 , . . . , Dk over {±1}n (the k “centers” of the distribution), and x is drawn from Di with probability pi . This distribution is called a product mixture over the hypercube. Given such a distribution, our goal is to recover from samples the parameters of the individual product distributions. That is, we would like to estimate the probability pi of drawing from each product distribution, and furthermore we would like to estimate the parameters of the product distribution itself. This problem has been studied extensively and approached with a variety of strategies (see e.g. [FM99, CR08, FOS08]). A canonical approach to problems of this type is to empirically estimate the moments of the distribution, from which it may be possible to calculate the distribution parameters using linearalgebraic tools (see e.g. [AM05, MR06, FOS08, AGH+ 14], and many more). For product distributions over the hypercube, this technique runs into the problem that the square moments are always 1, and so they provide no information. The seminal work of Feldman, O’Donnell and Servedio [FOS08] introduces an approach to this problem which compensates for the missing higher-order moment information using matrix completion. Via a restricted brute-force search, Feldman et al. check all possible square moments, resulting in an algorithm that is triply-exponential in the number of distribution centers. Continuing this line work, by giving an alternative to the brute-force search Jain and Oh [JO13] recently obtained a polynomial-time algorithm for a restricted class of product mixtures. In this paper we extend these ideas, giving a polynomial-time algorithm for a wider class of product mixtures, and a quasi-polynomial time algorithm for an even broader class of product mixtures (including product mixtures with centers which are not linearly independent). Our main tool is a matrix-completion-based algorithm for completing tensors of order m from their multilinear moments in time Õ(nm+1 ), which we believe may be of independent interest. There has been ample work in the area of noisy tensor decomposition (and completion), see e.g. [JO14, BKS15, TS15, BM15]. However, these works usually assume that the tensor is obscured by random noise, while in our setting the “noise” is the absence of all non-multilinear entries. An exception to this is the work of [BKS15], where to obtain a quasi-polynomial algorithm it suffices to have the injective tensor norm of the noise be bounded via a Sum-of-Squares proof.1 To our knowledge, our algorithm is the only nO(m) -time algorithm that solves the problem of completing a symmetric tensor when only multilinear entries are known. 1.1 Our Results Our main result is an algorithm for learning a large subclass of product mixtures with up to even Ω(n) centers in polynomial (or quasi-polynomial) time. The subclass of distributions on which our algorithm succeeds is described by characteristics of the subspace spanned by the bias vectors. Specifically, the rank and incoherence of the span of the bias vectors cannot simultaneously be too large. Intuitively, the incoherence of a subspace measures how close the subspace is to a coordinate subspace of Rn . We give a formal definition of incoherence later, in Definition 2.3. More formally, we prove the following theorem: 1 It may be possible that this condition is met for some symmetric tensors when only multilinear entries are known, but we do not know an SOS proof of this fact. 1 Theorem 1.1. Let D be a mixture over k product distributions on {±1}n , with bias vectors v1 , . . . , vk ∈ Rn and mixing weights w1 , . . . , wk > 0. Let span{vi } have dimension r and incoherence µ. Suppose we are given as input the moments of D. 1. If v1 , . . . , vk are linearly independent, then as long as 4 · µ · r < n, there is a poly(n, k) algorithm that recovers the parameters of D. 2. Otherwise, if |hvi , vj i| < kvi k · kvj k · (1 − η) for every i 6= j and η > 0, then as long as 4 · µ · r · log k/ log parameters of D. 1 1−η 1 < n, there is an nO(log k/ log 1−η ) time algorithm that recovers the Remark 1.2. In the case that v1 , . . . , vk are not linearly independent, the runtime depends on the separation between the vectors. We remark however that if we have some vi = vj for i 6= j, then the distribution is equivalently representable with fewer centers by taking the center vi with mixing weight wi + wj . If there is some vi = −vj , then our algorithm can be modified to work in that case as well, again by considering vi and vj as one center–we detail this in Section 4. In the main body of the paper we assume access to exact moments; in Appendix B we prove Theorem B.2, a version of Theorem 1.1 which accounts for sampling error. The foundation of our algorithm for learning product mixtures is an algorithm for completing a low-rank incoherent tensor of arbitrary order given access only to its multilinear entries: P Theorem 1.3. Let T be a symmetric tensor of order m, so that T = i∈[k] wi · vi⊗m for some vectors v1 , . . . , vk ∈ Rn and scalars w1 , . . . , wk 6= 0. Let span{vi } have incoherence µ and dimension r. Given perfect access to all multilinear entries of T , if 4·µ ·r ·m/n < 1, then there is an algorithm which returns the full tensor T in time Õ(nm+1 ). 1.2 Prior Work We now discuss in more detail prior work on learning product mixtures over the hypercube, and contextualize our work in terms of previous results. The pioneering papers on this question gave algorithms for a very restricted setting: the works of [FM99] and [C99, CGG01] introduced the problem and gave algorithms for learning a mixture of exactly two product distributions over the hypercube. The first general result is the work of Feldman, O’Donnell and Servedio, who give an algorithm 3 for learning a mixture over k product distributions in n dimensions in time nO(k ) with sample complexity nO(k) . Their algorithm relies on brute-force search to enumerate all possible product mixtures that are consistent with the observed second moments of the distribution. After this, they use samples to select the hypothesis with the Maximum Likelihood. Their paper leaves as an open question the more efficient learning of discrete mixtures of product distributions, with a smaller exponential dependence (or even a quasipolynomial dependence) on the number of centers.2 More recently, Jain and Oh [JO13] extended this approach: rather than generate a large number of hypotheses and pick one, they use a tensor power iteration method of [AGH+ 14] to find the right decomposition of the second- and third-order moment tensors. To learn these moment tensors in the first place, they use alternating minimization to complete the (block)-diagonal of the second moments matrix, and they compute a least-squares estimation of the third-order moment tensor. 2 We do not expect better than quasipolynomial dependence on the number of centers, as learning the parity distribution on t bits is conjectured to require at least nΩ(t) time, and this distribution can be realized as a product mixture over 2t−1 centers. 2 Learning Product Mixtures with k Centers over {±1}n Reference Feldman et al. [FOS08] Jain & Oh [JO14] lin. indep. Our Results lin. dep. Runtime Samples Largest k Dep. Centers? Incoherence? nO(k) n Allowed Not Required poly(n, k) poly(n, k) Not Allowed Required poly(n,k), nÕ(log k) poly(n,k), nÕ(log k) k ≤ O(n2/7 ) k ≤ O(n) Allowed Required nO(k 3) Figure 1: Comparison of our work to previous results. We compare runtime, sample complexity, and restrictions on the centers of the distribution: the maximum number of centers, whether linearly dependent centers are allowed, and whether the centers are required to be incoherent. The two subrows correspond to the cases of linearly independent and linearly dependent centers, for which we guarantee different sample complexity and runtime. Using these techniques, Jain and Oh were able to obtain a significant improvement for a restricted class of product mixtures, obtaining a polynomial time algorithm for linearly independent mixtures over at most k = O(n2/7 ) centers. In order to ensure the convergence of their matrix (and tensor) completion subroutine, they introduce constraints on the span of the bias vectors of the distribution (see Section 2.3 for a discussion of incoherence assumptions on product mixtures). Specifically, letting r the rank of the span, letting µ be the incoherence of the span, and letting n be the dimension of the samples, they require that Ω̃(µ5 r 7/2 ) ≤ n.3 Furthermore, in order to extract the bias vectors from the moment information, they require that the bias vectors be linearly independent. When these conditions are met by the product mixture, Jain and Oh learn the mixture in polynomial time. In this paper, we improve upon this result, and can handle as many as Ω(n) centers in some parameter settings. Similarly to [JO13], we use as a subroutine an algorithm for completing lowrank matrices with adversarially missing entries. However, unlike [JO13], we use an algorithm with more general guarantees, the algorithm of [HKZ11].4 These stronger guarantees allow us to devise an algorithm for completing low-rank higher-order tensors from their multilinear entries, and this algorithm allows us to obtain a polynomial time algorithm for a more general class of linearly independent mixtures of product distributions than [JO13]. Furthermore, because of the more general nature of this matrix completion algorithm, we can give a new algorithm for completing low-rank tensors of arbitrary order given access only to the multilinear entries of the tensor. Leveraging our multilinear tensor completion algorithm, we can reduce the case of linearly dependent bias vectors to the linearly independent case by going to higherdimensional tensors. This allows us to give a quasipolynomial algorithm for the general case, in which the centers may be linearly dependent. To our knowledge, Theorem 1.1 is the first quasipolynomial algorithm that learns product mixtures whose centers are not linearly independent. Restrictions on Input Distributions. We detail our restrictions on the input distribution. In the linearly independent case, if there are k bias vector and µ is the incoherence of their span, and n is the dimension of the samples, then we learn a product mixture in time n3 so long as 3 The conditions are actually more complicated, depending on the condition number of the second-moment matrix of the distribution. For precise conditions, see [JO13]. 4 A previous version of this paper included an analysis of a matrix completion algorithm almost identical to that of [HKZ11], and claimed to be the first adversarial matrix completion result of this generality. Thanks to the comments of an anonymous reviewer, we were notified of our mistake. 3 4µr < n. Compare this to the restriction that Ω̃(r 7/2 µ5 ) < n, which is the restriction of Jain and Oh–we are able to handle even a linear number of centers so long as the incoherence is not too large, while Jain and Oh can handle at most O(n2/7 ) centers. If the k bias vectors are not independent, but their span has rank r and if they have maximum pairwise inner product 1 − η (when scaled to unit vectors), then we learn the product mixture in time nO(log k·(− log 1−η)) so long 1 < n (we also require a quasipolynomial number of samples in this case). as 4µr log k · log 1−η While the quasipolynomial runtime for linearly dependent vectors may not seem particularly glamorous, we stress that the runtime depends on the separation between the vectors. To illustrate the additional power of our result, we note that a choice of random v1 , . . . , vk in an r-dimensional √ subspace meet this condition extremely well, as we have η = 1 − Õ(1/ r) with high probability– for, say, k = 2r, the algorithm of [JO13] would fail in this case, since v1 , . . . , vk are not linearly independent, but our algorithm succeeds in time nO(1) . This quasipolynomial time algorithm resolves an open problem of [FOS08], when restricted to distributions whose bias vectors satisfy our condition on their rank and incoherence. We do not solve the problem in full generality, for example our algorithm fails to work when the distribution can have multiple decompositions into few centers. In such situations, the centers do not span an incoherent subspace, and thus the completion algorithms we apply fail to work. In general, the completion algorithms fail whenever the moment tensors admit many different low-rank decompositions (which can happen even when the decomposition into centers is unique, for example parity on three bits). In this case, the best algorithm we know of is the restricted brute force of Feldman, O’Donnell and Servedio. Sample Complexity. One note about sample complexity–in the linearly dependent case, we require a quasipolynomial number of samples to learn our product mixture. That is, if there are k product centers, we require nÕ(log k) samples, where the tilde hides a dependence on the separation between the centers. In contrast, Feldman, O’Donnell, and Servedio require nO(k) samples. This dependence on k in the sample complexity is not explicitly given in their paper, as for their algorithm to be practical they consider only constant k. Parameter Recovery Using Tensor Decomposition. The strategy of employing the spectral decomposition of a tensor in order to learn the parameters of an algorithm is not new, and has indeed been employed successfully in a number of settings. In addition to the papers already mentioned which use this approach for learning product mixtures ([JO14] and in some sense [FOS08], though the latter uses matrices rather than tensors), the works of [MR06, AHK12, HK13, AGHK14, BCMV14], and many more also use this idea. In our paper, we extend this strategy to learn a more general class of product distributions over the hypercube than could previously be tractably learned. 1.3 Organization The remainder of our paper is organized as follows. In Section 2, we give definitions and background, then outline our approach to learning product mixtures over the hypercube, as well as put forth a short discussion on what kinds of restrictions we place on the bias vectors of the distribution. In Section 3, we give an algorithm for completing symmetric tensors given access only to their multilinear entries, using adversarial matrix completion as an algorithmic primitive. In Section 4, we apply our tensor completion result to learn mixtures of product distributions over the hypercube, assuming access to the precise second- and third-order moments of the distribution. Appendix A and Appendix B contain discussions of matrix completion and learning product 4 mixtures in the presence of sampling error, and Appendix C contains further details about the algorithmic primitives used in learning product mixtures. 1.4 Notation We use ei to denote the ith standard basis vector. For a tensor T ∈ Rn×n×n , we use T (a, b, c) to denote the entry of the tensor indexed by a, b, c ∈ [n], and we use T (i, ·, ·) to denote the ith slice of the tensor, or the subset of entries in m which the first coordinate is fixed to i ∈ [n]. For an order-m tensor T ∈ Rn , we use T (X) to represent the entry indexed by the string X ∈ [n]m , and we use T (Y, ·, ·) to denote the slice of T indexed by the string Y ∈ [n]m−2 . For a vector v ∈ Rn , we use the shorthand x⊗k to denote the k-tensor x ⊗ x · · · ⊗ x ∈ Rn×···×n . We use Ω ⊆ [m] × [n] for the set of observed entries of the hidden matrix M , and PΩ denotes the projection onto those coordinates. 2 Preliminaries In this section we present background necessary to prove our results, as well as provide a short discussion on the meaning behind the restrictions we place on the distributions we can learn. We start by defining our main problem. 2.1 Learning Product Mixtures over the Hypercube A distribution D over {±1}n is called a product distribution if every bit in a sample x ∼ D is independently chosen. Let D1 , . . . , Dk be a set of product distributions over {±1}n . Associate with each Di a vector vi ∈ [−1, 1]n whose jth entry encodes the bias of the jth coordinate, that is P [x(j) = 1] = x∼Di 1 + vi (j) . 2 Define the distribution D to be a convex combination of these product distributions, sampling P x ∼ D = {x ∼ Di with probability wi }, where wi > 0 and i∈[k] wi = 1. The distributions D1 , . . . , Dk are said to be the centers of D, the vectors v1 , . . . , vk are said to be the bias vectors, and w1 , . . . , wk are said to be the mixing weights of the distribution. Problem 2.1 (Learning a Product Mixture over the Hypercube). Given independent samples from a distribution D which is a mixture over k centers with bias vectors v1 , . . . , vk ∈ [−1, 1]n and mixing weights w1 , . . . , wk > 0, recover v1 , . . . , vk and w1 , . . . , wk . This framework encodes many subproblems, including learning parities, a notorious problem in learning theory; the best current algorithm requires time nΩ(k) , and the noisy version of this problem is a standard cryptographic primitive [MOS04, Fel07, Reg09, Val15]. We do not expect to be able to learn an arbitrary mixture over product distribution efficiently. We obtain a polynomial-time algorithm when the bias vectors are linearly independent, and a quasi-polynomial time algorithm in the general case, though we do require an incoherence assumption on the bias vectors (which parities do not meet), see Definition 2.3. In [FOS08], the authors give an nO(k)-time algorithm for the problem based on the following idea. With great accuracy in polynomial time we may compute the pairwise moments of D, X M = E [xxT ] = E2 + wi · vi viT . x∼D i∈[k] 5 The matrix E2 is a diagonal matrix which corrects for the fact that Mjj = 1 always. If we were P able to learn E2 and thus access i∈[k] wi vi viT , the “augmented second moment matrix,” we may hope to use spectral information to learn v1 , . . . , vk . The algorithm of [FOS08] performs a brute-force search to learn E2 , leading to a runtime exponential in the rank. By making additional assumptions on the input D and computing higherorder moments as well, we avoid this brute force search and give a polynomial-time algorithm for product distributions with linearly independent centers: If the bias vectors are linearly independent, a power iteration algorithm of [AGH+ 14] allows us to learn D given access to both the augmented second- and third-order moments.5 Again, sampling the third-order moments only gives access P to Ex∼D [x⊗3 ] = E3 + i∈[k] wi · vi⊗3 , where E3 is a tensor which is nonzero only on entries of multiplicity at least two. To learn E2 and E3 , Jain and Oh used alternating minimization and a least-squares approximation. For our improvement, we develop a tensor completion algorithm based on recursively applying the adversarial matrix completion algorithm of Hsu, Kakade and Zhang [HKZ11]. In order to apply these completion algorithms, we require an incoherence assumption on the bias vectors (which we define in the next section). In the general case, when the bias vectors are not linearly independent, we exploit the fact that high-enough tensor powers of the bias vectors are independent, and we work with the Õ(log k)th moments of D, applying our tensor completion to learn the full moment tensor, and then using [AGH+ 14] to find the tensor powers of the bias vectors, from which we can easily recover the vectors themselves. (the tilde hides a dependence on the separation between the bias vectors). Thus if the distribution is assumed to come from bias vectors that are incoherent and separated, then we can obtain a significant runtime improvement over [FOS08]. 2.2 Matrix Completion and Incoherence As discussed above, the matrix (and tensor) completion problem arises naturally in learning product mixtures as a way to compute the augmented moment tensors. Problem 2.2 (Matrix Completion). Given a set Ω ⊆ [m] × [n] of observed entries of a hidden rank-r matrix M , the Matrix Completion Problem is to successfully recover the matrix M given only PΩ (M ). However, this problem is not always well-posed. For example, consider the input matrix M = + en eTn . M is rank-2, and has only 2 nonzero entries on the diagonal, and zeros elsewhere. Even if we observe almost the entire matrix (and even if the observed indices are random), it is likely that every entry we see will be zero, and so we cannot hope to recover M . Because of this, it is standard to ask for the input matrix to be incoherent: e1 eT1 Definition 2.3. Let U ⊂ Rn be a subspace of dimension r. We say that U is incoherent with parameter µ if maxi∈[n] k projU (ei )k2 ≤ µ nr . If M is a matrix with left and right singular spaces U and V , we say that M is (µU , µV )-incoherent if U (resp. V ) is incoherent with parameter µU (resp µV ). We say that v1 , . . . , vk are incoherent with parameter µ if their span is incoherent with parameter µ. Incoherence means that the singular vectors are well-spread over their coordinates. Intuitively, this asks that every revealed entry actually gives information about the matrix. For a discussion on There are actually several algorithms in this space; we use the tensor-power iteration of [AGH+ 14] specifically. There is a rich body of work on tensor decomposition methods, based on simultaneous diagonalization and similar techniques (see e.g. Jenrich’s algorithm [Har70] and [LCC07]). 5 6 what kinds of matrices are incoherent, see e.g. [CR09]. Once the underlying matrix is assumed to be incoherent, there are a number of possible algorithms one can apply to try and learn the remaining entries of M . Much of the prior work on matrix completion has been focused on achieving recovery when the revealed entries are randomly distributed, and the goal is to minimize the number of samples needed (see e.g. [CR09, Rec09, GAGG13, Har14]). For our application, the revealed entries are not randomly distributed, but we have access to almost all of the entries (Ω(n2 ) entries as opposed to the Ω(nr log n) entries needed in the random case). Thus we use a particular kind of matrix completion theorem we call “adversarial matrix completion,” which can be achieved directly from the work of Hsu, Kakade and Zhang [HKZ11]: Theorem 2.4. Let M be an m×n rank-r matrix which is (µU , µV )-incoherent, and let Ω ⊂ [m]×[n] be the set of hidden indices. If there are at most κ elements per column and ρ elements per row of Ω, and if 2(κ µmU + ρ µnV )r < 1, then there is an algorithm that recovers M . For the application of learning product mixtures, note that the moment tensors are incoherent exactly when the bias vectors are incoherent. In Section 3 we show how to apply Theorem 2.4 recursively to perform a special type of adversarial tensor completion, which we use to recover the augmented moment tensors of D after sampling. Further, we note that Theorem 2.4 is almost tight. That is, there exist matrix completion instances with κ/n = 1 − o(1), µ = 1 and r = 3 for which finding any completion is NP-hard [HMRW14, Pee96] (via a reduction from three-coloring), so the constant on the right-hand side is necessarily at most six. We also note that the tradeoff between κ/n and µ in Theorem 2.4 is necessary because for a matrix of fixed rank, one can add extra rows and columns of zeros in an attempt to reduce κ/n, but this process increases µ by an identical factor. This suggests that improving Theorem 1.1 by obtaining a better efficient adversarial matrix completion algorithm is not likely. 2.3 Incoherence and Decomposition Uniqueness In order to apply our completion techniques, we place the restriction of incoherence on the subspace spanned by the bias vectors. At first glance this may seem like a strange condition which is unnatural for probability distributions, but we try to motivate it here. When the bias vectors are incoherent and separated enough, even high-order moment-completion problems have unique P solutions, and moreover that solution is equal to i∈[k] wi · vi⊗m . In particular, this implies that the distribution must have a unique decomposition into a minimal number of well-separated centers (otherwise those different decompositions would produce different minimum-rank solutions to a moment-completion problem for high-enough order moments). Thus incoherence can be thought of as a special strengthening of the promise that the distribution has a unique minimal decomposition. Note that there are distributions which have a unique minimal decomposition but are not incoherent, such as a parity on any number of bits. 3 Symmetric Tensor Completion from Multilinear Entries In this section we use adversarial matrix completion as a primitive to give a completion algorithm for symmetric tensors when only a special kind of entry in the tensor is known. Specifically, we call a string X ∈ [n]m multilinear if every element of X is distinct, and we will show how to complete m a symmetric tensor T ∈ Rn when only given access to its multilinear entries, i.e. T (X) is known 7 if X is multilinear. In the next section, we will apply our tensor completion algorithm to learn mixtures of product distributions over the boolean hypercube. Our approach is a simple recursion: we complete the tensor slice-by-slice, using the entries we learn from completing one slice to provide us with enough known entries to complete the next. The following definition will be useful in precisely describing our recursive strategy: Definition 3.1. Define the histogram of a string X ∈ [n]m to be the multiset containing the number of repetitions of each character making at least one appearance in X. For example, the string (1, 1, 2, 3) and the string (4, 4, 5, 6) both have the histogram (2, 1, 1). Note that the entries of the histogram of a string of length m always sum to m, and that the length of the histogram is the number of distinct symbols in the string. Having defined a histogram, we are now ready to describe our tensor completion algorithm. Algorithm 3.2 (Symmetric Tensor Completion from Multilinear Moments). Input: The P multilinear entries of the tensor T = i∈[k] wi · vi⊗m + E, for vectors v1 , . . . , vk ∈ Rn and scalars w1 , . . . , wk > 0 and some error tensor E. Goal: Recover the symmetric tensor P T ∗ = i∈[k] wi · vi⊗3m . 1. Initialize the tensor T̂ with the known multilinear entries of T . 2. For each subset Y ∈ [n]m−2 with no repetitions: • Let T̂ (Y, ·, ·) ∈ Rn×n be the tensor slice indexed by Y . • Remove the rows and columns of T̂ (Y, ·, ·) corresponding to indices present in Y . Complete the matrix using the algorithm of [HKZ11] from Theorem 2.4 and add the learned entries to T̂ . 3. For ℓ = m − 2, . . . , 1: (a) For • • • • each X ∈ [n]m with a histogram of length ℓ, if T̂ (X) is empty: If there is an element xi appearing at least 3 times, let Y = X \ {xi , xi }. Else there are elements xi , xj each appearing twice, let Y = X \ {xi , xj }. Let T̂ (Y, ·, ·) ∈ Rn×n be the tensor slice indexed by Y . Complete the matrix T̂ (Y, ·, ·) using the algorithm from Theorem 2.4 and add the learned entries to T̂ . 4. Symmetrize T̂ by taking each entry to be the average over entries indexed by the same subset. Output: T̂ . Observation 3.3. One might ask why we go through the effort of completing the tensor slice-byslice, rather than simply flattening it to an nm/2 × nm/2 matrix and completing that. The reason ⊗m/2 ⊗m/2 may have is that when span v1 , . . . , vk has incoherence µ and dimension r, span v1 , . . . , vk m incoherence as large as µr /k, which drastically reduces the range of parameters for which recovery is possible (for example, if k = O(r) then we would need r < n1/m ). Working slice-by-slice keeps the incoherence of the input matrices small, allowing us to complete even up to rank r = Ω̃(n). P Theorem 3.4. Let T be a symmetric tensor of order m, so that T = i∈[k] wi · vi⊗m for some vectors v1 , . . . , vk ∈ Rn and scalars w1 , . . . , wk 6= 0. Let span{vi } have incoherence µ and dimension r. Given perfect access to all multilinear entries of T (i.e. E = 0), if 4 · µ · r · m/n < 1, then Algorithm 3.2 returns the full tensor T in time Õ(nm+1 ). 8 In Appendix B, we give a version of Theorem 3.4 that accounts for error E in the input. Proof. We prove that Algorithm 3.2 successfully completes all the entries of T by induction on the length of the histograms of the entries. By assumption, we are given as input every entry with a histogram of length m. For an entry X with a histogram of length m−1, exactly one of its elements has multiplicity two, call it xi , and consider the set Y = X \ {xi , xi }. When P step 2 reaches Y , the  T algorithm attempts to complete a matrix revealed from T (Y, ·, ·) = PY i∈[k] wi · vi (Y ) · vi vi , Q where vi (Y ) = j∈Y vi (j), and PY is the projector to the matrix with the rows and columns corresponding to indices appearing in Y removed. Exactly the diagonal of T (Y, ·, ·) is missing since all other entries are multilinear moments, and the (i, i)th entry should be T (X). Because the rank of this matrix is equal to dim(span(vi )) = r and 4µr/n ≤ 4µrm/n < 1, by Theorem 2.4, we can successfully recover the diagonal, including T (X). Thus by the end of step 2, T̂ contains every entry with a histogram of length ℓ ≥ m − 1. For the inductive step, we prove that each time step 3 completes an iteration, T̂ contains every entry with a histogram of length at least ℓ. Let X be an entry with a histogram of length ℓ. When step 3 reaches X in the ℓth iteration, if T̂ does not already contain T (X), the algorithm attempts P to complete a matrix with entries revealed from T (Y, ·, ·) = i∈[k] wi · vi (Y ) · vi viT , where Y is a substring of X with a histogram of the same length. Since Y has a histogram of length ℓ, every entry of T (Y, ·, ·) corresponds to an entry with a histogram of length at least ℓ+1, except for the ℓ×ℓ principal submatrix whose rows and columns correspond to elements in Y . Thus by the inductive hypothesis, T̂ (Y ) is only missing the aforementioned submatrix, and since 4µrℓ/n ≤ 4µrm/n < 1, by Theorem 2.4, we can successfully recover this submatrix, including T (X). Once all of the entries of T̂ are filled in, the algorithm terminates. Finally, we note that the runtime is Õ(nm+1 ), because the algorithm from Theorem 2.4 runs in time Õ(n3 ), and we perform at most nm−2 matrix completions because there are nm−2 strings of length m − 2 over the alphabet [n], and we perform at most one matrix completion for each such string. 4 Learning Product Mixtures over the Hypercube In this section, we apply our symmetric tensor completion algorithm (Algorithm 3.2) to learning mixtures of product distributions over the hypercube, proving Theorem 1.1. Throughout this section we will assume exact access to moments of our input distribution, deferring finite-sample error analysis to Appendix B. We begin by introducing convenient notation. Let D be a mixture over k centers with bias vectors v1 , . . . , vk ∈ [−1, 1]n and mixing weights nm to be the tensor of order-m moments of the distribution D, so w1 , . . . , wk > 0. Define MD m ∈R ⊗m ]. Define T D ∈ Rnm to be the symmetric tensor given by the weighted bias that MD m = Ex∼D [x m P vectors of the distribution, so that TmD = i∈[k] wi · vi⊗m . Note that TmD and MD m are equal on their multilinear entries, and not necessarily equal elsewhere. For example, when m is even, entries of MD m indexed by a single repeating character (the “diagonal”) are always equal to 1. Also observe that if one can sample from distribution D, then estimating MD m is easy. Suppose that the bias vectors of D are linearly independent. Then by Theorem 4.1 (due to [AGH+ 14], with similar statements appearing in [AHK12, HK13, AGHK14]), there is a spectral 9 algorithm which learns D given T2D and T3D 6 (we give an account of the algorithm in Appendix C). Theorem 4.1 (Consequence of Theorem 4.3 and Lemma 5.1 in [AGH+ 14]). Let D be a mixture over k centers with bias vectors v1 , . . . , vk ∈ [−1, 1]n and mixing weights w1 , . . . , wk > 0. Suppose we are P P given access to T2D = i∈[k] wi · vi viT and T3D = i∈[k] wi · vi⊗3 . Then there is an algorithm which recovers the bias vectors and mixing weights of D within ε in time O(n3 + k4 · (log log ε√w1 min )). i D Because T2D and T3D are equal to MD 2 and M3 on their multilinear entries, the tensor compleD tion algorithm of the previous section allows us to find T2D and T3D from MD 2 and M3 (this is only D possible because T2D and T3D are low-rank, whereas MD 2 and M3 are high-rank). We then learn D by applying Theorem 4.1. A complication is that Theorem 4.1 only allows us to recover the parameters of D if the bias vectors are linearly independent. However, if the vectors v1 , . . . , vk are not linearly independent, we can reduce to the independent case by working instead with v1⊗m , . . . , vk⊗m for sufficiently large m. The tensor power we require depends on the separation between the bias vectors: Definition 4.2. We call a set of vectors v1 , . . . , vk η-separated if for every i, j ∈ [k] such that i 6= j, |hvi , vj i| ≤ kvi k · kvj k · (1 − η). Lemma 4.3. Suppose that v1 , . . . , vk ∈ Rn are vectors which are η-separated, for η > 0. Let m ≥ ⌈log 1 k⌉. Then v1⊗m , . . . , vk⊗m are linearly independent. 1−η Proof. For vectors u, w ∈ Rn and for an integer t ≥ 0, we have that hu⊗t , w⊗t i = hu, wit . If v1 , . . . , vk are η-separated, then for all i 6= j,   vj⊗m vi⊗m , ≤ |(1 − η)m | ≤ k1 . kvi km kvj km Now considering the Gram matrix of the vectors ( kvvii k )⊗m , we have a k × k matrix with diagonal entries of value 1 and off-diagonal entries with maximum absolute value k1 . This matrix is strictly diagonally dominant, and thus full rank, so the vectors must be linearly independent. Remark 4.4. We re-iterate here that in the case where η = 0, we can reduce our problem to one with fewer centers, and so our runtime is never infinite. Specifically, if vi = vj for some i 6= j, then we can describe the same distribution by omitting vj and including vi with weight wi + wj . If vi = −vj , in the even moments we will see the center vi with weight wi + wj , and in the odd moments we will see vi with weight wi − wj . So we simply solve the problem by taking m′ = 2m for the first odd m so that the v ⊗m are linearly independent, so that both the 2m′ - and 3m′ -order moments are even to learn wi + wj and ±vi , and then given the decomposition into centers we can extract wi and wj from the order-m moments by solving a linear system. Thus, in the linearly dependent case, we may choose an appropriate power m, and instead apply D D D the tensor completion algorithm to MD 2m and M3m to recover T2m and T3m . We will then apply ⊗m ⊗m Theorem 4.1 to the vectors v1 , . . . , vk in the same fashion. Here we give the algorithm assuming perfect access to the moments of D and defer discussion of the finite-sample case to Appendix B. We remark again that the result in [AGH+ 14] is quite general, and applies to a large class of probability distributions of this character. However the work deals exclusively with distributions for which M2 = T2 and M3 = T3 , and assumes access to T2 and T3 through moment estimation. 6 10 Algorithm 4.5 (Learning Mixtures of Product Distributions). Input: Moments of the distribution D. Goal: Recover v1 , . . . , vk and w1 , . . . , wk . Let m be the smallest odd integer such that v1⊗m , . . . , vk⊗m are linearly independent. Let M̂ = M2m + Ê2 and T̂ = M3m + Ê3 be approximations to the moment tensors of order 2m and 3m. 1. Set the non-multilinear entries of M̂ and T̂ to “missing,” and run Algorithm 3.2 on P P M̂ and T̂ to recover M ′ = i wi · vi⊗2m + E2′ and T ′ = i wi · vi⊗3m + E3′ . P 2. Flatten M ′ to the nm × nm matrix M = i wi · vi⊗m (vi⊗m )⊤ + E2 and similarly P flatten T ′ to the nm × nm × nm tensor T = i wi · (vi⊗m )⊗3 + E3 . 3. Run the “whitening” algorithm from Theorem 4.1 (see Appendix C) on (M, T ) to recover w1 , . . . , wk and v1⊗m , . . . , vk⊗m . 4. Recover v1 , . . . , vk entry-by-entry, by taking the mth root of the corresponding entry in v1⊗m , . . . , vk⊗m . Output: w1 , . . . , wk and v1 , . . . , vk . Now Theorem 1.1 is a direct result of the correctness of Algorithm 4.5: Proof of Theorem 1.1. The proof follows immediately by combining Theorem 4.1 and Theorem 3.4, and noting that the parameter m is bounded by m ≤ 2 + log 1 k. 1−η Acknowledgements We would like to thank Prasad Raghavendra, Satish Rao, and Ben Recht for helpful discussions, and Moritz Hardt and Samuel B. Hopkins for helpful questions. We also thank several anonymous reviewers for very helpful comments. References [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 (2014), no. 1, 2773–2832. 1, 2, 6, 9, 10, 15, 16, 17 [AGHK14] Animashree Anandkumar, Rong Ge, Daniel Hsu, and Sham M. Kakade, A tensor approach to learning mixed membership community models, Journal of Machine Learning Research 15 (2014), no. 1, 2239–2312. 4, 9, 17 [AGJ14a] Anima Anandkumar, Rong Ge, and Majid Janzamin, Analyzing tensor power method dynamics: Applications to learning overcomplete latent variable models, CoRR abs/1411.1488 (2014). 17 [AGJ14b] Animashree Anandkumar, Rong Ge, and Majid Janzamin, Guaranteed non-orthogonal tensor decomposition via alternating rank-1 updates, CoRR abs/1402.5180 (2014). 17 [AHK12] Animashree Anandkumar, Daniel Hsu, and Sham M. Kakade, A method of moments for mixture models and hidden markov models, COLT 2012 - The 25th Annual Conference on Learning Theory, June 25-27, 2012, Edinburgh, Scotland, 2012, pp. 33.1–33.34. 4, 9, 17 11 [AM05] Dimitris Achlioptas and Frank McSherry, On spectral learning of mixtures of distributions, Learning Theory, 18th Annual Conference on Learning Theory, COLT, 2005, pp. 458–469. 1 [BCMV14] Aditya Bhaskara, Moses Charikar, Ankur Moitra, and Aravindan Vijayaraghavan, Smoothed analysis of tensor decompositions, Symposium on Theory of Computing, STOC, 2014. 4 [BKS15] Boaz Barak, Jonathan A. Kelner, and David Steurer, Dictionary learning and tensor decomposition via the sum-of-squares method, Proceedings of the Forty-Seventh Annual ACM on Symposium on Theory of Computing, STOC 2015, Portland, OR, USA, June 14-17, 2015, 2015, pp. 143–151. 1 [BM15] Boaz Barak and Ankur Moitra, Tensor prediction, rademacher complexity and random 3-xor, CoRR abs/1501.06521 (2015). 1 [C99] Ph.D. thesis. 2 [CGG01] Mary Cryan, Leslie Ann Goldberg, and Paul W. Goldberg, Evolutionary trees can be learned in polynomial time in the two-state general markov model, SIAM J. Comput. 31 (2001), no. 2, 375–397. 2 [CP09] Emmanuel J. Candès and Yaniv Plan, Matrix completion with noise, CoRR abs/0903.3131 (2009). 14 [CR08] Kamalika Chaudhuri and Satish Rao, Learning mixtures of product distributions using correlations and independence, 21st Annual Conference on Learning Theory - COLT 2008, Helsinki, Finland, July 9-12, 2008, 2008, pp. 9–20. 1 [CR09] Emmanuel J. Candès and Benjamin Recht, Exact matrix completion via convex optimization, Foundations of Computational Mathematics 9 (2009), no. 6, 717–772. 7 [Fel07] Vitaly Feldman, Attribute-efficient and non-adaptive learning of parities and DNF expressions, Journal of Machine Learning Research 8 (2007), 1431–1460. 5 [FM99] Yoav Freund and Yishay Mansour, Estimating a mixture of two product distributions, Proceedings of the Twelfth Annual Conference on Computational Learning Theory, COLT Santa Cruz, CA, USA, July 7-9, 1999, pp. 53–62. 1, 2 [FOS08] Jon Feldman, Ryan O’Donnell, and Rocco A. Servedio, Learning mixtures of product distributions over discrete domains, SIAM J. Comput. 37 (2008), no. 5, 1536–1564. 1, 3, 4, 5, 6 [GAGG13] Suriya Gunasekar, Ayan Acharya, Neeraj Gaur, and Joydeep Ghosh, Noisy matrix completion using alternating minimization, Machine Learning and Knowledge Discovery in Databases (Hendrik Blockeel, Kristian Kersting, Siegfried Nijssen, and Filip Železný, eds.), Lecture Notes in Computer Science, vol. 8189, Springer Berlin Heidelberg, 2013, pp. 194–209 (English). 7 [GHJY15] Rong Ge, Furong Huang, Chi Jin, and Yang Yuan, Escaping from saddle points - online stochastic gradient for tensor decomposition, CoRR abs/1503.02101 (2015). 17 12 [Har70] Richard A. Harshman, Foundations of the PARFAC Procedure: Models and Conditions for and “Explanatory” Multimodal Factor Analysis. 6 [Har14] Moritz Hardt, Understanding alternating minimization for matrix completion, 55th IEEE Annual Symposium on Foundations of Computer Science, FOCS 2014, Philadelphia, PA, USA, October 18-21, 2014, 2014, pp. 651–660. 7 [HK13] Daniel Hsu and Sham M. Kakade, Learning mixtures of spherical gaussians: moment methods and spectral decompositions, Innovations in Theoretical Computer Science, ITCS ’13, Berkeley, CA, USA, January 9-12, 2013, 2013, pp. 11–20. 4, 9, 17, 18 [HKZ11] Daniel Hsu, Sham M. Kakade, and Tong Zhang, Robust matrix decomposition with sparse corruptions, IEEE Transactions on Information Theory 57 (2011), no. 11, 7221– 7234. 3, 6, 7, 8, 14 [HMRW14] Moritz Hardt, Raghu Meka, Prasad Raghavendra, and Benjamin Weitz, Computational limits for matrix completion, CoRR abs/1402.2331 (2014). 7 [JO13] Prateek Jain and Sewoong Oh, Learning mixtures of discrete product distributions using spectral decompositions., CoRR abs/1311.2972 (2013). 1, 2, 3, 4 [JO14] Prateek Jain and Sewoong Oh, Provable tensor factorization with missing data, Advances in Neural Information Processing Systems 27: Annual Conference on Neural Information Processing Systems 2014, December 8-13 2014, Montreal, Quebec, Canada, 2014, pp. 1431–1439. 1, 3, 4 [LCC07] Lieven De Lathauwer, Joséphine Castaing, and Jean-François Cardoso, Fourth-order cumulant-based blind identification of underdetermined mixtures, IEEE Transactions on Signal Processing 55 (2007), no. 6-2, 2965–2973. 6 [MOS04] Elchanan Mossel, Ryan O’Donnell, and Rocco A. Servedio, Learning functions of k relevant variables, J. Comput. Syst. Sci. 69 (2004), no. 3, 421–434. 5 [MR06] Elchanan Mossel and Sébastien Roch, Learning nonsingular phylogenies and hidden markov models, The Annals of Applied Probability 16 (2006), no. 2, 583–614. 1, 4 [Pee96] Renè Peeters, Orthogonal representations over finite fields and the chromatic number of graphs, Combinatorica 16 (1996), no. 3, 417–431 (English). 7 [Rec09] Benjamin Recht, A simpler approach to matrix completion, CoRR abs/0910.0651 (2009). 7 [Reg09] Oded Regev, On lattices, learning with errors, random linear codes, and cryptography, J. ACM 56 (2009), no. 6. 5 [TS15] Gongguo Tang and Parikshit Shah, Guaranteed tensor decomposition: A moment approach. 1 [Val15] Gregory Valiant, Finding correlations in subquadratic time, with applications to learning parities and the closest pair problem, J. ACM 62 (2015), no. 2, 13. 5 13 A Tensor Completion with Noise Here we will present a version of Theorem 3.4 which account for noise in the input to the algorithm. We will first require a matrix completion algorithm which is robust to noise. The work of [HKZ11] provides us with such an algorithm; the following theorem is a consequence of their work.7 Theorem A.1. Let M be an m×n rank-r matrix which is (µU , µV )-incoherent, and let Ω ⊂ [m]×[n] be the set of hidden indices. If there are at most κ elements per column and q ρ elements per row of µV µU µV µU 3 r U µV . In particular, Ω, and if 2(κ m + ρ n )r < 1, then let α = 2 (κ m + ρ n )r and β = 1−λ κρµmn α < 1 and β < 1. Then for every δ > 0, there is a semidefinite program that computes outputs M̂ satisfying p r 2δ min(n, m) 1 kM̂ − M kF ≤ 2δ + 1+ . 1−β 1−α We now give an analysis for the performance of our tensor completion algorithm, Algorithm 3.2, in the presence of noise in the input moments. This will enable us to use the algorithm on empirically estimated moments. P Theorem A.2. Let T ∗ be a symmetric tensor of order m, so that T ∗ = i∈[k] wi · vi⊗m for some vectors v1 , . . . , vk ∈ Rn and scalars w1 , . . . , wk 6= 0. Let span{vi } have incoherence µ and dimension r. Suppose we are given access to T = T ∗ + E, where E is a noise tensor with |E(Y )| ≤ ε for every Y ∈ [n]m . Then if 4 · k · µ · m ≤ n, Then Algorithm 3.2 recovers a symmetric tensor T̂ such that kT̂ (X, ·, ·) − T ∗ (X, ·, ·)kF ≤ 4 · ε · (5n3/2 )m−1 , for any slice T (X, ·, ·) indexed by a string X ∈ [n]m−2 , in time Õ(nm+1 ). In particular, the total 3 Frobenius norm error kT̂ − T ∗ kF is bounded by 4 · ε · (5n3/2 ) 2 m−2 . Proof. We proceed by induction on the histogram length of the entries: we will prove that an entry with a histogram of length ℓ has error at most ε(5n3/2 )m−ℓ . In the base case of ℓ = m, we have that by assumption, every entry of E is bounded by ε. Now, for the inductive step, consider an entry X with a histogram of length ℓ ≤ m − 1. In filling in the entry T (X), we only use information from entries with shorter histograms, which by the inductive hypothesis each have error at most α = ε(5n3/2 )m−ℓ−1 . Summing over the squared errors of the individual entries, the squared Frobenius norm error of the known entries in the slice in which T (X) was completed, pre-completion is at most n2 α2 . Due to the assumptions on k, µ, m, n, by Theorem A.1, matrix completion amplifies the Frobenius norm error of β to at most a Frobenius norm error of 5β · n1/2 . Thus, we have that the Frobenius norm of the slice T (X) was completed in, post-completion, is at most 5n3/2 α, and therefore that the error in the entry T (X) is as most ε · (5n3/2 )m−ℓ , as desired. This concludes the induction. Finally, as our error bound is per entry, it is not increased by the symmetrization in step 4. Any slice has at most one entry with a histogram of length one, 2n − 2 entries with a histogram of length two, and n2 − (2n − 1) entries with a histogram of length three. Thus the total error in a slice is at most 4 · ε · (5n3/2 )m−1 , and there are nm−2 slices. 7 In a previous version of this paper, we derive Theorem A.1 as a consequence of Theorem 2.4 and the work of [CP09]; we refer the interested reader to http://arxiv.org/abs/1506.03137v2 for the details. 14 B Empirical Moment Estimation for Learning Product Mixtures In Section 4, we detailed our algorithm for learning mixtures of product distributions while assuming access to exact moments of the distribution D. Here, we will give an analysis which accounts for the errors introduced by empirical moment estimation. We note that we made no effort to optimize the sample complexity, and that a tighter analysis of the error propagation may well be possible. Algorithm B.1 (Learning product mixture over separated centers). Input: N independent samples x1 , . . . , xN from D, where D has bias vectors with separation η > 0. Goal: Recover the bias vectors and mixing weights of D. Let m be the smallest odd integer for which v1⊗m , . . . , vk⊗m become linearly independent. ⊗m ⊗m ⊤ 1 P D 1. Empirically estimate MD 2m and M3m by calculation M := N i∈[N ] (xi )(xi ) P ⊗3 and T := N1 i∈[N ] (x⊗m i ) . 2. Run Algorithm 4.5 on M and T . Output: The approximate mixing weights ŵ1 , . . . , ŵk , and the approximate vectors v̂1 , . . . , v̂k . Theorem B.2 (Theorem 1.1 with empirical moment estimation). Let D be a product mixture over k centers with bias vectors v1 , . . . , vk ∈ [−1, 1]n and mixing weights w1 , . . . , wk > 0. Let m be the smallest odd integer for which v1⊗m , . . . , vk⊗m are linearly independent (if v1 , . . . , vk are η-separated P for η > 0, then m ≤ log 1 k). Define M2m = i∈[k] vi⊗m (vi⊗m )⊤ . Suppose 1−η 4 · m · r · µ ≤ n, where µ and r are the incoherence and dimension of the space span{vi } respectively. Furthermore,
√ 1 let β ≤ min O(1/k wmax ), 40 be suitably small, and let the parameter N in Algorithm B.1 satisfy N ≥ ε22 (4 log n + log 1δ ) for ε satisfying ! σk (M )1/2 β · σk (M2m ) 1 , ε≤ min √ 6 wmax (5n3/2 )3m/2 4 · (5n3/2 )3m−2 Finally, pick any η ∈ (0, 1). Then with probability at least 1 − δ − η, Algorithm B.1 returns vectors v̂1 , . . . , v̂k and mixing weights ŵ1 , . . . , ŵk such that kv̂i − vi k ≤ √  1/2 n · 10 · β + 60 · β · kM2m k β · σk (M2m ) + √ 6 wmax 1/m , and |ŵi − wi | ≤ 40β, and runs in time nO(m) · O(N · poly(k) log(1/η) · (log k + log log( wmax ε ))). In particular, a choice of Õ(m) N ≥n gives sub-constant error, where the tilde hides the dependence on wmin and σk (M2m ). Before proving Theorem B.2, we will state state the guarantees of the whitening algorithm of [AGH+ 14] on noisy inputs, which is used as a black box in Algorithm 4.5. We have somewhat modified the statement in [AGH+ 14] for convenience; for a breif account of their algorithm, as well as an account of our modifications to the results as stated in [AGH+ 14], we refer the reader to Appendix C. 15 Theorem B.3 (Corollary of Theorem 4.3 in [AGH+ 14]). Let v1 , . . . , vk ∈ [−1, 1]n be vectors and P P let w1 , . . . , wk > 0 be weights. Define M = i∈[k] wi · vi viT and T = i∈[k] wi · vi⊗3 , and suppose we are given M̂ = M + EM and T̂ = T + ET , where EM ∈ Rn×n and ET ∈ Rn×n×n are symmetric error terms such that √   √ 6kEM kF wmax 1 kkET kF <O √ 2β := + . σk (M ) wmax · k σk (M )3/2 Then there is an algorithm that recovers vectors v̂1 , . . . , v̂k and weights ŵ1 , . . . , ŵk such that for all i ∈ [n], kvi − v̂i k ≤ kEM k1/2 + 60kM k1/2 β + 10β, and |wi − ŵi | ≤ 40β, 1 with probability 1 − η in time O(L · k 3 · (log k + log log( √wmax ·ε ))), where L is poly(k) log(1/η). Having stated the guarantees of the whitening algorithm, we are ready to prove Theorem B.2. Proof of Theorem B.2. We account for the noise amplification in each step. Step 1: In this step, we empirically estimate the multilinear moments of the distribution. We will apply concentration inequalities on each entry individually. By a Hoeffding bound, each entry concentrates within ε of its expectation with probability 1 − exp(− 12 N · ε2 ). Taking a union n
n
bound over the 2m + 3m moments we must estimate, we conclude that with probability at least 1 − exp(− 21 N · ε2 + 4m log n), all moments concentrate to within ε of their expectation. Setting N = ε22 (4m log n + log 1δ ), we have that with probability 1 − δ, every entry concentrates to within ε of its expectation. Now, we run Algorithm 4.5 on the estimated moments. Step 1 of Algorithm 4.5: Applying Theorem A.2, we see that the error satisfies kE2′ kF ≤ 9 4 · ε · (5n3/2 )3m−2 and kE3′ kF ≤ 4 · ε · (5n3/2 ) 2 m−2 . Step 2 of Algorithm 4.5: No error is introduced in this step. Step 3 of Algorithm 4.5: Here, we apply Theorem B.3 out of the box, where our vectors are now the vi⊗m . The desired result now follows immediately for the estimated mixing weights, and for the estimated tensored vectors we have kui − vi⊗m k ≤ 10 · β + 60 · βkM k1/2 + kE2′ k, for √ β as defined in Theorem B.2. Note that kE2′ k ≤ kE2′ kF ≤ β · σk (M2m )/6 wmax , so let γ = k (M2m ) √ 10 · β + 60 · βkM k1/2 + β·σ 6 wmax . Step 4 of Algorithm 4.5: Let u∗i be the restriction of ui to the single-index entries, and let be the same restriction for vi⊗m . The bound on the error of the ui applies to restrictions, so we have kvi∗ − u∗i k ≤ γ. So the error in each entry is bounded by γ. By the concavity of the mth root, √ we thus have that kvi − v̂i k ≤ n · γ 1/m . vi∗ To see that choosing N ≥ nÕ(m) gives sub-constant error, calculations suffice; we only add that kM2m k ≤ rnm , where we have applied a bound on the Frobenius norm of kM2m k. The tilde hides the dependence on wmin and σk (M2m ). This concludes the proof. 16 C Recovering Distributions from Second- and Third-Order Tensors In this appendix, we give an account of the algorithm of [AGH+ 14] which, given access to estimates of MVD and TVD , can recover the parameters of D. We note that the technique is very similar to those of [AHK12, HK13, AGHK14], but we use the particular algorithm of [AGH+ 14]. In previous sections, we have given a statement that follows from their results; here we will detail the connection. In [AGH+ 14], the authors show that for a family of distributions with parameters v1 , . . . , vk ∈ Rn and w1 , . . . , wk > 0, if the v1 , . . . , vk are linearly independent and one has approximate access to P P MV := i∈[k] wi vi viT and TV := i∈[k] wi · vi⊗3 , then the parameters can be recovered. For this, they use two algorithmic primitives: singular value decompositions and tensor power iteration. Tensor power iteration is a generalization of the power iteration technique for finding matrix eigenvectors to the tensor setting (see e.g. [AGH+ 14]). The generalization is not complete, and the convergence criteria for the method are quite delicate and not completely understood, although there has been much progress in this area of late ([AGJ14b, AGJ14a, GHJY15]). However, it is wellknown that when the input tensor T ∈ Rn×n×n is decomposable into k < n symmetric orthogonal P rank-1 tensors, i.e. T = i∈[k] vi⊗3 where k < n and hvi , vj i = 0 for i 6= j, then it is possible to recover v1 , . . . , vk using tensor power iteration. The authors of [AGH+ 14] prove that this process is robust to some noising of T : Theorem C.1 (Theorem 5.1 in [AGH+ 14]). Let T̃ = T + E ∈ Rk×k×k be a symmetric tensor, P where T has the decomposition T = i∈[k] λi · ui ⊗ ui ⊗ ui for orthonormal vectors u1 , · · · , uk and λ1 , . . . , λk > 0, and E is a tensor such that kEkF ≤ β. Then there exist universal constants C1 , C2 , C3 > 0 such that the following holds. Choose η ∈ (0, 1), and suppose β ≤ C1 · and also r ln(L/ log2 (k/η)) · ln k λmin k ln(ln(L/ log2 (k/η))) + C3 1− − 4 ln(L/ log2 (k/η)) s ln 8 ln(L/ log2 (k/η)) ! ≥ 1.02 1 + r ln 4 ln k ! . Then there is a tensor power iteration based algorithm that recovers vectors û1 , . . . , ûk and coefficients λ̂1 , . . . , λ̂k with probability at least 1 − η such that for all i ∈ [n], kûi − ui k ≤ β 8 , λi and |λ̂i − λi | ≤ 5β, in O(L · k3 · (log k + log log( λmax β ))) time. The conditions are met when L = poly(k) log(1/η). The idea is then to take the matrix MV , and apply a whitening map W = (MV† )1/2 to orthogonalize the vectors. Because v1 , . . . , vk are assumed to be linearly independent, and because P T = Id , it follows that the √w · W v are orthogonal vectors. W MW = i i k i∈[k] wi (W vi )(W vi ) Now, applying the map W ∈ Rk×n to every slice of T in every direction, we obtain a new tensor P TW = i∈[k] wi (W vi )⊗3 , by computing each entry: T (W, W, W )a,b,c := TW (a, b, c) = X 1≤a′ ,b′ ,c′ ≤n W T (a′ , a) · W T (b′ , b) · W T (c′ , c) · T (a′ , b′ , c′ ). 17 From here on out we will use T (A, A, A) to denote this operation on tensors. The tensor TW P √ thus has an orthogonal decomposition. Letting ui = wi W vi , we have that T = i∈[k] √1wi · u⊗3 i . √ 1 Applying tensor power iteration allows the recovery of the ui = wi · W vi and the weights √wi , from which the vi are recoverable. The theorem Theorem B.3 is actually the consequence of Theorem C.1 and the following proposition, which controls the error propagation in the whitening step. P Proposition C.2 (Consequence of Lemma 12 of [HK13]). Let M2 = i∈[k] λi · ui uTi be a rank-k PSD matrix, and let M̂ be a symmetric matrix whose top k eigenvalues are positive. Let T = P ⊗3 i∈[k] λi · ui , and let T̂ = T + ET where ET is a symmetric tensor with kET kF ≤ γ. Suppose kM2 − M̂ kF ≤ εσk (M2 ), where σk (M ) is the kth eigenvalue of M2 . Let U be the square root of the pseudoinverse of M2 , and let Û be the square root of the pseudoinverse of the projection of M̂ to its top k eigenvectors. Then kT (U, U, U ) − T̂ (Û , Û , Û )k ≤ √ 6 ε + γ · kÛ k2 kÛ kF λmin Proof. We use the following fact, which is given as Lemma 12 in [HK13]. kT (U, U, U ) − T̂ (Û , Û , Û )k ≤ √ 6 ε + kET (Û , Û , Û )k2 . λmin The proof of this fact is straightforward, but requires a great deal of bookkeeping; we refer the reader to [HK13]. It remains to bound kET (Û , Û , Û )k2 . Some straightforward calculations yield the desired bound, X X kE(Û , Û , Û )k2 ≤ k(Û ei ) ⊗ Û T Ei Û k ≤ kÛ ei k2 kÛ T Ei Û k i i 2 ≤ kÛ k · 2 ≤ kÛ k · X i kÛ ei k2 kEi k ≤ kÛ k2 · sX i kÛ ei k22 sX i X i kÛ ei k2 kEi kF kEi k2F ≤ kÛ k2 · kÛ kF · kEkF , where we have applied the triangle inequality, the behavior of the spectral norm under tensoring, the submultiplicativity of the norm, and Cauchy-Schwarz. We now prove Theorem B.3. Proof of Theorem B.3. Let Û be the square the projection of M̂ to its top k eigenvectors. √ root of−1/2 −1/2 Note that kÛ k ≤ σk (M2 ) , kÛ kF ≤ kσk (M2 ) , and thus by Proposition C.2, the error E in Theorem C.1 satisfies √ 6kEM kF kET kF k √ . 2β := kEkF ≤ + σk (M2 ) λmin σk (M2 )3/2 Suppose 1/40 ≥ 2β ≥ kEkF . Applying Proposition C.2, we obtain vectors u1 , . . . , uk and √ √ scaling factors λ1 , . . . , λk such that kui − wi · M −1/2 vi k ≤ 16 · β · wi and | √1wi − λi | ≤ 5 · β. The wi are now recovered by taking the inverse square of the λi , so we have that when 10β < |ŵi − wi | = 2λi − 10β 1 1 1 ≤ 5β · 2 ≤ 40β, − wi ≤ 2 − 2 2 (λi ± 10β) λi λi λi (λi − 10β)2 18 1 4 ≤ 41 λi , where to obtain the second inequality we have taken a Taylor expansion, and in the final inequality we have used the fact that 10β < 41 λi . We now recover vi by taking v̂i = λi · Û ui , so we have √ √ kv̂i − vi k ≤ kλi · Û wi · M −1/2 vi − vi k + kλi · Û (ui − wi M −1/2 vi )k √ √ √ ≤ (λi · wi )k(Û · M −1/2 − I)vi k + (1 − λi wi )kvi k + kÛ k · 16βλi wi ≤ (1 + 10β)kÛ · M −1/2 − Ik + 10β + kÛ k · 16β(1 + 10β) ≤ (1 + 10β)kÛ · M −1/2 − Ik + 10β + kÛ k · 16β(1 + 10β). It now suffices to bound kÛ M −1/2 − Ik, for which it in turn suffices to bound kM −1/2 Û Û M −1/2 − Ik, since the eigenvalues of AAT are the square eigenvalues of A. Consider k(M −1/2 Πk (M + EM )Πk )M −1/2 − Ik, where Πk is the projector to the top k eigenvectors of M . Because both matrices are PSD, finally this reduces to bounding kM − Πk (M + EM )Πk k. Since M is rank k, we have that kM − Πk (M + E)Πk k = σk+1 (EM ) ≤ kEM k. Thus, taking loose bounds, we have kvi − v̂i k ≤ kEM k1/2 + 60β · kM2 k1/2 + 10β, as desired. 19 

Fast camera focus estimation for gaze-based focus control Wolfgang Fuhla , Thiago Santinia , Enkelejda Kasnecia a Eberhard Karls University Tübingen, Perception Engineering, Germany, 72076 Tübingen, Sand 14, Tel.: +49 70712970492, wolfgang.fuhl@uni-tuebingen.de, thiago.santini@uni-tuebingen.de, Enkelejda.Kasneci@uni-tuebingen.de arXiv:1711.03306v1 [cs.CV] 9 Nov 2017 Abstract Many cameras implement auto-focus functionality; however, they typically require the user to manually identify the location to be focused on. While such an approach works for temporally-sparse autofocusing functionality (e.g., photo shooting), it presents extreme usability problems when the focus must be quickly switched between multiple areas (and depths) of interest – e.g., in a gaze-based autofocus approach. This work introduces a novel, real-time auto-focus approach based on eye-tracking, which enables the user to shift the camera focus plane swiftly based solely on the gaze information. Moreover, the proposed approach builds a graph representation of the image to estimate depth plane surfaces and runs in real time (requiring ≈ 20ms on a single i5 core), thus allowing for the depth map estimation to be performed dynamically. We evaluated our algorithm for gaze-based depth estimation against state-of-the-art approaches based on eight new data sets with flat, skewed, and round surfaces, as well as publicly available datasets. Keywords: Shape from focus, Depth from focus, Delaunay Triangulation, Gaze based focus control, eye controlled autofocus, focus estimation 1. Introduction Human vision is a foveated vision, i.e., sharpest vision is possible only in the central 2◦ of the field of view. Therefore, in order to perceive our environment, we perform eye movements, which allow us to focus on and switch between regions of interest sequentially [23]. Modern digital microscopes and cameras share a similar problem, in the sense that the auto focus is only applied to the center of the field of view of the camera. For various applications – such as microsurgery, microscopical material inspection, human-robot collaborative settings, security cameras – it would be a significant usability improvement to allow users to adjust the focus in the image to their point of interest without reorienting the camera or requiring manual focus adjustments. This would not only generate a benefit for the user of the optical system, but also to non-users – for instance, patients would benefit from a faster surgery and a less strained surgeon. For a security camera, the security staff could watch over a complete hall with eye movement speed, quickly scanning the environment for suspicious activity. Applied to different monitors, the efficiency gain would be even greater. In this work, we used a commercial eye tracker to capture the subject’s gaze. This gaze is then mapped to a screen where the camera images (two images are overlaid in red and cyan for a 3D impression) are presented. The gaze position on the image is mapped to the estimated depth map, from which the focal length of the camera is automatically adjusted. This enables the user to quickly explore the scene without manually adjusting the camera’s focal length. The depth map creation for the complete scene takes ≈20ms (single core from i5), whereas the preprocessing for each image takes ≈15ms. For depth map Preprint submitted to Springer creation, we record 20 images with different focal lengths, although this process can be completed with fewer images by trading off accuracy. This leads to a system update time of 332ms based on our cameras frame rate (60Hz or 16.6ms per frame), but buffering delays increase this time to ≈450ms. It is worth noticing that this process can be sped-up through a faster camera, multiple cameras, and GPU/multicore processing but is limited by the reaction time of the optotune lens (2.5ms [17]), which is used to change the focus. In the following sections, we use depth as an index in relation to the acquired image set. The real world depth can be calculated using the resulting focal length (acquired from the index) and the camera parameters. 2. Related work To compute a 3D representation of an area given a set of images produced with different camera focal lengths, two steps have to be applied. The first step is measuring or calculating how focused each pixel in this set is. Afterwards, each pixel is assigned an initial depth value (usually the maximum response) on which the 3D reconstruction is performed. In the sequence, we describe the shape-from-focus measuring operators briefly, grouping then similarly to Pertuz et al. [19] – which we refer the reader to for a wider overview. • Gradient-based measure operators are first or higher order derivatives of the Gaussian and are commonly applied for edge detection. The idea here is that unfocused or blurred edges have a lower response than sharp or focused edges. The best performing representatives according to [19] are November 10, 2017 first-order derivatives [21, 8], second central moment on first order derivatives [18], and the second central moment on a Laplacian (or second order derivatives) [18]. • Statistics-based measurements are based on calculated moments of random variables. These random variables are usually small windows shifted over the image. The idea behind statistics for focus measurements is that the moments (especially the second central moment) reach their maximum at focused parts of the image. According to [19], the best representatives are [32] using Chebyshev moments, second central moment of the principal components obtained from the covariance matrix [28], second central moment of second central moments in a window [18], and second central moment from the difference between the image and a blurred counterpart [9, 24, 26, 12]. Figure 1: The system for image recording consisting of a digital camera (XIMEA mq013mge2) and an optotune lens (el1030). On the left side, the subject with eye tracker looking at the image visualization is shown. The same subject with 3D goggles is shown on the right. focus measure based on the result of a least squares optimization technique is computed. These results are combined and used as depth estimation [1, 2, 27, 14]. • Frequency-based measures transform the image to the frequency domain, which is usually used in image compression. These transformations are Fourier, wavelet or curvelet transforms. Afterwards, the coefficients of the base functions are summed [31, 30, 11] or the statistical measures are applied on the coefficients [19]. The idea behind frequency-based focus measure is that the need for many base functions (or non zero coefficients) to describe an image is a measure of complexity or structure in the image. This amount of structure or complexity is the measure of how focused the image is. • Surface fitting: Here the samples for the fitting procedure are volumes around a pixel of focus measures. The surface is fitted to those samples, and the value aligned (in direction of the set of measurements) to the pixel is used as a new value. The improvement to the Gaussian or polynomial fit is that the neighborhood of a pixel influences its depth estimation too. This approach together with a neuronal network for final optimization has been proposed by [4]. • Texture-based measures use recurrence of intensity values [10, 24, 26], computed locally binary patterns [13], or the distance of orthogonally computed gradients in a window [12]. The idea here is equivalent to the frequency-based approaches, meaning that the amount of texture present (complexity of the image) is the measure of how focused the image is. 3. Setup description As shown in Figure 1, the setup consists of a binocular Dikablis Professional eye tracker [6], a conventional computer visualizing the image from the camera, and an optotune lens. The optotune lens has a focal tuning range of 50mm to 120mm [17], which can be adjusted online over the lens driver (serial communication). The reaction time of the lens is 2.5ms [17]. We used the XIMEA mq013mge2 digital camera shown in Figure 1 with a frame rate of 60 Hz and resolution 1280x1024 (we used the binned image 640x512). For estimating the subjects gaze we used the software EyeRec [22] and the pupil center estimation algorithm ElSe [7]. The calibration was performed with a nine-point grid, fitting a second order polynomial to the pupil center in the least squares sense. Additionally, it has to be noted that our setup includes a set of fixed lenses between the optotune lens and the object, which could not be disclosed at the time of writing (due to a NonDisclosure Agreement). Regarding 3D reconstructions, common methods are: • Gaussian and polynomial fit: These techniques fit a Gaussian [16] or polynomial [25] to the set of focus measures. To accomplish this, samples are collected outgoing from the maximum response of a pixel in the set of measurements (for each image, one measurement) in both directions. The maximum of the resulting Gaussian or polynomial is then used as depth estimate. • Surface fitting: Here the samples for the fitting procedure are volumes around a pixel of focus measures. The surface is fitted to those samples, and the value aligned (in direction of the set of measurements) to the pixel is used as a new value. The improvement to the Gaussian or polynomial fit is that the neighborhood of a pixel influences its depth estimation too. This approach together with a neuronal network for final optimization has been proposed by [4]. 4. Application The Graphical User Interface (GUI) of the system can be seen in Figure 2; in this GUI, the subjects gaze is mapped to the viewing area through marker detection and transformation between the eye tracker coordinates to screen coordinates. The top two images are from two cameras with optotune lenses. Their • Dynamic programming: In this approach, the volume is divided into subvolumes. For each sub-volume, an optimal 2 (a) Input (b) Magnitude (c) Edges (d) Filtered magnitude Figure 2: The GUI of the system. In the top row, the images from the two cameras with optotune lenses are shown. The depth map for those is on the right. The correspondence is marked by an arrow. Markers on the left side are used to map the gaze coordinates from the head-mounted eye tracker to the subjects view area. For a 3D representation to the user, we overlay the images from both cameras in red and cyan, which can be seen in the subjects viewing area. Figure 4: Canny edge based in focus estimation for one input image 4a. In 4b and 4c the output of the canny edge filter is shown and the filtered magnitude image in 4d. in classical shape-from-focus methods. Therefore, we use the Canny edge detector [5] as focus measure. The applied filter is the first derivative of a Gaussian. The resulting edges are used to filter the magnitude response of the filter, allowing only values with assigned edges to pass. For each filtered pixel magnitude, a maximum map through the set of responses is collected. In this map, most of the pixels have no value assigned. Additionally, it has to be noticed that the same edge can be present in this map multiple times because the changing focal length influences the field of view of the camera. This leads to tracing edges in the maximum map. After computing and filtering the focus measures of the image set, they have to be separated into parts. Therefore, we need candidates representing a strong edge part and their corresponding edge trace to interpolate the depth estimation for the candidate pixel. The candidate selection is performed by only selecting local maxima using an eight-connected neighborhood in the maximum map. These local maxima are used to build a graph representing the affiliation between candidates. This graph is build using the Delaunay triangulation, connecting candidates without intersections. The separation of this graph into a maximum response and edge trace responses is performed by separating nodes that are maximal in their neighborhood from those that are not. For interpolation of the depth value of maximal nodes, nonmaximal nodes are assigned based on their interconnection to the maxima and to an additional nonmaximal node. Additionally the set of responses is searched for values at the same location since the influence of the field of view does not affect all values in the image, and, as a result, centered edges stay at the same location. The interpolation is performed fitting a Gaussian (as in [16]) to all possible triple assignments and using the median of all results. Figure 3: The algorithmic workflow. The gray boxes are in an output of the algorithm. White boxes with rounded corners are algorithmic steps. depth maps can be seen on the right side and the correspondence is indicated by an arrow. Based on a slight shift of both cameras it is possible to use the red-cyan technique, to achieve a 3D impression for the user too (Figure 1 on the right side). The focal length of both cameras is automatically set to the depth at the users gaze position. 5. Method All steps of the algorithm are shown in Figure 3. The input to the algorithm is a set of grayscale images recorded with different focal length. The images have to be in the correct order otherwise the depth estimation will assign wrong depth values to focused parts of the volume. The main idea behind the algorithm is to estimate the depth map only based on parts of the image in which the focus is measurable, and interpolate it to the surrounding pixels if possible. Regions, where the focus is measurable, are clear edges or texture in an image. Plain regions, for example, usually induce erroneous depth estimations, which have to be filtered out afterward, typically using a median filter 3 (a) Magnitude of maximum val- (b) Depth of maximum values ues (a) Magnitude trace (b) Depth trace , Figure 6: Maximum magnitude responses (6a) and the assigned depth index (6b) in the set of images. In comparison to figure 5 were only 19 images in the input set where used, here the set consist of 190 images to show the traces more clear. (c) Local maxima 5.2. Graph Representation After in each image the focus measure was applied and filtered, the maximum along z of each location is collected in a maximum map (Figure 5a, Equation 1). (d) Graph representation Figure 5: Maximum responses in the set of images. In 5a the maximum magnitude for each location collected in the image is shown (black means no measurement collected) and the corresponding depth values in 5b. 5c shows the local maxima (pixels increased for visualization) of the maximum magnitude map 5a on which a Delaunay triangulation is applied resulting in a graph representation 5d. M(x, y) = maxz (V (x, y, z)) ( z D(x, y) = 0 M(x,y) ∈ V(x,y,z) M(x,y)=0 (1) (2) Equation 1 calculates the maximum map M (Figure 5a) where V represents the volume of filtered focus measures (one for each image). The coordinates x, y correspond to the image pixel location, and z is the image index in the input image. Equation 2 is the corresponding depth or z-index map D (Figure 5b) where an image set position is assigned to each maximum value. In Figure 5a and its corresponding depth map 5b, it can be seen that not every pixel has a depth estimation. Additionally, most of the collected edges have traces, meaning that the edge was collected in images recorded with different focal length. The trace occurs because changes in focal length induce a scaling of the field of view of the camera. In Figure 6, these traces and their occurrences are shown more clear due to the increased amount of the input image set (190 images). For Figure 5, we used 19 images in the input set. The bottom right part of Figure 6a shows that the occurrence of those traces is not present as strongly as in the other parts. This is due to the lens center (in our setup bottom right) from which the field of view scale impact increases linearly in distance. The next step of the algorithm is the computation of local maxima (Figure 5c) and, based on those, setting up a graph by applying the Delaunay triangulation (Figure 5d). The idea behind this step is to abstract the depth measurements, making it possible to estimate the depth of plain surfaces (as long as their borders are present) without the need of specifying a window size. Additionally, this graph is used to assign a set of depth estimations to one edge by identifying connected traces. These values are important because the set of input images does not have to contain the optimal focus distance of an edge. Therefore, the set of depth values belonging to one edge are used to interpolate its depth value. The graph spanned by the maxima nodes and the corresponding interpolated depth values are now the representation of the depth map. For further error correction, an interdependent median filter is applied to each node and its direct neighbors in a non-iterative way to ensure convergence. The last part of the algorithm is the conversion of this graph into a proper depth map. It has to be noticed that this graph is a set of triangles spanned between maximal nodes. Therefore, each pixel in the resulting depth map can be interpolated using its barycentric coordinates between the three assigned node values of the triangle it is assigned to. Pixels not belonging to a triangle have no assigned depth value. All steps are described in the following subsections in more detail. 5.1. Focus Measurement Figure 4 shows the first step of the algorithm for one input image 4a. We used the canny edge filter [5] with the first derivax2 +y2 1 2σ 2 ∂ ∂ ). The tive of a Gaussian as kernel (N 0 (x, y) = 2πσ 2e ∂x ∂y response (magnitude) of the convolution with this kernel is visualized in the Figure 4b. For √ σ , which is the standard deviation of the Gaussian, we used 2. After adaptive threshold selection (95% are not edges) and nonmaximum suppression of the Canny edge filter, we use the resulting edges (Figure 4c) as filter mask. In other words, only magnitude values assigned to a valid edge are allowed to pass. The stored magnitude responses for the input image are shown in Figure 4d. The idea behind this step is to restrict the amount of information gathered per image, consequently reducing the impact of noise on the algorithm. These two parameters (σ and non-edge ratio) are the only variables of the proposed method. 4 Algorithm 1 Algorithm for candidate selection where CAN(a) are all candidates for node a, CN(a) are the connected neighbors to node a, V the set of focus measure responses for each input frame, z the frame index, D(a) the depth index of node a, Gall all local maxima, Gmax all maximal nodes and Gnonmax all not maximal nodes. Require: Gall , Gmax , Gnonmax ,V (a) Nodes Gmax function Selectcorrespondences (Gall , Gmax , Gnonmax ) for a ∈ Gmax do for b ∈ CN(Gall , a), b ∈ Gnonmax do if D(a) , D(b) then add(CAN(a), b) end if for c ∈ CN(Gall , b) AND c ∈ Gnonmax do if D(a) , D(c) then add(CAN(a), c) end if end for end for for z ∈ V (a), V (a, z) > 0 do if D(a) , z then add(CAN(a), z) end if end for end for return CAN end function (b) Nodes Gnonmax Figure 7: White dots represent node locations (pixels increased for visualization). In 7a the nodes which have an equal or larger magnitude value compared to their connected neighbors in Gall are shown. 7b show the remaining non maximal nodes of Gall after removing those in Gmax . The local maxima (5) are computed based on a eight connected neighborhood on the maximum magnitude map ( 5a). Based on those points, the Delaunay triangulation (5d) sets up a graph, where each triple of points creates a triangle if the circumcircle does not contain another point. This graph Gall (Figure 5) contains multiple maxima from the same edge on different depth plains. To separate those, we introduce two types of nodes: a maximal response set Gmax (Figure 7a) and a non maximal response set Gnonmax (Figure 7b). Gmax = ∀i ∈ Gall , ∀ j ∈ CN(Gall , i), V ( j) ≤ V (i) Gnonmax = ∀i ∈ Gall , i < Gmax (3) Equation 5 describes the correspondences collection based on the collected candidates (CAN(a)) belonging to node a. The equations after the large bracket are the conditions, where D(a) ~ ac is the angle between is the depth index of node a and ab]~ ~ the two vectors ab and a~c. In our implementation, we used −0.95 (-1 corresponds to π) because an image is a solid grid and floating point inaccuracy. (4) Equation 3 is used to build the maximal response set Gmax (Figure 7a) where i is a Node in Gall and CN(Gall , i) delivers all connected neighbors of i. Therefore only nodes with an equal or higher magnitude value compared to their connected neighbors belong to Gmax . Gnonmax (Figure 7b) consists of all nodes in Gall which are not in Gmax and specified in equation 4. 5.4. Interpolation For estimating the real depth of a node in Gmax , we used the three point Gaussian fit technique proposed by Willert and Gharib [29] and first used for depth estimation by Nayar et D−D̄ al. [16]. The assumed Gaussian is M = M peak e−0.5 σ where M is the focus measure response, D the depth, and σ the standard deviation of the Gaussian. This can be rewritten with the D̄ natural logarithm ln(M) = ln(M peak ) − 0.5 D− σ . D̄ is the depth value where the Gaussian has the highest focus measure (mean) and obtained using equation 6. 5.3. Node Correspondence Collection Algorithm 1 performs the candidate selection. Candidates are possible node correspondences and marked by CAN(a), where a is the index node for the assignment. For each node in Gmax , connected nodes in Gnonmax with a different depth value are collected. Since we want to collect all nodes that could possibly build a line over a trace and the maximum could be the last or first measurement, we have to collect the connected nodes to the node from Gnonmax too. In case the node from Gmax is close to the lens center, where the scaling has low to none impact, we have to search in the volume of responses (V ) as well. After all candidates are collected, each pair has to be inspected to be a possible line trace or, in other words, a valid pair of corresponding focus measures. COR(a) = ∀b, c ∈ CAN(a),   b , c D(a) , D(b) , D(c)  ~ ab]~ ac = π M + (a, b) = ln(M(a)) − ln(M(b)) M − (a, b, c) = M + (a, b) + M + (a, c) D2− (a, b) = D(a)2 − D(b)2 ∆D(a, c) = 2|D(a) − D(c)| D̄(a, b, c) = (6) M + (a, c) ∗ D2− (a, b) ∆D(a, c) ∗ M − (a, b, c) In equation 6, a, b, c are node triples obtained from COR(a), where M is the focus measure, and D is the depth value (we used the same letters as in equation 1 and 2 for simplification (5) 5 (a) Barycentric Figure 9: In 9a the barycentric coordinates of a triangle spanned by nodes A,B and C is shown. The gray dot P in the middle of this triangle has coordinates a,b and c which is related to its distance to A, B and C. 9b shows an exemplary interpolation in such a triangle, where the numbers next to each corner are the intesity value of the corner pixel. Figure 8: The graph build on Gmax using Delaunay triangulation. and want to note that it is not valid for nonmembers of Gall , which are obtained through the response volume (1) ). Since COR(a) can have more than one pair of possible interpolations, we use the median over all possible interpolation values (D(a) = Median({D̄(a, b, c)}), ∀b, c ∈ COR(a)). 5.5. Rebuild Graph For using those interpolated nodes in Gmax as image representation, we have to rebuild the graph. Again we use the Delaunay triangulation with the result shown in Figure 8. Due to possible errors from the interpolation or the focus measurement, we apply a median filter on the depth of the node and its neighborhood (D(a) = Median({D(CN(a)), D(a)})). This median interpolation is performed interdependently; in other words, the values are stored directly into the depth map D, therefore influencing the median filtering of its neighbors. We used this way of median filtering because it delivered slightly better results. A more time-consuming approach would be iteratively determining the median. However, such an iterative approach could lead to oscillation and therefore no convergence. (a) Depth map (b) 3D model Figure 10: In 10a (normalized) white is closer, dark gray is further away and black means that the depth measure could not estimate a depth value. The 3D model in 10b is generated with the depth map from 10a using matlab. triangle and a,b and c are the barycentric coordinates. An exemplary interpolated triangle can be seen in 9b. D(P) = a ∗ D(A) + b ∗ D(B) + c ∗ D(C) (8) For depth assignment to point P equation 8 is used where D is the depth value. The resulting depth map after linear interpolation of all triangles can be seen in Figure 10a. In this depth map, white is closer to the camera and dark is further away. Comparing Figure 10a to Figure 8 it can be seen that areas for which no enclosing triangle exists are treated as not measurable (black regions in Figure 10a). If an estimation for those regions is wanted, it is possible to assign those pixels the depth value of the closest pixel with depth information or to interpolate the depth value using barycentric coordinates of the enclosing polygon. The 3D reconstruction based on the depth map from Figure 10a can be seen in Figure 10b. 5.6. Depth Map Creation For depth map creation, the graph in Figure 8 has to be transformed into a surface. This is done by assigning each pixel in a triangle the weighted value of the depth estimations from the corner nodes. The weights are determined using the distance to each node. The idea behind this is to have linear transitions between regions with different depth values. This makes it more comfortable for the subject to slide over the scene with their gaze, without having an oscillatory effect of the focal length close to region borders. This can be achieved very fast using barycentric coordinates (Figure 9a) to linearly interpolate (Figure 9b) those three values, which is usually applied in computer graphics to 3D models. P(a, b, c) ∆PBC ∆PAC ∆PAB a= ,b = ,c = ∆ABC ∆ABC ∆ABC (b) Interpolation 6. Data sets In Figure 11, all objects of the new data sets are shown. We scanned each object in 191 steps over the complete range (focal tuning range of 50mm to 120mm [17]) of the optotune lens. Therefore each object set contains 191 images with a resolution of 640x512. The objects plastic towers, lego bevel, lego steps, and glass are coated with a grid of black lines which should simplify the (7) Equation 7 describes the transformation from Cartesian coordinates to barycentric coordinates, where A,B and C are the corner nodes of a triangle (Figure 9a), ∆ is the area of the spanned 6 Figure 11: Shows all objects used to generate the datasets. Below each object image stands the title which will be used further in this document. depth estimation. For the objects tape bevel and tape steps, we used regular package tape to reduce the focus information for different focal length. Objects cup, raspberry pi, foam, tin, and CPU cooler are real objects where tin and raspberry pi are scanned in an oblique position. All objects except CPU cooler are used for evaluation, whereas the said object is used in limitations because the laminate is interpolated to a flat surface (without modification of the algorithm). This is due to the oppression of the not maximal responses along the laminate (which do not represent real edges). For evaluation we also used zero motion from Suwajanakorn et al. [27] and balcony, alley and shelf from Möller et al. [14]. 7.1. Algorithm evaluation Since human perception of sharpness varies between subjects, it is very challenging to make exact depth measurements for all objects and to manually label those. Therefore, we decided to show the depth map for each algorithm (proposed, variational depth [14] and SFF) and the timings. In addition, we provide an evaluation against the best index in the image stack as seen by the authors. Furthermore, in the supplementary material, we provide the parameter settings for SFF. Since SFF was implemented in MATLAB, a comparison with regard to the runtime is not absolutely fair, but we argue, the results are close to what could be reached by a C implementation. For visualization and comparison purposes we normalized the depth maps, based on the input stack with a step size of 10 (between consecutive frames, i.e., the first frame has focus value 1 and the second would have 11). Invalid regions are colored red. The normalization of the algorithm variational depth [14] is always over the complete range because the implementation returns no corresponding scale. Figure 12 shows the results of our proposed method and the state-of-the-art. The first four rows represent the results of objects, where the surface is marked with a grid. The purpose here is to have a comparison based on more objects. For the plastic tower, the irregular transitions between the four regions are due to the triangle interpolation and, therefore, a result as accepted. The corresponding 3D map is shown in Figure 10b. In the third row (lego steps), it can be seen that our method is able to correctly detect not measurable regions. The closest result is WAVR with 60 times our runtime. For the tape bevel, we see a superior performance of GLVM, but our method is closest to its result in comparison to the others. For the tape steps object, our method outperforms related approaches in a fraction of runtime. The most difficult object in our data set is the cup, which contains a big reflection and only a small part is textured. The other methods estimate the rim to be closer to the camera than the body of the cup which (is not correct, our method labels large parts as not measurable). The 7. Evaluation For the evaluation, we used 20 images per object from the set of 191. Those images were equally spread, meaning that the change in focal length between consecutive images is constant. In addition to the objects in Section 6, we used the data sets provided by Suwajanakorn et al. [27] and Moeller et al. [14], which are recorded with a professional camera in the real world. For the evaluation, we did not change any parameter of our algorithm. We used the algorithm variational depth [14] with the parameters as specified by the authors on a GeForce GT 740 GPU. The shape-from-focus (SFF) measures we evaluate are modified gray level variance (GLVM), modified Laplacian [15] (MLAP), Laplacian in 3D window [3] (LAP3d), variance of wavelets coefficients [31] (WAVV) and ratio of wavelet coefficients [30] (WAVR) as Matlab implementation from Pertuz et al. [19, 20], since these are the best performing SFF measures in [19]. For optimal parameter estimation, we tried all focus measure filter sizes from 9 to 90 in a stepwise search of 9 as for the median filter size. Additionally, the Gauss interpolation was used. 7 Figure 12: The results on all new data sets in terms of the depth map and the runtime are shown. Red regions represent areas that are marked by the algorithm to be not measurable. Brighter means closer to the camera. 8 Figure 13: The results on the data sets alley [14], balcony [14], shelf [14] and zeromotion [27]. The first three data sets have a resolution of 1920x1080, whereas the last one has a resolution of 774x518 pixel. In the top row, each algorithm is named as shown in figure 12. In addition to the depth maps, we show the processing time of each algorithm. Red regions represent areas that were marked by the algorithm as not measurable. Brightness represents the distance to the camera (the brighter, the further away). valid region is estimated appropriately with a negligible error (white spot). The Raspberry PI is estimated correctly by all methods except for GLVM. For foam, all methods perform well. The last object of the new data set is a tin. The two bright spots on the left side of the result of our method are due to dust particles on the lens which get sharp on a closer focal length. The best performing algorithm for the tin is GLVM. Figure 13 presents the results on the data sets provided by Suwajanakorn et al. [27] and Moeller et al. [14]. In comparison to the related algorithms, the results achieved by our approach are not as smooth. For the data set balcony [14] our algorithm did not get the centered leaf correctly but the depth approximation of the remaining area is comparable to that achieved by the state of the art, while our result does not look as smooth. For the shelf [14] and zeromotion [27] datasets, our algorithm performs better because it detects the not measurable regions. For the algorithm variational depth [14], it was not possible to provide a depth map for zeromotion from [27] because of the algorithm crashes (we tried to crop the center out of the image with 640x512 and 640x480 but it is still used to crash). The purpose of this evaluation is not to show an algorithm which is capable of estimating correct depth maps for all scenarios but to show that our algorithm can produce comparable results in a fraction of runtime without changing any parameter. We provide all depth maps in the supplementary material and as a download, together with a Matlab script to show them in 3D. Method GLVM LAPM LAP3 WAVV WAVR VARDEPTH Proposed R1 R2 3.40 3.88 6.06 3.80 12.01 3.87 19.10 4.33 27.15 4.59 28.25 18.14 12.96 4.07 R3 6.76 11.48 13.47 19.87 32.53 10.29 6.96 R4 17.33 21.92 22.02 30.87 57.54 35.22 5.61 Table 1: Results for the data set tin. The values represent the mean absolute errors over the marked regions. The regions are named R1-4 and visualized as red for R1, green for R2, blue for R3 and cyan for R4. "plastic tower", "tape steps" and "tin" (see Figure 12). The same parameters as for the images in Figure 12 and 13 are used. The tabels 1, 2, 3, and 4 show the mean absolute error ( 1n |vi − vgt |) for the data sets "tin", "lego steps", "tape steps", and "plastic tower", respectively. Over each table we placed an image for the specific data set with the evaluated regions marked by different colors. Regions without depth estimation (marked red in Figure 12) are excluded in the calculation of the mean absolute error. Our methods shows similar performance to the state of the art with less computational time. 7.1.1. Best index evaluation For evaluation against the best index in the image stack, we selected a region and an index in which the authors see this region as best focused. It has to be mentioned that we recorded for each data set 191 images with different focal length. For the depth map reconstruction, we used 19 equally spaced images. For evaluation we used the image stacks from "lego steps", 9 Method GLVM LAPM LAP3 WAVV WAVR VARDEPTH Proposed R1 8.26 12.85 12.77 12.98 14.13 19.61 3.67 R2 4.65 5.28 5.27 5.33 4.92 0.72 4.14 R3 3.78 3.75 3.61 3.57 3.09 2.68 5.28 R4 14.86 11.68 11.80 11.77 11.76 12.98 12.90 Method GLVM LAPM LAP3 WAVV WAVR VARDEPTH Proposed R1 21.56 45.86 45.86 34.84 56.53 52.89 12.39 R2 R3 R4 14.05 25.83 2.18 62.75 39.90 63.64 62.75 39.90 63.64 59.93 24.33 43.74 65.62 95.51 112.41 61.42 104.43 109.43 19.14 6.27 15.93 Table 2: Results for the data set lego steps. The values represent the mean absolute errors over the marked regions. The regions are named R1-4 and visualized as red for R1, green for R2, blue for R3 and cyan for R4. Table 3: Results for the data set tape steps. The values represent the mean absolute errors over the marked regions. The regions are named R1-4 and visualized as red for R1, green for R2, blue for R3 and cyan for R4. 8. Limitations and a runtime 6.91sec on a GeForce GT 740 GPU. For calculation, we used 20 images as for the evaluation and did not change the Canny parameters for the proposed algorithm. The algorithm is capable of determining plain surfaces if valid measures surrounding this surface are present. This advantage comes at the same time with the drawback that a nonexisting surface is interpolated in case of invalid measures in a region. This effect is depicted in Figure 14d, where the laminate of the CPU cooler is interpreted as a plain surface. The proposed algorithm could determine the lower path in the center of the image but, as can be seen in Figure 14e, all Gmax nodes are on the top of the laminate resulting in a wrong interpretation. For a better reconstruction of the CPU cooler, we adapted the proposed algorithm by simply not removing the Gnonmax nodes for the second graph building step (meaning we used all nodes in Gall ). As can be seen in Figure 14f, the resulting depth map got the laminate correctly but still fails for the right-skewed part, which is because there was no valid focus measurement. This can be seen in the graph shown in Figure 14g. The disadvantages of using all nodes are the slightly increased runtime (296ms instead of 283ms) due to additional costs introduced by interpolation. Additionally, some nodes are faulty because they either belong to an edge trace or are an invalid measurement. To avoid this, an iterative median filter over should be applied resulting in even higher runtimes. Another approach for improvement would be filtering nodes which have been selected for interpolation. However, this would not reduce the number of faulty measurements. For comparison, we used the modified gray level variance with filter size 18, median filter size 63 and Gaussian interpolation, which delivered the best result. The runtime was 7.06sec (Matlab implementation from Pertuz et al. [19, 20]) and is shown in Figure 14b. The black regions are added to the depth map because of zero reliability, otherwise, the focus measure has everywhere a result. The algorithm variational depth [14] produced the result in Figure 14c with default parameters 9. Conclusion In this paper, we showed the use of eye tracking for automatically adapting, focus to the presented image. Due to the realtime capability, our methods can be beneficial in various applications where autofocus facilitates interaction (e.g. surgery, security, human-robot collaboration, etc.). The algorithm proposed here shows similar performance as the state of the art but requires minimal computational resources and requires no parameter adjustment. In our future work, we will further optimize our method and provide a GPU implementation to make it computationally possible to use more than only one focus measuring method (e.g. GLVM). For evaluation purpose, we also have to increase the data sets by different materials and structures. References [1] Ahmad, M. B., Choi, T.-S., 2005. A heuristic approach for finding best focused shape. IEEE Transactions on Circuits and Systems for Video Technology 15 (4), 566–574. [2] Ahmad, M. B., Choi, T.-S., 2006. Shape from focus using optimization technique. In: 2006 IEEE International Conference on Acoustics Speech and Signal Processing Proceedings. Vol. 2. IEEE, pp. II–II. [3] An, Y., Kang, G., Kim, I.-J., Chung, H.-S., Park, J., 2008. Shape from focus through laplacian using 3d window. In: 2008 Second International Conference on Future Generation Communication and Networking. Vol. 2. IEEE, pp. 46–50. [4] Asif, M., Choi, T.-S., et al., 2001. Shape from focus using multilayer feedforward neural networks. IEEE Transactions on Image Processing 10 (11), 1670–1675. [5] Canny, J., 1986. A computational approach to edge detection. IEEE Transactions on pattern analysis and machine intelligence (6), 679–698. 10 (a) Input Method GLVM LAPM LAP3 WAVV WAVR VARDEPTH Proposed R1 6.26 4.70 4.74 9.74 7.57 8.35 3.27 R2 2.08 2.29 2.20 2.96 3.79 2.06 4.02 R3 2.37 3.20 3.22 3.75 3.95 4.24 2.57 R4 8.06 10.82 10.73 13.35 13.31 12.79 2.70 (b) Addapted (c) Variational (d) Proposed Gmax (e) Gmax graph (f) Proposed Gall (g) Gall graph Table 4: Results for the data set plastic tower. The values represent the mean absolute errors over the marked regions. The regions are named R1-4 and visualized as red for R1, green for R2, blue for R3 and cyan for R4. [6] Ergoneers, 2016. Dikablis glasses. [7] Fuhl, W., Santini, T., Kuebler, T., Kasneci, E., 03 2016. Else: Ellipse selection for robust pupil detection in real-world environments. In: ACM Symposium on Eye Tracking Research and Applications, ETRA 2016. [8] Geusebroek, J.-M., Cornelissen, F., Smeulders, A. W., Geerts, H., 2000. Robust autofocusing in microscopy. Cytometry 39 (1), 1–9. [9] Groen, F. C., Young, I. T., Ligthart, G., 1985. A comparison of different focus functions for use in autofocus algorithms. Cytometry 6 (2), 81–91. [10] Hilsenstein, V., 2005. Robust autofocusing for automated microscopy imaging of fluorescently labelled bacteria. In: Digital Image Computing: Techniques and Applications (DICTA’05). IEEE, pp. 15–15. [11] Huang, J.-T., Shen, C.-H., Phoong, S.-M., Chen, H., 2005. Robust measure of image focus in the wavelet domain. In: 2005 International Symposium on Intelligent Signal Processing and Communication Systems. IEEE, pp. 157–160. [12] Huang, W., Jing, Z., 2007. Evaluation of focus measures in multi-focus image fusion. Pattern recognition letters 28 (4), 493–500. [13] Lorenzo, J., Castrillon, M., Méndez, J., Deniz, O., 2008. Exploring the use of local binary patterns as focus measure. In: Computational Intelligence for Modelling Control and Automation, 2008 International Conference on. IEEE, pp. 855–860. [14] Moeller, M., Benning, M., Schönlieb, C., Cremers, D., 2015. Variational depth from focus reconstruction. IEEE Transactions on Image Processing 24 (12), 5369–5378. [15] Nayar, S. K., Ikeuchi, K., Kanade, T., 1989. Surface reflection: physical and geometrical perspectives. Tech. rep., DTIC Document. [16] Nayar, S. K., Nakagawa, Y., 1994. Shape from focus. IEEE Transactions on Pattern analysis and machine intelligence 16 (8), 824–831. [17] Optotune, 2016. Datasheet: El-10-30-series. [18] Pech-Pacheco, J. L., Cristóbal, G., Chamorro-Martinez, J., FernándezValdivia, J., 2000. Diatom autofocusing in brightfield microscopy: a comparative study. In: Pattern Recognition, 2000. Proceedings. 15th International Conference on. Vol. 3. IEEE, pp. 314–317. [19] Pertuz, S., Puig, D., Garcia, M. A., 2013. Analysis of focus measure operators for shape-from-focus. Pattern Recognition 46 (5), 1415–1432. [20] Pertuz, S., Puig, D., Garcia, M. A., 2013. Reliability measure for shapefrom-focus. Image and Vision Computing 31 (10), 725–734. [21] Russell, M. J., Douglas, T. S., 2007. Evaluation of autofocus algorithms for tuberculosis microscopy. In: 2007 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE, pp. 3489–3492. [22] Santini, T., Fuhl, W., Geisler, D., Kasneci, E., 2017. Eyerectoo: Opensource software for real-time pervasive head-mounted eye tracking. In: VISIGRAPP (6: VISAPP). pp. 96–101. Figure 14: In 14a one image of the CPU cooler set can be seen. 14b is the result of SFF where we set the optimal parameters (modified gray level variance, filter size 18, median filter size 63, Gauss interpolation, runtime 7.06sec), where black means zero reliability. In 14c the result of the variational depth [14] with default parameters is shown (runtime 6.91sec on GPU). 14d is the output of the proposed method and the corresponding Gmax graph in 14e (runtime 283ms). Result of the proposed method withy not removing non maximal 14f and the corresponding graph 14g (runtime 296ms). [23] Santini, T., Fuhl, W., Kübler, T., Kasneci, E., 2016. Bayesian identification of fixations, saccades, and smooth pursuits. In: Proceedings of the Ninth Biennial ACM Symposium on Eye Tracking Research &amp; Applications. ACM, pp. 163–170. [24] Santos, A., Ortiz de Solórzano, C., Vaquero, J. J., Pena, J., Malpica, N., Del Pozo, F., 1997. Evaluation of autofocus functions in molecular cytogenetic analysis. Journal of microscopy 188 (3), 264–272. [25] Subbarao, M., Choi, T., 1995. Accurate recovery of three-dimensional shape from image focus. IEEE Transactions on pattern analysis and machine intelligence 17 (3), 266–274. [26] Sun, Y., Duthaler, S., Nelson, B. J., 2004. Autofocusing in computer microscopy: selecting the optimal focus algorithm. Microscopy research and technique 65 (3), 139–149. [27] Suwajanakorn, S., Hernandez, C., Seitz, S. M., 2015. Depth from focus with your mobile phone. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition. pp. 3497–3506. [28] Wee, C.-Y., Paramesran, R., 2007. Measure of image sharpness using eigenvalues. Information Sciences 177 (12), 2533–2552. [29] Willert, C. E., Gharib, M., 1991. Digital particle image velocimetry. Experiments in fluids 10 (4), 181–193. [30] Xie, H., Rong, W., Sun, L., 2006. Wavelet-based focus measure and 3-d surface reconstruction method for microscopy images. In: 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems. IEEE, pp. 229–234. [31] Yang, G., Nelson, B. J., 2003. Wavelet-based autofocusing and unsupervised segmentation of microscopic images. In: Intelligent Robots and Systems, 2003.(IROS 2003). Proceedings. 2003 IEEE/RSJ International Conference on. Vol. 3. IEEE, pp. 2143–2148. [32] Yap, P. T., Raveendran, P., 2004. Image focus measure based on cheby- 11 shev moments. IEE Proceedings-Vision, Image and Signal Processing 151 (2), 128–136. 12 

JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014 1 Dual theory of transmission line outages arXiv:1606.07276v2 [nlin.AO] 22 Jan 2017 Henrik Ronellenfitsch, Debsankha Manik, Jonas Hörsch, Tom Brown, Dirk Witthaut Abstract—A new graph dual formalism is presented for the analysis of line outages in electricity networks. The dual formalism is based on a consideration of the flows around closed cycles in the network. After some exposition of the theory is presented, a new formula for the computation of Line Outage Distribution Factors (LODFs) is derived, which is not only computationally faster than existing methods, but also generalizes easily for multiple line outages and arbitrary changes to line series reactance. In addition, the dual formalism provides new physical insight for how the effects of line outages propagate through the network. For example, in a planar network a single line outage can be shown to induce monotonically decreasing flow changes, which are mathematically equivalent to an electrostatic dipole field. Index Terms—Line Outage Distribution Factor, DC power flow, dual network, graph theory I. I NTRODUCTION The robustness of the power system relies on its ability to withstand disturbances, such as line and generator outages. The grid is usually operated with ‘n−1 security’, which means that it should withstand the failure of any single component, such as a transmission circuit or a transformer. The analysis of such contingencies has gained in importance with the increasing use of generation from variable renewables, which have led to larger power imbalances in the grid and more situations in which transmission lines are loaded close to their thermal limits [1]–[6]. A crucial tool for contingency analysis is the use of Line Outage Distribution Factors (LODFs), which measure the linear sensitivity of active power flows in the network to outages of specific lines [7]. LODFs are not only used to calculate power flows after an outage, but are also employed in security-constrained linear optimal power flow (SCLOPF), where power plant dispatch is optimized such that the network is always n − 1 secure [7]. LODF matrices can be calculated from Power Transfer Distribution Factors (PTDFs) [8], [9], which describe how power flows change when power injection is shifted from one node to another. In [10], a dual method for calculating PTDFs was presented. The dual method is based on an analysis of the flows around closed cycles (closed cycles are paths that are non-intersecting and start and end at the same node [11]) in the network graph; for a plane graph, a basis of these closed H. Ronellenfitsch and D. Manik are at the Max Planck Institute for Dynamics and Self-Organization (MPIDS), 37077 Göttingen, Germany. H. Ronellenfitsch is additionally at the Department of Physics and Astronomy, University of Pennsylvania, Philadelphia PA, USA. J. Hörsch and T. Brown are at the Frankfurt Institute for Advanced Study, 60438 Frankfurt am Main, Germany. D. Witthaut is at the Forschungszentrum Jülich, Institute for Energy and Climate Research - Systems Analysis and Technology Evaluation (IEK-STE), 52428 Jülich, Germany and the Institute for Theoretical Physics, University of Cologne, 50937 Köln, Germany. cycles corresponds to the nodes of the dual graph [11]. Seen as a planar polygon, each basis cycle corresponds to one facet of the polygon. (This notion of dual of a plane graph is called the weak dual). In this paper the dual formalism is applied to derive a new direct formula for the LODF matrices, which is not only computationally faster than existing methods, but has several other advantages. It can be easily extended to take account of multiple line outages and, unlike other methods, it also works for the case where a line series reactance is modified rather than failing completely. This latter property is relevant given the increasing use of controllable series compensation devices for steering network power flows. Moreover, the dual formalism is not just a calculational tool: it provides new insight into the physics of how the effects of outages propagate in the network, which leads to several useful results. Depending on network topology, the dual method can lead to a significant improvement in the speed of calculating LODFs. Thus it can be useful for applications where LODFs must be calculated repeatedly and in a time-critical fashion, for instance in ‘hot-start DC models’ or ‘incremental DC models’ [12]. The dual method we describe is particularly suited to these types of problems because unlike in the primal case, most of the involved matrices only depend on the network topology and can be stored and reused for each calculation run. II. T HE PRIMAL FORMULATION OF LINEARIZED NETWORK FLOWS In this paper the linear ‘DC’ approximation of the power flow in AC networks is used, whose usefulness is discussed in [13], [14]. In this section the linear power flow formulation for AC networks is reviewed and a compact matrix notation is introduced. In the linear approximation, the directed active power flow Fℓ on a line ℓ from node m to node n can be expressed in terms of the line series reactance xℓ and the voltage angles θm , θn at the nodes Fℓ = 1 (θm − θn ) = bℓ (θm − θn ), xℓ (1) where bℓ = 1/xℓ is the susceptance of the line. In the following we do not distinguish between transmission lines and transformers, which are treated the same. The power flows of all lines are written in vector form, F = (F1 , . . . , FL )t ∈ RL , and similarly for the nodal voltage angles θ = (θ1 , . . . , θN )t ∈ RN , where the superscript t denotes the transpose of a vector or matrix. Then equation (1) can be written compactly in matrix form F = B d K t θ, (2) JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014 2 where B d = diag(b1 , . . . , bL ) ∈ RL×L is the diagonal branch susceptance matrix. The incidence matrix K ∈ RN ×L encodes how the nodes of the directed network graph are connected by the lines [15]. It has components   1 if line ℓ starts at node n, −1 if line ℓ ends at node n, (3) Kn,ℓ =  0 otherwise. In homology theory K is the boundary operator from the vector space of lines ∼ = RL to the vector space of nodes ∼ = RN . The incidence matrix also relates the nodal power injections at each node P = (P1 , . . . , PN ) ∈ RN to the flows incident at the node P = KF . (4) This is Kirchhoff’s Current Law expressed in terms of the active power: the net power flowing out of each node must equal the power injected at that node. Combining (2) and (4), we obtain an equation for the power injections in terms of the voltage angles, P = Bθ. (5) Here we have defined the nodal susceptance matrix B ≡ KB d K t ∈ RN ×N with the components ( X bℓ if m = n; ℓ∈Λm Bm,n = (6) −bℓ if m is connected to n by ℓ, where Λm is the set of lines which are incident on m. The matrix B is a weighted Laplace matrix [15] and equation (5) is a discrete Poisson equation. Through equations (2) and (5), there is now a linear relation between the line flows F and the nodal power injections P . For a connected network, B has a single zero eigenvalue and therefore cannot be inverted directly. Instead, the MoorePenrose pseudo-inverse B ∗ can be used to solve (5) for θ and obtain the line flows directly as a linear function of the nodal power injections F = BdK tB∗P . (7) This matrix combination is taken as the definition of the nodal Power Transfer Distribution Factor (PTDF) PTDF ∈ RL×N PTDF = B d K t B ∗ . (8) Next, the effect of a line outage is considered. Suppose the flows before the outage are given by Fk and the line which (ℓ) fails is labeled ℓ. The line flows after the outage of ℓ, Fk are linearly related to the original flows by the matrix of Line Outage Distribution Factors (LODFs) [7], [16] (ℓ) Fk = Fk + LODFkℓ Fℓ , (9) where on the right hand side there is no implied summation over ℓ. It can be shown [8], [9] that the LODF matrix elements can be expressed directly in terms of the PTDF matrix elements as [PTDF · K]kℓ . (10) LODFkℓ = 1 − [PTDF · K]ℓℓ For the special case of k = ℓ, one defines LODFkk = −1. The matrix [PTDF · K]kℓ can be interpreted as the sensitivity of the flow on k to the injection of one unit of power at the from-node of ℓ and the withdrawal of one unit of power at the to-node of ℓ. III. C YCLES AND THE D UAL G RAPH The power grid defines a graph G = (V, E) with vertex set V formed by the nodes or buses and edge set E formed by all transmission lines and transformers. The orientation of the edges is arbitrary but has to be fixed because calculations involve directed quantities such as the real power flow. In the following we reformulate the theory of transmission line outages in terms of cycle flows. A directed cycle is a combination of directed edges of the graph which form a closed loop. All such directed cycles can be decomposed into a set of L−N +1 fundamental cycles, with N being the number of nodes, L being the number of edges and assuming that the graph is connected [11]. An example is shown in Fig. 1, where two fundamental cycles are indicated by blue arrows. The fundamental cyles are encoded in the cycle-edge incidence matrix C ∈ RL×(L−N +1)   1 if edge ℓ is element of cycle c, −1 if reversed edge ℓ is element of cycle c, Cℓ,c =  0 otherwise. (11) It is a result of graph theory, which can also be checked by explicit calculation, that the L − N + 1 cycles are a basis for the kernel of the incidence matrix K [11], KC = 0 . (12) Using the formalism of cycles, the Kirchoff Voltage Law (KVL) can be expressed in a concise way. KVL states that the sum of all angle differences along any closed cycle must equal zero, X (θi − θj ) = 0 . (13) (ij)∈cycle c Since the cycles form a vector space it is sufficient to check this condition for the L − N + 1 basis cycles. In matrix form this reads C tK tθ = 0 , (14) which is satisfied automatically by virtue of equation (12). Using equation (2), the KVL in terms of the flows reads C t X d F = 0, (15) where X d is the branch reactance matrix, defined by X d = diag(x1 , . . . , xL ) = diag(1/b1 , . . . , 1/bL ) ∈ RL×L . The results of Sections I through V apply for any graph. In the final Section VI, a special focus is made on planar graphs, i.e., graphs which can be drawn or ‘embedded’ in the plane R2 without edge crossings. Once such an embedding is fixed, the graph is called a plane graph. Power grids are not naturally embedded in R2 , but while line crossings are possible, they are sufficiently infrequent in large scale transmission grids (such as the high-voltage European transmission grid). The embedding (drawing) in the plane yields a very intuitive approach to the cycle flow formulation. The edges separate polygons, which are called the facets of the graph. We can JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014 3 now define a cycle basis which consists exactly of these facets. Then all edges are part of at most two basis cycles, which is called MacLane’s criterion for panarity [11]. This construction is formalized in the definition of the weak dual graph DG of G. The weak dual graph DG is formed by putting dual nodes in the middle of the facets of G as described above, and then connecting the dual nodes with dual edges across those edges where facets of G meet [11], [15]. DG has L − N + 1 nodes and its incidence matrix is given by C t . The simple topological properties of plane graphs are essential to derive some of the rigorous results obtained in section VI. For more complex networks, graph embeddings without line crossings can still be defined – but not on the plane R2 . More complex geometric objects (surfaces with a genus g > 0) are needed (see, e.g., [17] and references therein). IV. D UAL THEORY OF NETWORK FLOWS In this section the linear power flow is defined in terms of dual cycle variables following [10], rather than the nodal voltage angles. To do this, we define the linear power flow equations directly in terms of the network flows. The power conservation equation (4) KF = P , (16) provides N equations, of which one is linearly dependent, for the L components of F . The solution space is thus given by an affine subspace of dimension L − N + 1. In section III we discussed that the kernel of K is spanned by the cycle flows. Thus, we can write every solution of equation (16) as a particular solution of the inhomogeneous equation plus a linear combination of cycle flows: F = F (part) + Cf , f ∈ RL−N +1 . (17) The components fc of the vector f give the strength of the cycle flows for all basis cycles c = 1, 2, · · · , L − N + 1. A particular solution F (part) can be found by taking the uniquely-determined flows on a spanning tree of the network graph [10]. To obtain the correct physical flows we need a further condition to fix the L − N + 1 degrees of freedom fc . This condition is provided by the KVL in (15), which provides exactly L − N + 1 linear constraints on f C t X d Cf = −C t X d F (part) . (18) Together with equation (16), this condition uniquely determines the power flows in the grid. Equation (18) is the dual equation of (5). If the cycle reactance matrix A ∈ RL−N +1×L−N +1 is defined by A ≡ C t X d C, then A also has the form,  P xℓ  ℓ∈κP c Acc′ = ±xℓ  ℓ∈κc ∩κc′ (19) where κc is the set of edges around cycle c and the sign ambiguity depends on the orientation of the cycles. The construction of A is very similar to the weighted Laplacian in equation (6); for plane graphs where the cycles correspond to the faces of the graph, this analogy can be made exact (see Section VI). Unlike B, the matrix A is invertible, due to the fact that the outer boundary cycle of the network is not included in the cycle basis. This is analogous to removing the row and column corresponding to a slack node from B, but it is a natural feature of the theory, and not manually imposed. V. D UAL COMPUTATION OF LINE OUTAGE DISTRIBUTION FACTORS A. Single line outages The dual theory of network flows derived in the previous section can be used to derive an alternative formula for the LODFs. For the sake of generality we consider an arbitrary change of the reactance of a transmission line ℓ, xℓ → xℓ + ξℓ . The generalization to multiple line outages is presented in the following section. The change of the network structure is described in terms of the branch reactance matrix X̂ d = X d + ∆X d = X d + ξℓ uℓ utℓ , if c = c ; if c 6= c, (20) (22) where uℓ ∈ RL is a unit vector which is 1 at position ℓ and zero otherwise. In this section we use the hat to distinguish the line parameters and flows in the modified grid after the outage from the original grid before the outage. This perturbation of the network topology will induce a change of the power flows F̂ = F + ∆F . (23) We consider a change of the topology while the power injections remain constant. The flow change ∆F thus does not have any source such that it can be decomposed into cycle flows ∆F = C∆f . (24) The uniqueness condition (15) for the perturbed network reads C t (X d + ∆X d )(F + ∆F ) = 0. (25) Using condition (15) for the original network and the cycle flow decomposition equation (24) for the flow changes yields C t X̂ d C∆f = −C t ∆X d F ⇒ ∆f = −(C t X̂ d C)−1 C t uℓ ξℓ utℓ F (26) such that the flow changes are given by ∆F = C∆f = −C(C t X̂ d C)−1 C t uℓ ξℓ utℓ F . ′ (21) (27) This expression suggests that we need to calculate the inverse separately for every possible contingency case, which would require a huge computational effort. However, we can reduce JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014 240 MW 0 MW 40 MW 26.8 MW 226.5 MW (a) 466.5 MW 323.5 MW 300 MW 300 MW 362.4 MW 0 MW 40 MW 37.6 MW 104.1 MW (b) 466.5 MW 323.5 MW 300 MW 300 MW (c) 122.4 MW 5 4 6 Thus, the flow changes can be written according to equation (24) with 122.4 MW 122.4 MW 1 64.4 MW 2 3 1 64.4 MW 64.4 MW 2 4 64.4 MW 5 ∆f = (122.4 MW, 64.4 MW)t . it to the inverse of A = C t X d C describing the unperturbed grid using the Woodbury matrix identity [18],
−1 = A−1 − A−1 C t uℓ ξℓ−1 + utℓ CA−1 C t uℓ Thus we obtain
−1 utℓ CA−1 .
−1 . We then obtain the induced cycle flows and flow change by inserting this expression into equation (27). We summarize our results in the following proposition. Proposition 1. If the reactance of a single transmission line ℓ is changed by an amount ξℓ , the real power flows change as −ξℓ Fℓ M uℓ 1 + ξℓ utℓ M uℓ −1 (28) t with the matrix M = CA C . If the line ℓ fails, we have ξℓ → ∞. The line outage distribution factor for a transmission line k is thus given by LODFk,ℓ = (cf. also [22], [23] for a discussion of cycle flows in power grids). The dual approach to the LODFs can be computationally advantageous for sparse networks as discussed in section V-C. Furthermore, we will use it to prove some rigorous results on flow redistribution after transmission line failures in section VI. B. Multiple line outages (C t X̂ d C)−1 C t uℓ = A−1 C t uℓ 1 + ξℓ utℓ CA−1 C t uℓ ∆F = (32) 3 Fig. 1. (a) DC power flow in a 5-bus test grid from [21]. (b) DC power flow in the same network after the outage of one transmission line. (c) The change of power flows can be decomposed into two cycle flows. (C t X̂ d C)−1 = A + C t uℓ ξℓ utℓ C (30) The grid contains 2 independent cycles, which are chosen as cycle 1: line 2, line 6, line 3. cycle 2: line 1, line 4, line 5, reverse line 2 The cycle-edge incidence matrix thus reads   0 +1 +1 0 0 +1 t C = . (31) +1 −1 0 +1 +1 0 400 MW 14.1 MW 314.1 MW 170 MW example, the node-edge incidence matrix is given by   +1 −1 +1 0 0 0 −1 0 0 +1 0 0    0 0 −1 +1 0  K= 0 .  0 +1 0 0 −1 −1 0 0 −1 0 0 +1 400 MW 50.3 MW 249.7 MW 170 MW 4 ∆Fk ut M uℓ = − kt . Fℓ uℓ M uℓ (29) Note that the formula for an arbitrary change in series reactance of a line is useful for the assessment of the impact of flexible AC transmission (FACTS) devices, in particular series compensation devices [19] or adjustable inductors that clamp onto overhead lines [20]. Finally, an example of failure induced cycle flows and the corresponding flow changes is shown in Figure 1. In the The dual approach can be generalized to the case of multiple damaged or perturbed transmission lines in a straightforward way. Consider the simultaneous perturbation of the M transmission lines ℓ1 , ℓ2 , . . . , ℓM according to xℓ1 → xℓ1 + ξℓ1 , xℓ2 → xℓ2 + ξℓ2 , . . . , xℓM → xℓM + ξℓM . The change of the branch reactance matrix is then given by ∆X d = U Ξ U t , (33) where we have defined the matrices Ξ = diag(ξℓ1 , ξℓ2 , . . . , ξℓM ) ∈ RM×M , U = (uℓ1 , uℓ2 , . . . , uℓM ) ∈ RN ×M . The formula (27) for the flow changes then reads  −1 ∆F = −C C t X̂ d C C t U Ξ U tF . (34) To evaluate this expression we again make use of the Woodbury matrix identity [18], which yields  −1 C t X̂ d C =
−1 t A−1 − A−1 C t U Ξ−1 + U t M U U CA−1 . We then obtain the flow change by inserting this expression into equation (34) with the result
−1 Ξ U t F . (35) ∆F = −CA−1 C t U 1l + Ξ U t M U In case of a multiple line outages of lines ℓ1 , . . . , ℓm we have to consider the limit ξℓ1 , . . . , ξℓM → ∞. (36) JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014 5 TABLE I C OMPARISON OF CPU TIME FOR THE CALCULATION OF THE PTDF S IN SPARSE NUMERICS . Test Grid name source case300 case1354pegase GBnetwork case2383wp case2736sp case2746wp case2869pegase case3012wp case3120sp case9241pegase [21] [25] [26] [21] [21] [21] [25] [21] [21] [25] nodes N 300 1354 2224 2383 2736 2746 2869 3012 3120 9241 Grid Size cycles speedup ratio L−N+1 L−N+1 N tconv tdual 110 357 581 504 760 760 1100 555 565 4967 0.37 0.26 0.26 0.21 0.28 0.28 0.38 0.18 0.18 0.54 1.83 4.43 4.09 4.20 3.27 3.35 2.79 3.93 3.96 1.31 In this limit equation (35) reduces to ∆F = −M U U t M U
−1 U tF . (37) Specifically, for the case of two failing lines, we obtain ∆Fk = Mk,ℓ1 Mℓ2 ,ℓ2 − Mk,ℓ2 Mℓ2 ,ℓ1 Fℓ Mℓ1 ,ℓ1 Mℓ2 ,ℓ2 − Mℓ1 ,ℓ2 Mℓ2 ,ℓ1 1 Mk,ℓ2 Mℓ1 ,ℓ1 − Mk,ℓ1 Mℓ1 ,ℓ2 Fℓ . + Mℓ1 ,ℓ1 Mℓ2 ,ℓ2 − Mℓ1 ,ℓ2 Mℓ2 ,ℓ1 2 (38) C. Computational aspects The dual formula (29) for the LODFs can be computationally advantageous to the conventional approach. To calculate the LODFs via equation (10) we have to invert the matrix B ∈ RN ×N to obtain the PTDFs. Using the dual approach the most demanding step is the inversion of the matrix A = C t X d C ∈ R(L−N +1)×(L−N +1), which can be much smaller than B if the network is sparse. However, more matrix multiplications need to be carried out, which decreases the potential speed-up. We test the computational performance of the dual method by comparing it to the conventional approach, which is implemented in many popular software packages such as for instance in M ATPOWER [21]. Conventionally, one starts with the calculation of the nodal PTDF matrix defined in Eq. (8). In practice, one usually does not compute the full inverse but solves the linear system of equations PTDF· B = B d K instead. Furthermore, one fixes the voltage phase angle at a slack node s, such that one can omit the sth row and column in the matrix B and the sth column in matrix B f = B d K T while solving the linear system. The result is multiplied by the matrix K from the right to obtain the PTDFs between the endpoints of all lines. One then divides each column ℓ by the value 1 − PTDFℓℓ to obtain the LODFs via formula (10). An implementation of these steps in M ATLAB is listed in the supplement [24]. The dual approach yields the direct formula (29) for the LODFs. To efficiently evaluate this formula we first compute the matrix M = CA−1 C t . Again we do not compute the full matrix inverse but solve a linear system of equations instead. The full LODF matrix is then obtained by dividing every column ℓ by the factor Mℓℓ . We evaluate the runtime for various test grids from [21], [25], [26] using a M ATLAB program listed in the supplement [24]. All input matrices are sparse, such that the computation is faster when using sparse numerical methods (using the command sparse in M ATLAB and converting back to full at the appropriate time). Then M ATLAB employs the high-performance supernodal sparse Cholesky decomposition solver C HOLMOD 1.7.0 to solve the linear system of equations. We observe a significant speed-up of the dual method by a factor between 1.31 and 4.43 depending on how meshed the grid is (see Table I). VI. T OPOLOGY OF CYCLE FLOWS In this section the propagation of the effects of line outages are analyzed using the theory of discrete calculus and differential operators on the dual network graph. There is a wide body of physics and mathematics literature on discrete field theory (see, e.g., [27]). We turn back to the cycle flows themselves and derive some rigorous results. These results help to understand the effects of a transmission line outage and cascading failures in power grids, in particular whether the effects are predominatly local or affect also remote areas of the grid (cf. [28]–[31]). We start with a general discussion of the mathematical structure of the problem and show that line outages affect only parts of the grid which are sufficiently connected. Further results are obtained for planar graphs (graphs that can be embedded in the plane without line crossings, which approximately holds, e.g., for the European high voltage transmission grid). We characterize the direction of the induced cycle flows and show that the effect of failures decreases monotonically with the distance from the outage. Finally we proceed to discuss nonlocal effects in non-planar networks. A. General results The starting point of our analysis of the topology of cycle flows is a re-formulation of Proposition 1. Lemma 1. The outage of a single transmission line ℓ induces cycle flows which are determined by the linear system of equations A∆f = q (39) with q = Fℓ (utℓ CA−1 C t uℓ )−1 C t uℓ and A = C t X d C. ∆f , q ∈ R(L−N +1) and A ∈ R . It will now be shown that in some cases this equation can be interpreted as a discrete Poisson equation for ∆f with Laplacian operator A and inhomogeneity q. This formulation is convenient to derive some rigorous results on flow rerouting after a transmission line failure. We first note from the explicit construction of A in equation (20) that two cycles in the dual network are only coupled via their common edges. The coupling is given by the sum of the reactances of the common edges. Generally, the reactance of a line is proportional to its length. The coupling of two cycles is then directly proportional to the total length of their common boundary, provided that the lines are all of the same Note that (L−N +1)×(L−N +1) JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014 (a) real power flow 0 18.5 (c) flow change 37 (b) cycle flows 6 0 0 3 6 (d) predicted change dc: inf 4 3 2 1 0 ? Fig. 2. Flow changes and cycle flows after a transmission line failure in a small test grid. (a) Real power flows in the initial intact network in MW. (b) The failure of a transmission line (dashed) must be compensated by cycle flows indicated by cyclic arrows. The thickness of the lines indicates the approximate strength of the cycle flows. (c) The resulting flow changes after the failure of the marked transmission line. (d) The direction of the flow changes can be predicted using Propositions 3 and 4 for all edges and the magnitude decreases with the cycle distance dc . The power flow in (a,c) has been calculated using the standard software M ATPOWER for the 30-bus test case [21]. type. Since the inhomogeneity q is proportional to C t uℓ , it is non-zero only for the cycles which are adjacent to the failing edge ℓ: qc 6= 0 only if ℓ is an element of cycle c. (40) The matrix A typically has a block structure such that a failure in one block cannot affect the remaining blocks. The dual approach to flow rerouting gives a very intuitive picture of this decoupling. To see this, consider the example shown in Figure 2. The cycle at the top of the network is connected to the rest of the network via one node. However, it is decoupled in the dual representation because it shares no common edge with any other cycle. Thus, a failure in the rest of the grid will not affect the power flows in this cycle—the mutual LODFs vanish. This result is summarized in the following proposition, and a formal proof is given in the supplement [24]. Proposition 2. The line outage distribution factor LODFk,ℓ between two edges k = (i, j) and ℓ = (s, r) vanishes if there is only one independent path between the vertex sets {r, s} and {i, j}. B. Planar networks Some important simplifications can be made in the specific case of a plane network. We can then define the cycle basis in terms of the interior faces of the graph which allows for a intuitive geometric picture of induced cycle flows as in Figures 2 and 3. For the remainder of this section we thus restrict ourselves to such plane graphs and fix the cycle basis by the interior faces and fix the orientation of all basis cycles to be counter-clockwise. Thus equation (39) is formulated on the weak dual of the original graph. Fig. 3. Schematic representation of the flow changes after the damage of a single edge (dashed). According to Mac Lane’s planarity criterion [11], every edge in a plane graph belongs to at most two cycles such that q has at most two-nonzero elements: One non-zero element qc1 if ℓ is at the boundary and two non-zero elements qc1 = −qc2 if the line ℓ is in the interior of the network. Furthermore, the matrix A is a Laplacian matrix in the interior of the network [15]. That is, for all cycles c which are not at the boundary we have X Adc = −Acc . (41) d6=c Up to boundary effects, equation (39) is thus equivalent to a discretized Poisson equation on a complex graph with a dipole source (monopole source if the perturbation occurs on the boundary). For plane networks we now prove some rigorous results on the orientation of cycle flows (clockwise vs. counterclockwise) and on their decay with the distance from the failing edge. In graph theory, the (geodesic) distance of two vertices is defined as as the number of edges in a shortest path connecting them [11]. Similarly, the distance of two edges is defined as the number of vertices on a shortest path between the edges. Proposition 3. Consider the cycle flows ∆f induced by the failure of a a single line ℓ in a plane linear flow network described by equation (39). The weak dual graph can be decomposed into at most two connected subgraphs (‘domains’) D+ and D− , with ∆fc ≥ 0 ∀c ∈ D+ and ∆fc ≤ 0 ∀c ∈ D− . The domain boundary, if it exists, includes the perturbed line ℓ, i.e. the two cycles adjacent to ℓ belong to different domains. A proof is given in the supplement [24]. The crucial aspect of this proposition is that the two domains D+ and D− must be connected. The implications of this statement are illustrated in Figure 3 in panel (2), showing the induced cycle flows when the dashed edge is damaged. The induced cycle flows are JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014 7 oriented clockwise above the domain boundary and counterclockwise below the domain boundary. If the perturbed edge lies on the boundary of a finite plane network, then there is only one domain and all cycle flows are oriented in the same way. With this result we can obtain a purely geometric view of how the flow of all edges in the network change after the outage. For this, we need some additional information about the magnitude of the cycle flows in addition to the orientation. We consider the upper and lower bound for the cycle flows ∆fc at a given distance to the cycle c1 with qc1 > 0 and the cycles c2 with qc2 < 0, respectively: C̃ ∈ RL×(L−N +1−2g) and 2g topological non-contractible cycles encoded in the cycle adjacency matrix Ĉ ∈ RL×2g , t t which satisfy Ĉ X d C̃ = 0 and C̃ X d Ĉ = 0. (42) A proof is given in the supplement [24]. The main result of this proposition is that the cycle basis of any graph can be decomposed into two parts. The geometric cycles behave just as the facets in a planar graph. But for non-planar graphs there is a second type of cycles – the topological ones. For the simplest non-planar examples one can find an embedding without line-crossings on the surface of a torus, which has the genus g = 1. Two topological cycles have to be added to the cycle basis, which wind around the torus in the two distinct directions. These cycles are intrinsically non-local. The following corollary now shows that also the effects of a line outage can be decomposed. Here, dist denotes the graph-theoretic distance between two cycles or faces, i.e. the length of the shortest path between the two faces in the dual graph. We then find the following result. Corollary 1. Consider a general graph with embedding and cycle basis as in Proposition 5. Then the flow changes after the outage of a line ℓ are given by Proposition 4. The maximum (minimum) value of the cycle flows decreases (increases) monotonically with the distance d to the reference cycles c1 and c2 , respectively: ∆F = C̃∆f˜ + Ĉ∆fˆ ud = ℓd = max ∆fc min ∆fc . c,dist(c,c1 )=d c,dist(c,c2 )=d ud ≤ ud−1 , ℓd ≥ ℓd−1 , 1 ≤ d ≤ dmax . 1 ≤ d ≤ dmax . where the cycle flows are given by t t Fℓ C̃ uℓ (C̃ X d C̃)∆f˜ = Mℓ,ℓ t t Fℓ Ĉ uℓ . (Ĉ X d Ĉ)∆fˆ = Mℓ,ℓ (43) A proof is given in the supplement [24]. Strict monotonicity can be proven when some additional technical assumptions are satisfied, which are expected to hold in most cases. For a two-dimensional lattices with regular topology and constant weights the cycle flows are proportional to the inverse distance (see supplement [24] for details). However, irregularity of the network topology and line parameters can lead to a stronger, even exponential, localization [28]–[30]. Hence, the response of the grid is strong only in the ‘vicinity’ of the damaged transmission line, but may be non-zero everywhere in the connected component. However, it has to be noted that the distance is defined for the dual graph, not the original graph, and that the rigorous results hold only for plane graphs. The situation is much more involved in non-planar graphs, as a line can link regions which would be far apart otherwise. Examples for the failure induced cycles flows and the decay with the distance are shown in Figure 2. C. General, non-planar networks Here, we consider fully general, non-planar networks. Unlike in the previous section, we show that it is impossible to derive a simple monotonic decay of the effect of line failures. Instead, by decomposing the LODFs into a geometric and a topological part, we show that complex, non-local interactions result. We start with Proposition 5. Every connected graph G can be embedded into a Riemannian surface of genus g ∈ N0 without line crossings. The cycle basis can be chosen such that it consists of the boundaries of L − N + 1 − 2g geometric facets of the embedding encoded in the cycle adjacency matrix (44) (45) (46) Proof: According to proposition 5
the cycle incidence matrix is decomposed as C = C̃, Ĉ . Similarly, we can decompose the strength of the cycle flows after the line outage as   ∆f˜ ∆f = (47) ∆fˆ such that the flow changes are given by ∆F = C̃∆f˜+ Ĉ∆fˆ. Then Eq. (39) reads ! !   t t
∆f˜ Fℓ C̃ C̃ = (48) t X d C̃, Ĉ t uℓ . Mℓ,ℓ Ĉ ∆fˆ Ĉ t t Using that Ĉ X d C̃ = 0 and C̃ X d Ĉ = 0 the corollary follows. Remarkably, the corollary shows that the cycle flows around geometric and topological cycles can be decoupled. The matrix t à = C̃ X d C̃ has a Laplacian structure as in Eq. (41) because at each edge of the graph at most two facets meet. Thus, Eq. (45) is a discrete Poisson equation as for plane graphs and the propositions 3 and 4 also hold for for the flows ∆f˜ around the geometric cycles. However, Eq. (46) has no such interpretation and it is, in general, dense on both sides. Thus, the topological cycles represented by Eq. (46) are responsible for complicated, non-local effects of damage spreading in general power grids. VII. C ONCLUSIONS Line Outage Distribution Factors are important for assessing the reliability of a power system, in particular with the recent rise of renewables. In this paper, we described a new dual JOURNAL OF LATEX CLASS FILES, VOL. 13, NO. 9, SEPTEMBER 2014 formalism for calculating LODFs, that is based on using power flows through the closed cycles of the network instead of using nodal voltage angles. The dual theory yields a compact formula for the LODFs that only depends on real power flows in the network. In particular, the formula lends itself to a straightforward generalization for the case of multiple line outages. Effectively, using cycle flows instead of voltage angles changes the dimensionality of the matrices appearing in the formulae from N × N to (L − N + 1) × (L − N + 1). In cases where the network is very sparse (i.e., it contains few cycles but many nodes), this can lead to a significant speedup in LODF computation time, a critical improvement for quick assessment of real network contingencies. In addition, the formalism generalises easily to multiple outages and arbitrary changes in series reactance, which is important for the assessment of the impact of FACTS devices. Often, some of the quantities involved in power flow problems are not known exactly, i.e., they are random (see, e.g., [32]). Thus, extending our work to include effects of randomness will be an important next step. The dual theory not only yields improvements for numerical computations, it also provides a novel viewpoint of the underlying physics of power grids, in particular if they are (almost) planar. Within the dual framework for planar networks, it is easy to show that single line contingencies induce flow changes in the power grid which decay monotonically in the same way as an electrostatic dipole field. ACKNOWLEDGMENTS We gratefully acknowledge support from the Helmholtz Association (joint initiative ‘Energy System 2050 – a contribution of the research field energy’ and grant no. VH-NG1025 to D.W.) and the German Federal Ministry of Education and Research (BMBF grant nos. 03SF0472B, 03SF0472C and 03SF0472E). The work of H. R. was supported in part by the IMPRS Physics of Biological and Complex Systems, Göttingen. R EFERENCES [1] S. M. Amin and B. F. Wollenberg, “Toward a smart grid: power delivery for the 21st century,” IEEE Power and Energy Magazine, vol. 3, no. 5, p. 34, 2005. [2] D. Heide, L. von Bremen, M. Greiner, C. Hoffmann, M. Speckmann, and S. Bofinger, “Seasonal optimal mix of wind and solar power in a future, highly renewable europe,” Renewable Energy, vol. 35, p. 2483, 2010. [3] M. Rohden, A. Sorge, M. Timme, and D. Witthaut, “Self-organized synchronization in decentralized power grids,” Phys. Rev. Lett., vol. 109, p. 064101, 2012. [4] T. Pesch, H.-J. Allelein, and J.-F. Hake, “Impacts of the transformation of the german energy system on the transmission grid,” Eur. Phys. J. Special Topics, vol. 223, p. 2561, 2014. [5] Brown, T., Schierhorn, P., Tröster, E., and Ackermann, T., “Optimising the European transmission system for 77% renewable electricity by 2030,” IET Renewable Power Generation, vol. 10, no. 1, pp. 3–9, 2016. [6] D. Witthaut, M. Rohden, X. Zhang, S. Hallerberg, and M. Timme, “Critical links and nonlocal rerouting in complex supply networks,” Phys. Rev. Lett., vol. 116, p. 138701, 2016. [7] A. J. Wood, B. F. Wollenberg, and G. B. Sheblé, Power Generation, Operation and Control. New York: John Wiley & Sons, 2014. [8] T. Güler, G. Gross, and M. Liu, “Generalized line outage distribution factors,” IEEE Transactions on Power Systems, vol. 22, no. 2, pp. 879– 881, 2007. 8 [9] J. Guo, Y. Fu, Z. Li, and M. Shahidehpour, “Direct calculation of line outage distribution factors,” IEEE Transactions on Power Systems, vol. 24, no. 3, pp. 1633–1634, 2009. [10] H. Ronellenfitsch, M. Timme, and D. Witthaut, “A dual method for computing power transfer distribution factors,” IEEE Transactions on Power Systems, vol. PP, no. 99, pp. 1–1, 2016. [11] R. Diestel, Graph Theory. New York: Springer, 2010. [12] B. Stott, J. Jardim, and O. Alsac, “Dc power flow revisited,” IEEE Trans. Power Syst., vol. 24, no. 3, p. 1290, 2009. [13] K. Purchala, L. Meeus, D. V. Dommelen, and R. Belmans, “Usefulness of dc power flow for active power flow analysis,” in IEEE Power Engineering Society General Meeting, June 2005, pp. 454–459 Vol. 1. [14] D. Van Hertem, J. Verboomen, K. Purchala, R. Belmans, and W. Kling, “Usefulness of dc power flow for active power flow analysis with flow controlling devices,” in The 8th IEEE International Conference on AC and DC Power Transmission, 2006, pp. 58–62. [15] M. E. J. Newman, Networks – An Introduction. Oxford: Oxford University Press, 2010. [16] J. J. Grainger and W. D. Stevenson Jr., Power System Analysis. New York: McGraw-Hill, 1994. [17] C. D. Modes, M. O. Magnasco, and E. Katifori, “Extracting hidden hierarchies in 3d distribution networks,” Phys. Rev. X, vol. 6, p. 031009, Jul 2016. [Online]. Available: http://link.aps.org/doi/10.1103/PhysRevX.6.031009 [18] M. A. Woodbury, “Inverting modified matrices,” Memorandum report, vol. 42, p. 106, 1950. [19] X.-P. Zhang, C. Rehtanz, and B. Pal, Flexible AC Transmission Systems: Modelling and Control. Berlin: Springer, 206. [20] “Smart Wires,” http://www.smartwires.com/, accessed: 2016-11-30. [21] R. D. Zimmerman, C. E. Murillo-Sanchez, and R. J. Thomas, “Matpower: Steady-state operations, planning and analysis tools for power systems research and education,” IEEE Trans. Power Syst., vol. 26, p. 12, 2011. [22] T. Coletta, R. Delabays, I. Adagideli, and P. Jacquod, “Vortex flows in high voltage ac power grids,” preprint arxiv:1605.07925, 2016. [23] D. Manik, M. Timme, and D. Witthaut, “Cycle flows and multistabilty in oscillatory networks: an overview,” preprint arxiv:1611.09825, 2016. [24] Supplementary Information, available online at IEEE. [25] S. Fliscounakis, P. Panciatici, F. Capitanescu, and L. Wehenkel, “Contingency ranking with respect to overloads in very large power systems taking into account uncertainty, preventive and corrective actions,” IEEE Trans. Power Syst., vol. 28, p. 4909, 2013. [26] W. A. Bukhsh and K. McKinnon, “Network data of real transmission networks,” http://www.maths.ed.ac.uk/optenergy/NetworkData/, 2013, [Online; accessed 03-August-2015]. [27] L. J. Grady and J. R. Polimeni, Discrete Calculus - Applied Analysis on Graphs for Computational Science. Springer, 2010. [28] D. Labavić, R. Suciu, H. Meyer-Ortmanns, and S. Kettemann, “Long-range response to transmission line disturbances in DC electricity grids,” The European Physical Journal Special Topics, vol. 223, no. 12, pp. 2517–2525, oct 2014. [Online]. Available: http://link.springer.com/10.1140/epjst/e2014-02273-0 [29] S. Kettemann, “Delocalization of phase perturbations and the stability of ac electricity grids,” preprint arXiv:1504.05525, 2015. [30] D. Jung and S. Kettemann, “Long-range Response in AC Electricity Grids,” arXiv preprint, no. 3, pp. 1–8, 2015. [Online]. Available: http://arxiv.org/abs/1512.05391 [31] P. Hines, I. Dobson, and P. Rezaei, “Cascading power outages propagate locally in an influence graph that is not the actual grid topology,” IEEE Transactions on Power Systems, vol. PP, no. 99, pp. 1–1, 2016. [32] T. Mühlpfordt, T. Faulwasser, and V. Hagenmeyer, “Solving stochastic ac power flow via polynomial chaos expansion,” in 2016 IEEE Conference on Control Applications (CCA), Sept 2016, pp. 70–76. 

1 The Arbitrarily Varying Broadcast Channel with Degraded Message Sets with Causal Side Information at the Encoder Uzi Pereg and Yossef Steinberg arXiv:1709.04770v2 [cs.IT] 15 Sep 2017 Abstract In this work, we study the arbitrarily varying broadcast channel (AVBC), when state information is available at the transmitter in a causal manner. We establish inner and outer bounds on both the random code capacity region and the deterministic code capacity region with degraded message sets. The capacity region is then determined for a class of channels satisfying a condition on the mutual informations between the strategy variables and the channel outputs. As an example, we consider the arbitrarily varying binary symmetric broadcast channel with correlated noises. We show cases where the condition holds, hence the capacity region is determined, and other cases where there is a gap between the bounds. Index Terms Arbitrarily varying channel, broadcast channel, degraded message sets, causal state information, Shannon strategies, side information, minimax theorem, deterministic code, random code, symmetrizability. The arbitrarily varying channel (AVC) was first introduced by Blackwell et al. [5] to describe a communication channel with unknown statistics, that may change over time. It is often described as communication in the presence of an adversary, or a jammer, attempting to disrupt communication. The arbitrarily varying broadcast channel (AVBC) without side information (SI) was first considered by Jahn [13], who derived an inner bound on the random code capacity region, namely the capacity region achieved by encoder and decoders with a random experiment, shared between the three parties. As indicated by Jahn, the arbitrarily varying broadcast channel inherits some of the properties of its single user counterpart. In particular, the random code capacity region is not necessarily achievable using deterministic codes [5]. Furthermore, Jahn showed that the deterministic code capacity region either coincides with the random code capacity region or else, it has an empty interior [13]. This phenomenon is an analogue of Ahlswede’s dichotomy property [2]. Then, in order to apply Jahn’s inner bound, one has to verify whether the capacity region has nonempty interior or not. As observed in [12], this can be resolved using the results of Ericson [10] and Csiszár and Narayan [8]. Specifically, a necessary and sufficient condition for the capacity region to have a non-empty interior is that both user marginal channels are non-symmetrizable. Various models of interest involve SI available at the encoder. In [19], the arbitrarily varying degraded broadcast channel with non-causal SI is addressed, using Ahlswede’s Robustification and Elimination Techniques [1]. The single user AVC with causal SI is addressed in the book by Csiszár and Körner [7], while their approach is independent of Ahlswede’s work. A straightforward application of Ahlswede’s Robustification Technique (RT) would violate the causality requirement. In this work, we study the AVBC with causal SI available at the encoder. We extend Ahlswede’s Robustification and Elimination Techniques [2, 1], originally used in the setting of non-causal SI. In particular, we derive a modified version of Ahlswede’s RT, suited to the setting of causal SI. In a recent paper by the authors [15], a similar proof technique is applied to the arbitrarily varying degraded broadcast channel with causal SI. Here, we generalize those results, and consider a general broadcast channel with degraded message sets with causal SI. We establish inner and outer bounds on the random code and deterministic code capacity regions. Furthermore, we give conditions on the AVBC under which the bounds coincide, and the capacity region is determined. As an example, we consider the arbitrarily varying binary symmetric broadcast channel with correlated noises. We show that in some cases, the conditions hold and the capacity region is determined. Whereas, in other cases, there is a gap between the bounds. I. D EFINITIONS AND P REVIOUS R ESULTS A. Notation We use the following notation conventions throughout. Calligraphic letters X , S, Y, ... are used for finite sets. Lowercase letters x, s, y, . . . stand for constants and values of random variables, and uppercase letters X, S, Y, . . . stand for random variables. The distribution of a random variable X is specified by a probability mass function (pmf) PX (x) = p(x) over a finite set X . The set of all pmfs over X is denoted by P(X ). We use xj = (x1 , x2 , . . . , xj ) to denote a sequence of letters from X . A random sequence X n and its distribution PX n (xn ) = p(xn ) are defined accordingly. For a pair of integers i and j, 1 ≤ i ≤ j, we define the discrete interval [i : j] = {i, i + 1, . . . , j}. This research was supported by the Israel Science Foundation (grant No. 1285/16). 2 B. Channel Description A state-dependent discrete memoryless broadcast channel (X × S, WY1 ,Y2 |X,S , Y1 , Y2 ) consists of a finite input alphabet X , two finite output alphabets Y1 and Y2 , a finite state alphabet S, and a collection of conditional pmfs WY1 ,Y2 |X,S . The Qn channel is memoryless without feedback, and therefore WY1n ,Y2n |X n ,S n (y1n , y2n |xn , sn ) = i=1 WY1 ,Y2 |X,S (y1,i , y2,i |xi , si ). The marginals WY1 |X,S and WY2 |X,S correspond to user 1 and user 2, respectively. Throughout, unless mentioned otherwise, it is assumed that the users have degraded message sets. That is, the encoder sends a private message which is intended for user 1, and a public message which is intended for both users. For state-dependent broadcast channels with causal SI, the channel input at time i ∈ [1 : n] may depend on the sequence of past and present states si . The arbitrarily varying broadcast channel (AVBC) is a discrete memoryless broadcast channel WY1 ,Y2 |X,S with a state sequence of unknown distribution, not necessarily independent nor stationary. That is, S n ∼ q(sn ) with an unknown joint pmf q(sn ) over S n . In particular, q(sn ) can give mass 1 to some state sequence sn . We denote the AVBC with causal SI by B = {WY1 ,Y2 |X,S }. To analyze the AVBC with degraded message sets with causal SI, we consider the compound broadcast channel. Different models of compound broadcast channels have been considered in the literature, as e.g. in [18] and [3]. Here, we define the compound broadcast channel as a discrete memoryless broadcast channel with a discrete memoryless state, where the state distribution q(s) is not known in exact, but rather belongs to a family of distributions Q, with Q ⊆ P(S). That is, Qn S n ∼ i=1 q(si ), with an unknown pmf q ∈ Q over S. We denote the compound broadcast channel with causal SI by B Q . The random parameter broadcast channel is a special case of a compound broadcast channel where the set Q consists of a single distribution, i.e. when the state sequence is memoryless and distributed according to a given state distribution q(s). Hence, we denote the random parameter broadcast channel with causal SI by B q . In Figure 1, we set the basic notation for the broadcast channel families that we consider. The columns correspond to the channel families presented above, namely the random parameter broadcast channel, the compound broadcast channel and the AVBC. The rows indicate the role of SI, namely the case of no SI and causal SI. In the first row, and throughout, we use the subscript ‘0’ to indicate the case where SI is not available. C. Coding with Degraded Message Sets We introduce some preliminary definitions, starting with the definitions of a deterministic code and a random code for the AVBC B with degraded message sets with causal SI. Note that in general, the term ‘code’, unless mentioned otherwise, refers to a deterministic code. Definition 1 (A code, an achievable rate pair and capacity region). A (2nR0 , 2nR1 , n) code for the AVBC B with degraded message sets with causal SI consists of the following; two message sets [1 : 2nR0 ] and [1 : 2nR1 ], where it is assumed throughout that 2nR0 and 2nR1 are integers, a sequence of n encoding functions fi : [1 : 2nR0 ] × [1 : 2nR1 ] × S i → X , i ∈ [1 : n], and two decoding functions, g1 : Y1n → [1 : 2nR0 ] × [1 : 2nR1 ] and g2 : Y2n → [1 : 2nR0 ]. At time i ∈ [1 : n], given a pair of messages (m0 , m1 ) ∈[1 : 2nR0 ] × [1 : 2nR1 ] and a sequence si , the encoder transmits xi = fi (m0 , m1 , si ). The codeword is then given by
xn = f n (m0 , m1 , sn ) , f1 (m0 , m1 , s1 ), f2 (m0 , m1 , s2 ), . . . , fn (m0 , m1 , sn ) . (1) Decoder 1 receives the channel output y1n , and finds an estimate for the message pair (m̂0 , m̂1 ) = g1 (y1n ). Decoder 2 only estimates the common message with m e 0 = g2 (y2n ). We denote the code by C = (f n (·, ·, ·), g1 (·), g2 (·)). Define the conditional probability of error of C given a state sequence sn ∈ S n by (n) Pe|sn (C ) = 1 2n(R0 +R1 ) nR0 nR1 2X 2X X WY1n ,Y2n |X n ,S n (y1n , y2n |f n (m0 , m1 , sn ), sn ) , (2) m0 =1 m1 =1 D(m0 ,m1 )c where D(m0 , m1 ) ,  PP Channel PP PP SI P P Fig. 1. (y1n , y2n ) ∈ Y1n × Y2n : g1 (y1n ) = (m0 , m1 ) , g2 (y2n ) = m0 . Random Parameter Compound AVBC without SI – B0Q B0 causal SI Bq BQ B Notation of broadcast channel families. The columns correspond to the channel family, and the rows indicate the role of SI at the encoder. (3) 3 Now, define the average probability of error of C for some distribution q(sn ) ∈ P(S n ), X (n) Pe(n) (q, C ) = q(sn ) · Pe|sn (C ) . (4) sn ∈S n We say that C is a (2nR0 , 2nR1 , n, ε) code for the AVBC B if it further satisfies Pe(n) (q, C ) ≤ ε , for all q(sn ) ∈ P(S n ) . (5) We say that a rate pair (R0 , R1 ) is achievable if for every ε > 0 and sufficiently large n, there exists a (2nR0 , 2nR1 , n, ε) code. The operational capacity region is defined as the closure of the set of achievable rate pairs and it is denoted by C(B). We use the term ‘capacity region’ referring to this operational meaning, and in some places we call it the deterministic code capacity region in order to emphasize that achievability is measured with respect to deterministic codes. We proceed now to define the parallel quantities when using stochastic-encoder stochastic-decoders triplets with common randomness. The codes formed by these triplets are referred to as random codes. Definition 2 (Random code). A (2nR0 , 2nR1 , n) random code for the AVBC B consists of a collection of (2nR0 , 2nR1 , n) codes {Cγ = (fγn , g1,γ , g2,γ )}γ∈Γ , along with a probability distribution µ(γ) over the code collection Γ. We denote such a code by C Γ = (µ, Γ, {Cγ }γ∈Γ ). Analogously to the deterministic case, a (2nR0 , 2nR1 , n, ε) random code has the additional requirement X Pe(n) (q, C Γ ) = µ(γ)Pe(n) (q, Cγ ) ≤ ε , for all q(sn ) ∈ P(S n ) . (6) γ∈Γ The capacity region achieved by random codes is denoted by C⋆(B), and it is referred to as the random code capacity region. Next, we write the definition of superposition coding [4] using Shannon strategies [16]. See also [17], and the discussion after Theorem 4 therein. Here, we refer to such codes as Shannon strategy codes. Definition 3 (Shannon strategy codes). A (2nR0 , 2nR1 , n) Shannon strategy code for the AVBC B with degraded message sets with causal SI is a (2nR0 , 2nR1 , n) code with an encoder that is composed of two strategy sequences un0 :[1 : 2nR0 ] → U0n , un1 :[1 : 2 nR0 ] × [1 : 2 (7) nR1 ]→ U1n , (8) and an encoding function ξ(u0 , u1 , s), where ξ : U0 × U1 × S → X , as well as a pair of decoding functions g1 : Y1n → [1 : 2nR0 ] × [1 : 2nR1 ] and g2 : Y2n → [1 : 2nR0 ]. The codeword is then given by  n xn = ξ n (un0 (m0 ), un1 (m0 , m1 ), sn ) , ξ(un0,i (m0 ), un1,i (m0 , m1 ), si ) i=1 . (9) We denote the code by C = (un0 , un1 , ξ, g1 , g2 ). D. In the Absence of Side Information – Inner Bound In this subsection, we briefly review known results for the case where the state is not known to the encoder or the decoder, i.e. SI is not available. Consider a given AVBC with degraded message sets without SI, which we denote by B0 . Let   R0 ≤ Iq (U ; Y2 ) ,  [ \  (R0 , R1 ) : R1 ≤ Iq (X; Y1 |U ) , (10) R⋆0,in ,   R0 + R1 ≤ Iq (X; Y1 ) p(x,u) q(s) In [13, Theorem 2], Jahn introduced an inner bound for the arbitrarily varying general broadcast channel. In our case, with degraded message sets, Jahn’s inner bound reduces to the following. Theorem 1 (Jahn’s Inner Bound [13]). Let B0 be an AVBC with degraded message sets without SI. Then, R⋆0,in is an achievable rate region using random codes over B0 , i.e. C⋆(B0 ) ⊇ R⋆0,in . (11) Now we move to the deterministic code capacity region. Theorem 2 (Ahlswede’s Dichotomy [13]). The capacity region of an AVBC B0 with degraded message sets without SI
either coincides with the random code capacity region or else, its interior is empty. That is, C(B0 ) = C⋆(B0 ) or else, int C(B0 ) = ∅. By Theorem 1 and Theorem 2, we have that
R⋆0,in is an achievable rate region, if the interior of the capacity region is non-empty. That is, C(B0 ) ⊇ R⋆0,in , if int C(B0 ) 6= ∅.
Theorem 3 (see [10, 8, 12]). For an AVBC B0 without SI, the interior of the capacity region is non-empty, i.e. int C(B0 ) 6= ∅, if and only if the marginals WY1 |X,S and WY2 |X,S are not symmetrizable. 4 II. M AIN R ESULTS We present our results on the compound broadcast channel and the AVBC with degraded message sets with causal SI. A. The Compound Broadcast Channel with Causal SI We now consider the case where the encoder has access to the state sequence in a causal manner, i.e. the encoder has S i . 1) Inner Bound: First, we provide an achievable rate region for the compound broadcast channel with degraded message sets with causal SI. Consider a given compound broadcast channel B Q with causal SI. Let   R0 ≤ Iq (U0 ; Y2 ) ,  [ \  (R0 , R1 ) : R1 ≤ Iq (U1 ; Y1 |U0 ) , (12) Rin (B Q ) ,   R0 + R1 ≤ Iq (U0 , U1 ; Y1 ) p(u0 ,u1 ), ξ(u0 ,u1 ,s) q(s)∈Q subject to X = ξ(U0 , U1 , S), where U0 and U1 are auxiliary random variables, independent of S, and the union is over the pmf p(u0 , u1 ) and the set of all functions ξ : U0 × U1 × S → X . This can also be expressed as   R0 ≤ inf q∈Q Iq (U0 ; Y2 ) ,  (R0 , R1 ) :  [ R1 ≤ inf q∈Q Iq (U1 ; Y1 |U0 ) , . (13) Rin (B Q ) =   R0 + R1 ≤ inf q∈Q Iq (U0 , U1 ; Y1 ) p(u0 ,u1 ), ξ(u0 ,u1 ,s) Lemma 4. Let B Q be a compound broadcast channel with degraded message sets with causal SI available at the encoder. Then, Rin (B Q ) is an achievable rate region for B Q , i.e. C(B Q ) ⊇ Rin (B Q ) . (14) Specifically, if (R0 , R1 ) ∈ Rin (B Q ), then for some a > 0 and sufficiently large n, there exists a (2nR0 , 2nR1 , n, e−an ) Shannon strategy code over the compound broadcast channel B Q with degraded message sets with causal SI. The proof of Lemma 4 is given in Appendix A. 2) The Capacity Region: We determine the capacity region of the compound broadcast channel B Q with degraded message sets with causal SI available at the encoder. In addition, we give a condition, for which the inner bound in Lemma 4 coincides with the capacity region. Let   R0 ≤ Iq (U0 ; Y2 ) ,   (R0 , R1 ) : \ [ R1 ≤ Iq (U1 ; Y1 |U0 ) . (15) Rout (B Q ) ,   R0 + R1 ≤ Iq (U0 , U1 ; Y1 ) q(s)∈Q p(u0 ,u1 ), ξ(u0 ,u1 ,s) Now, our condition is defined in terms of the following. Definition 4. We say that a function ξ : U0 × U1 × S → X and a set D ⊆ P(U0 × U1 ) achieve both Rin (B Q ) and Rout (B Q ) if   R0 ≤ Iq (U0 ; Y2 ) ,  [ \  (R0 , R1 ) : R1 ≤ Iq (U1 ; Y1 |U0 ) , , (16a) Rin (B Q ) =   R0 + R1 ≤ Iq (U0 , U1 ; Y1 ) p(u0 ,u1 )∈D q(s)∈Q and Rout (B Q ) = \ [ q(s)∈Q p(u0 ,u1 )∈D   (R0 , R1 ) :   R0 ≤ Iq (U0 ; Y2 ) ,  R1 ≤ Iq (U1 ; Y1 |U0 ) , ,  R0 + R1 ≤ Iq (U0 , U1 ; Y1 ) (16b) subject to X = ξ(U0 , U1 , S). That is, the unions in (12) and (15) can be restricted to the particular function ξ(u0 , u1 , s) and set of strategy distributions D. Observe that by Definition 4, given a function ξ(u0 , u1 , s), if a set D achieves both Rin (B Q ) and Rout (B Q ), then every set D′ with D ⊆ D′ ⊆ P(U0 × U1 ) achieves those regions, and in particular, D′ = P(U0 × U1 ). Nevertheless, the condition defined below requires a certain property that may hold for D, but not for D′ . Definition 5. Given a convex set Q of state distributions, define Condition T Q by the following; for some ξ(u0 , u1 , s) and D that achieve both Rin (B Q ) and Rout (B Q ), there exists q ∗ ∈ Q which minimizes the mutual informations Iq (U0 ; Y2 ), Iq (U1 ; Y1 |U0 ), and Iq (U0 , U1 ; Y1 ), for all p(u0 , u1 ) ∈ D, i.e. T Q : For some q ∗ ∈ Q, q ∗ = arg min Iq (U0 ; Y2 ) = arg min Iq (U1 ; Y1 |U0 ) = arg min Iq (U0 , U1 ; Y1 ) , q∈Q ∀p(u0 , u1 ) ∈ D . q∈Q q∈Q (17) 5 Intuitively, when Condition T Q holds, there exists a single jamming strategy q ∗ (s) which is worst for both users simultaneously. That is, there is no tradeoff for the jammer. As the optimal jamming strategy is unique, this eliminates ambiguity for the users as well. Theorem 5. Let B Q be a compound broadcast channel with causal SI available at the encoder. Then, 1) the capacity region of B Q follows
C(B Q ) = Rout (B Q ) , if int C(B Q ) 6= ∅ , (18)
and it is identical to the corresponding random code capacity region, i.e. C⋆(B Q ) = C(B Q ) if int C(B Q ) 6= ∅. 2) Suppose that Q ⊆ P(S) is a convex set of state distributions. If Condition T Q holds, the capacity region of B Q is given by C(B Q ) = Rin (B Q ) = Rout (B Q ) , (19) and it is identical to the corresponding random code capacity region, i.e. C⋆(B Q ) = C(B Q ).
The proof of Theorem 5 is given in Appendix B. Regarding part 1,
we note that when int C(B Q ) = ∅, then the inner bound Rin (B Q ) has an empty interior as well (see (13)). Thus, int Rin (B Q ) 6= ∅ is also a sufficient condition for C(B Q ) = Rout (B Q ). 3) The Random Parameter Broadcast Channel with Causal SI: Consider the random parameter broadcast channel with causal SI. Recall that this is simply a special case of a compound broadcast channel, where the set of state distributions consists of a single member, i.e. Q = {q(s)}. Then, let   R0 ≤ Iq (U0 ; Y2 ) ,   (R0 , R1 ) : [ R1 ≤ Iq (U1 ; Y1 |U0 ) , , (20) C(B q ) ,   R0 + R1 ≤ Iq (U0 , U1 ; Y1 ) p(u0 ,u1 ),ξ(u0 ,u1 ,s) with |U0 | ≤|X ||S| + 2 , |U1 | ≤ |X ||S|(|X ||S| + 2) . (21) q Theorem 6. The capacity region of the random parameter broacast channel B with degraded message sets with causal SI is given by C(B q ) = C(B q ) . (22) Theorem 6 is proved in Appendix C. B. The AVBC with Causal SI We give inner and outer bounds, on the random code capacity region and the deterministic code capacity region, for the AVBC B with degraded message sets with causal SI. We also provide conditions, for which the inner bound coincides with the outer bound. 1) Random Code Inner and Outer Bounds: Define R⋆in (B) , Rin (B Q ) , R⋆out (B) , Rout (B Q ) Q=P(S) , (23) Q=P(S) and T =TQ Q=P(S) . (24) Theorem 7. Let B be an AVBC with degraded message sets with causal SI available at the encoder. Then, 1) the random code capacity region of B is bounded by R⋆in (B) ⊆ C⋆(B) ⊆ R⋆out (B) . (25) 2) If Condition T holds, the random code capacity region of B is given by C⋆(B) = R⋆in (B) = R⋆out (B) . (26) The proof of Theorem 7 is given in Appendix D. Before we proceed to the deterministic code capacity region, we need one further result. The following lemma is a restatement of a result from [2], stating that a polynomial size of the code collection {Cγ } is sufficient. This result is a key observation in Ahlswede’s Elimination Technique (ET), presented in [2], and it is significant for the deterministic code analysis. Lemma 8. Consider a given (2nR0 , 2nR1 , n, εn ) random code C Γ = (µ, Γ, {Cγ }γ∈Γ ) for the AVBC B, where limn→∞ εn = 0. Then, for every 0 < α < 1 and sufficiently large n, there exists a (2nR0 , 2nR1 , n, α) random code (µ∗ , Γ∗ , {Cγ }γ∈Γ∗ ) with the following properties: 6 1) The size of the code collection is bounded by |Γ∗ | ≤ n2 . 2) The code collection is a subset of the original code collection, i.e. Γ∗ ⊆ Γ. 3) The distribution µ∗ is uniform, i.e. µ∗ (γ) = |Γ1∗ | , for γ ∈ Γ∗ . The proof of Lemma 8 follows the same lines as in [2, Section 4] (see also [13, 19]). For completeness, we give the proof in Appendix E. 2) Deterministic Code Inner and Outer Bounds: The next theorem characterizes the deterministic code capacity region, which demonstrates a dichotomy property. Theorem 9. The capacity region of an AVBC B with degraded message sets with causal SI either
coincides with the random code capacity region or else, it has an empty interior. That is, C(B) = C⋆(B) or else, int C(B) = ∅. The proof of Theorem 9 is given in Appendix F. Let U = (U0 , U1 ), hence U = U0 × U1 . For every pair of functions ′ ξ : U × S → X and ξ ′ : U0 × S → X , define the DMCs VYξ1 |U,S and VYξ2 |U0 ,S specified by VYξ1 |U,S (y1 |u, s) = WY1 |X,S (y1 |ξ(u, s), s) , ′ VYξ2 |U0 ,S (y2 |u0 , s) = WY2 |X,S (y2 |ξ ′ (u0 , s), s) , (27a) (27b) respectively. Corollary 10. The capacity region of B is bounded by
C(B) ⊇ R⋆in (B) , if int C(B) 6= ∅ , C(B) ⊆ R⋆out (B) . (28) (29) ′ Furthermore, if VYξ1 |U,S and VYξ2 |U0 ,S are non-symmetrizable for some ξ : U × S → X and ξ ′ : U0 × S → X , and Condition T holds, then C(B) = R⋆in (B) = R⋆out (B). The proof of Corollary 10 is given in Appendix G. III. D EGRADED B ROADCAST C HANNEL WITH C AUSAL SI In this section, we consider the special case of an arbitrarily varying degraded broadcast channel (AVDBC) with causal SI, when user 1 and user 2 have private messages. A. Definitions We consider a degraded broadcast channel (DBC), which is a special case of the general broadcast channel described in the previous sections. Following the definitions by [17], a state-dependent broadcast channel WY1 ,Y2 |X,S is said to be physically degraded if it can be expressed as WY1 ,Y2 |X,S (y1 , y2 |x, s) = WY1 |X,S (y1 |x, s) · p(y2 |y1 ) , (30) i.e. (X, S) Y1 Y2 form a Markov chain. User 1 is then referred to as the stronger user, whereas user 2 is referred to P as the weaker user. More generally, a broadcast channel is said to be stochastically degraded if WY2 |X,S (y2 |x, s) = e(y2 |y1 ) for some conditional distribution pe(y2 |y1 ). We note that the definition of degradedness y1 ∈Y1 WY1 |X,S (y1 | x, s)· p here is stricter than the definition in [13, Remark IIB5]. Our results apply to both the physically degraded and the stochastically degraded broadcast channels. Thus, for our purposes, there is no need to distinguish between the two, and we simply say that the broadcast channel is degraded. We use the notation BD for an AVDBC with causal SI. We consider the case where the users have private messages. A deterministic code and a random code for the AVDBC BD with causal SI are then defined as follows. Definition 6 (A private-message code, an achievable rate pair and capacity region). A (2nR1 , 2nR2 , n) private-message code for the AVDBC BD with causal SI consists of the following; two message sets [1 : 2nR1 ] and [1 : 2nR2 ], where it is assumed throughout that 2nR1 and 2nR2 are integers, a set of n encoding functions fi : [1 : 2nR1 ] × [1 : 2nR2 ] × S i → X , i ∈ [1 : n], and two decoding functions, g1 : Y1n → [1 : 2nR1 ] and g2 : Y2n → [1 : 2nR2 ]. At time i ∈ [1 : n], given a pair of messages m1 ∈ [1 : 2nR1 ] and m2 ∈ [1 : 2nR2 ] and a sequence si , the encoder transmits xi = fi (m1 , m2 , si ). The codeword is then given by
xn = f n (m1 , m2 , sn ) , f1 (m1 , m2 , s1 ), f2 (m1 , m2 , s2 ), . . . , fn (m1 , m2 , sn ) . (31) Decoder k receives the channel output ykn , for k = 1, 2., and finds an estimate for the k th message, m̂k = gk (ykn ). Denote the code by C = (f n (·, ·, ·), g1 (·), g2 (·)). 7 Define the conditional probability of error of C given a state sequence sn ∈ S n by (n) Pe|sn (C ) = 1 2n(R1 +R2 ) nR1 nR2 2X 2X X m1 =1 m2 =1 D(m1 ,m2 WY1n ,Y2n |X n ,S n (y1n , y2n |f n (m1 , m2 , sn ), sn ) , (32) )c where D(m1 , m2 ) , We say that C is a (2 nR1 ,2 nR2  (y1n , y2n ) ∈ Y1n × Y2n : g1 (y1n ) = m1 , g2 (y2n ) = m2 . (33) , n, ε) code for the AVDBC B if it further satisfies X (n) Pe(n) (q, C ) = q(sn ) · Pe|sn (C ) ≤ ε , for all q(sn ) ∈ P(S n ) . (34) sn ∈S n An achievable private-message rate pair (R1 , R2 ) and the capacity region C(BD ) are defined as usual. We proceed now to define the parallel quantities when using stochastic-encoder stochastic-decoders triplets with common randomness. Definition 7 (Random code). A (2nR1 , 2nR2 , n) private-message random code for the AVDBC BD consists of a collection of (2nR1 , 2nR2 , n) codes {Cγ = (fγn , g1,γ , g2,γ )}γ∈Γ , along with a probability distribution µ(γ) over the code collection Γ. Analogously to the deterministic case, a (2nR1 , 2nR2 , n, ε) random code has the additional requirement X Pe(n) (q, C Γ ) = µ(γ)Pe(n) (q, Cγ ) ≤ ε , for all q(sn ) ∈ P(S n ) . (35) γ∈Γ The private-message capacity region achieved by random codes is denoted by C⋆(BD ), and it is referred to as the random code capacity region. By standard arguments, a private-message rate pair (R1 , R2 ) is achievable for the AVDBC BD if and only if (R0 , R1 ) is achievable with degraded message sets, with R0 = R2 . This immediately implies the following results. B. Results The results in this section are a straightforward consequence of the results in Section II. 1) Random Code Inner and Outer Bounds: Define [ \  (R1 , R2 ) : R2 ≤ Iq (U2 ; Y2 ) ,  ⋆ , Rin (BD ) , R1 ≤ Iq (U1 ; Y1 |U2 ) (36) p(u1 ,u2 ), ξ(u1 ,u2 ,s) q(s) and R⋆out (BD ) , \ [ q(s) p(u0 ,u1 ), ξ(u0 ,u1 ,s)  (R1 , R2 ) : R2 R1 ≤ Iq (U2 ; Y2 ) , ≤ Iq (U1 ; Y1 |U2 )  . (37) Now, we define a condition in terms of the following. Definition 8. We say that a function ξ : U1 × U2 × S → X and a set D⋆ ⊆ P(U1 × U2 ) achieve both R⋆in (BD ) and R⋆out (BD ) if \  (R1 , R2 ) : R2 ≤ Iq (U2 ; Y2 ) ,  [ ⋆ Rin (BD ) = , (38a) R1 ≤ Iq (U1 ; Y1 |U2 ) p(u0 ,u1 )∈D ⋆ q(s) and R⋆out (BD ) = \ [ q(s) p(u0 ,u1 )∈D ⋆  (R1 , R2 ) : R2 R1 ≤ Iq (U2 ; Y2 ) , ≤ Iq (U1 ; Y1 |U2 )  , (38b) subject to X = ξ(U1 , U2 , S). That is, the unions in (36) and (37) can be restricted to the particular function ξ(u1 , u2 , s) and set of strategy distributions D⋆ . Definition 9. Define Condition TD by the following; for some ξ(u1 , u2 , s) and D⋆ that achieve both R⋆in (BD ) and R⋆out (BD ), there exists q ∗ ∈ P(S) which minimizes both Iq (U2 ; Y2 ) and Iq (U1 ; Y1 |U2 ), for all p(u1 , u2 ) ∈ D⋆ , i.e. TD : For some q ∗ ∈ P(S), q ∗ = arg min Iq (U2 ; Y2 ) = arg min Iq (U1 ; Y1 |U2 ) ∀p(u1 , u2 ) ∈ D⋆ . q(s) q(s) Theorem 11. Let BD be an AVDBC with causal SI available at the encoder. Then, 8 1) the random code capacity region of BD is bounded by R⋆in (BD ) ⊆ C⋆(BD ) ⊆ R⋆out (BD ) . (39) 2) If Condition TD holds, the random code capacity region of BD is given by C⋆(BD ) = R⋆in (BD ) = R⋆out (BD ) . (40) Theorem 11 is a straightforward consequence of Theorem 7. 2) Deterministic Code Inner and Outer Bounds: The next theorem characterizes the deterministic code capacity region, which demonstrates a dichotomy property. Theorem 12. The capacity region of an AVDBC BD with causal SI either coincides with the random code capacity region or
else, it has an empty interior. That is, C(BD ) = C⋆(BD ) or else, int C(BD ) = ∅. 9. Now, Theorem 11 and Theorem 12 yield the following corollary. Theorem 12 is a straightforward consequence of Theorem ′ For every function ξ ′ : U2 × S → X , define a DMC VYξ2 |U2 ,S specified by ′ VYξ2 |U2 ,S (y2 |u2 , s) =WY2 |X,S (y2 |ξ ′ (u2 , s), s) . (41) Corollary 13. The capacity region of BD is bounded by
C(BD ) ⊇ R⋆in (BD ) , if int C(BD ) 6= ∅ , C(BD ) ⊆ R⋆out (BD ) . (42) (43) ′ Furthermore, if VYξ2 |U2 ,S is non-symmetrizable for some ξ ′ : U2 ×S → X , and Condition TD holds, then C(BD ) = R⋆in (BD ) = R⋆out (BD ). No SI q=0 q=0.25 q=0.5 q=0.75 q=1 0.3 R 0.2 2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 R1 Fig. 2. The private-message capacity region of the AVDBC in Example 1, the arbitrarily varying binary symmetric broadcast channel. The area under the thick blue line is the capacity region of the AVDBC BD with causal SI, with θ0 = 0.005, θ1 = 0.9, and α = 0.2. The black square at the origin stands q ) for q = 0, 0.25, 0.5, 0.75, 1, where the capacity for the capacity region of the AVDBC BD,0 without SI, C(BD,0 ) = {(0, 0)}. The curves depict C(BD q ⋆ region of BD is given by C(BD ) = Rout (BD ) = C(BD ) for q = 1 (see (37)). IV. E XAMPLES To illustrate the results above, we give the following examples. In the first example, we consider an AVDBC and determine the private-message capacity region. Then, in the second example, we consider a non-degraded AVBC and determine the capacity region with degraded message sets. Example 1. Consider an arbitrarily varying binary symmetric broadcast channel (BSBC), Y1 =X + ZS mod 2 , Y2 =Y1 + K mod 2 , 9 where X, Y1 , Y2 , S, ZS , K are binary, with values in {0, 1}. The additive noises are distributed according to Zs ∼Bernoulli(θs ) , for s ∈ {0, 1} , K ∼Bernoulli(α) , with θ0 ≤ 1 − θ1 ≤ 21 and α < 12 , where K is independent of (S, ZS ). It is readily seen the channel is physically degraded. Then, consider the case where user 1 and user 2 have private messages. We have the following results. Define the binary entropy function h(x) = −x log x − (1 − x) log(1 − x), for x ∈ [0, 1], with logarithm to base 2. The private-message capacity region of the arbitrarily varying BSBC BD,0 without SI is given by C(BD,0 ) = {(0, 0)} . (44) The private-message capacity region of the arbitrarily varying BSBC BD with causal SI is given by [  (R1 , R2 ) : R2 ≤ 1 − h(α ∗ β ∗ θ1 ) ,  C(BD ) = . R1 ≤ h(β ∗ θ1 ) − h(θ1 ) (45) 0≤β≤1 It will be seen in the achievability proof that the parameter β is related to the distribution of U1 , and thus the RHS of (45) can be thought of as a union over Shannon strategies. The analysis is given in Appendix H. It is shown in Appendix H that Condition TD holds and C(BD ) = R⋆in (BD ) = R⋆out (BD ). Figure 2 provides a graphical interpretation. Consider a DBC WY1 ,Y2 |X,S with random parameters with causal SI, governed by an i.i.d. state sequence, q distributed according to S ∼ Bernoulli(q), for a given 0 ≤ q ≤ 1, and let C(BD ) denote the corresponding capacity region. q∗ Then, the analysis shows that Condition TD implies that there exists 0 ≤ q ∗ ≤ 1 such that C(BD ) = C(BD ), where ∗ q q q C(BD ) ⊆ C(BD ) for every 0 ≤ q ≤ 1. Indeed, looking at Figure 2, it appears that the regions C(BD ), for 0 ≤ q ≤ 1, form a q∗ well ordered set, hence C(BD ) = C(BD ) with q ∗ = 1. Next, we consider an example of an AVBC which is not degraded in the sense defined above. Example 2. Consider a state-dependent binary symmetric broadcast channel (BSBC) with correlated noises, Y1 =X + ZS mod 2 , Y2 =X + NS mod 2 , where X, Y1 , Y2 , S, ZS , NS are binary, with values in {0, 1}. The additive noises are distributed according to Zs ∼Bernoulli(θs ) , Ns ∼ Bernoulli(εs ) , for s ∈ {0, 1} , where S, Z0 , Z1 , N0 , N1 are independent random variables, with θ0 ≤ ε0 ≤ 21 and 12 ≤ ε1 ≤ θ1 . Intuitively, this suggests that Y2 is a weaker channel. Nevertheless, observe that this channel is not degraded in the sense defined in Section III-A (see (30)). For a given state S = s, the broadcast channel WY1 ,Y2 |X,S (·, ·|·, s) is stochastically degraded. In particular, one can define the following random variables, As ∼ Bernoulli(πs ) , where πs , ε s − θs , 1 − 2θs Ye2 = Y1 + AS mod 2 .   Then, Ye2 is distributed according to Pr Ye2 = y2 |X = x, S = s = WY2 |X,S2 (y2 |x, s), and X (46) (47) (Y1 , S) Ye2 form a Markov chain. However, since X and AS depend on the state, it is not necessarily true that (X, S) Y1 Ye2 form a Markov chain, and the BSBC with correlated noises could be non-degraded. We have the following results. Random Parameter BSBC with Correlated Noises First, we consider the random parameter BSBC B q , with a memoryless state S ∼ Bernoulli(q), for a given 0 ≤ q ≤ 1. Define the binary entropy function h(x) = −x log x − (1 − x) log(1 − x), for x ∈ [0, 1], with logarithm to base 2. We show that the capacity region of the random parameter BSBC B q with degraded message sets with causal SI is given by  [  (R0 , R1 ) : R0 ≤ 1 − h(β ∗ δ (2) ) , q C(B q ) = C(B q ) = , (48) R1 ≤ h(β ∗ δq(1) ) − h(δq(1) ) 0≤β≤1 where δq(1) =(1 − q)θ0 + q(1 − θ1 ) , δq(2) =(1 − q)ε0 + q(1 − ε1 ) . (49) 10 0.5 q=0 q=1/3 q=2/3 q=1 0.4 0.3 R1 0.2 0.1 0 0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 R0 Fig. 3. The capacity region of the AVBC in Example 2, the arbitrarily varying binary symmetric broadcast channel with correlated noises, with parameters that correspond to case 1. The area under the thick blue line is the capacity region of the AVBC B with causal SI, with θ0 = 0.12, θ1 = 0.85, ε0 = 0.18 ⋆ (B) = C(Bq ) for q = 1 (see and ε1 = 0.78. The curves depict C(Bq ) for q = 0, 1/3, 2/3, 1, where the capacity region of B is given by C(B) = R out (48)). The proof is given in Appendix I-A. It can be seen in the achievability proof that the parameter β is related to the distribution of U1 , and thus the RHS of (48) can be thought of as a union over Shannon strategies. Arbitrarily Varying BSBC with Correlated Noises We move to the arbitrarily varying BSBC with correlated noises. As shown in Appendix I-B, the capacity region of the arbitrarily varying BSBC B0 with degraded message sets without SI is given by C(B0 ) = {(0, 0)}. For the setting where causal SI is available at the encoder, we consider two cases. Case 1: Suppose that θ0 ≤ 1 − θ1 ≤ ε0 ≤ 1 − ε1 ≤ 12 . That is, S = 1 is a noisier channel state than S = 0, for both users. The capacity region of the arbitrarily varying BSBC B with degraded message sets with causal SI is given by  [  (R0 , R1 ) : R0 ≤ 1 − h(β ∗ ε1 ) , C(B) = C(B q ) = . (50) R1 ≤ h(β ∗ θ1 ) − h(θ1 ) q=1 0≤β≤1 It is shown in Appendix I-B that Condition T holds and C(B) = R⋆in (B) = R⋆out (B). Figure 3 provides a graphical ∗ interpretation. The analysis shows that Condition T implies that there exists 0 ≤ q ∗ ≤ 1 such that C(B) = C(B q ), where ∗ C(B q ) ⊆ C(B q ) for every 0 ≤ q ≤ 1. Indeed, looking at Figure 3, it appears that the regions C(B q ), for 0 ≤ q ≤ 1, form a ∗ well ordered set, hence C(B) = C(B q ) with q ∗ = 1. Case 2: Suppose that θ0 ≤ 1 − θ1 ≤ 1 − ε1 ≤ ε0 ≤ 21 . That is, S = 1 is a noisier channel state for user 1, whereas S = 0 is noisier for user 2. The capacity region of the arbitrarily varying BSBC B with degraded message sets with causal SI is bounded by  [  (R0 , R1 ) : R0 ≤ 1 − h(β ∗ ε0 ) , C(B) ⊇R⋆in (B) = , (51) R1 ≤ h(β ∗ θ1 ) − h(θ1 ) 0≤β≤1 and C(B) ⊆R⋆out (B) ⊆ C(B q=0 ) ∩ C(B q=1 )    (R0 , R1 ) : R0 ≤ 1 − h(β0 ∗ ε0 ) , [  R0 ≤ 1 − h(β1 ∗ ε1 ) , = R1 ≤ h(β0 ∗ θ0 ) − h(θ0 ) ,   0≤β0 ≤1 ,  R1 ≤ h(β1 ∗ θ1 ) − h(θ1 ) 0≤β1 ≤1        . (52) 11 Fig. 4. The inner and outer bounds on the capacity region of the AVBC in Example 2, the arbitrarily varying binary symmetric broadcast channel with correlated noises, with parameters that correspond to case 2, namely, θ0 = 0.12, θ1 = 0.85, ε0 = 0.22 and ε1 = 0.88. 0.5 q=0 q=1 0.4 0.3 R1 0.2 0.1 0 0 0.1 0.2 0.3 R 0.4 0.5 0 (a) The dashed and dotted lines depict the boundaries of C(B q=0 ) and C(B q=1 ), respectively. The colored lines depict C(B q ) for a range of values of 0 < q < 1. 0.4 Inner bound Outer bound 0.3239 0.3 R1 0.2 0.1 0 0 0.05 0.0784 0.1 0.15 R 0.2 0.25 0 (b) The area under the thick blue line is the inner bound R⋆in (B), and the area under the thin line is the outer bound R⋆out (B). The analysis is given in Appendix I. Figure 4 provides a graphical interpretation. The dashed and dotted lines in Figure 4(a) depict the boundaries of C(B q=0 ) and C(B q=1 ), respectively. The colored lines depict C(B q ) for a range of values of 0 < q < 1. It appears that R⋆out (B) = ∩0≤q≤1 C(B q ) reduces to the intersection of the regions C(B q=0 ) and C(B q=1 ). Figure 4(b) demonstrates the gap between the bounds in case 2. A PPENDIX A P ROOF OF L EMMA 4 We show that every rate pair (R0 , R1 ) ∈ Rin (B Q ) can be achieved using deterministic codes over the compound broadcast channel B Q with causal SI. We construct a code based on superposition coding with Shannon strategies, and decode using joint typicality with respect to a channel state type, which is “close” to some q ∈ Q. We use the following notation. Basic method of types concepts are defined as in [7, Chapter 2]; including the definition of a type P̂xn of a sequence xn ; a joint type P̂xn ,yn and a conditional type P̂xn |yn of a pair of sequences (xn , y n ); and a δ-typical 12 set Aδ (PX,Y ) with respect to a distribution PX,Y (x, y). Define a set Q̂n of state types n o Q̂n = P̂sn : sn ∈ Aδ1 (q) , for some q ∈ Q , (53) where δ , 2 · |S| δ1 , (54) where δ > 0 is arbitrarily small. That is, Q̂n is the set of types that are δ1 -close to some state distribution q(s) in Q. Note that for any fixed δ (or δ1 ), for a sufficiently large n, the set Q̂n covers the set Q, and it is in fact a δ1 -blowup of Q. Now, a code for the compound broadcast channel with causal SI is constructed as follows. Codebook Generation: Fix the distribution PU0 ,U1 and the function ξ(u0 , u1 , s). Generate 2nR0 independent sequences at random, n Y (55) PU0 (u0,i ) , for m0 ∈ [1 : 2nR0 ] . un0 (m0 ) ∼ i=1 For every m0 ∈ [1 : 2 nR0 ], generate 2 nR1 sequences at random, un1 (m0 , m1 ) ∼ n Y PU1 |U0 (u1,i |u0,i (m0 )) , for m1 ∈ [1 : 2nR1 ] , (56) i=1 conditionally independent given un0 (m0 ). Encoding: To send a pair of messages (m0 , m1 ) ∈ [1 : 2nR0 ] × [1 : 2nR1 ], transmit at time i ∈ [1 : n], xi = ξ (u0,i (m0 ), u1,i (m0 , m1 ), si ) . Decoding: Let PUq 0 ,U1 ,Y1 ,Y2 (u0 , u1 , y1 , y2 ) = X (57) q(s)PU0 ,U1 (u0 , u1 )WY1 ,Y2 |X,S (y1 , y2 |ξ(u0 , u1 , s), s) . (58) s∈S Observing y2n , decoder 2 finds a unique m e 0 ∈ [1 : 2nR0 ] such that (un0 (m e 0 ), y2n ) ∈ Aδ (PU0 PYq2 |U0 ) , for some q ∈ Q̂n . (59) If there is none, or more than one such m e 0 ∈ [1 : 2nR0 ], then decoder 2 declares an error. n Observing y1 , decoder 1 finds a unique pair of messages (m̂0 , m̂1 ) ∈ [1 : 2nR0 ] × [1 : 2nR1 ] such that (un0 (m̂0 ), un1 (m̂0 , m̂1 ), y1n ) ∈ Aδ (PU0 ,U1 PYq1 |U0 ,U1 ) , for some q ∈ Q̂n . (60) If there is none, or more than such pair (m̂0 , m̂1 ), then decoder 1 declares an error. We note that using the set of types Q̂n instead of the original set of state distributions Q alleviates the analysis, since Q is not necessarily finite nor countable. Analysis of Probability of Error: Assume without loss of generality that the users sent the message pair (M0 , M1 ) = (1, 1). Let q(s) ∈ Q denote the actual state distribution chosen by the jammer. By the union of events bound,     f0 6= 1 + Pr (M̂0 , M̂1 ) 6= (1, 1) , Pe(n) (q, C ) ≤ Pr M (61) where the conditioning on (M0 , M1 ) = (1, 1) is omitted for convenience of notation. The error event for decoder 2 is the union of the following events. ′ E2,1 ={(U0n (1), Y2n ) ∈ / Aδ (PU0 PYq2 |U0 ) for all q ′ ∈ Q̂n } , (62) ′ E2,2 ={(U0n (m0 ), Y n ) ∈ Aδ (PU0 PYq2 |U0 ) for some m0 6= 1, q ′ ∈ Q̂n } . Then, by the union of events bound,   f2 = Pr M 6 1 ≤ Pr (E2,1 ) + Pr (E2,2 ) . (63) (64) ′′ Considering the first term, we claim that the event E2,1 implies that (U0n (1), Y2n ) ∈ / Aδ/2 (PU0 PYq2 |U0 ) for all q ′′ ∈ Q. Assume ′′ to the contrary that E2,1 holds, but there exists q ′′ ∈ Q such that (U0n (1), Y2n ) ∈ A /2 (PU0 PYq2 |U0 ). Then, for a sufficiently large n, there exists a type q ′ (s) such that |q ′ (s) − q ′′ (s)| ≤ δ1 for all s ∈ S. It can then be inferred that q ′ ∈ Q̂n (see (53)), and ′ ′′ δ (65) |PYq2 |U0 (y2 |u0 ) − PYq2 |U0 (y2 |u0 )| ≤ |S| · δ1 = , 2 δ 13 ′ for all u0 ∈ U0 and y2 ∈ Y2 (see (54) and (58)). Hence, (U0n (1), Y2n ) ∈ Aδ (PU0 PYq2 |U0 ), which contradicts the first assumption. Thus,   ′′ δ Pr (E2,1 ) ≤ Pr (U0n (1), Y2n ) ∈ / A /2 (PU0 PYq2 |U0 ) for all q ′′ ∈ Q   δ ≤ Pr (U0n (1), Y2n ) ∈ / A /2 (PU0 PYq2 |U0 ) . (66) The last expression tends to zero exponentially as n → ∞ by the law of large numbers and Chernoff’s bound. Moving to the second term in the RHS of (64), we use the classic method of types considerations to bound Pr (E2,2 ). By the union of events bound and the fact that the number of type classes in S n is bounded by (n + 1)|S| , we have that Pr (E2,2 )   ′ ≤(n + 1)|S| · sup Pr (U0n (m0 ), Y2n ) ∈ Aδ (PU0 PYq2 |U0 ) for some m0 6= 1 . (67) q′ ∈Q̂n For every m0 6= 1,     X ′ ′ Pr (U0n (m0 ), Y2n ) ∈ Aδ (PU0 PYq2 |U0 ) = PU0n (un0 ) · Pr (un0 , Y2n ) ∈ Aδ (PU0 PYq2 |U0 ) n un 0 ∈U0 = X X PU0n (un0 ) · n un 0 ∈U0 PYq n (y2n ) , q n δ y2n : (un 0 ,y2 )∈A (PU0 PY ′ 2 |U0 (68) 2 ) ′ where the last equality holds since U0n (m0 ) is independent of Y2n for every m0 6= 1. Let (un0 , y2n ) ∈ Aδ (PU0 PYq2 |U0 ). Then, ′ y2n ∈ Aδ2 (PYq2 ) with δ2 , |U0 | · δ. By Lemmas 2.6 and 2.7 in [7], PYq n (y2n ) = 2 2   −n H(P̂yn )+D(P̂yn ||PYq ) 2 2 2 ≤2 −nH(P̂yn ) 2 ≤ 2−n(Hq′ (Y2 )−ε1 (δ)) , (69) where ε1 (δ) → 0 as δ → 0. Therefore, by (67)−(69), Pr (E2,2 ) ≤(n + 1)|S|  · sup 2nR0 · q′ ∈Q̂n |S| ≤(n + 1) X n un 0 ∈U0  ′ PU0n (un0 ) · |{y2n : (un0 , y2n ) ∈ Aδ (PU0 PYq2 |U0 )}| · 2−n(Hq′ (Y2 )−ε1 (δ))  · sup 2−n[Iq′ (U0 ;Y2 )−R0 −ε2 (δ)] , (70) q′ ∈Q with ε2 (δ) → 0 as δ → 0, where the last inequality is due to [7, Lemma 2.13]. The RHS of (70) tends to zero exponentially as n → ∞, provided that R0 < inf q′ ∈Q Iq′ (U0 ; Y2 ) − ε2 (δ). Now, consider the error event of decoder 1. For every (m0 , m1 ) ∈ [1 : 2nR0 ] × [1 : 2nR1 ], define the event ′ E1,1 (m0 , m1 ) = {(U0n (m0 ), U1n (m0 , m1 ), Y1n ) ∈ Aδ (PU0 ,U1 PYq1 |U0 ,U1 ) , for some q ′ ∈ Q̂n } . Then, the error event is bounded by n o [ (M̂0 , M̂1 ) 6= (1, 1) ⊆ E1,1 (1, 1)c ∪ E1,1 (m1 , 1) ∪ m1 ∈[1:2nR1 ] , m0 6=1 Thus, by the union of events bound,   Pr (M̂0 , M̂1 ) 6= (1, 1) X ≤ Pr (E1,1 (1, 1)c ) + X m1 6=1 Pr (E1,1 (m0 , m1 )) + ≤2−θn + 2 q′ ∈Q E1,1 (m0 , m1 ) . (72) Pr (E1,1 (m1 , 1)) m1 6=1 nR1 m1 ∈[1:2 ], m0 6=1   −n inf Iq′ (U0 ,U1 ;Y1 )−R0 −R1 −ε3 (δ) [ (71) + X Pr (E1,1 (m1 , 1)) , (73) m1 6=1 where the last inequality follows from the law of large numbers and type class considerations used before, with ε3 (δ) → 0 as δ → 0. The middle term in the RHS of (73) exponentially tends to zero as n → ∞ provided that R0 + R1 < 14 inf Iq′ (U0 , U1 ; Y1 ) − ε3 (δ). It remains for us to bound the last sum. Using similar type class considerations, we have that for q′ ∈Q every q ′ ∈ Q̂n and m1 6= 1,   ′ Pr (U0n (1), U1n (m1 , 1), Y1n ) ∈ Aδ (PU0 ,U1 PYq1 |U0 ,U1 ) X PU0n (un0 ) · PU1n |U0n (un1 |un0 ) · PYq n |U n (y1n |un0 ) = q n n δ (un 0 ,u1 ,y1 )∈A (PU0 ,U1 PY 1 ′ 1 |U0 ,U1 0 ) ≤2n(Hq′ (U0 ,U1 ,Y1 )+ε4 (δ)) · 2−n(H(U0 )−ε4 (δ)) · 2−n(H(U1 |U0 )−ε4 (δ)) · 2−n(Hq′ (Y1 |U0 )−ε4 (δ)) =2−n(Iq′ (U1 ;Y1 |U0 )−4ε4 (δ)) , (74) where ε4 (δ) → 0 as δ → 0. Therefore, the sum term in the RHS of (73) is bounded by X Pr (E1,1 (m1 , 1)) m1 6=1 = X m1 6=1   ′ Pr (U0n (1), U1n (m1 , 1), Y1n ) ∈ Aδ (PU0 ,U1 PYq1 |U0 ,U1 ) , for some q ′ ∈ Q̂n } ≤(n + 1)|S| · 2   −n inf Iq′ (U1 ;Y1 |U0 )−R1 −ε5 (δ) q′ ∈Q , (75) where the last line follows from (74), and ε5 (δ) → 0 as δ → 0. The last expression tends to zero exponentially as n → ∞ and δ → 0 provided that R1 < inf q′ ∈Q Iq′ (U1 ; Y1 |U0 ) − ε5 (δ). The probability of error, averaged over the class of the codebooks, exponentially decays to zero as n → ∞. Therefore, there must exist a (2nR0 , 2nR1 , n, e−an ) deterministic code, for a sufficiently large n. A PPENDIX B P ROOF OF T HEOREM 5 PART 1
At the first part of the theorem it is assumed that the interior of the capacity region is non-empty, i.e. int C(B Q ) 6= ∅. Achievability proof. We show that every rate pair (R0 , R1 ) ∈ Rout (B Q ) can be achieved using a code based on Shannon strategies with the addition of a codeword suffix. At time i = n + 1, having completed the transmission of the messages, the type of the state sequence sn is known to the encoder. Following the assumption that the interior of the capacity region is non-empty, the type of sn can be reliably communicated to both receivers as a suffix, while the blocklength is increased by ν > 0 additional channel uses, where ν is small compared to n. The receivers first estimate the type of sn , and then use joint typicality with respect to the estimated type.
The details are provided below. Following the assumption that int C(B Q ) 6= ∅, we have that for every ε1 > 0 and sufficiently large blocklength ν, there e e e0 > 0 and R e1 > 0. g1 , e g2 ) for the transmission of a type P̂sn at positive rates R exists a (2ν R0 , 2ν R1 , ν, ε1 ) code Ce = (feν , e Since the total number of types is polynomial in n (see [7]), the type P̂sn can be transmitted at a negligible rate, with a blocklength that grows a lot slower than n, i.e. ν = o(n) . (76) We now construct a code C over the compound broadcast channel with causal SI, such that the blocklength is n + o(n), and the rate Rn′ approaches R as n → ∞. nR0 independent sequences Codebook Generation: Fix the distribution PU0 ,U1 and Qn the function ξ(u0 , u1 , s). GeneratenR20 n nR0 ], generate 2nR1 sequences ], at random, each according to i=1 PU0 (u0,i ). For every m0 ∈ [1 : 2 u0 (m0 ), m0 ∈ [1 : 2 at random, un1 (m0 , m1 ) ∼ n Y PU1 |U0 (u1,i |u0,i (m0 )) , for m1 ∈ [1 : 2nR1 ] , (77) i=1 conditionally independent given un0 (m0 ). Reveal the codebook of the message pair (m0 , m1 ) and the codebook of the type P̂sn to the encoder and the decoders. Encoding: To send a message pair (m0 , m1 ) ∈ [1 : 2nR0 ] × [1 : 2nR1 ], transmit at time i ∈ [1 : n], xi = ξ (u0,i (m0 ), u1,i (m0 , m1 ), si ) . (78) At time i ∈ [n + 1 : n + ν], knowing the sequence of previous states sn , transmit xi = fei (P̂sn , sn+1 , . . . , sn+i ) , (79) 15 where P̂sn is the type of the sequence (s1 , . . . , sn ). That is, the encoded type P̂sn is transmitted as a suffix of the codeword. We note that the type of the P̂sn , and it is irrelevant for that matter, since the
sequence (sn+1 , . . . , sn+i ) is notν Rnecessarily e1 e0 νR Q g2 ) for the transmission of assumption that int C(B ) 6= ∅ implies that there exists a (2 , 2 , ν, ε1 ) code Ce = (feν , ge1 , e e0 > 0 and R e1 > 0. P̂sn , with R Decoding: Let X PUq 0 ,U1 ,Y1 ,Y2 (u0 , u1 , y1 , y2 ) = (80) q(s)PU0 ,U1 (u0 , u1 )WY1 ,Y2 |X,S (y1 , y2 |ξ(u0 , u1 , s), s) . s∈S y2n+ν . Decoder 2 receives the output sequence As a pre-decoding step, the receiver decodes the last ν output symbols, and finds an estimate of the type of the state sequence, qb2 = e g2 (y2,n+1 , . . . , y2,n+ν ). Then, given the output sequence y2n , decoder nR0 ] such that 2 finds a unique m e 0 ∈ [1 : 2 (un0 (m e 0 ), y2n ) ∈ Aδ (PU0 PYqb22|U0 ) . (81) (un0 (m̂0 ), un1 (m̂0 , m̂1 ), y1n ) ∈ Aδ (PU0 ,U1 PYqb11|U0 ,U1 ) . (82) If there is none, or more than one such m e 0 ∈ [1 : 2nR0 ], then decoder 2 declares an error. n+ν Similarly, decoder 1 receives y1 and begins with decoding the type of the state sequence, qb1 = e g1 (y1,n+1 , . . . , y1,n+ν ). Then, decoder 1 finds a unique pair of messages (m̂0 , m̂1 ) ∈ [1 : 2nR0 ] × [1 : 2nR1 ] such that If there is none, or more than one such pair (m̂0 , m̂1 ) ∈ [1 : 2nR0 ] × [1 : 2nR1 ], then decoder 1 declares an error. Analysis of Probability of Error: By symmetry, we may assume without loss of generality that Qn the users sent (M0 , M1 ) = (1, 1). Let q(s) ∈ Q denote the actual state distribution chosen by the jammer, and let q(sn ) = i=1 q(si ). Then, by the union of events bound, the probability of error is bounded by     f0 6= 1 + Pr (M̂0 , M̂1 ) 6= (1, 1) , Pe(n) (q, C ) ≤ Pr M (83) where the conditioning on (M0 , M1 ) = (1, 1) is omitted for convenience of notation. Define the events E1,0 = {b q1 6= P̂S n } ′ E1,1 (m0 , m1 , q ) = (84) {(U0n (m0 ), U1n (m0 , m1 ), Y1n ) δ ∈A ′ (PU0 ,U1 PYq1 |U0 ,U1 )} (85) and E2,0 = {b q2 6= P̂S n } ′ E2,1 (m0 , q ) = (86) {(U0n (m0 ), Y2n ) δ ∈A ′ (PU0 PYq2 |U0 )} , (87) for every m0 ∈ [1 : 2nR0 ], m1 ∈ [1 : 2nR1 ], and q ′ ∈ P(S). The error event of decoder 2 is bounded by o n [ f2 6= 1 ⊆ E2,0 ∪ E2,1 (1, qb2 )c ∪ M E2,1 (m0 , qb2 ) m0 6=1 = E2,0   [
c c ∪ E2,0 ∩ E2,1 (1, qb2 )c ∪  E2,0 ∩ E2,1 (m0 , qb2 ) . m0 6=1 By the union of events bound,   f2 6= 1 Pr M   [
c c ≤ Pr (E2,0 ) + Pr E2,0 ∩ E2,1 (1, qb2 )c + Pr  E2,0 ∩ E2,1 (m0 , qb2 ) . (88) Pr (E1,0 ∪ E2,0 ) ≤ ε1 . (89) m0 6=1 e e Since the code Ce for the transmission of the type is a (2ν R0 , 2ν R1 , ν, ε1 ) code, where ε1 > 0 is arbitrarily small, we have that the probability of erroneous decoding of the type is bounded by 16 Thus, the first term in the RHS of (88) is bounded by ε1 . Then, we maniplute the last two terms as follows.   X
c f2 6= 1 ≤ q(sn ) Pr E2,0 ∩ E2,1 (1, qb2 )c | S n = sn Pr M sn ∈Aδ2 (q) X + sn ∈A / δ2 (q) X + sn ∈Aδ2 (q) + X sn ∈A / δ2 (q) where c q(sn ) Pr E2,0 ∩ E2,1 (1, qb2 )c | S n = sn  q(sn ) Pr  [ m0 6=1  q(sn ) Pr  [ m0 6=1 δ2 ,
 c E2,0 ∩ E2,1 (m0 , qb2 ) | S n = sn   c E2,0 ∩ E2,1 (m0 , qb2 ) | S n = sn  + ε1 , 1 ·δ. 2|S| (90) (91) Next we show that the first and the third sums in (90) tend to zero as n → ∞. Consider a given sn ∈ Aδ2 (q). For notational convenience, denote q ′′ = P̂sn . (92) Then, by the definition of the δ-typical set, we have that |q ′′ (s) − q(s)| ≤ δ2 for all s ∈ S. It follows that ′′ |PU0 (u0 )PYq2 |U0 (y|u0 ) − PU0 (u0 )PYq2 |U0 (y2 |u0 )| X δ PU1 |U0 (u1 |u0 )WY2 |X,S (y2 |ξ(u0 , u1 , s), s) ≤ δ2 · |S| = , ≤δ2 · 2 s,u (93) 1 for all u0 ∈ U0 and y2 ∈ Y2 , where the last equality follows from (91). Consider the first sum in the RHS of (90). Given a state sequence sn ∈ Aδ2 (q), we have that  
c c ∩ E2,1 (1, P̂sn )c | S n = sn Pr E2,0 ∩ E2,1 (1, qb2 )c | S n = sn = Pr E2,0
c = Pr E2,0 ∩ E2,1 (1, q ′′ )c | S n = sn
c = Pr E2,0 | E2,1 (1, q ′′ )c , S n = sn · Pr (E2,1 (1, q ′′ )c ) | S n = sn ) , where the first equality follows from (86), and the second equality follows from (92). Then,
c Pr E2,0 ∩ E2,1 (1, qb2 )c | S n = sn ≤ Pr (E2,1 (1, q ′′ )c | S n = sn )   ′′ = Pr (U0n (1), Y2n ) ∈ / Aδ (PU0 PYq2 |U0 ) S n = sn . (94) (95) Now, suppose that (U0n (1), Y2n ) ∈ Aδ/2 (PU0 PYq2 |U0 ), where q is the actual state distribution. By (93), in this case we have that ′′ (U0n (1), Y2n ) ∈ Aδ (PU0 PYq2 |U0 ). Hence, (95) implies that  
δ c / A /2 (PU0 PYq2 |U0 ) S n = sn . Pr E2,0 ∩ E2,1 (1, qb2 )c | S n = sn ≤ Pr (U0n (1), Y2n ) ∈ (96) The first sum in the RHS of (90) is then bounded as follows. X
c q(sn ) Pr E2,0 ∩ E2,1 (1, qb2 )c | S n = sn sn ∈Aδ2 (q) ≤ X sn ∈Aδ2 (q) ≤ X sn ∈S n   δ q(sn ) Pr (U0n (1), Y2n ) ∈ / A /2 (PU0 PYq2 |U0 ) S n = sn   δ q(sn ) Pr (U0n (1), Y2n ) ∈ / A /2 (PU0 PYq2 |U0 ) S n = sn   δ / A /2 (PU0 PYq2 |U0 ) ≤ ε2 , = Pr (U0n (1), Y2n ) ∈ for a sufficiently large n, where the last inequality follows from the law of large numbers. (97) 17 ′′ We bound the third sum in the RHS of (90) using similar arguments. If (U0n (m0 ), Y2n ) ∈ Aδ (PU0 PYq2 |U0 ), then (U0n (m0 ), Y2n ) ∈ 3δ A /2 (PU0 PYq2 |U0 ), due to (93). Thus, for every sn ∈ Aδ2 (q),   [ X c Pr  E2,0 ∩ E2,1 (m0 , qb2 ) | S n = sn  ≤ Pr (E2,1 (m0 , q ′′ ) | S n = sn ) m0 6=1 m0 6=1 = X m0 6=1 ≤ X m0 6=1   ′′ Pr (U0n (m0 ), Y2n ) ∈ Aδ (PU0 PYq2 |U0 ) S n = sn   3δ Pr (U0n (m0 ), Y2n ) ∈ A /2 (PU0 PYq2 |U0 ) S n = sn . (98) This, in turn, implies that the third sum in the RHS of (90) is bounded by   X [ c q(sn ) Pr  E2,0 ∩ E2,1 (m0 , qb ) | S n = sn  sn ∈Aδ2 (q) ≤ X X sn ∈S n m0 6=1 = X m0 6=1 m0 6=1   3δ q(sn ) · Pr (U0n (m0 ), Y2n ) ∈ A /2 (PU0 PYq2 |U0 ) S n = sn   3δ Pr (U0n (m0 ), Y2n ) ∈ A /2 (PU0 PYq2 |U0 ) ≤2−n[Iq (U0 ;Y2 )−R0 −ε2 (δ)] , (99) with ε2 (δ) → 0 as δ → 0. The last inequality follows from standard type class considerations. The RHS of (99) tends to zero as n → ∞, provided that R0 < Iq (U0 ; Y2 ) − ε2 (δ). Then, it follows from the law of large numbers that the second and fourth sums RHS of (90) tend to zero as n → ∞. Thus, by (97) and (99), we have that the probability of error of decoder 2,  in the  f Pr M2 6= 1 , tends to zero as n → ∞. Now, consider the error event of decoder 1, n o [ [ E1,1 (m0 , m1 , qb1 ) ∪ E1,1 (1, m1 , qb1 ) . (100) (M̂0 , M̂1 ) 6= (1, 1) ⊆ E1,0 ∪ E1,1 (1, 1, qb1 )c ∪ m1 6=1 m0 6=1 , m1 ∈[1:2nR1 ] Thus, by the union of events bound,    
c Pr (M̂0 , M̂1 ) 6= (1, 1) ≤ Pr (E1,0 ) + Pr E1,0 ∩ E1,1 (1, 1, qb1 )c + Pr    + Pr  [ m1 6=1  [ m0 6=1 , m1 ∈[1:2nR1 ]   c E1,0 ∩ E1,1 (m0 , m1 , qb1 )  c E1,0 ∩ E1,1 (m1 , 1, qb1 ) . (101) By (89), the first term is bounded by ε1 , and as done above, we write     X c Pr (M̂0 , M̂1 ) 6= (1, 1) ≤ q(sn ) Pr E1,0 ∩ E1,1 (1, 1, P̂sn )c | S n = sn sn ∈Aδ2 (q) +  X sn ∈Aδ2 (q) + X sn ∈Aδ2 (q)  q(sn ) Pr    q(sn ) Pr  [ m1 ∈[1:2nR1 ] m0 6=1 [ δ2 m1 6=1  c E1,0 ∩ E1,1 (m0 , m1 , P̂sn ) | S n = sn    c E1,0 ∩ E1,1 (m1 , 1, P̂sn ) | S n = sn 
+ 3 · Pr S ∈ / A (q) + ε1 , n  (102)
/ Aδ2 (q) tends to zero as n → ∞. As for where δ2 is given by (91). By the law of large numbers, the probability Pr S n ∈ the sums, we use similar arguments to those used above. 18 We have that for a given sn ∈ Aδ2 (q), ′′ |PU0 ,U1 (u0 , u1 )PYq1 |U0 ,U1 (y1 |u0 , u1 ) − PU0 ,U1 (u0 , u1 )PYq1 |U0 ,U1 (y1 |u0 , u1 )| X δ ≤ δ2 · WY1 |X,S (y1 |ξ(u0 , u1 , s) ≤ |S| · δ2 = , 2 (103) s∈S with q ′′ = P̂sn , where the last equality follows from (91). The first sum in the RHS of (102) is bounded by   X c q(sn ) Pr E1,0 ∩ E1,1 (1, 1, P̂sn )c | S n = sn sn ∈Aδ2 (q) ≤ X sn ∈S n   δ q(sn ) Pr (U0n (1), U1n (1, 1), Y1n ) ∈ / A /2 (PU0 ,U1 PYq1 |U0 ,U1 ) | S n = sn   δ / A /2 (PU0 ,U1 PYq1 |U0 ,U1 ) ≤ ε2 . = Pr (U0n (1), U1n (1, 1), Y1n ) ∈ The last inequality follows from the law of large numbers, for a sufficiently large n. The second sum in the RHS of (102) is bounded by   X sn ∈Aδ2 (q)  q(sn ) Pr   [ m1 ∈[1:2nR1 ] m0 6=1  −n(Iq (U0 ,U1 ;Y1 )−R0 −R1 −ε3 (δ) c , E1,0 ∩ E1,1 (m0 , m1 , P̂sn ) | S n = sn  ≤2 (104) (105) with ε3 (δ) → 0 as n → ∞ and δ → 0. This is obtained following the same analysis as for decoder 2. Then, the second sum tends to zero provided that R0 + R1 < Iq (U0 , U1 ; Y1 ) − ε3 (δ). The third sum in the RHS of (102) is bounded by     [ X X X c q(sn ) Pr  E1,0 ∩ E1,1 (m1 , 1, P̂sn ) | S n = sn  ≤ q(sn ) Pr E1,1 (m1 , 1, P̂sn ) | S n = sn . sn ∈Aδ2 (q) m1 6=1 sn ∈Aδ2 (q) m1 6=1 (106) For every sn ∈ Aδ2 (q), it follows from (103) that the event E1,1 (m1 , 1, P̂sn ) implies that (U0n (1), U1n (m1 , 1), Y1n ) ∈ A3δ/2 (PUq 0 ,U1 ,Y1 ). Thus, the sum is bounded by   X [ c q(sn ) Pr  E1,0 ∩ E1,1 (m1 , 1, P̂sn ) | S n = sn  ≤ 2−n(Iq (U1 ;Y1 |U0 )−R1 −δ3 ) , sn ∈Aδ2 (q) (107) m1 6=1 where δ3 → 0 as δ → 0. Then, the RHS of (107) tends to zero as n → ∞ provided that R1 < Iq (U1 ; Y1 |U0 ) − δ3 . We conclude that the RHS of both (90) and (102) tend to zero as n → ∞. Thus, the overall probability of error, averaged over the class of the codebooks, decays to zero as n → ∞. Therefore, there must exist a (2nR0 , 2nR1 , n, ε) deterministic code, for a sufficiently large n. Converse proof. First, we claim that it can be assumed that U0 U1 X form a Markov chain. Define the following region,   R0 ≤ Iq (U0 ; Y2 ) ,   (R0 , R1 ) : \ [ R1 ≤ Iq (U1 ; Y1 |U0 ) RM,out (B Q ) = , (108)   R + R ≤ I (U , U ; Y ) e 1 ,s) q(s)∈Q p(u0 ,u1 ), ξ(u 0 1 q 0 1 1 e 1 , S). Clearly, RM,out (B Q ) ⊆ Rout (B Q ), since RM,out (B Q ) is obtained by restriction of the function subject to X = ξ(U ξ(u0 , u1 , s) in the union on the RHS of (15). Moreover, we have that RM,out (B Q ) ⊇ Rout (B Q ), since, given some U0 , U1 e1 = (U0 , U1 ), and then X is a deterministic function of (U e1 , S). and ξ(u0 , u1 , s), we can define a new strategy variable U Q Q As Rout (B ) = RM,out (B ), it can now be assumed that U0 U1 X (Y1 , Y2 ) form a Markov chain, hence Iq (U0 , U1 ; Y1 ) = Iq (U1 ; Y1 ). Then, by similar arguements to those used in [14] (see also [7, Chapter 16]), we have that   R0 ≤ Iq (U0 ; Y2 ) ,  (R0 , R1 ) :  \ [ R0 + R1 ≤ Iq (U1 ; Y1 |U0 ) + Iq (U0 ; Y2 ) Rout (B Q ) = . (109)   R + R ≤ I (U ; Y ) e q(s)∈Q p(u0 ,u1 ), ξ(u1 ,s) 0 1 q 1 1 We show that for every sequence of (2nR0 , 2nR1 , n, θn ) codes, with limn→∞ θn = 0, we have that (R0 , R1 ) belongs to the set above. 19 Define the following random variables, n U0,i , (M0 , Y1i−1 , Y2,i+1 ) , U1,i , (M0 , M1 , S i−1 ) . (110) It follows that Xi is a deterministic function of (U1,i , Si ), and since the state sequence is memoryless, we have that Si is independent of (U0,i , U1,i ). Next, by Fano’s inquality, nR0 ≤ Iq (M0 ; Y2n ) + nεn , n(R0 + R1 ) ≤ Iq (M0 , M1 ; Y1n ) + nεn , (111) (112) n(R0 + R1 ) ≤ Iq (M1 ; Y1n |M0 ) + Iq (M0 ; Y2n ) + nεn , (113) where εn → 0 as n → ∞. Applying the chain rule, we have that (111) is bounded by Iq (M0 ; Y2n ) = n X n Iq (M0 ; Y2,i |Y2,i+1 )≤ Iq (M0 , M1 ; Y1n ) = n X Iq (U0,i ; Y2,i ) , Iq (M0 , M1 ; Y1,i |Y1i−1 ) ≤ n X Iq (U0,i , U1,i ; Y1,i ) = n X Iq (U1,i ; Y1,i ) , (115) i=1 i=1 i=1 (114) i=1 i=1 and (112) is bounded by n X where the last equality holds since U0,i U1,i Y1,i form a Markov chain. As for (113), we have that Iq (M1 ; Y1n |M0 ) + Iq (M0 , Y2n ) = n X Iq (M1 ; Y1,i |M0 , Y1i−1 ) + ≤ n Iq (M0 ; Y2,i |Y2,i+1 ) i=1 i=1 n X n X n Iq (M1 , Y2,i+1 ; Y1,i |M0 , Y1i−1 ) + = n Iq (M1 ; Y1,i |M0 , Y1i−1 , Y2,i+1 )+ n X n Iq (Y2,i+1 ; Y1,i |M0 , Y1i−1 ) i=1 i=1 + n Iq (M0 , Y2,i+1 ; Y2,i ) i=1 i=1 n X n X n X n Iq (M0 , Y1i−1 , Y2,i+1 ; Y2,i ) − n X n Iq (Y1i−1 ; Y2,i |M0 , Y2,i+1 ). (116) i=1 i=1 Then, the second and fourth sums cancel out, by the Csiszár sum identity [9, Section 2.3]. Hence, Iq (M1 ; Y1n |M0 ) + Iq (M0 ; Y2n ) ≤ n X n Iq (M1 ; Y1,i |M0 , Y1i−1 , Y2,i+1 )+ ≤ n Iq (M0 , Y1i−1 , Y2,i+1 ; Y2,i ) i=1 i=1 n X n X Iq (U1,i ; Y1,i |U0,i ) + i=1 Thus, by (111)–(113) and (115)–(117), we have that n X Iq (U0,i ; Y2,i ) . (117) i=1 n R0 ≤ 1X Iq (U0,i ; Y2,i ) + εn , n i=1 (118) n R0 + R1 ≤ 1X Iq (U1,i ; Y1,i ) + εn , n i=1 n R0 + R1 ≤ (119) n X 1X Iq (U0,i ; Y2,i ) + εn . Iq (U1,i ; Y1,i |U0,i ) + n i=1 i=1 (120) Introducing a time-sharing random variable K, uniformly distributed over [1 : n] and independent of (S n , U0n , U1n ), we have that R0 ≤Iq (U0,K ; Y2,K |K) + εn , R0 + R1 ≤Iq (U1,K ; Y1,K |K) + εn , R0 + R1 ≤Iq (U1,K ; Y1,K |U0,K , K) + Iq (U0,K ; Y2,K |K) + εn . (121) (122) (123) Define U0 , (U0,K , K) and U1 , (U1,K , K). Hence, PY1,K ,Y2,K |U0 ,U1 = PY1 ,Y2 |U0 ,U1 . Then, by (109) and (121)–(123), it follows that (R0 , R1 ) ∈ Rout (B Q ). 20 PART 2 We show that when the set of state distributions Q is convex, and Condition T Q holds, the capacity region of the compound broadcast channel B Q with causal SI is given by C(B Q ) = C⋆(B Q ) = Rin (B Q ) = Rout (B Q ) (and this holds regardless of whether the interior of the capacity region is empty or not). Due to part 1, we have that C⋆(B Q ) ⊆ Rout (B Q ) . (124) C(B Q ) ⊇ Rin (B Q ) . (125) By Lemma 4, Thus, Rin (B Q ) ⊆ C(B Q ) ⊆ C⋆(B Q ) ⊆ Rout (B Q ) . Q Q (126) Q To conclude the proof, we show that Condition T implies that Rin (B ) ⊇ Rout (B ), hence the inner and outer bounds coincide. By Definition 4, if a function ξ(u0 , u1 , s) and a set D achieve Rin (B Q ) and Rout (B Q ), then   R0 ≤ minq∈Q Iq (U0 ; Y2 ) ,  (R0 , R1 ) :  [ R1 ≤ minq∈Q Iq (U1 ; Y1 |U0 ) , , (127a) Rin (B Q ) =   R0 + R1 ≤ minq∈Q Iq (U0 , U1 ; Y1 ) p(u0 ,u1 )∈D and Rout (B Q ) = \ [ q(s)∈Q p(u0 ,u1 )∈D   (R0 , R1 ) :  R0 R1 R0 + R1 Hence, when Condition T Q holds, we have by Definition 5 that for some  R0  (R0 , R1 ) : [ R1 Rin (B Q ) =  R0 + R1 p(u0 ,u1 )∈D ⊇Rout (B Q ) ,  ≤ Iq (U0 ; Y2 ) ,  ≤ Iq (U1 ; Y1 |U0 ) , .  ≤ Iq (U0 , U1 ; Y1 ) (127b) ξ(u0 , u1 , s), D ⊆ P(U0 × U1 ), and q ∗ ∈ Q,  ≤ Iq∗ (U0 ; Y2 ) ,  ≤ Iq∗ (U1 ; Y1 |U0 ) ,  ≤ Iq∗ (U0 , U1 ; Y1 ) (128) where the last line follows from (127b). A PPENDIX C P ROOF OF T HEOREM 6 At first, ignore the cardinality bounds in (21). Then, it immediately follows from Theorem 5 that C(B q ) = C(B q ), by taking the set Q that consists of a single state distribution q(s). To prove the bounds on the alphabet sizes of the strategy variables U0 and U1 , we apply the standard Carathéodory techniques (see e.g. [7, Lemma 15.4]). Let L0 , (|X | − 1)|S| + 3 ≤ |X ||S| + 2 , (129) where the inequality holds since |S| ≥ 1. Without loss of generality, assume that X = [1 : |X |] and S = [1 : |S|]. Then, define the following L0 functionals, X ϕij (PU1 ,X|S ) = PU1 ,X|S (u1 , i|j) = PX|S (i|j) , i = 1, . . . |X | − 1, j = 1, . . . , |S| , (130) u1 ∈U1 ψ1 (PU1 ,X|S ) = − X s,u1 ,x,y1 ψ2 (PU1 ,X|S ) = − X s,u1 ,x,y2 ψ3 (PU1 ,X|S ) = − X u1 ,x,s − X u1 ,x,s  q(s)PU1 ,X|S (u1 , x|s)WY1 |X,S (y1 |x, s) log   q(s)PU1 ,X|S (u1 , x|s)WY2 |X,S (y2 |x, s) log   q(s)PU1 ,X|S (u1 , x|s) log  X x′ ,s′ " X s′ ,u′1 ,x′ X s′ ,u′1 ,x′  q(s′ )PU1 ,X|S (u′1 , x′ |s′ )WY1 |X,S (y1 |x′ , s′ ) , (131)  q(s′ )PU1 ,X|S (u′1 , x′ |s′ )WY2 |X,S (y2 |x′ , s′ ) , q(s′ )PU1 ,X|S (u1 , x′ |s′ ) q(s)PU1 ,X|S (u1 , x|s)WY1 |X,S (y1 |x, s) log P P x′ ,s′ ′′ ′′ u′′ 1 ,x ,s (132)  q(s′ )PU1 ,X|S (u1 , x′ |s′ )WY1 |X,S (y1 |x′ , s′ ) q(s′′ )PU1 ,X|S (u′′1 , x′′ |s′′ )WY1 |X,S (y1 |x′′ , s′′ ) # . (133) 21 Then, observe that X p(u0 )ϕi,j (PU1 ,X|S,U0 (·, ·|·, u0 )) = PX|S (i|j) , (134) p(u0 )ψ1 (PU1 ,X|S,U0 (·, ·|·, u0 )) = H(Y1 |U0 ) , (135) p(u0 )ψ1 (PU1 ,X|S,U0 (·, ·|·, u0 )) = H(Y2 |U0 ) , (136) p(u0 )ψ1 (PU1 ,X|S,U0 (·, ·|·, u0 )) = I(U1 ; Y1 |U0 ) . (137) u0 ∈U0 X u0 ∈U0 X u0 ∈U0 X u0 ∈U0 By [7, Lemma 15.4], the alphabet size of U0 can then be restricted to |U0 | ≤ L0 , while preserving PX,S,Y1 ,Y2 ; I(U0 ; Y2 ) = H(Y2 ) − H(Y2 |U0 ); I(U0 ; Y1 |U0 ); and I(U0 , U1 ; Y1 ) = I(U0 ; Y1 |U0 ) + H(Y1 ) − H(Y1 |U0 ). Fixing the alphabet of U0 , we now apply similar arguments to the cardinality of U1 . Then, less than |X ||S|L0 − 1 functionals are required for the joint distribution PU0 ,X|S , and an additional functional to preserve H(Y1 |U1 , U0 ). Hence, by [7, Lemma 15.4], the alphabet size of U0 can then be restricted to |U1 | ≤ |X ||S|L0 ≤ |X ||S|(|X ||S| + 2) (see (129)). A PPENDIX D P ROOF OF T HEOREM 7 A. Part 1 First, we explain the general idea. We devise a causal version of Ahlswede’s Robustification Technique (RT) [1, 19]. Namely, we use codes for the compound broadcast channel to construct a random code for the AVBC using randomized permutations. However, in our case, the causal nature of the problem imposes a difficulty, and the application of the RT is not straightforward. In [1, 19], the state information is noncausal and a random code is defined via permutations of the codeword symbols. This cannot be done here, because the SI is provided to the encoder in a causal manner. We resolve this difficulty using Shannon strategy codes for the compound broadcast channel to construct a random code for the AVBC, applying permutations to the strategy sequence (un1 , un0 ), which is an integral part of the Shannon strategy code, and is independent of the channel state. The details are given below. 1) Inner Bound: We show that the region defined in (23) can be achieved by random codes over the AVBC B with causal SI, i.e. C(B) ⊇ R⋆in (B). We start with Q Ahlswede’s RT [1], stated below. Let h : S n → [0, 1] be a given function. If, for some n n fixed αn ∈ (0, 1), and for all q(s ) = i=1 q(si ), with q ∈ P(S), X q(sn )h(sn ) ≤ αn , (138) sn ∈S n then, 1 X h(πsn ) ≤ βn , n! for all sn ∈ S n , (139) π∈Πn where Πn is the set of all n-tuple permutations π : S n → S n , and βn = (n + 1)|S| · αn . According to Lemma 4, for every (R0 , R1 ) ∈ R⋆in (B), there exists a (2nR0 , 2nR1 , n, e−2θn ) Shannon strategy code for the compound broadcast channel B P(S) with causal SI, for some θ > 0 and sufficiently large n. Given such a Shannon strategy (n) code C = (un0 (m0 ), un1 (m0 , m1 ), ξ(u0 , u1 , s), g1 (y1n ), g2 (y2n )), we have that (138) is satisfied with h(sn ) = Pe|sn (C ) and αn = e−2θn . As a result, Ahlswede’s RT tells us that 1 X (n) Pe|πsn (C ) ≤ (n + 1)|S| e−2θn ≤ e−θn , for all sn ∈ S n , (140) n! π∈Πn for a sufficiently large n, such that (n + 1)|S| ≤ eθn . On the other hand, for every π ∈ Πn , X X 1 (a) (n) WY1n ,Y2n |X n ,S n (πy1n , πy2n |ξ n (un0 (m0 ), un1 (m0 , m1 ), πsn ), πsn ) Pe|πsn (C ) = n(R +R ) 2 0 1 m ,m n n 0 (b) = (c) = 1 2n(R0 +R1 ) 1 2n(R0 +R1 ) X 1 (πy1 ,πy2 )∈D(m / 0 ,m1 ) X WY1n ,Y2n |X n ,S n (y1n , y2n |π −1 ξ n (un0 (m0 ), un1 (m0 , m1 ), πsn ), sn ) , m0 ,m1 (πy1n ,πy2n )∈D(m / 0 ,m1 ) X X m0 ,m1 (πy1n ,πy2n )∈D(m / 0 ,m1 ) WY1n ,Y2n |X n ,S n (y1n , y2n |ξ n (π −1 un0 (m0 ), π −1 un1 (m0 , m1 ), sn ), sn ) (141) 22 where (a) is obtained by plugging πsn and xn = ξ n (·, ·, ·) in (2) and then changing the order of summation over (y1n , y2n ); (b) holds because the broadcast channel is memoryless; and (c) follows from that fact that for a Shannon strategy code, xi = ξ(u0,i , u1,i , si ), i ∈ [1 : n], by Definition 3. The last expression suggests the use of permutations applied to the encoding strategy sequence and the channel output sequences. Then, consider the (2nR0 , 2nR1 , n) random code C Π , specified by fπn (m0 , m1 , sn ) = ξ n (π −1 un1 (m0 , m1 ), π −1 un0 (m0 ), sn ) , (142a) and g1,π (y1n ) = g1 (πy1n ) , 1 |Πn | g2,π (y2n ) = g(πy2n ) , 1 n! . (142b) n for π ∈ Πn , with a uniform distribution µ(π) = = Such permutations can be implemented without knowing s , hence this coding scheme does not violate the causality requirement. From (141), we see that X (n) (n) Pe|sn (C Π ) = µ(π)Pe|πsn (C ) , (143) π∈Πn n n for all s ∈ S , and therefore, together with (140), we have that the probability of error of the random code C Π is bounded by Pe(n) (q, C Π ) ≤ e−θn , (144) for every q(sn ) ∈ P(S n ). That is, C Π is a (2nR0 , 2nR1 , n, e−θn ) random code for the AVBC B with causal SI at the encoder. This completes the proof of the inner bound. 2) Outer Bound: We show that the capacity region of the AVBC B with causal SI is bouned by C⋆(B) ⊆ R⋆out (B) (see (23)). The random code capacity region of the AVBC is included within the random code capacity region of the compound broadcast channel, namely C⋆(B) ⊆ C⋆(B P(S) ) . (145) By Theorem 5 we have that C⋆(B Q ) ⊆ Rout (B Q ). Thus, with Q = P(S), C⋆(B P(S) ) ⊆ R⋆out (B) . (146) It follows from (145) and (146) that C⋆(B) ⊆ R⋆out (B). Since the random code capacity region always includes the deterministic code capacity region, we have that C(B) ⊆ R⋆ (B) as well. out Part 2 The second equality, R⋆in (B) = R⋆out (B), follows from part 2 of Theorem 5, taking Q = P(S). By part 1, R⋆in (B) ⊆ C⋆(B) ⊆ ⋆ Rout (B), hence the proof follows. A PPENDIX E P ROOF OF L EMMA 8 The proof follows the lines of [2, Section 4]. Let k > 0 be an integer, chosen later, and define the random variables L1 , L2 , . . . , Lk i.i.d. ∼ µ(ℓ) . (147) n Fix s , and define the random variables (n) Ωj (sn ) = Pe|sn (CLj ) , j ∈ [1 : k] , (148) which is the conditional probability of error of the code CLj given the state sequence sn . P P (n) Since C Γ is a (2nR1 , 2nR2 , n, εn ) code, we have that γ µ(γ) sn q(sn )Pe|sn (Cγ ) ≤ εn , for all q(sn ). In particular, for a kernel, we have that X (n) EΩj (sn ) = µ(γ) · Pe|sn (Cγ ) ≤ εn , (149) γ∈Γ for all j ∈ [1 : k]. 23 Now take n to be large enough so that εn < α. Keeping sn fixed, we have that the random variables Ωj (sn ) are i.i.d., due to (147). Next the technique known as Bernstein’s trick [2] is applied.       k k  X X (a)  Pr  Ωj (sn ) ≥ kα ≤ E exp β  Ωj (sn ) − kα (150)   j=1 j=1   k  Y n eβΩj (s ) (151) =e−βkα · E   j=1 (b) −βkα =e · k Y j=1 (c) ≤ e−βkα · k Y j=1 (d) n o n E eβΩj (s )  E 1 + eβ · Ωj (sn ) ≤ e−βkα · 1 + eβ εn
k (152) (153) (154) where (a) is an application of Chernoff’s inequality; (b) follows from the fact that Ωj (sn ) are independent; (c) holds since eβx ≤ 1 + eβ x, for β > 0 and 0 ≤ x ≤ 1; (d) follows from (149). We take n to be large enough for 1 + eβ εn ≤ eα to hold. Thus, choosing β = 2, we have that   k X 1 Ωj (sn ) ≥ α ≤e−αk , Pr  (155) k j=1 for all sn ∈ S n . Now, by the union of events bound, we have that     k k X X 1 1 Ωj (sn ) ≥ α = Pr ∃sn : Ωj (sn ) ≥ α Pr max sn k k j=1 j=1   k X X 1 ≤ Pr  Ωj (sn ) ≥ α k n n j=1 (156) (157) s ∈S ≤|S|n · e−αk . (158) Since |S|n grows only exponentially in n, choosing k = n2 results in a super exponential decay. ∗ Consider the code C Γ = (µ∗ , Γ∗ = [1 : k], {CLj }kj=1 ) formed by a random collection of codes, with µ∗ (j) = k1 . It follows that the conditional probability of error given sn , which is given by (n) ∗ Pe|sn (C Γ ) = k 1 X (n) P n (CLj ) , k j=1 e|s 2 (159) ∗ exceeds α with a super exponentially small probability ∼ e−αn , for all sn ∈ S n . Thus, there exists a random code C Γ = (µ∗ , Γ∗ , {Cγj }kj=1 ) for the AVBC B, such that X ∗ ∗ (n) Pe(n) (q, C Γ ) = q(sn )Pe|sn (C Γ ) ≤ α , for all q(sn ) ∈ P(S n ) . (160) sn ∈S n A PPENDIX F P ROOF OF T HEOREM 9 Achievability proof. To show achievability, we follow the lines of [2], with the required adjustments. We use the random code constructed in the proof of Theorem 7 to construct a deterministic
code. Let (R0 , R1 ) ∈ C⋆(B), and consider the case where int C(B) 6= ∅. Namely, C(W1 ) > 0 , and C(W2 ) > 0 , (161) where W1 = {WY1 |X,S } and W2 = {WY2 |X,S } denote the marginal AVCs with causal SI of user 1 and user 2, respectively. By Lemma 8, for
every ε1 > 0 and sufficiently large n, there exists a (2nR0 , 2nR1 , n, ε1 ) random code C Γ = µ(γ) = k1 , Γ = [1 : k], {Cγ }γ∈Γ , where Cγ = (fγn , g1,γ , g2,γ ), for γ ∈ Γ, and k = |Γ| ≤ n2 . Following (161), we have that for every 24 e e ε2 > 0 and sufficiently large ν, the code index γ ∈ [1 : k] can be sent over B using a (2ν R0 , 2ν R1 , ν, ε2 ) deterministic code e0 > 0, R e1 > 0. Since k is at most polynomial, the encoder can reliably convey γ to the receiver Ci = (feν , e g1 , ge0 ), where R with a negligible blocklength, i.e. ν = o(n). Now, consider a code formed by the concatenation of Ci as a prefix to a corresponding code in the code collection {Cγ }γ∈Γ . That is, the encoder sends both the index γ and the message pair (m0 , m1 ) to the receivers, such that the index γ is transmitted first by feν (γ, sν ), and then the message pair (m0 , m1 ) is transmitted by the codeword xn = fγn ( m0 , m1 , sν+1 , . . . , sν+n ). Subsequently, decoding is performed in two stages as well; decoder 1 estimates the index at first, with γ b1 = e g1 (y1,1 , . . . , y1,ν ), and the message pair (m0 , m1 ) is then estimated by (m b 0, m b 1 ) = g1,bγ1 (y1,ν+1 , . . . , y1,ν+n ). Similarly, decoder 2 estimates the index with γ b2 = e g0 (y2,1 , . . . , y2,ν ), and the message m0 is then estimated by m e 2 = g2,bγ2 (y2,ν+1 , . . . , y2,ν+n ). By the union of events bound, the probability of error is then bounded by ε = ε1 + ε2 , for every joint distribution in e e P ν+n (S ν+n ). That is, the concatenated code is a (2(ν+n)R1,n , 2(ν+n)R2,n , ν + n, ε) code over the AVBC B with causal SI, e1,n = n · R1 approach R0 and e0,n = n · R0 and R where ν = o(n). Hence, the blocklength is n + o(n), and the rates R ν+n ν+n R1 , respectively, as n → ∞. Converse proof. In general, the deterministic code capacity region is included within the random code capacity region. Namely, C(B) ⊆ C⋆(B). P ROOF A PPENDIX G OF C OROLLARY 10 First, consider the inner and outer bounds in (28) and (29). The bounds are obtained as a direct consequence of part 1 of Theorem 7 and Theorem 9. Note that the outer bound (29) holds regardless of any condition, since the deterministic code capacity region is always included within the random code capacity region, i.e. C(B) ⊆ C⋆(B) ⊆ R⋆out (B). ′ Now, suppose that the marginals VYξ1 |U,S and VYξ2 |U0 ,S are non-symmetrizable for some ξ : U × S → X and ξ ′ : U0 × S → X , and Condition T holds. Then,
based on [8, 7], both marginal (single-user) AVCs have positive capacity, i.e. C(W1 ) > 0 and C(W2 ) > 0. Namely, int C(B) 6= ∅. Hence, by Theorem 9, the deterministic code capacity region coincides with the random code capacity region, i.e. C(B) = C⋆(B). Then, the proof follows from part 2 of Theorem 7. A PPENDIX H A NALYSIS OF E XAMPLE 1 We begin with the case of an arbitrarily varying BSBC BD,0 without SI. We claim that the single user marginal AVC W1,0 without SI, corresponding to the stronger user, has zero capacity. Denote q , q(1) = 1 − q(0). Then, observe that the additive noise is distributed according to ZS ∼ Bernoulli(εq ), with ηq , (1 − q) · θ0 + q · θ1 , for 0 ≤ q ≤ 1. Based on [5], C(W1,0 ) ≤ C⋆(W1,0 ) = min0≤q≤1 [1 − h(ηq )]. Since θ0 < 12 ≤ θ1 , there exists 0 ≤ q ≤ 1 such that ηq = 12 , thus C(W1,0 ) = 0. The capacity region of the AVDBC BD,0 without SI is then given by C(BD,0 ) = {(0, 0)}. Now, consider the arbitrarily varying BSBC BD with causal SI. By Theorem 11, the random code capacity region is bounded q by R⋆in (BD ) ⊆ C⋆(BD ) ⊆ R⋆out (BD ). We show that the bounds coincide, and are thus tight. Let BD denote the random parameter DBC WY1 ,Y2 |X,S with causal SI, governed by an i.i.d. state sequence, distributed according to S ∼ Bernoulli(q). By [17], the corresponding capacity region is given by [  (R1 , R2 ) : R2 ≤ 1 − h(α ∗ β ∗ δq ) ,  q C(BD ) = , (162a) R1 ≤ h(β ∗ δq ) − h(δq ) 0≤β≤1 where δq , (1 − q) · θ0 + q · (1 − θ1 ) , (162b) T q q′ ) ⊆ C(BD ). Thus, taking q ′ = 1, we have for 0 ≤ q ≤ 1. For every given 0 ≤ q ′ ≤ 1, we have that R⋆out (BD ) = 0≤q≤1 C(BD that [  (R1 , R2 ) : R2 ≤ 1 − h(α ∗ β ∗ θ1 ) ,  ⋆ Rout (BD ) ⊆ , (163) R1 ≤ h(β ∗ θ1 ) − h(θ1 ) 1 0≤β≤ 2 where we have used the identity h(α ∗ (1 − δ)) = h(α ∗ δ). Now, to show that the region above is achievable, we examine the inner bound,   [ (R1 , R2 ) : R2 ≤ min0≤q≤1 Iq (U2 ; Y2 ) , . R⋆in (BD ) = R1 ≤ min0≤q≤1 Iq (U1 ; Y1 |U2 ) p(u1 ,u2 ),ξ(u1 ,u2 ,s) (164) 25 Consider the following choice of p(u1 , u2 ) and ξ(u1 , u2 , s). Let U1 and U2 be independent random variables,   1 U1 ∼ Bernoulli(β) , and U2 ∼ Bernoulli , 2 (165) for 0 ≤ β ≤ 21 , and let ξ(u1 , u2 , s) = u1 + u2 + s mod 2 . (166) Then, Hq (Y1 |U1 , U2 ) = Hq (S + ZS ) = h(δq ) , Hq (Y1 |U2 ) = Hq (U1 + S + ZS ) = h(β ∗ δq ) , Hq (Y2 |U2 ) = Hq (U1 + S + ZS + V ) = h(α ∗ β ∗ δq ) , Hq (Y2 ) = 1 , (167) where addition is modulo 2, and δq is given by (162b). Thus, Iq (U2 ; Y2 ) = 1 − h(α ∗ β ∗ δq ) , Iq (U1 ; Y1 |U2 ) = h(β ∗ δq ) − h(δq ) , hence R⋆in (BD ) ⊇ [ 0≤β≤ 12  (R1 , R2 ) : R2 R1 ≤ min0≤q≤1 1 − h(α ∗ β ∗ δq ) , ≤ min0≤q≤1 h(β ∗ δq ) − h(δq ) (168)  . (169) Note that θ0 ≤ δq ≤ 1 − θ1 ≤ 12 . For 0 ≤ δ ≤ 21 , the functions g1 (δ) = 1 − h(α ∗ β ∗ δ) and g2 (δ) = h(β ∗ δ) − h(δ) are monotonic decreasing functions of δ, hence the minima in (169) are both achieved with q = 1. It follows that [  (R1 , R2 ) : R2 ≤ 1 − h(α ∗ β ∗ θ1 ) ,  ⋆ ⋆ ⋆ C (BD ) = Rin (BD ) = Rout (BD ) = . (170) R1 ≤ h(β ∗ θ1 ) − h(θ1 ) 0≤β≤1 It can also be verified that Condition TD holds (see Definition 9), in agreement with part 2 of Theorem 11. First, we specify a function ξ(u1 , u2 , s) and a distributions set D⋆ that achieve R⋆in (BD ) and R⋆out (BD ) (see Definition 38). Let ξ(u1 , u2 , s) be as in (166), and let D⋆ be the set of distributions p(u1 , u2 ) such that U1 and U2 are independent random variables, distributed according to (165). By the derivation above, the requirement (38a) is satisfied. Now, by the derivation in [17, Section IV], we have that   [ (R1 , R2 ) : R2 ≤ Iq (U2 ; Y2 ) , q C(BD ) = . (171) R1 ≤ Iq (U1 ; Y1 |U2 ) p(u1 ,u2 )∈D ⋆ Then, the requirement (38b) is satisfied as well, hence ξ(u1 , u2 , s) and D⋆ achieve R⋆in (BD ) and R⋆out (BD ). It follows that Condition TD holds, as q ∗ = 1 satisfies the desired property with ξ(u1 , u2 , s) and D⋆ as described above. We move to the deterministic code capacity region of the arbitrarily varying BSBC BD with causal SI. If θ1 = 12 , the capacity region is given by C(BD ) = C⋆(BD ) = {(0, 0)}, by (170). Otherwise, θ0 < 12 < θ1 , and we now show that the ′ condition in Corollary 10 is met. Suppose that VYξ2 |U2 ,S is symmetrizable for all ξ ′ : U2 × S → X . That is, for every ξ ′ (u2 , s), there exists λu2 = J(1|u2 ) such that (1 − λub )WY2 |X,S (y2 |ξ ′ (ua , 0), 0) + λub WY2 |X,S (y2 |ξ ′ (ua , 1), 1) = (1 − λua )WY2 |X,S (y2 |ξ ′ (ub , 0), 0) + λua WY2 |X,S (y2 |ξ ′ (ub , 1), 1) (172) for all ua , ub ∈ U2 , y2 ∈ {0, 1}. If this is the case, then for ξ ′ (u2 , s) = u2 + s mod 2, taking ua = 0, ub = 1, y2 = 1, we have that (1 − λ1 ) · (α ∗ θ0 ) + λ1 · (1 − α ∗ θ1 ) = (1 − λ0 ) · (1 − α ∗ θ0 ) + λ0 · (α ∗ θ1 ) . (173) This is a contradiction. Since f (θ) = α∗θ is a monotonic increasing function of θ, and since 1−f (θ) = f (1−θ), we have that the value of the LHS of (173) is in [0, 12 ), while the value of the RHS of (173) is in ( 12 , 1]. Thus, there exists ξ ′ : U2 × S → X such ′ that VYξ2 |X,S is non-symmetrizable for θ0 < 21 < θ1 . As Condition TD holds, we have that C(BD ) = R⋆in (BD ) = R⋆out (BD ), due to Corollary 13. Hence, by (170), we have that the capacity region of the arbitrarily varying BSBC BD with causal SI is given by (45). 26 A PPENDIX I A NALYSIS OF E XAMPLE 2 A. Random Parameter BSBC with Correlated Noises Consider the random parameter BSBC B q with causal SI. By Theorem 6, the capacity region of B q with degraded message sets with causal SI is given by C(B q ) = C(B q ) (see (20)). Then, to show achievability, consider the following choice of p(u0 , u1 ) and ξ(u0 , u1 , s). Let U0 and U1 be independent random variables,   1 , and U1 ∼ Bernoulli(β) , (174) U0 ∼ Bernoulli 2 for 0 ≤ β ≤ 12 , and let ξ(u0 , u1 , s) = u0 + u1 + s mod 2 . (175) Then, Hq (Y1 |U0 , U1 ) = Hq (S + ZS ) = h(δq(1) ) , Hq (Y1 |U0 ) = Hq (U1 + S + ZS ) = h(β ∗ δq(1) ) , Hq (Y2 |U0 ) = Hq (U1 + S + NS ) = h(β ∗ δq(2) ) , Hq (Y2 ) = 1 , (176) where addition is modulo 2, and δq , δq are given by (49). Thus, (1) (2) Iq (U0 ; Y2 ) = 1 − h(β ∗ δq(2) ) , Iq (U1 ; Y1 |U0 ) = h(β ∗ δq(1) ) − h(δq(1) ) . The last inequality on the sum rate in (20) is redundant, as shown below. Since θ0 ≤ ε0 ≤ δq(1) ≤ δq(2) ≤ 12 . Hence, (177) 1 2 and 1 2 ≤ θ1 ≤ ε1 , we have that Iq (U0 ; Y2 ) = 1 − h(β ∗ δq(2) ) ≤ 1 − h(β ∗ δq(1) ) = Iq (U0 ; Y1 ) , (178) which implies that Iq (U0 ; Y2 ) + Iq (U1 ; Y1 |U0 ) ≤ Iq (U0 , U1 ; Y1 ). This completes the proof of the direct part. As for the converse, we need to show that if, R1 > h(β ∗ δq(1) ) − h(δq(1) ) , for some 0 ≤ β ≤ 1 2, (179) then it must follows that R0 ≤ 1 − h(β ∗ δq ). Indeed, by (20) and (179), (2) Hq (Y1 |U0 ) >h(β ∗ δq(1) ) − h(δq(1) ) + Hq (Y1 |U0 , U1 ) ≥h(β ∗ δq(1) ) − h(δq(1) ) + min Hq (ξ(u0 , u1 , S) + ZS ) u0 ,u1 =h(β ∗ δq(1) ) − h(δq(1) ) + min (Hq (ZS ), Hq (S + ZS ))
=h(β ∗ δq(1) ) − h(δq(1) ) + min h((1 − q)θ0 + qθ1 ), h(δq(1) ) =h(β ∗ δq(1) ) − h(δq(1) ) + h(δq(1) ) =h(β ∗ δq(1) ) . (180) Then, since δq(1) ≤ δq(2) ≤ 12 , there exists a random variable L ∼ Bernoulli(λq ), with δq(2) = δq(1) ∗ λq , (181)  for some 0 ≤ λq ≤ 12 , such that Ye2 = Y1 + L mod 2 is distributed according to Pr Ye2 = y2 | U0 = u0 , U1 = u1 = P s∈S q(s)WY2 |X,S (y2 |ξ(u0 , u1 , s), s). Thus,  (a)
(b) (c) Hq (Y2 |U0 ) =Hq (Ye2 |U0 ) ≥ h [h−1 (H(Y1 |U0 ))] ∗ λq ≥ h(β ∗ δq(1) ∗ λq ) = h(β ∗ δq(2) ) , (182) where (a) is due to Mrs. Gerber’s Lemma [20], and (b)-(c) follow from (180) and (181), respectively. B. Arbitrarily Varying BSBC with Correlated Noises 1) Without SI: We begin with the case of an arbitrarily varying BSBC B0 without SI. We claim that the single user marginal AVCs W1,0 and W2,0 without SI, corresponding to user 1 and user 2, respectively, have zero capacity. Denote q , q(1) = 1 − (1) (2) q(0). Then, observe that the additive noises are distributed according to ZS ∼ Bernoulli(ηq ) and NS ∼ Bernoulli(ηq ) , with (1) (2) (1) ⋆ ηq , (1−q)·θ0 +q·θ1 and ηq , (1−q)·ε0 +q·ε1 , for 0 ≤ q ≤ 1. Based on [5], C(W1,0 ) ≤ C (W1,0 ) = min0≤q≤1 [1−h(ηq )]. (1) Since θ0 < 12 ≤ θ1 , there exists 0 ≤ q1 ≤ 1 such that ηq1 = 12 , thus C(W1,0 ) = 0. Similarly, ε0 < 12 ≤ ε1 implies that (2) ηq2 = 21 , for some 0 ≤ q2 ≤ 1, thus C(W2,0 ) = 0 as well. The capacity region of the AVBC B0 without SI is then given by C(B0 ) = {(0, 0)}. 27 2) Causal SI – Case 1: Consider the arbitrarily varying BSBC B with causal SI, with θ0 ≤ 1 − θ1 ≤ ε0 ≤ 1 − ε1 ≤ 21 . By Theorem 7, the random code capacity region is bounded by R⋆in (B) ⊆TC⋆(B) ⊆ R⋆out (B). We show that the bounds coincide, ′ and are thus tight. By (15), (20) and (23), we have that R⋆out (B) = 0≤q≤1 C(B q ) ⊆ C(B q ), for every given 0 ≤ q ′ ≤ 1. Thus, taking q ′ = 1, we have by (48) that  [  (R1 , R2 ) : R2 ≤ 1 − h(β ∗ ε1 ) , , (183) R⋆out (B) ⊆ R1 ≤ h(β ∗ θ1 ) − h(θ1 ) 1 0≤β≤ 2 where we have used the identity h(α ∗ (1 − δ)) = h(α ∗ δ). Now, to show that the region above is achievable, we examine the inner bound,   R0 ≤ min0≤q≤1 Iq (U0 ; Y2 ) ,  (R0 , R1 ) :  [ R1 ≤ min0≤q≤1 Iq (U1 ; Y1 |U0 ) . R⋆in (B) =   R0 + R1 ≤ min0≤q≤1 Iq (U0 , U1 ; Y1 ) p(u0 ,u1 ),ξ(u0 ,u1 ,s) Consider the following choice of p(u0 , u1 ) and ξ(u0 , u1 , s). Let U0 and U1 be independent random variables,   1 U2 ∼ Bernoulli , and U1 ∼ Bernoulli(β) , 2 (184) (185) for 0 ≤ β ≤ 12 , and let ξ(u0 , u1 , s) = u0 + u1 + s mod 2 . Then, as in Subsection I-A above, this yields the following inner bound, ) ( (2) [ (R0 , R1 ) : R0 ≤ min0≤q≤1 1 − h(β ∗ δq ) , ⋆ Rin (B) ⊇ . (1) (1) R1 ≤ min0≤q≤1 h(β ∗ δq ) − h(δq ) 1 (186) (187) 0≤β≤ 2 (1) (2) Note that θ0 ≤ δq ≤ 1 − θ1 ≤ 12 and ε0 ≤ δq ≤ 1 − ε1 ≤ 21 . For 0 ≤ δ ≤ 21 , the functions g1 (δ) = 1 − h(β ∗ δ) and g2 (δ) = h(β ∗ δ) − h(δ) are monotonic decreasing functions of δ, hence the minima in (187) are both achieved with q = 1. It follows that  [  (R0 , R1 ) : R0 ≤ 1 − h(β ∗ ε1 ) , C⋆(B) = R⋆in (B) = R⋆out (B) = . (188) R1 ≤ h(β ∗ θ1 ) − h(θ1 ) 0≤β≤1 It can also be verified that Condition T holds (see Definition 5 and (24)), in agreement with part 2 of Theorem 7. First, we specify a function ξ(u0 , u1 , s) and a distribution set D⋆ that achieve R⋆in (B) and R⋆out (B) (see Definition 4). Let ξ(u0 , u1 , s) be as in (186), and let D⋆ be the set of distributions p(u0 , u1 ) such that U0 and U1 are independent random variables, distributed according to (185). By the derivation above, the first requirement in Definition 5 is satisfied with Q = P(S), and by our derivation in Subsection I-A,   R0 ≤ Iq (U0 ; Y2 ) ,   (R0 , R1 ) : [ R1 ≤ Iq (U1 ; Y1 |U0 ) . (189) C(B q ) =   R0 + R1 ≤ Iq (U0 , U1 ; Y1 ) p(u0 ,u1 )∈D ⋆ Then, the second requirement is satisfied as well, hence ξ(u0 , u1 , s) and D⋆ achieve R⋆in (B) and R⋆out (B). It follows that Condition T holds, as q ∗ = 1 satisfies the desired property with ξ(u0 , u1 , s) and D⋆ as described above. We move to the deterministic code capacity region of the arbitrarily varying BSBC B with causal SI. Consider the following cases. First, if θ1 = 12 , then ε1 = 21 as well, and the capacity region is given by C(B) = C⋆(B) = {(0, 0)}, by (188). Otherwise, θ0 < 12 < θ1 . Then, for the case where ε1 = 12 , we show that the random code capacity region, C⋆(B) = {(R0 , R1 ) : R0 = 0, R1 ≤ C ⋆ (W1 ) = 1 − h(θ1 )} can be achieved by deterministic codes as well. Based on [8, 7], it suffices to show that there exists a function ξ : U × S → X such that VYξ1 |U,S is non-symmetrizable. Indeed, assume to the contrary that θ0 < 12 < θ1 , yet VYξ1 |U,S is symmetrizable for all ξ : U × S → X . That is, for every ξ(u, s), there exists σu = J(1|u) such that (1 − σub )WY1 |X,S (y1 |ξ(ua , 0), 0) + σub WY1 |X,S (y1 |ξ(ua , 1), 1) = (1 − σua )WY1 |X,S (y1 |ξ(ub , 0), 0) + σua WY1 |X,S (y1 |ξ(ub , 1), 1) (190) for all ua , ub ∈ U, y1 ∈ {0, 1}. If this is the case, then for ξ(u, s) = u + s mod 2, taking ua = 0, ub = 1, y1 = 1, we have that (1 − σ1 )θ0 + σ1 (1 − θ1 ) = (1 − σ0 )(1 − θ0 ) + σ0 θ1 . (191) 28 This is a contradiction, since the value of the LHS of (191) is in [0, 12 ), while the value of the RHS of (191) is in ( 12 , 1]. Hence, VYξ1 |U,S is non-symmetrizable, and C(B) = C⋆(B). The last case to consider is when θ0 ≤ ε0 < 12 < ε1 ≤ θ1 . We now claim that the condition in Corollary 10 is met. Indeed, the contradiction in (191) implies that VYξ1 |U0 ,U1 ,S is non-symmetrizable with ξ(u0 , u1 , s) = u0 + u1 + s mod 2, given that ′ θ0 < 21 < θ1 . Similarly, VYξ2 |U0 ,S is non-symmetrizable with ξ ′ (u0 , s) = u0 + s mod 2, given that ε0 < 21 < ε1 . As Condition T holds, we have that C(B) = R⋆in (B) = R⋆out (B), due to Corollary 10. Hence, by (188), we have that the capacity region of the arbitrarily varying BSBC with correlated noises B with causal SI is given by (50). 3) Causal SI – Case 2: Consider the arbitrarily varying BSBC B with causal SI, with θ0 ≤ 1 − θ1 ≤ 1 − ε1 ≤ ε0 ≤ 12 . By Theorem 7, the random code capacity region is bounded by R⋆in (B) ⊆ C⋆(B) ⊆ R⋆out (B). Next, we show that the deterministic code capacity region is bounded by (51) and (52). Inner Bound. Denote [  (R0 , R1 ) : R0 R1 Ain , 0≤β≤1 ≤ 1 − h(β ∗ ε0 ) , ≤ h(β ∗ θ1 ) − h(θ1 )  . (192) We show that R⋆in (B) ⊆ Ain and R⋆in (B) ⊇ Ain , hence R⋆in (B) = Ain . As in the proof for case 1 above, consider the set of strategy distributions D⋆ and function ξ(u0 , u1 , s) as specified by (185) and (186). Then, this results in the following inner bound,   R0 ≤ min0≤q≤1 Iq (U0 ; Y2 ) ,  [  (R0 , R1 ) : R⋆in (B) ⊇ R1 ≤ min0≤q≤1 Iq (U1 ; Y1 |U0 )   R0 + R1 ≤ min0≤q≤1 Iq (U0 , U1 ; Y1 ) p∈D ⋆ ) ( (2) [ (R0 , R1 ) : R0 ≤ min0≤q≤1 1 − h(β ∗ δq ) , = (1) (1) R1 ≤ min0≤q≤1 h(β ∗ δq ) − h(δq ) 1 0≤β≤ 2 =Ain , (193) 1 2 1 2. where the last equality holds since in case 2, we assume that θ0 ≤ 1 − θ1 ≤ and 1 − ε1 ≤ ε0 ≤ Now, we upper bound R⋆in (B) by   R0 ≤ mins∈S Iq (U0 ; Y2 |S = s) ,  [  (R0 , R1 ) : R1 ≤ mins∈S Iq (X; Y1 |U0 , S = s) . R⋆in (B) ⊆   R0 + R1 ≤ mins∈S Iq (X; Y1 |S = s) p(u0 ,x) (194) We have replaced U1 with X since (U0 , U1 ) X Y1 form a Markov chain. Now, since X (Y1 , S) Y2 form a Markov chain, the third inequality in (194) is not necessary. Furthermore WY1 |X,S (y1 |x, 1) is degraded with respect to WY1 |X,S (y1 |x, 0), whereas WY2 |X,S (y2 |x, 0) is degraded with respect to WY2 |X,S (y2 |x, 1). Thus, [  (R0 , R1 ) : R0 ≤ Iq (U0 ; Y2 |S = 0) ,  . (195) R⋆in (B) ⊆ R1 ≤ Iq (X; Y1 |U0 , S = 1) p(u0 ,x) Observe that the RHS of (195) is the capacity region of a BSBC without a state, specified by Y1 = X + Z1 mod 2, Y2 = X + N0 mod 2 [4, 11]. This upper bound can thus be expressed as in the RHS of (192) (see e.g. [6, Example 15.6.5]). Hence, R⋆in (B) = Ain , which proves the equality in (51). To show that the inner bound is achievable by deterministic codes, i.e. C(B) ⊇ R⋆in (B), we consider the following cases. First, if θ1 = 12 , then ε0 = 12 as well, and R⋆in (B) = {0, 0}, by (192). Otherwise, θ0 < 12 < θ1 . In particular, for ε0 = 21 , we have that R⋆in (B) = {(R0 , R1 ) : R0 = 0, R1 ≤ 1 − h(θ1 )}. Then, as shown in the proof of case 1, there exists a function ξ : U × S → X such that VYξ1 |U,S is non-symmetrizable. Thus, based on [8, 7], the deterministic code capacity of user 1 marginal AVC is given by C(W1 ) = 1 − h(θ1 ), which implies that R⋆in (B) is achievable for ε0 = 21 . It remains to consider the case where θ0 ≤ ε0 < 21 < ε1 ≤ θ1 . By Corollary 10, in order to show that C(B) ⊇ R⋆in (B), it suffices to prove that the capacity region has non-empty interior. Following the same steps as in the proof of case 1 above, we ′ have that the channels VYξ1 |U,S and VYξ2 |U0 ,S are non-symmetrizable for ξ(u, s) = u + s mod 2 and ξ ′ (u0 , s) = u0 + s mod 2 (see (27)). Thus, based
on [8, 7], the deterministic code capacity of the marginal AVCs of user 1 and user 2 are positive, which implies that int C(B) 6= ∅, hence C(B) ⊇ R⋆in (B). 29 Outer Bound. Since the deterministic code capacity region is included within T the random code capacity region, it follows that C(B) ⊆ R⋆out (B). Based on (15), (20) and (23), we have that R⋆out (B) = 0≤q≤1 C(B q ). Thus, R⋆out (B) ⊆ C(B q=0 ) ∩ C(B q=1 )     [  (R0 , R1 ) : R0 [  (R0 , R1 ) : R0 ≤ 1 − h(β ∗ ε0 ) , ∩ = R1 ≤ h(β ∗ θ0 ) − h(θ0 ) R1 0≤β≤ 12 0≤β≤ 21    (R0 , R1 ) : R0 ≤ 1 − h(β0 ∗ ε0 ) ,     [  R0 ≤ 1 − h(β1 ∗ ε1 ) , = . R1 ≤ h(β0 ∗ θ0 ) − h(θ0 ) ,    0≤β0 ≤1 ,    R1 ≤ h(β1 ∗ θ1 ) − h(θ1 ) 0≤β1 ≤1 ≤ 1 − h(β ∗ ε1 ) , ≤ h(β ∗ θ1 ) − h(θ1 )    (196) R EFERENCES [1] R. Ahlswede. “Arbitrarily varying channels with states sequence known to the sender”. IEEE Trans. Inform. Theory 32.5 (Sept. 1986), pp. 621–629. [2] R. Ahlswede. “Elimination of correlation in random codes for arbitrarily varying channels”. Z. Wahrscheinlichkeitstheorie Verw. Gebiete 44.2 (June 1978), pp. 159–175. [3] M. Benammar, P. Piantanida, and S. Shamai. “On the compound broadcast channel: multiple description coding and interference decoding”. arXiv:1410.5187 (Oct. 2014). [4] P. Bergmans. “Random coding theorem for broadcast channels with degraded components”. IEEE Trans. Inform. Theory 19.2 (Mar. 1973), pp. 197–207. [5] D. Blackwell, L. Breiman, and A. J. Thomasian. “The capacities of certain channel classes under random coding”. Ann. Math. Statist. 31.3 (Sept. 1960), pp. 558–567. [6] T. M. Cover and J. A. Thomas. Elements of Information Theory. 2nd ed. Wiley, 2006. [7] I. Csiszár and J. Körner. Information Theory: Coding Theorems for Discrete Memoryless Systems. 2nd ed. Cambridge University Press, 2011. [8] I. Csiszár and P. Narayan. “The capacity of the arbitrarily varying channel revisited: positivity, constraints”. IEEE Trans. Inform. Theory 34.2 (Mar. 1988), pp. 181–193. [9] A. El Gamal and Y.H. Kim. Network Information Theory. Cambridge University Press, 2011. [10] T. Ericson. “Exponential error bounds for random codes in the arbitrarily varying channel”. IEEE Trans. Inform. Theory 31.1 (Jan. 1985), pp. 42–48. [11] R.G. Gallager. “Capacity and coding for degraded broadcast channels”. Probl. Inf. Transm. 10.3 (May 1974), pp. 3–14. [12] E. Hof and S. I. Bross. “On the deterministic-code capacity of the two-user discrete memoryless Arbitrarily Varying General Broadcast channel with degraded message sets”. IEEE Trans. Inform. Theory 52.11 (Nov. 2006), pp. 5023–5044. [13] J. H. Jahn. “Coding of arbitrarily varying multiuser channels”. IEEE Trans. Inform. Theory 27.2 (Mar. 1981), pp. 212–226. [14] J. Körner and K. Marton. “General broadcast channels with degraded message sets”. IEEE Trans. Inform. Theory 23.1 (Jan. 1977), pp. 60–64. [15] U. Pereg and Y. Steinberg. “The arbitrarily varying degraded broadcast channel with causal Side information at the encoder”. Proc. IEEE Int’l Symp. Inform. Theory (ISIT’2017). Aachen, Germany. [16] C. E. Shannon. “Channels with side Information at the transmitter”. IBM J. Res. Dev. 2.4 (Oct. 1958), pp. 289–293. [17] Y. Steinberg. “Coding for the degraded broadcast channel with random parameters, with causal and noncausal side information”. IEEE Trans. Inform. Theory 51.8 (July 2005), pp. 2867–2877. [18] H. Weingarten et al. “The capacity region of the degraded multiple-input multiple-output compound broadcast channel”. IEEE Trans. Inform. Theory 55.11 (Oct. 2009), pp. 5011–5023. [19] A. Winshtok and Y. Steinberg. “The arbitrarily varying degraded broadcast channel with states known at the encoder”. Proc. IEEE Int’l Symp. Inform. Theory (ISIT’2006). Seattle, Washington, July 2006, pp. 2156–2160. [20] A. D. Wyner and J. Ziv. “A theorem on the entropy of certain binary sequences and applications: Part I”. IEEE Trans. Inform. Theory 19.6 (Nov. 1973), pp. 769–772. 

The International Journal of Soft Computing and Software Engineering [JSCSE], Vol. 3, No. 3, Special Issue: The Proceeding of International Conference on Soft Computing and Software Engineering 2013 [SCSE’13], San Francisco State University, CA, U.S.A., March 2013 Doi: 10.7321/jscse.v3.n3.91 e-ISSN: 2251-7545 Modeling the Behavior of Reinforced Concrete Walls under Fire, Considering the Impact of the Span on Firewalls * Nadia Otmani-Benmehidi Meriem Arar Imene Chine Civil engineering, Laboratory: LMGE, University Badji Mokhtar Annaba, Algeria benmehidi_nadia1@yahoo.fr Civil engineering, Laboratory: LMGE, University Badji Mokhtar Annaba, Algeria maria.2a21@hotmail.fr Civil engineering, Laboratory: LMGE, University Badji Mokhtar Annaba, Algeria chineimene@yahoo.com Abstract — Numerical modeling using computers is known to a fire; which has developed fire rules. Regarding the fire behavior of bearing walls, among the authors working in this field, we mention Nadjai A [1], who performed a numerical study validated by an experimental investigation on masonry walls. Also, Cheer-Germ Go and Jun-Ren Tang [12] presented an experimental investigation. Our work presents a contribution to the study of the behavior of reinforced concrete walls, cast in place, exposed to fire, belonging to a residential building. These walls were studied under the rules of wind and earthquake, by engineers in preparation for their final project in study. The building is composed of a ground floor + 9 floors, located in the Prefecture of Annaba (Algeria) [2]. In a fire situation the temperature building rises as a function of the material combustibility and the present oxygen. The fire causes degradation in characteristics of the material, a deformation in structural elements, and cracks will appear; finally, the structure is in ruin. In order to prevent those phenomena and to minimize the spread of the disaster with controlling it as quickly as possible, we can use the firewall in buildings. In this paper we will study four concrete walls; two walls with a section 20470 cm2 reinforced with bars of Ø10 and two other walls having a section of 20350 cm 2 (reinforced with Ø12). We consider a strip of 20 cm to reduce the work. The thermal loading is defined by the standard fire ISO 834[3]. Three walls are exposed to fire on one side; the fourth wall is exposed on two of its sides. The mechanical loading (i.e. compressive load and moment) exerted on the walls in question was taken from a study conducted at cold. The thermal analysis gives the temperatures at every moment and at every point of the walls. These temperatures were used in the mechanical analysis. For the thermal analysis and the mechanical analysis we used the software Safir[4]. This software was developed by Franssen J M [4] in Belgium at the University of Liege, performed for the thermal and mechanical present several advantages compared to experimental testing. The high cost and the amount of time required to prepare and to perform a test were among the main problems on the table when the first tools for modeling structures in fire were developed. The discipline structures-in-fire modeling is still currently the subject of important research efforts around the word, those research efforts led to develop many software. In this paper, our task is oriented to the study of fire behavior and the impact of the span reinforced concrete walls with different sections belonging to a residential building braced by a system composed of porticoes and sails. Regarding the design and mechanical loading (compression forces and moments) exerted on the walls in question, we are based on the results of a study conducted at cold. We use on this subject the software Safir witch obeys to the Eurocode laws, to realize this study. It was found that loading, heating, and sizing play a capital role in the state of failed walls. Our results justify well the use of reinforced concrete walls, acting as a firewall. Their role is to limit the spread of fire from one structure to another structure nearby, since we get fire resistance reaching more than 10 hours depending on the loading considered. Keywords-fire; resistance; flame; behavior; walls I. INTRODUCTION A structure must be designed and calculated so that it remains fit for use as planned. It must resist to different degrees of reliability during execution as well as that during service. Finally, the structure must have adequate durability regarding the maintenance costs. To meet the requirements outlined above, we must: choose the materials appropriately, define a design and appropriate dimensioning. For this purpose, it is imperative to provide rules specific to each country. Various researches were performed by experts in the field of fire, to find out the behavior of the structures; as examples the separations and the bearer elements (concrete column, steel column…) of a building during 600 The International Journal of Soft Computing and Software Engineering [JSCSE], Vol. 3, No. 3, Special Issue: The Proceeding of International Conference on Soft Computing and Software Engineering 2013 [SCSE’13], San Francisco State University, CA, U.S.A., March 2013 Doi: 10.7321/jscse.v3.n3.91 e-ISSN: 2251-7545 study of structures subjected to fire, taking into account the material and geometrical nonlinearity and large displacements. Rules Relating to concrete Firewalls in Table I. These rules concern the walls with mechanical slenderness at most equal to 50 and are valid for a wall exposed to fire on one or two sides. The concrete cast in place can be used to make firewalls. Implementation of these structures must respond to rules and code of calculation which concerned them ( DTU fire concrete) [6]. A. Mechanical behavior relating to concrete: the Eurocode 2 model The division of the macroscopically measurable strains in heated concrete into individual strain components is done in the EC2 according to Eq (1)[7][14]: ε tot = ε th +ε  + ε tr + ε cr TABLE I. DEGREE FIREWALLS (1) Bearing wall Separating wall where ε th is the free thermal strain, ε  is the instantaneous stress-related strain, ε tr is the transient creep strain and ε cr is the basic creep strain. The mechanical strain is the sum of the instantaneous stressrelated strain and the transient creep strain. ε tot = ε th +ε m +ε cr Degree CF Depth (cm) Depth (cm) 1/2h 1h 1h30 2h 3h 4h 10 11 12 15 20 25 6 7 9 11 15 17.5 C. Fire walls according to Eurocode2 In section 5.4.3 of Eurocode 2[6], it is recommended that the minimum thickness for normal weight concrete, should not be less than: 200mm for unreinforced wall 140 mm for reinforced load-bearing wall 120 mm for reinforced non load bearing wall (2) where ε m is the mechanical strain. In implicit models, the stress is directly related to the mechanical strain, without calculation of the transient creep strain. In the EC2 model, the relationship at a given Temperature T between the stress and the mechanical strain is given for the ascending branch by Eq (3). D. Fire walls according to our numerical study (3) For more details we invite the lector to see [7]. B. Firewalls with elements in cellular concrete We take as an example, a firewall [5] composed of concrete columns of 45×45 cm and panels of 600×60×15 cm (Posed in front or between the columns) presents a degree firewall equal to 4 hours. We must also note that the PV CSTB n° 87-25851 dated 11 /07 /95 precise that : “an experimental wall of element with reinforced cellular concrete of 15 cm thickness with a nominal density of 450 KG / m3 mounted on flexible joints, has a degree firewall of 4 hours”. Depending on the thickness, the limit height of wall is: Wall thickness 15 cm corresponds to height: H = 17 m Wall thickness 20 cm corresponds to height: H = 22 m Wall thickness 25 cm corresponds to height: H = 28 m As a first approximation, the degree of a firewall composed of solid panels with pre-cast concrete can be deduced from simplified rules, coming from the norm P 92-701 [6] expressed (a) Length of wall superior to 3.5 m (b) The span wall must be reduced with forecast column According to results of numerical study; we recommend to add column, when length of the fire wall exceed 3.5 m, to reduce the span wall. 601 The International Journal of Soft Computing and Software Engineering [JSCSE], Vol. 3, No. 3, Special Issue: The Proceeding of International Conference on Soft Computing and Software Engineering 2013 [SCSE’13], San Francisco State University, CA, U.S.A., March 2013 Doi: 10.7321/jscse.v3.n3.91 e-ISSN: 2251-7545 20470cm2. Concerning the mechanical loading each wall is submitted to its ultimate moment (M-ult) and its ultimate compressive load (N-ult). 1 ,5 ≤ L1 ≤3,5 Kr = L1/L , → Kr=3,5/L (4) L1= Kr. L (5) K r: factor of reduction L1: reduced length of wall [m] L: initial length of wall [m] This recommendation should be added in the Eurocode. II. Figure 1. Discretization of walls The two other walls, "Mu (12) 20" and "MuCH (12) 20" have a section of 20350 cm2. They are armed with steel of diameter Ø 12 spaced with 20 cm. Mu (12) 20 is submitted to its ultimate loading (moment and load). The Wall "MuCH (12) 20" has the same mechanical loading (ultimate moment and ultimate compressive load) that Mu 20. For the thermal loading, "Mu (12) 20" and "MuCH (12) 20" are exposed to a fire ISO834 [3] on one side. The floor height (H) is equal to 2.86 m for the four walls. MODELING OF WALLS To begin the numerical study, it is necessary to model the walls considered. The Table II defines the geometrical characteristics and the loads. The reinforcement of each type of wall was calculated according to [2]. Walls: "Mu 20" and "MuF 2O" don’t have the same thermal load. The first is subjected to normalized fire ISO834 [3] on one side, for the second one, we apply the fire on two sides. They contain reinforcements Ø 10 spaced with 20 cm “Fig. 1”. They have the same section of TABLE II. GEOMETRICAL CHARACTERISTICS AND LOADING OF CONSIDERED WALLS Walls H(m) L (cm) e (cm) Ø (mm) thermal loading mechanical loading Mu 20 2,86 470 20 10 ISO834 on one side N-ult,M-ult of (Mu20) MuF 20 2,86 470 20 10 ISO834 on two sides N-ult,M-ult of (MuF20) Mu (12)20 2,86 350 20 12 ISO834 on one side N-ult,M-ult of (Mu (12)20) MuCH (12)20 2,86 350 20 12 ISO834 on one side N-ult,M-ult of (Mu 20) III. h : coefficient of convection, w/m2-K Tg : temperature of the gas, given in the data as a function of time, K Ts: temperature on the boundary of the structure, K THERMAL ANALYSIS A. Basic equation In the software Safir, the heat flux exchanged between a boundary and the hot gas in a fire compartment can be modeled according to the recommendation of Eurocode 1 with a linear convective term and radiation term, see Equation5. B. Temperatures in the wall « MuF20 » In this numerical study, the thermal analysis is a prerequisite for any result, so we start firstly by determining the temperatures at each point of the walls by using code "SAFIR". This code is based on norms [7] and [8]. We cite two cases of walls exposed to fire (MuF 20 and Mu (12) 20). In the case of the wall "MuF 20" which is exposed to fire(in red) in two faces “Fig. 2”, at failed time t = 8940sec or 149min (2.43 h), we get a temperature between 900.78 and 1048.50 ° C (6)  : Stefan-Boltzman coefficient, 5.67 10-8 ε* : relative emissivity of the material 602 The International Journal of Soft Computing and Software Engineering [JSCSE], Vol. 3, No. 3, Special Issue: The Proceeding of International Conference on Soft Computing and Software Engineering 2013 [SCSE’13], San Francisco State University, CA, U.S.A., March 2013 Doi: 10.7321/jscse.v3.n3.91 e-ISSN: 2251-7545 at the surfaces in contact with the fire. Away from both faces exposed to fire, temperature decreases to 457.60 ° C. failed time. Of course after a long period, the temperature rises considerably. IV. MECHANICAL ANALYSIS Figure 2. Temperatures of wall « MuF 20 » at failed time Units C. Temperatures in the wall« Mu (12)20 » Figure 4. Appearance of Mu (12) 20 at the failed time The mechanical data file is dependent on thermal analysis for the use of the elements temperatures in function of time. The file in question contains the dimensions of the wall (height and width), the number of nodes is equal to 21. The number of the beam element according to the discretization is taken equal to 10, each element contains 3 nodes. Mechanical loading is represented by a normal force and moment for each wall, the calculation is performed for a strip of 20 cm. Figure 4 shows the appearance of the wall Mu (12) 20 at failed time (t = 25664 sec). Figure 3. Temperatures of wall « Mu (12)20 » at failed time We note from Table III, that Mu (12) 20 has a better fire behavior compared to other walls, because of its good rigidity. MuF20 is identical to Mu 20; however MuF20 is exposed to fire at two sides which explains the good fire behavior of Mu 20 compared to the behavior of MuF 20. The results obtained from the numerical study concerning variations in temperature at the ruin time in the concrete section are in “Fig. 3”. For the failed time (25680s) or 7h 13min, observed temperature of the face exposed to fire (number 1) varies between 970,08 and 1222.20 ° C. In the side who is not exposed to fire (number 2), the temperature is 213,70 ° C at the TABLE III. FIRE RESISTANCE OF CONSIDERED WALLS 603 The International Journal of Soft Computing and Software Engineering [JSCSE], Vol. 3, No. 3, Special Issue: The Proceeding of International Conference on Soft Computing and Software Engineering 2013 [SCSE’13], San Francisco State University, CA, U.S.A., March 2013 Doi: 10.7321/jscse.v3.n3.91 e-ISSN: 2251-7545 Wall coating (cm) Height (m) Ø (mm) span Depth M-ult N-ult Rf (m) (m) (t.m) (t) (min) Mu 20 2,4 2,86 Ø 10 4,7 0,2 2,26 41,2 179,78 MuCH (12)20 3 2,86 Ø 12 3,5 0,2 2,26 41,2 278,93 Mu (12)20 3 2,86 Ø 12 3,5 0,2 0,14 26,32 427,75 MuF 20 2,4 2,86 Ø 10 4,7 0,2 2,26 41,2 148,76 A. Displacement and strain of Mu (12)20 In Figure 5, the curve represents the horizontal displacement of the wall "Mu (12) 20." There is a positive evolution (dilatation) during the exposition to fire. The maximum displacement of the node 11 (middle of the wall) at the collapse is 10 cm after a period of t = 25500sec (7h), which is representing 50% of the wall thickness. This displacement is the largest (buckling phenomenon). Given that Mu (12) 20 with section (20x 350) is exposed to fire on one side and mechanically loaded with an ultimate load of 15,040 N and an ultimate moment of 80 N.m according to [2] [9]. We can say that this wall has a good fire resistance. Figure 6. Vertical displacement of Mu (12)20 at the upper end In the case of the node 21 “Fig. 6” located at the upper extremity, the node 21 presents the maximum vertical displacement in sight of the boundary conditions. The vertical displacement is positive and equal to 1, 4 cm (there is an expansion due to thermal loading). This displacement is followed by the collapse of the wall at 25500sec (7h). B. Displacement and strain of Mu20 The mechanical analysis shows that Mu20 deforms with increasing temperature and with time. Curve of node 11“Fig. 7” represents a positive evolution throughout the time of exposition to fire. Figure5. Horizontal displacement of Mu (12)20 at half height Figure 7. Horizontal displacements of nodes 3 and 11 604 The International Journal of Soft Computing and Software Engineering [JSCSE], Vol. 3, No. 3, Special Issue: The Proceeding of International Conference on Soft Computing and Software Engineering 2013 [SCSE’13], San Francisco State University, CA, U.S.A., March 2013 Doi: 10.7321/jscse.v3.n3.91 e-ISSN: 2251-7545 The maximum displacement at the collapse is 9cm, given that Mu20 (20x470) is exposed to fire on one side and mechanically loaded with a force (17532 N) and a moment (961,7 N.m). But the node 3 has a small displacement, equal to 3cm. C. Displacement and strain of MuF 20 Figure 9. Vertical displacement of wall MuF20 at the upper end The Wall "MuF20" at its upper end (node 21), underwent dilatation (vertical displacement) of 2,5 cm after an estimated ruin time of 8900sec (148 min)“Fig. 9”. We note that this displacement is important compared to vertical displacements of previous walls, because MuF 20 is exposed to fire according two sides, its dilatation is considerable. In addition, loading has a considerable effect on the walls in case of fire. The fire acts indirectly on the structures (reinforced concrete walls), it destroys the mechanical properties of materials (concrete, steel), so that they become incapable of supporting the loads. Figure 8. Horizontal displacement of the nodes 3 and 11 “Fig.8” shows the horizontal displacements of nodes 3 and 11. MuF20 (20x470) is exposed to fire on both sides, in the case of node 11, whose curve has a greater displacement reaching 4,5cm in an estimated time of 8925sec (149 min). but node 3 has a small displacement equal to 1, 5 cm. The curves obtained in “Fig.11”; show the fire resistances of two walls exposed to fire on one side, and submitted identically to different rates of mechanical loading. These two walls does not have the same dimensions, but are subject to the same mechanical loading and to the same thermal loading (ISO834), as it was mentioned previously. Their sections are respectively, for Mu 20: 20470cm2 and for MuCH (12) 20: 20350 cm2. We note that the fire resistances of these two types of walls are considerably higher than preceding walls (Figure 10). We observe that MuCH (12) acts better than Mu 20. The section of MuCH (12)20 is lower than to that of Mu 20, thus his stress resistance (=N/S) is greater than the stress resistance of Mu 20. Otherwise the section of reinforcement (Ø12) of MuCH (12) 20 is greater than the section of reinforcement (Ø10) of Mu 20. 18000 16000 14000 12000 Load(N/20cm) 20000 N(Mu 20) N(MuF 20) 10000 8000 6000 4000 100 150 200 250 300 350 Rf(min) 400 Figure 10. Fire resistance of walls « Mu 20 » and « MuF 20 » depending on the load 605 The International Journal of Soft Computing and Software Engineering [JSCSE], Vol. 3, No. 3, Special Issue: The Proceeding of International Conference on Soft Computing and Software Engineering 2013 [SCSE’13], San Francisco State University, CA, U.S.A., March 2013 Doi: 10.7321/jscse.v3.n3.91 e-ISSN: 2251-7545 V. COMPARISON OF CONSIDERED WALLS Load(N/20cm) 20000 18000 16000 14000 In “Fig.10” the curves show the fire resistance of two walls exposed at fire; the first on two sides (grey curve) and the second (black curve) on one side, considering four rates of loading (100%, 70%, 50% and 30%). We find that the resistance of Mu 20 who was exposed on one side is larger than the resistance of MuF 20 which was exposed to fire on both sides. We also find that mechanical N(Mu 20) N(MuCH(12)20) 12000 10000 8000 The Analysis of these results shows that the increasing of the span wall causes a reduction in the fire resistance. In addition, this analysis justifies well the use of reinforced concrete walls, to limit the spread of fire from one structure to another structure nearby, since the resistances obtained are considerable (10 hours). Finally, we deduce that MuCH(12) 20 has a better fire behavior (stop fire) because it has a good fire resistance which reaches 10 hours. observed that the fire resistance of the wall with the little span is considerably higher than that fire resistance wall with the great span. We conclude that a significant span, more than 3 m is unfavorable for the firewall, since it leads to a reduction in fire resistance.  Walls studied have appreciable fire resistances, which justifies well the use of reinforced concrete walls (firewall), to limit the spread of fire from one structure to another structure nearby. Particularly “Mu (12) 20» has an appropriate size, allowing it to play the role of a firewall, because it has better fire resistance and good rigidity.  Furthermore, it would be interesting to carry out an experimental study on the walls considered to complete this work. 6000 Rf(min) 4000 100 200 300 400 500 600 700 Figure 11. Fire resistance of walls « Mu 20» and «MuCH (12)20 » depending on the load VI.       CONCLSIONS Mu 20 has a better fire behavior than MuF20 (exposed to fire at two sides) despite of their similarities “Fig.12”. The walls "Mu (12) 20 and" MuCH (12) 20 " are similar but the mechanical loading of the first is smaller than the second which gives that the resistance of Mu (12) 20 is equal to the double of the resistance of MuCH (12) 20. We note that the sizing has a significant effect on fire resistance, as well, Mu 20 with a section of 20x470 cm 2 is less resistant than Mu (12) 20 having a section : 20x350 cm2. It is recommended to forecast column, when length of the firewall exceed 3.5 m, to reduce the span wall. The fire resistances of walls considered are close to the fire resistances given by the norms [table1] [5] [6]. On the other hand the displacements of walls are in accordance with the appearance of the curves founded by (A Nadjai, 2006). We note that mechanical loading has a considerable effect on the walls in case of fire, experimental results of numerous researches of structures studied (for example, in university of Liege in BELGIUM), have previously demonstrated that the fire acts indirectly on the structures (in our case the reinforced concrete walls). The fire destroys the mechanical properties of materials (concrete, steel), so that they become unable to bearer the mechanical load. ACKNOWLEDGMENT The work presented was possible with the assistance of Professor Jean Marc Franssen and Mr. Thomas Gernay which we thank gratefully. REFERENCES [1] [2] In order to know the impact of dimensioning, more precisely of the span wall in case of fire we considered two walls (Mu 20 and MuCH (12) 20) not having the same dimensions exposed every two, to the same mechanical loading and to the same thermal loading. We 606 Nadjai A∗, O’Gara M, F. Ali and Jurgen R, Compartment Masonry Walls in Fire Situations, Belfast, 2006. Hacène chaouch A, Etude d’un bâtiment à usage d’habitation, RDC + 9 étages , 2010, [3] ENV 1991 1-2, Eurocode 1, Actions sur les structures – Partie 1-2 : Actions générales - Actions sur les structures exposées au feu, 2002. [4] Franssen J-M. SAFIR, A Thermal/Structural Program Modeling Structures under Fire, Engineering Journal, A.I.S.C., Vol 42, No. 3, 2005,143-158. The International Journal of Soft Computing and Software Engineering [JSCSE], Vol. 3, No. 3, Special Issue: The Proceeding of International Conference on Soft Computing and Software Engineering 2013 [SCSE’13], San Francisco State University, CA, U.S.A., March 2013 Doi: 10.7321/jscse.v3.n3.91 e-ISSN: 2251-7545 [5] 8.5 Les murs coupe-feu en béton, ( DTU feu béton, Eurocode 2, DAN partie 1-2, 1997)iut-tice.ujf-grenoble.fr/ticeespaces/GC/materiaux/mtx3/.../8.5.pdf [6] DTU feu béton, Eurocode 2, DAN partie 1-2, DTU23.1, la norme P 92-701 ), fascicule 62 du CCTG, 1997. [7] Eurocode 2. Calcul des structures en béton et Document d'Application Nationale Partie 1-2 : Règles générales, Calcul du comportement au feu, 2005 [8] NBN EN 1993-1-2, EUROCODE 3 : Calcul des structures en acier – Partie 1-2 : Règles générales – Calcul du comportement au feu, 2004. [9] Autodesk Robot Structural Analysis, 2009 [10] Otmani N, modélisation multi physique du comportement au feu des colonnes en acier et en béton armé, thèse de doctorat, Université d’ Annaba, Algérie, 2010. [11] Dotreppe J.C, Brüls A, Cycle de formation, Résistance au feu des constructions, application des Eurocodes dans le cadre de la formation, Fire Safety Engineering, 2000. [12] Cheer-Germ Go, Jun-Ren Tang, Jen-Hao Chi , Cheng-Tung Chen, Yue-Lin Huang, “Fire-resistance property of reinforced lightweight aggregate concrete wall” Construction and Building Materials 30 (2012) 725–733 [13] G.A. Khoury, B.N. Gringer, and P.J.E. Sullivan, “Transient thermal strain of concrete: Literature review, conditions within specimen and behaviour of individual constituents,” Magazine of Concrete Research, vol. 37, 1985, pp. 131–144. [14] Gernay T, A multiaxial constitutive model for concrete in the fire situation including transient creep and cooling down phases, thesis, University of Liege, Belgium, 2011/2012. [15] Dimia M .S, Guenfoud M, Gernay T, Franssen J-M, Collapse of concrete columns during and after the cooling phase of a fire, Journal of Fire Protection Engineering, (2011) 21(4) 245–263 TABLE IV. NOMENCLATURE Mu 20: MuF 20: Mu (12) 20: MuCH (12) 20: Firewall N-ult: M-ult: : Rf: H: h: L: e: N: S: Ø: Reinforced concrete wall with a thickness equal to 20cm and a span equal to 470cm (reinforcement :Ø10), this wall was exposed to fire on one side. Reinforced concrete wall with a thickness equal to 20cm and a span equal to 470cm (reinforcement :Ø10), this wall was exposed to fire on two sides. Reinforced concrete wall with a thickness equal to 20cm and a span equal to 350cm (reinforcement :Ø12), this wall was exposed to fire on one side. Reinforced concrete wall with a thickness equal to 20cm and a span equal to 350cm (reinforcement: Ø12), this wall was exposed to fire on one side. In this case, we use the ultimate mechanical loading of wall Mu 20. Reinforced concrete wall intended to limit the spread of fire from a structure to another nearby. Ultimate compressive load. Ultimate moment. Stress. Fire resistance. Height. Hour. Span of wall. Thickness of wall. Compressive load. Area of section. Diameter of used steel. 607 

Stochastic simulators based optimization by Gaussian process metamodels - Application to maintenance investments planning issues Short title: Metamodel-based optimization of stochastic simulators arXiv:1512.07060v2 [stat.ME] 3 May 2016 Thomas BROWNE, Bertrand IOOSS, Loı̈c LE GRATIET, Jérôme LONCHAMPT, Emmanuel REMY EDF Lab Chatou, Industrial Risk Management Department Corresponding author: Bertrand Iooss EDF R&D, 6 Quai Watier, F-78401 Chatou, France Phone: +33130877969 Email: bertrand.iooss@edf.fr Abstract This paper deals with the optimization of industrial asset management strategies, whose profitability is characterized by the Net Present Value (NPV) indicator which is assessed by a Monte Carlo simulator. The developed method consists in building a metamodel of this stochastic simulator, allowing to get, for a given model input, the NPV probability distribution without running the simulator. The present work is concentrated on the emulation of the quantile function of the stochastic simulator by interpolating well chosen basis functions and metamodeling their coefficients (using the Gaussian process metamodel). This quantile function metamodel is then used to treat a problem of strategy maintenance optimization (four systems installed on different plants), in order to optimize an NPV quantile. Using the Gaussian process framework, an adaptive design method (called QFEI) is defined by extending in our case the well known EGO algorithm. This allows to obtain an “optimal” solution using a small number of simulator runs. Keywords: Adaptive design, Asset management, Computer experiments, Gaussian process, Stochastic simulator. 1 Introduction The French company Electricité de France (EDF) looks for assessing and optimizing its strategic investments decisions for its electricity power plants by using probabilistic and optimization methods of “cost of maintenance strategies”. In order to quantify the technical and economic impact of a candidate maintenance strategy, some economic indicators are evaluated by Monte Carlo simulations using the VME software developed by EDF R&D (Lonchamp and Fessart [15]). The major output result of the 1 Monte Carlo simulation process from VME is the probability distribution of the Net Present Value (NPV) associated with the maintenance strategy. From this distribution, some indicators, such as the mean NPV, the standard deviation of NPV or the regret investment probability (P(N P V < 0)), can easily be derived. Once these indicators have been obtained, one is interested in optimizing the strategy, for instance by determining the optimal investments dates leading to the highest mean NPV and the lowest regret investment probability. Due to the discrete nature of the events to be optimized, the optimization method is actually based on genetic algorithms. However, during the optimization process, one of the main issues is the computational cost of the stochastic simulator to optimize (here VME), which leads to methods requiring a minimal number of simulator runs (Dellino and Meloni [7]). Optimization algorithms require often several thousands of simulator runs and, in some cases, are no more applicable. The solution investigated in this paper is a first attempt to break this computational cost by the way of using a metamodel instead of the simulator within the mathematical optimization algorithm. From a first set of simulator runs (called the learning sample and coming from a specific design of experiments), a metamodel consists in approximating the simulator outputs by a mathematical model (Fang et al. [8]). This metamodel can then be used to predict the simulator outputs for other input configurations, that can be served for instance for optimization issues (Forrester et al. [9]), safety assessment (de Rocquigny et al. [6]) or sensitivity analysis (Iooss and Lemaı̂tre [10]). Many metamodeling techniques are available in the computer experiments literature (Fang et al. [8]). Formally, the function G representing the computer model is defined as G: E → R (1) x 7→ G(x) where E ⊂ Rd (d ∈ N \ {0}) is the space of the input variables of the computer b from an model. Metamodeling consists in the construction of a statistical estimator G initial sample of G values corresponding to a learning set χ with χ ⊂ E and #χ < ∞ (with # the cardinal operator). In this paper, we will use the Gaussian process (also called Kriging) metamodel (Sacks et al. [23]) which has been shown relevant in many applicative studies (for example Marrel et al. [18]). Moreover, this metamodel is the base ingredient of the efficient global Optimization (EGO) algorithm (Jones et al. [11]), which allows to perform powerful optimization of CPU-time expensive deterministic computer code (Eq. (1)) as shown in Forrester et al. [9]. However, all these conventional methods are not suitable in our applicative context because of the stochastic nature of the VME simulator: the output of interest is not a single scalar variable but a full probability density function (or a cumulative distribution function, or a quantile function). The computer code G is therefore stochastic: G : E×Ω → R (x, ω) 7→ G(x, ω) (2) where Ω denotes the probability space. In the framework of robust design, robust optimization and sensitivity analysis, previous works with stochastic simulators consist 2 mainly in approximating the mean and variance of the stochastic output (Bursztyn and Steinberg [5], Kleijnen [12], Ankenman et al. [1], Marrel et al. [16], Dellino and Meloni [7]). Other specific characteristics of the distribution of the random variable G(x, ·) (as a quantile) can also be modeled to obtain, in some cases, more efficient algorithms (Picheny et al. [21]). In the following, G(x), for any x ∈ E, denotes the output random variable G(x, ·). In this paper, as first proposed by Reich et al. [22] (who used a simple Gaussian mixture model), we are interested in a metamodel which predicts the full distribution of the random variable G(x), ∀x ∈ E. We first focus on the output probability density functions of G (i.e. the densities of G(x) for any input x ∈ E) as they are a natural way to model a random distribution. Therefore we define the following deterministic function f : f : E → F (3) x 7→ fx density of the r.v. G(x) with F the family of probability density functions:  F = g ∈ L1 (R), positive, measurable, Z  g=1 . (4) R A first idea is to estimate the function f over the input set E. For a point x ∈ E, building fx with a kernel method requires a large number NMC of realizations of G(x). A recent work (Moutoussamy et al. [19]) has proposed kernel-regression based algorithms to build an estimator fb of the output densities, under the constraint that each call to f is CPU-time expensive. Their result stands as the starting point for the work presented in this paper. The next section presents the optimization industrial issues and our application case. In the third section, we re-define the functions of interest as the output quantile functions of G as they come with less constraints. We propose to use the Gaussian process metamodel and we develop an algorithm to emulate the quantile function instead of the probability density function. In the fourth section, this framework is used to develop a new quantile optimization algorithm, called Quantile Function Expected Improvement criterion and inspired from the EGO algorithm. The normality assumptions set for the metamodel imply that the function to maximize, a given level of quantile, is also the realization of a Gaussian process. In the following applicative section, this adaptive design method allows to obtain an “optimal” solution using a small number of VME simulator runs. Finally, a conclusion synthesizes the main results of this paper. 2 Maintenance investments planning issues and the VME tool 2.1 Engineering asset management Asset management processes, focused on realizing values from physical industrial assets, have been developed for years. However, for the last one or two decades, these methods 3 have been going from qualitative or semi-qualitative ones to quantitative management methods that are developed in the field of Engineering Assets Management (EAM). As a matter of fact, with budget issues becoming more and more constrained, the idea is not anymore to justify investments for the assets (maintenance, replacement, spare parts purchases . . . ) but to optimize the entire portfolio of investments made by a company or an organization. The value of an asset may be captured in its Life Cycle Cost (LCC) that monetizes all the events that may impact an asset throughout its useful life in all its dimensions (reliability, maintenance, supply chain . . . ). The investments that are made for these assets (for example preventive replacements or spare part purchases) are then valorized by the variation of the LCC created by these changes in the asset management strategy. This variation is made of positive impacts (usually avoided losses due to avoided failures or stock shortages) and negative ones (investments costs). If the cash-flows used to calculate the LCC are discounted to take into account the time value of money, the value of an investment is then equivalent to a Net Present Value (NPV) as described in Figure 1. Figure 1: Net Present Value of an investment in Engineering Asset Management. EAM is then about evaluating and optimizing the value of the asset management strategies in order to support investments decision making. 4 Input System replacement date for plant System replacement date for plant System replacement date for plant System replacement date for plant System recovering date 1 2 3 4 Name x1 x2 x3 x4 x5 min 41 41 41 41 11 max 50 50 50 50 20 Table 1: Minimal and maximal values (in years) of the five inputs used in the VME model. 2.2 VME case study As some of the cash-flows generated by the asset depend on stochastic events (failures dates depending on the probabilistic reliability), the NPV of an investment is also a stochastic variable that needs to be assessed. One of the tools developed by EDF to do so is a tool called VME that uses Monte Carlo simulation to evaluate various risk indicators needed to support decision making. The algorithm of VME consists in replicating the event model shown in Figure 2 with randomized failure dates for both a reference strategy and the strategy to be evaluated in order to get a good approximation of the NPV probabilistic distribution. Usually, the NPV takes some values in dollars or euros; for confidentiality reasons, no real cost is given in this paper and a fictive monetary unit is used. One replication of the Monte-Carlo simulation consists in creating random dates for failures (according to reliability input data) and propagating them to the occurrence of the various events in the model (maintenance task, spare part purchase or further failure). The delay between these events may be deterministic (delay between the purchase of a spare and its delivery) or probabilistic (time to failure after maintenance), but both of them are defined by input parameters. Each event of the model generates cash flows (depending on input data such as the cost of a component, the value of one day of forced outage, . . . ) that pile up to constitute the Life Cycle Cost of the assets under one asset management strategy. The result of one replication is then the NPV that is the difference of the two correlated life-cycle costs of the current strategy and the assessed new strategy. This evaluation is replicated in order to obtain a good estimation of the probabilistic distribution of the NPV. The test-case used in this paper is made of four systems installed on different plants, all these systems being identical (same reliability law) and using the same spare parts stock. The goal is to find the best replacement dates (in year) of these systems as it is wanted to prevent any failure event while replacements cannot be carried out too often. When to purchase a new spare part also needs to be taken into account to secure the availability of the plants fleet (see Lonchamp and Fessart [14] for more details). This is given by the date (in year) of recovering a new system. Therefore the whole strategy relies on these five dates which are taken as inputs for VME. These dates can take different values (only discrete and integer values), which are described in Table 1, and which have to be optimized. These five dates are the deterministic events displayed in Figure 2 in green. The random events, that make the NPV random, are the dates of failure of the plants. They 5 Figure 2: Event model used in VME. are illustrated in Figure 2 in red. For a given set of five dates, VME computes a possible NPV based on a realization of the date of failure, randomly generated, regarding the different steps of the event model. This input space of dates is denoted E and regroups the possible years for replacements and the supply: E= 4 O ! {41, 42, ..., 50} × {11, 12, ..., 20} . (5) i=1 E is therefore a discrete set (#E = 105 ). We have G : E×Ω → R (x, ω) = (x1 , x2 , x3 , x4 , x5 , ω) 7→ NPV(x, ω), (6) f : → F 7 → fx E x = (x1 , x2 , x3 , x4 , x5 ) (density of NPV(x)). Figure 3 provides examples of the output density of VME. The 10 input values inside E have been randomly chosen. It shows that there is a small variability between the curves. The optimization process consists in finding the point of E which gives the “maximal NPV” value. As NPV(x) is a random variable, we have to summarize its distribution by a deterministic operator H, for example: H(g) = E(Z) , Z ∼ g , ∀g ∈ F 6 (7) 0.06 0.04 FVME.Chi[i, ] 0.02 0.00 ●●●●●●●●●●●●●● ●●● ●●●●● ●●● ●● ●●● ●● ●●● ●● ●●●● ● ●●●●● ● ●●●●●●●● ● ●●●●●●●●●●●● ● ●●●●●●●●●●●● ● ●●●●●●●●●●●● ● ●●●●●●●●● ● ●●●●●●● ● ●●●●● ●●●● ● ●●●● ● ●●●● ● ●●●● ● ●●● ● ●●● ● ●●● ●●● ● ●●● ● ●●● ● ●●● ● ●●● ● ●●● ● ●●● ● ●●● ● ●●● ● ●●● ● ●●● ● ●●● ● ●●● ● ●●●● ● ●●●● ●● ●●●● ● ●●●●● ●● ●●●●● ● ●●●●●● ● ●●●●●●● ●● ● ●●●●●●●● ●●●●●●●●● ●●● ● ●●●●●●●●● ● ●●●●●●●●●●● ●●●● ● ● ● ● ●●●●●●●●●●●●●● ● ●● ●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●● ● ● ● ● ● ● ● ● ● ● ● ●●●●●●●●●●●●●●●●●●● ● ● ● ● ● ● ● ● ●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●●●●●●●●●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● −20 −10 0 10 20 Supp Figure 3: 10 output probability densities of the VME code (randomly sampled). In abscissa, the values of NPV are given using a fictive monetary unit. or H(g) = qZ (p) , Z ∼ g , ∀g ∈ F (8) with qZ (p) the p-quantile (0 < p < 1) of g. Our VME-optimization problem turns then to the determination of x∗ := arg max H(fx ). (9) x∈E However, several difficulties occur: • VME is a CPU-time costly simulator and the size of the set E is large. Computing (fx )x∈E , needing NMC × #E simulator calls (where NMC is worth several thousands), is therefore impossible. Our solution is to restrict the VME calls to a learning set χ ⊂ E (with #χ = 200 in our application case), randomly chosen inside E. We will then have (fx )x∈χ .  • Our metamodel, that we will denote ˆ fˆx  , cannot be computed on E due x∈E to its large size. A new set E is therefore defined, with #E = 2000. In this work, we limit our study to this restricted space (also denoted E for simplicity) instead of the full space. Other work will be performed in the future to extend our methodology to the study on the full space. 3 Gaussian process regression metamodel of a stochastic simulator 3.1 Basics on the Gaussian process regression model The Gaussian process regression model is introduced here in the framework of deterministic scalar computer codes, following closely the explanations of Marrel et al. [18]. 7 The aim of a metamodel is to build a mathematical approximation of a computer code denoted by G(x) ∈ R, x = (x1 , . . . , xd ) ∈ E ⊂ Rd from n of its evaluations. These evaluations are performed from an experimental design set denoted in this work by χ = (x(1) , . . . , x(n) ) where x(i) ∈ E, i = 1, . . . , n. The simulations on χ are denoted by yn = (y (1) , . . . , y (n) ) with y (i) = G(x(i) ), i = 1, . . . , n. Gaussian process regression models (Sacks et al. [23]), also called Kriging models (Stein [24]), suppose that the simulator response G(x) is a realization of a Gaussian random process Y of the following form: Y (x) = h(x) + Z(x), (10) where h(x) is a deterministic function and Z(x) is a centered Gaussian process. h(x) gives the mean approximation of the computer code. In this work, h(x) is considered to be a one-degree polynomial model: h(x) = β0 + d X βj xj , j=1 where β = [β0 , β1 , . . . , βd ]t is a regression parameter vector. The stochastic part Z(x) allows to model the non-linearities that h(x) does not take into account. Furthermore, the random process Z(x) is considered stationary with a covariance of the following form: Cov(Z(x), Z(u)) = σ 2 Kθ (x − u), where σ 2 is the variance of Z(x), Kθ is a correlation function and θ ∈ Rd is a vector of hyperparameters. This structure allows to provide interpolation and spatial correlation properties. Several parametric families of correlation functions can be chosen (Stein [24]). Now let us suppose that we want to predict the output of G(x) at a new point ∆ ∆ x∆ = (x∆ 1 , . . . , xd ) ∈ E. The predictive distribution for G(x ) is obtained by con∗ ditioning Y (x ) by the known values yn of Y (x) on χ (we use the notation Yn = Y (χ) = (Y (x(1) ), . . . , Y (x(n) )). Classical results on Gaussian processes imply that this distribution is Gaussian:   [Y (x∆ )|Yn = yn ] = N E[Y (x∆ )|Yn = yn ], Var[Y (x∆ )|Yn ] . (11) Finally, the metamodel is given by the mean of the predictive distribution (also called kriging mean): E[Y (x∆ )|Yn = yn ] = h(x∆ ) + k(x∆ )t Σ−1 n (Yn − h(χ)) with and k(x∆ ) = [Cov(Y (x(1) ), Y (x∆ )), . . . , Cov(Y (x(n) ), Y (x∆ ))]t = σ 2 [Kθ (x(1) , x∆ ), . . . , Kθ (x(n) , x∆ ))]t   Σn = σ 2 Kθ x(i) − x(j) 8  i,j=1...n . (12) Furthermore, the accuracy of the metamodel can be evaluated by the predictive mean squared error given by ∆ M SE(x∆ ) = Var[Y (x∆ )|Yn ] = σ 2 − k(x∆ )t Σ−1 n k(x ) (13) The conditional mean (12) is used as a predictor and its mean squared error (MSE) (13) – also called kriging variance – is a local indicator of the prediction accuracy. More generally, Gaussian process regression model defines a Gaussian predictive distribution for the output variables at any arbitrary set of new points. This property can be used for uncertainty and sensitivity analysis (see for instance Le Gratiet et al. [13]). Regression and correlation parameters β = (β0 , β1 , . . . , βd ), σ 2 and θ are generally estimated with a maximum likelihood procedure (Fang et al. [8]). The maximum likelihood estimates of β and σ 2 are given by the following closed form expressions:  β̂ = h(χ)t Σ−1 n h(χ)  σ̂ 2 = yn − h(χ)β̂ t −1 h(χ)t Σ−1 n yn ,  Σ−1 yn − h(χ)β̂ n  . n − (d + 1) However, such expression does not exist for the estimate of θ and it has to be evaluated by solving the following problem (see for instance Marrel et al. [18]): h i θ̂ = argminθ log(det(Σn )) + n log(σ̂ 2 ) . In practice, we have to rely on the estimators ĥ, β̂ and θ̂. Therefore we introduce some extra variability in the computation of the Kriging mean and the variance. For instance, this harms the normality distribution of [Y (x∆ )|Yn = yn ]. In [25] and [26], the authors offer an accurate computation of the kriging variance based on a correction term. 3.2 3.2.1 Emulation of the simulator quantile function General principles In our VME-optimization problem, we are especially interested by several quantiles (for example at the order 1%, 5%, 50%, 95%, 99%) rather than statistical moments. In Moutoussamy et al. [19] and Browne [4], quantile prediction with a density-based emulator has shown some deficiencies. Mainly, both the positivity and integral constraints make it impossible to apply the following approach to output densities. Therefore, instead of studying Eq. (6), we turn our modeling problem to G : E×Ω → R (x, ω) = (x1 , x2 , x3 , x4 , x5 , ω) 7→ NPV(x, ω), (14) Q : E x → Q 7 → Qx 9 quantile function of NPV(x) where Q is the space of increasing functions defined on ]0, 1[, with values in [a, b] (which is the support of the NPV output). For x ∈ E, a quantile function is defined by ∀p ∈]0, 1[, Qx (p) = t ∈ [a, b] Z t fx (ε)dε = p. such that (15) a For the same points as in Figure 3, Figure 4 shows the 10 quantile function outputs Q which present a rather low variability. 20 ● 10 ● 0 QVME.Chi[i, 6:96] 30 ● ● ● ● ● ● ●● ● ●● ●● ●● ●●● ●●● ●● ●●● ●●●● ●●●●●● ●●●● ●●●● ●●●● ●●●● ●●●● ●● ●●● ●●●●●●●●●●● ● ●●●●●●●●●●●●●● ●● 0.2 0.4 0.6 0.8 SuppQ[6:96] Figure 4: The 10 quantile functions Q of the VME code associated with the 10 VME pdf given in Figure 3. We consider  the learning set χ (n = #χ) and NMC × n G-simulator calls in order to N MC obtain Q̃x , the empirical quantile functions of (NPV(x))x∈χ . In this work, we x∈χ will use NMC = 104 , which is sufficiently large to obtain a precise estimator of Qx with MC . Therefore, we neglect this Monte Carlo error. In the following, we simplify the Q̃N x MC by Q . notations by replacing Q̃N x x The approach we adopt is similar to the one used in metamodeling a functional output of a deterministic simulator (Bayarri et al., [2], Marrel et. al. [17]). The first step consists in finding a low-dimensional functional basis in order to reduce the output dimension by projection, while the second step consists in emulating the coefficients of the basis functions. However, in our case, due to the nature of the functional outputs (quantile functions), some particularities will arise. 3.2.2 Projection of Qx by the Modified Magic Points (MMP) algorithm Adapted from the Magic Points algorithm (Maday et al. [20]) for probability density functions, the MMP algorithm has been proposed in Moutoussamy et al. [19]. It is a greedy algorithm that builds an interpolator (as a linear combination of basis functions) for a set of functions. Its principle is to iteratively pick a basis function in the learning sample output set and projecting the set of functions on the picked functions. Its 10 advantage over a more classical projection method (such as Fourier basis systems) is that it sequentially builds an optimal basis for this precise projection. The first step of the algorithm consists in selecting in the learning sample output set the functions which are the most correlated with the other ones. This function constitutes the first element of the functional basis. Then, at each step j ∈ {2, . . . , k} of the building procedure, the element of the learning sample output set that maximizes the L2 distance between itself and the interpolator — using the previous j − 1 basis functions — is added to the functional basis. The total number k of functions is chosen with respect to a convergence criterion. Mathematical details about this criterion are not provided in the present paper. In this paper, we apply the MMP algorithm on quantile functions. Therefore, any quantile function (Qx )x∈χ is expressed as follows: Q̂x = k X ψj (x)Rj , ∀x ∈ χ, j=1 where R = (R1 , ..., Rk ) are the quantile basis functions determined by MMP and ψ = (ψ1 , . . . , ψk ) are the coefficients obtained by the projection of Qx on R. In order to ensure the monotonic increase of Q̂x , we can restrict the coefficients to the following constrained space: n o C = ψ ∈ Rk , ψ1 , ..., ψk ≥ 0 . However, this restriction is sufficient but not necessary. That is why this constraint is not considered in Section 4. Indeed, it allows to preserve usefull properties of Gaussian process metamodels (such as any linear combinations of Gaussian process metamodels is Gaussian) for the optimization procedure. In practice, the monotonicity is verified afterwards. 3.2.3 Gaussian process metamodeling of the basis coefficients The estimation of the coefficients ψ(x) = (ψ1 (x), . . . , ψk (x)) (x ∈ E) is performed with k independent Gaussian process metamodels. According to (11), each ψj (x), j = 1, . . . , k, can be modeled by a Gaussian process of the following form:   ψj (x) ∼ N ψ̂j (x), M SEj (x) , ∀j ∈ {1, ..., k} , ∀x ∈ E (16) where ψ̂j (x) is the kriging mean (12) and M SEj (x) the kriging variance (13) obtained from the n observations ψj (χ). Finally, the following metamodel can be used for Q̂x : k X ˆ Q̂x = ψ̂j (x)Rj , ∀x ∈ E, (17) j=1 However, we have to ensure that ψ̂j ∈ C (j = 1 . . . k). The logarithmic transformation can then be used: T1 : C → Rk ψ 7→ (log(ψ1 + 1), ..., log(ψk + 1)) 11 and its inverse transformation: T2 : Rk → C φ 7→ (exp(φ1 ) − 1, ..., exp(φk ) − 1) . Then supposing that φ(x) := T1 (ψ(x)) , ∀x ∈ E, is a Gaussian process realization with k independent margins, it can be estimated by: φ̂(x) = E[φ(x) | φ(χ) = T1 (ψ(χ))] , ∀x ∈ E, with ψ(χ) the learning sample output. The following biased estimates of ψ(x) can then be considered: ψ̂(x) = T2 (φ̂(x)) , ∀x ∈ E, and (17) can be applied as our metamodel predictor of the quantile function. However, the positivity constraint is not necessary and in our application the monotonicity is respected without considering it. Therefore, these transformations have not be used in our work. 3.3 3.3.1 Numerical tests on a toy function Presentation ˆ In order to test the efficiency of the estimator Q̂, we first substitute the industrial stochastic simulator VME by the following toy function G: G(x) = sin (x1 + U1 ) + cos (x2 + U2 ) + x3 × U3 , (18) with x = (x1 , x2 , x3 ) ∈ {0.1; 0.2; ...; 1}3 , U1 ∼ N (0, 1), U2 ∼ Exp(1) and U3 ∼ U([−0.5, 0.5]), which are all independent. G(x) is a real random variable and we have #E = 103 . The goal is to estimate the output quantile functions: Q: E → Q x 7→ Qx quantile function of G(x). As G is a simple simulator and the input set E has a low dimension (d = 3), we can afford to compute NMC = 104 runs of G(x) for each input x ∈ E. Hence we can easily deduce all the output quantile functions (Qx )x∈E . We display in Figure 5 the output quantile functions (Qx ) for 10 different x ∈ E. 3.3.2 Applications to the MMP algorithm ˆ At present, we proceed at the first step for our estimator Q̂ by using the MMP algorithm. We randomly pick the learning set χ ⊂ E such that #χ = 150. As a result of the MMP algorithm, we get a basis of functions, (R1 , ..., Rk ), extracted from the output quantile functions, (Qx )x∈E , as well as the MMP-projections of the output quantile functions on χ: ∀x ∈ χ , Q̂x = k X j=1 12 ψj (x)Rj . 0 −2 −1 Quantiles 1 2 Sample of output quantile functions of G 0.0 0.2 0.4 0.6 0.8 1.0 [0;1] Figure 5: 10 ouput quantile functions of the simulator G (randomly sampled). In the example, we set k = 4. Since the metamodel of the stochastic simulator is based on the estimation of the MMP-projection Q̂, it is necessary to verify its relevance. This is why we compute the following MMP error rate: err1 = 1 X k Qx − Q̂x kL2 = 0.09%. #χ x∈χ k Qx kL2 This result shows that the MMP projection is very close to the real output quantile ˆ functions: if Q̂ is a good estimator of Q̂, then it is a good estimator of Q. It is important to recall that the basis (R1 , ..., Rk ) depends on the choice of χ, and so does Q̂. It probably approximates Q better on χ than on the whole input set E. Therefore it is natural to wonder whether the approximation Q̂ is still relevant on E. Since we could compute all the output quantile functions (Qx )x∈E , we could go further and compute 1 X k Qx − Q̂x kL2 err2 = = 0.13%. #E x∈E k Qx kL2 From this result, we infer that even though the approximation is a little less correct on E, it is still efficient. 3.3.3 Applications to the Gaussian process regression  We now have the MMP projections Q̂x  x∈χ , as well as the coefficients (ψ1 (x), ..., ψk (x))x∈χ . The Gaussian process regression in this model consists in assuming that the k coordinates (ψ1 (x), ..., ψk (x))x∈E are the realizations of k independent Gaussian processes whose values on χ are already known. From Eq. (17), we have the expression of the metamodel: k X ˆ ∀x ∈ E , Q̂x = ψ̂j (x)Rj . j=1 13 We verify the efficiency of the metamodel by computing the following error: err3 = ˆ 1 X k Qx − Q̂x kL2 = 1.42%. #E x∈E k Qx kL2 We conclude from this result, that the metamodel provides a good approximation for the output quantile functions of the simulator G as runs were performed only on χ. We display in Figure 6 the output quantile function Qx for a point x ∈ E, with its ˆ successive approximations Q̂x and Q̂x . 2−step estimation for the quantile function of G(x) 0 −1 Q_x(p) 1 2 ●● ●● ●● ●● ●● ●● ●● ●● ● ●● ●● ●● ●● ●● ●● ●● ●● ●● ● ● ●● ●● ●● ●● ●● ●● ●● ●● ●● ● ●● ●● ●● ●● ●●● ●●● ●●● ●●● ●● ●● ● ● ●● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● −2 ● ● 0.0 0.2 0.4 0.6 0.8 1.0 ]0;1[ ˆ Figure 6: For a point x ∈ E, Qx is in black points, Q̂x in red line and Q̂x in red dotted line. Figure 6 confirms that, for this point x ∈ E randomly picked, the MMP projection is really close to the initial quantile function. As for Q̂x , the error is larger but the whole distribution of G(x) is well estimated. 4 4.1 Application to an optimization problem Direct optimization on the metamodel A first solution would be to apply our quantile function metamodel with a quantilebased objective function: H : Q → R (19) q 7→ q(p) with p ∈]0, 1[. We look for: x∗ := arg max Qx (p) (20) x∈E but have only access to: ˆ x̃∗ := arg max Q̂x (p). x∈E 14 (21) ˆ We also study the relative error of H(Q̂) on E by computing: 1 err = × maxx∈E (Qx (p)) − minx∈E (Qx (p)) ! X ˆ | Qx (p) − Q̂x (p) | . (22) x∈E As an example for p = 0.5 (median estimation), with the toy-function G introduced in (18), we find ˆ maxx∈E (Qx (p)) = 0.82, maxx∈E (Q̂x (p)) = 0.42, err = 5.4%. As the estimated maximum value (0.42) has a large error with respect to the value of the exact solution (0.82), we strongly suspect that the quantile function metamodel cannot be directly applied to solve the optimization problem. In the next section, we will also see that this strategy does not work either on the VME application. We then propose an adaptive algorithm which consists in sequentially adding simulation points in order to capture interesting quantile functions to be added in our functional basis. 4.2 QFEI: an adaptive optimization algorithm ˆ After the choice of χ, E and the families (Qx )x∈χ , (Q̂x )x∈χ and (Q̂x )x∈E , our new algorithm will propose to perform new interesting (for our specific problem) calls to the VME simulator on E (outside of χ). With the Gaussian process metamodel, which provides a predictor and its uncertainty bounds, this is a classical approach used for example in black-box optimization problem (Jones et al. [11]) and rare event estimation (Bect et al. [3]). The goal is to provide some algorithms which mix global space exploration and local optimization. Our algorithm is based on the so-called EGO (Efficient Global Optimization) algorithm (Jones et al. [11]) which uses the Expected Improvement (EI) criterion to maximize a deterministic simulator. Our case is different as we want to maximize: H : E → R x 7→ Qx (p) p-quantile of NPV(x), (23) ie the p-quantile function, for p ∈]0, 1[, of the stochastic simulator. We will then propose a new algorithm called the QFEI (for Quantile Function Expected Improvement) algorithm. As previously, we use a restricted set E with #E = 5000 (E is a random sample in the full set), the initial learning set χ ⊂ E with #χ = 200 (initial design of experiment), ˆ (Qx )x∈χ , (Q̂x )x∈χ and (Q̂x )x∈E . We denote the current learning set by D (the initial learning set increased with additional points coming from QFEI). The Gaussianity on the components of ψ is needed for the EGO algorithm, that is why we do not perform the logarithmic transformation presented in Section 3.2.3. In our case, it has not implied negative consequences. 15 We apply the Gaussian process metamodeling on the k independent components P ψ1 , ..., ψk (see Equation (16)). As Q̂x (p) = kj=1 ψj (x)Rj (p), we have  ˆ Q̂x (p) ∼ N Q̂x (p), k X  Rj (p)2 M SEj (x) , ∀x ∈ E. (24) j=1 Then Q̂x (p) is a realization of the with UD Ûx σU2 |D (x) underlying Gaussian process Ux = Pk j=1 ψj (x)Rj (p) := (Ux )x∈D , := E[Ux | UD ], ∀x ∈ E, := Var[Ux | UD ], ∀x ∈ E. (25) The conditional mean and variance of Ux are directly obtained from the k Gaussian process metamodels of the ψ coefficients (16). At present, we propose to use the following improvement random function: I :  E → R x 7→ (Ux − max (UD ))+ , (26)  where x 7→ (x)+ = x.1x>0 is the positive part function. In our adaptive design, finding a new point consists in solving: xnew := arg max E[I(x)]. (27) x∈E Added points are those which have more chance to improve the current optimum. The expectation of the improvement function writes (the simple proof is given in Browne [4]): Ûx − max(UD ) σU |D (x) (28) where ϕ and φ correspond respectively to the density and distribution functions of the reduced centered Gaussian law. In practice, several iterations of this algorithm are performed, allowing to complete the experimental design D. At each iteration, a new projection functional basis is computed and the k Gaussian process metamodels are re-estimated. The stopping criterion of the QFEI algorithm can be a maximal number of iterations or a stabilization criterion on the obtained solutions. No guarantee on convergence of the algorithm can be given. In conclusion, this algorithm provides the following estimation of the optimal point x∗ : x̂∗ := arg max(UD ). (29) E[I(x)] = σU |D (x) (u(x)φ(u(x)) + ϕ(u(x))) , ∀x ∈ E, with u(x) = x∈D 4.3 Validation of the QFEI algorithm on the toy function We get back to the toy-function G introduced in (18). The goal is to determine x∗ ∈ E such that: x∗ = arg max Qx (p) x∈E 16 with p = 0.4. In other words, we try to find the input x ∈ E whose output distribution has the highest 40%-quantile. In this very case, we have: x∗ = (1, 0, 1, 0.5) with Qx∗ (p) = 0.884. We have also computed 1 X Qx (p) = −0.277, #E x∈E   Var (Qx (p))x∈E = 0.071. ˆ Let us first remember that we set an efficient metamodel Q̂ for Q in the section 3.3.3. Indeed, we had err3 = 1.42%. As before, we test the natural way to get an estimation for x∗ , by determining ˆ x̃∗ = arg maxQ̂x (p). x∈E Unfortunately, still in our previous example, we get x̃∗ = (0.9, 0, 2, 0.8) which is far from being satisfying. Besides, when we compute the real output distribution for x̃∗ , we have Qx̃∗ (p) = 0.739. ˆ Therefore only relying on Q̂ to estimate x∗ would lead to an important error. This is due to the high smoothness of the function x −→ Qx (p): even a small error in the estimator Q· (p) completely upsets the order of the family (Qx (p))x∈E . At present, we use the QFEI algorithm (Eq. (29)) in order to estimate x∗ by x̂∗ , with 20 iterations. At the end of the experiments, the design D is the learning set χ to which we have consecutively added 20 points of E. With QFEI, we finally get x̂∗ = (1, 0, 1, 0.2) with Qx̂∗ (p) = 0.878. It is important to mention that x̂∗ has the second highest output 40%-quantile. Overall we tested the whole procedure 30 times with, in each case, a new learning set χ. In the 30 trials, we made sure that the maximizer x∗ was not already in χ. In 22 cases, we obtain x̂∗ = x∗ , which is the best result that we could expect. For the remaining 8 times, we obtain x̂∗ = (1, 0, 1, 0.2). We can conclude that the QFEI algorithm is really efficient for this optimization problem. 5 Application to the VME case study We return to our VME application (cf Section 2). For the projection of Qx by the Modified Magic Points (MMP) algorithm, the choice k = 5 has shown sufficient approximation capabilities. For one example of quantile function output, a small relative 17 10 ● ● ● ● 0 −5 QVME.Chi[i, ] 5 ● ● ● ● ● ● ● ● ● ● ●● ●● ●● ●● ●● ●● ●● ● ●●● ●● ●●● ●● ●● ●● ●●● ● ● ● ●● ●● ●● ●●● ●●●● ●●●● ● ●●●●●●●●●●●● ●● 0.2 0.4 0.6 0.8 SuppQ Figure 7: For one point x ∈ χ, Q̂x (red line) and Qx (black points). QVME.E[i, ] −5 0 5 10 L2 -error (0.2%) between the observed quantile function and the projected quantile function is obtained. Figure 7 confirms also the relevance of the MMP method. We build the Gaussian process metamodel on the set E (with the choice k = 5). For one example of quantile function output, a small relative L2 -error (2.8%) between the observed quantile function and the emulated quantile function is obtained. Figure 8 confirms also the relevance of the metamodeling method. ● ● ● ● ● ● ● ● ●● ● ● ● ● ● ● ●● ● ● ● ● ●● ●● ● ●● ● ● ●● ● ●● ●●● ●● ●● ● ● ●● ●●● ●● ● ●● ●● ●●● ●● ●●●● ● ● ●●●●●●●●●●●●● 0.2 0.4 0.6 0.8 SuppQ ˆ Figure 8: For one point x ∈ χ, Q̂x (red line) and Qx (black points). As for the toy function, our optimization exercise is to determine x∗ ∈ E such that x∗ = arg max Qx (p) x∈E with p = 0.5. We first try to directly apply an optimization algorithm on the previously 18 obtained metamodel. As an example, for p = 0.5 (median estimation), we find: ˆ maxx∈E (Qx (p)) = 0.82, maxx∈E (Q̂x (p)) = 0.42, err = 5.4%. (30) ˆ If we define y = arg maxx∈E [Q̂x (p)] the best point from the metamodel, we obtain Qy (p) = 0.29 while maxx∈χ Qx (p) = 0.35. The exploration of E by our metamodel does not bring any information. We have observed the same result by repeating the experiments 100 times (changing the initial design each time). It means that the punctual errors on the quantile function metamodel are too large for this optimization algorithm. In fact, the basis functions R1 , ..., R5 that the MMP algorithm has chosen on χ are not able to represent the extreme parts of the quantile functions of E. As a conclusion of this test, the quantile function metamodel cannot be directly applied to solve the optimization problem. We now apply the QFEI algorithm. In our application case, we have performed all the simulations in order to know (Qx )x∈E , therefore the solution x∗ . Our first objective is to test our proposed algorithm for p = 0.4 which has the following solution: ( x∗ = (41, 47, 48, 45, 18) ∗ Qx (p) = −1.72. (31) We have also computed 1 X Qx (p) = −3.15, #E x∈E
Var (Qx (p))x∈E = 0.59. (32) We start with D := χ and we obtain max (Qx ) = −1.95. x∈χ (33) After 50 iterations of the QFEI algorithm, we obtain: ( x̂∗ = (41, 47, 45, 46, 19) Qx̂∗ (p) = −1.74. (34) We observe that x̂∗ and Qx̂∗ (p) are close to x∗ and ' Qx∗ (p). This is a first confirmation of the relevance of our method. With respect to the initial design solution, the QFEI has allowed to obtain a strong improvement of the proposed solution. 50 repetitions of this experiment (changing the initial design) has also proved the robustness of QFEI. The obtained solution is always one of the five best points on E. QFEI algorithm seems promising but a lot of tests remain to perform and will be pursued in future works: changing p (in particular testing extremal cases), increasing the size of E, increasing the dimension d of the inputs, . . . 6 Conclusion In this paper, we have proposed to build a metamodel of a stochastic simulator using the following key points: 19 1. Emulation of the quantile function which proves better efficiency for our problem than the emulation of the probability density function; 2. Decomposition of the quantile function in a sum of the quantile functions coming from the learning sample outputs; 3. Selection of the most representative quantile functions of this decomposition using an adaptive choice algorithm (called the MMP algorithm) in order to have a small number of terms in the decomposition; 4. Emulation of each coefficient of this decomposition by a Gaussian process metamodel, by taking into account constraints ensuring that a quantile function is built. The metamodel is then used to treat a simple optimization strategy maintenance problem using a stochastic simulator (VME), in order to optimize an output (NPV) quantile. Using the Gaussian process metamodel framework and extending the EI criterion to quantile function, the adaptive QFEI algorithm has been proposed. In our example, it allows to obtain an “optimal” solution using a small number of VME simulator runs. This work is just a first attempt and needs to be continued in several directions: • Consideration of a variable NMC whose decrease could help to fight against the computational cost of the stochastic simulator, • Improvement of the initial learning sample choice by replacing the random sample by a space filling design (Fang et al. [8]), • Algorithmic improvements to counter the cost of the metamodel evaluations and to increase the size of the study set E, • Multi-objective optimization (several quantiles to be optimized) in order to take advantage of our powerful quantile function emulator, • Including the estimation error induced in practice by ĥ, β̂ and θ̂ to define a more precise version of the QFEI algorithm, • Application to more complex real cases, • Consideration of a robust optimization problem where environmental input variables of the simulator has not to be optimized but just create an additional uncertainty on the output. References [1] B. Ankenman, B.L. Nelson, and J. Staum. Stochastic kriging for simulation metamodeling. Operations Research, 58:371–382, 2010. [2] M.J. Bayarri, J.O. Berger, J. Cafeo, G. Garcia-Donato, F. Liu, J. Palomo, R.J. Parthasarathy, R. Paulo, J. Sacks, and D. Walsh. Computer model validation with functional output. The Annals of Statistics, 35:1874–1906, 2007. 20 [3] J. Bect, D. Ginsbourger, L. Li, V. Picheny, and E. Vazquez. Sequential design of computer experiments for the estimation of a probability of failure. Statistics and Computing, 22:773–793, 2012. [4] T. Browne. Développement d’émulateurs de codes stochastiques pour la résolution de problèmes d’optimisation robuste par algorithmes génétiques. Rapport de stage de Master M2 de l’Université Paris Descartes, EDF R&D, Chatou, France, 2014. [5] D. Bursztyn and D.M. Steinberg. Screening experiments for dispersion effects. In A. Dean and S. Lewis, editors, Screening - Methods for experimentation in industry, drug discovery and genetics. Springer, 2006. [6] E. de Rocquigny, N. Devictor, and S. Tarantola, editors. Uncertainty in industrial practice. Wiley, 2008. [7] G. Dellino and C. Meloni, editors. Uncertainty management in simulationoptimization of complex systems. Springer, 2015. [8] K-T. Fang, R. Li, and A. Sudjianto. Design and modeling for computer experiments. Chapman & Hall/CRC, 2006. [9] A. Forrester, A. Sobester, and A. Keane, editors. Engineering design via surrogate modelling: a practical guide. Wiley, 2008. [10] B. Iooss and P. Lemaı̂tre. A review on global sensitivity analysis methods. In C. Meloni and G. Dellino, editors, Uncertainty management in SimulationOptimization of Complex Systems: Algorithms and Applications. Springer, 2015. [11] D.R. Jones, M. Schonlau, and W.J. Welch. Efficient global optimization of expensive black-box functions. Journal of Global Optimization, 13:455–492, 1998. [12] J.P.C. Kleijnen. Design and analysis of simulation experiments, Second Edition. Springer, 2015. [13] L. Le Gratiet, C. Cannamela, and B. Iooss. A Bayesian approach for global sensitivity analysis of (multifidelity) computer codes. SIAM/ASA Journal of Uncertainty Quantification, 2:336–363, 2014. [14] J. Lonchampt and K. Fessart. An optimization approach for life cycle management applied to large power transformers. EPRI Report, 1023033, 2012. [15] J. Lonchampt and K. Fessart. Investments portfolio optimal planning, a tool for an intergrated life management. ANS embedded meeting on Risk Management for Complex Socio-technical Systems (RM4CSS), Washington DC, USA, 2013. [16] A. Marrel, B. Iooss, S. Da Veiga, and M. Ribatet. Global sensitivity analysis of stochastic computer models with joint metamodels. Statistics and Computing, 22:833–847, 2012. [17] A. Marrel, B. Iooss, M. Jullien, B. Laurent, and E. Volkova. Global sensitivity analysis for models with spatially dependent outputs. Environmetrics, 22:383–397, 2011. [18] A. Marrel, B. Iooss, F. Van Dorpe, and E. Volkova. An efficient methodology for modeling complex computer codes with Gaussian processes. Computational Statistics and Data Analysis, 52:4731–4744, 2008. 21 [19] V. Moutoussamy, S. Nanty, and B. Pauwels. Emulators for stochastic simulation codes. ESAIM: PROCEEDINGS AND SURVEYS, 48:116–155, 2015. [20] Y. Maday N.C. Nguyen, A.T. Patera, and G.S.H. Pau. A general, multipurpose interpolation procedure: the magic points. In Proceedings of the 2nd International Conference on Scientific Computing and Partial Differential Equations, 2007. [21] V. Picheny, D. Ginsbourger, Y. Richet, and G. Caplin. Quantile-based optimization of noisy computer experiments with tunable precision. Technometrics, 55:2–13, 2013. [22] B.J. Reich, E. Kalendra, C.B. Storlie, H.D. Bondell, and M. Fuentes. Variable selection for high dimensional Bayesian density estimation: Application to human exposure simulation. Journal of the Royal Statistical Society: Series C (Applied Statistics), 61:47–66, 2012. [23] J. Sacks, W.J. Welch, T.J. Mitchell, and H.P. Wynn. Design and analysis of computer experiments. Statistical Science, 4:409–435, 1989. [24] M.L. Stein. Interpolation of spatial data. Springer, 1999. [25] E. Zhu and M. L. Stein. Spatial sampling design for prediction with estimated parameters. Journal of Agricultural, Biological, and Environmental Statistics, 11:24–44, 2006. [26] D. L. Zimmerman. Optimal network design for spatial prediction, covari- ance parameter estimation, and empirical prediction. Environmetrics, 17:635–652, 2006. 22 

Updated version of the paper published at the ICDL-Epirob 2017 conference (Lisbon, Portugal), with the same title and authors. Embodied Artificial Intelligence through Distributed Adaptive Control: An Integrated Framework Clément Moulin-Frier SPECS Lab Universitat Pompeu Fabra Barcelona, Spain Email: clement.moulinfrier@gmail.com Martì Sanchez-Fibla SPECS Lab Universitat Pompeu Fabra Barcelona, Spain Email: santmarti@gmail.com Jordi-Ysard Puigbò SPECS Lab Universitat Pompeu Fabra Barcelona, Spain Email: jordiysard.puigbo@upf.edu Xerxes D. Arsiwalla SPECS Lab Universitat Pompeu Fabra Barcelona, Spain Email: x.d.arsiwalla@gmail.com Paul FMJ Verschure SPECS Lab Universitat Pompeu Fabra ICREA Institute for Bioengineering of Catalonia (IBEC) Barcelona Institute of Science and Technology (BIST) Barcelona, Spain Email: paul.verschure@upf.edu Abstract—In this paper, we argue that the future of Artificial Intelligence research resides in two keywords: integration and embodiment. We support this claim by analyzing the recent advances in the field. Regarding integration, we note that the most impactful recent contributions have been made possible through the integration of recent Machine Learning methods (based in particular on Deep Learning and Recurrent Neural Networks) with more traditional ones (e.g. Monte-Carlo tree search, goal babbling exploration or addressable memory systems). Regarding embodiment, we note that the traditional benchmark tasks (e.g. visual classification or board games) are becoming obsolete as state-of-the-art learning algorithms approach or even surpass human performance in most of them, having recently encouraged the development of first-person 3D game platforms embedding realistic physics. Building on this analysis, we first propose an embodied cognitive architecture integrating heterogeneous subfields of Artificial Intelligence into a unified framework. We demonstrate the utility of our approach by showing how major contributions of the field can be expressed within the proposed framework. We then claim that benchmarking environments need to reproduce ecologically-valid conditions for bootstrapping the acquisition of increasingly complex cognitive skills through the concept of a cognitive arms race between embodied agents. Index Terms—Cognitive Architectures, Embodied Artificial Intelligence, Evolutionary Arms Race, Unified Theories of Cognition. I. INTRODUCTION In recent years, research in Artificial Intelligence has been primarily dominated by impressive advances in Machine Learning, with a strong emphasis on the so-called Deep Learning framework. It has allowed considerable achievements such as human-level performance in visual classification [1] and description [2], in Atari video games [3] and even in the highly complex game of Go [4]. The Deep Learning approach is characterized by supposing very minimal prior on the task to be solved, compensating this lack of prior knowledge by feeding the learning algorithm with an extremely high amount of training data, while hiding the intermediary representations. However, it is important noting that the most important contributions of Deep Learning for Artificial Intelligence often owe their success in part to their integration with other types of learning algorithms. For example, the AlphaGo program which defeated the world champions in the famously complex game of Go [4], is based on the integration of Deep Reinforcement Learning with a Monte-Carlo tree search algorithm. Without the tree search addition, AlphaGo still outperforms previous machine performances but is unable to beat high-level human players. Another example can be found in the original Deep Q-Learning algorithm (DQN, Mnih et al., 2015), achieving very poor performance in some Atari games where the reward is considerably sparse and delayed (e.g. Montezuma Revenge). Solving such tasks has required the Updated version of the paper published at the ICDL-Epirob 2017 conference (Lisbon, Portugal), with the same title and authors. integration of DQN with intrinsically motivated learning algorithms for novelty detection [5], or goal babbling [6]. such as the prisoner dilemma [13] and studying the emergence of cooperation and competition among agents [14]. A drastically different approach has also received considerable attention, arguing that deep learning systems are not able to solve key aspects of human cognition [7]. The approach states that human cognition relies on building causal models of the world through combinatorial processes to rapidly acquire knowledge and generalize it to new tasks and situations. This has led to important contributions through model-based Bayesian learning algorithms, which surpass deep learning approaches in visual classification tasks while displaying powerful generalization abilities in one-shot training [8]. This solution, however, comes at a cost: the underlying algorithm requires a priori knowledge about the primitives to learn from and about how to compose them to build increasingly abstract categories. An assumption of such models is that learning should be grounded in intuitive theories of physics and psychology, supporting and enriching acquired knowledge [7], as supported by infant behavioral data [9]. The above examples emphasize two important challenges in modern Artificial Intelligence. Firstly, there is a need for a unified integrative framework providing a principled methodology for organizing the interactions of various subfields (e.g. planning and decision making, abstraction, classification, reinforcement learning, sensorimotor control or exploration). Secondly, Artificial Intelligence is arriving at a level of maturation where more realistic benchmarking environments are required, for two reasons: validating the full potential of the state-of-the-art artificial cognitive systems, as well as understanding the role of environmental complexity in the shaping of cognitive complexity. Considering the pre-existence of intuitive physics and psychology engines as an inductive bias for Machine Learning is far from being a trivial assumption. It immediately raises the question: where does such knowledge come from and how is it shaped through evolutionary, developmental and cultural processes? All the aforementioned approaches are lacking this fundamental component shaping intelligence in the biological world, namely embodiment. Playing Atari video games, complex board games or classifying visual images at a human level are considerable milestones of Artificial Intelligence research. Yet, in contrast, biological cognitive systems are intrinsically shaped by their physical nature. They are embodied within a dynamical environment and strongly coupled with other physical and cognitive systems through complex feedback loops operating at different scales: physical, sensorimotor, cognitive, social, cultural and evolutionary. Nevertheless, many recent Artificial Intelligence benchmarks have focused on solving video games or board games, adopting a third-person view and relying on a discrete set of actions with no or poor environmental dynamics. A few interesting software tools have however recently been released to provide more realistic benchmarking environments. This for example, is the case of Project Malmo [10] which provides an API to control characters in the MineCraft video game, an open-ended environment with complex physical and environmental dynamics; or Deepmind Lab [11], allowing the creation of rich 3D environments with similar features. Another example is OpenAI Gym [12], providing access to a variety of simulation environments for the benchmarking of learning algorithms, especially reinforcement learning based. Such complex environments are becoming necessary to validate the full potential of modern Artificial Intelligence research, in an era where human performance is being achieved on an increasing number of traditional benchmarks. There is also a renewed interest for multi-agent benchmarks in light of the recent advances in the field, solving social tasks In this paper, we first propose an embodied cognitive architecture structuring the main subfields of Artificial Intelligence research into an integrated framework. We demonstrate the utility of our approach by showing how major contributions of the field can be expressed within the proposed framework, providing a powerful tool for their conceptual description and comparison. Then we argue that the complexity of a cognitive agent strongly depends on the complexity of the environment it lives in. We propose the concept of a cognitive arms race, where an ecology of embodied cognitive agents interact in a dynamic environment reproducing ecologically-valid conditions and driving them to acquire increasingly complex cognitive abilities in a positive feedback loop. II. AN INTEGRATED COGNITIVE ARCHITECTURE FOR EMBODIED ARTIFICIAL INTELLIGENCE Considering an integrative and embodied approach to Artificial Intelligence requires dealing with heterogeneous aspects of cognition, where low-level interaction with the environment interacts bidirectionally with high-level reasoning abilities. This reflects a historical challenge in formalizing how cognitive functions arise in an individual agent from the interaction of interconnected information processing modules structured in a cognitive architecture [15], [16]. On one hand, top-down approaches mostly rely on methods from Symbolic Artificial Intelligence (from the General Problem Solver [17] to Soar [18] or ACT-R [19] and their follow-up), where a complex representation of a task is recursively decomposed into simpler elements. On the other hand, bottom-up approaches instead emphasize lower-level sensory-motor control loops as a starting point of behavioral complexity, which can be further extended by combining multiple control loops together, as implemented in behavior-based robotics [20] (sometimes referred as intelligence without representation [21]). These two approaches thus reflect different aspects of cognition: high-level symbolic reasoning for the former and low-level embodied behaviors for the latter. However, both aspects are of equal importance when it comes to defining a unified theory of cognition. It is therefore a major challenge of cognitive science to unify both approaches into a single theory, where (a) reactive control allows an initial level Updated version of the paper published at the ICDL-Epirob 2017 conference (Lisbon, Portugal), with the same title and authors. of complexity in the interaction between an embodied agent and its environment and (b) this interaction provides the basis for learning higher-level representations and for sequencing them in a causal way for top-down goal-oriented control. For this aim, we adopt the principles of the Distributed Adaptive Control (DAC) theory of the mind and brain [22], [23]. Besides its biological grounding, DAC is an adequate modeling framework for integrating heterogeneous concepts of Artificial Intelligence and Machine Learning into a coherent cognitive architecture, for two reasons: (a) it integrates the principles of both the aforementioned bottom-up and top-down approaches into a coherent information processing circuit; (b) it is agnostic to the actual implementation of each of its functional modules. Over the last fifteen years, DAC has been applied to a variety of complex and embodied benchmark tasks, for example foraging [22], [24] or social humanoid robot control [16], [25]. A. The DAC-EAI cognitive architecture: Distributed Adaptive Control for Embodied Artificial Intelligence DAC posits that cognition is based on the interaction of interconnected control loops operating at different levels of abstraction (Figure 1). The functional modules constituting the architecture are usually described in biological or psychological terms (see e.g. [26]). Here we propose instead to describe them in purely computational term, with the aim of facilitating the description of existing Artificial Intelligence systems within this unified framework. We call this instantiation of the architecture DAC-EAI: Distributive Adaptive Control for Embodied Artificial Intelligence. Figure 1: The DAC-EAI architecture allows a coherent organization of heterogeneous subfields of Artificial Intelligence. DAC-EAI stands for Distributed Adaptive Control for Embodied Machine Learning. It is composed of three layers operating in parallel and at different levels of abstraction. See text for detail, where each module name is referred with italics. The first level, called the Somatic layer, corresponds to the embodiment of the agent within its environment. It includes the sensors and actuators, as well internal variables to be regulated (e.g. energy or safety levels). The self-regulation of these internal variables occurs in the Reactive layer and extends the aforementioned behavior-based approaches (e.g. the Subsumption architecture [20]) with drive reduction mechanisms through predefined sensorimotor control loops (i.e. reflexes). In Figure 1, this corresponds to the mapping from the Sensing to the Motor Control module through Self Regulation. The Reactive layer offers several advantages when analyzed from the embodied artificial intelligence perspective of this paper. First, reward is traditionally considered in Machine Learning as a scalar value associated with external states of the environment. DAC proposes instead that it should derive from the internal dynamics of multiple internal variables modulated by the body-environment real-time interaction, providing an embodied notion of reward in cognitive agents. Second, the Reactive layer generates a first level of behavioral complexity through the interaction of predefined sensorimotor control loops for self-regulation. This provides a notion of embodied inductive bias bootstrapping and structuring learning processes in the upper levels of the architecture. This is a departure from the model-based approaches mentioned in the introduction [7], where inductive biases are instead considered as intuitive core knowledge in the form of a pre-existent physics and psychology engine. Behavior generated in the Reactive layer bootstraps learning processes for acquiring a state space of the agent-environment interaction in the Adaptive layer. The Representation Learning module receives input from Sensing to form increasingly abstract representations. For example, unsupervised learning methods such as deep autoencoders [27] could be a possible implementation of this module. The resulting abstract states of the world are mapped to their associated values through the Value Prediction module, informed by the internal states of the agent from Self Regulation. This allows the inference of action policies maximizing value through Action Selection, a typical reinforcement learning problem [28]. We note that Deep QLearning [3] provides an integrated solution to the three processes involved in the Adaptive layer, based on Deep Convolutional Networks for Representation Learning, Q-value estimation for Value Prediction and an ε-greedy policy for Action Selection. However, within our proposed framework, the self-regulation of multiple internal variables in the Reactive layer requires the agent to switch between different action policies (differentiating e.g. between situation of low energy vs. low safety). A possible way to achieve this using the Deep Q-Learning framework is to extend it to multi-task learning (see e.g. [29]). Since it is likely that similar abstract features are relevant to various tasks, a promising solution is to share the representation learning part of the network (the convolutional layers in [3]) across tasks, while multiplying the fully-connected layers in a task-specific way. The state space acquired in the Adaptive layer then supports the acquisition of higher-level cognitive abilities such as goal selection, memory and planning in the Contextual layer. The abstract representations acquired in Representation Learning Updated version of the paper published at the ICDL-Epirob 2017 conference (Lisbon, Portugal), with the same title and authors. are linked together through Relational Learning. The availability of abstract representations in possibly multiple modalities provides the substrate for causal and compositional linking. Several state-of-the-art methods are of interest for learning such relations, such as Bayesian program learning [8] or Long Short Term Memory neural network (LSTM, [30]). Based on these higher-level representations, Goal Selection forms the basis of goal-oriented behavior by selecting valuable states to be reached, where value is provided by the Value Prediction module. Intrinsically-motivated methods maximizing learning progress can be applied here for an efficient exploration of the environment [31]. The selected goals are reached through Planning, where any adaptive method of this field can be applied [32]. The resulting action plans, learned from action-state-value tuples generated by the Adaptive layer, propagate down the architecture to modulate behavior. Finally, an addressable memory system registers the activity of the Contextual layer, allowing the persistence of the agent experience over the long term for lifelong learning abilities [33]. In psychological terms, this memory system is analog to an autobiographical memory. These high-level cognitive processes, in turn, modulate behavior at lower levels via top-down pathways shaped by behavioral feedback. The control flow is therefore distributed within the architecture, both from bottom-up and top-down interactions between layers, as well as from lateral information processing into the subsequent layers. B. Expressing existing Machine Learning systems within the DAC-EAI framework We now demonstrate the generality of the proposed DACEML architecture by describing how well-known Artificial Intelligence systems can be conceptually described as subparts of the DAC-EAI architecture (Figure 2). We start with behavior-based robotics [20], implementing a set of reactive controllers through low-level coupling between sensors to effectors. Within the proposed framework, there are described as the lower part of the architecture, spanning the Somatic and Reactive layers (Figure 2B). However, those approaches are not considering the self-regulation of internal variables but instead of exteroceptive variables, such as light quantity for example. In contrast, top-down robotic planning algorithms [34] correspond to the right column (Action) of the DAC-EAI architecture: spanning from Planning to Action Selection and Motor Control, where the current state of the system is typically provided by pre-processed sensory-related information along the Reactive or Adaptive layers (Figure 2C). More recent Deep Reinforcement Learning methods, such as the original Deep Q-Learning algorithm (DQN, [3]) typically span over all the Adaptive layer, They use convolutional deep networks learning abstract representation from pixel-level sensing of video game frames, Q-learning for predicting the cumulated value of the resulting states and competition among discrete actions as an action selection process (Figure 2D). Still, there is no real motor control in this system, given that most available benchmarks operate on a limited set of discrete (up-down-left-right) or continuous (forward speed, rotation speed) actions. Not shown in Figure 2, classical reinforcement learning [28] relies on the same architecture as Figure 2D, however not addressing the representation learning problem, since the state space is usually pre-defined in these studies (often considering a grid world). Several extensions based on the DQN algorithm exist. For example, intrinsically-motivated deep reinforcement learning [6] extends it with a goal selection mechanism (Figure 2E). This extension allows solving tasks with delayed and sparse reward (e.g. Montezuma Revenge) by encouraging exploratory behaviors. AlphaGo also relies on a Deep Reinforcement Learning method (hence spanning the Adaptive layer as in the last examples), coupled with a Monte-Carlo tree search algorithm which can be conceived as a planning process (see also [35]), as represented in Figure 2F. Another recent work, adopting a drastically opposite approach as compared to end-to-end deep learning, addresses the problem of learning highly abstract concepts from the perspective of the human ability to perform one-shot learning. The resulting model, called Bayesian Program Learning [8], relies on a priori knowledge about the primitives to learn from and about how to compose them to build increasingly abstract categories. In this sense, it is described within the DAC-EAI framework as addressing the pattern recognition problem from the perspective of relational learning, where primitives are causally linked for composing increasingly abstract categories (Figure 2G). Finally, the Differentiable Neural Computer [36], the successor of the Neural Turing Machine [37], couples a neural controller (e.g. based on an LSTM) with a content-addressable memory. The whole system is fully differentiable and is consequently optimizable through gradient descent. It can solve problems requiring some levels of sequential reasoning such has path planning in a subway network or performing inferences in a family tree. In DAC-EAI terms, we describe it as an implementation of the higher part of the architecture, where causal relations are learned from experience and selectively stored in an addressable memory, which can further by accessed for reasoning or planning operations (Figure 2H). An interesting challenge with such an integrative approach is therefore to express a wide range of Artificial systems within a unified framework, facilitating their description and comparison in conceptual terms. Updated version of the paper published at the ICDL-Epirob 2017 conference (Lisbon, Portugal), with the same title and authors. Figure 2: The DAC-EAI architecture allows a conceptual description of many Artificial Intelligence systems within a unified framework. A) The complete DAC-EAI architecture (see Figure 1 for a larger version). The other subfigures (B to E) show conceptual descriptions of different Artificial Intelligence systems within the DAC-EAI framework. B): Behavior-based Robotics [20]. C) Top-down robotic planning [34]. D) Deep Q-Learning [3]. E) Intrinsically-Motivated Deep Reinforcement Learning [6]. F) AlphaGo [4]. G) Bayesian Program Learning [8]. H) Differentiable Neural Computer [36]. III. THE COGNITIVE ARMS RACE: REPRODUCING ECOLOGICALLY-VALID CONDITIONS FOR DEVELOPING COGNITIVE COMPLEXITY A general-purpose cognitive architecture for Artificial Intelligence, as the one proposed in the previous section, tackles the challenge of general-purpose intelligence with the aim of addressing any kind of task. Traditional benchmarks, mostly based on datasets or on idealized reinforcement learning tasks, are progressively becoming obsolete in this respect. There are two reasons for this. The first one is that state-of-the-art learning algorithms are now achieving human performance in an increasing number of these traditional benchmarks (e.g. visual classification, video or board games). The second reason is that the development of complex cognitive systems is likely to depend on the complexity of the environment they evolve in1. For these two reasons, Machine Learning benchmarks have recently evolved toward first- 1 See also https://deepmind.com/blog/open-sourcing-deepmind-lab/: “It is possible that a large fraction of animal and human intelligence is a direct consequence of the richness of our environment, and unlikely to arise without it”. person 3D game platforms embedding realistic physics [10], [11] and likely to become the new standards in the field. It is therefore fundamental to figure out what properties of the environment act as driving forces for the development of complex cognitive abilities in embodied agents. We propose in this paper the concept of a cognitive arms race as a fundamental driving force catalyzing the development of cognitive complexity. The aim is to reproduce ecologicallyvalid conditions among embodied agents forcing them to continuously improve their cognitive abilities in a dynamic multi-agent environment. In natural science, the concept of an evolutionary arms race has been defined as follows: “an adaptation in one lineage (e.g. predators) may change the selection pressure on another lineage (e.g. prey), giving rise to a counter-adaptation” [38]. This process produces the conditions of a positive feedback loop where one lineage pushes the other to better adapt and vice versa. We propose that such a positive feedback loop is a key driving force for achieving an important step towards the development of machine general intelligence. A first step for achieving this objective is the computational modeling of two populations of embodied cognitive agents, preys and predators, each agent being driven by the cognitive architecture proposed in the previous section. Basic survival behaviors are implemented as sensorimotor control loops operating in the Reactive layer, where predators hunt preys, while preys escape predators and are attracted to other food sources. Since these agents adapt to environmental constraints through learning processes occurring in the upper levels of the architecture, they will reciprocally adapt to each other. A cognitive adaptation (in term of learning) of members of one population will perturb the equilibrium attained by the others for self-regulating their own internal variables, forcing them to re-adapt in consequence. This will provide an adequate setup for studying the conditions of entering in a cognitive arms race between populations, where both reciprocally improve their cognitive abilities against each other. A number of previous works have tackled the challenge of solving social dilemmas in multi-agent simulations (see e.g. [13] for a recent attempt using Deep Reinforcement Learning). Within these works, the modeling of wolf-pack hunting behavior ([13], [39], [40]) is of particular interest as a starting point for bootstrapping a cognitive arms race. Such behaviors are based both on competition between the prey and the wolf group, as well as cooperation between wolves to maximize hunting efficiency. This provides a complex structure of codependencies among the considered agents where adaptations of one’s behavior will have consequences on the equilibrium of the entire system. Such complex systems have usually been studied in the context of Evolutionary Robotics [41] where coadaptation is driven by a simulated Darwinian selection process. However complex co-adaptation can also be studied through coupled learning among agents endowed with the cognitive architecture presented in the previous section. It is interesting to note that there exist precursors of this concept of an arms race in the recent literature under a quite different angle. An interesting example is a Generative Adversarial Network [42], where a pattern generator and a pattern discriminator compete and adapt against each other. Another example is the AlphaGo program [4] which was partly trained by playing games against itself, consequently improving its performance in an iterative way. Both these systems owe their success in part to their ability to enter into a positive feedback loop of performance improvement. IV. CONCLUSION Building upon recent advances in Artificial Intelligence and Machine Learning, we have proposed in this paper a cognitive architecture, called DAC-EAI, allowing the conceptual description of many Artificial Intelligence systems within a unified framework. Then we have proposed the concept of a cognitive arms race between embodied agent population as a potentially powerful driving force for the development of cognitive complexity. We believe that these two research directions, summarized by the keywords integration and embodiment, are key challenges for leveraging the recent advances in the field toward the achievement of General Artificial Intelligence. This ambitious objective requires a cognitive architecture autonomously and continuously optimizing its own behavior through embodied interaction with the world. This is, however, not a sufficient condition for an agent to continuously learn increasingly complex skills. Indeed, in an environment of limited complexity with sufficient resources, the agent will rapidly converge towards an efficient strategy and there will be no need to further extend the repertoire of skills. However, if the environment contains other agents competing for the same, limited resources, the efficiency of one’s strategy will depend on the strategies adopted by the others. The constraints imposed by such a multi-agent environment with limited resources are likely to be a crucial factor in bootstrapping a positive-feedback loop of continuous improvement through competition among the agents, as described in the previous section. The main lesson of our integrative effort at the cognitive level, as summarized in Figure 2, is that powerful algorithms and control systems are existing which, taken together, span all the relevant aspects of cognition required to solve the problem of General Artificial Intelligence2. We see however that there is still a considerable amount of work to be done to integrate all the existing subparts into a coherent and complete cognitive system. This effort is central to the research program of our group and we have already demonstrated our ability to implement a complete version of the architecture (see [16], [24] for our most recent contributions). As we already noted in previous publications [15], [26], [43]– [45], there is, however, a missing ingredient in these systems preventing them to being considered at the same level as animal intelligence: they are not facing the constraint of the massively multi-agent world in which biological systems evolve. We propose here that a key constraint imposed by a multi-agent world is the emergence of positive feedback loops between competing agent populations, forcing them to continuously adapt against each other. Our approach is facing several important challenges. The first one is to leverage the recent advances in robotics and machine learning toward the achievement of general artificial intelligence, based on the principled methodology provided by the DAC framework. The second one is to provide a unified theory of cognition [46] able to bridge the gap between computational and biological science. The third one is to understand the emergence of general intelligence within its ecological substrate, i.e. the dynamical aspect of coupled physical and cognitive systems. ACKNOWLEDGMENT Work supported by ERC’s CDAC project: "Role of Consciousness in Adaptive Behavior" (ERC-2013-ADG 341196) & EU project Socialising Sensori-Motor Contingencies (socSMC-641321—H2020-FETPROACT2014), as well as the INSOCO Plan Nacional Project (DPI2016-80116-P). V. REFERENCES [1] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. 2 But this does not mean those aspects are sufficient to solve the problem. [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] Fei-Fei, “ImageNet Large Scale Visual Recognition Challenge,” Int. J. Comput. Vis., vol. 115, no. 3, pp. 211–252, 2015. A. Karpathy and L. Fei-Fei, “Deep Visual-Semantic Alignments for Generating Image Descriptions,” in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015, pp. 3128–3137. V. Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, D. Hassabis, others, S. Petersen, C. Beattie, A. Sadik, I. Antonoglou, H. King, D. Kumaran, D. Wierstra, S. Legg, and D. Hassabis, “Human-level control through deep reinforcement learning,” Nature, vol. 518, no. 7540, pp. 529– 533, Feb. 2015. 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, no. 7587, pp. 484–489, Jan. 2016. M. Bellemare, S. Srinivasan, G. Ostrovski, T. Schaul, D. Saxton, and R. Munos, “Unifying count-based exploration and intrinsic motivation,” in Advances in Neural Information Processing Systems, 2016, pp. 1471–1479. T. D. Kulkarni, K. R. Narasimhan, A. Saeedi, and J. B. Tenenbaum, “Hierarchical deep reinforcement learning: Integrating temporal abstraction and intrinsic motivation,” arXiv Prepr. arXiv1604.06057, 2016. B. M. Lake, T. D. Ullman, J. B. Tenenbaum, and S. J. Gershman, “Building Machines That Learn and Think Like People,” Behav. Brain Sci., Nov. 2017. B. M. Lake, R. Salakhutdinov, and J. B. Tenenbaum, “Human-level concept learning through probabilistic program induction,” Science (80-. )., vol. 350, no. 6266, pp. 1332–1338, Dec. 2015. A. E. Stahl and L. Feigenson, “Observing the unexpected enhances infants’ learning and exploration,” Science (80-. )., vol. 348, no. 6230, pp. 91–94, 2015. M. Johnson, K. Hofmann, T. Hutton, and D. Bignell, “The Malmo Platform for Artificial Intelligence Experimentation,” Int. Jt. Conf. Artif. Intell., pp. 4246–4247, 2016. C. Beattie, J. Z. Leibo, D. Teplyashin, T. Ward, M. Wainwright, H. Küttler, A. Lefrancq, S. Green, V. Valdés, A. Sadik, and others, “DeepMind Lab,” arXiv Prepr. arXiv1612.03801, 2016. G. Brockman, V. Cheung, L. Pettersson, J. Schneider, J. Schulman, J. Tang, and W. Zaremba, “OpenAI Gym,” in arXiv preprint arXiv:1606.01540, 2016. J. Z. Leibo, V. Zambaldi, M. Lanctot, J. Marecki, and T. Graepel, “Multi-agent Reinforcement Learning in Sequential Social Dilemmas,” Proceedings of the 16th Conference on Autonomous Agents and MultiAgent Systems. International Foundation for Autonomous Agents and Multiagent Systems, pp. 464–473, 2017. A. Tampuu, T. Matiisen, D. Kodelja, I. Kuzovkin, K. Korjus, J. Aru, J. Aru, and R. Vicente, “Multiagent Cooperation and Competition with Deep Reinforcement Learning,” Nov. 2015. C. Moulin-Frier, X. D. Arsiwalla, J.-Y. Puigbò, M. Sánchez-Fibla, A. Duff, and P. F. M. J. Verschure, “Top-Down and Bottom-Up Interactions between Low-Level Reactive Control and Symbolic Rule Learning in Embodied Agents,” in Proceedings of the Workshop on Cognitive Computation: Integrating neural and symbolic approaches. 30th Annual Conference on Neural Information Processing Systems (NIPS 2016), 2016. C. Moulin-Frier, T. Fischer, M. Petit, G. Pointeau, J.-Y. Puigbo, U. Pattacini, S. C. Low, D. Camilleri, P. Nguyen, M. Hoffmann, H. J. Chang, M. Zambelli, A.-L. Mealier, A. Damianou, G. Metta, T. Prescott, Y. Demiris, P.-F. Dominey, and P. Verschure, “DAC-h3: A Proactive Robot Cognitive Architecture to Acquire and Express Knowledge About the World and the Self,” Submitt. to IEEE Trans. Cogn. Dev. Syst., 2017. A. Newell, J. C. Shaw, and H. A. Simon, “Report on a general problem-solving program,” IFIP Congr., pp. 256–264, 1959. J. E. Laird, A. Newell, and P. S. Rosenbloom, “SOAR: An [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] [36] [37] [38] [39] [40] [41] [42] architecture for general intelligence,” Artificial Intelligence, vol. 33, no. 1. pp. 1–64, 1987. J. R. Anderson, The Architecture of Cognition. Harvard University Press, 1983. R. Brooks, “A robust layered control system for a mobile robot,” IEEE J. Robot. Autom., vol. 2, no. 1, pp. 14–23, 1986. R. A. Brooks, “Intelligence without representation,” Artif. Intell., vol. 47, no. 1–3, pp. 139–159, 1991. P. F. M. J. Verschure, T. Voegtlin, and R. J. Douglas, “Environmentally mediated synergy between perception and behaviour in mobile robots,” Nature, vol. 425, no. 6958, pp. 620– 624, 2003. P. F. M. J. Verschure, C. M. A. Pennartz, and G. Pezzulo, “The why, what, where, when and how of goal-directed choice: neuronal and computational principles,” Philos. Trans. R. Soc. B Biol. Sci., vol. 369, no. 1655, p. 20130483, 2014. G. Maffei, D. Santos-Pata, E. Marcos, M. Sánchez-Fibla, and P. F. M. J. Verschure, “An embodied biologically constrained model of foraging: From classical and operant conditioning to adaptive realworld behavior in DAC-X,” Neural Networks, 2015. V. Vouloutsi, M. Blancas, R. Zucca, P. Omedas, D. Reidsma, D. Davison, and others, “Towards a Synthetic Tutor Assistant: The EASEL Project and its Architecture,” in International Conference on Living Machines, 2016, pp. 353–364. P. F. M. J. Verschure, “Synthetic consciousness: the distributed adaptive control perspective,” Philos. Trans. R. Soc. Lond. B. Biol. Sci., vol. 371, no. 1701, pp. 263–275, 2016. G. E. Hinton and R. R. Salakhutdinov, “Reducing the Dimensionality of Data with Neural Networks,” Science (80-. )., vol. 313, no. 5786, 2006. R. S. Sutton and A. G. Barto, Reinforcement Learning: an introduction. The MIT Press, 1998. J. Oh, S. Singh, H. Lee, and P. Kohli, “Zero-Shot Task Generalization with Multi-Task Deep Reinforcement Learning,” Jun. 2017. S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,” Neural Comput., vol. 9, no. 8, pp. 1735–1780, Nov. 1997. A. Baranes and P.-Y. Oudeyer, “Active Learning of Inverse Models with Intrinsically Motivated Goal Exploration in Robots,” Rob. Auton. Syst., vol. 61, no. 1, pp. 49–73, 2013. S. M. LaValle, Planning Algorithms. 2006. Z. Chen and B. Liu, Lifelong Machine Learning. Morgan & Claypool Publishers, 2016. J.-C. Latombe, Robot motion planning, vol. 124. Springer Science & Business Media, 2012. X. Guo, S. Singh, H. Lee, R. L. Lewis, and X. Wang, “Deep learning for real-time Atari game play using offline Monte-Carlo tree search planning,” in Advances in neural information processing systems, 2014, pp. 3338–3346. A. Graves, G. Wayne, M. Reynolds, T. Harley, I. Danihelka, A. Grabska-Barwińska, S. G. Colmenarejo, E. Grefenstette, T. Ramalho, and others, “Hybrid computing using a neural network with dynamic external memory,” Nature, pp. 1476–4687, 2016. A. Graves, G. Wayne, and I. Danihelka, “Neural Turing Machines,” Arxiv, vol. abs/1410.5, 2014. R. Dawkins and J. R. Krebs, “Arms races between and within species.,” Proc. R. Soc. Lond. B. Biol. Sci., vol. 205, no. 1161, pp. 489–511, 1979. A. Weitzenfeld, A. Vallesa, and H. Flores, “A Biologically-Inspired Wolf Pack Multiple Robot Hunting Model,” in 2006 IEEE 3rd Latin American Robotics Symposium, 2006, pp. 120–127. C. Muro, R. Escobedo, L. Spector, and R. P. Coppinger, “Wolfpack (Canis lupus) hunting strategies emerge from simple rules in computational simulations,” Behav. Processes, vol. 88, no. 3, pp. 192–197, Nov. 2011. D. Floreano and L. Keller, “Evolution of adaptive behaviour in robots by means of Darwinian selection,” PLoS Biol., vol. 8, no. 1, pp. 1–8, 2010. I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. WardeFarley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,” in Advances in neural information processing systems, 2014, pp. 2672–2680. [43] [44] [45] [46] X. D. Arsiwalla, I. Herreros, C. Moulin-Frier, M. Sanchez, and P. F. M. J. Verschure, “Is Consciousness a Control Process?,” in International Conference of the Catalan Association for Artificial Intelligence, 2016, pp. 233–238. C. Moulin-Frier and P. F. M. J. Verschure, “Two possible driving forces supporting the evolution of animal communication. Comment on ‘Towards a Computational Comparative Neuroprimatology: Framing the language-ready brain’ by Michael A. Arbib,” Phys. Life Rev., 2016. X. D. Arsiwalla, C. Moulin-Frier, I. Herreros, M. Sanchez-Fibla, and P. Verschure, “The Morphospace of Consciousness,” May 2017. A. Newell, Unified theories of cognition. Harvard University Press, 1990. 

1 Techniques for Improving the Finite Length Performance of Sparse Superposition Codes arXiv:1705.02091v4 [cs.IT] 19 Nov 2017 Adam Greig, Student Member, IEEE, and Ramji Venkataramanan, Senior Member, IEEE Abstract—Sparse superposition codes are a recent class of codes introduced by Barron and Joseph for efficient communication over the AWGN channel. With an appropriate power allocation, these codes have been shown to be asymptotically capacityachieving with computationally feasible decoding. However, a direct implementation of the capacity-achieving construction does not give good finite length error performance. In this paper, we consider sparse superposition codes with approximate message passing (AMP) decoding, and describe a variety of techniques to improve their finite length performance. These include an iterative algorithm for SPARC power allocation, guidelines for choosing codebook parameters, and estimating a critical decoding parameter online instead of pre-computation. We also show how partial outer codes can be used in conjunction with AMP decoding to obtain a steep waterfall in the error performance curves. We compare the error performance of AMP-decoded sparse superposition codes with coded modulation using LDPC codes from the WiMAX standard. Index Terms—Sparse regression codes, Approximate Message Passing, Low-complexity decoding, Finite length performance, Coded modulation I. I NTRODUCTION W E consider communication over the memoryless additive white Gaussian noise (AWGN) channel given by y = x + w, where the channel output y is the sum of the channel input x and independent zero-mean Gaussian noise w of variance σ 2 . There is an average power constraint P on theP input, so a n length-n codeword (x1 , . . . , xn ) has to satisfy n1 i=1 x2i ≤ P . The goal is to build computationally efficient codes that have low probability of decoding error at rates close to the AWGN channel capacity C = 21 log(1+snr). Here snr denotes the signal-to-noise ratio P/σ 2 . Though it is well known that Shannon-style i.i.d. Gaussian codebooks can achieve very low probability of error at rates approaching the AWGN capacity [1], this approach has been largely avoided in practice due to the high decoding complexity of unstructured Gaussian codes. Current state of the art approaches for the AWGN channel such as coded modulation [2], [3] typically involve separate coding and modulation steps. In this approach, a binary error-correcting code such as an LDPC or turbo code is first used to generate a binary codeword from the information bits; the code bits are then modulated with a standard scheme such as quadrature amplitude modulation. A. Greig and R. Venkataramanan are with Department of Engineering, University of Cambridge, Cambridge CB2 1PZ, UK (e-mails: ag611@cam.ac.uk, rv285@cam.ac.uk). This work was supported in part by EPSRC Grant EP/N013999/1, and by an EPSRC Doctoral Training Award. Though these schemes have good empirical performance, they have not been proven to be capacity-achieving for the AWGN channel. Sparse Superposition Codes or Sparse Regression Codes (SPARCs) were recently proposed by Barron and Joseph [4], [5] for efficient communication over the AWGN channel. In [5], they introduced an efficient decoding algorithm called “adaptive successive decoding” and showed that it achieved near-exponential decay of error probability (with growing block length), for any fixed rate R < C. Subsequently, an adaptive soft-decision successive decoder was proposed in [6], [7], and Approximate Message Passing (AMP) decoders were proposed in [8]–[11]. The adaptive soft-decision decoder in [7] as well as the AMP decoder in [11] were proven to be asymptotically capacity-achieving, and have superior finite length performance compared to the original adaptive successive decoder of [5]. The above results mainly focused on characterizing the error performance of SPARCs in the limit of large block length. In this work, we describe a number of code design techniques for improved finite length error performance. Throughout the paper, we focus on AMP decoding due to its ease of implementation. However, many of the code design ideas can also be applied to the adaptive soft-decision successive decoder in [6], [7]. A hardware implementation of the AMP decoder was recently reported in [12], [13]. We expect that the techniques proposed in this paper can be used to reduce the complexity and optimize the decoding performance in such implementations. In the remainder of this section, we briefly review the SPARC construction and the AMP decoder from [11], and then list the main contributions of this paper. A word about notation before we proceed. Throughout the paper, we use log to denote logarithms with base 2, and ln to denote natural logarithms. For a positive integer N , we use [N ] to denote the set {1, . . . , N }. The transpose of a matrix A is denoted by A∗ , and the indicator function of an event E by 1{E}. A. The sparse superposition code A SPARC is defined in terms of a design matrix A of dimension n × M L. Here n is the block length, and M, L are integers which are specified below in terms of n and the rate R. As shown in Fig. 1, the design matrix A has L sections with M columns each. In the original construction of [4], [5] and in the theoretical analysis in [6], [7], [11], [14], the entries of A are assumed to be i.i.d. Gaussian ∼ N (0, 1/n). For our empirical results, we use a random Hadamard-based construction for A that leads to significantly lower encoding and decoding complexity [9]–[11]. 2 Section 1 M columns Section 2 M columns decoder produces iteratively refined estimates of the message vector, denoted by β 1 , β 2 , . . . , β T , where T is the (prespecified) number of iterations. Starting with β 0 = 0, for t = 0, 1, . . . , T − 1 the AMP decoder generates   kβ t k2 z t−1 t t P− , (2) z = y − Aβ + 2 τt−1 n Section L M columns A: βit+1 = ηit (β t + A∗ z t ), β: 0, √ 0, nP1 , √ 0, nP2 , 0, √ nPL , 0, ,0 T where ηit (s) = Fig. 1. A is the n×LM design matrix, β is an M L×1 sparse vector with one non-zero in each of the L sections. The length-n codeword is Aβ. The message determines the locations of the non-zeros in β, while P1 , . . . , PL are fixed a priori. Codewords are constructed as sparse linear combinations of the columns of A. In particular, a codeword is of the form Aβ, where β = (β1 , . . . , βM L )∗ is a length M L column vector with the property that there is exactly one non-zero βj for the section 1 ≤ j ≤ M , one non-zero βj for the section M + 1 ≤ j ≤ 2M , and so √ forth. The non-zero value of β in each section ` is set to nP` , where P P1 , . . . , PL are preL specified positive constants that satisfy `=1 P` = P , the average symbol power allowed. Both A and the power allocation {P1 , . . . , PL } are known to both the encoder and decoder in advance. The choice of power allocation plays a crucial role in determining the error performance of the decoder. Without loss of generality, we will assume that the power allocation is non-increasing across sections. Two examples of power allocation are: P • Flat power allocation, where P` = L for all `. This choice was used in [4] to analyze the error performance with optimal (least-squares) decoding. • Exponentially decaying power allocation, where P` ∝ 2−2C`/L . This choice was used for the asymptotically capacity-achieving decoders proposed in [5], [7], [11]. At finite block lengths both these power allocations could be far from optimal and lead to poor decoding performance. One of the main contributions of this paper is an algorithm to determine a good power allocation for the finite-length AMP decoder based only on R, P , σ 2 . Rate: As each of the L sections contains M columns, the total number of codewords is M L . With the block length being n, the rate of the code is given by log(M L ) L log M = . (1) n n In other words, a SPARC codeword corresponding to L log M input bits is transmitted in n channel uses. Encoding: The input bitstream is split into chunks of log M bits. A chunk of log M input bits can be used to index the location of the non-zero entry in one section of β. Hence L successive chunks determine the message vector β, with the `th chunk of log M input bits determining the non-zero location in section `, for 1 ≤ ` ≤ L. Approximate Message Passing (AMP) decoder: The AMP R= (3) p  √  ` exp si τnP 2 t  √ , nP` P nP` j∈sec(i) exp sj τ 2 1 ≤ i ≤ M L. t (4) Here the notation j ∈ sec(i) refers to all indices j in the same section as i. (Note that there are√M indices in each section.) At the end of each step t, βit / nP` may be interpreted as the updated posterior probability of the ith entry being the non-zero one in its section. The constants τt2 are specified by the following scalar recursion called “state evolution” (SE): τ02 = σ 2 + P, τt2 = σ 2 + P (1 − x(τt−1 )), t ≥ 1, (5) where x(τ ) := L X P` `=1 P √  √ E e e nP` τ (U1` + nP` τ √ √ (U1` + nP` τ  ) √ nP` τ ) + PM j=2 e nP` τ Uj` . (6) {Uj` } In (6), are i.i.d. N (0, 1) random variables for j ∈ [M ], ` ∈ [L]. The significance of the SE parameters τt2 is discussed in Section II. In Section IV, we use an online approach to accurately compute the τt2 values rather than precomputing them via (6). At the end of T iterations, the decoded message vector βb is produced by setting the maximum value in section ` of β T √ to nP` and the remaining entries to zero, for 1 ≤ ` ≤ L. Error rate of the AMP decoder: We measure the section error rate Esec as Esec = L o 1 X nb 1 β` 6= β` L (7) `=1 Assuming a uniform mapping between the input bitstream and the non-zero locations in each section, each section error will cause approximately half of the bits it represents to be incorrect, leading to a bit error rate Eber ≈ 21 Esec . Another figure of merit is the codeword error rate Ecw , which estimates the probability P(βb 6= β). If the SPARC is used to transmit a large number of messages (each via a length n codeword), Ecw measures the fraction of codewords that are decoded with one or more section errors. The codeword error rate is insensitive to where and how many section errors occur within a codeword when it is decoded incorrectly. At finite code lengths, the choice of a good power allocation crucially depends on whether we want to minimize Esec or Ecw . As we will see in the next section, a power allocation that yields reliably low section error rates may result in a 3 high codeword error rate, and vice versa. In this paper, we will mostly focus on obtaining the best possible section error rate, since in practical applications a high-rate outer code could readily correct a small fraction of section errors to give excellent codeword error rates as well. Further, the bit error rate (which is approximately half the section error rate) is useful to compare with other channel coding approaches, where it is a common figure of merit. B. Organization of the paper and main contributions In the rest of the paper, we describe several techniques to improve the finite length error performance and reduce the complexity of AMP decoding. The sections are organized as follows. • In Section II, we introduce an iterative power allocation algorithm that gives improved error performance with fewer tuning parameters than other power allocation schemes. • In Section III, we analyze the effects of the code parameters L, M and the power allocation on error performance and its concentration around the value predicted by state evolution. • In Section IV, we describe how an online estimate of the key SE parameter τt2 improves error performance and allows a new early-stopping criterion. Furthermore, the online estimate enables us to accurately estimate the actual section error rate at the end of the decoding process. • In Section V, we derive simple expressions to estimate Esec and Ecw given the rate and power allocation. • In Section VI we compare the error performance of AMP-decoded SPARCs to LDPC-based coded modulation schemes used in the WiMAX standard. • In Section VII, we describe how partial outer codes can be used in conjunction with AMP decoding. We propose a three-stage decoder consisting of AMP decoding, followed by outer code decoding, and finally, AMP decoding once again. We show that by covering only a fraction of sections of the message β with an outer code, the threestage decoder can correct errors even in the sections not covered by the outer code. This results in bit-error curves with a steep waterfall behavior. The main technical contributions of the paper are the iterative power allocation algorithm (Section II) and the threestage decoder with an outer code (Section VII). The other sections describe how various choices of code parameters influence the finite length error performance, depending on whether the objective is to minimize the section error rate or the codeword error rate. We remark that the focus in this paper is on improving the finite length performance using the standard SPARC construction with power allocation. Optimizing the finite length performance of spatially-coupled SPARCs considered in [9], [10] is an interesting research direction, but one that is beyond the scope of this paper. II. P OWER A LLOCATION Before introducing the power allocation scheme, we briefly give some intuition about the AMP update rules (2)–(4), and the SE recursion in (5)–(6). The update step (3) to generate each estimate of β is underpinned by the following key property: after step t, the “effective observation” β t + A∗ z t is approximately distributed as β + τt Z, where Z is standard normal random vector independent of β. Thus τt2 is the effective noise variance at the end of step t. Assuming that the above distributional property holds, β t+1 is just the Bayesoptimal estimate of β based on the effective observation. The entry βit+1 is proportional to the posterior probability of the ith entry being the non-zero entry in its section. We see from (5) that the effective noise variance τt2 is the sum of two terms. The first is the channel noise variance σ 2 . The other term P (1 − x(τt−1 )) can be interpreted as the interference due to the undecoded sections in β t . Equivalently, x(τt−1 ) is the expected power-weighted fraction of sections which are correctly decodable at the end of step t. The starting point for our power allocation design is the following result from [11], which gives analytic upper and lower bounds for x(τ ) of (5). Lemma 1. [14, Lemma 1(b)] Let ν` := sufficiently large M , and for any δ ∈ (0, 1), x(τ ) ≤ L X P` h `=1 P LP` Rτ 2 ln 2 . For i 2 1 {ν` > 2 − δ} + M −κ1 δ 1 {ν` ≤ 2 − δ} , (8) 2 x(τ ) ≥ M −κ2 δ 1− √ δ ln M ! L X `=1 P` 1 {ν` > 2 + δ} . P (9) where κ1 , κ2 are universal positive constants. As the constants κ1 , κ2 in (8)–(9) are not precisely specified, for designing power allocation schemes, we use the following approximation for x(τ ): x(τ ) ≈ L X P` `=1 P  1 LP` > 2Rτ 2 ln 2 . (10) This approximate version, which is increasingly accurate as L, M grow large, is useful for gaining intuition about suitable power allocations. Indeed, if the effective noise variance after step t is τt2 , then (10) says that any section ` whose normalized power LP` is larger than the threshold 2Rτt2 ln 2 is likely to be decodable correctly in step (t+1), i.e., in β t+1 , the probability mass within the section will be concentrated on the correct non-zero entry. For a given power allocation, we can iteratively estimate the SE parameters (τt2 , x(τt2 )) for each t using the lower bound in (10). This provides a way to quickly check whether or not a given power allocation will lead to reliable decoding in the large system limit. For reliable decoding at a given rate R, the effective noise variance given by τt2 = σ 2 + P (1 − x(τt−1 )) should decrease with t until it reaches a value close to σ 2 in a finite number of iterations. Equivalently, x(τt ) in (6) should increase to a value very close to 1. For a rate R < C, there are infinitely many power allocations for which (10) predicts successful decoding in the large system limit. However, as illustrated below, their finite length error performance may differ significantly. Thus the key question 4 Fig. 2. The dashed lines show the minimum required power in section for successful decoding when R = C (above), and R = 0.7C (below), where C = 2 bits. The solid line shows the exponentially-decaying power allocation in (11). addressed in this section is: how do we choose a power allocation that gives the lowest section error rate? The exponentially-decaying power allocation given by P (22C/L − 1) −2C`/L 2 , ` ∈ [L], (11) 1 − 2−2C was proven in [11] to be capacity-achieving in the large system limit, i.e., it was shown that the section error rate Esec of the AMP decoder converges almost surely to 0 as n → ∞, for any R < C. However, it does not perform well at practical block lengths, which motivated the search for alternatives. We now evaluate it in the context of (10) to better explain the development of a new power allocation scheme. P` = Given a power allocation, using (10) one can compute the minimum required power for any section ` ∈ [L] to decode, assuming that the sections with higher power have decoded correctly. The dashed lines in Figure 2 shows the minimum power required for each section to decode (assuming the exponential allocation of (11) for the previous sections), for R = C and R = 0.7C. The figure shows that the power allocation in (11) matches (up to order L1 terms) with the minimum required power when R = C. However, for R = 0.7C, we see that the exponentially-decaying allocation allocates significantly more power to the earlier sections than the minimum required, compared to later sections. This leads to relatively high section error rates, as shown in Figure 6. Figure 2 shows that the total power allocated by the minimal power allocation at R = 0.7C is significantly less than the available power P . Therefore, the key question is: how do we balance the allocation of available power between the various sections to minimize the section error rate? Allocating excessive power to the earlier sections ensures they decode reliably early on, but then there will not be sufficient power left to ensure reliable decoding in the final sections. This is the reason for the poor finite length performance of the exponentially-decaying allocation. Conversely, if the power is spread too evenly then no section particularly stands out against the noise, so it is hard for the decoding to get started, and early errors can cause cascading failures as subsequent sections are also decoded in error. This trade-off motivated the following modified exponential Fig. 3. The modified power allocation with a = f = 0.7 results in slightly more than the minimum power required for the first 70% of sections; the remaining available power is allocated equally among the last 30% of sections. The original allocation with P` ∝ 2−2C`/L is also shown for comparison. power allocation proposed in [11]: ( κ · 2−2aC`/L 1 ≤ ` ≤ f L, P` = κ · 2−2aCf f L + 1 ≤ ` ≤ L, (12) where PL the normalizing constant κ is chosen to ensure that `=1 P` = P . In (12), the parameter a controls the steepness of the exponential allocation, while the parameter f flattens the allocation after the first fraction f of the sections. Smaller choices of a lead to less power allocated to the initial sections, making a larger amount available to the later sections. Similarly, smaller values of f lead to more power allocated to the final sections. See Figure 3 for an illustration. While this allocation improves the section error rate by a few orders of magnitude (see [11, Fig. 4]), it requires costly numerical optimization of a and f . A good starting point is to use a = f = R/C, but further optimization is generally necessary. This motivates the need for a fast power allocation algorithm with fewer tuning parameters. A. Iterative power allocation We now describe a simple parameter-free iterative algorithm to design a power allocation. The L sections of the SPARC are divided into B blocks of L/B sections each. Each section within a block is allocated the same power. For example, with L = 512 and B = 32, there are 32 blocks with 16 sections per block. The algorithm sequentially allocates power to each of the B blocks as follows. Allocate the minimum power to the first block of sections so that they can be decoded in the first iteration when τ02 = σ 2 +P . Using (10), we set the power in each section of the first block to 2Rτ02 ln 2 L , 1≤`≤ . L B Using (10) and (5), we then estimate τ12 = σ 2 + (P − BP1 ). Using this value, allocate the minimum required power for the second block of sections to decode, i.e., P` = 2 ln 2Rτ12 /L for L 2L B + 1 ≤ ` ≤ B . If we sequentially allocate power in this manner to each of the B blocks, then the total power allocated by this scheme will be strictly less than P whenever R < C. We therefore modify the scheme as follows. For 1 ≤ b ≤ B, to allocate power to the bth block of sections assuming that the first (b − 1) blocks have been P` = 5 Fig. 4. Example illustrating the iterative power allocation algorithm with B = 5. In each step, the height of the light gray region represents the allocation that distributes the remaining power equally over all the remaining sections. The dashed red line indicates the minimum power required for decoding the current block of sections. The dark gray bars represent the power that has been allocated at the beginning of the current step. Algorithm 1 Iterative power allocation routine Require: L, B, σ 2 , P , R such that B divides L. L Initialise k ← B for b = 0 to B − P 1 do bk Premain ← P − `=1 P` 2 2 τ ← σ + Premain Pblock ← 2 ln(2)Rτ 2 /L if Premain /(L − bk) > Pblock then Pbk+1 , . . . , PL ← Premain /(L − bk) break else Pbk+1 , . . . , P(b+1)k ← Pblock end if end for return P1 , . . . , PL allocated, we compare the two options and choose the one that allocates higher power to the block: i) allocating the minimum required power (computed as above) for the bth block of sections to decode; ii) allocating the remaining available power equally to sections in blocks b, . . . , B, and terminating the algorithm. This gives a flattening in the final blocks similar to the allocation in (12), but without requiring a specific parameter that determines where the flattening begins. The iterative power allocation routine is described in Algorithm 1. Figure 4 shows a toy example building up the power allocation for B = 5, where flattening is seen to occur in step 4. Figure 5 shows a more realistic example with L = 512 and R = 0.7C. Choosing B: By construction, the iterative power allocation scheme specifies the number of iterations of the AMP decoder in the large system limit. This is given by the number of blocks with distinct powers; in particular the number of iterations (in the large system limit) is of the order of B. For finite code lengths, we find that it is better to use a termination criterion for the decoder based on the estimates generated by the algorithm. This criterion is described in Sec. IV. This datadriven termination criterion allows us to choose the number of Fig. 5. Iterative allocation, with L = 512, and B = 16 blocks. Flattening occurs at the 11th block. AMP section error rate Esec vs R at snr = 7, 15, 31, corresponding to C = 1.5, 2, 2.5 bits (shown with dashed vertical lines). At each snr, the section error rate is reported for rates R/C = 0.70, 0.75, 0.80, 0.85, 0.90. The SPARC parameters are M = 512, L = 1024. The top black curve shows the Esec with P` ∝ 2−2C`/L . The lower green curve shows Esec for the iterative power allocation, with B = L and RPA numerically optimized. (See Sec. III-B for a discussion of RPA .) Fig. 6. blocks B to be as large as L. We found that choosing B = L, together with the termination criterion in Sec. IV, consistently gives a small improvement in error performance (compared to other choices of B), with no additional time or memory cost. Additionally, with B = L, it is possible to quickly determine a pair (a, f ) for the modified exponential allocation in (12) which gives a nearly identical allocation to the iterative algorithm. This is done by first setting f to obtain the same flattening point found in the iterative allocation, and then searching for an a which matches the first allocation coefficient P1 between the iterative and the modified exponential allocations. Consequently, any simulation results obtained for the iterative power allocation could also be obtained using a suitable (a, f ) with the modified exponential allocation, without having to first perform a costly numerical optimization over (a, f ). Figure 6 compares the error performance of the exponential and iterative power allocation schemes discussed above for different values of R at snr = 7, 15, 31. The iterative power allocation yields significantly improved Esec for rates away from capacity when compared to the original exponential allocation, and additionally outperforms the modified exponential allocation results reported in [11]. 6 For the experiments in Figure 6, the value for R used in constructing the iterative allocation (denoted by RP A ) was optimized numerically. Constructing an iterative allocation with R = RP A yields good results, but due to finite length concentration effects, the RP A yielding the smallest average error rate may be slightly different from the communication rate R. The effect of RP A on the concentration of error rates is discussed in Section III-B. We emphasize that this optimization over RP A is simpler than numerically optimizing the pair (a, f ) for the modified exponential allocation. Furthermore, guidelines for choosing RP A as a function of R are given in Section III-B. Fig. 7. AMP error performance with increasing M , for L = 1024, III. E RROR C ONCENTRATION T RADE - OFFS In this section, we discuss how the choice of SPARC design parameters can influence the trade-off between the ‘typical’ value of section error rate and concentration of actual error rates around the typical values. The typical section error rate refers to that predicted by state evolution (SE). Indeed, running the SE equations (5)–(6) until convergence gives the following prediction for the section error rate:     √ √ SE Esec L 1X  E := 1−  L `=1 √ e nP` τT e  U1` + nP` τT √ nP` τT U1` + nP` τT √  + PM j=2 e nP` τT Uj`  ,  (13) where τT2 denotes the value in the final iteration. The conSE centration refers to how close the SE prediction Esec is to the observed section error rate. As we describe below, the choice of SPARC parameters (L, M ) and the power allocation both determine a trade-off SE , and concentration of between obtaining a low value for Esec SE . This trade-off is of the actual section error rate around Esec particular interest when applying an outer code to the SPARC, as considered in Section VII, which may be able to reliably handle only a small number of section errors. A. Effect of L and M on concentration Recall from (1) that the code length n at a given rate R is determined by the choice of L and M according to the relationship nR = L log M . In general, L and M may be chosen freely to meet a desired rate and code length. To understand the effect of increasing M , consider Figure 7 which shows the error performance of a SPARC with R = 1.5, L = 1024, as we increase the value of M . From (1), the code length n increases logarithmically with M . We observe that the section error rate (averaged over 200 trials) decreases with M up to M = 29 , and then starts increasing. This is in sharp contrast to the SE prediction (13) (plotted using a dashed line in Figure 7) which keeps decreasing as M is increased. This divergence between the actual section error rate and the SE prediction for large M is due to large fluctuations in the number of section errors across trials. Recent work on the error exponent of SPARCs with AMP decoding shows that the concentration of error rates near the SE prediction is strongly Eb R = 1.5, and N = 5.7 dB (2 dB from Shannon limit). See Section V 0 for details of Ēsec . dependent on both L and M . For R < C, [14, Theorem 1] shows that for any  > 0, the section error rate Esec satisfies 2 −κT L
(  ln(1+snr) −f (M )) SE P Esec > Esec +  ≤ KT e (log M )2T −1 4(1+snr) , (14) where T is the number of iterations until state evolution convergence, κT , KT are constants depending on T , and −κ2 δ 2 √ f (M ) = M is a quantity that tends to zero with growing δ ln M M . For any power allocation, T increases as R approaches C. For example, T ∝ 1/ log(C/R) for the exponential power allocation. We observe that the deviation probability bound on the RHS of (14) depends on the ratio L/(log M )2T −1 . In our experiments, T is generally on the order of a few tens. Therefore, keeping L constant, the probability of large SE increases with M . This deviations from the SE prediction Esec leads to the situation shown in Figure 7, which shows that the SE continues to decrease with M , but beyond SE prediction Esec a certain value of M , the observed average section error rate becomes progressively worse due to loss of concentration. This is caused by a small number of trials with a very large number of section errors, even as the majority of trials experience lower and lower error rates as M is increased. This effect can be clearly seen in Figure 8, which compares the histogram of section error rates over 200 trials for M = 64 and M = 4096. The distribution of errors is clearly different, but both cases have the same average section error rate due to the poorer concentration for M = 4096. To summarize, given R, snr, and L, there is an optimal M that minimizes the empirical section error rate. Beyond this value of M , the benefit from any further increase is outweighed by the loss of concentration. For a given R, values of M close to L are a good starting point for optimizing the empirical section error rate, but obtaining closed-form estimates of the optimal M for a given L is still an open question. For fixed L, R, the optimal value of M increases with snr. This effect can be seen in the results of Figure 12, where there is an inversion in the order of best-performing M values as Eb /N0 increases. This is because as snr increases, the number of iterations T for SE to converge decreases. A smaller T 7 Fig. 8. Histogram of AMP section errors over 200 trials M = 64 (top) Eb and M = 4096 (bottom), with L = 1024, R = 1.5, N = 5.7dB. 0 The left panels highlight distribution of errors around low section error counts, while the right panels show the distribution around higherror-count events. As shown in Figure 7, both cases have an average section error rate of around 10−2 . Fig. 9. Histogram of AMP section errors over 1000 trials for RPA = 0.98R (top) and RPA = 1.06R (bottom). The SPARC parameters are L = 1024, M = 512, R = 1.6, snr = 15. The left panels highlight distribution of trials with low section error counts (up to 8); the right panels indicate the distribution of infrequent but higherror-count trials. At lower RPA , many more trials have no section errors, but those that do often have hundreds. At higher RPA , at most 7 section errors were seen, but many fewer trials had zero section errors. mitigates the effect of larger M in the large deviations bound of (14). In other words, a larger snr leads to better error rate concentration around the SE prediction, so larger values of M are permissible before the performance starts degrading. B. Effect of power allocation on concentration The non-asymptotic bounds on x(τ ) in Lemma 1 indicate that at finite lengths, the minimum power required for a section ` to decode in an iteration may be slightly different than that indicated by the approximation in (10). Recall that the iterative power allocation algorithm in Section II-A was designed based on (10). We can compensate for the difference between the approximation and the actual value of x(τ ) by running the iterative power allocation in Algorithm 1 using a modified rate RPA which may be slightly different from the communication rate R. The choice of RPA directly affects the error concentration. We now discuss the mechanism for this effect and give guidelines for choosing RPA as a function of R. If we run the power allocation algorithm with RPA > R, from (10) we see that additional power is allocated to the initial blocks, at the cost of less power for the final blocks (where the allocation is flat). Consequently, it is less likely that one of the initial sections will decode in error, but more likely that some number of the later sections will instead. Figure 9 (bottom) shows the effect of choosing a large RPA = 1.06R: out of a total of 1000 trials, there were no trials with more than 7 sections decoded in error (the number of sections L = 1024); however, relatively few trials (29%) have zero section errors. Conversely, choosing RPA < R allocates less power to the initial blocks, and increases the power in the final sections which have a flat allocation. This increases the likelihood of the initial section being decoded in error; in a trial when this happens, there will be a large number of section errors. However, if the initial sections are decoded correctly, the additional power in the final sections increases the probability of the trial being completely error-free. Thus choosing RP A < R makes completely error-free trials more likely, but also increases the likelihood of having trials with a large number of sections in error. In Figure 9 (top), the smaller RPA = 0.98R gives zero or one section errors in the majority (81%) of cases, but the remaining trials typically have a large number of sections in error. To summarize, the larger the RP A , the better the concentration of section error rates of individual trials around the overall average. However, increasing RP A beyond a point just increases the average section error rate because of too little power being allocated to the final sections. For different values of the communication rate R, we empirically determined an RPA that gives the lowest average section error rate, by starting at RPA = R and searching the neighborhood in steps of 0.02R. Exceptionally, at low rates (for R ≤ 1), the optimal RPA is found to be 0, leading to a completely flat power allocation with P` = PL for all `. We note from (10) that for 1 ≥ R > 2τ 2Pln 2 , the large system 0 limit theory does not predict that we can decode any of the L sections — this is because no section is above the threshold in the first iteration of decoding. However, in practice, we observe that some sections will decode initially (due to the correct column being aligned favorably with the noise vector), and this reduces the threshold enough to allow subsequent decoding to continue in most cases. For R ≤ 1, when RPA closer to R is used, the lower power in later sections hinders the finite length decoding performance. PA We found that the value of RR that minimizes the average section error rate increases with R. In particular, the optimal RP A PA was 0 for R ≤ 1; the optimal RR for R = 1.5 was R PA close to 1, and for R = 2, the optimal RR was between 1.05 and 1.1. Though this provides a useful design guideline, a deeper theoretical analysis of the role of RP A in optimizing the finite length performance is an open question. Finally, a word of caution when empirically optimizing RPA to minimize the average section error rate. Due to the loss of concentration as RP A is decreased below R, care must be taken to run sufficient trials to ensure that a rare unseen trial 8 with many section errors will not catastrophically impact the overall average section error rate. For example, in one scenario with L = 1024, M = 512, snr = 15, R = 1.4, RPA = 1.316, we observed 192 trials with errors out of 407756 trials, but only 4 of these trials had more than one error, with between 400 to 600 section errors in those 4 cases. The average section error rate was 5.6 × 10−6 . With fewer trials, it is possible that no trials with a large number of section errors would be observed, leading to an estimated error rate an order of magnitude better, at around 4.6 × 10−7 . IV. O NLINE C OMPUTATION OF τt2 AND E ARLY T ERMINATION Recall that the update step (4) of the AMP decoder requires the SE coefficients τt2 , for t ∈ [T ]. In the standard implementation [11], these coefficients are computed in advance using the SE equations (5)–(6). The total number of iterations T is also determined in advance by computing the number of iterations required the SE to converge to its fixed point (to within a specified tolerance). This technique produced effective results, but advance computation is slow as each of the L expectations in (6) needs to be computed numerically via Monte-Carlo simulation, for each t. A faster approach is to compute the τt2 coefficients using the asymptotic expression for x(τ ) given in (10). This gives error performance nearly identical to the earlier approach with significant time savings, but still requires advance computation. Both these methods are referred to as “offline” as the τt2 values are computed a priori. A simple way to estimate τt2 online during the decoding process is as follows. In each step t, after producing z t as in (2), we estimate n τbt2 = 1X 2 kz t k2 z . = n n i=1 i (15) The justification for this estimate comes from the analysis of the AMP decoder in [11], [14], which shows that for large n, τbt2 is close to τt2 in (5) with high probability. In particular, [14] provides a concentration inequality for τbt2 similar to (14). We note that such a similar online estimate has been used previously in various AMP and GAMP algorithms [8]–[10], [15]. Here, we show that in addition to being fast, the online estimator permits an interpretation as a measure of SPARC decoding progress and provides a flexible termination criterion for the decoder. Furthermore, the error performance with the online estimator was observed to be the same or slightly better than the offline methods. Recall from the discussion at the beginning of Section II that in each step, we have st := β t + A∗ z t ≈ β + τt Z, (16) where Z is a standard normal random vector independent of β. Starting from τ02 = σ 2 + P , a judicious choice of power allocation ensures that the SE parameter τt2 decreases with t, until it converges at τT2 = σ 2 in a finite number of iterations T. However, at finite lengths there are deviations from this trajectory of τt2 predicted by SE, i.e., the variance of the Fig. 10. Comparison between offline and online trajectories of the effective noise variance, at L = 1024, M = 512, P = 15, σ 2 = 1, R = 1.6. The dashed line represents the pre-computed SE trajectory of τt2 . The plot shows 15 successful runs, and one uncommon run with many section errors. The true value of Var[st − β] during decoding tracks τbt2 too precisely to distinguish on this plot. effective noise vector (st −β) may deviate from τt2 . The online estimator τbt2 is found to track Var(st −β) = kst −βk2 /n very accurately, even when this variance deviates significantly from τt2 . This effect can be seen in Figure 10, where 16 independent decoder runs are plotted and compared with the SE trajectory for τt2 (dashed line). For the 15 successful runs, the empirical variance Var(st − β) approaches σ 2 = 1 along different trajectories depending on how the decoding is progressing. In the unsuccessful run, Var(st − β) converges to a value much larger than σ 2 . In all the runs, τbt2 is indistinguishable from Var(st − β). This indicates that we can use the final value τbT2 to accurately estimate the power of the undecoded sections — and thus the number of sections decoded correctly — at runtime. Indeed, (b τT2 − σ 2 ) is an accurate estimate of the total power in the incorrectly decoded sections. This, combined with the fact that the power allocation is non-increasing, allows the decoder to estimate the number of incorrectly decoded sections. Furthermore, we can use the change in τbt2 between iterations to terminate the decoder early. If the value τbt2 has not changed between successive iterations, or the change is within some small threshold, then the decoder has stalled and no further iterations are worthwhile. Empirically we find that a stopping criterion with a small threshold (e.g., stop when 2 |b τt2 − τbt−1 | < PL ) leads to no additional errors compared to running the decoder for the full iteration count, while giving a significant speedup in most trials. Allowing a larger threshold for the stopping criterion gives even better running time improvements. This early termination criterion based on τbt2 gives us flexibility in choosing the number of blocks B in the iterative power allocation algorithm of Section II-A. This is because the number of AMP iterations is no longer tied to B, hence B can be chosen as large as desired. To summarize, the online estimator τbt2 provides an estimate of the noise variance in each AMP iteration that accurately reflects how the decoding is progressing in that trial. It thereby enables the decoder to effectively adapt to deviations from the τt2 values predicted by SE. This explains the improved 9 performance compared to the offline methods of computing τt2 . More importantly, it provides an early termination criterion for the AMP decoder as well as a way to track decoding progress and predict the number of section errors at runtime. V. P REDICTING ESEC , EBER AND ECW For a given power allocation {P` } and reasonably large SPARC parameters (n, M, L), it is desirable to have a quick way to estimate the section error rate and codeword error rate, without resorting to simulations. Without loss of generality, we assume that the power allocation is asymptotically good, i.e., the large system limit SE parameters (computed using (10)) predict reliable decoding, i.e., the SE converges to xT = 1 and τT2 = σ 2 in the large system limit. The goal is to estimate the finite length section error rate Esec . One way to estimate Esec is via the state evolution prediction (13), using τT = σ. However, computing (13) requires computing L expectations, each involving a function of M independent standard normal random variables. The following result provides estimates of Esec and Ecw that are as accurate as the SE-based estimates, but much simpler to compute. Proposition 1. Let the power allocation {P` } be such that the state evolution iteration using the asymptotic approximation (10) converges to τT2 = σ 2 . Then, under the idealized assumption that β T +A∗ z T = β+τT Z (where Z is a standard normal random vector independent of β), we have the following. The probability of a section (chosen uniformly at random) being incorrectly decoded is M −1  √ L nP` 1X EU Φ +U . (17) Ēsec = 1 − L σ `=1 The probability of the codeword being incorrectly decoded is M −1  √ L Y nP` Ēcw = 1 − EU Φ +U . (18) σ Fig. 11. Comparison of codeword error rate between simulation results and Perr -based analysis, for Ecw with varying M . L = 1024, R = 1.5, Eb /N0 = 5.7dB. Results are well matched even when concentration is poor. error can be computed as  p nP` + σZ`,1 > σZ`,j , 2 ≤ j ≤ M Perr,` = 1 − P √   Z Y M nP` P Z`,j < =1− + u Z`,1 = u φ(u)du σ R j=2 M −1  √ nP` +U , = 1 − EU Φ σ (19) where φ and Φ denote the density and the cumulative distribution function of the standard normal distribution, respectively. In the second line of (19), we condition on Z`,1 and then use the fact that Z`,1 , . . . , Z`,M are i.i.d. ∼ N (0, 1). The probability of a section chosen uniformly at random PL being incorrectly decoded is L1 `=1 Perr,` . The probability of codeword error is one minus the QLprobability that no section is in error, which is given by 1− `=1 (1−Perr,` ). Substituting for Perr,` from (19) yields the expressions in (17) and (18). `=1 In both expressions above, U is a standard normal random variable, and Φ(.) is the standard normal cumulative distribution function. Proof: As τT2 = σ 2 , the effective observation in the final iteration has the representation β + σZ. The denoising function η T generates a final estimate based on this effective observation, and the index of the largest entry in each section b Consider is chosen to form the decoded message vector β. the decoding of section ` of β. Without loss of generality, we can assume that the first entry of the section is the non-zero one. Using the notation β`,j to denote √ the jth entry of the section β` , we therefore have β`,1 = nP` , and β`,j = 0 for 2 ≤ j ≤ M . As the effective observation for section ` has the representation (β T + A∗ z T )` = β` + σZ` , the section will be incorrectly decoded if and only if the following event occurs: np o np o nP` + σZ`,1 ≤ σZ`,2 ∪. . .∪ nP` + σZ`,1 ≤ σZ`,M . Therefore, the probability that the `th section is decoded in The section error rate and codeword error rate can be estimated using the idealized expressions in (17) and (18). This still requires computing L expectations, but each expectation is now a function of a single Gaussian random variable, rather than the M independent ones in the SE estimate. Thus we reduce the complexity by a factor of M over the SE approach; evaluations of Ēsec and Ēcw typically complete within a second. SE Figure 7 shows Ēsec alongside the SE estimate Esec for L = 1024, and various values of M . We see that both these estimates match the simulation results closely up to a certain value of M . Beyond this point, the simulation results diverge from theoretical estimates due to lack of concentration in section error rates across trials, as described in Sec. III-A. Figure 11 compares the idealized codeword error probability in (18) with that obtained from simulations. Here, there is a good match between the estimate and the simulation results as the concentration of section error rates across trials plays no role — any trial with one or more section errors corresponds to one codeword error. 10 Figure 12 shows that for L = 1024, the best value of M among those considered increases from M = 29 at lower snr values to M = 213 at higher snr values. This is due to the effect discussed in Section III-A, where larger snr values can support larger values of M , before performance starts degrading due to loss of concentration. At both R = 1 and R = 1.5, the SPARCs outperform the LDPC coded modulation at Eb /N0 values close to the Shannon limit, but the error rate does not drop off as quickly at higher values of Eb /N0 . One way to enhance SPARC performance at higher snr is by treating it as a high-dimensional modulation scheme and adding an outer code. This is the focus of the next section. Fig. 12. Comparison with LDPC coded modulation at R = 1 Fig. 13. Comparison with LDPC coded modulation at R = 1.5 VI. C OMPARISON WITH C ODED M ODULATION In this section, we compare the performance of AMPdecoded SPARCs against coded modulation with LDPC codes. Specifically, we compare with two instances of coded modulation with LDPC codes from the WiMax standard IEEE 802.16e: 1) A 16-QAM constellation with a rate 12 LDPC code for an overall rate R = 1 bit/channel use/real dimension, and 2) A 64-QAM constellation with a rate 21 LDPC code for an overall rate R = 1.5 bits/channel use/real dimension. (The spectral efficiency is 2R bits/s/Hz.) The coded modulation results, shown in dashed lines in Figures 12 and 13, are obtained using the CML toolkit [16] with LDPC code lengths n = 576 and n = 2304. Each figure compares the bit error rates (BER) of the coded modulation schemes with various SPARCs of the same rate, including a SPARC with a matching code length of n = 2304. Using P = Eb R and σ 2 = N20 , the signal-to-noise ratio of the b SPARC can be expressed as σP2 = 2RE N0 . The SPARCs are implemented using Hadamard-based design matrices, power allocation designed using the iterative algorithm in Sec. II-A with B = L, and online τbt2 parameters with the early termination criterion (Sec. IV). An IPython notebook detailing the SPARC implementation is available at [17]. VII. AMP WITH PARTIAL O UTER C ODES Figures 12 and 13 show that for block lengths of the order of a few thousands, AMP-decoded SPARCs do not exhibit a steep waterfall in section error rate. Even at high Eb /N0 values, it is still common to observe a small number of section errors. If these could be corrected, we could hope to obtain a sharp waterfall behavior similar to the LDPC codes. In the simulations of the AMP decoder described above, when M and RPA are chosen such that the average error rates are well-concentrated around the state evolution prediction, the number of section errors observed is similar across trials. Furthermore, we observe that the majority of sections decoded incorrectly are those in the flat region of the power allocation, i.e., those with the lowest allocated power. This suggests we could use a high-rate outer code to protect just these sections, sacrificing some rate, but less than if we naı̈vely protected all sections. We call the sections covered by the outer code protected sections, and conversely the earlier sections which are not covered by the outer code are unprotected. In [4], it was shown that a Reed-Solomon outer code (that covered all the sections) could be used to obtain a bound the probability of codeword error from a bound on the probability of excess section error rate. Encoding with an outer code (e.g., LDPC or Reed-Solomon code) is straightforward: just replace the message bits corresponding to the protected sections with coded bits generated using the usual encoder for the chosen outer code. To decode, we would like to obtain bit-wise posterior probabilities for each codeword bit of the outer code, and use them as inputs to a soft-information decoder, such as a sum-product or min-sum decoder for LDPC codes. The output of the AMP decoding algorithm permits this: it yields β T , which contains weighted section-wise posterior probabilities; we can directly transform these into bit-wise posterior probabilities. See Algorithm 2 for details. Moreover, in addition to correcting AMP decoding errors in the protected sections, successfully decoding the outer code also provides a way to correct remaining errors in the unprotected sections of the SPARC codeword. Indeed, after decoding the outer code we can subtract the contribution of the protected sections from the channel output sequence y, and re-run the AMP decoder on just the unprotected sections. The key point is that subtracting the contribution of the later 11 T β: ··· ··· ··· L sections Luser Lparity Lunprotected Lprotected LLDP C Fig. 14. Division of the L sections of β for an outer LDPC code (protected) sections eliminates the interference due to these sections; then running the AMP decoder on the unprotected sections is akin to operating at a much lower rate. Thus the decoding procedure has three stages: i) first round of AMP decoding, ii) decoding the outer code using soft outputs from the AMP, and iii) subtracting the contribution of the sections protected by the outer code, and running the AMP decoder again for the unprotected sections. We find that the final stage, i.e., running the AMP decoder again after the outer code recovers errors in the protected sections of the SPARC, provides a significant advantage over a standard application of an outer code, i.e., decoding the final codeword after the second stage. We describe this combination of SPARCs with outer codes below, using an LDPC outer code. The resulting error rate curves exhibit sharp waterfalls in final error rates, even when the LDPC code only covers a minority of the SPARC sections. We use a binary LDPC outer code with rate RLDP C , block length nLDP C and code dimension kLDP C , so that kLDP C /nLDP C = RLDP C . For clarity of exposition we assume that both nLDP C and kLDP C are multiples of log M (and consequently that M is a power of two). As each section of the SPARC corresponds to log M bits, if log M is an integer, then nLDP C and kLDP C bits represent an integer number of SPARC sections, denoted by LLDP C = nLDP C log M and Lprotected = kLDP C , log M respectively. The assumption that kLDP C and nLDP C are multiples of log M is not necessary in practice; the general case is discussed at the end of the next subsection. We partition the L sections of the SPARC codeword as shown in Fig 14. There are Luser sections corresponding to the user (information) bits; these sections are divided into unprotected and protected sections, with only the latter being covered by the outer LDPC code. The parity bits of the LDPC codeword index the last Lparity sections of the SPARC. For convenience, the protected sections and the parity sections together are referred to as the LDPC sections. For a numerical example, consider the case where L = 1024, M = 256. There are log M = 8 bits per SPARC section. For a (5120, 4096) LDPC code (RLDP C = 4/5) we obtain the following relationships between the number of the sections of each kind: (5120 − 4096) nLDP C − kLDP C = = 128, Lparity = log M 8 Luser = L − Lparity = 1024 − 128 = 896, Algorithm 2 Weighted position posteriors β` to bit posteriors p0 , . . . , plog M −1 for section ` ∈ [L] Require: β` = [β`,1 , . . . , β`,M ], for M a power of 2 Initialise bit posteriors p0 , . . . , plog M −1 ← 0 PM Initialise normalization constant c ← i=1 β`,i for log i = 0, 1, . . . , log M − 1 do b ← log M − log i − 1 k←i while k < M do for j = k + 1, k + 2, . . . , k + i do pb ← pb + β`,j /c end for k ← k + 2i end while end for return p0 , . . . , plog M −1 kLDP C 4096 = = 512, log M 8 = Lprotected + Lparity = 512 + 128 = 640, Lprotected = LLDP C Lunprotected = Luser − Lprotected = L − LLDP C = 384. There are Luser log M = 7168 user bits, of which the final kLDP C = 4096 are encoded to a systematic nLDP C = 5120bit LDPC codeword. The resulting L log M = 8192 bits (including both the user bits and the LDPC parity bits) are encoded to a SPARC codeword using the SPARC encoder and power allocation described in previous sections. We continue to use R to denote the overall user rate, and n to denote the SPARC code length so that nR = Luser log M . The underlying SPARC rate (including the overhead due to the outer code) is denoted by RSP ARC . We note that nRSP ARC = L log M , hence RSP ARC > R. For example, with R = 1 and L, M and the outer code parameters as chosen above, n = Luser (log M )/R = 7168, so RSP ARC = 1.143. A. Decoding SPARCs with LDPC outer codes At the receiver, we decode as follows: 1) Run the AMP decoder to obtain β T . Recall that entry j within section ` of β T is proportional to the posterior probability of the column j being the transmitted one for section `. Thus the AMP decoder gives section-wise posterior probabilities for each section ` ∈ [L]. 2) Convert the section-wise posterior probabilities to bitwise posterior probabilities using Algorithm 2, for each of the LLDP C sections. This requires O(LLDP C M log M ) time complexity, of the same order as one iteration of AMP. 3) Run the LDPC decoder using the bit-wise posterior probabilities obtained in Step 2 as inputs. 4) If the LDPC decoder fails to produce a valid LDPC codeword, terminate decoding here, using β T to produce β̂ by selecting the maximum value in each section (as per usual AMP decoding). 5) If the LDPC decoder succeeds in finding a valid codeword, we use it to re-run AMP decoding on the unprotected sections. For this, first convert the LDPC codeword 12 bits to a partial β̂LDP C as follows, using a method similar to the original SPARC encoding: a) Set the first Lunprotected M entries of β̂LDP C to zero, b) The remaining LLDP C sections (with M entries per section) of β̂LDP C will have exactly one-non zero entry per section, with the LDPC codeword determining the location of the non-zero in each section. Indeed, noting that nLDP C = LLDP C log M , we consider the LDPC codeword as a concatenation of LLDP C blocks of log M bits each, so that each block of bits indexes the location of the non-zero entry in one section of β̂LDP √ C . The value of the non-zero in section ` is set to nP` , as per the power allocation. Now subtract the codeword corresponding to β̂LDP C from the original channel output y, to obtain y 0 = y − Aβ̂LDP C . 6) Run the AMP decoder again, with input y 0 , and operating only over the first Lunprotected sections. As this operation is effectively at a much lower rate than the first decoder (since the interference contribution from all the protected sections is removed), it is more likely that the unprotected bits are decoded correctly than in the first AMP decoder. We note that instead of generating y 0 , one could run the AMP decoder directly on y, but enforcing that in each AMP iteration, each of the LLDP C sections has all its non-zero mass on the entry determined by β̂LDP C , i.e., consistent with Step 5.b). 7) Finish decoding, using the output of the final AMP decoder to find the first Lunprotected M elements of β̂, and using β̂LDP C for the remaining LLDP C M elements. In the case where nLDP C and kLDP C are not multiples of log M , the values LLDP C = nLDP C / log M and Lprotected = kLDP C / log M will not be integers. Therefore one section at the boundary of Lunprotected and Lprotected will consist of some unprotected bits and some protected bits. Encoding is not affected in this situation, as the LDPC encoding happens prior to SPARC codeword encoding. When decoding, conversion to bit-wise posterior probabilities is performed for all sections containing LDPC bits (including the intermediate section at the boundary) and only the nLDP C bit posteriors corresponding to the LDPC codeword are given to the LDPC decoder. When forming β̂LDP C , the simplest option is to treat the intermediate section as though it were unprotected and set it to zero. It is also possible to compute column posterior probabilities which correspond to the fixed LDPC bits and probabilities arising from y, though doing so is not covered in this paper. B. Simulation results The combined AMP and outer LDPC setup described above was simulated using the (5120, 4096) LDPC code (RLDP C = 4/5) specified in [18] with a min-sum decoder. Bit error rates were measured only over the user bits, ignoring any bit errors in the LDPC parity bits. Figure 15 plots results at overall rate R = 45 , where the underlying LDPC code (modulated with BPSK) can be compared to the SPARC with LDPC outer code, and to a plain SPARC with rate 45 . In this case RP A = 0, giving a flat power Fig. 15. Comparison to plain AMP and to BPSK-modulated LDPC at overall rate R = 0.8. The SPARCs are both L = 768, M = 512. The underlying SPARC rate when the outer code is included is RSP ARC = 0.94. The BPSK-modulated LDPC is the same CCSDS LDPC code [18] used for the outer code. For this configuration, Luser = 654.2, Lparity = 113.8, Lunprotected = 199.1, Lprotected = 455.1, and LLDP C = 568.9. Fig. 16. Comparison to plain AMP and to the QAM-64 WiMAX LDPC of Section VI at overall rate R = 1.5 The SPARCs are both L = 1024, M = 512. The underlying SPARC rate including the outer code is RSP ARC = 1.69. For this configuration, Luser = 910.2, Lparity = 113.8, Lunprotected = 455.1, Lprotected = 455.1, and LLDP C = 455.1. allocation. Figure 16 plots results at overall rate R = 1.5, where we can compare to the QAM-64 WiMAX LDPC code, and to the plain SPARC with rate 1.5 of Figure 13. The plots show that protecting a fraction of sections with an outer code does provide a steep waterfall above a threshold Eb value of N . Below this threshold, the combined SPARC 0 + outer code has worse performance than the plain rate R SPARC without the outer code. This can be explained as follows. The combined code has a higher SPARC rate RSP ARC > R, which leads to a larger section error rate for the first AMP decoder, and consequently, to worse bit-wise Eb below the posteriors at the input of the LDPC decoder. For N 0 threshold, the noise level at the input of the LDPC decoder is beyond than the error-correcting capability of the LDPC code, so the LDPC code effectively does not correct any section 13 errors. Therefore the overall performance is worse than the performance without the outer code. Above the threshold, we observe that the second AMP decoder (after subtracting the contribution of the LDPC-protected sections) is successful at decoding the unprotected sections that were initially decoded incorrectly. This is especially apparent in the R = 45 case (Figure 15), where the section errors are uniformly distributed over all sections due to the flat power allocation; errors are just as likely in the unprotected sections as in the protected sections. experimentation is necessary to determine the best tradeoff. An interesting direction for future work would be to develop an EXIT chart analysis to jointly optimize the design of the SPARC and the outer LDPC code. ACKNOWLEDGEMENT The authors thank the Editor and the anonymous referees for several helpful comments which improved the paper. R EFERENCES C. Outer code design choices In addition to the various SPARC parameters discussed in previous sections, performance with an outer code is sensitive to what fraction of sections are protected by the outer code. When more sections are protected by the outer code, the overhead of using the outer code is also higher, driving RSP ARC higher for the same overall user rate R. This leads to worse performance in the initial AMP decoder, which has to operate at the higher rate RSP ARC . As discussed above, if RSP ARC is increased too much, the bit-wise posteriors input to the LDPC decoder are degraded beyond its ability to successfully decode, giving poor overall performance. Since the number of sections covered by the outer code depends on both log M and nLDP C , various trade-offs are possible. For example, given nLDP C , choosing a larger value of log M corresponds to fewer sections being covered by the outer code. This results in smaller rate overhead, but increasing log M may also affect concentration of the error rates around the SE predictions, as discussed in Section III-A. We conclude with two remarks about the choice of parameters for the SPARC and the outer code. 1) When using an outer code, it is highly beneficial to have good concentration of the section error rates for the initial AMP decoder. This is because a small number of errors in a single trial can usually be fully corrected by the outer code, while occasional trials with a very large number of errors cannot. 2) Due to the second AMP decoder operation, it is not necessary for all sections with low power to be protected by the outer code. For example, in Figure 15, all sections have equal power, and around 30% are not protected by the outer code. Consequently, these sections are often not decoded correctly by the first decoder. Only once the protected sections are removed is the second decoder able to correctly decode these unprotected sections. In general the aim should be to cover all or most of the sections in the flat region of the power allocation, but [1] R. G. Gallager, Information theory and reliable communication. Springer, 1968. [2] A. Guillén i Fàbregas, A. Martinez, and G. Caire, Bit-interleaved coded modulation. Now Publishers Inc, 2008. [3] G. Böcherer, F. Steiner, and P. Schulte, “Bandwidth efficient and rate-matched low-density parity-check coded modulation,” IEEE Trans. Commun., vol. 63, no. 12, pp. 4651–4665, 2015. [4] A. Barron and A. Joseph, “Least squares superposition codes of moderate dictionary size are reliable at rates up to capacity,” IEEE Trans. Inf. Theory, vol. 58, no. 5, pp. 2541–2557, Feb 2012. [5] A. Joseph and A. R. Barron, “Fast sparse superposition codes have near exponential error probability for R < C,” IEEE Trans. Inf. Theory, vol. 60, no. 2, pp. 919–942, Feb. 2014. [6] A. R. Barron and S. Cho, “High-rate sparse superposition codes with iteratively optimal estimates,” in Proc. IEEE Int. Symp. Inf. Theory, 2012. [7] S. Cho and A. Barron, “Approximate iterative Bayes optimal estimates for high-rate sparse superposition codes,” in Sixth Workshop on Information-Theoretic Methods in Science and Engineering, 2013. [8] J. Barbier and F. Krzakala, “Replica analysis and approximate message passing decoder for superposition codes,” in Proc. IEEE Int. Symp. Inf. Theory, 2014, pp. 1494–1498. [9] J. Barbier, C. Schülke, and F. Krzakala, “Approximate message-passing with spatially coupled structured operators, with applications to compressed sensing and sparse superposition codes,” Journal of Statistical Mechanics: Theory and Experiment, no. 5, 2015. [10] J. Barbier and F. Krzakala, “Approximate message-passing decoder and capacity-achieving sparse superposition codes,” IEEE Trans. Inf. Theory, vol. 63, no. 8, pp. 4894 – 4927, 2017. [11] C. Rush, A. Greig, and R. Venkataramanan, “Capacity-achieving sparse superposition codes via approximate message passing decoding,” IEEE Trans. Inf. Theory, vol. 63, no. 3, pp. 1476–1500, March 2017. [12] C. Condo and W. J. Gross, “Sparse superposition codes: A practical approach,” in Proc. IEEE Workshop on Signal Processing Systems (SiPS), 2015. [13] ——, “Implementation of sparse superposition codes,” IEEE Trans. Signal Process., vol. 65, no. 9, pp. 2421–2427, 2017. [14] C. Rush and R. Venkataramanan, “The error exponent of sparse regression codes with AMP decoding,” in Proc. IEEE Int. Symp. Inf. Theory, 2017. [15] S. Rangan, “Generalized approximate message passing for estimation with random linear mixing,” in Proc. IEEE Int. Symp. Inf. Theory, 2011. [16] (2008) The coded modulation library. [Online]. Available: http://www. iterativesolutions.com/Matlab.htm [17] Python script for SPARC with AMP decoding. [Online]. Available: https: //github.com/sigproc/sparc-amp/blob/master/sparc amp.ipynb [18] 131.0-B-2 TM Synchonization and Channel Coding, CCSDS, August 2011. [Online]. Available: https://public.ccsds.org/Pubs/131x0b2ec1.pdf 

IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, No 2, July 2011 ISSN (Online): 1694-0814 www.IJCSI.org Performance Analysis Cluster and GPU Computing Environment on Molecular Dynamic Simulation of BRV-1 and REM2 with GROMACS Heru Suhartanto1, Arry Yanuar2 and Ari Wibisono3 1 2 Faculty of Computer Science, Universitas Indonesia Depok, 16424, Indonesia Department of Pharmacy, Faculty of Mathematics and Natural Science, Universitas IndonesiaDepok, 16424, Indonesia 3 Faculty of Computer Science, Universitas Indonesia Depok, 16424, Indonesia Abstract One of application that needs high performance computing resources is molecular d ynamic. There is some software available that perform molecular dynamic, one of these is a well known GROMACS. Our previous experiment simulating molecular dynamics of Indonesian grown herbal compounds show sufficient speed up on 32 n odes Cluster computing environment. In order to obtain a reliable simulation, one usually needs to run the experiment on the scale of hundred nodes. But this is expensive to develop and maintain. Since the invention of Graphical Processing Units that is also useful for general programming, many applications have been developed to run on this. This paper reports our experiments that evaluate the performance of GROMACS that runs on two different environment, Cluster computing resources and GPU based PCs. We run the experiment on BRV-1 and REM2 compounds. Four different GPUs are installed on the same type of PCs of quad cores; they are Gefore GTS 250, GTX 465, GTX 470 and Quadro 4000. We build a cluster of 16 nodes based on these four quad cores PCs. The preliminary experiment shows that those run on GTX 470 is the best among the other type of GPUs and as well as the cluster computing resource. A speed up around 11 and 12 is gained, while the cost of computer with GPU is only about 25 percent that of Cluster we built. Keywords: Dynamic, GROMACS, Computing, Performance Analysis. GPU, Cluster 1. Introduction Virus is as one of the cause of illness. It is the smallest natural organism. Because of its simplicity and its small size, biologists choose virus as the first effort to simulate life forms with computer, and then choose one of the smallest, called mpsac tobacco satellite for further investigation. Researchers simulate viruses in a d rop of salt water use software – NAMD (Nanoscale Molecular Dynamics) built by the University of Illinois at UrbanaChampaign [13] Molecular Dynamic (MD) shows molecule structures, movement and function of molecules. MD performs the computation of atom movement in molecular system using molecular mechanics. The dynamic of a protein molecule is affected by protein structure and is an important element of special function and also general function of the protein. The understanding of relation between the dynamical three dimension structures of a protein is very important to know how a protein works. However, further real experiment of protein dynamic is very difficult to be done. Thus, people develop molecular dynamic simulation as a virtual experimental method which is able to analyze the relation between structure and protein dynamic. The simulation explores conformation energy of a protein molecule itself. Up to today, the development of MD simulation is still in progress. MD simulation in general is used to gain information on the movement and the changes of structure conformation of a protein as well as other biological macromolecules. Through this simulation, thermodynamic and kinetic information of a protein can be explored [2]. GROMACS is one of a computer program which is able to run MD simulation and energy minimization. It simulates Newton movement equation of system with hundreds to million molecules. The use of GROMACS can be implemented on biochemistry molecule such as protein which is dynamic and owns complicated binding. In addition, research on non biological molecules, such as polymers can be done with the program [7]. The property of a protein that is dynamic and changing in time is one of 131 IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, No 2, July 2011 ISSN (Online): 1694-0814 www.IJCSI.org the reasons why MD is necessary to be done [6]. With this simulation, the protein molecular movement and interaction that occurs in molecular level on a certain time can be estimated [8]. Our previous experiment using GROMACS on Cluster computing environment produced analysis on MD simulation results with three proteins of RGK sub family; they are Rad, Gem and RemGTpase. MD simulation against protein of subfamily RGK will help us to analyze the activities difference and the regulation mechanism among protein that is member of such subfamily. [9]. As we will mention in the following, we still face some limitation of computing resources. The parallel computation technique with MD using GROMACS c an visualize atomic details on the formation of s mall DPPC (dipalmitoyl phosphatidyl choline) with time order up to 90 ns [2]. #CPU 1 4 8 16 32 0 0 0 0 0 41.7 10.4 5.2 2.6 1.3 Days 206.3 52.1 26.0 13.0 6.5 1250 512.5 156.3 78.1 39.1 3750 937.5 468.8 234.4 117.2 Figure 1. Simulation in 90 ns which shows the formation of DPPC vesicle (dipalmitoyl phosphatidyl choline) [2] The figure 1 shows that, the simulation requires huge amount of computing resources in the scale of hundreds processors in order to complete the processes within an acceptable short time. However, for some people, to build and maintain a cl uster of hundreds computer acting as hundreds processors requires a lot of other resources such as electricity supplies and space to accommodate the machines. The invention of General Programming o n Graphical Processing Units (GPGPU) which provides hundreds of processors encourage many scientists to investigate the possibility of running their experiments on GPU based computers. GPU computing with CUDA (Compute Unified Driver Architecture) brings parallel data computation into wider societies. At the end of 2007, there were more than 40 millions application based CUDA GPU developers. In term of price, GPU graphical card is relative cheap. In the year 2007, one unit of the card that is capable of performing 500 G igaFlops costs less than 200 U SD. Currently, one unit of GPU of GEForce GTX 470 that consists of 448 processors cost less than 300 USD. System based on GPU runs relative fast. The bandwidth and the computation reach about 10 times of regular CPU. The micro-benchmark performance reaches mathematical instruction about 472 G FLOPS (for GPU of 8800 U ltra type); the raw bandwidth reaches 86 GB per second (for GPU of Tesla C870 type). Some application can ran faster such as N-body computation that reaches 240 GFLOPS. This is about 12 billion interaction per second. This case study was conducted on molecular dynamic and seismic data processing [2]. The General Purpose Programming on GPU (GPGPU) is the programming processes for general non graphical application on GPU. Initially, the technique is relatively difficult. The problems to be solved should be considered related to graphics. Data has to be mapped into texture maps and the algorithms should be adjusted for image synthesis. Even though this technique is promising, but it is delicate particularly for non graphic developers. There are also some constraints such as the overheads of API graphics, the memory limitation, and the needs for data passes to fulfill the bandwidth. The libraries for GPGPU is developed and provided in a ready to use programming languages . Many libraries are developed for CUDA such as CUBLAS (Basic Linear Algebra Subprograms in CUDA) and CUFFT (Fast Fourier Transform in CUDA) [3,4]. With the availability of GPGPU, CUDA and the libraries, one of difficulties in developing application run on GPU is solved. With the increasing performance of GPU, it is expected that the GPU computing support the development new application or new techniques in many areas of researches and industries. Many applications have been developed to run on GPU based computer. The capability of a machine to provide images in highly detailed within a very fast time unit is needed in breast cancer scanning processes. Techniscan, an industry that develops imaging system for automatically ultrasound, switched t heir implementation from CPU based into CUDA and NVIDIA Tesla GPUs. CUDA based system is capable of processing Techniscan algorithm two time faster. Once the appliance obtain operational permit from the government, then patients will know his or her cancer detection results within one visit [5]. On the limited cluster resources that one have and current development of a pplications which run nicely on GPU; motivate us to study the performance of molecular dynamic simulation run on two different computing environment. This paper reports our experiments that evaluate the performance of GROMACS that run on cluster computing resources and GPU based PC. 2.The experiment and results 132 IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, No 2, July 2011 ISSN (Online): 1694-0814 www.IJCSI.org We choose four different types of GPU, namely GeForce GTS 450, GeForce GTX 465, GeForce GTX 470, and Quadro 4000 ( For simplicity, the phrase GeForce will not be mentioned af ter this). Their specifications are in the following table Table 1. The specification of PC installed with GPU GTX 470 GTX Quadro GTS Description Cuda Cores 448 465 352 4000 256 250 128 Memory 1280 MB 1024 2GB 1 GDDR5 MB GDDR5 DDR3 1674 1603 MHz MHz 320 Bit 256 Bit 256 Bit 256-bit 133.9 102.6 89,6 70.4 GB/sec GB/sec GB/sec GB GDDR 5 Mem. Clock Mem. interface 1100 Width Mem.Bandwidth We install GTS 250 into a PC that based on Intel® Pentium ( R ) 4 C PU 3.20 GHz, 4 G B RAM, a 80 G B SATA HDD. While the other GPUs, GTX 465, GTX 470 and Quadro 4000 a re installed into PCs with their specification in the following Table 2. We also built 16 cores Cluster computing environment from four PCs with these specification. Table 2. The specification of PC installed with GPU INTEL Core i5 760 Processor for Desktop Quad Core, 2.8GHz, 8MB Cache, Socket LGA1156 Memory DDR3 2x 2GB, DDR3, PC-10600 Kingston Main board ASUS P7H55M-LX Thermaltake LitePower 700Watt DVD±RW Internal DVD-RW, SATA, Black, 2MB, 22x DVD+R Write SAMSUNG/LITEON Thermaltake VH8000BWS Armor+ MX WESTERN DIGITAL Caviar Black HDD Desktop 3.5 inch SATA 500GB, 7200RPM, SATA II, 64MB Cache, 3.5", include SATA cables and mounting screws We installed SDK CUDA 2.3 and s ome application on our GPU PCs, such as GROMACS 4.0.5 [13] which run on GPU but not fully using processors on the GPU for parallel computation. GROMACS is a program that is designed to run serially or in parallel. The parallel features in GROMACS are provided to speed up the simulation processes. GROMACS has parallel algorithm related to domain decomposition. The domain decomposition algorithm is useful to solve boundary value problems by decomposing t he main problem into some smaller boundary value problems. In GROMACS, the domain decomposition algorithm divides the simulation works into different processors so that the processes can be done in parallel. Each of the processor will collect and coordinate the motion of simulated particles. 133 In the parallel simulation processes, processors will communicate to each others. Workload imbalance is caused by different particle distribution on each of the processors as well as the different particle interaction among the processors. In order to have work balance, GROMACS uses dynamic load balancing algorithm where the volume of each domain decomposition can be adjusted independently . The mdrun script of GROMACS will automatically start dynamic load balancing when there is instability energy computation within 5 percent or more. Load imbalance is recorded on output log created by GROMACS. As there is no test performance record yet for GROMACS 4.5 [13] running on Quadro 4000, so when a user runs MD simulation, a warning message appear saying that the graphic card used has not been tested to use GROMACS 4.5 software. A further investigation is necessary in order to see how GROMACS can run nice much better on Quadro 4000. In the experiment, we use two different compounds. Breda Virus 1 abbreviated as BRV-1 or 1 br v (PDB id 1BRV) [14]. Its scientific name is B ovine respiratory syncytial virus [Biochemistry, 1996, 35 (47), pp 14684– 14688]. The second one is REM2 (PDB id: 3CBQ). REM2 is the molecular function induces angiogenesis (essential for growth, invasion and metastasis tumor). So REM2 a potential target to overcome condition of angiogenesis. REM2 was known to regulate p53 to the nature of immortal in somatic cells. REM2 also cause the stem cells immortal, REM2 alleged role in the mechanism self-renewal in stem cells to hESC (human Embryonic Stem Cell which protects from the process of apoptosis (programmed cell death). [12] In the following table 3, we provide the time required in the simulation of 1BRV compound in various values of time steps Table 3. The time of simulation processes of 1 BRV Time steps Resources 200ps 400ps 600ps 800ps 1000ps Cluster 16 39m:22s 1h:19m:13s 1h:59m:32s 2h:39m:37s 3h:19m:08s 23m:26s 46m:57s 1h:10m:23s 1h:33m:58s 1h:57m:18s 6m:45s 13m:35s 20:23 27m:05s 33m:29s 8m:48s 8m:32s 12m:46s 16m:57s 21m:21s 3m:30s 7m:00 10m:29s 14m:07s 17m:14s GTS 250 Quaddro 4000 GTX 465 GTX 470 It is obvious that the experiments with GTX 470 outperform simulations in other environment. For example, in 1000 ps time steps with GTX 470, it requires 17 minutes and 14 second to finish the simulation, while with IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, No 2, July 2011 ISSN (Online): 1694-0814 www.IJCSI.org GTS 250, it requires 1 h our 27 m inutes and 32 s econd which is almost seven times of GTX 470. The performance degrades from GTX 470, GTX 465, Quadro 4000, and to GTS 250. This perhaps that each of these GPUs has different number of GPU cores, 448, 352, 256 and 128 cores respectively. The interesting one is that all the experiment on GPU outperforms those of Cluster 16. about 25 percent that of Cluster 16 pr ocessors. Even though our preliminary findings are interesting, but it needs further investigation as the development of Molecular Dynamics codes on GPU is still in progress. We next simulate REM2, and the time spent in simulation is provided in the following table 4. It is also obvious that the similar pattern happen in this simulation that GTX 470 performs best, and all GPU outperform those of Cluster 16. This research is supported by Universitas Indonesia Multi Discipline Research Grant in the year 2010. The Final stage of the writing process of the paper was conducted when the first author visited the School of ITEE – the University of Queensland. Table 4 The time of simulation processes of REM2 Acknowledgments References Time steps Resources 200ps 400ps 600ps 800ps 1000ps Cluster 16 45m:21s 1h:27m:27s 2h:13m:05s 2h:57m:01s 3h:43m:15s GTS 250 27m:56s 56m:11s 1h:24m:26s 1h:52m:23s 2h:23m:32s 4000 6m:45s 13m:35s 20m:23s 27m:05s 33m:29s GTX 465 4m:13s 8m:32s 12m:46s 16m:57s 21m:21s GTX 470 3m:30s 7m:00s 10m:29s 14m:07s 17m:14s Quaddro We finally provide the speed up of GTX 470 compare with Cluster 16, the following table 5 summarizes the result from each simulation of 1BRV (SU-1BRV) a nd REM2 (SU-REM2). Table 5. Speed up of GTX 470 relative to Cluster 16 SU\tsteps 200ps 400ps 600ps 800ps 1000ps SU-REM2 12.95714 12.49286 12.69475 12.53955 12.95455 SU-1BRV 11.24762 11.31667 11.40223 11.30697 11.55513 It is obvious that GTX470 gain speed up at about 11 to Cluster16 on 1BRV simulation, and about 12 on REM2 simulation. In term of hardware price, the PC with GTX 470 is about 25 percent that of Cluster 16 with the same specification. 3. Conclusion Molecular dynamics simulation using BRV-1 and REM2 is conducted on t wo different computing environment, cluster computing o f 16 nodes and GPU computing of various types of NVIDIA – GTS 250, GTX 465, GTX 470 and Quadro 4000. Both cluster and GPU computing has the same hardware specification, except that of GTS 250. The experiment show the efficacy of GTX 470 compare to the others. It gain speed up around 11 and 12, the cost of this GTX 470 computer is only [1] Luebke, David, The Democratization of Parallel computing: High Performance Computing with CUDA, t he International Conference for High Performance Computing, Networking, Storage and Analysis, 2007, http://sc07.supercomputing.org/ [2] de Vries, A.H., A. E. Mark, and S. J. Marrink Molecular Dynamics Simulation of the Spontaneous Formation of a Small DPPC Vesicle in Water in Atomistic Detail, J. Am. Chem. Soc. 2004, 126, 4488-448 [3] Buck, Ian, Cuda Programming, the International Conference for High Performance Computing, Networking, Storage and Analysis, 2007, http://sc07.supercomputing.org/ [4] Fatica, Massimiliano, CUDA Libraries, the International Conference for High Performance Computing, Networking, Storage and Analysis, 2007, http://sc07.supercomputing.org/ [5] Cuda Medicine, Aplikasi Medicine, http://www.nvidia.co.uk/object/cuda_medical_uk.html [akses 13 Feb 2010] [6]Karplus, M. & J. Kuriyan. Molecular Dynamics and Protein Function. PNAS, 2005. 102 (19): 6679-6685 [7]Spoel DVD, Erick L, Berk H. Gerit G, Alan EmM & Herman JCB., Gomacs: Fast, Flexible and Free., J. Comput Chem, 2005, 26(16): 1701-1707 [8]Adcock SA dan JA McCammon. Molecular Dynamics: Survey Methods for Simulating tha Activity of Protein. Chem Rev 2006. 105(5):1589-1615 [9]Correll,RN., Pang C, Niedowicz, DM, Finlin, BS and. Andres, DA., The RGK family of GTP-binding Proteins: Regulators of Voltage-dependent Calcium Channels and Cytoskeleton Remodeling [10]Kutzner, C, D. Van Der Spoel, M Fechner, E Lindahl, U W. Schmitt, B L. De Groot, H Grubmüller, Speeding up parallel GROMACS on high-latency networks J. C omp. Chem. 2007. 28(12): 2075-2084. [11]J.C. Phillips, R. Braun,W.Wang, J. Gumbart, E. Tajkhorshid, E. Villa, C. Chipot, R.D. Skeel, L. Kale, K. Schulten, Scalable molecular dynamics with NAMD, J. Comp. Chem. 26 (2005) 1781–1802. [12]Ruben Bierings, Miguel Beato, Michael J. Edel, an Endothelial Cell Genetic Screen Identifies the GTPase Rem2 as a Suppresor of P19 Expression that Promotes Endothelial Cell Proliferation and Angiogenesis, the 134 IJCSI International Journal of Computer Science Issues, Vol. 8, Issue 4, No 2, July 2011 ISSN (Online): 1694-0814 www.IJCSI.org Journal of Biological Chemistry, vol 283, no. 7, pp. 4408 – 4416 [13]GROMACS, http://www.gromacs.org [14][http://www.uniprot.org/taxonomy/360393]. Heru Suhartanto is a Professor in Faculty of Computer Science, Universitas Indonesia (Fasilkom UI). He has been w ith Fasilkom UI since 1986. Previously he hel d some positions such as Post doctoral fellow at Advanced Computational Modelling Centre, the University of Queensland, Australia in 1998 – 2000; two periods vice Dean for General Affair at Fasilkom UI since 2000. He graduated from undergraduate study at Department of Mathematics, UI in 1986. He holds Master of S cience, from Department of Computer Science, The University of Toronto, Canada since 1990. He also holds Ph.D in Parallel Computing from Department of Mathematics, The University of Queensland since 1998. His main research interests are Numerical, Parallel, Cloud and Grid computing. He is also a member of reviewer of several referred international journal such as journal of Computational and Applied Mathematics, International Journal of Computer Mathematics, and Journal of Universal Computer Science. Furthermore, he ha s supervised some Master and PhD students; he ha s won some research grants; holds several software copyrights; published a number of boo ks in Indonesian and i nternational papers in proceeding and j ournal. H e is also member of IEEE and ACM. Arry Yanuar is an as sistant Professor of Department Pharmacy, Universitas Indonesia since 1990. He graduated from undergraduate program Department of Pharmacy, Universitas Indonesia in 1990. He also holds Apoteker Profession certificate in 1991. In 1997, he finished his Master Program from Department of Pharmacy, Universitas Gadjah Mada. He holds PhD in 2006 from Nara Institute of Science and Technology (NAIST), Jepang, with Structure Biology/protein Cristalography laboratorium. In 19992003 he w orked as pharmacy expert ini ISO certification for pharmacy industries at Llyod Register Quality Assurance. In 2002, he visited National Institute of Health (NIH), Bethesda, USA. He won several research grants and published some paper at international journals and conferences. Ari Wibisono is a r esearch Assistant at Faculty of Computer Science Universitas Indonesia (Fasilkom UI). He graduated from undergraduate program Fasilkom UI and currently takes Master program at Fasilkom UI. 135 

1 A Note on the Information-Theoretic-(in)Security of Fading Generated Secret Keys arXiv:1705.07533v1 [cs.IT] 22 May 2017 Robert Malaney∗ Abstract—In this work we explore the security of secret keys generated via the electromagnetic reciprocity of the wireless fading channel. Identifying a new sophisticated colluding attack, we explore the information-theoretic-security for such keys in the presence of an all-powerful adversary constrained only by the laws of quantum mechanics. Specifically, we calculate the reduction in the conditional mutual information between transmitter and receiver that can occur when an adversary with unlimited computational and communication resources places directionalantenna interceptors at chosen locations. Such locations, in principal, can be arbitrarily far from the intended receiver yet still influence the secret key rate. The fast generation in real-world communication systems of an information-theoretic-secure key remains an ongoing endeavor. Currently, almost all key generation systems being considered for commercial deployment scenarios are those based on the quantum mechanical properties of the information carriers. However, although great progress has been made over the years with respect to quantum key distribution (QKD) systems, significant practical challenges remain before their deployment becomes ubiquitous (see [1] for a recent review). The technical reasons for such a circumstance are many-fold, but they do give rise to a search for other sources of shared randomness that can be exploited for secret key generation. Transceivers connected via a wireless channel offer up one possibility - via purely classical means.1 Indeed, for many years it has been known that the random fading inherent in the general wireless environment (coupled with electromagnetic reciprocity) is a potential source of secret key generation (for recent reviews see [4–7]). Conditioned on the reasonable (and trivial to verify) assumption that Eve (the adversary) is not in the immediate vicinity (a few cm at GHz frequencies) of Bob (the legitimate receiver), it is often stated that fading can lead to information-theoretic-secure keys. Here we clarify this is not the case when Me → ∞, Me being the number of receiving devices held by Eve. An all-powerful adversary constrained only by nature herself is used in almost all security analyses of QKD systems [1] - and it is the protection from such an unearthly adversary that lends quantum-based key systems their acclaimed security status. Such acclaimed security is information-theoretic-secure (conditioned on the classical channel being authenticated) and remains in place irrespective of the computational resources or energy afforded to the adversary, even as Me → ∞. ∗ Robert Malaney is with the School of Electrical Engineering and Telecommunications at the University of New South Wales, Sydney, NSW 2052, Australia. email: r.malaney@unsw.edu.au 1 Such a classical route is particularly important since QKD directly over wireless channels in the GHz range is constrained to short distances [2, 3]. Here, we explore an attack by such an all-powerful Eve on wireless fading generated key systems. More specifically, we quantify how this Eve can place directional-antenna receivers (e.g. apertures, linear-arrays, phased-arrays, etc.) at multiple locations in real-word scattering environments, and in principal drive to arbitrary low levels the conditional mutual information between Bob and Alice (the transmitter). Practical scenarios invoking limited forms of the attack are discussed, showing (at the very least) how current fading-generated key systems are partially susceptible to the attack. We will take the term ‘information-theoretic-secure’ to specifically mean the following. Conditioned on some welldefined assumption (or restriction) on the system model, but independent of the capabilities of an adversary (other than being constrained by natural law), the key information accessible to an adversary can be driven to arbitrary small levels for increasing use of some system resource. Specifically, consider some series of observations of the random variables X = (X1 , X2 , . . . Xn ,), Y = (Y1 , Y2 , . . . Yn ,), and Z = (Z1 , Z2 , . . . Zn ,) of a shared random resource by Alice, Bob and Eve, respectively. We assume a scheme with unlimited message exchanges between Alice and Bob (available to Eve) whereby for some sufficiently large n, keys computed by Alice and Bob (KA and KB , respectively) are made to satisfy the following requirements for some ǫ > 0, (i) Pr (KA 6= KB ) ≤ ε, (ii) n−1 I (KA ; Z) ≤ ε, (iii) n−1 I (KA ) ≥ rK − ε, and (iv) n−1 log |C| ≤ n−1 H (KA ) + ε, where H (·) is the entropy, I (·;·) is the mutual information between two random variables, I (·; ·| ·) is the mutual information between two random variables conditioned on another random variable, |C| is the cardinality of the key’s alphabet (C), and rK is an achievable secret key rate of the scheme. In general, the secret key rate is not known, but an upper limit can be given by [8, 9] rK ≤ min (I (X; Y ) , I (X; Y | Z)), where the mutual information between X and Y conditioned on Z can be written I (X; Y | Z) = H(X, Z) + H(Y, Z) − H(Z) − H(X, Y, Z), where H (·, ·, . . .) is the joint entropy of a sequence of random variables.2 If we introduce the Kullback-Leibler information divergence between two probability mass functions p(w) and q(w) with sample space i W , viz., D ( pk q) = h P q(w) p(w) p(w) log q(w) = E − log p(w) then the conditional muw∈W tual information can be estimated via a discrete formulation, , p(x) being viz., I (X; Y | Z) = D p(x, y, z)k p(x,z)p(y,z) p(z) the probability X = x, and p (·, ·, . . .) the corresponding joint probability. 2 Here, I(X; Y | Z) ≤ I(X; Y ) for all our calculations. 2 channel coherence time (and therefore the key rate). The movement within the scattering environment ultimately manifests itself at the receiver through variation in received amplitudes and delays. However, to enable clarity of exposition, we will make some simplifications to our scattering model - noting that the attack we describe can in fact be applied to any scattering environment scenario. The simplifications are that we assume equal amplitudes for all received signals, and adopt random uniform distributions for all AoA and all phases as measured by the receiver. This is in fact the celebrated 2D isotropic scattering model of Clarke [11]. Moving to the baseband signal henceforth for convenience, we note in Clarke’s model the signal of a transmitted symbol can be N P written as g(t) = √1N exp (j (wd t cos αn + φn )), where n=1 Cn (t) sinϕn (t). In the Rayleigh channel these wd is now the maximum Doppler frequency in rad/s, and φn is the phase of each path. Assuming both αn and φn are independent and uniformly distributed in [−π, π), then in the limit of large N the amplitude of the signal g(t) is distributed as a Rayleigh distribution. Of particular interest to us here will be the statistics of g(t) at low values of N , since in such circumstances the potential for an adversary to intercept all signals is larger. The higher order statistics of the distribution within Clarke’s model at finite N have been explored in [12], showing that the following autocorrelation functions are in place; Υgg (τ ) = J0 (ωd τ ), J 2 (ω τ ) and Υ|g|2 |g|2 (τ ) = 1 + J02 (ωd τ ) − 0 Nd , where J0 (·) is the zero-order Bessel function of the first kind. For large N these functions approach those of the Rayleigh distribution. Importantly, the Υ|g|2 |g|2 function is well approximated by an exponentiated sinc function at values of N ≥ 6, meaning that (as per the usual assumption for any fading generated key), Eve must be several wavelengths away from Bob for the secret key rate to be non-zero.3 Many refinements on Clarke’s model exist with perhaps the most widely used being that of [12] in which the main differn , entiator is a constraint placed on the AoA, viz., αn = 2πn+θ N where θn is independently (relative to φn ) and uniformly distributed in [−π, π). This latter simulator is wide sense stationary, is more efficient, and has improved second-order statistics. In consideration of all statistical measures, it is noted that for this refined model any differences between N & 8 and the N = ∞ model (pure Rayleigh distribution) are largely inconsequential [12].4 In real-world channels, therefore, we have to be aware that even in cases where the channel appears to be consistent with a Rayleigh channel, the number of propagation paths contributing to the received signal can be relatively small. This can be seen more clearly from Fig. (1) where the probability density functions formed from six and five propagation paths quadratures q are independent Gaussian processes. Writing 2 2 |A| = rI (t) + rQ (t) and ϑ = arctan (rQ (t)/rI (t)) we then have r(t) = |A| cos (2πfc t + ϑ), where |A| is Rayleigh distributed and ϑ is uniformly distributed. In such a channel, |A| and/or ϑ can be used for secret key construction. Ultimately the secret key is dependent on movement in the scattering environment, and it this movement that sets the 3 A relaxation of this requirement may be obtained in specific correlated channel scenarios applicable to a distance of order 10 wavelengths (∼ meters at GHz frequencies) away from the receiver [13]. The attack we describe here is unrelated to the special case of correlated channels. It is a general attack. Eve’s receivers can in principal be positioned anywhere (e.g. kms away from the intended receiver) yet still mount a successful attack. 4 In the calculations to follow, we find that the key rates computed are, in effect, independent of whether this refined model or Clarke’s original model is the adopted Rayleigh simulator. Fig. 1. Probability density functions (pdf) for different path settings. The inset graph shows the pdf at Bob for an infinite number of paths (dashed), 6 paths (dot-dashed), and a colluding Eve who has perfectly intercepted (see text) 5 paths (solid). The sketch on left illustrates the nature of the colluding attack where the solid (black) lines indicate some of the potentially many rays towards Bob that Eve intercepts, and the (red) dashed line indicates one of the potentially many ‘interference’ paths to a directional antenna held by Eve. We adopt the narrow-band flat-fading channel, and take the far-field approximation with wave propagation confined to a plane geometry. We assume the electric field vector is orthogonal to the plane and that isotropic gain antennas are held by Alice and Bob. If we consider, at the carrier frequency fc (wavelength λc ), a bandpass transmitted signal s(t) = Re s̃(t)ej2πfc t , where s̃(t) is the complex envelope, the received signal can then be written  bandpass  (e.g. [10]), N P D [t−τn ]) j2π ([fc +fn ] r(t) = Re C e s̃(t − τ ) . Here N is n n n=1 the number of propagation paths reaching the receiver, Cn and τn are the amplitude and time delay, respectively, and fnD is the Doppler frequency (n indicates the nth path). This latter quantity can be expressed as fnD = (v/λc ) cos αn , v being the velocity of the receiver and αn being the angle of arrival (AoA) of the nth path at the receiver, relative to the velocity vector. Similar to the transmit signal, a complex envelope for the reN P ceived signal can be written, r̃(t) = Cn e−jϕn (t) s̃(t − τn ), n=1 
 where ϕn (t) = 2π fc + fnD τn − fnD t . Therefore, we have r(t) = Re r̃(t)ej2πfc t . In the case of a transmitted single tone this can be written as r(t) = rI (t) cos 2πfc t − N P rQ (t) sin 2πfc t, where rI (t) = Cn (t) cosϕn (t) and n=1 rQ (t) = N P n=1 3 are shown in comparison to the infinite path limit. The five path model corresponds to a case where Eve is missing one of the propagation paths used to construct Bob’s signal. For the cases shown the Kullback-Leibler divergence between the Rayleigh distribution and the lower-path models is very small. Let us assume the communications obtained by Bob consist of the combined signals from N last-scattering events. We are interested in determining the effect, on some secret key generation scheme, caused by Eve’s interception of all (or some fraction of) the N last-scattered paths received by Bob. We assume Eve has Me >> N directional-antenna receivers, and has placed them at multiple locations with the aim of continuously intercepting all of the last-scattered signals towards Bob with high probability.5 We assume that these locations are much greater than λc from Bob. Beyond our assumption of 2D geometry, and that the amplitude of each last-scattered ray entering any receiver is equal, we also assume that the number of paths reaching each of Eve’s antennas is equal to N .6 Extension of our analysis to cover these issues is cumbersome, but straightforward. To make our mathematical notation less unwieldy, we will artificially set Me = N in our equations, with the understanding that we are then simply ignoring all of Eve’s devices which (at any given time) are not intercepting any scattered rays towards Bob. For added focus, we will assume Eve uses circular apertures of diameter d as her directional receivers - the physics and properties of which can be found elsewhere, e.g. [14]. Eve configures her nth aperture at each location so as to maximize signal gain for the signal directed by the last scatterer in the direction of Bob (i.e. the nth of N rays reaching Bob is centered in the main lobe of Eve’s nth aperture). In such circumstances the signal collected by Eve’s nth receiver can be approximated as,   exp  (j (wde t + φen )) +        1 exp (j (wde t cos αek + φek )) × N P gcn (t) = √    2λc J1 ( λπd sin(βke )) N  c     k=2 πd sin(βke ) where the superscript e applied to any previously used variable means it is now applied to Eve (but same meaning), where βke represents the angle between the kth propagation path (side lobe ‘interference’ path) arriving at Eve’s detector and ray n (i.e. βne = 0), and where J1 (·) is the Bessel function of the first kind of order one. Note that the maximum Doppler shift wde on Eve’s detector is included so as to cover the general case. However, for focus we will assume all of Eve’s detectors are stationary, and in the following always set wde = 0. To reduce the mathematical complexity further we have not 5 Such a possibility can be enhanced in some scenarios by additional actions on Eve’s part. For example, a scenario in which Eve has conducted an a priori ray-tracing measurement (or analysis) campaign between a given pair of transmit and receive locations thereby obtaining probabilistic information on likely last scattering points (for that given pair of locations). Of course in the limit ME → ∞ her probability of intercepting all paths approaches one in any case. 6 As an aside, we find a doubling of this number of paths at each of Eve’s detectors has negligible impact on the results. Also note, as the number of paths reaching Eve approach infinity, the size of her aperture must be made to approach infinity for the attack to remain viable. Neither limits are ever in place of course. included in our analysis an obliquity factor (1 + cos βke ) /2, which makes our calculations conservative (i.e. results in higher key rates). Upon receipt of the signals gcn (t) Eve will adjust the signals for the known distance offset between each detector and Bob, and the known motion of Bob. This entails a phase adjustment at each detector which manifests itself as an ‘adjusting’ phase φna . The combined adjusted signal obtained by Eve after such signal processing, can then be written as N P g(t) = gcn (t) exp(jφna ). n=1 Assuming Eve’s different apertures intercept all paths that are received by Bob, the above relations lead us to conclude that, in principle, by increasing her aperture size Eve can determine Bob’s received signal to arbitrary accuracy. In practice this accuracy will be limited by any receiver noise on Eve’s antennas, and error due to imprecise location information on Bob. However, with regard to these accuracy limitations (which we include in our Monte Carlo simulations below), we note the following two points that favor Eve. (i) Given her all-powerful status, Eve can set her noise to be at the quantumlimit (quantum noise). (ii) Beyond any other means available to her, an unlimited Eve can determine the location of Bob at time t to any required accuracy through signal acquisition. More specifically in regard to (ii), the √ minimum position error via signal processing varies as 1/ Me - a result that holds even if some of Eve’s devices are affected by shadowing in which the path-loss exponents are unknown [15]. To make further progress we must introduce an actual scheme for generating a secret key. Although there are many such schemes (e.g. [4–7]) we will adopt here a generic formulation that covers the conceptual framework of the widely used signal threshold schemes. The basic concept of such schemes is to quantize a received signal metric, say amplitude, into a 1 or 0 value. For some parameter T > 0, and for some median value m of the expected amplitude distribution, the decision value can then be set dependent on whether the amplitude is below m − T or above m + T . Such schemes offer many pragmatic advantages and compensate to a large extent errors introduced through a lack of exact reciprocity between transceiver configurations. Assuming a given level of Gaussian noise at Bob and Eve’s receivers, an appropriate value of T can be chosen. Further, so as to maximize the entropy of the final key, we introduce an ‘entropy’ factor s. For a given T and probability density function R′ (r) for the received of s can be determined through R m−T ′amplitudeRr,∞the value ′ R (r)dr = R (r)dr. Note, in general R′ is the 0 m+T +s distribution for the amplitudes in the presence of non-zero Gaussian receiver noise. When r is measured by Alice and/or Bob to be between the two ‘allowed’ regions, as defined by the integrals of this relation, it is agreed by both parties that the measurement be dropped. Clearly, in practice larger values of T will minimize mismatches in the key at the cost of a reduced key generation rate. Ultimately, in any real scheme a period of reconciliation and privacy amplification will be pursued in order to obtain the final key. However, here we will simply investigate the upper limit of the key rate through a numerical evaluation 4 I(X;Y| Z) 1 Actual Paths=6 & Adversary Paths=4 Actual Paths=6 & Adversary Paths=5 Actual Paths=6 & Adversary Paths=6 0.5 0 0 0.1 0.2 0.3 0.4 0.5 0.6 Diameter (m) I(X;Y| Z) 1 Actual Paths=20 & Adversary Paths=15 Actual Paths=20 & Adversary Paths=19 Actual Paths=20 & Adversary Paths=20 0.5 0 0 0.1 0.2 0.3 0.4 0.5 0.6 Diameter (m) I(X;Y| Z) 1 Actual Paths=20 & Adversary Paths=15 Actual Paths=20 & Adversary Paths=19 Actual Paths=20 & Adversary Paths=20 0.5 0 0 0.1 0.2 0.3 0.4 0.5 0.6 Diameter (m) Fig. 2. Change in the conditional mutual information between Alice and Bob as function of the diameter of Eve’s directional antenna (a circular aperture) for different path conditions. Six paths (top figure) and 20 paths (middle figure) are used to construct the approximate Rayleigh distribution. One calculation (bottom figure) on the 20 path scenario assumes zero receiver noise at Eve and zero location error on Bob. Results shown are for 1 million Monte Carlo runs. of the conditional mutual information as defined earlier. We assume Eve’s strategy on detection is to decide on the binary number in the ‘disallowed’ region by setting s = T = 0. We also assume all issues on the decision strategy of the scheme and all communications between Alice and Bob (e.g. which measurements to drop) are available to Eve. Fig. (2) (top) displays a calculation of the conditional mutual information as a function of aperture diameter (all of Eve’s circular apertures are assumed to be the same size) in which a receiver noise contribution (on all receivers) is set so that the signal-to-noise ratio (SNR) is equal to 17dB. The maximum Doppler shift of Bob is set to 10Hz, λc is set to 0.1m, and a Gaussian error on the pointing of Eve’s apertures (due to location error on Bob) is set to a standard deviation of 0.002 radians. The threshold is set at three times the receiver noise. We can see that if all signals are intercepted the key rate can be driven to almost zero over the range of aperture diameters probed. For fewer signals intercepted we see that useful key rates are still possible, albeit at significantly diminished values relative to a no-attack scenario. For comparison, Fig. (2) (middle) displays similar calculations but for 20 propagation paths forming the Rayleigh distribution, and Fig. (2) (bottom) shows the same calculation when Eve’s detectors are operating with zero receiver noise, and location errors on Bob are assumed to be zero. The specific key scheme discussed here is limited in scope relative to the large number of possible key generation schemes available. More sophisticated schemes, such as those based on multi-antenna transceiver configurations, the use of optimal coding techniques, and the use of channel state information, are possible. However, straightforward extensions of the attack described here would still apply to all of these more sophisticated schemes - only the quantitative details on how the key rate is diminished under the attack will be different. Indeed, we note the attack described here can be generalized further so as to always drive the secret key rate to zero, even if we relax the assumption that it is only the last-scattering rays that are intercepted. An all-powerful Eve, with ME → ∞, can intercept all propagation paths (of any energy) at all points in space, and in principal possess knowledge on all characteristics of all scatterers. With the unlimited computational resources afforded to her the classical Maxwell equations can then be solved exactly, thereby providing information on any of Bob’s received signals at an accuracy limited only by quantum mechanical effects. Of course, such an attack whilst theoretically possible, is not tenable. The calculations described here can be considered a limited form of such an attack, tenable in a real-world scattering environment. In conclusion, we have described a new attack on classical schemes used to generate secret keys via the shared randomness inherent in wireless fading channels. Although the attack we have described will be difficult to implement in a manner that drives the secret key rate to zero, our work does illustrate how such a rate can at least be partially reduced. As such, all schemes for secret key generation via the fading channel must invoke a new restriction - a limitation on the combined information received by a colluding Eve. [1] E. Diamanti, H.-K. Lo, B. Qi, and Z. Yuan, “Practical challenges in quantum key distribution,” npj Quantum Information, 2, 16025, 2016. [2] C. Weedbrook, S. Pirandola, S. Lloyd, and T. C. Ralph, “Quantum cryptography approaching the classical limit,” Phys. Rev. Lett. 105, 110501, 2010. [3] N. Hosseinidehaj and R. Malaney, “Quantum entanglement distribution in next-generation wireless communication systems,” IEEE VTC Spring, Sydney, Australia (arXiv:1608.05188), 2017. [4] L. Lai, Y. Liang, H. V. Poor, and W. Du, “Key generation from wireless channels,” Physical Layer Security in Wireless Comms., (CRC), 2014. [5] P. Narayan and H. Tyagi, “Multiterminal secrecy by public discussion,” Found. and Trends in Comms. and Info. Theory,” 13, pp. 129–275, 2016. [6] J. Zhang, T. Q. Duong, A. Marshall, and R. Woods, “Key generation from wireless channels: A review,” IEEE Access, 10.1109/Access.2016.2521718, 2016. [7] H. V. Poor and R. F. Schaeferb, “Wireless physical layer security,” PNAS, vol. 114, no. 1, pp. 19-26, 2017. [8] U. Maurer, “Secret key agreement by public discussion from common information,” IEEE Trans. Inf. Theory, vol. 39, no. 3, pp. 733-742, 1993. [9] R. Ahlswede and I. Csiszar,“Common randomness in information theory and cryptography-Part I: Secret sharing” IEEE Trans. Inf. Theory, vol. 39, no. 4, pp. 1121-1132, 1993. [10] G. L. Stüber, “Principles of Mobile Communication,” (Kluwer) 2002. [11] R. H. Clarke, “A statistical theory of mobile-radio reception” Bell Syst. Tech. J., pp. 957-1000, Jul.-Aug. 1968. [12] C. Xiao,Y. R. Zheng, and N. C. Beaulieu, “Novel sum-of-sinusoids simulation models for Rayleigh and Rician fading channels” IEEE Transactions on Wireless Communications, vol. 5, no. 12, 2006. [13] X. He, H. Dai, W. Shen, and P. Ning,“Is link signature dependable for wireless security,” in Proc. IEEE Infocom, Turin, Italy, pp. 200–204, 2013. [14] S. J. Orfanidis, “Electromagnetic Waves and Antennas,” 2016. [15] R. Malaney, “Nuisance parameters and location accuracy in log-normal fading models,” IEEE Trans. Wireless Communications, vol 6, issue 3, pp. 937–947, 2007. 

arXiv:1802.05541v1 [cs.CE] 30 Jan 2018 Novel weak form quadrature elements for non-classical higher order beam and plate theories Md Ishaquddin∗, S.Gopalakrishnan† Department of Aerospace Engineering, Indian Institute of Science Bengaluru 560012, India Abstract Based on Lagrange and Hermite interpolation two novel versions of weak form quadrature element are proposed for a non-classical Euler-Bernoulli beam theory. By extending these concept two new plate elements are formulated using Lagrange-Lagrange and mixed Lagrange-Hermite interpolations for a non-classical Kirchhoff plate theory. The non-classical theories are governed by sixth order partial differential equation and have deflection, slope and curvature as degrees of freedom. A novel and generalize way is proposed herein to implement these degrees of freedom in a simple and efficient manner. A new procedure to compute the modified weighting coefficient matrices for beam and plate elements is presented. The proposed elements have displacement as the only degree of freedom in the element domain and displacement, slope and curvature at the boundaries. The Gauss-Lobatto-Legender quadrature points are considered as element nodes and also used for numerical integration of the element matrices. The framework for computing the stiffness matrices at the integration points is analogous to the conventional finite element method. Numerical examples on free vibration analysis of gradient beams and plates are presented to demonstrate the efficiency and accuracy of the proposed elements. Keywords: Quadrature element, gradient elasticity theory, weighting coefficients, non-classical dofs, frequencies, mixed interpolation ∗ † Corresponding author: E-mail address: ishaquddinmd@iisc.ac.in E-mail address: krishnan@iisc.ac.in; Phone: +91-80-22932048 1 1.0 INTRODUCTION In recent decades the research in the field of computational solid and fluid mechanics focused on developing cost effective and highly accurate numerical schemes. Subsequently, many numerical schemes were proposed and applied to various engineering problems. The early research emphasized on the development of finite element and finite difference methods [1–3], these methodologies had limitations related to the computational cost. Alternatively, differential quadrature method (DQM) was proposed by Bellman [4] which employed less number of grid points. Later, many enriched versions of differential quadrature method were developed, for example, differential quadrature method [5–10], harmonic differential quadrature method [11, 12], strong form differential quadrature element method (DQEM) [13–19], and weak form quadrature element method [20–23]. The main theme in these improved DQ versions was to develop versatile models to account for complex loading, discontinuous geometries and generalized boundary conditions. Lately, much research inclination is seen towards the strong and weak form DQ methods due their versatality [13–23]. The strong form differential quadrature method which is built on governing equations, require explicit expressions for interpolation functions and their derivatives, and yield unsymmetric matrices. In contrast, the weak form quadrature method is fomulated using variation principles, and the weighting coefficients are computed explicitly at integration points using the DQ rule, leading to symmetric matrices. The aforementioned literature forcussed on developing DQ schemes for classcial beam and plate theories which are governed by fourth order partial differential equations. The DQ solution for the sixth and eighth order differential equations using GDQR technique is due to Wu et al. [24, 25]. In their research, they employed strong form of governing equation in conjunction with Hermite interpolation function to compute the weighting coefficients and demonstrated the capability for structural and fluid mechanics problems. Recently, Wang et al. [26] proposed a strong form differential quadrature element based on Hermite interpolation to solve a sixth order partial differential equation associated with a non-local Euler-Bernoulli beam. The capability of the element was demonstrated through free vibration analysis. In this article the main focus is to propose a weak form quadrature beam and plate element for non-classical higher order theories, which are characterized by sixth order partial differential equations. As per the authors knowledge no such work is reported in the literature till date. The non-classical higher order theories unlike classical continuum theories are governed by sixth order partial differential equations [27–32]. These nonclassical continuum theories are modified versions of classical continuum the2 ories incorporating higher order gradient terms in the constitutive relations. The higher order terms consists of stress and strain gradients accompanied with intrinsic length which accounts for micro and nano scale effects [27–32]. These scale dependent non-classical theories are efficient in capturing the micro and nano scale behaviours of structural systems [29–31]. One such class of non-classical gradient elasticity theory is the simplified theory by Mindlin 0 et al. [29], with one gradient elastic modulus and two classical lame constant for structural applications [32–34]. This simlified theory was applied earlier to study the static, dynamic and buckling behaviour of gradient elastic beams [35–37] and plates [38–40] by developing analytical solutions. Pegios et al. [41] developed a finite element model for static and stability analysis of gradient beams. The numerical solution of 2-D and 3-D gradient elastic structural problems using finite element and boundary element methods can be found in [42]. In this paper, we propose for the first time, two novel versions of weak form quadrature beam elements to solve a sixth order partial differential equation encountered in higher order non-classical elasticity theories. The two versions of quadrature beam element are based on Lagrange and C 2 continuous Hermite interpolations, respectively. Further, we extend this concept and develop two new types of quadrature plate elements for gradient elastic plate theories. The first element employs Lagrange interpolation in x and y direction and second element is based on Lagrange-Hermite mixed interpolation with Lagrange interpolation in x and Hermite in y direction. These elements are formulated with the aid of variation principles, differential quadrature rule and Gauss Lobatto Legendre (GLL) quadrature rule. Here, the GLL points are used as element nodes and also to perform numerical integration to evaluate the stiffness and consistent mass matrices. The proposed elements have displacement, slope and curvature as the degrees of freedom at the element boundaries and only displacement in the domain. A new way to incorporate the non-classical boundary conditions associated with the gradient elastic beam and plate theory is proposed and implemented. The novelty in the proposed scheme is the way the classical and non-classical boundary conditions are represented accurately and with ease. It should be noted that the higher order degrees of freedom at the boundaries are built into the formulation only to enforce the boundary conditions. The paper is organized as follows, first the theoretical basis of gradient elasticity theory required to formulate the quadrature elements is presented. Next, the quadrature elements based on Lagrange and Hermite interpolations functions for an Euler-Bernoulli gradient beam are formulated. Later, the formulation for the quadrature plate elements are given. Finally, numerical results on free vibration analysis of gradient beams and plates are 3 presented to demonstrate the capability of the proposed elements followed by conclusions. 1 Strain gradient elasticity theory In this study, we consider Mindlin’s [29] simplified strain gradient microelasticity theory with two classical and one non-classical material constants. 0 The two classical material constants are Lame constants and the non-classical one is related to intrinsic bulk length g. The theoretical basis of gradient elastic theory required to formulate the quadrature beam and plate elements are presented in this section. 1.1 Gradient elastic beam theory The stress-strain relation for a 1-D gradient elastic theory is given as [35, 43] τ = 2 µ ε + λ trε I ς = g 2 [2 µ ∇ε + λ ∇(trε) I] (1) 0 ∂ ∂ + ∂y is the Laplacian operator and I where λ, µ are Lame constants.∇ = ∂x is the unit tensor. τ , ς denotes Cauchy and higher order stress respectively, ε and (tr ε) are the classical strain and its trace which are expressed in terms of displacement vector w as 1 ε = (∇w + w∇) , 2 trε = ∇w (2) From the above equations the constitutive relations for an Euler-Bernoulli gradient beam can be defined as τx = Eεx , 0 ςx = g 2 εx , εx = −z ∂ 2 w(x, t) ∂x2 (3) For the above state of stress and strain the strain energy in terms of displacements for a beam defined over a domain −L/2 ≤ x ≤ L/2 can be written as [43] Z  00  1 L/2 000 U= EI (w )2 + g 2 (w )2 dx (4) 2 −L/2 The kinetic energy is given as 1 K= 2 Z t1 t0 Z L/2 ρAẇ2 dxdt −L/2 4 (5) where E, A, I and ρ are the Young’s modulus, area, moment of inertia, and density, respectively. w(x, t) is transverse displacement and over dot indicates differentiation with respect to time. Using the The Hamilton’s principle [45]: Z t1 (U − K) dt = 0 δ (6) t0 we get the following weak form expression for elastic stiffness matrix ‘K’ and consistent mass matrix ‘m’ as Z L/2  000 000  K= EI w00 δw00 + g 2 w δw dx (7) −L/2 Z L/2 ˙ dx ρA ẇ δw m= (8) −L/2 The governing equation of motion for a gradient elastic Euler-Bernoulli beam is obtained as EI(wiv − g 2 wvi ) + mẅ = 0 (9) The above sixth order equation of motion yields three independent vari0 00 ables related to deflection w, slope w and curvature w and six boundary conditions in total, as given below Classical boundary conditions : 000 V = EI[w − g 2 wv ] = 0 or w = 0, at x = (−L/2, L/2) 00 0 M = EI[w − g 2 wiv ] = 0 or w = 0, at x = (−L/2, L/2) (10) Non-classical boundary conditions : 000 00 M̄ = [g 2 EIw ] = 0 or w = 0, at x = (−L/2, L/2) (11) where V , M and M̄ are shear force, bending moment and higher order moment, respectively. 5 1.2 Gradient elastic plate theory The strain-displacement relations for a Kirchhoff’s plate theory are defined as [46] εxx = −z w̄xx , εyy = −z w̄yy , γxy = −2z w̄xy (12) where w̄(x, y, t) is transverse displacement of the plate. The stress-strain relations for a gradient elastic Kirchhoff plate are given by [31, 43]: Classical: E (εxx + νεyy ) 1 − ν2 E (εyy + νεxx ) = 1 − ν2 E = εxy 1+ν (13) E ∇2 (εxx + νεyy ) 1 − ν2 E ∇2 (εyy + νεxx ) =g 2 1 − ν2 E ∇2 εxy =g 2 1+ν (14) τxx = τyy τxy Non-classical: ςxx =g 2 ςyy ςxy where τxx , τyy ,τxy , are the classical Cauchy stresses and ςxx ,ςyy , ςxy denotes higher order stresses related to gradient elasticity. The strain energy for a gradient elastic Kirchhoff plate is gven by [31, 40] Up = Ucl + Usg (15) where Ucl and Usg are the classical and gradient elastic strain energy given by Z Z h i 1 2 2 2 2 w̄xx + w̄yy + 2w̄xy + 2 ν (w̄xx w̄yy − w̄xy ) dxdy (16) Ucl = D 2 A Usg 1 = g2D 2 Z Z h 2 2 2 2 w̄xxx + w̄yyy + 3(w̄xyy + w̄xxy ) A i 2 2 − w̄xxy dxdy + 2 ν (w̄xyy w̄xxx + w̄xxy w̄yyy − w̄xyy 6 (17) where, D = Eh3 . 12(1−ν 2 ) The kinetic energy is given by Z 1 K= 2 ρ h w̄˙ 2 dx dy (18) Using the The Hamilton’s principle: Z t1 (U − K) dt = 0 δ (19) A t0 we obtain the following expression for elastic stiffness and mass matrix for a gradient elastic plate Elastic stif f ness matrix : [K] = [K]cl + [K]sg (20) where [K]cl , [K]sg are classical and non-classical elastic stiffness matrix defined as Z h [K]cl =D w̄xx δ w̄xx + w̄yy δ w̄yy + 2w̄xy δ w̄xy + A i ν (δ w̄xx w̄yy + w̄xx δ w̄yy − 2w̄xy δ w̄xy ) dxdy (21) 2 [K]sg =g D Z h w̄xxx δ w̄xxx + w̄yyy δ w̄yyy + 3(w̄xyy δ w̄xyy + A w̄xxy δ w̄xxy ) + ν (w̄xyy δ w̄xxx + w̄xxx δ w̄xyy + i w̄xxy δ w̄yyy + w̄yyy δ w̄xxy − 2 w̄xyy δ w̄xyy − 2 w̄xxy δ w̄yxx ) dxdy (22) Consistent mass matrix : Z [M ] = ρ h w̄˙ δ w̄˙ dx dy (23) A The equation of motion for a gradient elastic Kirchhoff plate considering the inertial effect is obtained as: D∇4 w̄ − g 2 D∇6 w̄ + ρh 7 ∂ 2 w̄ =0 ∂t2 (24) where, ∂ 4 w̄ ∂ 4 w̄ ∂ 4 w̄ + + 2 , ∂x4 ∂y 4 ∂x2 ∂y 2 ∂ 6 w̄ ∂ 6 w̄ ∂ 6 w̄ ∂ 6 w̄ ∇6 = + + 3 + 3 ∂x6 ∂y 6 ∂x4 ∂y 2 ∂x2 ∂y 4 ∇4 = the associated boundary conditions for the plate with origin at (0, 0) and domain defined over (−lx /2 ≤ x ≤ lx /2), (−ly /2 ≤ y ≤ ly /2), are listed below. Classical boundary conditions : ∂ 3 w̄ ∂ 3 w̄ + (2 − ν) Vx = −D ∂x3 ∂x∂y 2   ∂ 5 w̄ ∂ 5 w̄ ∂ 5 w̄ + (3 − ν) +3 2 3 =0 +g D ∂x5 ∂x∂y 4 ∂y ∂x or w̄ = 0 at x = (−lx /2, lx /2) ∂ 3 w̄ ∂ 3 w̄ + (2 − ν) Vy = −D ∂y 3 ∂y∂x2   ∂ 5 w̄ ∂ 5 w̄ ∂ 5 w̄ + (3 − ν) +3 2 3 =0 +g D ∂y 5 ∂y∂x4 ∂x ∂y or w̄ = 0, at y = (−ly /2, ly /2) (25)   2 2   ∂ 2 w̄ ∂ 2 w̄ Mx = −D + ν ∂x2 ∂y 2   ∂ 4 w̄ ∂ 4 w̄ ∂ 4 w̄ +g D + ν 4 + (3 − ν) 2 2 = 0 ∂x4 ∂y ∂x ∂y or w̄x = 0, at x = (−lx /2, lx /2) ∂ 2 w̄ ∂ 2 w̄ + ν My = −D ∂y 2 ∂x2   ∂ 4 w̄ ∂ 4 w̄ ∂ 4 w̄ +g D + ν 4 + (3 − ν) 2 2 = 0 ∂y 4 ∂x ∂x ∂y or w̄y = 0, at y = (−ly /2, ly /2) (26)   2 2   8 Non-classical boundary conditions :   3 ∂ w̄ ∂ 3 w̄ 2 = 0 or w̄xx = 0, at x = (−lx /2, lx /2) M̄x = −g D +ν ∂x3 ∂x∂y 2   3 ∂ w̄ ∂ 3 w̄ 2 = 0 or w̄yy = 0, at y = (−ly /2, ly /2) M̄y = −g D +ν ∂y 3 ∂y∂x2 (27) where lx and ly are the length and width of the plate. Vx ,Vy are the shear force, Mx ,My are the bending moment and M̄x ,M̄y are the higher order moment. The different boundary conditions employed in the present study for a gradient elastic Kirchhoff plate are Simply supported on all edges SSSS : w̄ = Mx = w̄xx = 0 at x = (−lx /2, lx /2) w̄ = My = w̄yy = 0 at y = (−ly /2, ly /2) Free on all edges FFFF : Vx = Mx = M̄x = 0 at x = (−lx /2, lx /2) Vy = My = M̄y = 0 at y = (−ly /2, ly /2) Simply supported and free on adjacent edges SSFF : w̄ = My = w̄yy = 0 at y = −ly /2 w̄ = Mx = w̄xx = 0 at x = lx /2 Vx = Mx = M̄x = 0 at x = −lx /2 Vy = My = M̄y = 0 at y = ly /2 for the SSFF plate at (−lx /2, −ly /2) and (lx /2, ly /2), w̄ = 0 condition is enforced. The above boundary conditions are described by a notation, for example, consider a SSFF plate, the first and second letter correspond to y = −ly /2 and x = lx /2 edges, similarly, the third and fourth letter correspond to the edges y = ly /2 and x = −lx /2, respectively. Further, the letter S, C and F correspond to simply supported, clamped and free edges of the plate. 2 Quadrature element for a gradient elastic Euler-Bernoulli beam Two novel quadrature elements for a gradient Euler-Bernoulli beam are presented in this section. First, the quadrature element based on Lagrangian 9 interpolation is formulated. Later, the quadrature element based on C 2 continuous Hermite interpolation is developed. The procedure to modify the DQ rule to implement the classical and non-classical boundary conditions are explained. A typical N-node quadrature element for an Euler-Bernoulli gradient beam is shown in the Figure 1. 𝑥, 𝜉 = 2𝑥 𝐿 𝑤1′′ 𝑤5′′ 𝑤1′ 𝑤1 𝑤2 𝑤4 𝑤3 𝑤5′ 𝑤5 𝐿 Figure 1: A typical quadrature element for a gradient Euler-Bernoulli beam. It can be noticed from the Figure 1, each interior node has only displacement w as degrees of freedom and the boundary has 3 degrees of freedom 0 00 w, w , w . The new displacement vector now includes the slope and curvature as additional degrees of freedom at the element boundaries given by: 0 0 00 00 wb = {w1 , · · · , wN , w1 , wN , w1 , wN }. The procedure to incorporate these extra boundary degrees of freedom in to the formulation will be presented next for Lagrange and C 2 continuous Hermite interpolation based quadrature elements. 2.1 Lagrange interpolation based quadrature beam element The displacement for a N-node quadrature beam is assumed as [10]: w(x, t) = N X Lj (x)wjb j=1 = N X L̄j (ξ)wjb (28) j=1 Lj (x) and L̄j (ξ) are Lagrangian interpolation functions in x and ξ co-ordinates respectively, and ξ = 2x/L with ξ ∈ [−1, 1]. The Lagrange interpolation 10 functions are defined as [7, 10] N Y β(ξ) (ξ − ξk ) Lj (ξ) = = β(ξj ) (ξj − ξk ) k=1 (29) (k6=j) where β(ξ) = (ξ − ξ1 )(ξ − ξ2 ) · · · (ξ − ξj−1 )(ξ − ξj+1 ) · · · (ξ − ξN ) β(ξj ) = (ξj − ξ1 )(ξj − ξ2 ) · · · (ξj − ξj−1 )(ξj − ξj+1 ) · · · )(ξj − ξN ) The first order derivative of the above interpolation function can be written as,  N N Y Y    (ξi − ξk )/ = (ξj − ξk ) (i 6= j)     k=1 k=1   (k6=j) (k6=i,j) 0 (30) Aij = Lj (ξi )   N  X   1   (ξi −ξk )    k=1 (k6=i) The conventional higher order weighting coefficients are computed as Bij = N X k=1 Aik Akj , Cij = N X Bik Akj (i, j = 1, 2, ..., N ) (31) k=1 Here, Bij and Cij are weighting coefficients for second and third order derivatives, respectively. The sixth order partial differential equation given in Equation (9) ren0 00 ders slope w and curvature w as extra degrees of freedom at the element boundaries. To account for these extra boundary degrees of freedom in the formulation, the derivatives of conventional weighting function Aij , Bij , and Cij are modified as follows: First order derivative matrix :   Aij (i, j = 1, 2, · · · , N ) Āij =   0 (i, j = 1, 2, · · · , N, j = N + 1, · · · , N + 4) 11 (32) Second order derivative matrix :   Bij (j = 1, 2, · · · , N ) B̄ij =   0 (j = N + 1, · · · , N + 4)(i = 2, 3, · · · , N − 1) B̄ij = N −1 X (33) Aik Akj (j = 1, 2, · · · , N, i = 1, N ) k=2 B̄i(N +1) = Ai1 ; B̄i(N +2) = AiN (i = 1, N ) Third order derivative matrix :   Cij (j = 1, 2, · · · , N ) C̄ij =   0 (j = N + 1, · · · , N + 4, i = 2, 3, · · · , N − 1) C̄ij = N −1 X (34) (35) Bik Akj (j = 1, 2, · · · , N, i = 1, N ) k=2 C̄i(N +3) = Ai1 ; C̄i(N +4) = AiN (i = 1, N ) (36) Using the above Equations (32)-(36), the element matrices can be expressed in terms of weighting coefficients as Elastic stif f ness matrix : Kij = N N X 8EI X 2 32EI H B̄ B̄ + g Hk C̄ki C̄kj k ki kj L3 k=1 L5 k=1 (i, j = 1, 2, ..., N, N + 1, · · · , N + 4) (37) ρAL Hk δij (i, j = 1, 2, ..., N ) 2 (38) Consistent mass matrix : Mij = Here ξ and H are the coordinate and weights of GLL quadrature. δij is the Dirac-delta function. 12 2.2 Hermite interpolation based quadrature beam element For the case of quadrature element based on C 2 continuous Hermite interpolation the displacement for a N-node gradient beam element is assumed as w(ξ, t) = N X 0 0 00 00 φj (ξ)wj + ψ1 (ξ)w1 + ψN (ξ)wN + ϕ1 (ξ)w1 + ϕN (ξ)wN = j=1 N +4 X Γj (ξ)wjb j=1 (39) φ, ψ and ϕ are Hermite interpolation functions defined as [24, 26] ϕj (ξ) = 1 Lj (x)(x − xj )2 (x − xN −j+1 )2 (j = 1, N ) 2(ξj − ξN −j+1 )2 1 Lj (ξ)(ξ − ξj )(ξ − ξN −j+1 )2 (ξj − ξN −j+1 )2   4 1 − 2Lj (ξj ) + ϕj (ξ) (j = 1, N ) ξj − ξN −j+1 (40) ψj (ξ) = (41)   1 2 2 1 φj (ξ) = Lj (ξ)(ξ − ξN −j+1 ) − Lj (ξj ) + ψj (ξ) (ξj − ξN −j+1 )2 ξj − ξN −j+1   4L1j (ξj ) 2 2 + − Lj (ξj ) + ϕj (ξ) (j = 1, N ) ξj − ξN −j+1 (ξj − ξN −j+1 )2 (42) φj (ξ) = 1 Lj (ξ)(ξ − ξ1 )2 (ξ − ξN )2 (j = 2, 3, ..., N − 1) 2 2 (ξj − ξ1 ) (ξj − ξN ) (43) The kth order derivative of w(ξ) with respect to ξ is obtained from Equation (39) as k w (ξ) = N X φkj (ξ)wj + 0 ψ1k (ξ)w1 + 0 k ψN (ξ)wN j=1 + 00 ϕk1 (ξ)w1 + 00 ϕkN (ξ)wN = N +4 X Γkj (ξ)wjb j=1 (44) Using the above Equation (40)-(44), the element matrices can be expressed in terms of weighting coefficients as 13 Elastic stif f ness matrix : N N X 8EI X (2) (2) (3) (3) 2 32EI Kij = 3 Hk Γki Γkj + g Hk Γki Γkj 5 L k=1 L k=1 (i, j = 1, 2, ..., N, N + 1, · · · , N + 4) (45) here ξ and H are the coordinate and weights of GLL quadrature. The consistent mass matrix remains the same as given by Equation (38). Combining the stiffness and mass matrix, the system of equations after applying the boundary conditions can be expressed as       kbb kbd  ∆ I 0   b    fb            = (46)  kdb 0 kdd   ∆d  ω 2 Mdd   ∆d          where the vector ∆b contains the boundary related non-zero slope and curvature dofs. Similarly, the vector ∆d includes all the non-zero displacement dofs of the beam. In the present analysis the boundary force vector is assumed to be zero, fb = 0. Now, expressing the ∆b dofs in terms of ∆d , the system of equations reduces to in o h h in o −1 (47) kdd − kdb kbb kbd ∆d = ω 2 Mdd wd h i −1 Here, K̄ = kdd − kdb kbb kbd is the modified stiffness matrix associated with ∆d dofs. The above system of equations leads to an Eigenvalue problem and its solutions renders frequencies and corresponding mode shapes. 3 Quadrature element for gradient elastic Kirchhoff plate In this section, we formulate two novel quadrature elements for non-classical gradient Kirchhoff plate. First, the quadrature element based on Lagrange interpolation in ξ and η direction is presented. Next, the quadrature element based on Lagrange-Hermite mixed interpolation, with Lagrangian interpolation is ξ direction and Hermite interpolation assumed in η direction is formulated. The GLL points in ξ and η directions are used as element nodes. Similar to the beam elements discussed in the section 2, the plate element also has displacement w̄ as the only degrees of freedom in the domain and at the edges it has 3 degrees of freedom w̄, w̄x or w̄y , w̄xx or w̄yy depending 14 upon the edge. At the corners the element has five degrees of freedom, w̄, w̄x , w̄y , w̄xx and w̄yy . The new displacement vector now includes the slope and curvature as additional degrees of freedom at the element boundaries j j , · · · }, where , · · · , w̄yy given by: wp = {w̄i , · · · , w̄N ×N , w̄xj , · · · , w̄yj , · · · , w̄xx (i = 1, 2, · · · , N × N ; j = 1, 2, · · · , 4N ). A quadrature element for a gradient Kirchhoff plate with Nx × Ny grid is shown in the Figure 2. 𝑦, 𝜂 = 2𝑦 𝑙𝑦 𝑑𝑤 𝑑 2 𝑤 Dofs related to 𝑑𝑦 , 𝑑𝑦2 on edge 𝑦 = 𝑙𝑦 /2 Dofs related to 𝑑𝑤 𝑑 2 𝑤 , 𝑑𝑥 𝑑𝑥 2 on edge 𝑥 = −𝑙𝑥 /2 𝒘′′ 𝟓𝟏 𝒘′′ 𝟓𝟐 𝒘′′ 𝟓𝟑 𝒘′′ 𝟓𝟓 𝒘′′ 𝟓𝟒 Dofs related to 𝒘′𝟓𝟏 𝒘′𝟓𝟐 𝒘′𝟓𝟑 𝒘′𝟓𝟒 𝒘′𝟓𝟓 𝑑𝑤 𝑑 2 𝑤 , 𝑑𝑥 𝑑𝑥 2 𝑥 = 𝑙𝑥 /2 𝒘′′ 𝟓𝟏 𝒘′𝟓𝟏 𝒘𝟓𝟏 𝒘𝟓𝟐 𝒘𝟓𝟑 𝒘𝟓𝟒 𝒘𝟓𝟓 𝒘′𝟓𝟓 𝒘′′ 𝟓𝟓 𝒘′′ 𝟒𝟏 𝒘′𝟒𝟏 𝒘𝟒𝟏 𝒘𝟒𝟐 𝒘𝟒𝟑 𝒘𝟒𝟒 𝒘𝟒𝟓 𝒘′𝟒𝟓 𝒘′′ 𝟒𝟓 𝒘′′ 𝟑𝟏 𝒘′𝟑𝟏 𝒘𝟑𝟏 𝒘𝟑𝟐 𝒘𝟑𝟑 𝒘𝟑𝟒 𝒘𝟑𝟓 𝒘′𝟑𝟓 𝒘′′ 𝟑𝟓 𝒘′′ 𝟐𝟏 𝒘′𝟐𝟏 𝒘𝟐𝟏 𝒘𝟐𝟐 𝒘𝟐𝟑 𝒘𝟐𝟒 𝒘𝟐𝟓 𝒘′𝟐𝟓 𝒘′′ 𝟐𝟓 𝒘′′ 𝟏𝟏 𝒘′𝟏𝟏 𝒘𝟏𝟏 𝒘𝟏𝟐 𝒘𝟏𝟑 𝒘𝟏𝟒 𝒘𝟏𝟓 𝒘′𝟏𝟓 𝒘′′ 𝟏𝟓 𝒘′𝟏𝟏 𝒘′𝟏𝟐 𝒘′𝟏𝟑 𝒘′𝟏𝟒 𝒘′𝟏𝟓 Dofs related to 𝒘′′ 𝟏𝟏 𝒘′′ 𝟏𝟐 𝒘′′ 𝟏𝟑 𝒘′′ 𝟏𝟓 𝒘′′ 𝟏𝟒 on edge 𝑥, 𝜉 = 𝒍𝒚 𝑑𝑤 𝑑 2 𝑤 , 𝑑𝑦 𝑑𝑦 2 2𝑥 𝑙𝑥 on edge 𝑦 = −𝑙𝑦 /2 𝒍𝒙 Figure 2: A typical quadrature element for a gradient elastic Kirchhoff plate with N = Nx = Ny = 5. Here, N = Nx = Ny = 5 are the number of grid points in ξ and η directions, respectively. It can be seen from the Figure 2, the plate element has three degrees of freedom on each edge, five degrees of freedom at the 0 corners and only displacement in the domain. The slope w̄ and curvature 00 w̄ dofs related to each edge of the plate are highlighted by the boxes. The transformation used for the plate is ξ = 2x/lx and η = 2y/ly with −1 ≤ 15 (ξ, η) ≤ 1. 3.1 Lagrange interpolation based quadrature element for gradient elastic plates The displacement for a Nx × Ny node quadrature plate element is assumed as N X N N X N X X p p w̄(x, y, t) = Li (x)Lj (y)wij (t) = L̄i (ξ)L̄j (η)wij (t) (48) i=1 j=1 i=1 j=1 p w̄ij (t) where is the nodal displacement vector for the plate and L̄i (ξ), L̄j (η) are the Lagrange interpolation functions in ξ and η directions, respectively. The slope and curvature degrees of freedom at the element boundaries are accounted while computing the weighting coefficients of higher order derivatives as discussed in section 2.1. Substituting the above Equation (48) in Equation (20) we get the stiffness matrix for a gradient elastic quadrature plate element as N N  T ab X X Hi Hj F (ξi , ηj ) cl [D]cl [F (ξi , ηj )]cl [K]cl = 4 i=1 j=1 [K]sg = g 2 ab 4 N X N X Hi Hj [F (ξi , ηj )]Tsg [D]sg [F (ξi , ηj )]sg (49) (50) i=1 j=1 where (ξi ηj ) and (Hi , Hj ) are abscissas and weights of GLL quadrature rule. [F (ξi , ηj )]cl and [F (ξi , ηj )]sg are the classical and non-classical strain matrices at location (ξi , ηj ) for gradient elastic plate. [D]cl and [D]sg are the constitutive matrices corresponding to classical and gradient elastic plate. The classical and non-classical strain matrices are defined as   N +4 X 4 ξ p   B̄ik w̄kj   2 a   k=1       N +4   X     4 η p p  (i, j = 1, 2, .., N )  (51) F (ξi , ηj ) cl {w̄ } =  B̄ik w̄ik  2 b   k=1       N +4 N +4   X X  8 ξ η p  Āil Ājk w̄lk  ab l=1 k=1 16   N +4 X 2 8 ξ p   g 3 C̄ik w̄kj   a   k=1       N +4   X   η p 2 8   g 3 C̄ik w̄ik   b   k=1     p  F (ξi , ηj ) sg {w̄ } =    N +4 N +4   X X  2 8  p ξ η g  w̄ B̄ Ā lk il jk  a2 b    l=1 k=1       N +4 N +4    2 8 XX ξ η p  g Āil B̄jk w̄lk  ab2 l=1 (i, j = 1, 2, .., N ) k=1 The classical and non-classical constitutive matrices are given as   1 µ 0        1 0 D]cl =  µ    0 0 2(1 − µ)   D]sg 1   0  =  0   0 (52) 0 µ 1 0 µ (3 − 2µ) µ 0 0 (53)    µ      0   (3 − 2µ) (54) The diagonal mass matrix is given by Mkk = 3.2 ρhab Hi Hj 4 (i, j = 1, 2, ..., N ) (k = (i − 1) × N + j) (55) Mixed interpolation based quadrature element for gradient elastic plates The quadrature element presented here is based on mixed Lagrange-Hermite interpolation, with Lagrangian interpolation is assumed in ξ direction and 17 Hermite interpolation in η direction. The displacement for a Nx × Ny node mixed interpolation quadrature plate element is assumed as w̄(x, y, t) = N N +4 X X p Li (x)Γj (y)wij (t) i=1 j=1 = N N +4 X X p L̄i (ξ)Γ̄j (η)wij (t) (56) i=1 j=1 p where w̄ij (t) is the nodal displacement vector of the plate and L̄i (ξ) and Γ̄j (η) are the Lagrange and Hermite interpolation functions in ξ and η directions, respectively. The formulations based on mixed interpolation methods have advantage in excluding the mixed derivative dofs at the free corners of the plate [10]. The modified weighting coefficient matrices derived in section 2.1, using Lagrange interpolations and those given in section 2.2, for Hermite interpolations are used in forming the element matrices. Substituting the above Equation (56) in Equation (20), we get the stiffness matrix for gradient elastic quadrature plate element based on mixed interpolation as N [K]cl = [K]sg = g N  T ab X X Hi Hj G(ξi , ηj ) cl [D]cl [G(ξi , ηj )]cl 4 i=1 j=1 2 ab 4 N X N X Hi Hj [G(ξi , ηj )]Tsg [D]sg [G(ξi , ηj )]sg (57) (58) i=1 j=1 where (ξi ηj ) and (Hi , Hj ) are abscissas and weights of GLL quadrature rule. [D]cl and [D]sg are the classical and gradient elastic constitutive matrices for the plate defined in the section 3.1. The classical [G(ξi , ηj )]cl and non-classical [G(ξi , ηj )]sg strain matrices at the location (ξi , ηj ) are defined as,.   N +4 X 4 (ξ) p   B̄ik w̄kj   2 a   k=1       N +4   X     4 2(η) p p  (i, j = 1, 2, .., N ) (59)  G(ξi , ηj ) cl {w̄ } =  Γ̄ w̄ jk ik  2 b   k=1       N +4 N +4   X X  8 (ξ) 1(η) p  Āil Γ̄jk w̄lk  ab l=1 k=1 18  N +4 X 2 8  (ξ) p   g 3 C̄ik w̄kj   a   k=1       N +4   X   3(η) p 2 8   g 3 Γ̄jk w̄ik   b   k=1     p  F (ξi , ηj ) sg {w̄ } =    N +4 N +4   X X  2 8  2(ξ) (η) p g  Γ̄ Ā w̄ il lk jk  a2 b    l=1 k=1       N +4 N +4    2 8 X X 1(ξ) (η) p  g Γ̄il B̄jk w̄lk  ab2 l=1 (i, j = 1, 2, .., N ) (60) k=1 The diagonal mass matrix remains the same as Equation (55). Here, Ā, B̄ and C̄ are the first, second and third order derivatives of Lagrange interpolation functions along the ξ direction. Similarly, Γ̄1 , Γ̄2 and Γ̄3 are the first, second and third order derivatives of Hermite interpolation functions in the η direction . 4 Numerical Results and Discussion The efficiency of the proposed quadrature beam and plate element is demonstrate through free vibration analysis. Initially, the convergence study is performed for an Euler-Bernoulli gradient beam, followed by frequency comparisons for different boundary conditions and g values. Similar, study is conducted for a Kirchhoff plate and the numerical results are tabulated and compared with available literature. Four different values of length scale parameters, g = 0.00001, 0.05, 0.1, and 0.5 are considered in this study. Single element is used with GLL quadrature points as nodes to generate all the results reported herein. For results comparison the proposed gradient quadrature beam element based on Lagrange interpolation is designated as SgQE-L and the element based on Hermite interpolation as SgQE-H. Similarly, the plate element based on Lagrange interpolation in ξ and η directions as SgQE-LL and the element based on mixed interpolation as SgQE-LH. In this study, the rotary inertia related to slope and curvature degrees of freedom is neglected. 19 4.1 Quadrature beam element for gradient elasticity theory The numerical data used for the analysis of beams is as follows: Length L = 1, Young’s modulus E = 3 × 106 , Poission’s ratio ν = 0.3 and density ρ = p 1. All the frequencies reported for beams are nondimensional as 2 ω̄ = ωL ρA/EI. Where A and I are area and moment of inertia of the beam and ω is the natural frequency. The analytical solutions for gradient elastic Euler-Bernoulli beam with different boundary conditions are obtained by following the approach given in [44] and the associated frequency equations are presented in Appendix-I. The classical and non-classical boundary conditions used in the free vibration analysis for different end support are: Simply supported : 00 classical : w = M = 0 , non-classical : w = 0 at x = (− L2 , L2 ) Clamped : 0 00 classical : w = w = 0 , non-classical : w = 0 at x = (− L2 , L2 ) Free-free : classical : Q = M = 0 , non-classical : M̄ = 0 at x = (− L2 , L2 ) Cantilever : 0 classical : w = w = 0 00 non-classical : w = 0 at x = − L2 , Q = M = 0 at x = at x = − L2 , M̄ = 0 at x = L2 Propped cantilever : 0 classical : w = w = 0 at x = − L2 , w = M = 0 at x = 00 00 non-classical : w = 0 at x = − L2 , w = 0 at x = L2 L 2 L 2 The size of the displacement vector ∆d defined in Equation (46) remains as N − 2 for all the boundary conditions of the beam except for free-free and cantilever beam which are N and N −1, respectively. However, the size of the ∆b vector depends upon the number of non-zero slope and curvature dofs at the element boundaries. The non-classical boundary conditions employed for 00 simply supported gradient beam are w = 0 at x = (− L2 , L2 ), the equations related to curvature degrees of freedom are eliminated and the size of ∆b is 2. For the gradient cantilever beam the non-classical boundary conditions 00 used are w = 0 at x = − L2 and M̄ = 0 at x = L2 . The equation related to curvature degrees of freedom at x = − L2 is eliminated and the equation related to higher order moment at x = L2 is retained and the size of ∆b = 2. 20 In the case of clamped beam the non-classical boundary conditions read 00 w = 0 at x = (− L2 , L2 ) and the ∆b is zero. Similarly, the size for the propped 00 cantilever beam will be 3 as w = 0 at x = (− L2 , L2 ). Finally, for a free-free beam the size of ∆b vector is 4 due to M̄ = 0 at x = (− L2 , L2 ). 4.1.1 Frequency convergence for gradient elastic quadrature beam elements In this section, the convergence behaviour of frequencies obtained using proposed SgQE-L and SgQE-H elements for simply supported and free-free Euler-Bernoulli beam are compared. Figure 3, shows the comparison of first three frequencies for a simply supported gradient beam and their convergence trends for g/L = 0.1. The convergence is seen faster for both SgQE-L and SgQE-H elements for all the three frequencies with solution converging to analytical values with 10 nodes. Similar trend is noticed in the the Figure 4, for free-free beam. It is to be noted that, the proposed SgQE-L and SgQE-H elements are efficient in capturing the rigid body modes associated with the generalized degrees of freedom. The frequencies reported for free-free beam are related to elastic modes and the rigid mode frequencies are not reported here, which are zeros. Hence, single SgQE-L or SgQE-H element with fewer number of nodes can produce accurate solutions even for higher frequencies. SgQE−L SgQE−H Analytical 100 Mode−III Nondimensional frequency 80 60 Mode−II 40 20 Mode−I 0 4 6 8 10 12 14 16 18 20 22 Number of nodes N Figure 3: Convergence behaviour of frequencies for a simply supported gradient beam (g/L = 0.1). 21 SgQE−L SgQE−H Analytical 140 Mode−III Nondimensional frequency 120 100 80 Mode−II 60 40 Mode−I 20 0 4 6 8 10 12 14 16 18 20 22 Number of nodes N Figure 4: Convergence behaviour of frequencies for a free-free gradient beam (g/L = 0.1). 4.1.2 Free vibration analysis of gradient beams using SgQE-L and SgQE-H elements To demonstrate the applicability of the SgQE-L and SgQE-H elements for different boundary conditions the frequencies are compared with the analytical solutions in Tables 1-5. The comparison is made for first six frequencies obtained for different values of g/L = 0.00001, 0.05, 0.1, 0.5. 22 Freq. g/L ω̄1 ω̄2 ω̄3 ω̄4 ω̄5 ω̄6 0.00001 0.05 0.1 0.5 SgQE-L 9.869 9.870 9.874 9.984 SgQE-H 9.869 9.871 9.874 9.991 Analytical 9.870 9.871 9.874 9.991 SgQE-L 39.478 39.498 39.554 41.302 SgQE-H 39.478 39.498 39.556 41.381 Analytical 39.478 39.498 39.556 41.381 SgQE-L 88.826 88.923 89.207 97.725 SgQE-H 88.826 88.925 89.220 98.195 Analytical 88.826 88.925 89.220 98.195 SgQE-L 157.914 158.221 159.125 185.378 SgQE-H 157.915 158.225 159.156 186.497 Analytical 157.914 158.226 159.156 186.497 SgQE-L 246.740 247.480 249.655 310.491 SgQE-H 247.305 247.475 249.760 313.741 Analytical 246.740 247.500 249.765 313.743 SgQE-L 355.344 357.039 361.805 486.229 SgQE-H 355.306 356.766 361.564 488.302 Analytical 355.306 356.880 361.563 488.240 Table 1: Comparison of first six frequencies for a simply supported gradient beam 23 Freq. g/L ω̄1 ω̄2 ω̄3 ω̄4 ω̄5 ω̄6 0.00001 0.05 0.1 0.5 SgQE-L 22.373 22.376 22.387 22.691 SgQE-H 22.373 22.377 22.387 22.692 Analytical 22.373 22.377 22.387 22.692 SgQE-L 61.673 61.708 61.814 64.841 SgQE-H 61.673 61.708 61.814 64.856 Analytical 61.673 61.708 61.814 64.856 SgQE-L 120.903 121.052 121.496 133.627 SgQE-H 120.904 121.052 121.497 133.710 Analytical 120.903 121.052 121.497 133.710 SgQE-L 199.859 202.864 201.553 234.596 SgQE-H 199.876 200.287 201.556 234.875 Analytical 199.859 200.286 201.557 234.875 SgQE-L 298.550 299.528 302.422 374.535 SgQE-H 298.556 299.365 302.403 375.234 Analytical 298.555 299.537 302.443 375.250 SgQE-L 417.217 419.418 425.469 562.869 SgQE-H 416.991 418.438 424.747 562.758 Analytical 416.991 418.942 424.697 562.536 Table 2: Comparison of first six frequencies for a free-free gradient beam In the Table 1, the comparison of first six frequencies for a simply supported gradient beam are shown. For g/L = 0.00001, all the frequencies match well with the exact frequencies of classical beam. Good agreement with analytical solutions is noticed for all the frequencies obtained using SgQE-L and SgQE-H elements for higher values of g/L. In Table 2, the frequencies corresponding to elastic modes are tabulated and compared for a free-free beam. Similarly, in Tables 3-5, comparison in made for cantilever, clamped and propped cantilever gradient beams, respectively. The frequencies obtained using SgQE-L and SgQE-H elements are in close agreement with the analytical solutions for different g/L values. Hence, based on the 24 above findings it can be stated that the SgQE-I and SgQE-II elements can be applied for free vibration analysis of gradient Euler-Bernoulli beam for any choice of boundary conditions and g/L values. Freq. g/L ω̄1 ω̄2 ω̄3 ω̄4 ω̄5 ω̄6 0.00001 0.05 0.1 0.5 SgQE-L 22.324 22.801 23.141 27.747 SgQE-H 22.590 22.845 23.310 27.976 Analytical 22.373 22.831 23.310 27.976 SgQE-L 61.540 62.720 63.984 79.450 SgQE-H 62.276 63.003 64.365 79.970 Analytical 661.673 62.961 64.365 79.970 SgQE-L 120.392 122.916 125.542 162.889 SgQE-H 122.094 123.594 126.512 164.927 Analytical 120.903 123.511 126.512 164.927 SgQE-L 199.427 203.581 208.627 286.576 SgQE-H 201.843 204.502 209.887 289.661 Analytical 199.859 204.356 209.887 289.661 SgQE-L 297.282 304.138 312.503 455.285 SgQE-H 301.541 305.843 314.956 462.238 Analytical 298.555 305.625 314.956 462.238 SgQE-L 421.194 427.786 442.299 681.749 SgQE-H 421.092 427.787 442.230 691.292 Analytical 416.991 427.461 442.230 691.292 Table 3: Comparison of first six frequencies for a clamped gradient beam 25 Freq. g/L ω̄1 ω̄2 ω̄3 ω̄4 ω̄5 ω̄6 0.00001 0.05 0.1 0.5 SgQE-L 3.532 3.545 3.584 3.857 SgQE-H 3.534 3.552 3.587 3.890 Analytical 3.532 3.552 3.587 3.890 SgQE-L 21.957 22.188 22.404 24.592 SgQE-H 22.141 22.267 22.497 24.782 Analytical 22.141 22.267 22.496 24.782 SgQE-L 61.473 62.150 62.822 71.207 SgQE-H 61.997 62.375 63.094 71.863 Analytical 61.997 62.375 63.094 71.863 SgQE-L 120.465 121.867 123.424 146.652 SgQE-H 121.495 122.313 123.966 148.181 Analytical 121.495 122.313 123.966 148.181 SgQE-L 199.141 201.636 204.752 257.272 SgQE-H 200.848 202.377 205.658 260.336 Analytical 202.377 205.658 260.336 200.847 SgQE-L 297.489 301.551 307.229 410.222 SgQE-H 300.043 302.667 308.605 415.802 Analytical 300.043 302.667 308.605 415.802 Table 4: Comparison of first six frequencies for a cantilever gradient beam 26 Freq. g/L ω̄1 ω̄2 ω̄3 ω̄4 ω̄5 ω̄6 0.00001 0.05 0.1 0.5 SgQE-L 15.413 15.520 15.720 17.351 SgQE-H 15.492 15.581 15.740 17.324 Analytical 15.492 15.581 15.740 17.324 SgQE-L 49.869 50.313 51.026 57.767 SgQE-H 50.207 50.512 51.089 58.197 Analytical 50.207 50.512 51.089 58.197 SgQE-L 104.044 105.043 106.557 127.127 SgQE-H 104.758 105.457 106.865 128.005 Analytical 104.758 105.457 106.865 128.005 SgQE-L 177.922 179.778 182.822 231.247 SgQE-H 179.149 180.500 183.389 233.357 Analytical 179.149 180.500 183.389 233.357 SgQE-L 271.502 274.654 280.210 378.692 SgQE-H 273.383 275.749 281.089 382.058 Analytical 273.383 275.749 281.089 382.058 SgQE-L 384.785 389.746 399.154 575.841 SgQE-H 387.463 391.341 400.509 582.607 Analytical 387.463 391.341 400.509 582.607 Table 5: Comparison of first six frequencies for a propped cantilever gradient beam 4.2 Quadrature plate element for gradient elasticity theory Three different boundary conditions of the plate, simply supported on all edges (SSSS), free on all edges (FFFF) and combination of simply supported and free (SSFF) are considered. The converge behaviour of SgQE-LL and SgQE-LH plate elements is verified first, later numerical comparisons are made for all the three plate conditions for various g/lx values. Allp the fre2 quencies reported herein for plate are non-dimensional as ω̄ = ωlx ρh/D. 27 The numerical data used for the analysis of plates is: length lx = 1, width ly = 1, thickness h = 0.01, Young’s modulus E = 3 × 106 , Poission’s ratio ν = 0.3 and density ρ = 1. The number of nodes in either direction are assumed to be equal, N = Nx = Ny . The choice of the essential and natural boundary conditions for the above three plate problems are given in section 1.2. The size of the displacement vector ∆d defined in Equation (46) remains as (N − 2) × (N − 2) for all the boundary conditions of the gradient plate except for free-free and cantilever plate which are N × N and (N × N ) − N , respectively. However, the size of the ∆b vector depends upon the number of non-zero slope and curvature dofs along the element boundaries. The nonclassical boundary conditions employed for SSSS gradient plate are w̄xx = 0 at x = (− l2x , l2x ) and w̄yy = 0 at y = (− l2y , l2y ), the equations related to curvature degrees of freedom are eliminated and the size of ∆b will be 4N −8, as the w̄x = w̄y = 0 at the corners of the plate. For a FFFF plate the nonclassical boundary conditions employed are M̄x = 0 at x = (− l2x , l2x ) and M̄y = 0 at y = (− l2y , l2y ), and the size of ∆b is 8N . Finally, for SSFF plate ∆b = 6N − 4. 4.2.1 Frequency convergence of gradient elastic quadrature plate elements In Figure 5, convergence of first three frequencies for a SSSS plate obtained using SgQE-LL and SgQE-LH elements for g/lx = 0.05 is plotted and compared with analytical solutions [38]. Both SgQE-LL and SgQE-LH elements show excellent convergence behaviour for all the three frequencies. Figures 6 and 7, illustrate the frequency convergence for FFFF and SSFF plates, respectively, for g/lx = 0.05. Only the SgQE-LL and SgQE-LH element frequencies are shown, as the gradient solution are not available in literature for comparison. It is observed that SgQE-LL and SgQE-LH elements exhibit identical convergence characteristics. 28 SgQE−LL SgQE−LH Analytical [36] 100 Mode−III Nondimensional frequency 80 60 Mode−II 40 Mode−I 20 0 4 6 8 10 12 14 16 18 20 22 Number of nodes N Figure 5: Convergence behaviour of frequencies for a SSSS gradient plate (g/lX = 0.05). 30 SgQE−LL SgQE−LH Mode−III Nondimensional frequency 25 Mode−II 20 15 Mode−I 10 4 6 8 10 12 14 16 18 20 22 Number of nodes N Figure 6: Convergence behaviour of frequencies for a FFFF gradient plate (g/lX = 0.05). 4.2.2 Free vibration analysis of gradient plate using SgQE-LL and SgQE-LH elements The first six frequencies for SSSS, FFFF and SSFF plates obtained using SgQE-LL and SgQE-LH elements are compared and tabulated. The comparison is made for different length scale parameter: g/lx = 0.00001, 0.05, 0.1, 0.5. 29 25 SgQE−LL SgQE−LH Mode−III 20 Nondimensional frequency Mode−II 15 10 5 Mode−I 0 4 6 8 10 12 14 16 18 20 22 Number of nodes N Figure 7: Convergence behaviour of frequencies for a SSFF gradient plate (g/lX = 0.05). All the tabulated reslts are generated using Nx = Ny = 11 nodes. In the Table 6, the comparison of first six frequencies for SSSS gradient plate are shown. Good agreement with analytical solutions [38] is noticed for all the frequencies obtained using SgQE-LL and SgQE-LH elements for different g/lx . Tables 7 and 8 contains the frequency comparison for FFFF and SSFF plates for various g/lx values. As the exact solutions for gradient elastic plate are not available in the literature for FFFF and SSFF support conditions, the frequencies obtained using SgQE-LL and SgQE-LH are compared. Both elements show identical performance for all g/lx values. The frequencies obtained for lower values of g/lx = 0.00001 match well with the classical plate frequencies for all support conditions. In the above findings, SgQE-LL and SgQE-LH elements demonstrate excellent agreement with analytical results for all frequencies and g/lx values for a SSSS plate. For FFFF and SSFF plates, SgQE-LL and SgQE-LH elements produce similar frequencies for g/lx values considered. Hence, a single SgQE-LL or SgQE-LH element with few nodes can be used efficiently to study the free vibration behaviour of gradient plates with different support conditions and g/lx values. 30 Freq. g/lx ω̄1 ω̄2 ω̄3 ω̄4 ω̄5 ω̄6 0.00001 0.05 0.1 0.5 SgQE-LL 19.739 20.212 21.567 47.693 SgQE-LH 19.739 20.286 21.791 49.940 Analyt. [38] (m=1,n=1) 19.739 20.220 21.600 48.087 SgQE-LL 49.348 52.249 60.101 178.418 SgQE-LH 49.348 52.365 60.429 180.895 Analyt. [38] (m=1,n=2) 49.348 52.303 60.307 180.218 SgQE-LL 78.957 86.311 105.316 357.227 SgQE-LH 78.957 86.720 106.321 363.156 Analyt. [38] (m=2,n=2) 78.957 86.399 105.624 359.572 SgQE-LL 98.696 109.863 137.940 491.447 SgQE-LH 98.696 109.950 138.193 493.131 Analyt. [38] (m=1,n=3) 98.696 110.201 139.121 500.088 SgQE-LL 128.305 147.119 192.759 730.346 SgQE-LH 128.305 147.639 193.950 736.599 Analyt. [38] (m=2,n=3) 128.305 147.454 193.865 737.906 SgQE-LL 167.783 199.133 272.173 1084.136 SgQE-LH 167.783 199.262 272.486 1085.930 Analyt. [38] (m=1,n=4) 167.783 199.897 274.562 1099.535 Table 6: Comparison of first six frequencies for a gradient SSSS plate 31 Freq. g/lx ω̄1 ω̄2 ω̄3 ω4 ω̄5 ω̄6 0.00001 0.05 0.1 0.5 SgQE-LL 13.468 13.546 13.681 14.118 SgQE-LH 13.468 13.551 13.713 15.628 Classical [10] (g/lx =0) 13.468 —— —— —— SgQE-LL 19.596 19.820 20.313 22.113 SgQE-LH 19.596 19.820 20.315 22.129 Classical [10] (g/lx =0) 19.596 —— —— —— SgQE-LL 24.270 24.699 25.681 29.745 SgQE-LH 24.270 24.700 25.686 29.785 Classical [10] (g/lx =0) 24.270 —— —— —— SgQE-LL 34.8001 35.780 37.929 73.986 SgQE-LH 34.8001 35.722 38.015 76.161 Classical [10] (g/lx =0) 34.8001 —— —— —— SgQE-LL 61.093 64.314 71.238 145.033 SgQE-LH 61.093 64.317 71.244 145.065 Classical [10] (g/lx =0) 61.093 —— —— —— SgQE-LL 63.686 67.059 75.114 193.940 SgQE-LH 63.686 67.123 75.509 200.707 Classical [10] (g/lx =0) 63.686 —— —— —— Table 7: Comparison of first six frequencies for a gradient FFFF plate 32 Freq. g/lx ω̄1 ω̄2 ω̄3 ω̄4 ω̄5 ω̄6 0.00001 0.05 0.1 0.5 SgQE-LL 3.367 3.373 3.386 3.491 SgQE-LH 3.367 3.382 3.413 3.950 Classical [47,48] (g/lx =0) 3.367 —— —— —— SgQE-LL 17.316 17.598 18.370 32.579 SgQE-LH 17.316 17.634 18.474 33.927 Classical [47,48] (g/lx =0) 17.316 —— —— —— SgQE-LL 19.292 19.645 20.585 35.825 SgQE-LH 19.292 19.664 20.649 36.852 Classical [47,48] (g/lx =0) 19.292 —— —— —— SgQE-LL 38.211 39.671 39.162 105.800 SgQE-LH 38.211 39.775 43.851 109.959 Classical [47,48] (g/lx =0) 38.211 —— —— —— SgQE-LL 51.035 53.714 60.400 153.000 SgQE-LH 51.035 53.739 60.493 153.980 Classical [47,48] (g/lx =0) 51.035 —— —— —— SgQE-LL 53.487 56.431 63.699 158.557 SgQE-LH 53.487 56.537 64.000 161.072 Classical [47,48] (g/lx =0) 53.487 —— —— —— Table 8: Comparison of first six frequencies for a gradient SSFF plate 33 5 Conclusion Two novel versions of weak form quadrature elements for gradient elastic beam theory were proposed. This methodology was extended to construct two new and different quadrature plate elements based on Lagrange-Lagrange and mixed Lagrange-Hermite interpolations. A new way to account for the non-classical boundary conditions associated with the gradient elastic beam and plate theories were introduced and implemented. The capability of the proposed four elements was demonstrated through free vibration analysis. Based on the findings it was concluded that, accurate solutions can be obtained even for higher frequencies and for any choice of length scale parameter using single beam or plate element with fewer number of nodes. The new results reported for gradient plate for different boundary conditions can be a reference for the research in this field. References [1] Zienkiewicz.O.C, Taylor. R.L., The finite element method,Vol.1. Basic formulations and linear problems, London: McGraw-Hill, 1989,648p. [2] Zienkiewicz.O.C, Taylor. R.L., The finite element method, Vol.2. Solid and fluid mechanics: dynamics and non-linearity, London: McGraw-Hill, 1991, 807p. [3] K.J. Bathe., Finite element procedures in engineering analysis. PrenticeHall, 1982. [4] Bellman RE, Casti J., Differential quadrature and long-term integration. Journal of Mathematical Analysis and Applications 1971; 34:235–238. [5] Bert, C. W., and Malik, M., 1996,Differential Quadrature Method in Compu-tational Mechanics: A Review,. ASME Appl. Mech. Rev., 49(1), pp. 1–28. [6] Bert, C. W., Malik, M., 1996,The differential quadrature method for irregular domains and application to plate vibration.. International Journal of Mechanical Sciences 1996; 38:589–606. [7] C. Shu, Differential Quadrature and Its Application in Engineering,. Springer-Verlag, London, 2000. 34 [8] H. Du, M.K. Lim, N.R. Lin, Application of generalized differential quadrature method to structural problems,. Int. J. Num. Meth.Engrg. 37 (1994) 1881–1896. [9] H. Du, M.K. Lim, N.R. Lin, Application of generalized differential quadrature to vibration analysis,. J. Sound Vib. 181 (1995) 279–293. [10] Xinwei Wang, Differential Quadrature and Differential Quadrature Based Element Methods Theory and Applications,.Elsevier, USA, 2015 [11] O. Civalek, O.M. Ulker., Harmonic differential quadrature (HDQ) for axisymmetric bending analysis of thin isotropic circular plates,. Struct. Eng. Mech.17 (1) (2004) 1–14. [12] O. Civalek, Application of differential quadrature (DQ) and harmonic differential quadrature (HDQ) for buckling analysis of thin isotropic plates and elastic columns. Eng. Struct. 26 (2) (2004) 171–186. [13] X. Wang, H.Z. Gu, Static analysis of frame structures by the differential quadrature element method,. Int. J. Numer. Methods Eng. 40 (1997) 759–772. [14] X. Wang, Y. Wang, Free vibration analysis of multiple-stepped beams by the differential quadrature element method, Appl. Math. Comput. 219 (11) (2013) 5802–5810. [15] Wang Y, Wang X, Zhou Y., Static and free vibration analyses of rectangular plates by the new version of differential quadrature element method, International Journal for Numerical Methods in Engineering 2004; 59:1207–1226. [16] Y. Xing, B. Liu, High-accuracy differential quadrature finite element method and its application to free vibrations of thin plate with curvilinear domain, Int. J. Numer. Methods Eng. 80 (2009) 1718–1742. [17] Malik M., Differential quadrature element method in computational mechanics: new developments and applications. Ph.D. Dissertation, University of Oklahoma, 1994. [18] Karami G, Malekzadeh P., A new differential quadrature methodology for beam analysis and the associated differential quadrature element method. Computer Methods in Applied Mechanics and Engineering 2002; 191:3509–3526. 35 [19] Karami G, Malekzadeh P., Application of a new differential quadrature methodology for free vibration analysis of plates. Int. J. Numer. Methods Eng. 2003; 56:847–868. [20] A.G. Striz, W.L. Chen, C.W. Bert, Static analysis of structures by the quadrature element method (QEM),. Int. J. Solids Struct. 31 (1994) 2807–2818. [21] W.L. Chen, A.G. Striz, C.W. Bert, High-accuracy plane stress and plate elements in the quadrature element method, Int. J. Solids Struct. 37 (2000) 627–647. [22] H.Z. Zhong, Z.G. Yue, Analysis of thin plates by the weak form quadrature element method, Sci. China Phys. Mech. 55 (5) (2012) 861–871. [23] Chunhua Jin, Xinwei Wang, Luyao Ge, Novel weak form quadrature element method with expanded Chebyshev nodes, Applied Mathematics Letters 34 (2014) 51–59. [24] T.Y. Wu, G.R. Liu, Application of the generalized differential quadrature rule to sixth-order differential equations, Comm. Numer. Methods Eng. 16 (2000) 777–784. [25] G.R. Liu a , T.Y. Wu b, Differential quadrature solutions of eighth-order boundary-value differential equations, Journal of Computational and Applied Mathematics 145 (2002) 223–235. [26] Xinwei Wang, Novel differential quadrature element method for vibration analysis of hybrid nonlocal EulerBernoulli beams, Applied Mathematics Letters 77 (2018) 94–100. [27] Mindlin, R.D., 1965.1964. Micro-structure in linear elasticity. Arch. Rat. Mech. Anal. 16, 52–78. [28] Mindlin, R., Eshel, N., 1968. On first strain-gradient theories in linear elasticity. Int. J. Solids Struct. 4, 109–124. [29] Mindlin, R.D., 1965.1964. Micro-structure in linear elasticity. Arch. Rat. Mech. Anal. 16, 52–78. [30] Fleck, N.A., Hutchinson, J.W., A phenomenological theory for strain gradient effects in plasticity. 1993. J. Mech. Phys. Solids 41 (12), 1825e1857. 36 [31] Koiter, W.T., 1964.Couple-stresses in the theory of elasticity, I & II. Proc. K. Ned.Akad. Wet. (B) 67, 17–44. [32] Harm Askes, Elias C. Aifantis, Gradient elasticity in statics and dynamics: An overview of formulations,length scale identification procedures, finite element implementations and new results Int. J. Solids Struct. 48 (2011) 1962–1990 [33] Aifantis, E.C., Update on a class of gradient theories. 2003.Mech. Mater. 35,259e280. [34] Altan, B.S., Aifantis, E.C., On some aspects in the special theory of gradient elasticity. 1997. J. Mech. Behav. Mater. 8 (3), 231e282. [35] Papargyri-Beskou, S., Tsepoura, K.G., Polyzos, D., Beskos, D.E., Bending and stability analysis of gradient elastic beams. 2003. Int. J. Solids Struct. 40, 385e400. [36] S. Papargyri - Beskou, D. Polyzos, D. E. Beskos, Dynamic analysis of gradient elastic flexural beams. Structural Engineering and Mechanics, Vol. 15, No. 6 (2003) 705–716. [37] A.K. Lazopoulos, Dynamic response of thin strain gradient elastic beams, International Journal of Mechanical Sciences 58 (2012) 27–33. [38] Papargyri-Beskou, S., Beskos, D., Static, stability and dynamic analysis of gradient elastic flexural Kirchhoff plates. 2008. Arch. Appl. Mech. 78, 625e635. [39] Lazopoulos, K.A., On the gradient strain elasticity theory of plates. 2004. Eur. J.Mech. A/Solids 23, 843e852. [40] Papargyri-Beskou, S., Giannakopoulos, A.E., Beskos, D.E., Variational analysis of gradient elastic flexural plates under static loading. 2010. International Journal of Solids and Structures 47, 2755e2766. [41] I. P. Pegios S. Papargyri-Beskou D. E. Beskos, Finite element static and stability analysis of gradient elastic beam structures, Acta Mech 226, 745–768 (2015). [42] Tsinopoulos, S.V., Polyzos, D., Beskos, D.E, Static and dynamic BEM analysis of strain gradient elastic solids and structures, Comput. Model. Eng. Sci. (CMES) 86, 113–144 (2012). 37 [43] Vardoulakis, I., Sulem, J., Bifurcation Analysis in Geomechanics. 1995. Blackie/Chapman and Hall, London. [44] Kitahara, M. (1985), Boundary Integral Equation Methods in Eigenvalue Problems of Elastodynamics and Thin Plates, Elsevier, Amsterdam. [45] J.N. Reddy, Energy Principles and Variational Methods in Applied Mechanics, Second Edition, John Wiley, NY, 2002. [46] S.P. Timoshenko, D.H. Young, Vibration Problem in Engineering, Van Nostrand Co., Inc., Princeton, N.J., 1956. [47] B.Singh, S. Chakraverty, Flexural vibration of skew plates using boundary characteristics orthogonal polynomials in two variables, J. Sound Vib. 173 (2) (1994) 157–178. [48] A.W. Leissa, The free vibration of rectangular plates, J. Sound Vib. 31 (3) (1973) 257–293. APPENDIX 5.1 Analytical solutions for free vibration analysis of gradient elastic Euler-Bernoulli beam To obtain the natural frequencies of the gradient elastic Euler-Bernoulli beam which is governed by Equation (9), we assume a solution of the form w(x, t) = w̄(x)eiωt substituting the above solution in the governing equation (9), we get w̄iv − g 2 w̄vi − ω2 w̄ = 0 β2 here, β 2 = EI/m, and the above equation has the solution of type w̄(x) = 6 X ci eki x j=1 where, ci are the constants of integration which are determined through boundary conditions and the ki are the roots of the characteristic equation 38 k iv − g 2 k vi − ω2 =0 β2 After applying the boundary conditions listed in section 1.1 we get, [F (ω)]{C} = {0} For non-trivial solution, following condition should be satisfied det[F (ω)] = 0 The above frequency equation renders all the natural frequencies for a gradient elastic Euler-Bernoulli beam. The following are the frequency equations for different boundary conditions. (a) Simply supported beam :   1 1 1 1 1 1     (k1 L) e(k2 L) e(k3 L) e(k4 L) e(k5 L) e(k6 L)   e         2 2 2 2 2   k2 k2 k3 k4 k5 k6   1        2 (k L) 2 (k L) 2 (k L) 2 (k L) 2 (k L) 2 (k L)  2 3 4 5 6 [F (ω)] = k1 e 1  k2 e k3 e k4 e k5 e k6 e           4 4 4 4 4 4 k6  k3 k4 k5 k2  k1         k 4 e(k1 L) k 4 e(k2 L) k 4 e(k3 L) k 4 e(k4 L) k 4 e(k5 L) k 4 e(k6 L)  2 3 4 5 6   1 39 (b) Cantilever beam :   1 1 1 1 1 1     k2 k3 k4 k5 k6   k1         2 2 2 2 2   k2 k k k k k 1 2 3 4 5 6           [F (ω)] =  t1 t2 t3 t4 t5 t6            p2 p3 p4 p5 p6   p1         k 3 e(k1 L) k 3 e(k2 L) k 3 e(k3 L) k 3 e(k4 L) k 3 e(k5 L) k 3 e(k6 L)  6 5 4 3 2   1 (c) clamped beam :   1 1 1 1 1 1     k2 k3 k4 k5 k6   k1         2 2 2 2 2   k2 k k k k k 2 3 4 5 6  1         (k L)  (k L) (k L) (k L) (k L) (k L) 2 3 4 5 6 [F (ω)] =  e 1  e e e e e          (k1 L) (k2 L) (k3 L) (k4 L) (k5 L) (k6 L)  k e k e k e k e k e k e  1  2 3 4 5 6         k 2 e(k1 L) k 2 e(k2 L) k 2 e(k3 L) k 2 e(k4 L) k 2 e(k5 L) k 2 e(k6 L)  2 3 4 5 6  1  40 (d) Propped cantilever beam :   1 1 1 1 1 1     k2 k3 k4 k5 k6   k1         2 2 2 2 2   k2 k k k k k 1 2 3 4 5 6           (k2 L) (k3 L) (k4 L) (k5 L) (k6 L)  [F (ω)] =  e(k1 L) e e e e e          2 (k1 L) 2 (k2 L) 2 (k3 L) 2 (k4 L) 2 (k5 L) 2 (k6 L)  k2 e k3 e k4 e k5 e k6 e k1 e           p  p p p p p 1 2 3 4 5 6   (e) Free-free beam :   q2 q3 q4 q5 q6   q1         r2 r3 r4 r5 r6   r1         3 3 3 3 3   k3 k k k k k 1 2 5 6 4 3         [F (ω)] =     t1 t t t t t 2 3 4 5 6           p p p p p p   1 2 3 4 5 6         k 3 e(k1 L) k 3 e(k2 L) k 3 e(k3 L) k 3 e(k4 L) k 3 e(k5 L) k 3 e(k6 L)  1 2 3 4 5 6   Where, 41 t1 = (k13 − g 2 k1 5 )e(k1 L) , t2 = (k23 − g 2 k2 5 )e(k2 L) , t3 = (k33 − g 2 k3 5 )e(k3 L) t4 = (k43 − g 2 k4 5 )e(k4 L) , t5 = (k53 − g 2 k5 5 )e(k5 L) p1 = (k12 − g 2 k1 4 )e(k1 L) , p2 = (k22 − g 2 k2 4 )e(k2 L) , p3 = (k32 − g 2 k3 4 )e(k3 L) p4 = (k42 − g 2 k4 4 )e(k4 L) , p5 = (k52 − g 2 k5 4 )e(k5 L) , p6 = (k62 − g 2 k6 4 )e(k6 L) t6 = (k63 − g 2 k6 5 )e(k6 L) q1 = (k13 − g 2 k1 5 ), q2 = (k23 − g 2 k2 5 ), q4 = (k43 − g 2 k4 5 ), q5 = (k53 − g 2 k5 5 ) q6 = (k63 − g 2 k6 5 ) r1 = (k12 − g 2 k1 4 ), r2 = (k22 − g 2 k2 4 ), r3 = (k32 − g 2 k3 4 ) r4 = (k42 − g 2 k4 4 ), r5 = (k52 − g 2 k5 4 ), r6 = (k62 − g 2 k6 4 ) 42 q3 = (k33 − g 2 k3 5 ) 

Unmatched Perturbation Accommodation for an Aerospace Launch Vehicle Autopilot Using Dynamic Sliding Manifolds M. R. Saniee Department of Aerospace Engineering, Indian Institute of Science, Bangalore, India Abstract- Sliding mode control of a launch vehicle during its atmospheric flight phase is studied in the presence of unmatched disturbances. Linear time-varying dynamics of the aerospace vehicle is converted into a systematic formula and then dynamic sliding manifold as an advanced method is used in order to overcome the limited capability of conventional sliding manifolds in minimizing the undesired effects of unmatched perturbations on the control system. At the end, simulation results are evaluated and the performance of two approaches are compared in terms of stability and robustness of the autopilot. Keywords: Launch vehicle control, Variable structures, Unmatched perturbation, Dynamic sliding mode. I. INTRODUCTION To reach the aim of a mission, an aerospace launch vehicle (ALV) must move along the specified trajectory and has a required attitude. These tasks are fulfilled by the vehicle flight control system or autopilot. It forms control actions such as forces and torques to the ALV during its powered path providing the best fulfillment of the specified requirements to the vehicle terminal state vector [1]. Since dynamic equations of an ALV system can be mathematically modeled using estimated and time-varying coefficients, the most critical problem arises due to the variable characteristics of such vehicles [2]. So the attitude control systems are confronting dynamics with uncertain parameters in addition to nonlinearities and disturbances. In order to achieve an acceptable performance, robust controllers are proposed to follow the nominal trajectory [3]. One of the nonlinear robust control techniques is variable structure controls (VSC). It utilizes beneficial characteristics of various control configurations and delivers robust performance and new features that none of those structures possess on their own. The central feature of VSC is the so-called sliding mode control (SMC) on the switching manifold which the system remain insensitive to plant parameter variations and external disturbances [4]. Sliding mode control first introduces in the framework of VSC and soon became the principle operational mode for this class of control systems. Due to practical advantages such as order reduction, low sensitivity to turbulences and plant parameter variations, SMC has been known a very efficient technique to control complicated dynamic plants functioning under variable situations which are common for many processes of modern technologies. The main shortcoming of SMC namely chattering phenomenon arises due to switching in the both sides of sliding manifold. This dilemma can be treated by continuous approximation of discontinuous control or by continuous SMC design [6]. SMC also can not accommodate for unmatched disturbances unlike its powerful application for matched disturbance rejection. To obliterate this problem encountered in practice for flight dynamics and timescale separation of central loops in multi-loop systems, dynamic sliding mode (DSM) has received considerable attention [7]. DSM exploits benefits of dynamic compensators to offer a switching manifold in order to provide the systems with robustness characteristics against unmatched perturbations [8]. This technique can be applied in variety of complex control systems and even automated passive acoustic monitoring devices used for studying marine mammal vocalizations and their behaviors [9-12]. In this analysis enhanced properties of a control, designed based on DSM for longitudinal channel output tracking of a time varying ALV will be demonstrated in comparison to that of CSM control. Section 2 and 3 presents CSM and DSM control theory, respectively. ALV dynamics are offered in Sec. 4. CSM and DSM autopilot designed and simulation results demonstrated in Sec. 5. Section 6 devoted for conclusion. II. CSM CONTROL THEORY Consider the following dynamic model: x  Αx  Βu  F(x, t ) (1) where x(t )  R and u (t ) are the state vector are the control vector, respectively; A and B are constant matrices; F(x,t) is a bounded disturbance; It is assumed that {A,B} is a n controllable pair and rank(B)=m. The conventional sliding surface S(x,t) can be defined as: d S ( x, t )  (  λ) n1 e (2) dt where η is a small positive constant. Therefore, a control input can be chosen as: where λ is a positive real constant and e is tracking error as: (3) e  x  xd where ρ is positive real constant, ueq is the continuous u  ( λB) 1 (e  λAx  xd  ρsign( s ))  u eq  u disc control which is called “equivalent control” and udisc is the discontinuous control that switches around the sliding surface and so, system state synchronously moves on this manifold and toward the origin. where xd is the nominal trajectory and e(0)=0. In this study, it is assumed that n=2 and sliding surface can be determined in terms of error as follows: (4) S  e  λe III. S ( x, t )  0 S ( x, t )  0 (5) So, the system dynamics moves from any initial states toward the sliding hyperplanes in a certain amount of time and maintains on it hereafter [4]. In other words, the existence of the conventional sliding mode in the designed sliding manifold is provided. This two-stage design becomes simpler for systems in socalled regular form. Since rank{B}=m, matrix B in Eq. (1) can be partitioned as : B  (6) B   1 B 2  ( x2 , e)  x2  W (s)e  x1  A 11 x1  A 12 x 2  F1   x 2  A 21 x1  A 22 x 2  u  F2 P( s ) and P(s), Q(s) are Q( s ) polynomials of s . The operator W (s) has to be specified in order to provide the desired plant behavior as well as rejecting effects of unmatched disturbance F1 . To ensure the occurance of the sliding mode on the sliding manifold (12), the discontinuous control should be designed. By using the Lyapunov’s function (9) and the reachibility condition (10), the control law can be given as: u   A21 x1  A22 x2  W ( s)e  sign() (7)  u eq  u disc (13) The existence of the sliding mode in the dynamic sliding surface can be proven if ( x2 , e)  0 , derivative of (9) can be identified as: (8) where x1  R nm , x2  R m , A ij are constant matrices for   x  W (s)e V  T  2 i, j  1,2 , F1 is unmatched disturbance and F2 is matched disturbance. Lyapunov direct method is used to obtain the control law. A candidate function is selected as: V  1 sT s 2 (12) where x2  R m , s  d / dt , W ( s)  where B1  R ( nm )m and B 2  R mm with det(B 2 )  0 . The nonsingular coordinate transformation,  x1  , T  R nn  x   Tx  2 converts the system Equation (1) to regular form: DSM CONTROL APPROACH The main characteristic of the dynamic sliding mode is being compensator. It means that DSM control designs control law of each step based on previous step data and so, the system may need some additional dynamics to improve the system and sliding mode stability besides the desired system response. Dynamic sliding manifold can be modeled as a linear function in terms of some states and tracking error as follows [7]: The tracking error will asymptotically reach zero with a control law of bellow form: u  , u  u , (11) (14) Substituting (8) and (13) into (14) should yield the following expression: V   ρsign()  0 ρ  0 (15) (9) Consequently, the surface   0 is attractive and DSM control provides asymptotic stability to the states of tracking error dynamics. with V (0)  0 and V (s)  0 . The condition, guaranteeing an ideal sliding motion, is the η-reachibility condition given by: 1 d 2 s  η s (10) 2 dt IV. 2 THE EQUATIONS OF MOTION FOR ALV Newton’s second law can be employed to extract the motion equations of an ALV. Assuming rigid airframe for the vehicle, the 6DOF equations of motion obtained as follows [2]: Fx  m(U  qW  rV ) F  m(V  rU  pW ) The trust vector deflection of servo dynamics can be described as: TF servo  δ  1 (16) δc 0.1s  1 with a rate limit of | ddtδ | 25 deg/ sec . Because reference signal is pitch rate, a rate gyro is selected as follows: y Fz  m(W  pV  qU ) M x  I x p (14) TF gyro  M y  I y q  ( I x  I y ) pr M z  I z r  ( I y  I x ) pq V. Since altitude control systems of an ALV are usually simplified into a linear set of dynamical equations, a linear ALV model is developed in this article. Considering small perturbations, linearized equations of motion can be obtained as follows [1]: (17) AUTOPILOT DESIGN AND SIMULATION RESULTS In this section, both CSM and DSM control designed for the time varying ALV pitch longitudinal channel and excellent performance of DSM in comparison to that of CSM are demonstrated. v z  Z v v z  Z q q  Z θ θ  Z δe δe q  M vz v z  M q q  M δe δe v y  Z v v y  Z r r  Z θ θ  Z δr δr (80π ) 2 s  (40π ) s  (80π ) 2 2 A. CSM Control Design The goal is to generate the control δe to enforce state motion on CSM: S  θe  Kθ e (18) (15) r  M vy v y  M r r  M δr δr p  M p p  M δa δa where θe  θc  θ and K  const. is chosen in order to make ALV track the commanded pitch rate qc and so, the states trajectory of system asymptotically converge to the sliding manifold S  0 . Using Lyapunov function (9), its derivative is derived as: where Z, M are dynamic coefficients and δ is deflection of trust vector. Since the control objective is to track guidance command in pitch channel, thus the two first equations will be regarded as required dynamics and the other three ones are waved belonging to yaw and roll channels. Time varying coefficients of pitch dynamics in Eq. (15) are shown as in Figure 1. V  S T S  S T [q c  M vz v z  M q q  M δe δe  Kθe ] (19) By utilizing the equality form of (10) for ensuring asymptotic stability of the system, the necessary control is given as: δe  M δe1[q c  Kq e  M vz v z  M q q  ρsign(S )] (20) where K  1 and ρ  0.01 have been selected. The control law (20) is discontinuous and will cause chattering on the manifold (18). To solve this undesired phenomenon, the discontinuous term sign(S ) in Eq. (20) is replaced by the continuous term sat(S / ε) , where ε is a real small constant whose value is chosen 103 in this research. B. DSM Control Design Following the procedure in [7], the design procedure for dynamic sliding manifold is presented. Note that to transform the ALV longitudinal equations of motion to regular form (8) and avoiding singularity, the servo dynamic equation (16) is added to plant equations. Thus, δ is converted to one of system states and δc will be the control effort . Figure 1. Longitudinal dynamic coefficient 3 Based on Eq. (12), the following expression for dynamic sliding manifold is defined:   δ  W (s)e properly accommodate for matched disturbances [4], simulation was run in presence of unmatched disturbances depicted in Figure 3 such that f11 and f12 are exerted to the first and second expressions in Eq. (15), respectively. The simulation results with dynamic and conventional SMC are illustrated in Figure 4 and Figure 5, respectively. It is shown that DSM unlike the CSM can follow the nominal trajectory very closely and withstands the unmatched perturbations. (21) which W (s) can be selected as bellow: a1s 2  a2 s  a3 (22) b1s 2  b2 s where a1 , a2 , a3 , b1 , b2 are real indices determined for each iteration. In order to obtain these coefficients, tracking error achieved as: W ( s)  qc (23) 1  W ( s)G( s) where G(s) is the transfer function of q(s) relative to δ (s) . By comparing characteristics equation for (23) and integral of time multiplied by absolute tracking error criterion with: e s 5  2.8wn s 4  5wn2 s 3  5.5wn3 s 2  3.4wn4 s  wn5  0 (24) where wn  10 Hz is chosen and related parameters identified at each moment. Also, the control δc is given as follows: (25) δc   ρsat ( / ε) Figure 3. Unmatched disturbances profiles where ρ  1 and ε  103 is considered. Figure 2. Desired pitch and pitch rate to be tracked Figure 4. Pitch angle error, pitch rate error and controller command obtained from CSM autopilot In this paper, pitch rate program has been designed offline as shown in Figure 2 and it is desired to be tracked during the flight envelope. Since sliding mode control can In this research, it is assumed that a1  b1  1 and the other three coefficients are determined by the 4 corresponding algorithm at each moment during the ALV flight time whose variations are illustrated in Figure 6. conventional and dynamic SMC was designed and closedloop system operations of these methods were compared. Results show that dynamic SMC can accommodates unmatched disturbances and output tracking errors is much less than those of CSM, while conventional SMC does not operate properly and cannot satisfy requirements of system performance. The simple and straightforward design procedure, together with the encouraging robustness against unmatched disturbances; invite further application of this approach. REFERENCES [1] V. V. Malyshev, “Aerospace vehicle Control: Modern Theory and Applications,” IAE and MAI Russia Cooperation, Feb. 1985. [2] J. H. Blacklock, “Automatic Control of Aircraft and Missiles,” Wiley, 1991. [3] V. V. Malyshev, and M. N. Krasilshikov, “Aerospace vehicle Navigation and Control,” FAPESP Publication, Sao-Paulo, Brazil, 1996. [4] V. I. Utkin, “Sliding Modes in Control and Optimization,” 111-130, Springer-Verlag, Berlin, 1992. [5] M. Esfahanian, B. Ebrahimi, J. Roshanian, and M. Bahrami, “Time-Varying Dynamic Sliding Mode Control of a Nonlinear Aerospace Launch Vehicle,” Canadian Aeronautics and Space Journal, Vol. 57, No. 2, 2011. [6] Y. B. Shtessel, “Nonlinear Sliding Manifolds in Nonlinear Output Tracking Problem,” In the Proceeding of the American Control Conference Seattle, Washington, June, 1995. [7] Y. B. Shtessel, “Nonlinear Output Tracking in Conventional and Dynamic Sliding Manifolds,” IEEE Transactions on Automatic Control, Vol. 42, No. 9, 1997, pp. 1282-1286. [8] A. J. Koshkouei, K. J. Burnham and A. S. I. Zinober, “Dynamic Sliding Mode Control Design,” IEEE Proceedings-Control Theory and Applications, Vol. 152, No. 4, pp. 392_396, 2005. [9] M. Esfahanian, H. Zhuang, and N. Erdol, “On Contour-Based Classification of Dolphin Whistles by Type,” Journal of Applied Acoustics, Vol. 76, pp. 274-279, Feb 2014. [10] W. C. L. Filho and L. Hsu, “Adaptive Control of Missile Attitude,” IFAC Adaptive Systems in Control and Signal Processing, Lund, Sweden, 1995. [11] M. Esfahanian, H. Zhuang, and N. Erdol, and E. Gerstein, “Comparison of two methods for detection of north Atlantic right whale upcalls,” 2015 IEEE European Signal Processing Conference, Aug 31-Sep 4, Nice, France, 2015. [12] A. Nayir, E. Rosolowski, and L. Jedut, “New trends in wind energy modeling and wind turbine control,” IJTPE journal, Vol. 2, No. 4, pp. 51-59, September 2010. Figure 5. Pitch angle error, pitch rate error and controller command obtained from DSM autopilot . Figure 6. Indices variations of W (s) calculated on-line VI. CONCLUSION In this research, commanded pitch rate tracking with unmatched disturbances for the atmospheric flight of a time-varying ALV is considered in SMC. Both 5 

GRADED INTEGRAL CLOSURES FRED ROHRER arXiv:1302.1659v2 [math.AC] 8 Apr 2013 Abstract. It is investigated how graded variants of integral and complete integral closures behave under coarsening functors and under formation of group algebras. Introduction Let G be a group, let R be a G-graded ring, and let S be a G-graded R-algebra. (Throughout, monoids, groups and rings are understood to be commutative, and algebras are understood to be commutative, unital and associative.) We study a graded variant of (complete) integral closure, defined as follows: We denote by Int(R, S) (or CInt(R, S), resp.) the G-graded sub-R-algebra of S generated by the homogeneous elements of S that are (almost) integral over R and call this the (complete) integral closure of R in S. If R is entire (as a G-graded ring, i.e., it has no homogeneous zero-divisors) we consider its graded field of fractions Q(R), i.e., the G-graded R-algebra obtained by inverting all nonzero homogeneous elements, and then Int(R) = Int(R, Q(R)) (or CInt(R) = CInt(R, Q(R)), resp.) is called the (complete) integral closure of R. These constructions behave similar to their ungraded relatives, as long as we stay in the category of G-graded rings. But the relation between these constructions and their ungraded relatives, and more generally their behaviour under coarsening functors, is less clear; it is the main object of study in the following. For an epimorphism of groups ψ : G ։ H we denote by •[ψ] the ψ-coarsening functor from the category of G-graded rings to the category of H-graded rings. We ask for conditions ensuring that ψ-coarsening commutes with relative (complete) integral closure, i.e., Int(R, S)[ψ] = Int(R[ψ] , S[ψ] ) or CInt(R, S)[ψ] = CInt(R[ψ] , S[ψ] ), or – if R and R[ψ] are entire – that ψ-coarsening commutes with (complete) integral closure, i.e., Int(R)[ψ] = Int(R[ψ] ) or CInt(R)[ψ] = CInt(R[ψ] ). Complete integral closure being a more delicate notion than integral closure, it is not astonishing that the questions concerning the former are harder than the other ones. Furthermore, the case of integral closures of entire graded rings in their graded fields of fractions turns out to be more complicated than the relative case, because Q(R)[ψ] almost never equals Q(R[ψ] ), hence in addition to the coarsening we also change the overring in which we form the closure. The special case H = 0 of parts of these questions was already studied by several authors. Bourbaki ([3, V.1.8]) treats torsionfree groups G, Van Geel and Van Oystaeyen 2010 Mathematics Subject Classification. Primary 13B22; Secondary 13A02, 16S34. Key words and phrases. Graded ring, coarsening, integral closure, integrally closed ring, complete integral closure, completely integrally closed ring, group algebra. The author was supported by the Swiss National Science Foundation. 1 2 FRED ROHRER ([13]) consider G = Z, and Swanson and Huneke ([11, 2.3]) discuss the case that G is of finite type. Our main results, generalising these partial results, are as follows. Theorem 1 Let ψ : G ։ H be an epimorphism of groups and let R be a G-graded ring. a) If Ker(ψ) is contained in a torsionfree direct summand of G then ψ-coarsening commutes with relative (complete) integral closure. b) Suppose that R is entire. If G is torsionfree, or if Ker(ψ) is contained in a torsionfree direct summand of G and the degree support of R generates G, then ψ-coarsening commutes with integral closure. The questions above are closely related to the question of how (complete) integral closure behaves under formation of group algebras. If F is a group, there is a canonical G ⊕ F -graduation on the algebra of F over R; we denote the resulting G ⊕ F -graded ring by R[F ]. We ask for conditions ensuring that formation of graded group algebras commutes with relative (complete) integral closure, i.e., Int(R, S)[F ] = Int(R[F ], S[F ]) or CInt(R, S)[F ] = CInt(R[F ], S[F ]), or – if R is entire – that formation of graded group algebras commutes with (complete) integral closure, i.e., Int(R)[F ] = Int(R[F ]) or CInt(R)[F ] = CInt(R[F ]). Our main results are the following. Theorem 2 Let G be a group and let R be a G-graded ring. Formation of graded group algebras over R commutes with relative (complete) integral closure. If R is entire then formation of graded group algebras over R commutes with (complete) integral closure. It is maybe more interesting to consider a coarser graduation on group algebras, namely the G-graduation obtained from R[F ] by coarsening with respect to the canonical projection G ⊕ F ։ G; we denote the resulting G-graded R-algebra by R[F ][G] and call it the coarsely graded algebra of F over R. We ask for conditions ensuring that formation of coarsely graded group algebras commutes with relative (complete) integral closure, i.e., Int(R, S)[F ][G] = Int(R[F ][G] , S[F ][G] ) or CInt(R, S)[F ][G] = CInt(R[F ][G] , S[F ][G]), or – if R and R[F ][G] are entire – that formation of coarsely graded group algebras commutes with (complete) integral closure, i.e., Int(R)[F ][G] = Int(R[F ][G] ) or CInt(R)[F ][G] = CInt(R[F ][G] ). Ungraded variants of these questions (i.e., for G = 0) for a torsionfree group F were studied extensively by Gilmer ([6, §12]). On use of Theorems 1 and 2 we will get the following results. Theorem 3 Let G and F be groups and let R be a G-graded ring. Formation of the coarsely graded group algebra of F over R commutes with relative (complete) integral closure if and only if F is torsionfree. If R is entire and F is torsionfree then formation of the coarsely graded group algebra of F over R commutes with integral closure. Some preliminaries on graded rings, coarsening functors, and algebras of groups are collected in Section 1. Relative (complete) integral closures are treated in Section 2, and (complete) integral closures of entire graded rings in their graded fields of fractions are treated in Section 3. Our notation and terminology follows Bourbaki’s Éléments de mathématique. GRADED INTEGRAL CLOSURES 3 Before we start, a remark on notation and terminology may be appropriate. Since we try to never omit coarsening functors (and in particular forgetful functors) from our notations it seems conceptually better and in accordance with the general yoga of coarsening to not furnish names of properties of G-graded rings or symbols denoting objects constructed from G-graded rings with additional symbols that highlight the dependence on G or on the graded structure. For example, if R is a G-graded ring then we will denote by Nzd(R) (instead of, e.g., NzdG (R)) the monoid of its homogeneous non-zerodivisors, and we call R entire (instead of, e.g., G-entire) if Nzd(R) consists of all homogeneous elements of R different from 0. Keeping in mind that in this setting the symbol “R” always denotes a G-graded ring (and never, e.g., its underlying ungraded ring), this should not lead to confusions (whereas mixing up different categories might do so). Throughout the following let G be a group. 1. Preliminaries on graded rings First we recall our terminology for graded rings and coarsening functors. (1.1) By a G-graded ring we mean a pair (R, (Rg )g∈G ) consisting of a ring R and a family (Rg )g∈G of subgroups of the additive group of R whose direct sum equals the additive group of R such that Rg Rh ⊆ Rg+h for g, h ∈ G. If no confusion can arise we denote a G-graded ring (R, (Rg )g∈G ) just by R. Accordingly, for a G-graded ring R and S g ∈ G we denote by Rg the component of degree g of R. We set Rhom := g∈G Rg and call degsupp(R) := {g ∈ G | Rg 6= 0} the degree support of R. We say that R has full support if degsupp(R) = G and that R is trivially G-graded if degsupp(R) = {0}. Given G-graded rings R and S, by a morphism of G-graded rings from R to S we mean a morphism of rings u : R → S such that u(Rg ) ⊆ Sg for g ∈ G. By a G-graded R-algebra we mean a G-graded ring S together with a morphism of G-graded rings R → S. We denote by GrAnnG the category of G-graded rings with this notion of morphism. This category has inductive and projective limits. In case G = 0 we canonically identify GrAnnG with the category of rings. (1.2) Let ψ : G ։ H be an epimorphism of groups. For a G-graded ring R we define an H-graded ring R[ψ] , called the ψ-coarsening of R; its underlying ring is the ring underlying L R, and its H-graduation is given by (R[ψ] )h = g∈ψ−1 (h) Rg for h ∈ H. A morphism u : R → S of G-graded rings can be considered as a morphism of H-graded rings R[ψ] → S[ψ] , and as such it is denoted by u[ψ] . This gives rise to a functor •[ψ] : GrAnnG → GrAnnH . This functor has a right adjoint, hence commutes with inductive limits, and it has a left adjoint if and only if Ker(ψ) is finite ([10, 1.6; 1.8]). For a further epimorphism of groups ϕ : H ։ K we have •[ϕ◦ψ] = •[ϕ] ◦ •[ψ] . (1.3) We denote by that G = limU ∈F U. −→ G FG the set of subgroups of finite type of G, ordered by inclusion, so 4 FRED ROHRER (1.4) Let F ⊆ G be a subgroup. For a G-graded ring R we define an F -graded ring R(F ) L with underlying ring the subring g∈F Rg ⊆ R and with F -graduation (Rg )g∈F . For an F -graded ring S we define a G-graded ring S (G) with underlying ring the ring underlying (G) (G) S and with G-graduation given by Sg = Sg for g ∈ F and Sg = 0 for g ∈ G \ F . If R is a G-graded ring and F is a set of subgroups of G, ordered by inclusion, whose inductive limit is G, then R = limF ∈F ((R(F ) )(G) ). −→ The next remark recalls the two different notions of graded group algebras and, more general, of graded monoid algebras. (1.5) Let M be a cancellable monoid, let F be its group of differences, and let R be a G-graded ring. The algebra of M over R, furnished with its canonical G ⊕ F -graduation, is denoted by R[M] and called the finely graded algebra of M over R, and we denote by (ef )f ∈F its canonical basis. So, for (g, f ) ∈ G ⊕ F we have R[M](g,f ) = Rg ef . Denoting by π : G ⊕ F ։ G the canonical projection we set R[M][G] := R[M][π] and call this the coarsely graded algebra of M over R. If S is a G-graded R-algebra then S[M] is a G ⊕ F -graded R[M]-algebra, and S[M][G] is a G-graded R[M][G] -algebra. We have R[F ] = limU ∈F R[U](G⊕F ) and R[F ][G] = limU ∈F R[U][G] (1.3). −→ −→ F F We will need some facts about graded variants of simplicity (i.e., the property of “being a field”) and entirety. Although they are probably well-known, we provide proofs for the readers convenience. Following Lang we use the term “entire” instead of “integral” (to avoid confusion with the notion of integrality over some ring which is central in this article) or “domain” (to avoid questions as whether a “graded domain” is the same as a “domain furnished with a graduation”), and we use the term “simple” (which is more common in noncommutative algebra) instead of “field” for similar reasons. (1.6) Let R be a G-graded ring. We denote by R∗ the multiplicative group of invertible homogeneous elements of R and by Nzd(R) the multiplicative monoid of homogeneous non-zerodivisors of R. We call R simple if R∗ = Rhom \ 0, and entire if Nzd(R) = Rhom \ 0. If R is entire then the G-graded ring of fractions Nzd(R)−1 R is simple; we denote it by Q(R) and call it the (graded) field of fractions of R. If ψ : G ։ H is an epimorphism of groups and R[ψ] is simple or entire, then R is so. Let F ⊆ G be a subgroup. If R is simple or entire, then R(F ) is so, and an F -graded ring S is simple or entire if and only if S (G) is so. (1.7) Let I be a nonempty right filtering preordered set, and let ((Ri )i∈I , (ϕij )i≤j ) be an inductive system in GrAnnG over I. Analogously to [2, I.10.3 Proposition 3] we see that if Ri is simple or entire for every i ∈ I, then limi∈I Ri is simple or entire. If Ri is entire for −→ every i ∈ I and ϕij is a monomorphism for all i, j ∈ I with i ≤ j, then by [8, 0.6.1.5] we get an inductive system (Q(Ri ))i∈I in GrAnnG over I with limi∈I Q(Ri ) = Q(limi∈I Ri ). −→ −→ (1.8) Let F ⊆ G be a subgroup and let ≤ be an ordering on F that is compatible with its structure of group. The relation “g − h ∈ F≤0 ” is the finest ordering on G that is compatible with its structure of group and induces ≤ on F ; we call it the canonical GRADED INTEGRAL CLOSURES 5 extension of ≤ to G. If ≤ is a total ordering then its canonical extension to G induces a total ordering on every equivalence class of G modulo F . (1.9) Lemma Let ψ : G ։ H be an epimorphism of groups such that Ker(ψ) is torhom sionfree, let R be an entire G-graded ring, and let x, y ∈ R[ψ] \ 0 with xy ∈ Rhom . Then, x, y ∈ Rhom and xy 6= 0. Proof. (cf. [2, II.11.4 Proposition 8]) By [2, II.11.4 Lemme 1] we can choose a total ordering on Ker(ψ) that is compatible with its structure of group. Let ≤ denote its canonical extension to G (1.8). Let h := deg(x) and h′ := deg(y). There exist strictly increasing finite sequences (gi )ni=0 in ψ −1 (h) and (gj′ )m in ψ −1 (h′ ), xi ∈ Rgi \ 0 for i ∈ [0, n], and j=0 Pn P yj ∈ Rgj′ \ 0 for j ∈ [0, m] such that x = i=0 xi and y = m j=0 yj . If k ∈ [0, n] and ′ ′ l ∈ [0, m] with gk + gl = gn + gm then k = n and l = m by [2, VI.1.1 Proposition 1]. This ′ implies that the component of xy of degree gn + gm equals xn ym 6= 0, so that xy 6= 0. As hom ′ ′ , hence n = m = 0 and therefore x0 y0 6= 0 and xy ∈ R it follows g0 + g0 = gn + gm hom x, y ∈ R .  (1.10) Corollary Let ψ : G ։ H be an epimorphism of groups such that Ker(ψ) is ∗ . torsionfree, and let R be an entire G-graded ring. Then, R∗ = R[ψ] hom hom \ 0 with \ 0 then there exists y ∈ R[ψ] Proof. Clearly, R∗ ⊆ (R[ψ] )∗ . If x ∈ (R[ψ] )∗ ⊆ R[ψ] hom hom ∗ xy = 1 ∈ R , so 1.9 implies x ∈ R , hence x ∈ R .  (1.11) Let R be a G-graded ring and let F be a group. It is readily checked that R[F ] is simple or entire if and only if R is so. Analogously to [6, 8.1] it is seen that R[F ][G] is entire if and only if R is entire and F is torsionfree. Together with 1.10 it follows that R[F ][G] is simple if and only if R is simple and F = 0. (1.12) Proposition Let ψ : G ։ H be an epimorphism of groups. a) The following statements are equivalent: (i) ψ is an isomorphism; (ii) ψ-coarsening preserves simplicity. b) The following statements are equivalent: (i) Ker(ψ) is torsionfree; (ii) ψ-coarsening preserves entirety; (iii) ψ-coarsening maps simple G-graded rings to entire H-graded rings.1 Proof. If K is a field and R = K[Ker(ψ)](G) , then R is simple and R[ψ] is trivially Hgraded, hence R[ψ] is simple or entire if and only if R[0] is so (1.11, 1.6). If Ker(ψ) 6= 0 then R[0] is not simple, and if Ker(ψ) is not torsionfree then R[0] is not entire (1.11). This proves a) and the implication (iii)⇒(i) in b). The remaining claims follow from 1.9.  (1.13) A G-graded ring R is called reduced if 0 is its only nilpotent homogeneous element. With arguments similar to those above one can show that statements (i)–(iii) in 1.12 b) are also equivalent to the following: (iv) ψ-coarsening preserves reducedness; (v) ψ-coarsening maps simple G-graded rings to reduced H-graded rings. We will make no use of this fact. 1In case H = 0 the implication (i)⇒(ii) is [2, II.11.4 Proposition 8]. 6 FRED ROHRER Finally we make some remarks on a graded variant of noetherianness. (1.14) Let R be a G-graded ring. We call R noetherian if ascending sequences of Ggraded ideals of R are stationary, or – equivalently – if every G-graded ideal of R is of finite type. Analogously to the ungraded case one can prove a graded version of Hilbert’s Basissatz: If R is noetherian then so are G-graded R-algebras of finite type. If ψ : G ։ H is an epimorphism of groups and R[ψ] is noetherian, then R is noetherian. Let F ⊆ G be a subgroup. It follows from [4, 2.1] that if R is noetherian then so is R(F ) . Moreover, an F -graded ring S is noetherian if and only if S (G) is so. If F is a group then it follows from [4, 2.1] and the fact that ef ∈ R[F ]∗ for f ∈ F that R[F ] is noetherian if and only if R is so. Analogously to [6, 7.7] one sees that R[F ][G] is noetherian if and only if R is noetherian and F is of finite type. More general, it follows readily from a result by Goto and Yamagishi ([4, 1.1]) that G is of finite type if and only if ψ-coarsening preserves noetherianness for every epimorphism of groups ψ : G ։ H. (This was proven again two years later by Nǎstǎsescu and Van Oystaeyen ([9, 2.1]).) 2. Relative integral closures We begin this section with basic definitions and first properties of relative (complete) integral closures. (2.1) Let R be a G-graded ring and let S be a G-graded R-algebra. An element x ∈ S hom is called integral over R if it is a zero of a monic polynomial in one indeterminate with coefficients in Rhom . This is the case if and only if x, considered as an element of S[0] , is integral over R[0] , as is seen analogously to the first paragraph of [3, V.1.8]. Hence, using [3, V.1.1 Théorème 1] we see that for x ∈ S hom the following statements are equivalent: (i) x is integral over R; (ii) the G-graded R-module underlying the G-graded R-algebra R[x] is of finite type; (iii) there exists a G-graded sub-R-algebra of S containing R[x] whose underlying G-graded R-module is of finite type. An element x ∈ S hom is called almost integral over R if there exists a G-graded sub-Rmodule T ⊆ S of finite type containing R[x]. This is the case if and only if x, considered as an element of S[0] , is almost integral over R[0] . Indeed, this condition is obviously necessary. It is also sufficient, for if T ⊆ S[0] is a sub-R[0] -module of finite type containing R[0] [x] then the G-graded sub-R-module T ′ ⊆ S generated by the set of homogeneous components of elements of T is of finite type and contains T , hence R[x]. It follows from the first paragraph that if x ∈ S hom is integral over R then it is almost integral over R; analogously to [3, V.1.1 Proposition 1 Corollaire] it is seen that the converse is true if R is noetherian (1.14). (2.2) Let R be a G-graded ring and let S be a G-graded R-algebra. The G-graded sub-R-algebra of S generated by the set of elements of S hom that are (almost) integral over R is denoted by Int(R, S) (or CInt(R, S), resp.) and is called the (complete) integral closure of R in S. We have Int(R, S) ⊆ CInt(R, S), with equality if R is noetherian (2.1). For an epimorphism of groups ψ : G ։ H we have Int(R, S)[ψ] ⊆ Int(R[ψ] , S[ψ] ) and CInt(R, S)[ψ] ⊆ CInt(R[ψ] , S[ψ] ) (2.1). GRADED INTEGRAL CLOSURES 7 Let R′ denote the image of R in S. We say that R is (completely) integrally closed in S if R′ = Int(R, S) (or R′ = CInt(R, S), resp.), and that S is (almost) integral over R if Int(R, S) = S (or CInt(R, S) = S, resp.). If R is completely integrally closed in S then it is integrally closed in S, and if S is integral over R then it is almost integral over R; the converse statements are true if R is noetherian. If ψ : G ։ H is an epimorphism of groups, then S is (almost) integral over R if and only if S[ψ] is (almost) integral over R[ψ] , and if R[ψ] is (completely) integrally closed in S[ψ] then R is (completely) integrally closed in S. If G ⊆ F is a subgroup then Int(R, S)(F ) = Int(R(F ) , S (F ) ) and CInt(R, S)(F ) = CInt(R(F ) , S (F ) ), hence R is (completely) integrally closed in S if and only if R(F ) is (completely) integrally closed in S (F ) . From [3, V.1.1 Proposition 4 Corollaire 1] and [12, §135, p. 180]2 we know that sums and products of elements of S[0] that are (almost) integral over R[0] are again (almost) integral over R[0] . Hence, Int(R, S)hom (or CInt(R, S)hom , resp.) equals the set of homogeneous elements of S that are (almost) integral over R, and thus Int(R, S) (or CInt(R, S), resp.) is (almost) integral over R by the above. Moreover, Int(R, S) is integrally closed in S by [3, V.1.2 Proposition 7]. One should note that CInt(R, S) is not necessarily completely integrally closed in S, not even if R is entire and S = Q(R) ([7, Example 1]). (2.3) Suppose we have a commutative diagram of G-graded rings R /   R S / h S. hom is (almost) integral over R If x ∈ S hom is (almost) integral over R, then h(x) ∈ S (2.1, [3, V.1.1 Proposition 2], [5, 13.5]). Hence, if the diagram above is cartesian and R is (completely) integrally closed in S, then R is (completely) integrally closed in S. (2.4) Let R be a G-graded ring, let S be a G-graded R-algebra, and let T ⊆ Rhom be a subset. Analogously to [3, V.1.5 Proposition 16] one shows that T −1 Int(R, S) = Int(T −1 R, T −1S). Hence, if R is integrally closed in S then T −1 R is integrally closed in T −1 S. Note that there is no analogous statement for complete integral closures. Although −1 T CInt(R, S) ⊆ CInt(T −1 R, T −1 S) by 2.3, this need not be an equality. In fact, by [3, V.1 Exercice 12] there exists an entire ring R that is completely integrally closed in Q(R) and a subset T ⊆ R \ 0 such that Q(R) is the complete integral closure of T −1 R. (2.5) Let I be a right filtering preordered set, and let (ui )i∈I : (Ri )i∈I → (Si )i∈I be a morphism of inductive systems in GrAnnG over I. By 2.3 we have inductive systems (Int(Ri , Si ))i∈I and (CInt(Ri , Si ))i∈I in GrAnnG over I, and we can consider the sublimi∈I Ri -algebras −→ limi∈I Int(Ri , Si ) ⊆ limi∈I CInt(Ri , Si ) ⊆ limi∈I Si −→ −→ −→ 2Note that van der Waerden calls “integral” what we call “almost integral”. 8 FRED ROHRER and compare them with the sub-limi∈I Ri -algebras −→ Int(limi∈I Ri , limi∈I Si ) ⊆ CInt(limi∈I Ri , limi∈I Si ) ⊆ limi∈I Si . −→ −→ −→ −→ −→ Analogously to [8, 0.6.5.12] it follows limi∈I Int(Ri , Si ) = Int(limi∈I Ri , limi∈I Si ), hence if −→ −→ −→ Ri is integrally closed in Si for every i ∈ I then limi∈I Ri is integrally closed in limi∈I Si . −→ −→ Note that there is no analogous statement for complete integral closures. Although limi∈I CInt(Ri , Si ) ⊆ CInt(limi∈I Ri , limi∈I Si ) by 2.3, this need not be an equality (but −→ −→ −→ cf. 3.2). In fact, by [3, V.1 Exercice 11 b)] there exist a field K and an increasing family (Rn )n∈N of subrings of K such that Rn is completely integrally closed in Q(Rn ) = K for every n ∈ N and that limn∈N Rn is not completely integrally closed in Q(limn∈N Rn ) = K. −→ −→ Now we turn to the behaviour of finely and coarsely graded group algebras with respect to relative (complete) integral closures. (2.6) Theorem Let R be a G-graded ring. a) Formation of finely graded group algebras over R commutes with relative (complete) integral closure. b) Let S be a G-graded R-algebra, and let F be a group. Then, R is (completely) integrally closed in S if and only if R[F ] is (completely) integrally closed in S[F ]. Proof. a) Let F be a group, and let S be a G-graded R-algebra. Let x ∈ S[F ]hom . There are s ∈ S hom and f ∈ F with x = sef . If x ∈ Int(R, S)[F ]hom then s ∈ Int(R, S)hom , hence s ∈ Int(R[F ], S[F ]) (2.3), and as ef ∈ Int(R[F ], S[F ]) it follows x = sef ∈ Int(R[F ], S[F ]). This shows Int(R, S)[F ] ⊆ Int(R[F ], S[F ]). Conversely, suppose that x ∈ Int(R[F ], S[F ])hom . As ef ∈ R[F ]∗ it follows s ∈ Int(R[F ], S[F ])hom . So, there is a finite subset E ⊆ S[F ]hom such that the G ⊕F -graded sub-R[F ]-algebra of S[F ] generated by E contains R[F ][s]. As eh ∈ R[F ]∗ for every h ∈ F we can suppose E ⊆ S. If n ∈ N then sn is an R[F ]-linear combination of products in E, and comparing the coefficients of e0 shows that sn is an R-linear combination of products in E. Thus, R[s] is contained in the G-graded sub-R-algebra of S generated by E, hence s ∈ Int(R, S), and therefore x ∈ Int(R, S)[F ]. This shows Int(R[F ], S[F ]) ⊆ Int(R, S)[F ]. The claim for complete integral closures follows analogously. b) follows immediately from a).  (2.7) Theorem Let F be a group. The following statements are equivalent:3 (i) Formation of coarsely graded algebras of F over G-graded rings commutes with relative integral closure; (i’) Formation of coarsely graded algebras of F over G-graded rings commutes with relative complete integral closure; (ii) If R is a G-graded ring and S is a G-graded R-algebra, then R is integrally closed in S if and only if R[F ][G] is integrally closed in S[F ][G] ; (ii’) If R is a G-graded ring and S is a G-graded R-algebra, then R is completely integrally closed in S if and only if R[F ][G] is completely integrally closed in S[F ][G] ; 3In case G = 0 the implication (iii)⇒(i) is [3, V.1 Exercice 24]. GRADED INTEGRAL CLOSURES 9 (iii) F is torsionfree. Proof. “(i)⇒(ii)” and “(i’)⇒(ii’)”: Immediately from 2.3. “(ii)⇒(iii)” and “(ii’)⇒(iii)”: Suppose that F is not torsionfree. It suffices to find a noetherian ring R and an R-algebra S such that R is integrally closed in S and that R[F ][0] is not integrally closed in S[F ][0] , for then furnishing R and S with trivial G-graduations it follows that R[F ][G] is not integrally closed in S[F ][G] (2.2). The ring Z is noetherian and integrally closed in Q. By hypothesis there exist g ∈ F \ 0 and n ∈ N>1 with ng = 0, Pn−1 1 i so that eng = 1 ∈ Q[F ][0] . It is readily checked that f := i=0 e ∈ Q[F ][0] \ Z[F ][0] is n g n−1 idempotent. Setting c := 1 + (n − 1)eg ∈ Z[F ][0] we get Pn−1 n−1 i+n−1 Pn−1 n−1 i−1 d := f c = f + (n − 1)f egn−1 = f + i=0 eg = f + i=0 eg = n n f+ n−1 −1 e n g + Pn−2 i=0 n−1 i eg n + n−1 n−1 e n g − n−1 n−1 e n g = f + (n − 1)f = nf ∈ Z[F ][0] . Therefore, f 2 + (c − 1)f − d = f + d − f − d = 0 yields an integral equation for f over Z[F ][0] . Thus, Z[F ][0] is not integrally closed in Q[F ][0] . “(iii)⇒(i)” and “(iii)⇒(i’)”: Without loss of generality suppose G = 0 (2.2). Suppose that F is torsionfree, let R be a ring, and let S be an R-algebra. If n ∈ N then Int(R, S)[Nn ][0] = Int(R[Nn ][0] , S[Nn ][0] ) ([3, V.1.3 Proposition 12]), hence Int(R[Zn ][0] , S[Zn ][0] ) = Int(R, S)[Zn ][0] (2.4). This proves (i) in case F is of finite type, and so we get (i) in general by 1.5, 2.2 and 2.5. It remains to show (i’). The inclusion CInt(R, S)[F ][0] ⊆ CInt(R[F ][0] , S[F ][0] ) follows immediately from 2.2 and 2.3. We prove the converse inclusion analogously to [7, Proposition 1]. Since F is torsionfree we can choose a total ordering ≤ on F that is compatible with its structure of group ([2, II.11.4 Lemme 1]). Let x ∈ CInt(S[F ][0] , R[F ][0] ). There are n ∈ N, a family (xi )ni=1 in S \ 0, and a strictly increasing family (fi )ni=1 in F P with x = ni=1 xi efi . We prove by induction on n that x ∈ CInt(R, S)[F ][0] . If n = 0 this is clear. Suppose that n > 0 and that the claim is true for strictly smaller values of n. There exists a finite subset P ⊆ S[F ][0] with R[F ][0] [x] ⊆ hP iR[F ]. Let Q denote the finite set of coefficients of elements of P . Let k ∈ N. There exists a family (sp )p∈P in R[F ] with P xk = p∈P sp p. By means of the ordering of F we see that xkn is the coefficient of efkn in P xk , hence the coefficient of efkn in p∈P sp p. This latter being an R-linear combination of Q we get xkn ∈ hQiR . It follows R[xn ] ⊆ hQiR , and thus xn ∈ CInt(R, S). So, we get xn efn ∈ CInt(R[F ][0] , S[F ][0] ) (2.2, 2.3), hence x − xn efn ∈ CInt(R[F ][0] , S[F ][0] ), thus x − xn efn ∈ CInt(R, S)[F ][0] by our hypothesis, and therefore x ∈ CInt(R, S)[F ][0] as desired.  (2.8) The proof above shows that 2.7 remains true if we replace “If R is a G-graded ring” by “If R is a noetherian G-graded ring” in (ii) and (ii’). The rest of this section is devoted to the study of the behaviour of relative (complete) integral closures under arbitrary coarsening functors. Although we are not able to characterise those coarsenings with good behaviour, we identify two properties of the coarsening, 10 FRED ROHRER one that implies good behaviour of (complete) integral closures, and one that is implied by good behaviour of (complete) integral closures. (2.9) Let ψ : G ։ H be an epimorphism of groups. We say that ψ-coarsening commutes with relative (complete) integral closure if Int(R, S)[ψ] = Int(R[ψ] , S[ψ] ) (or CInt(R, S)[ψ] = CInt(R[ψ] , S[ψ] ), resp.) for every G-graded ring R and every G-graded R-algebra S. (2.10) Proposition Let ψ : G ։ H be an epimorphism of groups. We consider the following statements: (1) ψ-coarsening commutes with relative integral closure; (1’) ψ-coarsening commutes with relative complete integral closure; hom (2) If R is a G-graded ring, S is a G-graded R-algebra, and x ∈ S[ψ] , then x is integral over R[ψ] if and only if all its homogeneous components (with respect to the G-graduation) are integral over R; hom (2’) If R is a G-graded ring, S is a G-graded R-algebra, and x ∈ S[ψ] , then x is almost integral over R[ψ] if and only if all its homogeneous components (with respect to the G-graduation) are almost integral over R; (3) If R is a G-graded ring and S is a G-graded R-algebra, then R is integrally closed in S if and only if R[ψ] is integrally closed in S[ψ] . (3’) If R is a G-graded ring and S is a G-graded R-algebra, then R is completely integrally closed in S if and only if R[ψ] is completely integrally closed in S[ψ] . Then, we have (1)⇔(2)⇔(3) and (1’)⇔(2’)⇒(3’). Proof. The implications “(1)⇔(2)⇒(3)” and “(1’)⇔(2’)⇒(3’)” follow immediately from the definitions. Suppose (3) is true, let R be a G-graded ring R, and let S be a Ggraded R-algebra. As Int(R, S) is integrally closed in S (2.2) it follows that Int(R, S)[ψ] is integrally closed in S[ψ] , implying Int(R[ψ] , S[ψ] ) ⊆ Int(Int(R, S)[ψ] , S[ψ] ) = Int(R, S)[ψ] ⊆ Int(R[ψ] , S[ψ] ) (2.2) and thus the claim.  The argument above showing that (3) implies (1) cannot be used to show that (3’) implies (1’), as CInt(R, S) is not necessarily completely integrally closed in S (2.2). (2.11) Let ψ : G ։ H be an epimorphism of groups, suppose that there exists a section π : H → G of ψ in the category of groups, and let R be a G-graded ring. For g ∈ G there is a morphism of groups π jR,g : Rg → R[0] [Ker(ψ)], x 7→ xeg+π(ψ(g)) . L π The family (jR,g )g∈G induces a morphism of groups jRπ : g∈G Rg → R[0] [Ker(ψ)][0] that π is readily checked to be a morphism jR : R[ψ] → R[ψ] [Ker(ψ)][H] of H-graded rings. (2.12) Theorem Let ψ : G ։ H be an epimorphism of groups. a) If Ker(ψ) is contained in a torsionfree direct summand of G then ψ-coarsening commutes with relative (complete) integral closure. GRADED INTEGRAL CLOSURES 11 b) If ψ-coarsening commutes with relative (complete) integral closure then Ker(ψ) is torsionfree. Proof. a) First, we consider the case that Ker(ψ) itself is a torsionfree direct summand hom of G. Let R be a G-graded ring, let S be a G-graded R-algebra, and let x ∈ S[ψ] be (almost) integral over R[ψ] . As Ker(ψ) is a direct summand of G there exists a section π of ψ in the category of groups. So, we have a commutative diagram R[ψ] / S[ψ] π jS π jR   R[ψ] [Ker(ψ)][H] / S[ψ] [Ker(ψ)][H] of H-graded rings (2.11). Since Ker(ψ) is torsionfree it follows jSπ (x) ∈ Int(R[ψ] [Ker(ψ)][H] , S[ψ] [Ker(ψ)][H] ) = Int(R[ψ] , S[ψ] )[Ker(ψ)][H] (and similarly for complete integral closures) by 2.3 and 2.7. By the construction of jSπ this implies xg ∈ Int(R[ψ] , S[ψ] ) (or xg ∈ CInt(R[ψ] , S[ψ] ), resp.) for every g ∈ G, and thus the claim (2.10). Next, we consider the general case. Let F be a torsionfree direct summand of G containing Ker(ψ), let χ : G ։ G/F be the canonical projection and let λ : H ։ G/F be the induced epimorphism of groups, so that λ ◦ ψ = χ. Let R be a G-graded ring, and let S be a G-graded R-algebra. By 2.9 and the first paragraph, Int(R[χ] , S[χ] ) = Int(R, S)[χ] = (Int(R, S)[ψ] )[λ] ⊆ Int(R[ψ] , S[ψ] )[λ] ⊆ Int((R[ψ] )[λ] , (S[ψ] )[λ] ) = Int(R[χ] , S[χ] ), hence (Int(R, S)[ψ] )[λ] = Int(R[ψ] , S[ψ] )[λ] and therefore Int(R, S)[ψ] = Int(R[ψ] , S[ψ] ) (or the analogous statement for complete integral closures) as desired. b) Suppose K := Ker(ψ) is not torsionfree. By 2.7 and 2.8 there exist a noetherian ring R and an R-algebra S such that R is integrally closed in S (hence completely integrally closed in S) and that R[K][0] is not integrally closed in S[K][0] (hence not completely integrally closed in S[K][0] (2.2)). Then, R[K] is completely integrally closed in S[K] (2.6). Extending the K-graduations of R and S trivially to G-graduations it follows that R[K][G] is completely integrally closed in S[K][G] , while (R[K][G] )[ψ] is not integrally closed in (S[K][G] )[ψ] . This proves the claim.  (2.13) Corollary Let ψ : G ։ H be an epimorphism of groups. If G is torsionfree then ψ-coarsening commutes with relative (complete) integral closure.4 Proof. Immediately from 2.12.  (2.14) Supposing that the torsion subgroup T of G is a direct summand of G, it is readily checked that a subgroup F ⊆ G is contained in a torsionfree direct summand of G if and only if the composition of canonical morphisms T ֒→ G ։ G/F has a retraction. 4In case H = 0 the statement about relative integral closures is [3, V.1 Exercice 25]. 12 FRED ROHRER A torsionfree subgroup F ⊆ G is not necessarily contained in a torsionfree direct summand of G, not even if G is of finite type. Using the criterion above one checks that a counterexample is provided by G = Z ⊕ Z/nZ for n ∈ N>1 and F = h(n, 1)iZ ⊆ G. (2.15) Questions Let ψ : G ։ H be an epimorphism of groups. The above result gives rise to the following questions: a) If Ker(ψ) is torsionfree, does ψ-coarsening commute with (complete) integral closure? b) If ψ-coarsening commutes with (complete) integral closure, is then Ker(ψ) contained in a torsionfree direct summand of G? Note that, by 2.14, at most one of these questions has a positive answer. 3. Integral closures of entire graded rings In this section we consider (complete) integral closures of entire graded rings in their graded fields of fractions. We start with the relevant definitions and basic properties. (3.1) Let R be an entire G-graded ring. The (complete) integral closure of R in Q(R) is denoted by Int(R) (or CInt(R), resp.) and is called the (complete) integral closure of R. We say that R is (completely) integrally closed if it is (completely) integrally closed in Q(R). Keep in mind that Int(R) is integrally closed, but that CInt(R) is not necessarily completely integrally closed (2.2). If ψ : G ։ H is an epimorphism of groups and R is a G-graded ring such that R[ψ] (and hence R) is entire, then Q(R)[ψ] is entire and R[ψ] ⊆ Q(R)[ψ] ⊆ Q(Q(R)[ψ] ) = Q(R[ψ] ). From 2.9 it follows Int(R)[ψ] ⊆ Int(R[ψ] ) and CInt(R)[ψ] ⊆ CInt(R[ψ] ). Hence, if R[ψ] is (completely) integrally closed then so is R. We say that ψ-coarsening commutes with (complete) integral closure if whenever R is an entire G-graded ring then R[ψ] is entire and Int(R)[ψ] = Int(R[ψ] ) (or CInt(R)[ψ] = CInt(R[ψ] ), resp.). Clearly, if ψ-coarsening commutes with (complete) integral closure then Ker(ψ) is torsionfree (1.12 b)). Let F ⊆ G be a subgroup. An F -graded ring S is entire and (completely) integrally closed if and only if S (G) is so. (3.2) Let I be a nonempty right filtering preordered set, and let ((Ri )i∈I , (ϕij )i≤j ) be an inductive system in GrAnnG over I such that Ri is entire for every i ∈ I and that ϕij is a monomorphism for all i, j ∈ I with i ≤ j. By 2.5 and 1.7 we have inductive systems (Int(Ri ))i∈I and (CInt(Ri ))i∈I in GrAnnG over I, and we can consider the sub-limi∈I Ri −→ algebras lim Int(Ri ) ⊆ lim CInt(Ri ) ⊆ Q(lim Ri ) −→ −→ −→ i∈I i∈I i∈I and compare them with the sub-limi∈I Ri -algebras −→ Int(lim Ri ) ⊆ CInt(lim Ri ) ⊆ Q(lim Ri ). −→ −→ −→ i∈I i∈I i∈I GRADED INTEGRAL CLOSURES 13 It follows immediately from 2.5 and 1.7 that limi∈I Int(Ri ) = Int(limi∈I Ri ). Hence, if Ri −→ −→ is integrally closed for every i ∈ I then limi∈I Ri is integrally closed, and limi∈I CInt(Ri ) ⊆ −→ −→ CInt(limi∈I Ri ). −→ Suppose now in addition that (limi∈I Ri ) ∩ Q(Ri ) = Ri for every i ∈ I. Then, analo−→ gously to [1, 2.1] one sees that limi∈I CInt(Ri ) = CInt(limi∈I Ri ), hence if Ri is completely −→ −→ integrally closed for every i ∈ I then so is limi∈I Ri . This additional hypothesis is fulfilled −→ in case R is a G-graded ring, F is a group, I = FF (1.3), and RU equals R[U](G⊕F ) or R[U][G] for U ∈ FF . (3.3) Let R, S and T be G-graded rings such that R ⊆ S ⊆ T as graded subrings. Clearly, CInt(R, S) ⊆ CInt(R, T ) ∩ S. Gilmer and Heinzer found an (ungraded) example showing that this is not necessarily an equality ([7, Example 2]), not even if R, S and T are entire and have the same field of fractions. In [7, Proposition 2] they also presented the following criterion for this inclusion to be an equality, whose graded variant is proven analogously: If for every G-graded sub-S-module of finite type M ⊆ T with R ⊆ M there exists a G-graded S-module N containing M such that R is a direct summand of N, then CInt(R, S) = CInt(R, T ) ∩ S. In [7, Remark 2] Gilmer and Heinzer claim (again in the ungraded case) that this criterion applies if S is principal. As this would be helpful to us later (3.6, 3.10) we take the opportunity to point out that it is wrong. Namely, suppose that S is not simple, that T = Q(S), and that the hypothesis of the criterion is fulfilled. Let x ∈ S \ 0. There exists an S-module N containing hx−1 iS ⊆ T such that S is a direct summand of N. The tensor product with S/hxiS of the canonical injection S ֒→ N has a retraction, but it also factors over the zero morphism S/hxiS → hx−1 iS /S. This implies x ∈ S ∗ , yielding the contradiction that S is simple. Now we consider graded group algebras. We will show that both variants behave well with integral closures and that the finely graded variant behaves also well with complete integral closure. (3.4) Theorem a) Formation of finely graded group algebras over entire G-graded rings commutes with (complete) integral closure. b) If F is a group, then an entire G-graded ring R is (completely) integrally closed if and only if R[F ] is so. Proof. Keeping in mind that Q(R)[F ] = Q(R[F ]) (1.11) this follows immediately from 2.6.  (3.5) Lemma If R is a simple G-graded ring then R[Z][G] is entire and completely integrally closed. Proof. First we note that S := R[Z][G] is entire (1.11). The argument in [2, IV.1.6 Proposition 10] shows that S allows a graded version of euclidean division, i.e., for f, g ∈ S hom with f 6= 0 there exist unique u, v ∈ S hom with g = uf + v and degZ (v) < degZ (f ), where degZ denotes the usual Z-degree of polynomials over R. Using this we see analogously to 14 FRED ROHRER [2, IV.1.7 Proposition 11] that every G-graded ideal of S has a homogeneous generating set of cardinality 1. Next, developing a graded version of the theory of divisibility in entire rings along the line of [2, VI.1], it follows analogously to [2, VI.1.11 Proposition 9 (DIV); VII.1.2 Proposition 1] that for every x ∈ Q(S)hom there exist coprime a, b ∈ S hom with x = ab . So, the argument in [3, V.1.3 Proposition 10] shows that S is integrally closed. As it is noetherian the claim is proven (2.2).  (3.6) Theorem a) Formation of coarsely graded algebras of torsionfree groups over entire G-graded rings commutes with integral closure. b) If F is a torsionfree group, then an entire G-graded ring R is integrally closed if and only if R[F ][G] is so.5 Proof. It suffices to prove the first claim. We can without loss of generality suppose that F is of finite type, hence free of finite rank (1.3, 1.5, 3.2). By induction on the rank of F we can furthermore suppose F = Z. We have Int(R[Z][G] ) ∩ Q(R)[Z][G] = Int(R[Z][G] , Q(R)[Z][G] ). Since Q(R)[Z][G] is integrally closed (3.5) we get Int(R[Z][G] ) ⊆ Int(Q(R)[Z][G] , Q(R[Z][G] )) = Q(R)[Z][G] . It follows Int(R[Z][G] ) = Int(R[Z][G] ) ∩ Q(R)[Z][G] = Int(R[Z][G] , Q(R)[Z][G] ) = Int(R)[Z][G] (2.7) and thus the claim.  (3.7) Let F be a torsionfree group, let R be an entire G-graded ring, and suppose that CInt(R[Z][G] ) ∩ Q(R)[Z][G] = CInt(R[Z][G] , Q(R)[Z][G] ). Then, the same argument as in 3.6 (keeping in mind 3.2) yields CInt(R)[F ][G] = CInt(R[F ][G] ), hence R is completely integrally closed if and only if R[F ][G] is so. However, although R[Z][G] is principal by the proof of 3.5, we have seen in 3.3 that it is unclear whether CInt(R[Z][G] ) ∩ Q(R)[Z][G] and CInt(R[Z][G] , Q(R)[Z][G] ) are equal in general. (3.8) Corollary a) Formation of coarsely graded algebras of torsionfree groups over noetherian entire G-graded rings commutes with complete integral closure. b) If F is a torsionfree group, then a noetherian entire G-graded ring R is completely integrally closed if and only if R[F ][G] is so. Proof. We can without loss of generality suppose that F is of finite type (1.3, 1.5, 3.2). Then, R[F ][G] is noetherian (1.14), and the claim follows from 3.6 and 2.2.  In the rest of this section we study the behaviour of (complete) integral closures under arbitrary coarsening functors, also using the results from Section 2. (3.9) Proposition Let ψ : G ։ H be an epimorphism of groups. a) ψ-coarsening commutes with integral closure if and only if a G-graded ring R is entire and integrally closed if and only if R[ψ] is so. 5In case G = 0 the statement that integral closedness of R implies integral closedness of R[F ][G] is [3, V.1 Exercice 24]. GRADED INTEGRAL CLOSURES 15 b) If ψ-coarsening commutes with complete integral closure, then a G-graded ring R is entire and completely integrally closed if and only if R[ψ] is so. Proof. If ψ-coarsening commutes with (complete) integral closure then it is clear that a G-graded ring R is entire and (completely) integrally closed if and only if R[ψ] is so. Conversely, suppose that ψ-coarsening preserves the property of being entire and integrally closed. Let R be an entire G-graded ring. Since simple G-graded rings are entire and integrally closed, R[ψ] is entire (1.12 b)). As Int(R) is integrally closed (2.2) the same is true for Int(R)[ψ] , implying Int(R[ψ] ) = Int(R[ψ] , Q(R[ψ] )) ⊆ Int(Int(R)[ψ] , Q(R[ψ] )) = Int(Int(R)[ψ] ) = Int(R)[ψ] ⊆ Int(R[ψ] ) (3.1) and thus the claim.  The argument used in a) cannot be used to prove the converse of b), as CInt(R) is not necessarily completely integrally closed (3.1). (3.10) Proposition Let ψ : G ։ H be an epimorphism of groups. Suppose that ψcoarsening commutes with relative integral closure and maps simple G-graded rings to entire and integrally closed H-graded rings. Then, ψ-coarsening commutes with integral closure. Proof. If R is an entire G-graded ring, then R[ψ] is entire (2.12 b), 1.12 b)) and Q(R)[ψ] is integrally closed, and as Q(Q(R)[ψ] ) = Q(R[ψ] ) (3.1) it follows Int(R[ψ] ) = Int(R[ψ] , Q(R[ψ] )) ⊆ Int(Q(R)[ψ] , Q(R[ψ] )) = Int(Q(R)[ψ] ) = Q(R)[ψ] , hence Int(R[ψ] ) = Int(R[ψ] , Q(R)[ψ] ) = Int(R, Q(R))[ψ] = Int(R)[ψ] .  (3.11) We have seen in 3.3 that it is (in the notations of the proof of 3.10) not clear that CInt(R[ψ] ) ⊆ Q(R)[ψ] implies CInt(R[ψ] ) = CInt(R[ψ] , Q(R)[ψ] ). Therefore, the argument from that proof cannot be used to get an analogous result for complete integral closures. (3.12) Lemma Let F be a free direct summand of G, let H be a complement of F in G, let ψ : G ։ H be the canonical projection, let R be a simple G-graded ring, and suppose that ψ(degsupp(R)) ⊆ degsupp(R). Then, R[ψ] ∼ = R(H) [degsupp(R) ∩ F ][H] in GrAnnH . Proof. We set D := degsupp(R). As F is free the same is true for D ∩ F . Let E be a basis of D ∩ F . If e ∈ E then Re 6= 0, so that we can choose ye ∈ Re \ 0 ⊆ R∗ . For f ∈ D ∩ F P there exists a unique family (re )e∈E of finite support in Z with f = e∈E re e, and we set Q yf := e∈E yere ∈ Rf \ 0; in case f ∈ E we recover the element yf defined above. As (R(H) )[0] is a subring of R[0] there exists a unique morphism of (R(H) )[0] -algebras p : R(H) [D ∩ F ][0] → R[0] with p(ef ) = yf for f ∈ D ∩ F . If h ∈ H, then for f ∈ D ∩ F and x ∈ Rh we have p(xef ) = xyf ∈ Rh+f ⊆ (R[ψ] )h , so that p((R(H) [D ∩ F ][H] )h ) ⊆ (R[ψ] )h , and therefore we have a morphism p : R(H) [D ∩ F ][H] → R[ψ] in GrAnnH . 16 FRED ROHRER Let χ : G ։ F denote the canonical projection. For g ∈ G with χ(g) ∈ D there is a morphism of groups x eχ(g) , qg : Rg → R(H) [D ∩ F ], x 7→ yχ(g) and for g ∈ G with χ(g) ∈ / D we denote by qg the zero morphism of groups Rg → R(H) [D ∩ F ]. For h ∈ H the morphisms qg with g ∈ ψ −1 (h) induce a morphism of groups qh : (R[ψ] )h → R(H) [D ∩ F ]. So, we get a morphism of groups M qh : R[ψ] → R(H) [D ∩ F ]. q := h∈H Let g ∈ G and x ∈ Rg . If χ(g) ∈ / D then g ∈ / D, hence x = 0, and therefore x x x p(q(x)) = x. Otherwise, p(q(x)) = p( yχ(g) eχ(g) ) = yχ(g) p(eχ(g) ) = yχ(g) yχ(g) = x. This shows that q is a right inverse of p. If x ∈ R(H) then q(p(x)) = x, and if f ∈ D ∩ F then y q(p(ef )) = q(yf ) = yff ef = ef , hence q is a left inverse of p. Therefore, q is an inverse of p, and thus p is an isomorphism.  (3.13) Proposition Let ψ : G ։ H be an epimorphism of groups, let R be a simple G-graded ring, and suppose that one of the following conditions is fulfilled: i) G is torsionfree; ii) Ker(ψ) is contained in a torsionfree direct summand of G and R has full support. Then, R[ψ] is entire and completely integrally closed. Proof. First, we note that R[ψ] is entire (1.12 b)). In case i) it suffices to show that R[0] is integrally closed, so we can replace H with 0 and hence suppose Ker(ψ) = G. In case ii), by the same argument as in the proof of 2.12 (and keeping in mind 3.1) we can suppose without loss of generality that K := Ker(ψ) itself is a torsionfree direct summand of G and hence consider H as a complement of K in G. In both cases, as K = limL∈F L (1.3) −→ K we have G = K ⊕ H = limL∈F (L ⊕ H), hence R = limL∈F ((R(U ⊕H) )(G) ). Setting ψL := −→ K −→ K ψ ↾L⊕H : L ⊕ H ։ H we get R[ψ] = limL∈F ((R(L⊕H) )[ψL ] ) (1.2). Hence, if (R(L⊕H) )[ψ] −→ K is integrally closed for every L ∈ FK then R[ψ] is integrally closed (3.2). Therefore, as R(L⊕H) is simple for every L ∈ FK (1.6) we can suppose that K is of finite type, hence free. As R is simple it is clear that D := degsupp(R) ⊆ G is a subgroup, hence D∩K ⊆ K is a subgroup, and thus D ∩ K is free. In both cases, our hypotheses ensure ψ(D) ⊆ D, so that 3.12 implies R[ψ] ∼ = R(H) [D ∩ K][H] . As R is simple it is completely integrally closed, hence R(H) is completely integrally closed (3.1), thus R(H) [D ∩ K][H] is completely integrally closed (3.6), and so the claim is proven.  (3.14) Theorem Let ψ : G ։ H be an epimorphism of groups, let R be an entire G-graded ring, and suppose that one of the following conditions is fulfilled: i) G is torsionfree; ii) Ker(ψ) is contained in a torsionfree direct summand of G and hdegsupp(R)iZ = G. Then, Int(R)[ψ] = Int(R[ψ] ), and R is integrally closed if and only if R[ψ] is so.6 6In case i) and H = 0 this is [3, V.1 Exercice 25]. GRADED INTEGRAL CLOSURES 17 Proof. As degsupp(Q(R)) = hdegsupp(R)iZ this follows immediately from 3.10, 3.13 and 2.12.  (3.15) Questions Let R be an entire G-graded ring. The above, especially 3.7 and 3.11, gives rise to the following questions: a) Let ψ : G ։ H be an epimorphism of groups such that Ker(ψ) is torsionfree. Do we have CInt(R[ψ] ) ∩ Q(R)[ψ] = CInt(R[ψ] , Q(R)[ψ] )? b) Do we have CInt(R[Z][G] ) ∩ Q(R)[Z][G] = CInt(R[Z][G] , Q(R)[Z][G] )? If both these questions could be answered positively, then the same arguments as above would yield statements for complete integral closures analogous to 3.6, 3.10, and 3.14. Acknowledgement: I thank Benjamin Bechtold and the reviewer for their comments and suggestions. The remarks in 2.14 were suggested by Micha Kapovich and Will Sawin on http://mathoverflow.net/questions/108354. The counterexample in 3.3 is due to an anonymous user on http://mathoverflow.net/questions/110998. References [1] D. F. Anderson, Graded Krull domains. Comm. Algebra 7 (1979), 79–106. [2] N. Bourbaki, Éléments de mathématique. Algèbre. Chapitres 1 à 3. Masson, 1970; Chapitres 4 à 7. Masson, 1981. [3] N. Bourbaki, Éléments de mathématique. Algèbre commutative. Chapitres 5 à 7. Hermann, 1975. [4] S. Goto, K. Yamagishi, Finite generation of noetherian graded rings. Proc. Amer. Math. Soc. 89 (1983), 41–44. [5] R. Gilmer, Multiplicative ideal theory. Pure Appl. Math. 12, Marcel Dekker, 1972. [6] R. Gilmer, Commutative semigroup rings. Chicago Lectures in Math., Univ. Chicago Press, 1984. [7] R. Gilmer, W. J. Heinzer, On the complete integral closure of an integral domain. J. Aust. Math. Soc. 6 (1966), 351–361. [8] A. Grothendieck, J. A. Dieudonné, Éléments de géométrie algébrique. I: Le langage des schémas (Seconde édition). Grundlehren Math. Wiss. 166, Springer, 1971. [9] C. Nǎstǎsescu, F. Van Oystaeyen, Graded rings with finiteness conditions II. Comm. Algebra 13 (1985), 605–618. [10] F. Rohrer, Coarsenings, injectives and Hom functors. Rev. Roumaine Math. Pures Appl. 57 (2012), 275–287. [11] I. Swanson, C. Huneke, Integral closure of ideals, rings, and modules. London Math. Soc. Lecture Note Ser. 336, Cambridge Univ. Press, 2006. [12] B. L. van der Waerden, Algebra. Zweiter Teil. (Fünfte Auflage). Heidelb. Taschenb. 23, Springer, 1967. [13] J. Van Geel, F. Van Oystaeyen, About graded fields. Indag. Math. (N.S.) 84 (1981), 273–286. Universität Tübingen, Fachbereich Mathematik, Auf der Morgenstelle 10, 72076 Tübingen, Germany E-mail address: rohrer@mail.mathematik.uni-tuebingen.de 

Hygienic Source-Code Generation Using Functors Karl Crary arXiv:1801.01579v1 [cs.PL] 4 Jan 2018 Carnegie Mellon University Abstract Existing source-code-generating tools such as Lex and Yacc suffer from practical inconveniences because they use disembodied code to implement actions. To prevent this problem, such tools could generate closed functors that are then instantiated by the programmer with appropriate action code. This results in all code being type checked in its appropriate context, and it assists the type checker in localizing errors correctly. We have implemented a lexer generator and parser generator based on this technique for Standard ML, OCaml, and Haskell. 1 Introduction Compiler implementers have a love-hate relationship with source-code-generating tools such as Lex [9] (which generates lexers from regular expressions) and Yacc [7] (which generates shift-reduce parsers from context-free grammars). These tools automate the some of the most tedious parts of implementing a parser, but they can be awkward to use. One of the main awkward aspects of such tools is the disembodied code problem. To build a lexer or a parser, these tools cobble together snippets of code (each implementing an action of the lexer/parser) supplied by the programmer in a lexer/parser specification file. Unfortunately, the code snippets, as they appear in the specification file, are divorced from their ultimate context. The tools manipulate them as simple strings.1 This makes programming awkward in several ways. Functions and other values are passed into the snippets using identifiers that are bound nowhere in the programmer’s code, nor even introduced by a pseudo-binding such as open. Rather, the snippet is copied into a context in which such identifiers are in scope. This can make code difficult to read. More importantly, disembodied code makes debugging challenging, because the code seen by the compiler bears little resemblance to the code written by the programmer. For example, consider the following line from an ML-Lex [1] specification: {whitespace}+ => ( lex () ); This line tells the lexer to skip any whitespace it encounters by matching it and then calling itself recursively to continue. 1 Such strings may even include syntax errors, which are duly copied into the output code. Typically the tool does not even ensure that delimiters are matched. (Note that lex is an example of an identifier introduced implicitly when the snippet is copied.) ML-ULex2 [13] converts the line into the Standard ML code: fun yyAction0 (strm, lastMatch : yymatch) = (yystrm := strm; ( lex () )) This output code already is not very easy to read. However, the problem is greatly exacerbated by the familiar phenomenon in typed functional languages that type checkers are often bad at identifying the true source of a type error. Suppose we introduce an error into the specification by omitting the argument to lex: {whitespace}+ => ( lex ); We now obtain3 several pages of error messages looking like: foo.lex.sml:1526.25-1526.70 Error: operator and operand don’t agree [tycon mismatch] operator domain: yyInput.stream * action * yymatch operand: yyInput.stream * (yyInput.stream * yymatch -> unit -> (?.svalue,int) ?.token) * yymatch in expression: yyMATCH (strm,yyAction0,yyNO_MATCH) or like: foo.lex.sml:1686.20-1692.47 Error: types of if branches do not agree [tycon mismatch] then branch: (?.svalue,int) ?.token else branch: unit -> (?.svalue,int) ?.token in expression: if inp = #"\n" then yyQ38 (strm’,lastMatch) else if inp < #"\n" then if inp = #"\t" then yyQ37 (<exp>,<exp>) else yyQ36 (<exp>,<exp>) else yyQ37 (strm’,lastMatch) 2 The lexer generator (compatible with ML-Lex) that Standard ML of New Jersey uses. 3 Using Standard ML of New Jersey v100.68. and none of the errors is anywhere near the copied snippet containing the error. functor is already known, so an error in one action will not be misinterpreted as an error in all the other actions. The problem is related to the issue of variable hygiene in macro expansion [8]. In both cases, the programmer writes code (a lexer/parser action, or macro argument) divorced from its ultimate context and then—after processing—that code is dropped verbatim into its ultimate context. In the setting of macros, this sets the scene for variable capture to occur, which is nearly always erroneous. In lexer generators, variable capture often is actually desired (consider the lex identifier), but, as observed above, it is nevertheless difficult to reason about and to debug. Accordingly, we are interested in source-code generation in which all code is type-checked in the same context in which it is written. We call this hygienic source-code generation by analogy to hygienic macro expansion, which ensures the same thing for macros. An obvious way to accomplish hygienic source-code generation is to have the tool type-check every snippet before it assembles them into output code. But, this approach is unattractive in practice, because it necessitates including all the apparatus of parsing, elaboration, and type-checking as part of a tool that does not otherwise need all that apparatus. We propose a simpler and cleaner alternative: Rather than type-check disembodied code in context, we dispense with disembodied code altogether. To accomplish this, the tool—rather than assembling snippets of source code into a program—generates a functor that abstracts over the code that used to reside in snippets. The programmer then applies the functor in order to instantiate the lexer or parser with specific action implementations. A third alternative, arguably more principled than ours, is to implement the lexer/parser generator in a type-safe metaprogramming language such as MetaML [12] or its cousins. With such an approach, as in ours, the action implementations would be type-checked in context, without any need to duplicate compiler apparatus. Furthermore, it would remove the need to write the lexer/parser specification and action implementations in two separate places, as our proposal requires. On the other hand, this alternative requires one to use a special programming language. We want an approach compatible with pre-existing, conventional functional programming languages, specifically ML and Haskell. Finally, in some problem domains one may consider avoiding generated source code entirely. For example, in parsing, some programmers find parser combinators [5, 6] to be a suitable or even preferable alternative to Yacc-like tools. Nevertheless, many programmers prefer traditional LR parser generators for various reasons including error reporting and recovery, and ambiguity diagnostics. In this work we take it as given that source-code generation is preferred, for whatever reason. Employing this design, we have implemented a lexer generator, called CM-Lex, and a parser generator, called CM-Yacc. Each tool supports Standard ML, OCaml, and Haskell.4 Both tools are available on-line at: www.cs.cmu.edu/~crary/cmtool/ In the remainder of the paper we describe how the tools work, taking the lexer generator as our primary example. 2 Lexing Functor Generation The following is a very simple CM-Lex specification: sml name LexerFun alphabet 128 function f : t = (seq ’a ’a) => aa (seq ’a (* ’b) ’c) => abc The specification’s first line indicates that CM-Lex should generate Standard ML code. The next two lines indicate that CM-Lex should produce a functor named LexerFun, and that it should generate a 7-bit parser (any symbols outside the range 0 . . . 127 will be rejected automatically). The remainder gives the specification of a lexing function named f. The function will return a value of type t, and it is defined by two regular expressions. Regular expressions are given as S-expressions using the Scheme Shell’s SRE notation5 [11]. Thus, the first arm activates an action named aa when the regular expression aa is recognized. The second activates an action named abc when the regular expression ab∗ c is recognized. Observe that the specification contains no disembodied code. The actions are simply given names, which are instantiated when the resulting functor is applied. From this specification, CM-Lex generates the following Standard ML code:6 functor LexerFun (structure Arg : sig type t type info = { match : char list, follow : char stream } val aa : info -> t val abc : info -> t end) At first blush, our proposal might seem to replace one sort of disembodied code with another. This is true in a sense, but there is a key difference. The code in which the functor is applied is ordinary code, submitted to an ordinary compiler. That compiler then type checks the action code (that formerly resided in snippets) in the context in which it now appears, which is the functor’s argument. As a practical matter, each action becomes a distinct field of the functor argument, and consequently each action is type-checked independently, as desired. The type of the :> sig val f : char stream -> Arg.t end = . . . implementation . . . 4 The tool is implemented in Standard ML. Although SREs are less compact than some other notations, we find their syntax is much easier to remember. 6 We simplify here and in the following examples for the sake of exposition. 5 2 fun aa ({follow, ...}:info) = ( print "matched aa\n" ) signature STREAM = sig type ’a stream datatype ’a front = Nil | Cons of ’a * ’a stream The type checker is able to identify the source of the error precisely, finding that aa has the type unit instead of t: example.sml:8.4-29.12 Error: value type in structure doesn’t match signature spec name: aa spec: ?.Arg.info -> ?.Arg.t actual: ?.Arg.info -> unit val front : ’a stream -> ’a front val lazy : (unit -> ’a front) -> ’a stream end Figure 1: Lazy Streams 2.1 An expanded specification We may add a second function to the lexer by simply adding another function specification: When the programmer calls the functor, he provides the type t and the actions aa and abc, both of which produce a t from a record of matching information. The functor then returns a lexing function f, which produces a t from a stream of characters. Although the programmer-supplied actions can have side effects, the lexer itself is purely functional. The input is processed using lazy streams (the signature for which appears in Figure 1). Each action is given the portion of the stream that follows the matched string as part of the matching information. function g : u = (or (seq ’b ’c) (seq ’b ’d)) => bcbd epsilon => error In the parlance of existing lexer generators, multiple functions are typically referred to as multiple start conditions or start states, but we find it easier to think about them as distinct functions that might or might not share some actions. In this case, the function g is specified to return a value of type u. Since u might not be the same type as t, g cannot share any actions with f. The first arm activates an action named bcbd when the regular expression bc + bd is recognized. The second arm activates an action named error when the empty string is recognized. Like other lexer generators, CM-Lex prefers the longest possible match, so an epsilon arm will only be used when the input string fails to match any other arm. Thus, the latter arm serves as an error handler.7 From the expanded specification, CM-Lex generates the functor: As an illustration of how the functor might be applied, the following program processes an input stream, printing a message each time it recognizes a string: structure Lexer = LexerFun (structure Arg = struct type t = char stream type info = { match : char list, follow : char stream } functor LexerFun (structure Arg : sig type t type u fun aa ({follow, ...}:info) = ( print "matched aa\n"; follow ) fun abc ({follow, ...}:info) = ( print "matched ab*c\n"; follow ) end) type info = { match : char list, follow : char stream } val val val val end) fun loop (strm : char stream) = (case front strm of Nil => () | Cons _ => loop (Lexer.f strm)) The function Lexer.f matches its argument against the two regular expressions and calls the indicated action, each of which prints a message and returns the remainder of the stream. Observe that the implementations of the actions (the fields aa and abc of the argument structure) are ordinary ML code. As one consequence, the action code faces the standard type checker. Moreover, each action’s required type is unambiguously given by LexerFun’s signature and the type argument t, so error identification is much more accurate. For example, suppose we replace the aa action with an erroneous implementation that fails to return the remainder of the stream: aa : info -> t abc : info -> t bcbd : info -> u error : info -> u :> sig val f : char stream -> Arg.t val g : char stream -> Arg.u end = . . . implementation . . . 3 Recursion in actions One important functionality for a lexer generator is the ability for actions to invoke the lexer recursively. For example, 7 In contrast, the specification for f was inexhaustive, so CM-Lex added a default error handler that raises an exception. 3 it is common for a lexer, upon encountering whitespace, to skip the whitespace and call itself recursively (as in the example in Section 1).8 This can be problematic because it requires recursion between the lexer functor’s argument and its result. For example, consider a lexer that turns a stream of characters into a stream of words. The CM-Lex specification is: structure rec Arg = struct type t = string stream type info = { match : char list, follow : char stream } fun whitespace ({follow, ...}:info) = Words.f follow sml name WordsFun alphabet 128 fun word ({match, follow, ...}:info) = lazy (fn () => Cons (String.implode match, Words.f follow)) end) set whitechar = (or 32 9 10) /* space, tab, lf */ set letter = (range ’a ’z) function f : t = (+ whitechar) => whitespace (+ letter) => word and Words = WordsFun (structure Arg = Arg) Unfortunately, recursive modules bring about a variety of thorny technical issues [2, 10, 4]. Although some dialects of ML support recursive modules, Standard ML does not. As a workaround, CM-Lex provides recursive access to the lexer via a self field passed to each action. The info type is extended with a field self : self, where the type self is a record containing all of the lexing functions being defined. In this case: CM-Lex generates the functor: functor WordsFun (structure Arg : sig type t type info = { match : char list, follow : char stream } type self = { f : char stream -> t } Using the self-augmented functor, we can implement the lexer as follows: val whitespace : info -> t val word : info -> t end) structure Words = WordsFun (structure Arg = struct type t = string stream :> sig val f : char stream -> Arg.t end = . . . implementation . . . type self = { f : char stream -> t } type info = { match : char list, follow : char stream, self : self } A natural way9 to implement the desired lexer would be with a recursive module definition: 8 One way to accomplish this would be to structure the lexer with a driver loop (such as the function loop in Section 2), and for the action to signal the driver loop to discard the action’s result and recurse. However, the earlier example notwithstanding, this is usually not the preferred way to structure a lexer. 9 This simple implementation does not result in the best behavior from the lazy streams, because forcing the output stream causes the lexer to examine more of the input stream than is necessary to determine the output stream’s first element. We illustrate a better way to manage laziness in Appendix A. In any case, laziness is orthogonal to the issue being discussed here. fun whitespace ({match, follow, self, ...}:info) = #f self follow fun word ({match, follow, self, ...}:info) = lazy (fn () => Cons (String.implode match, #f self follow)) end) 4 Parsing Functor Generation The parser generator, CM-Yacc, works in a similar fashion to CM-Lex. A CM-Yacc specification for a simple arithmetic parser is: 4 at which a syntax error is detected and returns an exception for the parser to raise. For example: sml name ArithParseFun terminal terminal terminal terminal terminal datatype terminal = NUMBER of t | PLUS | TIMES | LPAREN | RPAREN NUMBER of t PLUS TIMES LPAREN RPAREN nonterminal Term : t = 1:NUMBER => number_term 1:Term PLUS 2:Term => plus_term 1:Term TIMES 2:Term => times_term LPAREN 1:Term RPAREN => paren_term structure Parser = ArithParseFun (structure Arg = struct type t = int start Term fun fun fun fun The specification says that the functor should be named ArithParseFun, and it declares five terminals, one of which carries a value of type t. The specification then declares one nonterminal called Term, indicates that a term carries a value of type t, and gives four productions that produce terms.10 Numbers are attached to the symbols on the left-hand-side of a production that carry values that should be passed to the production’s action. The number itself indicates the order in which values should be passed. Thus plus term is passed a pair containing the first and third symbols’ values. The final line specifies that the start symbol is Term. From this specification, CM-Yacc generates the following Standard ML code: datatype terminal = datatype terminal fun error _ = Fail "syntax error" end) Then our parser is Parser.parse : terminal -> int. Note that we use datatype copying (a little-known feature of Standard ML) to copy the terminal datatype into the Arg structure. If the datatype were defined within the Arg structure, there would be no way to use it outside. OCaml does not support datatype copying, but one can get the same effect by including a module that contains the datatype. functor ArithParseFun (structure Arg : sig type t val val val val number_term x = x plus_term (x, y) = x + y times_term (x, y) = x * y paren_term x = x 5 Functors in Haskell In broad strokes the Haskell versions of CM-Lex and CMYacc are similar to the ML versions. In one regard, they are simpler: In Haskell all definitions are mutually recursive, so no special functionality is required to allow lexer actions to reinvoke the lexer. However, Haskell does not support functors, the central mechanism we exploit here. Instead, we built an ersatz functor from polymorphic functions. Recall the CM-Lex specification from Section 2.1, reprised in Figure 2. From that specification, CM-Lex generates a module (in the Haskell sense) named LexerFun with the following exports: number_term : t -> t plus_term : t * t -> t times_term : t * t -> t paren_term : t -> t datatype terminal = NUMBER of t | PLUS | TIMES | LPAREN | RPAREN val error : terminal stream -> exn end) data LexInfo = LexInfo { match :: [Char], follow :: [Char] } :> sig val parse : Arg.terminal stream -> Arg.t end = . . . implementation . . . data Arg t u = Arg { t :: Proxy t, u :: Proxy u, As before, the programmer supplies the type t and the actions. (The actions need not be passed a self argument, because parser actions do not commonly need to reinvoke the parser.) He also supplies the terminal datatype and an error action, the latter of which takes the terminal stream aa :: LexInfo -> t, abc :: LexInfo -> t, bcbd :: LexInfo -> u, error :: LexInfo -> u } 10 This grammar is ambiguous, resulting in shift-reduce conflicts. The ambiguity can be resolved in either of the standard manners: by specifying operator precedences, or by refactoring the grammar. f :: Arg t u -> [Char] -> t g :: Arg t u -> [Char] -> u 5 Compare this with the ML version, also reprised in Figure 2. The type Arg represents the argument to the functor. It contains implementations for the four actions aa, abc, bcbc, and error. It also contains implementations for the two types t and u. Haskell does not support type fields like an ML structure, but we can get a similar effect by including proxy fields with the types Proxy t and Proxy u. The programmer then fills them in with the term Proxy :: Proxy T for some T.11 Proxy [3] is a type constructor in the Haskell standard library that is designed for this sort of use. For any type constructor C, the type Proxy C has a single data constructor Proxy. The Proxy type constructor is poly-kinded, so C need not have kind *. An alternative would be to leave out the type fields altogether and allow type inference to fill them automatically. We believe it would be a misstep to do so. The type implementations are critical documentation that should be given explicitly in the program. Moreover, leaving out the type implementations would bring back the possibility that the type checker would misattribute the source of a type error. The functor’s output is factored into two separate polymorphic functions that each take the functor argument as an argument. Since the type arguments t and u are propagated to the result types of the lexing functions, they must also appear as explicit parameters of the type Arg. sml name LexerFun alphabet 128 function f : t = (seq ’a ’a) => aa (seq ’a (* ’b) ’c) => abc function g : u = (or (seq ’b ’c) (seq ’b ’d)) => bcbd epsilon => error . . . became . . . The Haskell version of CM-Yacc builds an ersatz functor in a similar fashion. However, while the ML version specified the terminal type as an input to the parser functor, there is no way to specify a datatype as an input to an ersatz functor. Instead, the parsing module defines the terminal datatype and passes it out. In the example above, CM-Lex was used in purely functional mode. Consequently, the input stream was simply a character list, since Haskell lists are lazy already. Alternatively, CM-Lex and CM-Yacc can be directed to generate monadic code, which allows the lexer or parser to deal with side effects, either in the generation of the input stream (e.g., input read from a file) or in the actions. Doing so incurs some complications — it is important that the input stream be memoizing and not every monad is capable of supporting the proper sort of memoization12 — but these complications are orthogonal to the functor mechanism discussed here and are beyond the scope of this paper. functor LexerFun (structure Arg : sig type t type u type info = { match : char list, follow : char stream } val val val val end) aa : info -> t abc : info -> t bcbd : info -> u error : info -> u :> sig val f : char stream -> Arg.t val g : char stream -> Arg.u end = . . . implementation . . . 6 Conclusion We argue that functor generation is a cleaner mechanism for source-code-generating tools than assembling snippets of disembodied code. The resulting functor makes no demands on the surrounding code (other than a few standard libraries), and so it is guaranteed to type check.13 The programmer never need look at the generated code. Figure 2: The example from Section 2.1 11 Alternatively, one could give the proxy fields the bare types t and u and fill them in with undefined :: T, but that approach would be more awkward in the monadic case in which we also need to specify a monad. A monad has kind * -> * and therefore does not have elements. 12 Monads such as IO and ST that support references also support memoization, and Identity supports it trivially (since no memoization need ever be done), but most others do not. 13 More precisely, it is guaranteed to type check in an initial context containing standard libraries and other module definitions. Unfortunately, Standard ML does not quite enjoy the weakening property, so the resulting functor is not guaranteed to type check in any context. Pollution of the namespace with datatype constructors and/or 6 and the parser specification is: In contrast, with a snippet-assembling tool, an error in any snippet will — even in the best case — require the programmer to look at generated code containing the snippet. More commonly, the programmer will need to look at lots of generated code having nothing to do with the erroneous snippet. We have demonstrated the technique for lexer and parser generation, but there do not seem to be any limitations that would preclude its use for any other application of sourcecode generation. A sml name CalcParseFun terminal terminal terminal terminal terminal nonterminal Atom : t = 1:NUMBER => number_atom LPAREN 1:Term RPAREN => paren_atom A Full Example As a more realistic example, we implement a calculator that processes an input stream and returns its value. For simplicity, the calculator stops at the first illegal character (which might be the end of the stream). The lexer specification is: nonterminal Factor : t = 1:Atom => atom_factor 1:Atom TIMES 2:Factor => times_factor sml name CalcLexFun alphabet 128 nonterminal Term : t = 1:Factor => factor_term 1:Factor PLUS 2:Term => plus_term set digit = (range ’0 ’9) set whitechar = (or 32 9 10) /* space, tab, lf */ start Term which generates: function lex : t = (+ digit) => number ’+ => plus ’* => times ’( => lparen ’) => rparen (+ whitechar) => whitespace functor CalcParseFun (structure Arg : sig type t val val val val val val /* Stop at the first illegal character */ epsilon => eof which generates: functor CalcLexFun (structure Arg : sig type t number_atom : t -> t paren_atom : t -> t atom_factor : t -> t times_factor : t * t -> t factor_term : t -> t plus_term : t * t -> t datatype terminal = NUMBER of t | PLUS | TIMES | LPAREN | RPAREN type self = { lex : char stream -> t } type info = { match : char list, follow : char stream, self : self } val val val val val val val end) NUMBER of t PLUS TIMES LPAREN RPAREN val error : terminal stream -> exn end) :> sig val parse : Arg.terminal stream -> Arg.t end = . . . implementation . . . number : info -> t plus : info -> t times : info -> t lparen : info -> t rparen : info -> t whitespace : info -> t eof : info -> t We then assemble the calculator as follows: structure Calculator :> sig val calc : char stream -> int end = struct datatype terminal = NUMBER of int | PLUS | TIMES | LPAREN :> sig val lex : char stream -> Arg.t end = . . . implementation . . . infix declarations for identifiers that are used within the generated functor will prevent it from parsing correctly. This is one reason why it is considered good practice in SML for all code to reside within modules. 7 References | RPAREN [1] Andrew W. Appel, James S. Mattson, and David R. Tarditi. A lexical analyzer generator for Standard ML, October 1994. Available at www.smlnj.org/doc/ML-Lex/manual.html. structure Lexer = CalcLexFun (structure Arg = struct type t = terminal front [2] Karl Crary, Robert Harper, and Sidd Puri. What is a recursive module? In 1999 SIGPLAN Conference on Programming Language Design and Implementation, pages 50–63, Atlanta, May 1999. type self = { lex : char stream -> t } type info = { match : char list, follow : char stream, self : self } [3] Data.Proxy. Online documentation at https://hackage.haskell.org/package/base-4.7.0.0/docs/Data2014. Retrieved October 21, 2017. fun number ({ match, follow, self }:info) = Cons (NUMBER (Option.valOf (Int.fromString (String.implode match))), lazy (fn () => #lex self follow)) [4] Derek Dreyer. Understanding and Evolving the ML Module System. PhD thesis, Carnegie Mellon University, School of Computer Science, Pittsburgh, Pennsylvania, May 2005. [5] Richard Frost and John Launchbury. Constructing natural language interpreters in a lazy functional language. The Computer Journal, 32(2), 1989. fun simple terminal ({ follow, self, ... }:info) = Cons (terminal, lazy (fn () => #lex self follow)) val val val val [6] Graham Hutton. Higher-order functions for parsing. Journal of Functional Programming, 2(3), 1992. plus = simple PLUS times = simple TIMES lparen = simple LPAREN rparen = simple RPAREN [7] Stephen C. Johnson. Yacc — yet another compiler compiler. Technical Report 32, Bell Laboratories Computing Science, Murray Hill, New Jersey, July 1975. [8] Eugene Kohlbecker, Daniel P. Friedman, Matthias Felleisen, and Bruce Duba. Hygienic macro expansion. In 1986 ACM Conference on Lisp and Functional Programming, pages 161–161, 1986. fun whitespace ({ follow, self, ... }:info) = #lex self follow [9] M. E. Lesk. Lex — a lexical analyzer generator. Technical Report 39, Bell Laboratories Computing Science, Murray Hill, New Jersey, October 1975. fun eof _ = Nil end) structure Parser = CalcParseFun (structure Arg = struct type t = int [10] Claudio V. Russo. Types for Modules. Number 60 in Electronic Notes in Theoretical Computer Science. Elsevier, January 2003. [11] Olin Shivers. The SRE regular-expression notation, August 1998. Post to the comp.lang.scheme newsgroup, now archived at http://www.scsh.net/docu/post/sre.html. fun id x = x val val val fun val fun number_atom = id paren_atom = id atom_factor = id times_factor (x, y) = x * y factor_term = id plus_term (x, y) = x + y [12] Walid Taha and Tim Sheard. MetaML and multi-stage programming with explicit annotations. Theoretical Computer Science, 248(1–2), 2000. [13] Aaron Turon and John Reppy. SML/NJ Language Processing Tools: User Guide, September 2015. Available at www.smlnj.org/doc/ml-lpt/manual.pdf. datatype terminal = datatype terminal fun error _ = Fail "syntax error" end) fun calc strm = Parser.parse (lazy (fn () => Lexer.lex strm)) end 8 

Words and characters in finite p-groups by arXiv:1406.5395v3 [math.GR] 14 Jun 2016 Ainhoa Iñiguez Mathematical Institute University of Oxford Andrew Wiles Building Woodstock Road OX2 6GG, Oxford UNITED KINGDOM E-mail: ainhoa.iniguez@maths.ox.ac.uk and Josu Sangroniz1 Departamento de Matemáticas Facultad de Ciencia y Tecnologı́a Universidad del Paı́s Vasco UPV/EHU 48080 Bilbao SPAIN E-mail: josu.sangroniz@ehu.es Abstract Given a group word w in k variables, a finite group G and g ∈ G, we consider the number Nw,G (g) of k-tuples g1 , . . . , gk of elements of G such that w(g1 , . . . , gk ) = g. In this work we study the functions Nw,G for the class of nilpotent groups of nilpotency class 2. We show that, for the groups in this class, Nw,G (1) ≥ |G|k−1, an inequality that can be improved to Nw,G (1) ≥ |G|k /|Gw | (Gw is the set of values taken by w on G) if G has odd order. This last result is explained by the fact that the functions Nw,G are characters of G in this case. For groups of even order, all that can be said is that Nw,G is a generalized character, something that is false in general for groups of nilpotency class greater than 2. We characterize group theoretically when Nxn ,G is a character if G is a 2group of nilpotency class 2. Finally we also address the (much harder) problem of studying if Nw,G (g) ≥ |G|k−1 for g ∈ Gw , proving that this is the case for the free p-groups of nilpotency class 2 and exponent p. Keywords: p-groups; words; characters MSC: 20D15, 20F10 Both authors are supported by the MINECO (grants MTM2011-28229-C02-01 and MTM2014-53810-C22-P). The second author is also supported by the Basque Government (grants IT753-13 and IT974-16). 1 Corresponding author 1 Introduction Given a group word w in k variables x1 , . . . , xk , that is an element in the free group Fk on x1 , . . . , xk , we can define for any k elements g1 , . . . , gk in a group G the element w(g1, . . . , gk ) ∈ G by applying to w the group homomorphism from Fk to G sending xi to gi . For G a finite group and g ∈ G we consider the number Nw,G (g) = |{(g1 , . . . , gk ) ∈ G(k) | w(g1, . . . , gk ) = g}|. (1) (G(k) denotes the k-fold cartesian product of G with itself.) So Nw,G (g) can be thought of as the number of solutions of the equation w = g. The set of word values of w on G, i. e., the set of elements g ∈ G such that the equation w = g has a solution in G(k) , is denoted Gw . There is an extensive literature on the functions Nw,G , sometimes expressed in terms of the probabilistic distribution Pw,G = Nw,G /|G|k . For example Nikolov and Segal gave in [9] a characterization of the finite nilpotent (and also solvable) groups based on these probabilities: G is nilpotent if and only if inf w,g Pw,G (g) > p−|G| , where w and g range over all words and Gw , respectively, and p is the largest prime dividing |G|. One of the implications is easy: if G is not nilpotent the infimum is in fact zero. Indeed, we can consider the k-th left-normed lower central word γk = [[[x1 , x2 ], x3 ], ..., xk ]. Since G is not nilpotent, there exists some non-trivial element g ∈ Gγk (for any k). Since γk (g1 , . . . , gk ) = 1 if some gi = 1, we have that Pγk ,G (g) ≤ (|G| − 1)k /|G|k , which can be made arbitrarily small. On the other hand the estimation Pw,G (g) > p−|G| for g ∈ Gw seems to be far from sharp and Amit in an unpublished work has conjectured that if G is nilpotent, Pw,G (1) ≥ 1/|G|. We prefer to give our results in terms of the functions Nw,G . In this paper we focus our attention on finite nilpotent groups of nilpotency class 2, which we take to be p-groups right from the outset, so all the results quoted here are referred to this family of groups. In Section 2 we consider a natural equivalence relation for words that enable us to assume that they have a special simple form. Then it is not difficult to prove Amit’s conjecture Nw,G (1) ≥ |G|k−1 for w ∈ Fk . This result has been proved independently by Levy in [6] using a similar procedure, although our approach to the concept of word equivalence is different. In the next two sections we show that the functions Nw,G are generalized characters, a result that is false for nilpotent groups of nilpotency class greater than 2, and even more, if G has odd order, they are genuine characters. In particular we obtain an improvement of Amit’s conjectured bound, namely, Nw,G (1) ≥ |G|k /|Gw |. For 2-groups, there are easy examples where Nx2 ,G fails to be a character and we actually characterize group-theoretically when this happens for the power words w = xn (always within the class of nilpotent 2-groups of nilpotency class 2). In the last section we consider briefly the conjecture Nw,G (g) ≥ |G|k−1 for w ∈ Fk and g ∈ Gw . This problem is much harder than the case g = 1 and only some partial results have been obtained, for instance confirming the conjecture if G is a free nilpotent p-group of nilpotency class 2 and exponent p. 2 2 Words in p-groups of nilpotency class 2 Since we are going to work with p-groups of nilpotency class 2, it is more convenient for us to define a word in the variables x1 , . . . , xk as an element in the free pro-p group of nilpotency class 2 on the variables x1 , . . . , xk , Fk . Thus, if w ∈ Fk is a word, it can be represented in a unique way as Y w = xz11 . . . xzkk [xi , xj ]zij , 1≤i<j≤k where the exponents zl , zij are p-adic integers. Of course, if G is a finite p-group (or pro-p group) of nilpotency class 2 and g1 , . . . , gk ∈ G, it makes sense to evaluate w on g1 , . . . , gk by applying the homomorphism π : Fk −→ G given by xi 7→ gi . As in the introduction, we denote this element w(g1 , . . . , gk ) and define the function Nw = Nw,G by (1). If σ is an automorphism of Fk , σ is determined by the images of the generators x1 , . . . , xk , which we denote w1 , . . . , wk . Then the image of w ∈ Fk is the word w(w1 , . . . , wk ), the result of evaluating w on w1 , . . . , wk . Since σ is an automorphism, there exist w1′ , . . . , wk′ ∈ Fk such that wi′ (w1 , . . . , wk ) = xi , 1 ≤ i ≤ k, and the inverse automorphism is given by xi 7→ wi′ . If G is a finite p-group (or pro-p group) of nilpotency class 2, we can define the map ϕ : G(k) −→ G(k) by ϕ(g1 , . . . , gk ) = (w1 (g1 , . . . , gk ), . . . , wk (g1 , . . . , gk )) and it is clear that this map is a bijection with the inverse map given by (g1 , . . . , gk ) 7→ (w1′ (g1 , . . . , gk ), . . . , wk′ (g1 , . . . , gk )). If w ′ = w(w1 , . . . , wk ), it is clear that w ′ (g1 , . . . , gk ) = g if and only if w(ϕ(g1, . . . , gk )) = g, thus ϕ is a bijection between the solutions of w ′ = g and w = g and in partic ular, Nw,G = Nw′ ,G . Definition 2.1. We will say that two words w, w ′ ∈ Fk are equivalent if they belong to the same orbit under the action of the automorphism group of Fk . Therefore we have proved the following result. Proposition 2.1. If w, w ′ ∈ Fk are two equivalent words, Nw,G = Nw′ ,G for any finite p-group G of nilpotency class 2. Next we want to find a set of representatives of the equivalence classes of words. There are natural homomorphisms Aut(Fk ) ։ Aut(Fk /Fk′ ) ∼ = GLk (Zp ) → Aut(Fk′ ), (2) where the composite map is the restriction. For the middle isomorphism, given an automorphism, we write the coordinates of the images of the vectors in a basis of the (k) Zp -module Fk /Fk′ ∼ = Zp as rows of the corresponding matrix. The last homomorphism comes from the identification of Fk′ with the exterior square of Fk /Fk′ . More explicitly, ′ we identify additive group of the k × k antisymmetric matrices over Zp , Ak , Q Fk with the zij via w = i<j [xi , xj ] 7→ A, where A ∈ Ak has entries zij for 1 ≤ i < j ≤ k. Then, if X ∈ GLk (Zp ), the action of X on Ak is given by A 7→ X t AX. This action is better (k) understood if we interpret A as an alternating bilinear form on the free Zp -module Zp . Notice however that, under a change of basis, the matrix A is now transformed into P AP t , 3 where P is the matrix associated to the change of basis, writing the coordinates of the vectors in the new basis as rows of P . We start by analyzing the action of GLk (Zp ) and the affine subgroup    1 0 (k−1) Aff k−1 (Zp ) = | u ∈ Zp , X ∈ GLk−1 (Zp ) ut X (t means transposition), on Ak , giving a set of representatives of the orbits. The result about the action of GLk (Zp ) generalizes naturally if we replace Zp by any principal ideal domain (see for example [8, Theorem IV.1]), but a more elaborate proof is required. For completeness we include a proof here that takes advantage of the fact that Zp is a discrete valuation ring and can be later adapted to the case when the acting group is Aff k−1 (Zp ). (i) Each orbit of the action of GLk (Zp ) on Akcontains  a unique block 0 1 , 1 ≤ i ≤ r and diagonal matrix with diagonal non-zero blocks psi H, H = −1 0 0 ≤ s1 ≤ · · · ≤ sr (0 ≤ r ≤ k/2). Lemma 2.2. (ii) Each orbit of the action of the affine group Aff k−1 (Zp ) on Ak contains a unique tridiagonal matrix A (that is, all the entries aij of A with |i − j| > 1 are zero) with the non-zero entries above the main diagonal ai,i+1 = psi , 1 ≤ i ≤ r, and 0 ≤ s1 ≤ s2 ≤ · · · ≤ sr (0 ≤ r < k). Proof. Given A in Ak , we consider a basis {e1 , . . . , ek } (for instance, the canonical basis) (k) in the free Zp -module Zp and the alternating bilinear form ( , ) defined by the matrix A with respect to this basis. There is nothing to prove if A is the zero matrix, so we can suppose that (ei , ej ) 6= 0 for some 1 ≤ i < j ≤ k and we can assume that its p-adic valuation is minimum among the valuations of all the (non-zero) (er , es ). After reordering the basis, we can suppose that i = 1 and j = 2 and moreover, by multiplying e1 or e2 by a p-adic unit, we can suppose that (e1 , e2 ) = ps1 for some s1 ≥ 0. Notice that any (non-zero) (u, v) has p-adic valuation greater than or equal to ps1 . Now for each i ≥ 3 we set e′i = ei + αi e1 + βi e2 , where αi , βi ∈ Zp are chosen so that (e′i , e1 ) = (e′i , e2 ) = 0. The elements αi , βi exist because the valuation of (e1 , e2 ) is less than or equal to the valuations of (ei , e2 ) and (ei , e1 ). By replacing ei by e′i we can suppose that he1 , e2 i is orthogonal to he3 , . . . , ek i. Proceeding inductively, we obtain a basis {e′1 , . . . , e′k } such that, for some 0 ≤ r ≤ k/2, the subspaces he′2i−1 , e′2i i are pairwise orthogonal for 1 ≤ i ≤ r, the remaining vectors are in the kernel of the form and (e′2i−1 , e′2i ) = psi , 1 ≤ i ≤ r, with 0 ≤ s1 ≤ · · · ≤ sr . It is clear that with respect to this new basis the matrix associated to the form ( , ) has the desired form. To prove uniqueness suppose that A and A′ are block diagonal matrices with (non-zero) ′ ′ diagonal blocks ps1 H, . . . , psr H, 0 ≤ s1 ≤ · · · ≤ sr , and ps1 H, . . . , pst H, 0 ≤ s′1 ≤ · · · ≤ s′t , respectively, and A′ = X t AX for some X ∈ GLk (Zp ). The matrices A, A′ and X can be viewed as endomorphisms of the abelian groups Rn = (Z/pn Z)(k) , n ≥ 1. Since X defines ′ in fact an automorphism of Rn the image subgroups of A and P A (as endomorphisms of 2s Rn ) have the same order. For A this order is p , where s = si ≤n (n − si ), and similarly 4 P P for A′ . We conclude that, for any n ≥ 1, si ≤n (n − si ) = s′ ≤n (n − s′i ), whence r = t i and si = s′i for all 1 ≤ i ≤ r, that is, A = A′ . For the existence part in (ii) we have to show that, given an alternating form ( , ) (k) on Zp and a basis {e1 , . . . , ek }, there exists another basis {e′1 , . . . , e′k } such that e′1 ∈ e1 + he2 , . . . , ek i, he′2 , . . . , e′k i = he2 , . . . , ek i and (e′i , e′j ) = 0 for |i − j| > 1, (ei , ei+1 ) = psi , 0 ≤ s1 ≤ · · · ≤ sr , (ei , ei+1 ) = 0, r < i < k. We can suppose that ( , ) is not the trivial form and then consider the minimum valuation s1 of all the (non-zero) (ei , ej ). If this minimum is attained for some (e1 , ej ) we interchange e2 and ej . Otherwise this minimum is attained for some (ei , ej ), 2 ≤ i < j ≤ k and (e1 + ei , ej ) has still valuation s1 (because the valuation of (e1 , ej ) is strictly greater than s1 ). By replacing e1 by e1 + ei , interchanging e2 and ej , and adjusting units, we can suppose that (e1 , e2 ) = ps1 . Now we can replace ei , i geq3, by e′i = ei + αi e2 , where αi is chosen so that (e1 , e′i ) = 0. Thus we can assume (e1 , ei ) = 0 for i ≥ 3. Now we iterate the procedure with the basis elements e2 , . . . , ek . We prove uniqueness with a similar counting argument as in (i) but this time considering the order of the images of the subgroup of Rn , Sn = {0} × (Z/pn Z)(k−1) . So we assume that A′ = X t AX with A and A′ as in (ii) and X ∈ Aff k−1(Zp ). Notice that, as an automorphism of Rn , X fixes Sn , so the images of Sn by A and A′ must have the P ′ same order. These orders are ps and ps , where s = si ≤n (n − si ) and similarly for s′ , so s = s′ and, since this must happen for any n ≥ 1, it follows that si = s′i for all i, that is, A = A′ . Proposition 2.3. The following words are a system of representatives of the action of Aut Fk on Fk : s sr (3) [x1 , x2 ]p 1 . . . [x2r−1 , x2r ]p , 0 ≤ r ≤ k/2, 0 ≤ s1 ≤ · · · ≤ sr , ps1 ps2 ps3 psr x1 [x1 , x2 ] [x2 , x3 ] . . . [xr−1 , xr ] , 1 ≤ r ≤ k, s1 ≥ 0, 0 ≤ s2 ≤ · · · ≤ sr . (4) Proof. As explained above, the action of Aut Fk on Fk′ can be suitably identified with the action of GLk (Zp ) on Ak , thus it follows directly from the part (i) of the previous lemma that the words (3) are representatives for the orbits contained in Fk′ . s Q Now suppose w ∈ Fk \Fk′ . Then w = (xz11 . . . xzkk )p 1 1≤i<j≤k [xi , xj ]zij , where s1 ≥ 0 and some zi is a p-adic unit. After applying the inverse of the automorphism x1 7→ xz11 . . . xzkk , xi 7→ x1 , xj 7→ xj , for j 6= 1, i, we can assume that xz11 . . . xzkk = x1 . Now we consider the action of the stabilizer of x1 , Autx1 Fk . The image of this subgroup by the first map in (2) is Aff k−1 (Zp ), so we can identify the action of Autx1 Fk on Fk′ with the action of Aff k−1 (Zp ) on Ak . It follows from Lemma 2.2 (ii) that w is equivalent to a word in (4). Notice also that if w ′ = σ(w) for two of these words w and w ′ and some s′1 σ ∈ Aut Fk , it would follow by passing to Fk /Fk′ that σ(x1 )p 1 = xp1 . Since σ induces s automorphisms of (Fk /Fk′ )p and this chain of subg roups of Fk /Fk′ is strictly decreasing, we conclude that s1 = s′1 . But Fk /Fk′ is torsion-free, so σ fixes x1 , that is, σ(x1 ) = x1 z for some z ∈ Fk′ . Composing σ with the automorphism x1 7→ x1 z −1 , xi 7→ xi , i ≥ 2, we s1 can suppose that σ ∈ Autx1 Fk . Thus, the two matrices in Ak associated to x−p w and 1 −ps1 ′ x1 w are in the same orbit by Aff k−1 (Zp ), and so they coincide by Lemma 2.2 (ii). We conclude that w = w ′. s 5 Theorem 2.4. Let G be a finite p-group of nilpotency class 2 and w a word in k variables. Then Nw (1) ≥ |G|k−1. Proof. We can suppose that w is as in the last proposition. Write k0 = ⌊k/2⌋ and fix g2 , g4 . . . , g2k0 ∈ G. Then the map G′ ×G(k−k0 −1) −→ G′ given by (x1 , x3 , . . . , x2(k−k0 )−1 ) 7→ w(x1 , g2 , x3 , . . . ) is a group homomorphism whose kernel has size at least |G|k−k0−1 . Since there are |G|k0 choices for g2 , g4, . . . , g2k0 , we get at least |G|k−1 solutions to the equation w(x1 , . . . , xk ) = 1. 3 The functions Nw from a character-theoretical point of view In this section, unless otherwise stated, we consider an arbitrary finite group G and a word w that is thought now as an element in the free group with, say, free generators x1 , . . . , xk . The function Nw = Nw,G is of course a class function, so it can be written as a linear combination of the irreducible characters of G: X Nw = Nwχ χ, (5) χ∈Irr(G) where Nwχ = (Nw , χ) = 1 X 1 Nw (g)χ(g) = |G| g∈G |G| X χ(w(g1, . . . , gk )). (6) (g1 ,...,gk )∈G(k) It is a natural question to study when the function Nw is a character of G. Notice that in this case Nw (g) ≤ Nw (1) for any element g ∈ G, so Nw (1) will be at least the average of the non-zero values of the function Nw , that is Nw (1) ≥ 1 X |G|k Nw (g) = , |Gw | g∈G |Gw | (7) w which of course improves the bound conjectured by Amit, Nw (1) ≥ |G|k−1. It is easy to find examples where Nw is not a character. Probably the simplest is Q8 , the quaternion group of order 8, and w = x2 : Nx2 ,Q8 (1) = 2 but Nx2 ,Q8 (z) = 6 for the unique involution z ∈ Q8 . For p odd we can construct a p analogue to Q8 as p G = hg1 , . . . , gp−1i ⋊ hgpi/hgp−1 gp−p i, where hg1 , . . . , gp−1 i ∼ = Cp 2 = Cp × · · · × Cp × Cp2 , hgp i ∼ gp gp p (Cn denotes a cyclic group of order n) and gi = gi gi+1 , 1 ≤ i < p − 2, gp−2 = gp−2 gp−1 gp and gp−1 = gp−1 g1−1. It can be checked that Nxp ,G (1) = pp−1 but Nxp ,G (z) = pp + pp−1 for any non-trivial element z ∈ Z(G) = Gxp = hgpp i. Notice that |G| = pp+1 and this is the smallest order for which Nxp ,G can fail to be a character, since p-groups of order at most pp are regular and, for regular p-groups, Nxp ,G is the regular character of G/Gp . Also, for the quaternion group and its p analogues, (7) is false. However, it is a well known result that in general the functions Nxn ,G are generalized characters (i. e., Z-linear combinations of irreducible characters), see [4, Problem 4.7]. 6 For words w in more than one variable the situation is different and there are examples of groups G and words w where Nw,G is not a generalized character, even among nilpotent groups. As for non-solvable examples one can take G = P SL2 (11) and the 2-Engel word w = [x, y, y] (see [10] for another choice of w). Then the coefficients Nwχ for the two √ irreducible characters χ of degree 12 are 305 ± 23 5. More examples can be obtained using the following result by A. Lubotzky [7]: if G is a simple group and 1 ∈ A ⊆ G is a subset invariant under the group of automorphisms of G, then A = Gw for some word w in two variables. Notice that if A contains an element a such that ai 6∈ A, for some i coprime with the order of a, then Nw (ai ) = 0 but Nw (a) 6= 0, something that cannot happen if Nw is a generalized character. Examples of p-groups where Nw is not a generalized character are provided by the free p-groups of nilpotency class 4 and exponent p with p > 2 and p ≡ 1 (mod 4) and w = [x, y, x, y] (which settles in the negative a question of Parzanchevski who had asked if the functions Nw were always generalized characters for solvable or nilpotent groups). We realize these groups as as 1 + J, where J = I/I 4 and I is the ideal generated by X and Y in the algebra of the polynomials in the non-commuting indeterminates X and Y with coefficients in Fp . If x = 1 + X and y = 1 + Y and u = [x, y, x, y], then certainly u ∈ Gw but we claim that if i is not a quadratic residue modulo p, then ui 6∈ Gw (so Nw (u) 6= 0 but Nw (ui ) = 0). Indeed, one can check directly (or by appealing to the Lazard correspondence, but in this case, only for p > 3, see [5, §10.2]) that γ4 (G) = 1 + γ4 (J), where γ4 (J) is the fourth term in the descending central series of the Lie algebra J. Now ui ∈ Gw if and only if i[X, Y, X, Y ] = [aX + bY, cX + dY, aX + bY, cX + dY ] for some a, b, c, d ∈ Fp , and one can see that this equation has no solution if i is not a quadratic residue modulo p. In contrast to the previous example, we shall show that the functions Nw are always generalized characters for p-groups of nilpotency class 2 and, in the next section, that they are actually genuine characters for odd p. Before that we recall briefly that for some words w the functions Nw are known to be characters for any group, notably, for the commutator word w = [x, y] (this is basically [4, Problem 3.10]). This classical result due to Frobenius can be extended in various ways: when w is an admissible word (i. e., a word in which all the variables appear exactly twice, once with exponent 1 and once with −1) [3] or when w = [w ′ , y], where y is a variable which does not occur in w ′ (so in particular, for γk = [x1 , x2 , . . . , xk ], the k-th left-normed lower central word) [12]. It is also clear that if Nw and Nw′ are characters (or generalized characters), so is Nww′ if w and w ′ have no common variables. The reason is of course that Nww′ = Nw ∗ Nw′ is the convolution of the two functions Nw and Nw′ . More information in this direction is given in [10]. As promised we finish this section by proving that the functions Nw are generalized characters for p-groups of nilpotency class 2. We give first a characterization of when Nw is a generalized character. We have already used before that a necessary condition is that Nw (g) = Nw (g i ) for any i coprime with the order of g and we are going to see that this condition is in fact sufficient. The proof is standard once we know that the coefficients Nwχ in (5) are algebraic integers (as Amit and Vishne point out in [1] this follows immediately from (6) and the result of Solomon’s in [11] that Nw (g) is always a multiple of |CG (g)|). 7 Lemma 3.1. Let G be a group and w a word. Then Nw = Nw,G is a generalized character of G if and only if Nw (g) = Nw (g i ) for any g ∈ G and i coprime with the order of G. Proof. We only need to prove sufficiency. Since the coefficients Nwχ are algebraic integers it suffices to see that they are rational numbers. But, by elementary character and Galois theory, if f is a rational-valued class function of a group G, f is a Q-linear combination i of irreducible characters if and P only if f (g) = f (g ) for any g ∈ G and i coprime with the order of G. Indeed, if f = χ∈Irr(G) aχ χ with aχ ∈ Q, and σ is the automorphism of the cyclotomic extension Q(ε)/Q sending ε to εi , where ε is a primitive |G|-th root of unity, we have X X f (g) = f (g)σ = aχ χσ (g) = aχ χ(g i ) = f (g i ). χ∈Irr(G) χ∈Irr(G) i And conversely, if f (g) = f (g ), f (g) = f (g)σ −1 = f (g i )σ −1 =( X aχ χ(g i ))σ −1 X = χ∈Irr(G) −1 aσχ χ(g). χ∈Irr(G) −1 By the linear independence of the irreducible characters, we conclude that aχ = aσχ any automorphism σ, so aχ ∈ Q. for Theorem 3.2. Let G be a p-group of nilpotency class 2 and w a word. Then the function Nw = Nw,G is a generalized character of G. Proof. By Proposition 2.3 we can suppose that w has the form (3) or (4). Now we observe that, if i is not a multiple of p, the map (g1 , g2 , . . . , gk ) 7→ (g1i , g2 , g3i , . . . ) is a bijection from the set of solutions of w = g to the set of solutions of w = g i , so in particular Nw (g) = Nw (g i ) and the result follows from the previous lemma. 4 The functions Nw for odd p-groups of nilpotency class 2 The goal of this section is to show that Nw,G is a genuine character of a p-group G of nilpotency class 2 if p is odd. We begin with a general result. Lemma 4.1. Let χ ∈ Irr(G) with kernel K and w a word in k variables. Then Nwχ = χ , where χ is the character of G/K defined naturally by χ. |K|k−1Nw,G/K Proof. We have Nwχ = (Nw , χ) = (Ñw , χ)G/K , where Ñw is the average function defined by Ñw (g) = 1 |K| P function on G/K. As such a function it is clear that Ñw = |K| clear. n∈K k−1 Nw (gn) viewed as a Nw,G/K , so the result is We assume now that G is a p-group of nilpotency class 2. The following technical result characterizes when Nw is a character. 8 Lemma 4.2. Let G be a p-group of nilpotency class 2. Then Nw is a character of G if and only if for any (non-trivial) epimorphic image of G, say G1 , with cyclic center, Nw,G1 (1) ≥ Nw,G1 (z), where z is a central element of G1 of order p. Proof. By Theorem 3.2 we know that Nw is a generalized character, that is, all the numbers Nwχ are integers, so Nw is a character of G if and only if these numbers are non-negative. We recall that a group G with a faithful irreducible character χ has cyclic center. We claim that if G is a p-group of nilpotency class 2 and χ is a faithful irreducible character then Nwχ ≥ 0 if and only if Nw (1) ≥ Nw (z), where z ∈ Z(G) = Z has order p. Indeed, it is well known that χ(1)χ = η G , where η is a (faithful) linear character of Z(χ) = Z (see for instance [4, Theorem 2.31 and Problem 6.3]). Then Nwχ = 1 1 (Nw , η G ) = (Nw|Z , η)Z . χ(1) χ(1) If the order of Z is pr and z1 is a generator with z = z1p (Nw|Z , η)Z = r−1 we have X j 1 X 1 X j i (εp ) ), Nw (z1p )( Nw (z1i )εi = |Z| 0<i≤pr |Z| 0≤j≤r r−j (8) 0<i≤p (p,i)=1 where ε = η(z1 ) is a primitive pr -th root of unity and we have used Lemma 3.1 for the second equality. Notice that the innermost sum of (8) is the sum of all the primitive pr−j -th roots of unity, which is always zero except in the cases pr−j = 1 or p, when it is 1 or −1, respectively. We conclude that (Nw|Z , η)Z = r−1 1 1 (Nw (1) − Nw (z1p ) = (Nw (1) − Nw (z)) |Z| |Z| and our claim follows. Now we prove the sufficiency part in the lemma. Let χ ∈ Irr(G), K = ker(χ) and G1 = G/K. Of course we can suppose that χ 6= 1G (because Nw1G = |G|k−1 ≥ 0, k is the number of variables of w). By hypothesis Nw,G1 (1) ≥ Nw,G1 (z), where z is a central element of G1 of order p. We can view χ as a faithful character χ of G1 and then our χ ≥ 0. By Lemma 4.1, Nwχ ≥ 0, which shows that Nw is previous claim implies that Nw,G 1 a character. Conversely, suppose that Nw is a character, that is, all the numbers Nwχ are nonnegative, and consider an epimorphic image G1 = G/N with cyclic center and a central element z ∈ G1 of order p. Then G has an irreducible character χ with kernel N that is faithful when considered as a character χ of G1 . Since Nwχ ≥ 0, again by Lemma 4.1, χ ≥ 0 and, by our initial claim, the inequality Nw,G1 (1) ≥ Nw,G1 (z) follows. Nw,G 1 Theorem 4.3. Let G be a p-group of nilpotency class 2, p odd, and w a word. Then Nw is a character of G. 9 Proof. By the last result it suffices to show that if G has cyclic center Z and z ∈ Z has order p, Nw (1) ≥ Nw (z). Also we can assume that w has the form (4) (if w is as in (3), skip the next two paragraphs). s1 If Z p 1 6= 1, we can write z = z1p for some z1 ∈ Z and then the map (g1 , g2 , . . . , gk ) 7→ (g1 z1 , g2 , . . . , gk ) is a bijection between the sets of solutions of w = 1 and w = z, so Nw (1) = Nw (z). s s Now we suppose that Z p 1 = 1 and notice that, since G has nilpotency class 2 and p is odd, ps1 s s s s s (xy)p 1 = xp 1 y p 1 [y, x]( 2 ) = xp 1 y p 1 . (9) Therefore if we fix g2 , g4 , · · · ∈ G, the map (g1 , g3 , . . . ) 7→ w(g1 , g2 , . . . , gk ) is a group homomorphism ϕg2 ,g4 ,.... Obviously there is a bijection between the kernel of this homomorphism and the set of solutions of w = 1 with x2i = g2i . As for the solutions of w = z with x2i = g2i , either this set is empty or else its elements are in one-to-one correspondence with the elements in a coset of the kernel of ϕg2 ,g4,... . In any case, considering only solutions with x2i = g2i , the number of solutions of w = 1 is greater than or equal to the number of solutions of w = z. Varying g2 , g4 ,. . . , we conclude Nw (1) ≥ Nw (z), as desired. 5 The functions Nxn for 2-groups of nilpotency class 2 In this section we study the functions Nxn for 2-groups of nilpotency class 2 and characterize when this function is a character. As we already pointed out in Section 3, the function Nx2, Q8 is not a character and in fact for each r ≥ 1 we can define a 2-group Q23r of order 23r , which is the usual quaternion group Q8 when r = 1, such that Nx2r, Q23r is not a character. We shall see that this group is in some sense involved in G whenever Nx2r, G is not a character. We shall also need to introduce another family of groups, denoted D23r , that, for r = 1, is the usual dihedral group of order 8. Definition 5.1. Let r ≥ 1. We define the quasi dihedral and quasi quaternion group, D23r and Q23r , as r r r D23r = hx, z1 , y | x2 = z12 = y 2 = 1, [x, z1 ] = 1, [x, y] = z1 , [z1 , y] = 1i, r r r−1 Q23r = hx, z1 , y | z12 = 1, x2 = z12 r = y 2 , [x, z1 ] = 1, [x, y] = z1 , [z1 , y] = 1i. One can check that, if G = D23r or Q23r , G has order 23r , exponent 2r+1 and G′ = r−1 is the central involution, in the Z(G) = hz1 i is cyclic of order 2r . Moreover, if z = z12 3r−2 and Nx2r (z) = 23r−2 , whereas in the quaternion (quasi) dihedral case Nx2r (1) = 3 × 2 case the numbers are in the reverse order: Nx2r (1) = 23r−2 and Nx2r (z) = 3 × 23r−2 (and so Nx2r is not a character of Q23r ). If T and H are 2-groups with cyclic center and |Z(T )| ≤ |Z(H)|, we shall denote T ∗ H the central product of T and H with Z(T ) amalgamated with the corresponding subgroup 10 of Z(H). Notice that if all the generators of Z(T ) are in the same orbit under the action of the automorphism group of T (or if a similar situation holds in H), the group T ∗ H is unique up to isomorphism and this is what happens if T = D23r or Q23r . Also, for a p-group G, Ωr (G) is the subgroup generated by the elements of order at most pr . Lemma 5.1. Let G be a 2-group of nilpotency class 2 and cyclic center Z of order 2r . Suppose that Ωr+1 (G)′ = Z. Then G = T ∗ H, where T is isomorphic to D23r or Q23r . Proof. Since G has nilpotency class 2, Ωr+1 (G)′ is generated by the commutators of elements of order at most 2r+1 and it is cyclic, because it is contained in Z, which is cyclic, so it is generated by one of these commutators, say [x, y]. The orders of x and y have to be 2r or 2r+1 (because [x, y] has order 2r ). If both have order 2r it is clear that T = hx, yi is isomorphic to D23r and, if both have order 2r+1 , is isomorphic to Q23r (notice that r r r G2 ⊆ Z, so x2 = y 2 ). On the other hand, if one is of order 2r and the other of order 2r+1 , their product has order 2r and T is isomorphic to D23r again. Now it suffices to check that G = T CG (T ) (because obviously T ∩ CG (T ) = Z(T ) has order 2r , and so is the center of G). Indeed, the conjugacy class of x has order |[x, G]| = |G′| = |Z| = pr and the same for y, so |G : CG (T )| = |G : CG (x) ∩ CG (y)| ≤ |G : CG (x)||G : CG (y)| = 22r . But |T CG (T ) : CG (T )| = |T : T ∩ CG (T )| = |T : Z| = 22r , so G = T CG (T ), as claimed. One can check that, as it happens with the usual dihedral and quaternion groups, D23r ∗ D23r and Q23r ∗ Q23r are isomorphic. Using this result and iterating Lemma 5.1 we get the following. Proposition 5.2. Let G be a 2-group of nilpotency class 2 and cyclic center Z of order 2r . Then G is isomorphic to a group D23r ∗ . n. . ∗D23r ∗ H or D23r ∗ . n. . ∗D23r ∗ Q23r ∗ H, n ≥ 0, where H has cyclic center of order 2r and Ωr+1 (H)′ is properly contained in the center of H. Now suppose that G = T ∗H, where T = D23r or Q23r and H is a 2-group of nilpotency r r class 2 and cyclic center of order 2r . For any g ∈ T , g 2 = 1 or z, thus if h ∈ H, (gh)2 = 1 r r r r r if and only if g 2 = h2 = 1 or z. Similarly, (gh)2 = z if and only if g 2 = 1 and h2 = z or the other way round. This means that Nx2r, G (1) = (Nx2r, T (1)Nx2r, H (1) + Nx2r, T (z)Nx2r, H (z))/2r Nx2r, G (z) = (Nx2r, T (1)Nx2r, H (z) + Nx2r, T (z)Nx2r, H (1))/2r , whence Nx2r, G (1) = 22r−2 (3Nx2r, H (1) + Nx2r, H (z)) or 22r−2 (Nx2r, H (1) + 3Nx2r, H (z)) Nx2r, G (z) = 22r−2 (3Nx2r, H (z) + Nx2r, H (1)) or 22r−2 (Nx2r, H (z) + 3Nx2r, H (1)), 11 depending on whether T = D23r or Q23r , respectively. It follows that, in the former case, Nx2r, G (1) ≥ Nx2r, G (z) if and only if Nx2r, H (1) ≥ Nx2r, H (z) but, in the latter case, this holds if and only if Nx2r, H (1) ≤ Nx2r, H (z) (and the same equivalences hold if inequalities are replaced by equalities). The same happens if T = D23r ∗ . n. . ∗D23r or D23r ∗ . n. . ∗D23r ∗Q23r , n ≥ 0. The combination of this with Proposition 5.2 basically reduces our problem to the groups G with Ωr+1 (G)′ properly contained in the center, which is the situation considered in the next lemma. Lemma 5.3. Let G be a 2-group of nilpotency class 2 and cyclic center of order 2r and let z be the unique central involution. Suppose that Ωr+1 (G)′ is properly contained in Z = Z(G). Then Nx2r (1) > Nx2r (z) if and only if G has exponent 2r . Otherwise Nx2r (1) = Nx2r (z). r r Proof. Since (G′ )2 ⊆ Z 2 = 1, raising to the 2r+1-th power is a group endomorphism r+1 of G by (9) and Ωr+1 (G) = {x ∈ G | x2 = 1}. Moreover, Ωr+1 (G)′ is contained in r−1 Z 2 , so (Ωr+1 (G)′ )2 = 1 and raising to the 2r -th power is a group endomorphism of r Ωr+1 (G) with kernel Ωr (G) = {x ∈ G | x2 = 1}. It is clear now that Nx2r (1) = |Ωr (G)| r and Nx2r (z) = |Ωr+1 (G)| − |Ωr (G)| (for any element x in Ωr+1 (G)\Ωr (G), x2 is a central involution, so it is z), so Nx2r (1) > Nx2r (z) if and only if |Ωr+1 (G) : Ωr (G)| < 2, that r r is Ωr+1 (G) = Ωr (G) = G, i. e., G has exponent 2r . Otherwise 1 6= Ωr+1 (G)2 = {x2 | r x ∈ Ωr+1 (G)}, thus z ∈ Ωr+1 (G)2 is in the image of the 2r -th power endomorphism of Ωr+1 (G) and Nx2r (z) = |Ωr (G)| = Nx2r (1). Proposition 5.4. Let G be a 2-group of nilpotency class 2, cyclic center of order 2r and central involution z. Then Nx2r (1) < Nx2r (z) if and only if G is isomorphic to a group D23r ∗ . n. . ∗D23r ∗ Q23r ∗ H, n ≥ 0, where H has cyclic center of order 2r and exponent 2r . Proof. By Proposition 5.2, G has two possible decompositions as a central product with one factor H satisfying the hypotheses of Lemma 5.3. If Q23r does not occur in the decomposition of G, we know that Nx2r (1) < Nx2r (z) if and only if Nx2r, H (1) < Nx2r, H (z), something that, according to Lemma 5.3, never happens. Therefore Q23r does occur in the decomposition of G. In this case, we know that Nx2r (1) < Nx2r (z) if and only if Nx2r, H (1) > Nx2r, H (z), which, by Lemma 5.3, is equivalent to H having exponent 2r . It is not difficult to classify the groups H in the previous proposition. Lemma 5.5. Let H be a 2-group of nilpotency class 2, cyclic center of order 2r and exponent 2r . Then H ∼ = D23r1 ∗ · · · ∗ D23rn ∗ C2r with r1 ≤ · · · ≤ rn < r. Proof. The result is trivially true if H is abelian, so suppose H ′ = h[x, y]i = 6 1 has order s 2r 2s 2s 2 . Since (xy) = 1, (9) yields s < r. The elements x and y are central with orders s at most 2r−s , so they lie in hz12 i, where Z(H) = hz1 i, and, for suitable i and j, xz1i and yz1j have order exactly 2s . By replacing x and y by these elements, we can suppose that T = hx, yi ∼ = D23s . Arguing as in the last part of the proof of Lemma 5.1, H = T CH (T ) and T ∩ CH (T ) = Z(T ) is cyclic of order 2s . Since Z(H) ≤ CH (T ), the hypotheses still hold in CH (T ), so we can apply induction. 12 The last two results show that, with the hypotheses of Proposition 5.4, Nx2r (1) < Nx2r (z) if and only if G ∼ = D23r1 ∗ · · · ∗ D23rn ∗ Q23r , n ≥ 0, r1 ≤ · · · ≤ rn ≤ r. Notice simply that the cyclic factor of H is absorbed by Q23r . Theorem 5.6. Let G be a 2-group of nilpotency class 2. Then Nx2r is a character of G if and only if G has no epimorphic image isomorphic to D23r1 ∗ · · · ∗ D23rn ∗ Q23r , n ≥ 0, r1 ≤ · · · ≤ rn ≤ r. Proof. If G has an epimorphic image G1 of the indicated type, then by the last remark, Nx2r, G1 (1) < Nx2r, G1 (z) (z is the central involution of G1 ) and, by Lemma 4.2, Nx2r is not a character of G. Conversely, if Nx2r is not a character of G, by the same lemma, G has an epimorphic image G1 with cyclic center such that Nx2r, G1 (1) < Nx2r, G1 (z). We claim that the center Z of G1 has order 2r (and then, again by the last remark, G1 is the r desired epimorphic image). The map x 7→ x2 cannot be a group endomorphism of G1 (this would immediately imply that either Nx2r, G1 (z) = 0 or else Nx2r, G1 (1) = Nx2r, G1 (z)), r−1 r−1 hence (G′1 )2 6= 1 and Z 2 6= 1, that is |Z| ≥ 2r (here Z is the center of G1 ). If r |Z| > 2r , z = z12 for some z1 ∈ Z and then Nx2r, G1 (1) = Nx2r, G1 (z) because x 7→ xz1 maps r r bijectively the solutions of x2 = 1 to the solutions of x2 = z. Thus |Z| = 2r and the proof is complete. 6 p-groups with central Frattini subgroup In this section we consider a d-generated p-group G such that Φ(G) ≤ Z(G), that is, G has nilpotency class 2 and elementary abelian derived subgroup. We shall show that for the words wk = [x1 , x2 ] . . . [x2k−1 , x2k ], with k ≥ d0 = ⌊d/2⌋, Nwk (g) ≥ |G|2k−1 for all g ∈ Gwk . (10) For k = 1, (10) can be easily proved for any p-group of nilpotency class 2. Indeed, if g = [x, y], we can multiply x by any element commuting with y and y by any central element, thus Nw1 (g) ≥ |CG (y)||Z(G)| = |G||Z(G) : [y, G]| ≥ |G|. If V = G/Φ(G), as a vector space over Fp , there is a natural surjective linear V2 Vviewed 2 map π from = V , the exterior square of V , onto G′ given by x ∧ y 7→ [x, y]. For a V2 fixed ω ∈ and k ≥ 0, it is then natural to consider the number Nwk (ω) = Nwk ,V (ω) of solutions in V (2k) of the equation wk (x1 , . . . , x2k ) = x1 ∧ x2 + · · · + x2k−1 ∧ x2k = ω. The setVof values of wk , that is, the set {wk (v1 , . . . , v2k ) | vi ∈ V } ⊆ simply 2wk . (11) V2 will be denoted Similarly as in Section V22, if we fix a basis {e1 , . . . , ed } of V there is a one-to-one correspondence between and Ad , the set of d × d antisymmetric matrices over the P field Fp , given by 1≤i<j≤d aij (ei ∧ ej ) 7→ A, where A ∈ Ad has entries aij for 1 ≤ i < j ≤ d. Then solving (11) amounts to solving the matrix equation X t Jk X = A, where X represents a 2k × d matrix, Jk is the 2k × 2k block diagonal matrix with repeated diagonal 13 block  0 1 −1 0  and A ∈ Ad . Moreover, the number of solutions only depends on the   Jr 0 with r ≤ k (otherwise there rank of A, so there is no loss to assume that A = 0 0 are no solutions). This number, which is denoted N(K2k , Kd,2r ) in [13], V2wasVoriginally considered by Carlitz [2] and later on by other authors (see [13]). If ω ∈ wr \ 2wr−1 , the corresponding antisymmetric matrix A has rank 2r, so the number N(K2k , Kd,2r ) is, in our notation, Nwk (ω). An explicit formula for this number is given in [13, Theorems 3,4, 5] that, for r ≤ k ≤ d0 = ⌊d/2⌋, can be written as N(K2k , Kd,2r ) = p k Y r(2k−r) 2i (p − 1) i=k−r+1 k−r X j=0   j d − 2r ) ( p2 j p k−r Y (p2i − 1), (12) i=k−r−j+1     n n n n−j+1 j = 1. If we = (p − 1) . . . (p − 1)/(p − 1) . . . (p − 1) for j > 0 and where 0 p j p just consider in (12) the summand corresponding to j = k − r we get the inequality N(K2k , Kd,2r ) ≥ p r(2k−r)   k Y k−r d − 2r 2i . (p − 1)p( 2 ) k−r p (13) i=1   Q Q n > ( ni=n−j+1 pi−1 )/( ji=1 pi ) = pnj , = p , and But i=1 (p − 1) > i=1 p j p therefore k−r 2 N(K2k , Kd,2r ) > pr(2k−r)+k +( 2 )+(d−2r)(k−r) . Qk 2i Qk 2i−1 k2 The quadratic function of r in the exponent of p in this formula is decreasing in the interval 0 ≤ r ≤ k, so we conclude that V 2 Nwk (ω) = N(K2k , Kd,2r ) > p2k for any ω ∈ 2wk . (14) V V Since the rank of a d × d antisymmetric matrix is at most 2d0 we have that 2wk = 2 V for k ≥ d0 . But for any k, π maps 2wk onto Gwk , thus Gwk = G′ for k ≥ d0 . Now it is easy to show that if (10) holds for w = wd0 , it also holds for w = wk for any k ≥ d0 . Indeed, if k ≥ d0 , we have Nwk = Nwd0 ∗ Nwk−d0 and, since we are assuming that Nwd0 (x) ≥ |G|2d0 −1 for any x ∈ G′ , if g ∈ G′ , X X Nwk−d0 (y) Nwd0 (gy −1)Nwk−d0 (y) ≥ |G|2d0 −1 Nwk (g) = y∈Gwk−d = |G| y∈Gwk−d 0 2d0 −1 |G| 2(k−d0 ) = |G| 2k−1 0 . So in order to prove (10) for wk we can always assume that 1 ≤ k ≤ d0 . It is clear that if g ∈ G′ and π(ω) = g, the solutions of wk (x1 , . . . , x2k ) = ω in V can be lifted to solutions of wk (x1 , . . . , x2k ) = g in G and of course, all solutions of the equation in G occur in this way, so X Nwk ,V (ω) Nwk (g) = |Φ(G)|2k ω∈π −1 (g) 14 and (10) can be written now as X |Φ(G)| Nwk ,V (ω) ≥ |G : Φ(G)|2k−1 = pd(2k−1) (15) ω∈π −1 (g) V for g ∈ Gwk . Obviously only the ω’s in π −1 (g) ∩ 2wk contribute to this sum and for them we can use the estimation (14). Thus, since |G′ | = pd(d−1)/2 /|ker(π)| ≤ |Φ(G)|, the inequality V d(d−1) 2 (16) p 2 +2k −d(2k−1) |π −1 (g) ∩ 2wk | ≥ |ker(π)| V implies (15). If k = d0 , |π −1 (g) ∩ 2wk | = |π −1 (g)| = |ker(π)| and (16) holds because the exponent of p is positive (it is positive for any k). We conclude the following result. Proposition 6.1. Let G be a d-generated p-group with Φ(G) ≤ Z(G). Then for any k ≥ ⌊d/2⌋ and g ∈ G′ , Nwk (g) ≥ |G|2k−1. Another situation in which |π −1 (g) ∩ we also have the following result. V2 wk | = |ker(π)| is when π is an isomorphism, so Proposition 6.2. Let G be a d-generated p-group with Φ(G) ≤ Z(G) and |G′ | = pd(d−1)/2 . Then for any k ≥ 1 and g ∈ Gwk , Nwk (g) ≥ |G|2k−1. Notice that the last proposition applies in particular to the free p-groups of nilpotency class 2 and exponent p and, in this case the inequality Nw (g) ≥ |G|k−1, g ∈ Gw , is in fact true for any word w ∈ Fk . This is clear if w ∈ Fk′ and otherwise we can suppose that s1 = 0 in (4). But in this case all equations w = g have the same number of solutions, namely, |G|2k−1 . References [1] A. Amit, U. Vishne, Characters and solutions to equations in finite groups, J. Algebra Appl. 10, no. 4, (2011) 675–686. [2] L. Carlitz, Representations by skew forms in a finite field, Arch. Math. 5 (1954) 19–31. [3] A. K. Das, R. K. Nath, On solutions of a class of equations in a finite group, Comm. Algebra 37, no. 11, (2009) 3904–3911. [4] I. M. Isaacs, Character Theory of Finite Groups, Dover Publications INC., New York, 1994. [5] E. Khukhro, p-automorphisms of finite p-groups, Cambridge University Press, Cambridge, 1998. [6] M. Levy, On the probability of satisfying a word in nilpotent groups of class 2, arXiv:1101.4286 (2011). 15 [7] A. Lubotzky, Images of word maps in finite simple groups, arXiv:1211.6575 (2012). [8] M. Newman, Integral matrices, Academic Press, New York, 1972. [9] N. Nikolov, D. Segal, A characterization of finite soluble groups, Bull. Lond. Math. Soc. 39 (2007) 209–213. [10] O. Parzanchevski, G. Schul, On the Fourier expansion of word maps, Bull. Lond. Math. Soc. 46 (2014) 91–102. [11] L. Solomon, The solution of equations in groups, Arch. Math. 20 (1969) 241–247 [12] S. P. Strunkov, On the theory of equations on finite groups, Izv. Math. 59, no. 6, (1995) 1273–1282. [13] J. Wei, Y. Zhang, The number of solutions to the alternate matrix equation over a finite field and a q-identity, J. Statist. Plann. Inference 94, no. 2, (2001) 349–358. 16 

1-String B2-VPG Representation of Planar Graphs∗ Therese Biedl1 and Martin Derka1 1 David R. Cheriton School of Computer Science, University of Waterloo 200 University Ave W, Waterloo, ON N2L 3G1, Canada {biedl,mderka}@uwaterloo.ca arXiv:1411.7277v2 [cs.CG] 3 Dec 2015 Abstract In this paper, we prove that every planar graph has a 1-string B2 -VPG representation—a string representation using paths in a rectangular grid that contain at most two bends. Furthermore, two paths representing vertices u, v intersect precisely once whenever there is an edge between u and v. We also show that only a subset of the possible curve shapes is necessary to represent 4-connected planar graphs. 1998 ACM Subject Classification I.3.5 Computational Geometry and Object Modeling Keywords and phrases Graph drawing, string graphs, VPG graphs, planar graphs Digital Object Identifier 10.4230/LIPIcs.xxx.yyy.p 1 Preliminaries One way of representing graphs is to assign to every vertex a curve so that two curves cross if and only if there is an edge between the respective vertices. Here, two curves u, v cross means that they share a point s internal to both of them and the boundary of a sufficiently small closed disk around s is crossed by u, v, u, v (in this order). Such a representation of graphs using crossing curves is referred to as a string representation, and graphs that can be represented in this way are called string graphs. In 1976, Ehrlich, Even and Tarjan showed that every planar graph has a string representation [8]. It is only natural to ask if this result holds if one is restricted to using only some “nice” types of curves. In 1984, Scheinerman conjectured that all planar graphs can be represented as intersection graphs of line segments [12]. This was proved first for bipartite planar graphs [7, 10] with the strengthening that every segment is vertical or horizontal. The result was extended to triangle-free planar graphs, which can be represented by line segments with at most three distinct slopes [6]. Since Scheinerman’s conjecture seemed difficult to prove for all planar graphs, interest arose in possible relaxations. Note that any two line segments can intersect at most once. Define 1-String to be the class of graphs that are intersection graphs of curves (of arbitrary shape) that intersect at most once. We also say that graphs in this class have a 1-string representation. The original construction of string representations for planar graphs given in [8] requires curves to cross multiple times. In 2007, Chalopin, Gonçalves and Ochem showed that every planar graph is in 1-String [3, 4]. With respect to Scheinerman’s conjecture, while the argument of [3, 4] shows that the prescribed number of intersections can be achieved, it provides no idea on the complexity of curves that is required. ∗ Research supported by NSERC. The second author was supported by the Vanier CGS. A preliminary version appeared at the Symposium on Computational Geometry 2015. © Therese Biedl and Martin Derka; licensed under Creative Commons License CC-BY Conference title on which this volume is based on. Editors: Billy Editor and Bill Editors; pp. 1–28 Leibniz International Proceedings in Informatics Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany 2 1-String B2 -VPG Representation of Planar Graphs Another way of restricting curves in string representations is to require them to be orthogonal, i.e., to be paths in a grid. Call a graph a VPG-graph (as in “Vertex-intersection graph of Paths in a Grid”) if it has a string representation with orthogonal curves. It is easy to see that all planar graphs are VPG-graphs (e.g. by generalizing the construction of Ehrlich, Even and Tarjan). For bipartite planar graphs, curves can even be required to have no bends [7, 10]. For arbitrary planar graphs, bends are required in orthogonal curves. Recently, Chaplick and Ueckerdt showed that two bends per curve always suffice [5]. Let B2 -VPG be the graphs that have a string representation where curves are orthogonal and have at most two bends; the result in [5] then states that planar graphs are in B2 -VPG. Unfortunately, in Chaplick and Ueckerdt’s construction, curves may cross each other twice, and so it does not prove that planar graphs are in 1-String. The conjecture of Scheinerman remained open until 2009 when it was proved true by Chalopin and Gonçalves [2]. Our Results: In this paper, we show that every planar graph has a string representation that simultaneously satisfies the requirements for 1-String (any two curves cross at most once) and the requirements for B2 -VPG (any curve is orthogonal and has at most two bends). Our result hence re-proves, in one construction, the results by Chalopin et al. [3, 4] and the result by Chaplick and Ueckerdt [5]. I Theorem 1. Every planar graph has a 1-string B2 -VPG representation. In addition to Theorem 1, we show that for 4-connected planar graphs, only a subset of orthogonal curves with 2 bends is needed: I Theorem 2. Every 4-connected planar graph has a 1-string B2 -VPG representation where all curves have a shape of C or Z (including their horizontal mirror images). Our approach is inspired by the construction of 1-string representations from 2007 [3, 4]. The authors proved the result in two steps. First, they showed that triangulations without separating triangles admit 1-string representations. By induction on the number of separating triangles, they then showed that a 1-string representation exists for any planar triangulation, and consequently for any planar graph. In order to show that triangulations without separating triangles have 1-string representations, Chalopin et al. [4] used a method inspired by Whitney’s proof that 4-connected planar graphs are Hamiltonian [13]. Asano, Saito and Kikuchi later improved Whitney’s technique and simplified his proof [1]. Our paper uses the same approach as [4], but borrows ideas from [1] and develops them further to reduce the number of cases. 2 Definitions and Basic Results Let us begin with a formal definition of a 1-string B2 -VPG representation. I Definition 3 (1-string B2 -VPG representation). A graph G has a 1-string B2 -VPG representation if every vertex v of G can be represented by a curve v such that: 1. Curve v is orthogonal, i.e., it consists of horizontal and vertical segments. 2. Curve v has at most two bends. 3. Curves u and v intersect at most once, and u intersects v if and only if (u, v) is an edge of G. T. Biedl and M. Derka We always use v to denote the curve of vertex v, and write vR if the representation R is not clear from the context. We also often omit “1-string B2 -VPG” since we do not consider any other representations. Our technique for constructing representations of a graph uses an intermediate step referred to as a “partial 1-string B2 -VPG representation of a W-triangulation that satisfies the chord condition with respect to three chosen corners.” We define these terms, and related graph terms, first. A planar graph is a graph that can be embedded in the plane, i.e., it can be drawn so that no edges intersect except at common endpoints. All graphs in this paper are planar. We assume throughout the paper that one combinatorial embedding of the graph has been fixed by specifying the clockwise (CW) cyclic order of incident edges around each vertex. Subgraphs inherit this embedding, i.e., they use the induced clockwise orders. A facial region is a connected region of R2 − Γ where Γ is a planar drawing of G that conforms with the combinatorial embedding. The circuit bounding this region can be read from the combinatorial embedding of G and is referred to as a facial circuit. We sometimes refer to both facial circuit and facial region as a face when the precise meaning is clear from the context. The outer-face is the one that corresponds to the unbounded region; all others are called interior faces. The outer-face cannot be read from the embedding; we assume throughout this paper that the outer-face of G has been specified. Subgraphs inherit the outer-face by using as outer-face the one whose facial region contains the facial region of the outer-face of G. An edge of G is called interior if it does not belong to the outer-face. A triangulated disk is a planar graph G for which the outer-face is a simple cycle and every interior face is a triangle. A separating triangle is a cycle C of length 3 such that G has vertices both inside and outside the region bounded by C (with respect to the fixed embedding and outer-face of G). Following the notation of [4], a W-triangulation is a triangulated disk that does not contain a separating triangle. A chord of a triangulated disk is an interior edge for which both endpoints are on the outer-face. Let X, Y be two vertices on the outer-face of a connected planar graph so that neither of them is a cut vertex. Define PXY to be the counter-clockwise (CCW) path on the outer-face from X to Y (X and Y inclusive). We often study triangulated disks with three specified distinct vertices A, B, C called the corners. A, B, C must appear on the outer-face in CCW order. We denote PAB = (a1 , a2 , . . . , ar ), PBC = (b1 , b2 , . . . , bs ) and PCA = (c1 , c2 , . . . , ct ), where ct = a1 = A, ar = b1 = B and bs = c1 = C. I Definition 4 (Chord condition). A W-triangulation G satisfies the chord condition with respect to the corners A, B, C if G has no chord within PAB , PBC or PCA , i.e., no interior edge of G has both ends on PAB , or both ends on PBC , or both ends on PCA .1 I Definition 5 (Partial 1-string B2 -VPG representation). Let G be a connected planar graph and E 0 ⊆ E(G) be a set of edges. An (E 0 )-1-string B2 -VPG representation of G is a 1-string B2 -VPG representation of the subgraph (V (G), E 0 ), i.e., curves u, v cross if and only if (u, v) is an edge in E 0 . If E 0 consists of all interior edges of G as well as some set of edges F on the outer-face, then we write (int ∪ F ) representation instead. In our constructions, we use (int ∪ F ) representations with F = ∅ or F = {e}, where e is an outer-face edge incident to corner C of a W-triangulation. Edge e is called the special 1 For readers familiar with [4] or [1]: A W-triangulation that satisfies the chord condition with respect to corners A, B, C is called a W-triangulation with 3-boundary PAB , PBC , PCA in [4], and the chord condition is the same as Condition (W2b) in [1]. 3 4 1-String B2 -VPG Representation of Planar Graphs edge, and we sometimes write (int ∪ e) representation, rather than (int ∪ {e}) representation. 2.1 2-Sided, 3-Sided and Reverse 3-Sided Layouts To create representations where vertex-curves have few bends, we need to impose geometric restrictions on representations of subgraphs. Unfortunately, no one type of layout seems sufficient for all cases, and we will hence have three different layout types illustrated in Figure 1. I Definition 6 (2-sided layout). Let G be a connected planar graph and A, B be two distinct outer-face vertices neither of which is a cut vertex in G. Furthermore, let G be such that all cut vertices separate it into at most two connected components. An (int ∪ F ) B2 -VPG representation of G (for some set F ) has a 2-sided layout (with respect to corners A, B) if: 1. There exists a rectangle Θ that contains all intersections of curves and such that (i) the top of Θ is intersected, from right to left in order, by the curves of the vertices of PAB , (ii) the bottom of Θ is intersected, from left to right in order, by the curves of the vertices of PBA . 2. Any curve v of an outer-face vertex v has at most one bend. (By 1., this implies that A and B have no bends.) I Definition 7 (3-sided layout). Let G be a W -triangulation and A, B, C be three distinct vertices in CCW order on the outer-face of G. Let F be a set of exactly one outer-face edge incident to C. An (int ∪ F ) B2 -VPG representation of G has a 3-sided layout (with respect to corners A, B, C) if: 1. There exists a rectangle Θ containing all intersections of curves so that (i) the top of Θ is intersected, from right to left in order, by the curves of the vertices on PAB ; (ii) the left side of Θ is intersected, from top to bottom in order, by the curves of the vertices on PBbs−1 , possibly followed by C; 2 (iii) the bottom of Θ is intersected, from right to left in order, by the curves of vertices on Pc2 A in reversed order, possibly followed by C;2 (iv) curve bs = C = c1 intersects the boundary of Θ exactly once; it is the bottommost curve to intersect the left side of Θ if the special edge in F is (C, c2 ), and C is the leftmost curve to intersect the bottom of Θ if the special edge in F is (C, bs−1 ). 2. Any curve v of an outer-face vertex v has at most one bend. (By 1., this implies that B has precisely one bend.) 3. A and C have no bends. We also need the concept of a reverse 3-sided layout, which is similar to the 3-sided layout except that B is straight and A has a bend. Formally: I Definition 8 (Reverse 3-sided layout). Let G be a connected planar graph and A, B, C be three distinct vertices in CCW order on the outer-face of G. Let F be a set of exactly one outer-face edge incident to C. An (int ∪ F ) B2 -VPG representation of G has a reverse 3-sided layout (with respect to corners A, B, C) if: 2 Recall that (bs−1 , C) and (C, c2 ) are the two incident edges of C on the outer-face. T. Biedl and M. Derka 5 1. There exists a rectangle Θ containing all intersections of curves so that (i) the right side of Θ is intersected, from bottom to top in order, by the curves of the vertices on PAB ; (ii) the left side of Θ is intersected, from top to bottom in order, by the curves of the vertices on PBbs−1 , possibly followed by C; (iii) the bottom of Θ is intersected, from right to left in order, by the curves of vertices on Pc2 A in reversed order, possibly followed by C; (iv) curve bs = C = c1 intersects the boundary of Θ exactly once; it is the bottommost curve to intersect the left side of Θ if the special edge in F is (C, c2 ), and C is the leftmost curve to intersect the bottom of Θ if the special edge in F is (C, bs−1 ). 2. Any curve v of an outer-face vertex v has at most one bend. (By 1., this implies that A has precisely one bend.) 3. B and C have no bends. B other curves of PAB B other curves of PBC other curves of PBC bs-1 bs-1 C? B other curves of PBA A C? c2 other curves of PAB other curves of PAB other curves of PCA A C? C? c2 other curves of PCA A Figure 1 Illustration of a 2-sided layout, 3-sided layout, and reverse 3-sided layout. We sometimes refer to the rectangle Θ for these representations as a bounding box. Figure 2 (which will serve as base case later) shows such layouts for a triangle and varying choices of F. 2.2 Private Regions Our proof starts by constructing a representation for triangulations without separating triangles. The construction is then extended to all triangulations by merging representations of subgraphs obtained by splitting at separating triangles. To permit the merge, we apply the technique used in [4] (and also used independently in [9]): With every triangular face, create a region that intersects the curves of vertices of the face in a predefined way and does not intersect anything else, specifically not any other such region. Following the notation of [9], we call this a private region (but we use a different shape). I Definition 9 (Chair-shape). A chair-shaped area is a region bounded by a 10-sided orthogonal polygon with CW (clockwise) or CCW (counter-clockwise) sequence of interior angles 90°, 90°, 270°, 270°, 90°, 90°, 90°, 90°, 270°, 90°. See also Figure 3. I Definition 10 (Private region). Let G be a planar graph with a partial 1-string B2 -VPG representation R and let f be a facial triangle in G. A private region of f is a chair-shaped area Φ inside R such that: 1. Φ is intersected by no curves except for the ones representing vertices on f . 2. All the intersections of R are located outside of Φ. 6 1-String B2 -VPG Representation of Planar Graphs B B B A A Θ C A B A Θ B C A B C A B A Θ B C Θ B C A A Θ B A B C A B B A B C B A Θ A Θ C A Figure 2 (int ∪ F ) representations of a triangle: (Top) 2-sided representations for F ∈ {{(A, C)}, {(B, C)}, ∅}. (Bottom) 3-sided and reverse 3-sided representations for F ∈ {{(A, C)}, {(B, C)}}. Private regions are shaded in grey. 3. For a suitable labeling of the vertices of f as {a, b, c}, Φ is intersected by two segments of a and one segment of b and c. The intersections between these segments and Φ occur at the edges of Φ as depicted in Figure 3. 2.3 The Tangling Technique Our constructions will frequently use the following “tangling technique”. Consider a set of k vertical downward rays s1 , s2 , s3 , . . . , sk placed beside each other in left to right order. The operation of bottom-tangling from s1 to sk rightwards stands for the following (see also Figure 4): 1. For 1 < i ≤ k, stretch si downwards so that it ends below si−1 . 2. For 1 ≤ i < k, bend si rightwards and stretch it so that it crosses si+1 , but so that it does not cross si+2 . We similarly define right-tangling upwards, top-tangling leftwards and left-tangling downwards as rotation of bottom-tangling rightwards by 90°, 180° and 270° CCW. We define bottom-tangling leftwards as a horizontal flip of bottom-tangling rightwards, and right-tangling downwards, top-tangling rightwards and left-tangling upwards as 90°, 180° and 270° CCW rotations of bottom-tangling leftwards. 3 2-Sided Constructions for W-Triangulations We first show the following lemma, which is the key result for Theorem 2, and will also be used as an ingredient for the proof of Theorem 1. The reader is also referred to the appendix, where we give an example of a (3-sided) construction for a graph, which in the recursive cases uses some of the cases of the proof of Lemma 11. I Lemma 11. Let G be a W-triangulation. Let A, B, C be any three corners with respect to which G satisfies the chord condition, and let F be a set of at most one outer-face edge incident to C. Then G has an (int ∪ F ) 1-string B2 -VPG representation with 2-sided layout T. Biedl and M. Derka 7 a b b a c c c c a b a b a b b c c c a a c b a b Figure 3 The chair-shaped private region of a triangle a, b, c with possible rotations and flips. Note that labels of a, b, c can be arbitrarily permuted—the curve intersecting the “base” of the does not need to be named c. ... s1 ... sk s1 sk Figure 4 Bottom-tangling from s1 to sk rightwards. with respect to A, B. Furthermore, this representation has a chair-shaped private region for every interior face of G. We prove Lemma 11 by induction on the number of vertices. First, let us make an observation that will greatly help to reduce the number of cases of the induction step. Define G rev to be the graph obtained from graph G by reversing the combinatorial embedding, but keeping the same outer-face. This effectively switches corners A and B, and replaces special edge (C, c2 ) by (C, bs−1 ) and vice versa. If G satisfies the chord condition with respect to corners (A, B, C), then G rev satisfies the chord condition with respect to corners (B, A, C). (With this new order, the corners are CCW on the outer-face of G rev , as required.) Presume we have a 2-sided representation of G rev . Then we can obtain a 2-sided representation of G by flipping the 2-sided one of G rev horizontally, i.e., along the y-axis. Hence for all the following cases, we may (after possibly applying the above flipping operation) make a restriction on which edge the special edge is. Now we begin the induction. In the base case, n = 3, so G is a triangle, and the three corners A, B, C must be the three vertices of this triangle. The desired (int ∪ F ) representations for all possible choices of F are depicted in Figure 2. 8 1-String B2 -VPG Representation of Planar Graphs The induction step for n ≥ 4 is divided into three cases which we describe in separate subsections. 3.1 C has degree 2 Since G is a triangulated disk with n ≥ 4, (bs−1 , c2 ) is an edge. Define G0 := G − {C} and F 0 := {(bs−1 , c2 )}. We claim that G0 satisfies the chord condition for corners A0 := A, B 0 := B and a suitable choice of C 0 ∈ {bs−1 , c2 }, and argue this as follows. If c2 = A or c2 is incident to a chord that ends on PBC other than (bs−1 , c2 ) (thus bs−1 6= B), then set C 0 := bs−1 . The chord condition holds for G0 as bs−1 cannot be incident to a chord by planarity and the chord condition for G. Otherwise, c2 is not incident to a chord that ends in an interior vertex of PBC other than bs−1 , so set C 0 := c2 ; clearly the chord condition holds for G0 . Thus in either case, we can apply induction to G0 . To create a 2-sided representation of G, we use a 2-sided (int ∪ F 0 ) representation R0 of G0 constructed with respect to the aforementioned corners. We introduce a new vertical curve C placed between bs−1 and c2 below R0 . Add a bend at the upper end of C and extend it leftwards or rightwards. If the special edge e exists, then extend C until it hits the curve of the other endpoint of e; else extend it only far enough to allow for the creation of the private region. See also Figure 5. B bs-1 G' R' B bs-1 C A c2 C R' c2 A B bs-1 C R' c2 A B bs-1 C c2 A Figure 5 Case 1: 2-sided construction if C has degree 2 and (left) F = {}, (middle) F = {(C, c2 )} and (right) F = {(bs−1 , C)}. 3.2 G has a chord incident to C We may (after applying the reversal trick) assume that the special edge, if it exists, is (C, bs−1 ). By the chord condition, the chord incident to C has the form (C, ai ) for some 1 < i < r. The graph G can be split along the chord (C, ai ) into two graphs G1 and G2 . Both G1 and G2 are bounded by simple cycles, hence they are triangulated disks. No edges were added, so neither G1 nor G2 contains a separating triangle. So both of them are W-triangulations. T. Biedl and M. Derka 9 ai R2 B bs-1 G2 G1 C B ai bs-1 C ai A R1 C A Figure 6 Case 2(a): Constructing an (int ∪ (C, bs−1 )) representation when C is incident to a chord, in 2-sided (middle) and 3-sided (right) layout. We select (C, A, ai ) as corners for G1 and (ai , B, C) as corners for G2 and can easily verify that G1 and G2 satisfy the chord condition with respect to those corners: G1 has no chords on PAai or PCA as they would violate the chord condition in G. There is no chord on Pai C as it is a single edge. G2 has no chords on Pai B or PBC as they would violate the chord condition in G. There is no chord on Pai C as it is a single edge. Inductively construct a 2-sided (int ∪ (C, ai )) representation R1 of G1 and a 2-sided (int ∪ F ) representation R2 of G2 , both with the aforementioned corners. Note that CR2 and aiR2 are on the bottom side of R2 with CR2 to the left of aiR2 . Rotate R1 by 180°, and translate it so that it is below R2 with aiR1 in the same column as aiR2 . Stretch R1 and R2 horizontally as needed until CR1 is in the same column as CR2 . Then aiR and CR for R ∈ {R1 , R2 } can each be unified without adding bends by adding vertical segments. The curves of outer-face vertices of G then cross (after suitable lengthening) the bounding box in the required order. See also Figure 6. Every interior face f of G is contained in G1 or G2 and hence has a private region in R1 or R2 . As our construction does not make any changes inside the bounding boxes of R1 and R2 , the private region of f is contained in R as well. 3.3 G has no chords incident to C and deg(C) ≥ 3 We may (after applying the reversal trick) assume that the special edge, if it exists, is (C, c2 ). In this case we split G in a more complicated fashion illustrated in Figure 7. Let u1 , . . . , uq be the neighbours of vertex C in clockwise order, starting with bs−1 = u1 and ending with c2 = uq . We know that q = deg(C) ≥ 3 and that u2 , . . . , uq−1 are not on the outer-face, since C is not incident to a chord. Let uj be a neighbour of C that has at least one neighbour other than C on PCA , and among all those, choose j to be minimal. Such a j exists because G is a triangulated disk and therefore uq−1 is adjacent to both C and uq . We distinguish two sub-cases. Case 3(a): j 6= 1. Denote the neighbours of uj on Pc2 A by t1 , . . . , tx in the order in which they appear on Pc2 A . Separate G into subgraphs as follows (see also Figure 7): The right graph GR is bounded by (A, P.AB . . , B, P.Bu . .1 , u1 , u2 , . . . , uj , tx , P.t.x.A , A). 10 1-String B2 -VPG Representation of Planar Graphs B B GR bs-1 = u1 u2 bs-1 = u1 uj u2 GR uj=uq-1 GQ G0 C uq = c2 G1 t1 G3 G2 t2 t3 GQ G2 G1 t4=tx A C uq = c2 = t1 t2 G3 t3 t4=tx A Figure 7 Case 3(a): Splitting the graph when deg(C) ≥ 3, no chord is incident to C, and j > 1. (Left) j < q − 1; G0 is non-trivial. (Right) j = q − 1; G0 = {c2 }. Let GB be the graph bounded by (uj , t1 , P.t1. t.x , tx , uj ). We are chiefly interested in its subgraph GQ := GB − uj . Let GL be the graph bounded by (C, P.Ct . .1 , t1 , uj , C). We are chiefly interested in its subgraph G0 := GL − {uj , C}. The idea is to obtain representations of these subgraphs and then to combine them suitably. We first explain how to obtain the representation RR used for GR . Clearly GR is a W-triangulation, since u2 , . . . , uj are interior vertices of G, and hence the outer-face of GR is a simple cycle. Set AR := A and BR := B. If B 6= u1 then set CR := u1 and observe that GR satisfies the chord condition with respect to these corners: GR does not have any chords with both ends on PAR BR = PAB , PBR u1 ⊆ PBC , or Ptx AR ⊆ PCA since G satisfies the chord condition. If there were any chords between a vertex in u1 , . . . , uj and a vertex on PCR AR , then by CR = u1 the chord would either connect two neighbours of C (hence giving a separating triangle of G), or connect some ui for i < j to PCA (contradicting the minimality of j), or connect uj to some other vertex on Ptx A (contradicting that tx is the last neighbour of uj on PCA ). Hence no such chord can exist either. If B = u1 , then set CR := u2 (which exists by q ≥ 3) and similarly verify that it satisfies the chord condition as PBR CR is the edge (B, u2 ). Since CR ∈ {u1 , u2 } in both cases, we can apply induction on GR and obtain a 2-sided (int ∪ (u1 , u2 )) representation RR with respect to the aforementioned corners. Next we obtain a representation for the graph G0 , which is bounded by uj+1 , . . . , uq , Pc2 t1 and the neighbours of uj in CCW order between t1 and uj+1 . We distinguish two cases: (1) j = q − 1, and hence t1 = uq = c2 and G0 consists of only c2 . In this case, the representation of R0 consists of a single vertical line segment c2 . (2) j < q − 1, so G0 contains at least three vertices uq−1 , uq and t1 . Then G0 is a Wtriangulation since C is not incident to a chord and by the choice of t1 . Also, it satisfies the chord condition with respect to corners A0 := c2 , B0 := t1 and C0 := uj+1 since the three paths on its outer-face are sub-paths of PCA or contained in the neighbourhood of C or uj . In this case, construct a 2-sided (int ∪ (uj+1 , uj+2 )) representation R0 of G0 with respect to these corners inductively. Finally, we create a representation RQ of GQ . If GQ is a single vertex or a single edge, then simply use vertical segments for the curves of its vertices (recall that there is no special edge here). Otherwise, we can show: T. Biedl and M. Derka 11 I Claim 12. GQ has a 2-sided (int ∪ ∅) 1-string B2 -VPG representation with respect to corners t1 and tx . Proof. GQ is not necessarily 2-connected, so we cannot apply induction directly. Instead we break it into x − 1 graphs G1 , . . . , Gx−1 , where for i = 1, . . . , x − 1 graph Gi is bounded by Pti ti+1 as well as the neighbours of uj between ti and ti+1 in CCW order. Note that Gi is either a single edge, or it is bounded by a simple cycle since uj has no neighbours on PCA between ti and ti+1 . In the latter case, use Bi := ti , Ai := ti+1 , and Ci an arbitrary third vertex on Pti ti+1 ⊆ PCA , which exists since the outer-face of Gi is a simple cycle and (ti , ti+1 , uj ) is not a separating triangle. Observe that Gi satisfies the chord condition since all paths on the outer-face of Gi are either part of PCA or in the neighbourhood of uj . Hence by induction there exists a 2-sided (int ∪ ∅) representation Ri of Gi with respect to the corners of Gi . If Gi is a single edge (ti , ti+1 ), then let Ri consists of two vertical segments ti and ti+1 . Since each representation Ri has at its leftmost end a vertical segment ti and at its rightmost end a vertical segment ti+1 , we can combine all these representations by aligning Ri+1 i tR horizontally and filling in the missing segment. See also Figure 8. One easily i and ti verifies that the result is a 2-sided (int ∪ ∅) representation of GQ . J G1 uj G2 (empty) → t1 G1 G2 t2 G3 t3 t4 = tx G3 Θ t1 t2 t3 t4 = t x Figure 8 Left: Graph GB . The boundary of GQ is shown bold. Right: Merging 2-sided (int ∪ ∅) representations of Gi , 1 ≤ i ≤ 3, into a 2-sided (int ∪ ∅) representation of GQ . We now explain how to combine these three representations RR , RQ and R0 ; see also R R Figure 9. Translate RQ so that it is below RR with tR and tx Q in the same column; then x connect these two curves with a vertical segment. Rotate R0 by 180° and translate it so RQ 0 that it is below RR and to the left and above RQ , and tR are in the same column; 1 and t1 then connect these two curves with a vertical segment. Notice that the vertical segments of u2RR , . . . , ujRR are at the bottom left of RR . Horizontally stretch R0 and/or RR so that R0 u2RR , . . . , ujRR are to the left of the vertical segment of uj+1 , but to the right (if j < q − 1) R0 of the vertical segment of uj+2 . There are such segments by j > 1. Introduce a new horizontal segment C and place it so that it intersects curves uq , . . . , uj+2 , u2 , . . . , uj , uj+1 (after lengthening them, if needed). Attach a vertical segment to C. If j < q − 1, then top-tangle uq , . . . , uj+2 rightwards. (Recall from Section 2.3 that this creates intersections among all these curves.) Bottom-tangle u2 , . . . , uj rightwards. The construction hence creates intersections for all edges in the path u1 , . . . , uq , except for (uj+2 , uj+1 ) (which was represented in R0 ) and (u2 , u1 ) (which was represented in RR ). Bend and stretch ujRR rightwards so that it crosses the curves of all its neighbours in G0 ∪ GQ . Finally, consider the path between the neighbours of uj CCW from uj+1 to tx . Top-tangle curves of these vertices rightwards, but omit the intersection if the edge is on the outer-face (see e.g. (t2 , t3 ) in Figure 9). B B RR RR bs-1 bs-1 12 u uj uj 1-Stringu2B2 -VPG Representation ofA Planar Graphs 2 A uj+2 uj+2 C One verifies that the curves intersect the bounding boxes as desired. The constructed representations contain private regions for all interior faces of GR , GQ and G0 by induction. uj+1 uj+1 are of the form (C, u , u 0 The remaining and (uj , wk , wk+1 ) where wk and R0 faces i i+1 ), 1 ≤ i < q, R u =c q 2 uq=c 2 wk+1 are two consecutive neighbours of uj on the outer-face of G0 or GQ . Private regions RQ RQ Figure 9. for those faces are shown in C t1 t2 t3 t1 t4 = t x RR u2 B bs-1 t2 t3 t4 = t x RR uj u2 B bs-1 A uj uj+2 uj+2 R0 ... uj+1 R0 uq=c2 uj+1 uq=c2 RQ C t1 t3 t2 RQ C t4 = t x t1 Figure 9 Combining subgraphs in Case 3(a). 2-sided construction, for F = {(C, c2 )} and F = ∅. The construction matches the graph depicted in Figure 7 left. Case 3(b): j = 1, i.e., there exists a chord (bs−1 , ci ). In this case we cannot use the above construction directly since we need to bend uj = u1 = bs−1 horizontally rightwards to create intersections, but then it no longer extends vertically downwards as required for bs−1 . Instead we use a different construction. Edge (bs−1 , ci ) is a chord from PBC to PCA . Let (bk , c` ) be a chord from PBC to PCA that maximizes k − `, i.e., is furthest from C (our construction in this case actually works for any chord from PBC to PCA —it is not necessary that k = s − 1). Note that possibly ` = t (i.e., the chord is incident to A) or k = 1 (i.e., the chord is incident to B), but not both by the chord condition. We assume here that ` < t, the other case is symmetric. B G2 B bk C → cl A bk bk G1 C cl G2 cl G1 A B bk C c2 cl A Figure 10 Case 3(b): Construction of a 2-sided (int ∪ (C, c2 )) representation of G with a chord (bk , c` ). In order to construct a 2-sided (int ∪ F ) representation of G, split the graph along (bk , c` ) into two W-triangulations G1 (which includes C and the special edge, if any) and G2 (which includes A). Set (A, B, c` ) as corners for G1 (these are three distinct vertices by c` 6= A) and set (c` , bk , C) as corners for G2 and verify the chord condition: G1 has no chords on either PCc` ⊆ PCA or Pbk C ⊆ PBC as they would contradict the T. Biedl and M. Derka 13 chord condition in G. The third side is a single edge (bk , c` ) and so it does not have any chords either. G2 has no chords on either Pc` A ⊆ PCA or PAB as they would violate the chord condition in G. It does not have any chords on the path PBc` due to the selection of the chord (bk , c` ) and by the chord condition in G and by the chord condition in G. Thus, by induction, G1 has a 2-sided (int ∪ F ) representation R1 and G2 has a 2-sided (int ∪ (bk , c` )) representation R2 with respect to the aforementioned corners. Translate and 1 1 2 2 horizontally stretch R1 and/or R2 so that bR and cR are aligned with bR and cR ` ` , k k R1 R1 respectively, and connect each pair of curves with a vertical segment. Since bk and c` have no bends, this does not increase the number of bends on any curve and produces a 2-sided (int ∪ F ) representation of G. All the faces in G have a private region inside one of the representations of G1 or G2 . This ends the description of the construction in all cases, and hence proves Lemma 11. We now show how Lemma 11 implies Theorem 2: Proof of Theorem 2. Let G be a 4-connected planar graph. Assume first that G is triangulated, which means that it is a W -triangulation. Let (A, B, C) be the outer-face vertices and start with an (int ∪ (B, C))-representation of G (with respect to corners (A, B, C)) that exists by Lemma 11. The intersections of the other two outer-face edges (A, C) and (A, B) can be created by tangling B, A and C, A suitably (see Figure 11). Theorem 2 also stipulates that every curve used in a representation has at most one vertical segment. This is true for all curves added during the construction. Furthermore, we join two copies of a curve only by aligning and connecting their vertical ends, so all curves have at most one vertical segment. This proves Theorem 2 for 4-connected triangulations. To handle an arbitrary 4-connected planar graph, stellate the graph, i.e., insert into each non-triangular face f a new vertex v and connect it to all vertices on f . By 4-connectivity this creates no separating triangle and the graph is triangulated afterwards. Finding a representation of the resulting graph and deleting the curves of all added vertices yields the result. J Θ B C A Figure 11 Completing a 2-sided (int ∪ (B, C)) representation by adding intersections for (A, B) and (A, C). 4 3-Sided Constructions for W-Triangulations Our key tool for proving Theorem 1 is the following lemma: I Lemma 13. Let G be a W-triangulation and let A, B, C be any three corners with respect to which G satisfies the chord condition. For any e ∈ {(C, bs−1 ), (C, c2 )}, G has an (int ∪ e) 14 1-String B2 -VPG Representation of Planar Graphs 1-string B2 -VPG representation with 3-sided layout and an (int ∪ e) 1-string B2 -VPG representation with reverse 3-sided layout. Both representations have a chair-shaped private region for every interior face. The proof of Lemma 13 will use induction on the number of vertices. To combine the representations of subgraphs, we sometimes need them to have a 2-sided layout, and hence we frequently use Lemma 11 proved in Section 3. Also, notice that for Lemma 13 the special edge must exist (this is needed in Case 1 to find private regions), while for Lemma 11, F is allowed to be empty. We again reduce the number of cases in the proof of Lemma 13 by using the reversal trick. Define G rev as in Section 3. Presume we have a 3-sided/reverse 3-sided representation of G rev . We can obtain a 3-sided/reverse 3-sided representation of G by flipping the reverse 3-sided/3-sided representation of G rev diagonally (i.e., along the line defined by (x = y)). Again, this effectively switches corners A and B (corner C remains the same), and replaces special edge (C, c2 ) by (C, bs−1 ) and vice versa. If G satisfies the chord condition with respect to corners (A, B, C), then G rev satisfies the chord condition with respect to corners (B, A, C). Hence for all the following cases, we may again (after possibly applying the above flipping operation) make a restriction on which edge the special edge is. Alternatively, we only need to give the construction for the 3-sided, but not for the reverse 3-sided layout. So let G and a special edge e be given, and set F = {e}. In the base case, n = 3, so G is a triangle, and the three corners A, B, C must be the three vertices of this triangle. The desired (int ∪ F ) representations for all possible choices of F are depicted in Figure 2. The induction step for n ≥ 4 uses the same case distinctions as the proof of Lemma 11; we describe these cases in separate subsections. 4.1 C has degree 2 Since G is a triangulated disk with n ≥ 4, (bs−1 , c2 ) is an edge. Define G0 as in Section 3.1 to be G − {C} and recall that G0 satisfies the chord condition for corners A0 := A, B 0 := B and a suitable choice of C 0 ∈ {bs−1 , c2 } Thus, we can apply induction to G0 . To create a 3-sided representation of G, we use a 3-sided (int ∪ F 0 ) representation R0 of G0 , where F 0 = {(bs−1 , c2 )}. Note that regardless of which vertex is C 0 , we have bs−1 as bottommost curve on the left and c2 as leftmost curve on the bottom. Introduce a new horizontal segment representing C which intersects c2 if F = {(C, c2 )}, or a vertical segment which intersects bs−1 if F = {(C, bs−1 )}. After suitable lengthening, the curves intersect the bounding box in the required order. One can find the chair-shaped private region for the only new face {C, c2 , bs−1 } as shown in Figure 12. Observe that no bends were added to the curves of R0 and that C has the required number of bends. Since we have given the constructions for both possible special edges, we can obtain the reverse 3-sided representation by diagonally flipping a 3-sided representation of G rev . 4.2 G has a chord incident to C Let (C, ai ) be a chord that minimizes i (i.e., is closest to A). Define W-triangulations G1 and G2 with corners (C, A, ai ) for G1 and (ai , B, C) for G2 as in Section 3.2, and recall that they satisfy the chord condition. So, we can apply induction to both G1 and G2 , obtain representations R1 and R2 (with respect to the aforementioned corners) for them, and combine them suitably. We will do so for both possible choices of special edge, and hence need not give the constructions for reverse 3-sided layout due to the reversal trick. T. Biedl and M. Derka 15 B B bs-1 R' bs-1 G' R' bs-1 A c2 C B C c2 C A c2 A Figure 12 Case 1: 3-sided representation if C has degree 2. Using Lemma 11, construct a 2-sided (intR' ∪ (C, ai )) Case 2(a):R' F = {(C, bs−1 )}. R' representation R1 of G1 with respect to the aforementioned corners of G1 . Inductively, construct a 3-sided (int ∪ F ) representation R2 of G2 with respect to the corners of G2 . Note that CR2 and aiR2 are on the bottom side of R2 with CR2 to the left of aiR2 . First,b rotate R by 180°. We can now merge R and A R2 as describedBin Section 3.1 since bs-1 C 1 c2 bs-1 C B B A c2 1 A c2 s-1 C all relevant curves end vertically in R1 and R2 . The curves of outer-face vertices of G then cross (after suitable lengthening) the bounding box in the required order. See also Figure 13. ai B B bs-1 C bs-1 C ai G2 G1 R2 ai A R1 C A Figure 13 Case 2(a): Constructing a 3-sided (int ∪ (C, bs−1 )) representation when C is incident to a chord. Case 2(b): F = {(C, c2 )}. For the 3-sided construction, it does not seem possible to merge suitable representations of G1 and G2 directly, since the geometric restrictions imposed onto curves A, B, C, c2 and ai by the 3-sided layout cannot be satisfied using 3-sided and 2-sided representations of G1 and G2 . We hence use an entirely different approach that splits the graph further; it resembles Case 1 in [1, Proof of Lemma 2]. Let GQ = G1 − C, and observe that it is bounded by Pc2 A , PA,ai , and the path formed by the neighbours c2 = u1 , u2 , . . . , uq = ai of C in G1 in CCW order. We must have q ≥ 2, but possibly G1 is a triangle {C, A, ai } and GQ then degenerates into an edge. If GQ contains at least three vertices, then u2 , . . . , uq−1 are interior since chord (C, ai ) was chosen closest to A, and so GQ is a W-triangulation. We divide the proof into two subcases, depending on whether A 6= c2 or A = c2 . See also Figures 14 and 15. Case 2(b)1: A 6= c2 . Select the corners of GQ as (AQ := c2 , BQ := A, CQ := ai = uq ), and observe that it satisfies the chord condition since the three corners are distinct and the three 16 1-String B2 -VPG Representation of Planar Graphs outer-face paths are sub-paths of PCA and PAB or in the neighbourhood of C, respectively. Apply Lemma 11 to construct a 2-sided (int ∪ (uq , uq−1 )) representation RQ of GQ with respect to the corners of GQ . Inductively, construct a 3-sided (int ∪ (C, ai )) representation R2 of G2 with respect to the corners of G2 . To combine RQ with R2 , rotate RQ by 180°. Appropriately stretch RQ and translate it so R that it is below R2 with ai Q and aiR2 in the same column, and so that the vertical segment of each of the curves uq−1 , . . . , u1 = c2 is to the left of the bounding box of R2 . Then R ai Q and aiR2 can be unified without adding bends by adding a vertical segment. Curves uq−1 , . . . , u1 = c2 in the rotated RQ can be appropriately stretched upwards, intersected by CR2 after stretching it leftwards, and then top-tangled leftwards. All the curves of outer-face vertices of G then cross (after suitable lengthening) a bounding box in the required order. All faces in G that are not interior to GQ or G2 are bounded by (C, uk , uk+1 ), 1 ≤ k < q. The chair-shaped private regions for such faces can be found as shown in Figure 14. Case 2(b)2: A = c2 . In this case the previous construction cannot be applied since the corners for GQ would not be distinct. We give an entirely different construction. If GQ has at least 3 vertices, then q ≥ 3 since otherwise by A = c2 = u1 edge (A, uq ) would be a chord on PAB . Choose as corners for GQ the vertices AQ := A, BQ := ai = uq and CQ := uq−1 and observe that the chord condition holds since all three paths on the outer-face belong to PAB or are in the neighbourhood of C. By Lemma 11, GQ has a 2-sided (int ∪ (uq , uq−1 )) representation RQ with the respective corners and private region for every interior face of GQ . If GQ has at most 2 vertices, then GQ consists of edge (A, a2 ) only, and we use as representation R2 two parallel vertical segments a2 and A. We combine RQ with a representation R1 of G1 that is different from the one used in the previous cases; in particular we rotate corners. Construct a reverse 3-sided layout R2 of G2 with respect to corners C2 := ai , A2 := B and B2 := C. Rotate R2 by 180°, and translate R it so that it is situated below RQ with ai Q and aiR2 in the same column. Then, extend R R Q Q CR2 until it crosses uq−1 , . . . , u1 (after suitable lengthening), and then bottom-tangle R R Q uq−1 , . . . , u1 Q rightwards. This creates intersections for all edges in path uq , uq−1 , . . . , u1 , except for (uq , uq−1 ), which is either on the outer-face (if q = 2) or had an intersection in RQ . One easily verifies that the result is a 3-sided layout, and private regions can be found for the new interior faces as shown in Figure 15. B B R2 G2 C ai = u5 = uq u4 u3 u2 GQ c2=u1 C ai = u5 = uq A u2 u3 u4 RQ c2 = u1 Figure 14 Case 2(b)1: C is incident to a chord, F = {(C, c2 )}, and c2 6= A. A T. Biedl and M. Derka 17 u4 B B ai = u5 = uq G2 u4 u3 G u2 C RQ u3 u2 ai = u 5 = u q A = c2 = u1 R2 Q A = c2 = u 1 C B B A = c2 = u1 a2 = u2 = uq G2 R2 GQ C a2 = u 2 = u q A = c2 = u 1 C Figure 15 Case 2(b)2: Construction when C is incident to a chord, c2 = A, F = {(C, c2 )} and (A, ai , C) is not a face (top), or when (A, ai , C) is a face (bottom). 4.3 G has no chords incident to C and deg(C) ≥ 3 We will give explicit constructions for 3-sided and reverse 3-sided layout, and may hence (after applying the reversal trick) assume that the special edge is (C, c2 ). As in Section 3.3, let u1 , . . . , uq be the neighbours of C and let j be minimal such that uj has another neighbour on PAC . We again distinguish two sub-cases. Case 3(a): j 6= 1. As in Section 3.3, define t1 , . . . , tx , GR , GB , GQ , GL and G0 . See also Figure 7. Recall that GR satisfies all conditions with respect to corners AR := A, BR := B and CR ∈ {u1 , u2 }. Apply induction on GR and obtain an (int ∪ (u1 , u2 )) representation RR with respect to the corners of GR . We use as layout for RR the type that we want for G, i.e., use a 3-sided/reverse 3-sided layout if we want G to have a 3-sided/reverse 3-sided representation. For G0 and GQ , we use exactly the same representations R0 and RQ as in Section 3.3. Combine now these three representations RR , RQ and R0 as described in Section 3.3, Case 3(a); this can be done since the relevant curves u2RR , . . . , utRR all end vertically in RR . See also Figure 16. The only change occurs at curve C; in Section 3.3 this received a bend and a downward segment, but here we omit this bend and segment and let C end horizontally as desired. One easily verifies that the curves intersect the bounding boxes as desired. The constructed representations contain private regions for all interior faces of GR , GQ and G0 by induction. The remaining faces are of the form (C, ui , ui+1 ), 1 ≤ i < q, and (uj , wk , wk+1 ) where wk and wk+1 are two consecutive neighbours of uj on the outer-face of G0 or GQ . Private regions for those faces are shown in Figure 16. Case 3(b): j = 1, i.e., there exists a chord (bs−1 , ci ). In this case we cannot use the above construction directly since we need to bend uj = u1 = bs−1 horizontally rightwards to create intersections, but then it no longer extends vertically downwards as required for bs−1 . The simple construction described in Section 3.3, Case 3(b) does not apply either. However, if we use a different vertex as uj (and argue carefully that the chord condition holds), then 18 1-String B2 -VPG Representation of Planar Graphs B B RR RR bs-1 bs-1 u2 uj u2 A uj A uj+2 uj+2 C C R0 uj+1 uj+1 R0 uq=c2 uq=c2 RQ RQ t1 t2 t3 t1 t4 = t x t2 t3 t4 = t x Figure 16 Case 3(a): 3-sided representation when deg(C) ≥ 3, there is no chord incident to C, RR RR 7 left. F = {(C, c2 )}, and j > 1. The construction matches the graph depicted in Figure u2 B bs-1 uj uj+2 uj uj+2 B R0 uq=c2 u2 B bs-1 A the same construction works. ... GR uj+1 u2 bs-1 = u1 RQ C t1 R0 uj' t2 t3 C t4 = t x G0 C uj+1 uq=c2 uq = c2 GQ G1 t1 G3 G2 t2 RQ t1 t3 t4=tx A Figure 17 Case 3(b): Splitting the graph when deg(C) ≥ 3, no chord is incident to C, and j = 1. Recall that u1 , . . . , uq are the neighbours of corner C in CW order starting with bs−1 and ending with c2 . We know that q ≥ 3 and u2 , . . . , uq−1 are not on the outer-face. Now define j 0 as follows: Let uj 0 , j 0 > 1 be a neighbour of C that has at least one neighbour on PCA other than C, and choose uj 0 so that j 0 is minimal while satisfying j 0 > 1. Such a j 0 exists since uq−1 has another neighbour on PCA , and by q ≥ 3 we have q − 1 > 1. Now, separate G as in the previous case, except use j 0 in place of j. Thus, define t1 , . . . , tx to be the neighbours of uj 0 on Pc2 A , in order, and separate G into three graphs as follows: The right graph GR is bounded by (A, P.AB . . , B, P.Bu . .1 , u1 , u2 , . . . , uj 0 , tx , P.t.x.A , A). Let GB be the graph bounded by (uj 0 , t1 , P.t1. t.x , tx , uj 0 ). Define GQ := GB − uj 0 . Let GL be the graph bounded by (C, P.Ct . .1 , t1 , uj 0 , C). Define G0 := GL − {uj 0 , C}. Observe that the boundaries of all the graphs are simple cycles, and thus they are W-triangulations. Select (AR := A, BR := B, CR := u2 ) to be the corners of GR and argue the chord condition as follows: GR does not have any chords on PCR AR as such chords would either contradict the minimality of j 0 , or violate the chord condition in G. GR does not have any chords on PAR BR = PAB . T. Biedl and M. Derka 19 GR does not have any chords on PBbs−1 as it is a sub-path of PBC and they would violate the chord condition in G. It also does not have any chords in the form (CR = u2 , b` ), 1 ≤ ` < s − 1 as they would have to intersect the chord (bs−1 , ci ), violating the planarity of G. Hence, GR does not have any chords on PCR AR . Notice in particular that the chord (u1 , ci ) of GR is not a violation of the chord condition since we chose u2 as a corner. Hence, we can obtain a representation RR of GR with 3-sided or reverse 3-sided layout and special edge (u1 = bs−1 , u2 ). For graphs GQ and G0 the corners are chosen, the chord condition is verified, and the representations are obtained exactly as in Case 3(a). Since the special edge of GR is (u1 , u2 ) as before, curves u1 and u2 are situated precisely as in Case 3(a), and we merge representations and find private regions as before. This ends the description of the construction in all cases, and hence proves Lemma 13. 5 From 4-Connected Triangulations to All Planar Graphs In this section, we prove Theorem 1. Observe that Lemma 13 essentially proves it for 4-connected triangulations. As in [4] we extend it to all triangulations by induction on the number of separating triangles. B C Θ A Figure 18 Completing a 3-sided (int ∪ (B, C)) representation by adding intersections for (A, B) and (A, C). I Theorem 14. Let G be a triangulation with outer-face (A, B, C). G has a 1-string B2 -VPG representation with a chair-shaped private region for every interior face f of G. Proof. Our approach is exactly the same as in [4], except that we must be careful not to add too many bends when merging subgraphs at separating triangles, and hence must use 3-sided layouts. Formally, we proceed by induction on the number of separating triangles. In the base case, G has no separating triangle, i.e., it is 4-connected. As the outer-face is a triangle, G clearly satisfies the chord condition. Thus, by Lemma 13, it has a 3-sided (int ∪ (B, C)) representation R with private region for every face. R has an intersection for every edge except for (A, B) and (A, C). These intersections can be created by tangling B, A and C, A suitably (see Figure 18). Recall that A initially did not have any bends, so it has 2 bends in the constructed representation of G. The existence of private regions is guaranteed by Lemma 13. Now assume for induction that G has k + 1 separating triangles. Let ∆ = (a, b, c) be an inclusion-wise minimal separating triangle of G. It was shown in [4] that the subgraph G2 induced by the vertices inside ∆ is either an isolated vertex, or a W-triangulation with corners (A, B, C) such that the vertices on PAB are adjacent to b, the vertices on PBC are adjacent to c, and the vertices on PCA are adjacent to a. Furthermore, G2 satisfies the chord 20 1-String B2 -VPG Representation of Planar Graphs condition. Also, graph G1 = G − G2 is a W-triangulation that satisfies the chord condition and has k separating triangles. By induction, G1 has a representation R1 (with respect to the corners of G1 ) with a chair-shaped private region for every interior face f . Let Φ be the private region for face ∆. Permute a, b, c, if needed, so that the naming corresponds to the one needed for the private region and, in particular, the vertical segment of c intersects the private region of ∆ as depicted in Figure 19. Case 1: G2 is a single vertex v. Represent v by inserting into Φ an orthogonal curve v with 2 bends that intersects a, b and c. The construction, together with private regions for the newly created faces (a, b, v), (a, c, v) and (b, c, v), is shown in Figure 19. Case 2: G2 is a W-triangulation. Recall that G2 satisfies the chord condition with respect to corners (A, B, C). Apply Lemma 13 to construct a 3-sided (int ∪ (C, bs−1 )) representation R2 of G2 with respect to the corners of G2 . Let us assume that (after possible rotation) Φ has the orientation shown in Figure 19 (right); if it had the symmetric orientation then we would do a similar construction using a reverse 3-sided representation of G2 . Place R2 inside Φ as shown in Figure 19 (right). Stretch the curves representing vertices on PCA , PAB and PBbs−1 downwards, upwards and leftwards respectively so that they intersect a, b and c. Top-tangle leftwards the curves A = a1 , a2 , . . . , ar = B. Left-tangle downwards the curves B = b1 , b2 , . . . , bs−1 and bend and stretch C downwards so that it intersects a. Bottom-tangle leftwards the curves C = c1 , . . . , ct = A. It is easy to verify that the construction creates intersections for all the edges between vertices of ∆ and the outer-face of G2 . The tangling operation then creates intersections for all the outer-face edges of G2 except edge (C, bs−1 ), which is already represented in R2 . Every curve that receives a new bend represents a vertex on the outer-face of G2 , which means that it initially had at most 1 bend. Curve A is the only curve that receives 2 new bends, but this is allowed as A does not have any bends in R2 . Hence, the number of bends for every curve does not exceed 2. Private regions for faces formed by vertices a, b, c and vertices on the outer-face of G2 can be found as shown in Figure 19 right. J With Theorem 14 in hand, we can show our main result: every planar graph has a 1-string B2 -VPG representation. Proof of Theorem 1. If G is a planar triangulated graph, then the claim holds by Theorem 14. To handle an arbitrary planar graph, repeatedly stellate the graph (recall that this means inserting into each non-triangular face a new vertex connected to all vertices of the face). It is easily shown that one stellation makes the graph connected, a second one makes it 2-connected, and a third one makes it 3-connected and triangulated. Thus after 3 stellations we have a 3-connected triangulated graph G0 such that G is an induced subgraph of G0 . Apply Theorem 14 to construct a 1-string B2 -VPG representation R0 of G0 (with the three outer-face vertices chosen as corners). By removing curves representing vertices that are not in G, we obtain a 1-string B2 -VPG representation of G. J 6 Conclusions and Outlook We showed that every planar graph has a 1-string B2 -VPG representation, i.e., a representation as an intersection graph of strings where strings cross at most once and each string is orthogonal with at most two bends. One advantage of this is that the coordinates to describe such a representation are small, since orthogonal drawings can be deformed easily T. Biedl and M. Derka 21 R1 b c c v a v a b R1 b c A c C G2 B R2 C B a A a b Figure 19 A separating triangle enclosing one vertex and the construction (top), and a separating triangle enclosing a W-triangulation and the corresponding construction (bottom). such that all bends are at integer coordinates. Every vertex curve has at most two bends and hence at most 3 segments, so the representation can be made to have coordinates in an O(n) × O(n)-grid with perimeter at most 3n. Note that none of the previous results provided an intuition of the required size of the grid. Following the steps of our proof, it is not hard to see that our representation can be found in linear time, since the only non-local operation is to test whether a vertex has a neighbour on the outer-face. This can be tested by marking such neighbours whenever they become part of the outer-face. Since no vertex ever is removed from the outer-face (updating the outer-face markers upon removing such vertices could increase the time complexity), this takes overall linear time. The representation constructed in this paper uses curves of 8 possible shapes for planar graphs. For 4-connected planar graphs, the shapes that have at most one vertical segment suffice. A natural question is if one can restrict the number of shapes required to represent all planar graphs. Bringing this effort further, is it possible to restrict the curves even more? The existence of 1-string B1 -VPG representations for planar graphs is open. Furthermore, Felsner et al. [9] asked the question whether every planar graph is the intersection graph of only two 22 1-String B2 -VPG Representation of Planar Graphs shapes, namely {L, Γ}. As they point out, a positive result would provide a different proof of Scheinerman’s conjecture (see [11] for details). Somewhat inbetween: is every planar graph the intersection graph of xy-monotone orthogonal curves, preferably in the 1-string model and with few bends? References 1 2 3 4 5 6 7 8 9 10 11 12 13 A Takao Asano, Shunji Kikuchi, and Nobuji Saito. A linear algorithm for finding Hamiltonian cycles in 4-connected maximal planar graphs. Discr. Applied Mathematics, 7(1):1 – 15, 1984. Jérémie Chalopin and Daniel Gonçalves. Every planar graph is the intersection graph of segments in the plane: extended abstract. In ACM Symposium on Theory of Computing, STOC 2009, pages 631–638. ACM, 2009. Jérémie Chalopin, Daniel Gonçalves, and Pascal Ochem. Planar graphs are in 1-string. In ACM-SIAM Symposium on Discrete Algorithms, SODA ’07, pages 609–617. SIAM, 2007. Jérémie Chalopin, Daniel Gonçalves, and Pascal Ochem. Planar graphs have 1-string representations. Discrete & Computational Geometry, 43(3):626–647, 2010. Steven Chaplick and Torsten Ueckerdt. Planar graphs as VPG-graphs. J. Graph Algorithms Appl., 17(4):475–494, 2013. Natalia de Castro, Francisco Javier Cobos, Juan Carlos Dana, Alberto Márquez, and Marc Noy. Triangle-free planar graphs and segment intersection graphs. J. Graph Algorithms Appl., 6(1):7–26, 2002. Hubert de Fraysseix, Patrice Ossona de Mendez, and János Pach. Representation of planar graphs by segments. Intuitive Geometry, 63:109–117, 1991. Gideon Ehrlich, Shimon Even, and Robert Endre Tarjan. Intersection graphs of curves in the plane. J. Comb. Theory, Ser. B, 21(1):8–20, 1976. Stefan Felsner, Kolja B. Knauer, George B. Mertzios, and Torsten Ueckerdt. Intersection graphs of L-shapes and segments in the plane. In Mathematical Foundations of Computer Science (MFCS’14), Part II, volume 8635 of Lecture Notes in Computer Science, pages 299–310. Springer, 2014. Irith Ben-Arroyo Hartman, Ilan Newman, and Ran Ziv. On grid intersection graphs. Discrete Mathematics, 87(1):41–52, 1991. Matthias Middendorf and Frank Pfeiffer. The max clique problem in classes of string-graphs. Discrete Mathematics, 108(1-3):365–372, 1992. Edward R. Scheinerman. Intersection Classes and Multiple Intersection Parameters of Graphs. PhD thesis, Princeton University, 1984. Hassler Whitney. A theorem on graphs. The Annals of Mathematics, 32(2):387–390, 1931. Example Here we provide an example of constructing an (int ∪(18, 16)) 1-string B2 -VPG representation R of the W-triangulation shown in Figure 20. We use numbers and colors to distinguish vertices. We use letters to indicate special vertices such as corners; note that the designation as such a corner may change as the subgraph gets divided further. The special edge is marked with hatches. One can verify that the graph with the chosen corners (1,4,18) satisfies the chord condition. Vertex C has degree 3, but it is not incident to a chord, so one applies the construction from Section 4.3. Finding vertex uj , we can see that j > 1, so Case 3(a) applies. Figure 20 shows the graphs GR , GQ and G0 , and how to construct R from their representations RR , RQ and R0 . T. Biedl and M. Derka The construction of RQ is shown in Figure 21. The representation should have a 2-sided layout and no special edge. Graph GQ decomposes into three subgraphs G1 , G2 , G3 . Their 2-sided representations are found separately and combined as described in the proof of Claim 12. The construction of RR is shown in Figure 22 (decomposition of GR ) and 23 (combining the representations). Representation RR is supposed to be 3-sided. We first apply Case 1 (Section 4.1) twice, since corner C has degree 2. Then corner C becomes incident to a chord, so we are in Case 2, and use sub-case Case 2(a) (Section 4.2) since the special edge is (C, bs−1 = B). This case calls for a 3-sided representation of a G2 (which is a triangle in this case, so the base case applies). It also calls for a 2-sided representation of G1 with special edge (C, A = c2 ). This is Case 2 (Section 3.2) and we need to apply the reversal trick—we flip the graph and relabel the corners. After obtaining the representation, it must be flipped horizontally in order to undo the reversal. The construction decomposes the graph further, using Case 2 repeatedly, which breaks the graphs into elementary triangles. Their 2-sided representations are obtained using the base case and composed as stipulated by the construction. Figure 24 shows the complete 3-sided (int ∪ (18, 16)) representation of the graph. 23 24 1-String B2 -VPG Representation of Planar Graphs 4 5 3 6 17 14 18 16 2 15 13 12 11 10 7 8 9 1 B:=4 B u1=bs-1:=5 C := 5 u2=uj :=6 GR c2 := 6 C: c t =1 2 := 1 := 16 13 8 t2 : = t3 : t4 = = 11 10 tx := 7 A:=1 A:=1 t4=tx GQ B u1=bs-1 Case 3(a) A1=B2:=14 RR u2=uj B1:=13 C1=12 A3:=7 A2=B3:=10 A C R0 G0 c2 RQ t1 A0=16 t2 C0=17 B0=B1:=13 t3 t4=tx Figure 20 Illustration of the example. The goal is to find an (int ∪ (18, 16)) 1-string B2 -VPG representation of the W-triangulation shown on top, using corners (1,4,18). T. Biedl and M. Derka 25 A1 B1 t1=B0=B1 t2=A1=B2 A2 C1 t3=A2=B3 C1 A3 B2 A3=t4=tx B3 ai:=15 B1 B2:=11 A2:=10 B1:=13 C1:=12 B3:=10 A3:=7 ai A1:=11 B2 C:=15 A:=15 A1 C1 ai:=14 180° ai B A:=12 B:=11 B:=13 C:=12 C A:=14 B:=13 C:=12 B C A 180° B:=15 C:=14 A:=12 BC A B C Figure 21 Illustration of the example: Finding RQ (top right). A A A2 B3 A3 26 1-String B2 -VPG Representation of Planar Graphs B:=4 Case 3(a) C:=5 c2:=6 A:=1 Case 1, 3-sided B:=4 B:=4 Case 1, 3-sided C:=6 A:=1 C:=7 A:=1 c2:=8 Case 2, 3-sided B:=4 C:=3 C:=3 Case 2, 2-sided, reversing trick A:=1 ai:=8 B:=7 A:=3 A:=7 B:=1 C:=7 Case 2, 2-sided C:=3 A:=3 B:=3 Case 2, 2-sided, a :=2 ai:=2 reversing trick i B:=1 A:=1 C:=8 C:=8 A:=8 B:=7 C:=2 Case 2, 2-sided Case 2, 2-sided B:=3 C:=2 A:=2 Case 2, 2-sided, reversing trick A:=9 C:=2 B:=8 A:=2 C:=8 A:=8 ai:=9 B:=1 A:=1 ai:=9 B:=8 B:=1 Figure 22 Illustration of the example: Decomposing graph GR . C:=9 T. Biedl and M. Derka 27 B:=4 B Case 3(a) C:=5 c2:=6 A:=1 A C B B Case 1, 3-sided Case 1, 3-sided C A A C Case 2, 3-sided 180° B B Case 2, 2-sided, reversing trick C flip A A C B C Case 2, 2-sided 180° ai ai Case 2, 2-sided, reversing trick B C A A B B flip A C C Case 2, 2-sided Case 2(a), 2-sided 180° ai B C A B A Case 2, 2-sided, reversing trick B C B flip C ai C A A 180° Figure 23 Illustration of the example: Composing representation RR . B C A A 28 1-String B2 -VPG Representation of Planar Graphs 4 5 3 6 17 14 18 16 2 15 7 13 12 11 10 8 9 1 B RR u1=bs-1 u2=uj A C R0 RQ c2 t1 t2 t3 t4=tx Figure 24 Illustration of the example: Complete 3-sided (int ∪ (18, 16)) representation. 

arXiv:1803.00673v1 [math.CO] 2 Mar 2018 An efficient algorithm to test forcibly-connectedness of graphical degree sequences Kai Wang∗ March 5, 2018 Abstract We present an algorithm to test whether a given graphical degree sequence is forcibly connected or not and prove its correctness. We also outline the extensions of the algorithm to test whether a given graphical degree sequence is forcibly k-connected or not for every fixed k ≥ 2. We show through experimental evaluations that the algorithm is efficient on average, though its worst case run time is probably exponential. We also adapt Ruskey et al’s classic algorithm to enumerate zero-free graphical degree sequences of length n and Barnes and Savage’s classic algorithm to enumerate graphical partitions of even integer n by incorporating our testing algorithm into theirs and then obtain some enumerative results about forcibly connected graphical degree sequences of given length n and forcibly connected graphical partitions of given even integer n. Based on these enumerative results we make some conjectures such as: when n is large, (1) almost all zero-free graphical degree sequences of length n are forcibly connected; (2) almost none of the graphical partitions of even n are forcibly connected. Keywords— graphical degree sequence, graphical partition, forcibly connected, forcibly k-connected, co-NP 1 Introduction A graphical degree sequence of finite length n is a non-increasing sequence of non-negative integers d1 ≥ d2 ≥ · · · ≥ dn such that it is the vertex degree sequence of some simple graph (i.e. a finite undirected graph without loops or multiple edges). Given an arbitrary nonincreasing sequence of non-negative integers a1 ≥ a2 ≥ · · · ≥ an , it is easy to test whether it is a graphical degree sequence by using the Erdős-Gallai criterion [7] or the Havel-Hakimi algorithm [11, 9]. Seven equivalent criteria to characterize graphical degree sequences are summarized by Sierksma and Hoogeveen [19]. The notion of partition of an integer is well known in number theory, and is defined to be a non-increasing sequence of positive integers whose sum is the given integer. An integer partition is called a graphical partition if it is Department of Computer Sciences, Georgia Southern University, Statesboro, GA 30460, USA kwang@georgiasouthern.edu ∗ 1 the vertex degree sequence of some simple graph. Essentially a zero-free graphical degree sequence and a graphical partition are the same thing. It is often interesting to observe the properties of all the graphs having the same vertex degree sequence. A graph G with degree sequence d = (d1 ≥ d2 ≥ · · · ≥ dn ) is called a realization of d. Let P be any property of graphs (e.g. being bipartite, connected, planar, triangle-free, Hamiltonian, etc). A degree sequence d is called potentially P-graphic if it has at least one realization having the property P and forcibly P-graphic if all its realizations have the property P [15]. In this paper we only consider the property of k-connectedness (k ≥ 1 is fixed). Wang and Cleitman [20] give a simple characterization of potentially k-connected graphical degree sequences of length n, through which we can easily test whether a given graphical degree sequence is potentially connected. However, to the best of our knowledge no simple characterization of forcibly k-connected graphical degree sequences has been found so far and no algorithm has been published to test whether a given graphical degree sequence is forcibly connected or forcibly k-connected with given k. Some sufficient (but unnecessary) conditions are known for a graphical degree sequence to be forcibly connected or forcibly k-connected [5, 3, 6]. In the rest of this paper we will present a straight-forward algorithm to characterize forcibly connected graphical degree sequences and outline the extensions of the algorithm to test forcibly k-connectedness of graphical degree sequences for fixed k ≥ 2. We will demonstrate the efficiency of the algorithm through some computational experiments and then present some enumerative results regarding forcibly connected graphical degree sequences of given length n and forcibly connected graphical partitions of given even integer n. Base on the observations on these available enumerative results we make some conjectures about the relative asymptotic behavior of considered functions and the unimodality of certain associated integer sequences. 2 2.1 The testing algorithm Preliminaries Based on a result of Wang and Cleitman [20], a graphical · · · ≥ dn Pndegree sequence kd
1 ≥ d2 ≥ P k−1 di . is potentially k-connected if and only if dn ≥ k and i=1 di ≥ 2n − 2 2 − 2 + 2 i=1 Taking k = 1, we get that a zero-free P graphical degree sequence d1 ≥ d2 ≥ · · · ≥ dn is potentially connected if and only if ni=1 di ≥ 2n − 2. Note that any graphical degree sequence with a 0 in it can be neither potentially nor forcibly connected. We will design an algorithm to test whether a zero-free graphical degree sequence d is forcibly connected based on the simple observation that d is forcibly connected if and only if it is not potentially disconnected, i.e., it does not have any disconnected realization. Equivalently we need to test whether d can be decomposed into two sub graphical degree sequences. For example, 3,3,3,3,2,2,2 is a potentially connected graphical degree sequence of length 7. It is not forcibly connected since it can be decomposed into two sub graphical degree sequences 3,3,3,3 and 2,2,2. Note also that when a graphical degree sequence can be decomposed into two sub graphical degree sequences, the terms in each sub sequence need not be consecutive in the original sequence. For example, the graphical 2 degree sequence 4,4,3,3,3,2,2,2,1 can be decomposed into two sub graphical degree sequences 4,4,3,3,3,1 and 2,2,2 or into 4,4,3,3,2 and 3,2,2,1. We say that the graphical degree sequence 3,3,3,3,2,2,2 has a natural decomposition because it has a decomposition in which the terms in each sub sequence are consecutive in the original sequence. The graphical degree sequence 4,4,3,3,3,2,2,2,1 is not forcibly connected but does not have a natural decomposition. On the other hand, the graphical degree sequence 6,6,6,5,5,5,5,4,4 is forcibly connected since there is no way to decompose it into two sub graphical degree sequences. 2.2 Pseudo-code and the proof of its correctness In this section we will present the pseudo-code of our Algorithm 1 to test forcibly-connectedness of a given zero-free graphical degree sequence. We then give a proof why it correctly identifies such graphical degree sequences. We assume the input is a zero-free graphical degree sequence already sorted in nonincreasing order. In case an input that does not satisfy this condition is given, we can still easily test whether it is graphical by the Erdős-Gallai criterion [7] or the Havel-Hakimi algorithm [11, 9]. The output will be True if the input is forcibly connected and False otherwise. The output can also include a way to decompose the input in case it is not forcibly connected and such a decomposition is desired. Algorithm 1: Pseudo-code to test forcibly-connectedness of a graphical degree sequence Input: A zero-free graphical degree sequence d = (d1 ≥ d2 ≥ · · · ≥ dn ) Output: True or False, indicating whether d is forcibly connected or not 1 if d1 ≥ n − 2 or dn ≥ ⌊n/2⌋ then 2 return True 3 if d1 = dn then 4 return False 5 su ← max{s : s < n − ds+1 }; // 2 ≤ su ≤ n − dn − 1 6 if there exists an s such that d1 + 1 ≤ s ≤ su and d1 = (d1 ≥ d2 ≥ · · · ≥ ds ) and d2 = (ds+1 ≥ ds+2 ≥ · · · ≥ dn ) are both graphical then 7 return False 8 for l ← dn + 1 to min{⌊n/2⌋, n − d1 − 1} do 9 if dn+1−l < l then 10 m ← min{i : di < l}; // 1 ≤ m ≤ n − l + 1 11 if l ≤ n − m then 12 Form all candidate decompositions of d into s1 and s2 such that s1 is taken from dL = (dm ≥ dm+1 ≥ · · · ≥ dn ) of length l and s2 = d − s1 is of length n − l and both with even sum. If both s1 and s2 are graphical, return False 13 return True Now we show why Algorithm 1 correctly identifies whether d is forcibly connected or not. The conditional test on line 1 works as follows. 3 • If d1 ≥ n − 2, then in any realization G of d the vertex v1 with degree d1 will be in a connected component with at least n − 1 vertices, leaving at most 1 vertex to be in any other connected component should the input d be non forcibly connected. However, a graph with a single vertex has the degree sequence 0, which contradicts the assumption that d is zero-free. Thus in this case d must be forcibly connected. • If dn ≥ ⌊n/2⌋, then the vertex vn with degree dn will be in a connected component with at least 1 + ⌊n/2⌋ vertices. Should the input d be non forcibly connected, there will be another connected component not containing vn and also having at least 1 + ⌊n/2⌋ vertices since each vertex in that connected component has degree at least dn ≥ ⌊n/2⌋. This will result in a realization with at least 2 + 2⌊n/2⌋ > n vertices, a contradiction. Thus in this case d must also be forcibly connected. The conditional test on line 3 works as follows. Let d1 = dn = d. • If n is odd, then d must be even since the degree sum of a graphical degree sequence must be even. When we reach line 3, we must have d < ⌊n/2⌋. Now d can be decomposed into two sub graphical degree sequences of length n−1 and n+1 respectively 2 2 n−1 since d < 2 = ⌊n/2⌋. Thus it is not forcibly connected. • If n is even, we consider two cases. Case (A): n/2 is even, i.e. n ≡ 0 mod 4. In this case d can be decomposed into two graphical degree sequences of length n/2 since d < ⌊n/2⌋ = n/2. Thus it is not forcibly connected. Case (B): n/2 is odd, i.e. n ≡ 2 mod 4. Further consider two sub cases. (B1): if d is odd, then d can be decomposed into two graphical degree sequences of length n/2 − 1 and n/2 + 1 respectively since d < n/2 − 1 as a result of d and n/2 being both odd and d < n/2. (B2): if d is even, then d can be decomposed into two graphical degree sequences of length n/2 since d < n/2. Thus in this case d is not forcibly connected. Lines 5 to 7 try to find if d can be decomposed into two sub graphical degree sequences such that each sub sequence contains terms consecutive in the original sequence d, i.e. if the input d has a natural decomposition. For each given s, the two sub sequences d1 = (d1 ≥ d2 ≥ · · · ≥ ds ) and d2 = (ds+1 ≥ ds+2 ≥ · · · ≥ dn ) can be tested whether they are graphical by utilizing a linear time algorithm [12] that is equivalent to the Erdős-Gallai criterion. The smallest s that need to be tested is d1 + 1 since d1 can only be in a graphical degree sequence of length at least d1 + 1 and d1 has length s. The largest s that need to be tested is at most n − dn − 1 since dn can only be in a graphical degree sequence of length at least dn + 1 and d2 has length n − s. Actually the upper bound of the tested s can be chosen to be at most the largest s such that s < n − ds+1 since ds+1 can only be in a graphical degree sequence of length at least ds+1 + 1. Let su be the largest integer that satisfies this inequality. Note su ≥ 2 since s = 2 satisfies the inequality at the point of line 5. Also note that su ≤ n − dn − 1 because if su ≥ n − dn then n − dn ≤ su < n − dsu +1 , which leads to dsu +1 < dn , a contradiction. Therefore the upper bound of tested s is chosen to be su . Additional data structures can be maintained to skip the tests of those s for which each of the two sub sequences d1 and d2 has odd sum. Clearly a necessary condition for the input 4 d to have a natural decomposition is su ≥ d1 + 1. A weaker necessary condition easier to check is n − dn − 1 ≥ d1 + 1, i.e. d1 + dn ≤ n − 2. The for loop starting from line 8 is to test whether the input d can be decomposed into two sub graphical degree sequences of length l and n − l respectively, whether the decomposition is natural or not. At first glance we need to test the range of l in 2 ≤ l ≤ n−2 since the shortest zero-free graphical degree sequence has length 2. By symmetry we do not need to test those l beyond ⌊n/2⌋. Actually we only need to test the range of l from dn + 1 to min{⌊n/2⌋, n − d1 − 1}. We can start the loop with l = dn + 1 since, should the input d be decomposable, dn must be in a sub graphical degree sequence of length at least dn + 1 and the other sub graphical degree sequence not containing dn must also be of length at least dn + 1 due to all its terms being at least dn . There is no need to test those l > n − d1 − 1 since, should the input d be decomposable, d1 must be in a sub graphical sequence of length at least d1 + 1 and the other sub graphical sequence not containing d1 must have length at most n − d1 − 1. The condition tested on line 9 (dn+1−l < l) is necessary for d to be decomposable into two sub graphical degree sequences of length l and n − l respectively. A zero-free graphical degree sequence of length l must have all its terms less than l. If d is decomposable into two sub graphical degree sequences of length l and n − l respectively, d must have at least l terms less than l and n − l terms less than n − l. Therefore, the l smallest terms of d (dn−l+1 ≥ dn−l+2 ≥ · · · ≥ dn ) must be all less than l and the n − l smallest terms of d (dl+1 ≥ dl+2 ≥ · · · ≥ dn ) must be all less than n − l. These translate to the necessary conditions dn−l+1 < l and dl+1 < n − l for d to be decomposable. The condition dl+1 < n − l has already been satisfied since d1 < n − l based on the loop range of l on line 8. Lines 10 to 12 first find out the sub sequence dL of d consisting exactly of those terms less than l and then exhaustively enumerate all sub sequences s1 of dL with length l and even sum, trying to find a valid decomposition of d into s1 and s2 = d − s1 with length n−l, consisting of the terms of d not in s1 . Note that the l terms of s1 need not be consecutive in dL . The motivation for the construction of m and dL = (dm ≥ dm+1 ≥ · · · ≥ dn ) is that, should the input d be decomposable into two sub graphical degree sequences of length l and n − l respectively, the sub graphical degree sequence with length l must have all its terms coming from dL . For each such sub sequence s1 of dL with length l (we can always choose such an s1 since dL has length n − m + 1 ≥ l due to the definition of m on line 10), let the remaining terms of d form a sub sequence s2 = d − s1 of length n − l. If both s1 and s2 are graphical degree sequences, then the input d is not forcibly connected since we have found a valid decomposition of d into s1 and s2 and we may return False on line 12. The conditional test on line 11 (l ≤ n − m) is added because at this point we know d cannot be naturally decomposed and we can therefore exclude the consideration of l = n − m + 1 since under this condition there is only one possible choice of s1 from dL and consequently only one possible decomposition of d into two sub sequences of length l and n − l respectively, which is also a natural decomposition. If we remove the natural decomposition test from lines 5 to 7 and also remove the conditional test on line 11, the algorithm would obviously still be correct. If in the for loop from lines 8 to 13 we never return False on line 12, this means there is no way to decompose the input d into two sub graphical degree sequences whatsoever and we should return True on line 14. If we return False on line 4, 7, or 12 then a valid decomposition can also be returned if desired. 5 Later we will show that there is a computable threshold M(n) given the length n of the input d such that if d1 is below this threshold the algorithm can immediately return False without any exhaustive enumerations. However, our computational experiences suggest that if the input satisfies d1 < M(n) then Algorithm 1 already runs fast and it might not be worthwhile to add the computation of the additional threshold M(n) into it. 2.3 Extensions of the algorithm In this section we show how to extend Algorithm 1 to perform additional tasks such as listing all possible decompositions of a graphical degree sequence and testing forcibly kconnectedness of a graphical degree sequence for fixed k ≥ 2. 2.3.1 Enumeration of all possible decompositions Algorithm 1 can be easily extended to give all possible decompositions of the input d into two sub graphical degree sequences in case it is not forcibly connected. We simply need to report a valid decomposition found on line 3, 6 and 12 and continue without returning False immediately. Such an enumerative algorithm to find all valid decompositions of the input d can be useful when we want to explore the possible realizations of d and their properties. 2.3.2 Testing forcibly k-connectedness of d when k ≥ 2 It is also possible to extend Algorithm 1 to decide whether a given graphical degree sequence d is forcibly biconnected or not. We know that a connected graph is biconnected (nonseparable) if and only if it does not have a cut vertex. This characterization leads us to observe that if in any forcibly connected graphical degree sequence d the removal of any term di and the reduction of some collection dS of di elements from the remaining sequence d − {di } by 1 each results in a non forcibly connected graphical degree sequence d′ then d is not forcibly biconnected. If no such term and a corresponding collection of elements from the remaining sequence can be found whose removal/reduction results in a non forcibly connected graphical degree sequence, then d is forcibly biconnected. We give a pseudo-code framework in Algorithm 2 to decide whether a given graphical degree sequence d is forcibly biconnected or not. To simplify our description, we call the above mentioned combination of removal/reduction operations a generalized Havel-Hakimi (GHH) operation, notationally d′ = GHH(d, di , dS ). We remark that if the d′ obtained on line 4 of Algorithm 2 is not a graphical degree sequence then the condition on line 5 is not satisfied and the algorithm will not return False at the moment. Similarly we can test whether a given graphical degree sequence d is forcibly k-connected or not for k ≥ 3 iteratively as long as we already have a procedure to test whether a graphical degree sequence is forcibly (k − 1)-connected or not. Suppose we already know an input d is potentially k-connected and forcibly (k − 1)-connected. We can proceed to choose a term di and a collection dS of size di from the remaining sequence d − {di } and perform a GHH operation on d. If the resulting sequence d′ is a graphical degree sequence and is non forcibly (k − 1)-connected, then d is not forcibly k-connected. If no such term and a corresponding collection of elements from the remaining sequence can be found whereby a GHH operation 6 Algorithm 2: Pseudo-code to test whether a graphical degree sequence is forcibly biconnected (See text for the description of GHH operation) Input: A zero-free graphical degree sequence d = (d1 ≥ d2 ≥ · · · ≥ dn ) Output: True or False, indicating whether d is forcibly biconnected or not 1 if d is not potentially biconnected or forcibly connected then 2 return False 3 for each di and each collection dS of size di from d − {di } do 4 d′ ← GHH(d, di, dS ); 5 if d′ is a non forcibly connected graphical degree sequence then 6 return False 7 return True can be performed on d to result in a non forcibly (k−1)-connected graphical degree sequence, then d is forcibly k-connected. We give a pseudo-code framework in Algorithm 3 to decide whether a given graphical degree sequence d is forcibly k-connected or not. Algorithm 3: Pseudo-code to test whether a graphical degree sequence is forcibly k-connected (See text for the description of GHH operation) Input: A zero-free graphical degree sequence d = (d1 ≥ d2 ≥ · · · ≥ dn ) and an integer k≥2 Output: True or False, indicating whether d is forcibly k-connected or not 1 if d is not potentially k-connected or forcibly (k − 1)-connected then 2 return False 3 for each di and each collection dS of size di from d − {di } do 4 d′ ← GHH(d, di, dS ); 5 if d′ is a non forcibly (k − 1)-connected graphical degree sequence then 6 return False 7 3 return True Complexity analysis We conjecture that Algorithm 1 runs in time polynomial in n on average. The worst case run time complexity is probably still exponential in n. We are unable to provide a rigorous proof at this time, but we will later show through experimental evaluations that it runs fast on randomly generated long graphical degree sequences most of the time. Now we give a discussion of the run time behavior of Algorithm 1. Observe that lines 1 to 4 take constant time. Lines 5 to 7 take O(n2 ) time if we use the linear time algorithm from [12] to test whether an integer sequence is graphical. Lines 9 to 11 combined take O(n) time and they are executed O(n) times. So the overall time complexity is O(n2 ) excluding the time on line 12. 7 Next consider all candidate decompositions of d into s1 and s2 on line 12. The sub sequence s1 is taken from dL = (dm ≥ dm+1 ≥ · · · ≥ dn ) whose length could be as large as n and the length
l of s1 could be as large as ⌊n/2⌋. Therefore in the worst case we may n have up to n/2 candidate decompositions, which could make the run time of Algorithm 1 exponential in n. A careful implementation of Algorithm 1 will help reduce running time by noting that dL is a multi-set and provides us an opportunity to avoid duplicate enumerations of s1 because different l combinations of the indices (m, m + 1, · · · , n) could produce the same sub sequence s1 . For this purpose, we can assume the input d is also provided in another format (e1 , f1 ), (e2 , f2 ), · · · , (eq , fq ) where d contains fi copies of ei for i = 1, · · · , q and e1 > e2 > · · · > eq > 0. (Clearly d1 = e1 and dn = eq .) Now enumerating s1 of length l from dL can be equivalently translated to the following problem of enumerating all non-negative integer solutions of Equation (1) subject to constraints (2), k X xi = l, (1) i=1 0 ≤ xi ≤ fq−k+i , for i = 1, · · · , k, (2) where k is the number of distinct elements in dL = (dm ≥ dm+1 ≥ · · · ≥ dn ) which can also be represented as (eq−k+1, fq−k+1 ), (eq−k+2, fq−k+2 ), · · · , (eq , fq ) and k satisfies k ≤ q and k ≤ l − dn since all the elements of dL are < l and ≥ dn . In this context m and k vary with l as the for loop from lines 8 to 12 progresses. Each solution of Equation (1) represents a candidate choice of s1 out of dL with length l by taking xi copies of eq−k+i . Further improvement could be achieved by noting the odd terms among eq−k+1 , eq−k+2, · · · , eq since we must have an even number of odd terms in s1 for it to have even sum. We can categorize the xi variables of Equation (1) into two groups based on the parity of the corresponding ei and enumerate only its solutions having an even sum of the xi ’s belonging to the odd group. The number of solutions of Equation (1) can be exponential in n. For example, let l = n/2, k = n/4 and let fj = 4 for j = q − k + 1, · · · , q. Then the number of solutions of
Equation (1) will be at least n/4 by taking half of all x1 , · · · , xk to be 4 and the remaining n/8 half to be 0. However in practice we rarely find such a large number of solutions are actually all enumerated before Algorithm 1 returns. To the best of our knowledge, the computational complexity of the decision problem of whether a given graphical degree sequence is forcibly connected is unknown. The problem is clearly in co-NP since a short certificate to prove that a given input is not forcibly connected is a valid decomposition of the input sequence. But is it co-NP-hard? As far as we know, this is an open problem. The time complexity of the extension Algorithms 2 and 3 to test whether a given graphical degree sequence d is forcibly k-connected or not for k ≥ 2 is apparently exponential due to the exhaustive enumeration of the candidate collection dS of size di from the remaining sequence d − {di } and the ultimate calls to Algorithm 1 possibly an exponential number of times. The computational complexity of the decision problem of whether a given graphical degree sequence is forcibly k-connected (k ≥ 2) is also unknown to us. Clearly the problem 8 is still in co-NP when k is fixed as to prove that a graphical degree sequence d is not forcibly k-connected is as easy as using a sequence of k − 1 certificates each consisting of a pair (di , dS ) and a k th certificate being a valid decomposition to show that the final resulting sequence is decomposable after a sequence of k − 1 GHH operations on d, but we do not know if it is inherently any harder than the decision problem for k = 1. 4 Computational results In this section we will first present some results on the experimental evaluation on the performance of Algorithm 1 on randomly generated long graphical degree sequences. We will then provide some enumerative results about the number of forcibly connected graphical degree sequences of given length and the number of forcibly connected graphical partitions of a given even integer. Based on the available enumerative results we will make some conjectures about the asymptotic behavior of related functions and the unimodality of certain associated integer sequences. 4.1 Performance evaluations of algorithm 1 In order to evaluate how efficient Algorithm 1 is, we aim to generate long testing instances with length n in the range of thousands and see how Algorithm 1 performs on these instances. Our experimental methodology is as follows. Choose a constant ph in the range [0.1,0.95] and a constant pl in the range of [0.001,min{ph − 0.01, 0.49}] and generate 100 random graphical degree sequences of length n with largest term around ph n and smallest term around pl n. Each such graphical degree sequence is generated by first uniformly random sampling integer partitions with the specified number of parts n and the specified largest part and smallest part and then accept it as input for Algorithm 1 if it is a graphical degree sequence. We run Algorithm 1 on these random instances and record the average performance and note the proportion of them that are forcibly connected. Table 1 lists the tested ph and pl . The largest tested pl is 0.49 since any graphical degree sequence of length n and smallest term at least 0.5n will cause Algorithm 1 to return True on line 2. We implemented our Algorithm 1 using C++ and compiled it using g++ with optimization level -O3. The experimental evaluations are performed on a common Linux workstation. We summarize our experimental results for the length n = 1000 as follows. 1. For those instances with ph in the range from 0.1 to 0.5, Algorithm 1 always finishes instantly (run time < 0.01s) and all the tested instances are non forcibly connected. This does not necessarily mean that there are no forcibly connected graphical degree sequences of length n = 1000 with largest term around ph n with ph in this range. It only suggests that forcibly connected graphical degree sequences are relatively rare in this range. 2. For each ph in the range from 0.55 to 0.95, we observed a transition interval It of pl for each fixed ph . See Table 2 for a list of observed transition intervals. All those instances with pl below the range It are non forcibly connected and all those instances with pl above the range It are forcibly connected. Those instances with pl in the range It exhibit the behavior that the proportion of forcibly connected among all tested 100 instances gradually increases from 0 to 1 as pl increases in the range It . For example, based on the results of Table 2, 9 Table 1: Chosen ph and pl in the experimental performance evaluation of Algorithm 1. ph 0.10 0.20 0.30 0.40 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 pl 0.001,0.002,0.003,...,0.01,0.02,0.03,...,0.09 0.001,0.002,0.003,...,0.01,0.02,0.03,...,0.19 0.001,0.002,0.003,...,0.01,0.02,0.03,...,0.29 0.001,0.002,0.003,...,0.01,0.02,0.03,...,0.39 0.001,0.002,0.003,...,0.01,0.02,0.03,...,0.49 0.001,0.002,0.003,...,0.01,0.02,0.03,...,0.49 0.001,0.002,0.003,...,0.01,0.02,0.03,...,0.49 0.001,0.002,0.003,...,0.01,0.02,0.03,...,0.49 0.001,0.002,0.003,...,0.01,0.02,0.03,...,0.49 0.001,0.002,0.003,...,0.01,0.02,0.03,...,0.49 0.001,0.002,0.003,...,0.01,0.02,0.03,...,0.49 0.001,0.002,0.003,...,0.01,0.02,0.03,...,0.49 0.001,0.002,0.003,...,0.01,0.02,0.03,...,0.49 0.001,0.002,0.003,...,0.01,0.02,0.03,...,0.49 Table 2: Transition interval It of pl for each ph (for n = 1000). ph 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 It of pl 0.30 to 0.40 0.20 to 0.30 0.15 to 0.24 0.09 to 0.17 0.05 to 0.12 0.03 to 0.09 0.01 to 0.07 0.003 to 0.04 0.001 to 0.03 the proportions of forcibly connected graphical degree sequences of length 1000 with largest term around 850 (ph = 0.85) and smallest term below around 10 (pl = 0.01) are close to 0. If the smallest term is above around 70 (pl = 0.07) then the proportion is close to 1. When the smallest term is between 10 and 70 (pl in the range from 0.01 to 0.07) then the proportion transitions from 0 to 1. Again these results should be interpreted as relative frequency instead of absolute law. 3. Algorithm 1 is efficient most of the time but encounters bottlenecks at occasions. For ph from 0.80 to 0.95 and pl near the lower end of the transition interval It , Algorithm 1 does perform poorly on some of the tested instances with run time from a few seconds to more than a few hours (time out). The exact range of pl near the lower end of It where Algorithm 1 could perform poorly varies. We observed that this range of pl for which the algorithm could perform poorly is quite narrow. For example, when n = 1000, ph = 0.9, this range of pl we observed is from 0.001 to 0.01. We observed that the frequency at which the 10 algorithm performs poorly also varies. We believe that this is because among all possible instances with given length and given largest and smallest terms there is still great variety in terms of difficulty of testing their property of forcibly connectedness using Algorithm 1. In particular, some instances will trigger the exhaustive behavior of Algorithm 1 on line 12, making it enumerate a lot of candidate decompositions without returning. We have also performed experimental evaluations of Algorithm 1 for the length n = 2000, 3000, ..., 10000 without being able to finish all the same ph , pl choices as for n = 1000 because of shortage of time. The behavior of Algorithm 1 on inputs of these longer lengths is similar to the case of n = 1000 but with different transition intervals It and varied range of pl near the lower end of the transition interval It for which it could perform poorly. To sum up, we believe that the average case run time of Algorithm 1 is polynomial. We estimate that more than half of all zero-free graphical degree sequences of length n can be tested in constant time on line 1. However, its worst case run time should be exponential. As mentioned above, the computational complexity of the decision problem itself is unknown to us. Currently we have a very rudimentary implementation of Algorithm 2 and do not have an implementation of Algorithm 3 for any k ≥ 3 yet. Algorithm 2 can start to encounter bottlenecks for input length n around 40 to 50, which is much shorter than the input lengths Algorithm 1 can handle. We suspect that to handle input length n ≥ 100 when k = 3 will be very difficult unless significant enhancement to avoid many of those exhaustive enumerations can be introduced. 4.2 Enumerative results In this section we will present some enumerative results related to forcibly connected graphical degree sequences of given length and forcibly connected graphical partitions of given even integer. We also make some conjectures based on these enumerative results. For the reader’s convenience, we summarize the notations used in this section in Table 3. In a previous manuscript [21] we have presented efficient algorithms for counting the number of graphical degree sequences of length n and the number of graphical degree sequences of k-connected graphs with n vertices (or graphical degree sequences of length n that are potentially k-connected). It is proved there that the asymptotic orders of the number D(n) of zero-free graphical degree sequences of length n and the number Dc (n) of potentially conc (n) nected graphical degree sequences of length n are equivalent. That is, limn→∞ DD(n) = 1. In order to investigate how the number Df (n) of forcibly connected graphical degree sequences of length n grows compared to D(n) we conduct computations to count such graphical degree sequences. We do not have any algorithm that can get the counting result without actually generating the sequences. The fastest algorithm we know of that can generate all zero-free graphical degree sequences of length n is from Ruskey et al [18]. We adapted this algorithm to incorporate the test in Algorithm 1 to count those that are forcibly connected. Since D(n) grows as an exponential √ function of n based on the bounds given by Burns [4] (4n /(c1 n) ≤ D(n) ≤ 4n /((log n)c2 n)) for all sufficiently large n with c1 , c2 positive constants), it is unlikely to get the value of Df (n) for large n using an exhaustive generation algorithm. We only have counting results of Df (n) for n up to 26 due to the long running 11 Table 3: Terminology used in this section Term D(n) Dc (n) Df (n) Cn [N] Fn [N] Ln [j] M(n) g(n) gc (n) gf (n) cn [j] fn [j] ln [j] m(n) Meaning number of zero-free graphical sequences of length n number of potentially connected graphical sequences of length n number of forcibly connected graphical sequences of length n number of potentially connected graphical degree sequences of length n with degree sum N number of forcibly connected graphical degree sequences of length n with degree sum N number of forcibly connected graphical degree sequences of length n with largest term j minimum largest term in any forcibly connected graphical sequence of length n number of graphical partitions of even n number of potentially connected graphical partitions of even n number of forcibly connected graphical partitions of even n number of potentially connected graphical partitions of even n with j parts number of forcibly connected graphical partitions of even n with j parts number of forcibly connected graphical partitions of n with largest term j minimum largest term of forcibly connected graphical partitions of n time of our implementation. The results together with the proportion of them in all zerofree graphical degree sequences are listed in Table 4. From the table it seems reasonable to conclude that the proportion Df (n)/D(n) will increase when n ≥ 8 and it might tend to the limit 1. Since our adapted algorithm from Ruskey et al [18] for computing Df (n) actually generates all forcibly connected graphical degree sequences of length n it is trivial to also output the individual counts based on the degree sum N or the largest degree ∆. That is, we can output the number of forcibly connected graphical degree sequences of length n with degree sum N or largest term ∆. In Table 5 we show itemized potentially and forcibly connected graphical degree sequences of length 7 based on the degree sum N. The counts for N < 12 are not shown because those counts are all 0. The highest degree sum is 42 for any graphical degree sequence of length 7. From the table we see that the individual counts based on the degree sum N that contribute to Dc (7) (row C7 [N]) and Df (7) (row F7 [N]) both form a unimodal sequence. Counts for other degree sequence lengths from 5 to 26 exhibit similar behavior. Based on the available enumerative results we find that for any given n the range of even N for Cn [N] and Fn [N] to be nonzero respectively are exactly the same (between 2n − 2 and n(n − 1)). In fact, this can be proved as the following Proposition 4.1. An even N has a potentially connected graphical partition with n parts if and only if it has a forcibly connected graphical partition with n parts. 12 Table 4: Number of forcibly connected graphical degree sequences of length n and their proportions in zero-free graphical degree sequences of length n. n 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 D(n) Df (n) 7 6 20 18 71 63 240 216 871 783 3148 2843 11655 10535 43332 39232 162769 147457 614198 556859 2330537 2113982 8875768 8054923 33924859 30799063 130038230 118098443 499753855 454006818 1924912894 1749201100 7429160296 6752721263 28723877732 26114628694 111236423288 101153550972 431403470222 392377497401 1675316535350 1524043284254 6513837679610 5926683351876 25354842100894 23073049582134 Df (n)/D(n) 0.857143 0.900000 0.887324 0.900000 0.898967 0.903113 0.903904 0.905382 0.905928 0.906644 0.907079 0.907518 0.907861 0.908182 0.908461 0.908717 0.908948 0.909161 0.909356 0.909537 0.909705 0.909860 0.910006 Table 5: Number of potentially (row C7 [N]) and forcibly (row F7 [N]) connected graphical degree sequences of length 7 with given degree sum N. degree sum N C7 [N] F7 [N] 12 7 3 14 11 5 16 15 10 18 22 19 20 26 25 22 29 28 24 29 29 26 26 26 28 23 23 30 18 18 32 13 13 34 8 8 36 5 5 38 2 2 40 1 1 42 1 1 Table 6: Number L15 [j] of forcibly connected graphical degree sequences of length 15 with given largest term j. largest part j L15 [j] 14 3166852 13 2624083 12 1398781 11 600406 10 201128 9 52903 8 9718 7 1031 6 21 Proof. Sufficiency is obvious by definition. In the following we show the necessity. Suppose an even N has a potentially connected graphical partition with n parts. From the Wang and Cleitman characterization [20] we know that N must be between 2n − 2 and 13 n(n − 1) for it to have a potentially connected graphical partition of n parts. Now construct a partition π of N with n parts as follows. Let the largest part be n−1 and let the remaining −n+1 n − 1 parts be as evenly as possible. That is, let b = ⌊ N n−1 ⌋ and a = N − (n − 1)(b + 1). Then the smallest n − 1 parts of π consist of a copies of b + 1 and n − 1 − a copies of b. With 2n − 2 ≤ N ≤ n(n − 1), we have 0 < b ≤ n − 1 and 0 ≤ a < n − 1. Based on the Nash-Williams condition it is easy to verify that π is a graphical partition of N with n parts and it is forcibly connected since its largest part is n − 1.  In Table 6 we show itemized numbers of forcibly connected graphical degree sequences of length 15 based on the largest degree. The counts for largest degrees less than 6 are not shown because those counts are all 0. From the table we can see that the counts decrease with the largest degree. For other degree sequence lengths from 5 to 26 we observed similar behavior. The table also indicates that there are no forcibly connected graphical degree sequences of length 15 with largest degree less than 6. In fact, if we define M(n) to be the minimum largest term in any forcibly connected graphical sequence of length n. This is, . M(n) = min{∆: ∆ is the largest term of some forcibly connected graphical degree sequence of length n}. Clearly we have M(n) ≤ n/2 since for even n the sequence n/2, n/2, · · · , n/2 of length n is forcibly connected. We can show a lower bound of M(n) as follows. √ Theorem 4.2. For M(n) defined√above, we have M(n) = Ω( n). That is, there is a constant c > 0 such that M(n) > c n for all sufficiently large n. Proof. For the purpose of deriving a contradiction assume there is a forcibly connected graphical degree√sequence ß = (d1 ≥ d2 ≥ · · · ≥ dn ) of length n with the largest term d1 = M(n) = o( n). Let us first consider the case that n is even. Let ßH be the higher half (of length n/2) of ß and ßL be the lower half (of length n/2) of ß. If both ßH and ßL have even sums, then they can shown to be both graphical degree sequences based on the Nash-Williams condition [17, 16, 19] as follows. Suppose the Durfee square size of ß is s where s ≤ d1 = M(n) by the definition of Durfee square. Since ß is graphical it satisfies the Nash-Williams condition, which can be represented as s inequalities: j X i=1 d′1 , · · · (d′i − di ) ≥ j, j = 1, · · · , s, , d′s ′ where are the largest s parts of the conjugate partition of the partition √ ß (d1 = n). Now ß and ßH have the same Durfee square size by our assumption that s = o( n) and they have the same s largest parts. Let the s largest parts of the conjugate of ßH be d′′1 , · · · , d′′s with d′′1 = n/2 by our construction. To show that ßH is graphical, we only need to show that the following s inequalities hold: j X i=1 (d′′i − di ) ≥ j, j = 1, · · · , s. (3) ′′ The √ first of these inequalities d1 − d1 = n/2 − M(n) ≥ 1 is clearly satisfied since M(n) = o( n). We also have the following inequalities, d′′j ≥ s and dj ≤ M(n), j = 2, · · · , s, 14 so we have d′′j − dj ≥ s − M(n), j = 2, · · · , s. Even√if d′′j − dj , j = 2, · · · , s are all negative, their sum will be of order o(n) since s ≤ M(n) = o( n). Clearly the s inequalities in (3) are all satisfied since d′′1 − d1 = n/2 − M(n) is of order Ω(n). This shows that ßH is graphical. By the same argument ßL is graphical and we have found that ß can be decomposed into two sub graphical degree sequences ßH and ßL . This contradicts our assumption that ß is forcibly connected. If both ßH and ßL have odd sums (we cannot have one of them having even sum and the other having odd sum since the sum of all the terms of ß is even), then it must be the case that both ßH and ßL have an odd number of odd terms. Construct two new sequences ß′H and ß′L from ßH and ßL by removing the largest odd term from ßL and adding it to ßH . Now clearly ß′H and ß′L is a decomposition of ß into two sub sequences of length n/2 + 1 and n/2 − 1 respectively and both having even sums. Again they are guaranteed to be graphical degree sequences by the Nash-Williams condition using a similar argument as above, which contradicts the assumption that ß is forcibly connected. The case for n odd √ can be proved in a similar way. The conclusion that M(n) cannot be of lower order than n then follows.  We do not have any theory or algorithm to efficiently obtain M(n) for any given n. Other than recording the minimum largest term while enumerating all forcibly connected graphical degree sequences of length n, a naive approach would be to let ∆ start from 3 upward and test if there is a forcibly connected graphical degree sequence of length n with largest term ∆ and stop incrementing ∆ when we have found one. Obviously this works but is not efficient. Any efficient algorithm for M(n) might be worthwhile to be added into Algorithm 1 so that it can immediately return False if d1 < M(n). However, while we conduct performance evaluations of Algorithm 1 we do find that a random graphical degree sequence of length n with d1 ≤ n/2 most likely can be decided instantly by Algorithm 1. Therefore we believe that an efficient algorithm for M(n) will not help much on average. We show the values of M(n) based on our enumerative results in Table 7. The fact that M(15) = 6 agrees with the results of Table 6 where the counts L15 [j] = 0 for all j < 6. As a side note, the minimum largest term in any potentially connected graphical sequence of length n is clearly 2 since the degree sequence 2, 2, · · · , 2 (n copies) is potentially connected while 1, 1, · · · , 1 (n copies) is not potentially connected. Table 7: Minimum largest term M(n) of forcibly connected graphical sequences of length n. n M(n) n M(n) 3 2 15 6 4 2 16 6 5 2 17 7 6 3 18 7 7 3 19 7 8 3 20 7 9 4 21 8 10 4 22 8 11 5 23 8 12 5 24 8 13 5 25 8 14 6 26 9 We also investigated the number gf (n) of forcibly connected graphical partitions of a given even integer n. There is a highly efficient Constant Amortized Time (CAT) algorithm of Barnes and Savage [2] to generate all graphical partitions of a given even n. And there are efficient counting algorithms of Barnes and Savage [1] and Kohnert [13] to count the number g(n) of graphical partitions of even n without generating them. It is known from 15 Erdős and Richmond [8] that the number gc (n) of potentially connected graphical partitions c (2n) = 1. It is also known from Pittel of n and g(n) are of equivalent order, i.e. limn→∞ gg(2n) [14] that the proportion of graphical partitions among all partitions of an integer n tends to 0. Although the order of the number p(n) of unrestricted partitions of n is long known since Hardy and Ramanujan [10], the exact asymptotic order of g(n) is still unknown. We know of no algorithm to count the number gf (n) of forcibly connected graphical partitions of n without generating them. Using a strategy similar to that we employed in computing Df (n), we adapted the algorithm of Barnes and Savage [2] and incorporated the test of forcibly connectedness from Algorithm 1 and then count those that are forcibly connected. The growth of g(n) is quick and we only have numerical results of gf (n) for n up to 170. The results together with the proportion of them in all graphical partitions are listed in Table 8. For the purpose of saving space we only show the results in increments of 10 for n. From the table it seems reasonable to conclude that the proportion gf (n)/g(n) will decrease when n is beyond some small threshold and it might tend to the limit 0. Table 8: Number of forcibly connected graphical partitions of n and their proportions in all graphical partitions of n n 10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 170 g(n) 17 244 2136 14048 76104 357635 1503172 5777292 20614755 69065657 219186741 663394137 1925513465 5383833857 14555902348 38173235010 97368672089 gf (n) 8 81 586 3308 15748 66843 256347 909945 3026907 9512939 28504221 81823499 226224550 604601758 1567370784 3951974440 9714690421 gf (n)/g(n) 0.470588 0.331967 0.274345 0.235478 0.206927 0.186903 0.170537 0.157504 0.146832 0.137738 0.130045 0.123341 0.117488 0.112299 0.107679 0.103527 0.099772 Table 9: Number of potentially (row c20 [j]) and forcibly (row f20 [j]) connected graphical partitions of 20 with given number of parts j. number of parts j c20 [j] f20 [j] 5 1 1 6 9 9 16 7 26 25 8 38 22 9 37 10 10 36 9 11 30 5 Table 10: Number l20 [j] of forcibly connected graphical partitions of 20 with given largest term j. largest part j l20 [j] 3 1 4 14 5 26 6 20 7 12 8 5 9 2 10 1 Table 11: Minimum largest term m(n) of forcibly connected graphical partitions of n. n m(n) 10 2 20 3 30 4 40 5 50 5 60 6 70 6 80 6 90 7 100 7 Like the situation for Df (n) the adapted algorithm from Barnes and Savage [2] to compute gf (n) actually generates all forcibly connected graphical partitions of n so it is trivial to also output the individual counts based on the number of parts or the largest part. In Table 9 we show the individual counts of potentially and forcibly connected graphical partitions of 20 based on the number of parts. Counts for the number of parts less than 5 or greater than 11 are not shown since those counts are all 0. The ranges of the number of parts j for which the number cn [j] of potentially connected graphical partitions of n with j parts and the number fn [j] of forcibly connected graphical partitions of n with j parts are nonzero are exactly the same based on Proposition 4.1. The smallest number of parts j for which cn [j] and fn [j] are both nonzero is the smallest positive integer t(n) such that t(n)(t(n) − 1) ≥ n and this is also the smallest number of parts for which a partition of n with this many parts might be graphical. The largest number of parts j for which cn [j] and fn [j] are both nonzero is n/2 + 1 based on the Wang and Cleitman characterization [20]. In Table 10 we show the individual counts of forcibly connected graphical partitions of 20 based on the largest part. Counts for the largest part less than 3 or greater than 10 are not shown since those counts are all 0. Clearly n/2 is the maximum largest part of any forcibly connected graphical partition of n since n/2, 1, 1, · · · , 1 (n/2 copies of 1) is a forcibly connected graphical partition of n and no graphical partition of n has its largest part greater than n/2. However, similar to the case of M(n), the minimum largest part, m(n), of any forcibly connected graphical partition of √ n does not seem to be easily obtainable. Clearly m(n)√grows at most like n since for √ every large even n it has a graphical partition with about n parts and all parts about n − 1 and this graphical partition is forcibly connected. In Table 11 we show several values of m(n). They are obtained while we exhaustively generate all graphical partitions of n and keep a record of the minimum largest part. The fact that m(20) = 3 agrees with the results of Table 10 where l20 [j] = 0 for all j < 3. As a side note, the minimum largest part of any potentially connected graphical partition of n is clearly 2 since 2, 2, · · · , 2 (n/2 copies) is a potentially connected graphical partition of n while 1, 1, · · · , 1 (n copies) is not. 4.3 Questions and conjectures Based on the available enumerative results we ask the following questions and make certain conjectures: 1. What is the growth order of Df (n) relative to D(n)? Note that the exact asymptotic 17 order of D(n) is unknown yet. (Some upper and lower bounds of D(n) are known. See Df (n) = 1. That is, almost all zero-free graphical degree Burns [4]). We conjecture limn→∞ D(n) sequences of length n are forcibly connected. If this is true, then it is a stronger result than c (n) the known result (see [21]) limn→∞ DD(n) = 1 since Df (n) ≤ Dc (n) ≤ D(n). Furthermore, we conjecture that Df (n)/D(n) is monotonously increasing when n ≥ 8. Let Dc k (n) and Df k (n) denote the number of potentially and forcibly k-connected graphical degree sequences c k (n) 6= 1 when k ≥ 2. of length n respectively. It is already known from [21] that limn→∞ DD(n) D (n) f k Clearly we also have limn→∞ D(n) 6= 1 when k ≥ 2. What can be said about the relative orders of Dc k (n), Df k (n) and D(n) when k ≥ 2? 2. What is the growth order of gf (2n) relative to g(2n)? Note that the exact asymptotic gf (2n) order of g(2n) is unknown yet. We conjecture limn→∞ g(2n) = 0. That is, almost none of the graphical partitions of 2n are forcibly connected. Furthermore, we conjecture that gf (2n)/g(2n) is monotonously decreasing when n ≥ 5. Let gc k (n) and gf k (n) denote the number of potentially and forcibly k-connected graphical partitions of n respectively. What can be said about the relative orders of gc k (n), gf k (n) and g(n) when k ≥ 2? 3. We conjecture that the numbers of forcibly connected graphical partitions of N with exactly n parts, when N runs through 2n − 2, 2n, · · · , n(n − 1), give a unimodal sequence. 4. Let t(n) be the smallest positive integer such that t(n)(t(n) − 1) ≥ n. We conjecture that the numbers of forcibly connected graphical partitions of n with j parts, when j runs through t(n), t(n) + 1, · · · , n/2 + 1, give a unimodal sequence. 5. What is the growth order of M(n), the minimum largest term in any forcibly connected graphical sequence of length n? Is there a constant C > 0 such that limn→∞ Mn(n) = C? Is there an efficient algorithm to compute M(n)? 6. What is the growth order of m(n), the minimum largest term in any forcibly connected √ = C? Is there an graphical partition of n? Is there a constant C > 0 such that limn→∞ m(n) n efficient algorithm to compute m(n)? 7. We conjecture that the numbers of forcibly connected graphical partitions of an even n with the largest part exactly ∆, when ∆ runs through m(n), m(n) + 1, · · · , n/2, give a unimodal sequence. 8. We showed all these decision problems to test whether a given graphical degree sequence is forcibly k-connected to be in co-NP for fixed k ≥ 1. Are they co-NP-hard? Is the decision problem for k + 1 inherently harder than for k? 5 Conclusions In this paper we presented an efficient algorithm to test whether a given graphical degree sequence is forcibly connected or not and its extensions to test forcibly k-connectedness of graphical degree sequences for fixed k ≥ 2. Through performance evaluations on a wide range of long random graphical degree sequences we demonstrate its average case efficiency and we believe that it runs in polynomial time on average. We then incorporated this testing algorithm into existing algorithms that enumerate zero-free graphical degree sequences of length n and graphical partitions of an even integer n to obtain some enumerative results about the number of forcibly connected graphical degree sequences of length n and forcibly 18 connected graphical partitions of n. We proved some simple observations related to the available numerical results and made several conjectures. We are excited that there are a lot for further research in this direction. References [1] Tiffany M. Barnes and Carla D. Savage. A Recurrence for Counting Graphical Partitions. Electronic Journal of Combinatorics, 2, 1995. [2] Tiffany M. Barnes and Carla D. Savage. Efficient generation of graphical partitions. Discrete Applied Mathematics, 78(13):17–26, 1997. [3] F. T. Boesch. The strongest monotone degree condition for n-connectedness of a graph. Journal of Combinatorial Theory, Series B, 16(2):162–165, 1974. [4] Jason Matthew Burns. The Number of Degree Sequences of Graphs. PhD thesis, Massachusetts Institute of Technology. Dept. of Mathematics, 2007. [5] Gary Chartrand, S. F. Kapoor, and Hudson V. Kronk. A sufficient condition for nconnectedness of graphs. Mathematika, 15(1):5152, 1968. [6] S. A. Choudum. On forcibly connected graphic sequences. Discrete Mathematics, 96(3):175–181, 1991. [7] P. Erdős and T. Gallai. Graphs with given degree of vertices. Mat. Lapok, 11:264–274, 1960. [8] P. Erdős and L. B. Richmond. On graphical partitions. Combinatorica, 13(1):57–63, 1993. [9] S. L. Hakimi. On realizability of a set of integers as degrees of the vertices of a linear graph. I. Journal of the Society for Industrial and Applied Mathematics, 10(3):496–506, 1962. [10] G. H. Hardy and S. Ramanujan. Asymptotic formulae in combinatory analysis. Proc. London Math. Soc., 17:75–115, 1918. [11] Václav Havel. A remark on the existence of finite graphs. Časopis pro pěstovánı́ matematiky, 80(4):477–480, 1955. [12] Antal Iványi, Gergő Gombos, Loránd Lucz, and Tamás Matuszka. Parallel enumeration of degree sequences of simple graphs II. Acta Universitatis Sapientiae, Informatica, 5(2):245–270, 2013. [13] Axel Kohnert. Dominance Order and Graphical Partitions. Electronic Journal of Combinatorics, 11(1), 2004. [14] Boris Pittel. Confirming Two Conjectures About the Integer Partitions. Journal of Combinatorial Theory, Series A, 88(1):123–135, 1999. 19 [15] S. B. Rao. A survey of the theory of potentially P-graphic and forcibly P-graphic degree sequences, pages 417–440. Springer Berlin Heidelberg, 1981. [16] C. C. Rousseau and F. Ali. A Note on Graphical Partitions. Journal of Combinatorial Theory, Series B, 64(2):314–318, 1995. [17] E. Ruch and I. Gutman. The Branching Extent of Graphs. Journal of Combinatorics, Information and System Sciences, 4(4):285–295, 1979. [18] Frank Ruskey, Robert Cohen, Peter Eades, and Aaron Scott. Alley CATs in search of good homes. In 25th S.E. Conference on Combinatorics, Graph Theory, and Computing, volume 102, pages 97–110. Congressus Numerantium, 1994. [19] Gerard Sierksma and Han Hoogeveen. Seven Criteria for Integer Sequences Being Graphic. Journal of Graph Theory, 15(2):223–231, 1991. [20] D. L. Wang and D. J. Kleitman. On the Existence of n-Connected Graphs with Prescribed Degrees (n≥2). Networks, 3(3):225–239, 1973. [21] Kai Wang. Efficient counting of degree sequences. https://arxiv.org/abs/1604.04148, Under Review. 20 

Disruptive Event Classification using PMU Data in Distribution Networks I. Niazazari and H. Livani, Member, IEEE  Abstract— Proliferation of advanced metering devices with high sampling rates in distribution grids, e.g., micro-phasor measurement units (μPMU), provides unprecedented potentials for wide-area monitoring and diagnostic applications, e.g., situational awareness, health monitoring of distribution assets. Unexpected disruptive events interrupting the normal operation of assets in distribution grids can eventually lead to permanent failure with expensive replacement cost over time. Therefore, disruptive event classification provides useful information for preventive maintenance of the assets in distribution networks. Preventive maintenance provides wide range of benefits in terms of time, avoiding unexpected outages, maintenance crew utilization, and equipment replacement cost. In this paper, a PMUdata-driven framework is proposed for classification of disruptive events in distribution networks. The two disruptive events, i.e., malfunctioned capacitor bank switching and malfunctioned regulator on-load tap changer (OLTC) switching are considered and distinguished from the normal abrupt load change in distribution grids. The performance of the proposed framework is verified using the simulation of the events in the IEEE 13-bus distribution network. The event classification is formulated using two different algorithms as; i) principle component analysis (PCA) together with multi-class support vector machine (SVM), and ii) autoencoder along with softmax classifier. The results demonstrate the effectiveness of the proposed algorithms and satisfactory classification accuracies. Index Terms—Classification, PMU, Event Detection, SVM I. INTRODUCTION With the advent of advanced metering devices, e.g., phasor measurement units (PMUs), and micro-PMUs (μPMUs), power transmission and distribution networks have become more intelligent, reliable and efficient [1], [2]. Most of the early measurement devices have had the limited measuring capacity, while the multi-functional capabilities of PMUs have made them important monitoring assets to power networks. PMUs have been used for several monitoring and control applications in transmission grids, e.g., state estimation [3], dynamic stability assessment [4], event diagnostics and classification [5]. Recently, the use of PMUs and μPMUs for several monitoring and control applications in distribution networks have been introduced. In [6], a fault location algorithm in distribution networks have been proposed using PMU data. In [7], PMUs are utilized in distribution networks with distributed generation for three different applications, namely, state estimation, protection and instability prediction. In [8], PMUs are deployed in distribution grids for measurement of synchronized harmonic phasors. Additionally, PMUs can help power grids to restore quicker in case of cutting off the energy supply by providing voltage, current and frequency measurements for reclosing the circuit breakers, e.g., 2008 Florida blackout [9]. In [10], PMU deployment for state estimation in active distribution networks is discussed. Reference [11] presents the use of PMU data for abnormal situation detection in distribution networks. Event detection is an ongoing field of study, and can be a challenging task due to different types of correlated events in distribution grids, e.g., switching vs. load changing. Therefore, distinguishing the disruptive events from one another, and differentiating them from a normal condition of the network, requires advanced data-driven frameworks. Several different techniques are proposed in the literature for classification of events in power networks. In [12], a probabilistic neural network along with S-transform is utilized to classify power quality disturbances. Partial discharge pattern recognition is conducted by applying fuzzy decision tree method [13], and sparse representation and artificial neural network [14]. Support vector machine (SVM) is a widely used method for event classification, e.g., fault location [15], power system security assessment [16], and transient stability analysis [17]. Accurate distribution event detection and classification results in accurate preventive maintenance scheduling of the critical assets based on the warning signs of the pending electrical failures. Preventive maintenance is a beneficial task in terms of time, equipment replacement cost, maintenance crew utilization, avoiding unexpected outages, and consequently, extending the life of the critical assets. Real-time data analytics can help to detect multiple failures, along with offering online monitoring of feeder operations. Therefore, it can provide utilities with useful information about faulty equipment in particular parts of the network. In [18], the authors have used highly sensitive waveform recorders for gathering data and improving feeders’ visibility and operational efficiency. The collected data from waveform recorders is used for incipient equipment failures detection on distribution feeders [19]. In [20], a data-driven methodology is presented for classification of five disruptive events, i.e., cable failure, hot-line clamp failure, vegetation intrusion resulting in frequent fault, fault induced conductor slap, and capacitor controller malfunction. In this paper, we propose a framework for classification of two disruptive events from the normal abrupt load change in distribution networks using PMU data. These classes are malfunctioned capacitor bank switching, malfunctioned regulator on-load tap changer (OLTC) switching, and abrupt load changing. The disruptive events may not cause immediate failure, however, they will cause permanent equipment failure and expensive replacement cost over time. The classification of these events prioritizes the preventive maintenance scheduling and leads to higher life expectancy of distribution assets. In this paper, the classification algorithms are developed using (1) principal component analysis (PCA) along with SVM, and (2) autoencoder along with softmax layer. The rest of this paper is organized as follows, in section II, the problem statement and the proposed methodology is presented. Section III presents the simulation results, and the conclusion and future works are presented in section IV. using two different classification methods as; (1) PCA along with multi-class SVM, and (2) a neural network-based toolbox, i.e., autoencoder, along with a softmax layer classifier. Figure 1 illustrates the flowchart of these two methods which is discussed in detail in the next subsections. II. PROBLEM STATEMENT AND METHODOLOGY Data analytics plays a major role in power system post-event analysis such as event diagnosis and preventive maintenance scheduling. These applications are helpful in terms of saving maintenance time and cost, and leads to preventing unexpected outages due to permanent failure of critical assets. In this paper, two different disruptive events, i.e., malfunctioned capacitor bank switching and malfunctioned regulator OLTC switching, along with a normal abrupt load changing, are categorized as three different classes. Malfunctioned capacitor bank switching events are caused by failure in mechanical switches and can happen in less than 2 cycles, i.e., 32 msec. The malfunctioned regulator OLTC switching can be caused due to ageing and degradation of the selector switches. In a malfunctioned regulator OLTC switching, the tap is dislocated, and after a while relocated to its original position. In this paper, we propose a PMU-data-driven classification framework to distinguish these two disruptive events from the normal abrupt load changing in distribution networks. The rationale is that normal abrupt load changing has similar signatures on PMU data and advanced data analytics is required to distinguish the disruptive events from the normal load changing. The classification input are six different features that have been extracted from PMU data as; change of voltage magnitude between two consecutive samples (v(n+1)-v(n)), change of voltage angle between two consecutive samples (δv(n+1)-δv(n), current magnitude (p.u.), current angle (degree), change of current magnitude between two consecutive samples (i(n+1)i(n)), and change of current angle between two consecutive samples (δi(n+1)-δi(n)). Moreover, since these features change over time, and we have a feature matrix, shown in (1), as opposed to a feature vector. (1) 𝑓1 𝑋=[ ⋮ (1) 𝑓𝑛 (𝑝) ⋯ 𝑓1 ⋱ ⋮ ]  (𝑝) ⋯ 𝑓𝑛 𝑛×𝑝 (𝑡) Fig. 1. a) PCA+SVM method (left), and b) Autoencoder+Softmax (right) events classification flowcharts A. PCA+SVM event classification algorithm The extracted features from PMU data change over time, thus the features are presented as the feature matrix using (1). On the other hand the input to the SVM classifier is a vector and the matrices must be transformed into vectors. In this paper, PCA is utilized to obtain the dominant Eigen values of the feature matrix and used as the input to a multi-class SVM. The summary of PCA and SVM is presented below. A.1 Principal Component Analysis (PCA) Principal component analysis (PCA) is a technique for reducing the dimensionality of the problem, and extracting the dominant features of the systems. In addition, it can be used in pattern recognition in high-dimensional data and it has a wide  range of applications, e.g. image processing and face recognition, and data mining. For further study refer to [25]. where X is the feature matrix and 𝑓𝑖 is the value of feature i at time t. In this paper, we consider the PMU data with two reporting rates as i) 60 samples per second (sps), e.g. SEL 651 [21], and ii) 120 sps, e.g. micro PMUs (μPMUs) developed at University of California, Berkeley [22]. Additionally, it is assumed that the capacitor bank switching takes about 1 cycle (16.67 ms) [23], and on-load tap changer switching takes about 30-200 ms [24] and the PMUs are capable of capturing this event. In this paper, we present disruptive events classification A.2 Support Vector Machine (SVM) Support vector machine (SVM) is a supervised classification algorithm that uses linear or nonlinear hyper planes for separating classes from each other. The goal is to maximize the margin between the hyper planes, and therefore, the problem is formulated as a constrained optimization problem. Moreover, SVM can be applied to non-linear sets of data incorporating a method called kernel trick which maps original data into a higher-dimensional space. We used Gaussian kernel function in this paper. Additional discussion can be found in [26]. For multi-class classification, several algorithms have been proposed in the past. In this paper, a one-against-all algorithm is used to classify the events with respect to all the other classes using several binary SVMs. B. Autoencoder+ Softmax event classification algorithm As the second method, the event classification is carried out using a neural network-based toolbox, i.e., autoencoder, and softmax layer. In this method, the new feature matrix is first normalized using the mean and standard deviation of the historical feature matrices. The rows of the normalized feature matrix are then stacked on top of each other to create an input vector. The feature vector is then used as the input to the autoencoder layer for feature compression of the input vector. The softmax layer is then carried out using the compressed vector to classify the events. Fig. 1.b shows the flowchart and the summary is presented in the following. B.1 Autoencoder An autoencoder belongs to the artificial neural network family and is used as an unsupervised learning algorithm. It takes the data as input and tries to reconstruct the data through two layers of coding and decoding. The learning process is carried out using back propagation algorithms and the goal of training is to minimize the reconstruction error. An autoencoder takes the 𝑥 ∈ 𝑅𝑑 as input and maps it onto ′ 𝑧 ∈ 𝑅𝑑 as 𝑧 = 𝑠1 (𝑊𝑥 + 𝑏) (3) Where s1, W, and b are the element-wise sigmoid function, the weight matrix, and the bias term, respectively. Then, 𝑧 is mapped onto the 𝑅𝑑 to reconstruct the input using 𝑥′ = 𝑠2 (𝑊′ 𝑧 + 𝑏′ ) (4) Where 𝑠2 , 𝑊 ′ , and 𝑏 ′ are the element-wise sigmoid function, the weight matrix, and the bias term, respectively. [27]. B.2 Softmax Classifier Softmax function is a generalization of logistic function that output a multiclass probability distribution as opposed to a binary probability distribution. It serves as the output layer of the autoencoder. It takes an m-vector x as input and outputs a y-vector of real values between 0 and 1. It is defined as 𝑓𝑗 (𝑥) = 𝑥 𝑒 𝑗 𝑥𝑖 ∑𝑚 𝑖=1 𝑒 for i=1,2,..m (5) Where 𝑓𝑗 (𝑥) is the predicted probability for class j. Further discussion can be found in [28]. III. SIMULATION AND RESULTS In this paper, the proposed disruptive event classification is evaluated using the simulation studies. The PMU data for classification is generated by simulating the IEEE 13-node distribution system, as shown in Fig. 2. This distribution network has three different voltage levels as 115 kV, 4.16 kV, and 0.48 kV. The downstream network is connected via a 5000 kVA transformer to the upstream network. In this network there are (1) three single-phase regulators between bus 650 and bus 632, (2) a transformer between 633 bus and 634 bus, (3) a three-phase capacitor bank at bus 675, (4) a single-phase capacitor at bus 611, and (5) 15 distributed loads at different buses. We assume that one PMU is placed at bus 671 measuring voltage at this bus and current from bus 637 to bus 671. A Gaussian noise with zero mean and standard deviation of 1% of the measured values is then added to the voltage and current measurements (magnitudes and angles) to model the PMU inaccuracies. Fig. 2. The IEEE 13 Node Test Feeder with one PMU [29] Figure 3 shows the PMU voltage magnitude over one second, corresponding to three different classes, i.e. malfunctioned capacitor bank switching, malfunctioned OLTC switching of the voltage regulator and abrupt load changing. These figures demonstrate the voltage magnitude measurement of phase a. (a) (b) (c) Fig. 3. PMU voltage magnitudes for three different classes a) malfunctioned capacitor bank switching, b) malfunctioned OLTC switching c) abrupt load changing In order to create enough experiments for class 1, the malfunctioned three-phase capacitor bank switching event at bus 671 is simulated at different loading level. There are 15 different loads in the system, and for each of these loads, 10 different loading ranging from 50% up to 95%, with 5% intervals is considered. Therefore, 150 different experiments are simulated for class 1. Similarly, the same number of experiments is simulated for the malfunctioned OLTC switching event as the second class. Class three corresponds to normal abrupt load changing and it is assumed that one of the loads has a sudden change at a time. The abrupt load changing is simulated using 5%, 10%, 15%, 20%, and 25% increase or decrease of active and reactive power. Therefore, the total number of all class 3 experiments is 150, and we have generated 450 total number of experiments in all three classes. The proposed multi-class classification algorithms are then trained and evaluated using the simulated PMU data. The classifiers are trained using x percent of the data, i.e., selected randomly x  (10, 90), and the remaining data set is used to evaluate the classification accuracies as 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑎𝑐𝑐𝑢𝑟𝑎𝑡𝑒 𝑐𝑙𝑎𝑠𝑠𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛 𝑡𝑜𝑡𝑎𝑙 𝑛𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑡𝑒𝑠𝑡 𝑐𝑎𝑠𝑒𝑠 (6) accuracies increase as more training data is used. Additionally, PMUs with 60 sps results in (a) 62% accuracy with 20% of data used for training, and (b) 84% accuracy with 90% of data used for training. While PMUs with 120 sps results in higher accuracies for the same percentage of training experiments, As the results verify, higher sampling rates leads to better capturing of the events, and consequently, better classification of the disruptive events. In this paper, we have gradually increased the percentage of the training data set and evaluated the confusion matrices and the classification accuracies. Table 1 and 2 demonstrate the confusion matrices for the scenario with 50% of data used for training and the rest for evaluation, using PCA+SVM and autoencoder+softmax, respectively. Table 1. Confusion matrix in PCA+SVM method, with 50% of data used for training, and 60 sps PMU Predicted Class 1 Actual Class 1 Class 2 Class 3 53 (23.56%) 8 (3.56%) 10 (4.44%) Class 2 7 (3.11%) 54 (24%) 2 (0.88%) Class 3 12 (5.33%) 6 (2.67%) 62 (27.56%) Nonclassified 3 (1.33%) 7 (3.11%) 1 (0.44%) Table 2. Confusion matrix in Autoencoder+Softmax method, with 50% of data used for training, and 60 sps PMU Predicted Class 1 Actual Class 1 Class 2 Class 3 63 (28%) 9 (4%) 5 (2.22%) Class 2 Class 3 6 (2.66%) 59 (26.22%) 3 (1.33%) 6 (2.66%) 7 (3.11%) 67 (29.77%) Nonclassified 0 (0%) 0 (%) 0 (%) As it is observed in the first row of Table 1, from 75 experiments corresponding to class 1, only 53 experiment are accurately classified, and 7, 12, and 3 experiments are misclassified as class 2, class 3, and non-classified, respectively. The percentage next to each number is the percentage of each number with respect to all the test cases. The classification accuracy is then calculated using leaveone-out scenario which is a standard test for evaluation of any machine learning technique [30]. In this scenario it is assumed that all the experiments except one, i.e., 449 of the experiments, are used for training the classifiers and the only remaining experiment is tested using the trained classifiers. This process is carried out for the number of all experiments, i.e., 450 times, starting from the first experiment up to the last one. The accuracy is then calculated as the number of accurately classified experiment divided by the total number of experiments. The leave-one-out accuracies are 86.1% and 91.2% using PCA + SVM, and Autoencoder + Softmax, respectively which shows the better performance of the later method for disruptive event classification in distribution grids. Finally the classification accuracies are calculated for different percentages of training cases, starting from 20% to 90%, with 10% intervals. Fig. 4 shows the results for PCA+SVM method for the two different sampling rates 60 sps and 120 sps. As it is observed from Fig. 4, the classification Fig. 4. Classification accuracy of PCA + SVM method for different training percentage and two different sampling rates Fig. 5 shows the results using autoencoder+softmax method for two different sampling rates, i.e., 60 and 120 sps. Similar to PCA+SVM, as the training percentage increases, the accuracy increases. Additionally, as the sampling rate gets higher, the classification accuracies improve for the same percentage of the training cases. Furthermore, Figs. 4 and 5 are compared and it is verified that autoencoder+softmax method outperforms the PCA+SVM method in all different scenarios. Fig. 5. Classification accuracy of autoencoder and softmax method for different training percentage and two different sampling rates IV. CONCLUSION AND FUTURE WORKS Data-driven event detection in distribution grids provides essential operational and maintenance tools for next-generation smart grids with advanced measurement devices, e.g., microphasor measurement units (μPMUs). This paper presents a new framework for classification of disruptive events using PMU data in distribution grids. Two disruptive events are defined as malfunctioned capacitor bank switching and malfunctioned regulator on-load tap changer (OLTC) switching which provide similar PMU data pattern to normal abrupt load changes. The end result of this paper will provide a new framework that can be used for preventive maintenance scheduling of critical assets in distribution grids. In this paper, the event classification is developed using two multi-class classification algorithms for distinguishing between the two disruptive events and the normal load changing event. The first method is based on principal component analysis (PCA) along with multi-class support vector machine (SVM), and the second method is designed using autoencoder accompanied by the softmax layer classifier. The data for training and testing is provided using simulating the IEEE 13-bus distribution network with a PMU with 60 or 120 samples per second (SPS). The classification results verify the acceptable performance of the proposed framework for distinguishing the two disruptive events from the normal load change. The results also show the superiority of the second method over the first method. For future work, the authors will implement the proposed framework on larger standard networks, with more disruptive events, e.g. external voltage disturbance, voltage sag, reconfiguration. Furthermore, early-fusion, late-fusion techniques will be utilized for concatenation of several PMU data for wide-area disruptive event classification and localization frameworks. [14] [15] [16] [17] [18] [19] [20] REFERENCES [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] B. Singh, N.K. Sharma, A.N. Tiwari, K.S. Verma, and S.N. Singh, “Applications of phasor measurement units (PMUs) in electric power system networks incorporated with FACTS controllers” International Journal of Engineering, Science and Technology, Vol. 3, no. 3, pp.64-82, Jun 2011 N.H. Abbasy, and H.M. Ismail, “A unified approach for the optimal PMU location for power system state estimation” IEEE Transactions on power systems, Vol. 24, no. 2, pp.806-813, May 2009 R. Jalilzadeh Hamidi, H. Khodabandehlou, H. Livani, and M. Sami Fadali, “Application of distributed compressive sensing to power system state estimation,” NAPS, pp. 1-6, Oct., 2015. M. Zhou, V.A. Centeno, J.S. Thorp, A.G. Phadke, “An alternative for including phasor measurements in state estimators”, IEEE transactions on power systems,Vol. 21, no. 4,1930-7, Nov 2006. A.L. Liao, E.M. Stewart, E.C. Kara, “Micro-synchrophasor data for diagnosis of transmission and distribution level events”, Transmission and Distribution Conference and Exposition (T&D), IEEE PES, pp. 1-5, May 2016 M. Majidi, M. Etezadi-Amoli, and M.S. Fadali, “A Novel Method for Single and Simultaneous Fault Location in Distribution Networks,” IEEE Transactions on Power System. Vol. 30, no. 6, pp. 3368-3376, Nov. 2015. F. Ding, C.D. Booth, “Applications of PMUs in Power Distribution Networks with Distributed Generation”, Universities' Power Engineering Conference (UPEC), Proceedings of 2011 46th International, pp. 1-5, Sep 2011. A. Carta, N. Locci, C. Muscas, “A PMU for the measurement of synchronized harmonic phasors in three-phase distribution networks”, IEEE Transactions on instrumentation and measurement, Vol. 58, no. 10, pp. 3723-30, Oct. 2009. M. Wache, D.C. Murray, “Application of synchrophasor measurements for distribution networks”, IEEE Power and Energy Society General Meeting, pp. 1-4, Jul 24, 2011. J. Liu, J. Tang, F. Ponci, A. Monti, C. Muscas, P.A. Pegoraro, “Trade-offs in PMU deployment for state estimation in active distribution grids” IEEE transactions on Smart Grid, Vol.3, no. 2, pp. 915-24, Jun, 2012. D. Phillips, T. Overbye, “Distribution system event detection and classification using local voltage measurements”, Power and Energy Conference at Illinois (PECI), pp. 1-4, Feb 28, 2014. S. Mishra, C.N. Bhende, B.K. Panigrahi, “Detection and classification of power quality disturbances using S-transform and probabilistic neural network”, IEEE Transactions on power delivery, Vol. 23, no. 1, pp. 2807, Jan 2008. T.K. Abdel-Galil, R.M. Sharkawy, M.M. Salama, R. Bartnikas, “Partial discharge pattern classification using the fuzzy decision tree approach”, [21] [22] [23] [24] [25] [26] [27] [28] [29] [30] IEEE Transactions on Instrumentation and Measurement, Vol. 54, no.6, pp. 2258-63, Dec. 2005. M.Majidi, M.S.Fadali, M.Etezadi-Amoli, and M.Oskuoee, “Partial Discharge Pattern Recognition via Sparse Representation and ANN,” IEEE Transactions on Dielectrics and Electrical Insulation. Vol. 22, no. 2, pp. 1061-1070, April 2015. H. Livani, C.Y. Evrenosoğlu, “A fault classification and localization method for three-terminal circuits using machine learning”, IEEE Transactions on Power Delivery, Vol. 28, no. 4, pp. 2282-90, Oct 2013. S. Kalyani, K.S. Swarup, “Classification and assessment of power system security using multiclass SVM”, IEEE Transactions on Systems, Man, and Cybernetics, Part C (Applications and Reviews), Vol. 41, no. 5, pp. 753-8, Sep. 2011. L.S. Moulin, A.A. Da Silva, M.A. El-Sharkawi, R.J. Marks, “Support vector machines for transient stability analysis of large-scale power systems”, IEEE Transactions on Power Systems, Vol. 19, no. 2, pp. 81825, May 2004. J.A. Wischkaemper, C.L. Benner, B.D. Russell, K.M. Manivannan, “Waveform analytics-based improvements in situational awareness, feeder visibility, and operational efficiency”, IEEE PES T&D Conference and Exposition, pp. 1-5, Apr 14, 2014. J.S. Bowers, A. Sundaram, C.L. Benner, B.D. Russell, “Outage avoidance through intelligent detection of incipient equipment failures on distribution feeders”, Power and Energy Society General MeetingConversion and Delivery of Electrical Energy in the 21st Century, pp. 17, Jul 20, 2008. J.A. Wischkaemper, C.L. Benner, B.D. Russell, “A new monitoring architecture for distribution feeder health monitoring, asset management, and real-time situational awareness”, IEEE PES Innovative Smart Grid Technologies (ISGT), pp. 1-7, Jan 16, 2012. SEL 611 relay (Online) Available: https://selinc.com/ Y. Zhou, R. Arghandeh, I. Konstantakopoulos I, S. Abdullah, A. von Meier, C.J. Spanos, “Abnormal event detection with high resolution micro-PMU data”, Power Systems Computation Conference (PSCC), pp. 1-7, Jun 2016. L.P.Hayes, (2005) . First Energy Corp, Akron, OH, [Online] Available: http://castlepowersolutions.biz/Components/Joslyn%20HiVoltage/Technical%20Papers/First_Energy_Corporation_Paper_Doble_ Conference_2005_VBM_Failures.pdf R. Jongen, E. Gulski, J. Erbrink, (2014), “Degradation Effects and Diagnosis of On-load Tap Changer in Power Transformers”, [Online] Available: http://www.unistuttgart.de/ieh/forschung/veroeffentlichungen/2014_Wil d_CMD_Korea_Degradation_Effects_and_Diagnosis_of_Onload_Tap_Changer....pdf. S. Wold, K. Esbensen, P. Geladi, “Principal component analysis”, Chemometrics and intelligent laboratory systems, Vol. 2, no. 1-3, pp. 3752, Aug 1, 1987. H. Livani, C.Y. Evrenosoğlu, “A fault classification method in power systems using DWT and SVM classifier”, Transmission and Distribution Conference and Exposition (T&D), IEEE PES, pp. 1-5, May 7, 2012. P. Vincent, H. Larochelle, Y. Bengio, P.A. Manzagol, “Extracting and composing robust features with denoising autoencoders”, Proceedings of the 25th international conference on Machine learning, pp. 1096-1103, Jul 5, 2008. R.S. Sutton, A.G. Bart, “Reinforcement learning: An introduction” Vol. 1, no. 1, Cambridge: MIT press, Mar 1, 1998. W.H. Kersting, “Radial distribution test feeders”, Power Engineering Society Winter Meeting, Vol. 2, pp. 908-912, 2001. G.C. Cawley, N.L. Talbot, “Efficient leave-one-out cross-validation of kernel fisher discriminant classifiers”, Pattern Recognition, Vol. 36, no. 11, pp. 2585-92, Nov. 30, 2003. 

Implicit Regularization in Nonconvex Statistical Estimation: Gradient Descent Converges Linearly for Phase Retrieval, Matrix Completion and Blind Deconvolution arXiv:1711.10467v2 [cs.LG] 14 Dec 2017 Cong Ma∗ Kaizheng Wang∗ November 2017; Yuejie Chi† Yuxin Chen‡ Revised December 2017 Abstract Recent years have seen a flurry of activities in designing provably efficient nonconvex procedures for solving statistical estimation problems. Due to the highly nonconvex nature of the empirical loss, stateof-the-art procedures often require proper regularization (e.g. trimming, regularized cost, projection) in order to guarantee fast convergence. For vanilla procedures such as gradient descent, however, prior theory either recommends highly conservative learning rates to avoid overshooting, or completely lacks performance guarantees. This paper uncovers a striking phenomenon in nonconvex optimization: even in the absence of explicit regularization, gradient descent enforces proper regularization implicitly under various statistical models. In fact, gradient descent follows a trajectory staying within a basin that enjoys nice geometry, consisting of points incoherent with the sampling mechanism. This “implicit regularization” feature allows gradient descent to proceed in a far more aggressive fashion without overshooting, which in turn results in substantial computational savings. Focusing on three fundamental statistical estimation problems, i.e. phase retrieval, low-rank matrix completion, and blind deconvolution, we establish that gradient descent achieves near-optimal statistical and computational guarantees without explicit regularization. In particular, by marrying statistical modeling with generic optimization theory, we develop a general recipe for analyzing the trajectories of iterative algorithms via a leave-one-out perturbation argument. As a byproduct, for noisy matrix completion, we demonstrate that gradient descent achieves near-optimal error control — measured entrywise and by the spectral norm — which might be of independent interest. Contents 1 Introduction 1.1 Nonlinear systems and empirical loss minimization . . . 1.2 Nonconvex optimization via regularized gradient descent 1.3 Regularization-free procedures? . . . . . . . . . . . . . . 1.4 Numerical surprise of unregularized gradient descent . . 1.5 This paper . . . . . . . . . . . . . . . . . . . . . . . . . 1.6 Notations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 4 5 6 6 7 8 2 Implicit regularization – a case study 2.1 Gradient descent theory revisited . . . . . . . . . . . . 2.2 Local geometry for solving random quadratic systems 2.3 Which region enjoys nicer geometry? . . . . . . . . . . 2.4 Implicit regularization . . . . . . . . . . . . . . . . . . 2.5 A glimpse of the analysis: a leave-one-out trick . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 9 10 11 12 12 . . . . . ∗ Department of Operations Research and Financial Engineering, Princeton University, Princeton, NJ 08544, USA; Email: {congm, kaizheng}@princeton.edu. † Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, PA 15213, USA; Email: yuejiechi@cmu.edu. ‡ Department of Electrical Engineering, Princeton University, Princeton, NJ 08544, USA; Email: yuxin.chen@princeton.edu. 1 3 Main results 13 3.1 Phase retrieval . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 3.2 Low-rank matrix completion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 3.3 Blind deconvolution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 4 Related work 19 5 A general recipe for trajectory analysis 20 5.1 General model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 5.2 Outline of the recipe . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 6 Analysis for phase retrieval 6.1 Step 1: characterizing local geometry in the RIC . . . . . . 6.1.1 Local geometry . . . . . . . . . . . . . . . . . . . . . 6.1.2 Error contraction . . . . . . . . . . . . . . . . . . . 6.2 Step 2: introducing the leave-one-out sequences . . . . . . . 6.3 Step 3: establishing the incoherence condition by induction 6.4 The base case: spectral initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 23 23 23 24 24 26 7 Analysis for matrix completion 7.1 Step 1: characterizing local geometry in the RIC . . . . . . 7.1.1 Local geometry . . . . . . . . . . . . . . . . . . . . . 7.1.2 Error contraction . . . . . . . . . . . . . . . . . . . . 7.2 Step 2: introducing the leave-one-out sequences . . . . . . . 7.3 Step 3: establishing the incoherence condition by induction 7.4 The base case: spectral initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 27 27 27 28 29 31 8 Analysis for blind deconvolution 8.1 Step 1: characterizing local geometry in the RIC . . . . . . 8.1.1 Local geometry . . . . . . . . . . . . . . . . . . . . . 8.1.2 Error contraction . . . . . . . . . . . . . . . . . . . . 8.2 Step 2: introducing the leave-one-out sequences . . . . . . . 8.3 Step 3: establishing the incoherence condition by induction 8.4 The base case: spectral initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31 32 32 33 34 35 36 9 Discussions A Proofs for phase retrieval A.1 Proof of Lemma 1 . . . A.2 Proof of Lemma 2 . . . A.3 Proof of Lemma 3 . . . A.4 Proof of Lemma 4 . . . A.5 Proof of Lemma 5 . . . A.6 Proof of Lemma 6 . . . 38 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46 46 47 48 49 50 50 B Proofs for matrix completion B.1 Proof of Lemma 7 . . . . . B.2 Proof of Lemma 8 . . . . . B.3 Proof of Lemma 9 . . . . . B.3.1 Proof of Lemma 22 . B.3.2 Proof of Lemma 23 . B.4 Proof of Lemma 10 . . . . . B.5 Proof of Lemma 11 . . . . . B.5.1 Proof of Lemma 24 . B.5.2 Proof of Lemma 25 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 52 55 56 61 62 64 66 68 69 . . . . . . 2 B.6 Proof of Lemma 12 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . B.7 Proof of Lemma 13 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . C Proofs for blind deconvolution C.1 Proof of Lemma 14 . . . . . . . . . . C.1.1 Proof of Lemma 26 . . . . . . C.1.2 Proof of Lemma 27 . . . . . . C.2 Proofs of Lemma 15 and Lemma 16 C.3 Proof of Lemma 17 . . . . . . . . . . C.4 Proof of Lemma 18 . . . . . . . . . . C.4.1 Proof of Lemma 28 . . . . . . C.4.2 Proof of Lemma 29 . . . . . . C.4.3 Proof of Claim (224) . . . . . C.5 Proof of Lemma 19 . . . . . . . . . . C.6 Proof of Lemma 20 . . . . . . . . . . C.7 Proof of Lemma 21 . . . . . . . . . . 70 74 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78 . 78 . 79 . 81 . 87 . 89 . 94 . 98 . 99 . 99 . 101 . 102 . 106 D Technical lemmas D.1 Technical lemmas for phase retrieval . . . . D.1.1 Matrix concentration inequalities . . D.1.2 Matrix perturbation bounds . . . . . D.2 Technical lemmas for matrix completion . . D.2.1 Orthogonal Procrustes problem . . . D.2.2 Matrix concentration inequalities . . D.2.3 Matrix perturbation bounds . . . . . D.3 Technical lemmas for blind deconvolution . D.3.1 Wirtinger calculus . . . . . . . . . . D.3.2 Discrete Fourier transform matrices D.3.3 Complex-valued alignment . . . . . . D.3.4 Matrix concentration inequalities . . D.3.5 Matrix perturbation bounds . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 108 108 108 108 109 109 111 117 120 120 121 125 129 131 1 Introduction 1.1 Nonlinear systems and empirical loss minimization A wide spectrum of science and engineering applications calls for solutions to a nonlinear system of equations. Imagine we have collected a set of data points y = {yj }1≤j≤m , generated by a nonlinear sensing system,
yj ≈ Aj x\ , 1 ≤ j ≤ m, where x\ is the unknown object of interest, and the Aj ’s are certain nonlinear maps known a priori. Can we reconstruct the underlying object x\ in a faithful yet efficient manner? Problems of this kind abound in information and statistical science, prominent examples including low-rank matrix recovery [KMO10a,CR09], robust principal component analysis [CSPW11, CLMW11], phase retrieval [CSV13, JEH15], neural networks [SJL17, ZSJ+ 17], to name just a few. In principle, it is possible to attempt reconstruction by searching for a solution that minimizes the empirical loss, namely, m X 2 (1) minimizex f (x) = yj − Aj (x) . j=1 Unfortunately, this empirical loss minimization problem is, in many cases, nonconvex, making it NP-hard in general. This issue of non-convexity comes up in, for example, several representative problems that epitomize the structures of nonlinear systems encountered in practice.1 • Phase retrieval / solving quadratic systems of equations. Imagine we are asked to recover an unknown object x\ ∈ Rn , but are only given the square modulus of certain linear measurements about the object, with all sign/phase information of the measurements missing. This arises, for example, in X-ray crystallography [CESV13], and in latent-variable models where the hidden variables are captured by the missing signs [CYC14]. To fix ideas, assume we would like to solve for x\ ∈ Rn in the following quadratic system of m equations
\ 2 yj = a> , 1 ≤ j ≤ m, j x where {aj }1≤j≤m are the known design vectors. One strategy is thus to solve the following problem minimize x∈Rn m
2 i2 1 Xh yj − a> . f (x) = j x 4m j=1 (2) • Low-rank matrix completion. In many scenarios such as collaborative filtering, we wish to make predictions about all entries of an (approximately) low-rank matrix M \ ∈ Rn×n (e.g. a matrix consisting of users’ ratings about many movies), yet only a highly incomplete subset of the entries are revealed to us [CR09]. For clarity of presentation, assume M \ to be rank-r (r  n) and positive semidefinite (PSD), i.e. M \ = X \ X \> with X \ ∈ Rn×r , and suppose we have only seen the entries \ Yj,k = Mj,k = (X \ X \> )j,k , (j, k) ∈ Ω within some index subset Ω of cardinality m. These entries can be viewed as nonlinear measurements about the low-rank factor X \ . The task of completing the true matrix M \ can then be cast as solving minimizeX∈Rn×r f (X) = n2 X 4m (j,k)∈Ω > Yj,k − e> j XX ek
2 , (3) where the ej ’s stand for the canonical basis vectors in Rn . 1 Here, we choose different pre-constants in front of the empirical loss in order to be consistent with the literature of the respective problems. In addition, we only introduce the problem in the noiseless case for simplicity of presentation. 4 • Blind deconvolution / solving bilinear systems of equations. Imagine we are interested in estimating two signals of interest h\ , x\ ∈ CK , but only get to collect a few bilinear measurements about them. This problem arises from mathematical modeling of blind deconvolution [ARR14, LLSW16], which frequently arises in astronomy, imaging, communications, etc. The goal is to recover two signals from their convolution. Put more formally, suppose we have acquired m bilinear measurements taking the following form yj = b∗j h\ x\∗ aj , 1 ≤ j ≤ m, where aj , bj ∈ CK are distinct design vectors (e.g. Fourier and/or random design vectors) known a priori. In order to reconstruct the underlying signals, one asks for solutions to the following problem minimizeh,x∈CK f (h, x) = m X j=1 1.2 2 yj − b∗j hx∗ aj . Nonconvex optimization via regularized gradient descent First-order methods have been a popular heuristic in practice for solving nonconvex problems including (1). For instance, a widely adopted procedure is gradient descent, which follows the update rule
xt+1 = xt − ηt ∇f xt , t ≥ 0, (4) where ηt is the learning rate (or step size) and x0 is some proper initial guess. Given that it only performs a single gradient calculation ∇f (·) per iteration (which typically can be completed within near-linear time), this paradigm emerges as a candidate for solving large-scale problems. The concern is: whether xt converges to the global solution and, if so, how long it takes for convergence, especially since (1) is highly nonconvex. Fortunately, despite the worst-case hardness, appealing convergence properties have been discovered in various statistical estimation problems; the blessing being that the statistical models help rule out ill-behaved instances. For the average case, the empirical loss often enjoys benign geometry, in a local region (or at least along certain directions) surrounding the global optimum. In light of this, an effective nonconvex iterative method typically consists of two stages: 1. a carefully-designed initialization scheme (e.g. spectral method); 2. an iterative refinement procedure (e.g. gradient descent). This strategy has recently spurred a great deal of interest, owing to its promise of achieving computational efficiency and statistical accuracy at once for a growing list of problems (e.g. [KMO10a, JNS13, CW15, SL16, CLS15,CC17,LLSW16,LLB17]). However, rather than directly applying gradient descent (4), existing theory often suggests enforcing proper regularization. Such explicit regularization enables improved computational convergence by properly “stabilizing” the search directions. The following regularization schemes, among others, have been suggested to obtain or improve computational guarantees. We refer to these algorithms collectively as Regularized Gradient Descent. • Trimming/truncation, which discards/truncates a subset of the gradient components when forming the descent direction. For instance, when solving quadratic systems of equations, one can modify the gradient descent update rule as

xt+1 = xt − ηt T ∇f xt , (5) where T is an operator that effectively drops samples bearing too much influence on the search direction. This strategy [CC17, ZCL16, WGE17] has been shown to enable exact recovery with linear-time computational complexity and optimal sample complexity. • Regularized loss, which attempts to optimize a regularized empirical risk


xt+1 = xt − ηt ∇f xt + ∇R xt , (6) where R(x) stands for an additional penalty term in the empirical loss. For example, in low-rank matrix completion R(·) imposes penalty based on the `2 row norm [KMO10a, SL16] as well as the Frobenius norm [SL16] of the decision matrix, while in blind deconvolution, it penalizes the `2 norm as well as certain component-wise incoherence measure of the decision vectors [LLSW16, HH17, LS17]. 5 Table 1: Prior theory for gradient descent (with spectral initialization) Vanilla gradient descent Regularized gradient descent sample iteration step sample iteration type of complexity complexity size complexity complexity regularization Phase trimming 1 n log n n n log 1 log 1 n retrieval [CC17, ZCL16] regularized loss n 1 nr7 r log  Matrix [SL16] n/a n/a n/a projection completion 1 2 2 nr r log  [CW15, ZL16] regularized loss & Blind 1 n/a n/a n/a Kpoly log m m log  deconvolution projection [LLSW16] • Projection, which projects the iterates onto certain sets based on prior knowledge, that is,

xt+1 = P xt − ηt ∇f xt , (7) where P is a certain projection operator used to enforce, for example, incoherence properties. This strategy has been employed in both low-rank matrix completion [CW15, ZL16] and blind deconvolution [LLSW16]. Equipped with such regularization procedures, existing works uncover appealing computational and statistical properties under various statistical models. Table 1 summarizes the performance guarantees derived in the prior literature; for simplicity, only orderwise results are provided. Remark 1. There is another role of regularization commonly studied in the literature, which exploits prior knowledge about the structure of the unknown object, such as sparsity to prevent overfitting and improve statistical generalization ability. This is, however, not the focal point of this paper, since we are primarily pursuing solutions to (1) without imposing additional structures. 1.3 Regularization-free procedures? The regularized gradient descent algorithms, while exhibiting appealing performance, usually introduce more algorithmic parameters that need to be carefully tuned based on the assumed statistical models. In contrast, vanilla gradient descent (cf. (4)) — which is perhaps the very first method that comes into mind and requires minimal tuning parameters — is far less understood (cf. Table 1). Take matrix completion and blind deconvolution as examples: to the best of our knowledge, there is currently no theoretical guarantee derived for vanilla gradient descent. The situation is better for phase retrieval: the local convergence of vanilla gradient descent, also known as Wirtinger flow (WF), has been investigated in [CLS15, WWS15]. Under i.i.d. Gaussian design and with near-optimal sample complexity, WF (combined with spectral initialization) provably achieves -accuracy
(in a relative sense) within O n log (1/ε) iterations. Nevertheless, the computational guarantee is significantly outperformed by the regularized version (called truncated Wirtinger flow [CC17]), which only requires
O log (1/ε) iterations to converge with similar per-iteration cost. On closer inspection,
the high computational cost of WF is largely due to the vanishingly small step size ηt = O 1/(nkx\ k22 ) — and hence slow movement — suggested by the theory [CLS15]. While this is already the largest possible step size allowed
in the theory published in [CLS15], it is considerably more conservative than the choice ηt = O 1/kx\ k22 theoretically justified for the regularized version [CC17, ZCL16]. The lack of understanding and suboptimal results about vanilla gradient descent raise a very natural question: are regularization-free iterative algorithms inherently suboptimal when solving nonconvex statistical estimation problems of this kind? 1.4 Numerical surprise of unregularized gradient descent To answer the preceding question, it is perhaps best to first collect some numerical evidence. In what follows, we test the performance of vanilla gradient descent for phase retrieval, matrix completion, and blind 6 100 100 100 10-5 10-5 10-5 10-10 10-10 10-10 10-15 0 100 200 300 (a) phase retrieval 400 500 10-15 50 100 150 200 250 300 350 400 (b) matrix completion 450 500 10-15 20 40 60 80 100 120 140 160 180 200 (c) blind deconvolution Figure 1: (a) Relative `2 error of xt (modulo the global phase) vs. iteration count for phase retrieval under i.i.d. Gaussian design, where m = 10n and ηt = 0.1. (b) Relative error of X t X t> (measured by k·kF , k·k , k·k∞ ) vs. iteration count for matrix completion, where n = 1000, r = 10, p = 0.1, and ηt = 0.2. (c) Relative error of ht xt∗ (measured by k·kF ) vs. iteration count for blind deconvolution, where m = 10K and ηt = 0.5. deconvolution, using a constant step size. For all of these experiments, the initial guess is obtained by means of the standard spectral method. Our numerical findings are as follows: • Phase retrieval. For each n, set m = 10n, take x\ ∈ Rn to be a random vector with unit norm, and i.i.d. generate the design vectors aj ∼ N (0, In ), 1 ≤ j ≤ m. Figure 1(a) illustrates the relative `2 error t \ t \ min{kx − x k2 , kx + x k2 }/kx\ k2 (modulo the unrecoverable global phase) vs. the iteration count. The results are shown for n = 20, 100, 200, 1000, with the step size taken to be ηt = 0.1 in all settings. • Matrix completion. Generate a random PSD matrix M \ ∈ Rn×n with dimension n = 1000, rank r = 10, and all nonzero eigenvalues equal to one. Each entry of M \ is observed independently with probability p = 0.1. Figure 1(b) plots the relative error X t X t> − M \ / M \ vs. the iteration count, where |||·||| can either be the Frobenius norm k·kF , the spectral norm k · k, or the entrywise `∞ norm k · k∞ . Here, we pick the step size as ηt = 0.2. i.i.d. • Blind deconvolution. For each K ∈ {20, 100, 200, 1000} and m = 10K, generate the design vectors aj ∼ N (0, 21 IK ) + iN (0, 12 IK ) for 1 ≤ j ≤ m independently,2 and the bj ’s are drawn from a partial Discrete Fourier Transform (DFT) matrix (to be described in Section 3.3). The underlying signals h\ , x\ ∈ CK are produced as random vectors with unit norm. Figure 1(c) plots the relative error kht xt∗ −h\ x\∗ kF /kh\ x\∗ kF vs. the iteration count, with the step size taken to be ηt = 0.5 in all settings. In all of these numerical experiments, vanilla gradient descent enjoys remarkable linear convergence, always yielding an accuracy of 10−5 (in a relative sense) within around 200 iterations. In particular, for the phase retrieval problem, the step size is taken to be ηt = 0.1 although we vary the problem size from n = 20 to n = 1000. The consequence is that the convergence rates experience little changes when the problem sizes vary. In comparison, the theory published in [CLS15] seems overly pessimistic, as it suggests a diminishing step size inversely proportional to n and, as a result, an iteration complexity that worsens as the problem size grows. In short, the above empirical results are surprisingly positive yet puzzling. Why was the computational efficiency of vanilla gradient descent unexplained or substantially underestimated in prior theory? 1.5 This paper The main contribution of this paper is towards demystifying the “unreasonable” effectiveness of regularizationfree nonconvex iterative methods. As asserted in previous work, regularized gradient descent succeeds by properly enforcing/promoting certain incoherence conditions throughout the execution of the algorithm. In contrast, we discover that 2 Here and throughout, i represents the imaginary unit. 7 Table 2: Prior theory vs. our theory for vanilla Prior theory sample iteration complexity complexity Phase retrieval n log n n log (1/ε) Matrix completion n/a n/a Blind deconvolution n/a n/a gradient descent (with spectral initialization) Our theory step sample iteration step size complexity complexity size 1/n n log n log n log (1/ε) 1/ log n n/a nr3 poly log n log (1/ε) 1 n/a Kpoly log m log (1/ε) 1 Vanilla gradient descent automatically forces the iterates to stay incoherent with the measurement mechanism, thus implicitly regularizing the search directions. This “implicit regularization” phenomenon is of fundamental importance, suggesting that vanilla gradient descent proceeds as if it were properly regularized. This explains the remarkably favorable performance of unregularized gradient descent in practice. Focusing on the three representative problems mentioned in Section 1.1, our theory guarantees both statistical and computational efficiency of vanilla gradient descent under random designs and spectral initialization. With near-optimal sample complexity, to attain -accuracy,
• Phase retrieval (informal): vanilla gradient descent converges in O log n log 1 iterations;
• Matrix completion (informal): vanilla gradient descent converges in O log 1 iterations;
• Blind deconvolution (informal): vanilla gradient descent converges in O log 1 iterations. In words, gradient descent provably achieves (nearly) linear convergence in all of these examples. Throughout this paper, an algorithm is said to converge (nearly) linearly to x\ in the noiseless case if the iterates {xt } obey dist(xt+1 , x\ ) ≤ (1 − c) dist(xt , x\ ), ∀t ≥ 0 for some 0 < c ≤ 1 that is (almost) independent of the problem size. Here, dist(·, ·) can be any appropriate discrepancy measure. As a byproduct of our theory, gradient descent also provably controls the entrywise empirical risk uniformly across all iterations; for instance, this implies that vanilla gradient descent controls entrywise estimation error for the matrix completion task. Precise statements of these results are deferred to Section 3 and are briefly summarized in Table 2. Notably, our study of implicit regularization suggests that the behavior of nonconvex optimization algorithms for statistical estimation needs to be examined in the context of statistical models, which induces an objective function as a finite sum. Our proof is accomplished via a leave-one-out perturbation argument, which is inherently tied to statistical models and leverages homogeneity across samples. Altogether, this allows us to localize benign landscapes for optimization and characterize finer dynamics not accounted for in generic gradient descent theory. 1.6 Notations Before continuing, we introduce several notations used throughout the paper. First of all, boldfaced symbols are reserved for vectors and matrices. For any vector v, we use kvk2 to denote its Euclidean norm. For any matrix A, we use σj (A) and λj (A) to denote its jth largest singular value and eigenvalue, respectively, and let Aj,· and A·,j denote its jth row and jth column, respectively. In addition, kAk, kAkF , kAk2,∞ , and kAk∞ stand for the spectral norm (i.e. the largest singular value), the Frobenius norm, the `2 /`∞ norm (i.e. the largest `2 norm of the rows), and the entrywise `∞ norm (the largest magnitude of all entries) of a matrix A. Also, A> , A∗ and A denote the transpose, the conjugate transpose, and the entrywise conjugate of A, respectively. In denotes the identity matrix with dimension n × n. The notation On×r represents the set of all n × r orthonormal matrices. The notation [n] refers to the set {1, · · · , n}. Also, we use Re(x) to denote the real part of a complex number x. Throughout the paper, we use the terms “samples” and “measurements” interchangeably. 8 Additionally, the standard notation f (n) = O (g(n)) or f (n) . g(n) means that there exists a constant c > 0 such that |f (n)| ≤ c|g(n)|, f (n) & g(n) means that there exists a constant c > 0 such that |f (n)| ≥ c |g(n)|, and f (n)  g(n) means that there exist constants c1 , c2 > 0 such that c1 |g(n)| ≤ |f (n)| ≤ c2 |g(n)|. Besides, f (n)  g(n) means that there exists some large enough constant c > 0 such that |f (n)| ≥ c |g(n)|. Similarly, f (n)  g(n) means that there exists some sufficiently small constant c > 0 such that |f (n)| ≤ c |g(n)|. 2 Implicit regularization – a case study To reveal reasons behind the effectiveness of vanilla gradient descent, we first examine existing theory of gradient descent and identify the geometric properties that enable linear convergence. We then develop an understanding as to why prior theory is conservative, and describe the phenomenon of implicit regularization that helps explain the effectiveness of vanilla gradient descent. To facilitate discussion, we will use the problem of solving random quadratic systems (phase retrieval) and Wirtinger flow as a case study, but our diagnosis applies more generally, as will be seen in later sections. 2.1 Gradient descent theory revisited In the convex optimization literature, there are two standard conditions about the objective function — strong convexity and smoothness — that allow for linear convergence of gradient descent. Definition 1 (Strong convexity). A twice continuously differentiable function f : Rn 7→ R is said to be α-strongly convex for α > 0 if ∇2 f (x)  αIn , ∀x ∈ Rn . Definition 2 (Smoothness). A twice continuously differentiable function f : Rn 7→ R is said to be β-smooth for β > 0 if ∇2 f (x) ≤ β, ∀x ∈ Rn . It is well known that for an unconstrained optimization problem, if the objective function f is both αstrongly convex and β-smooth, then vanilla gradient descent (4) enjoys `2 error contraction [Bub15, Theorem 3.12], namely, xt+1 − x\ k2 ≤  1− 2 β/α + 1  xt − x\ , 2 xt − x\ k2 ≤ and  1− 2 β/α + 1 t x0 − x\ 2 , t ≥ 0, (8) as long as the step size is chosen as ηt = 2/(α + β). Here, x\ denotes the global minimum. This immediately reveals the iteration complexity for gradient descent: the number of iterations taken to attain -accuracy (in a relative sense) is bounded by   β 1 O log . α  In other words, the iteration complexity is dictated by and scales linearly with the condition number — the ratio β/α of smoothness to strong convexity parameters. Moving beyond convex optimization, one can easily extend the above theory to nonconvex problems with local strong convexity and smoothness. More precisely, suppose the objective function f satisfies ∇2 f (x)  αI and ∇2 f (x) ≤ β over a local `2 ball surrounding the global minimum x\ :  Bδ (x) := x | kx − x\ k2 ≤ δkx\ k2 . (9) Then the contraction result (8) continues to hold, as long as the algorithm is seeded with an initial point that falls inside Bδ (x). 9 2.2 Local geometry for solving random quadratic systems To invoke generic gradient descent theory, it is critical to characterize the local strong convexity and smoothness properties of the loss function. Take the problem of solving random quadratic systems (phase retrieval) i.i.d. as an example. Consider the i.i.d. Gaussian design in which aj ∼ N (0, In ), 1 ≤ j ≤ m, and suppose without loss of generality that the underlying signal obeys kx\ k2 = 1. It is well known that x\ is the unique minimizer — up to global phase — of (2) under this statistical model, provided that the ratio m/n of equations to unknowns is sufficiently large. The Hessian of the loss function f (x) is given by i
2 1 Xh 3 a> − yj aj a> j x j . m j=1 m ∇2 f (x) = (10) • Population-level analysis. Consider the case with an infinite number of equations or samples, i.e. m → ∞, where ∇2 f (x) converges to its expectation. Simple calculation yields that  

E ∇2 f (x) = 3 kxk22 In + 2xx> − In + 2x\ x\> . It it straightforward to verify that for any sufficiently small constant δ > 0, one has the crude bound   In  E ∇2 f (x)  10In , ∀x ∈ Bδ (x) : x − x\ 2 ≤ δ x\ 2 , meaning that f is 1-strongly convex and 10-smooth within a local ball around x\ . As a consequence, when we have infinite samples and an initial guess x0 such that kx0 − x\ k2 ≤ δ x\ 2 , vanilla gradient descent with a constant step size converges to the global minimum within logarithmic iterations. • Finite-sample regime with m  n log n. Now that f exhibits favorable landscape in the population level, one thus hopes that the fluctuation can be well-controlled so that the nice geometry carries over to the finite-sample regime. In the regime where m  n log n (which is the regime considered in [CLS15]), the local strong convexity is still preserved, in the sense that ∇2 f (x)  (1/2) · In , ∀x : x − x\ 2 ≤ δ x\ 2 occurs with high probability, provided that δ > 0 is sufficiently small (see [Sol14, WWS15] and Lemma 1). The smoothness parameter, however, is not well-controlled. In fact, it can be as large as (up to logarithmic factors)3 ∇2 f (x) . n even when we restrict attention to the local `2 ball (9) with δ > 0 being a fixed small constant. This means that the condition number β/α (defined in Section 2.1) may scale as O(n), leading to the step size recommendation ηt  1/n,
and, as a consequence, a high iteration complexity O n log(1/) . This underpins the analysis in [CLS15]. In summary, the geometric properties of the loss function — even in the local `2 ball centering around the global minimum — is not as favorable as one anticipates, in particular in view of its population counterpart. A direct application of generic gradient descent theory leads to an overly conservative step size and a pessimistic convergence rate, unless the number of samples is enormously larger than the number of unknowns. Remark 2. Notably, due to Gaussian designs, the phase retrieval problem enjoys more favorable geometry compared to other nonconvex problems. In matrix completion and blind deconvolution, the Hessian matrices are rank-deficient even at the population level. In such cases, the above discussions need to be adjusted, e.g. strong convexity is only possible when we restrict attention to certain directions. 3 To demonstrate this, take x = x\ + (δ/ka1 k2 ) · a1 in (10), one can easily verify that, with high probability, ∇2 f (x) ≥ 2 2 2 − y1 a1 a> 1 /m − O(1) & δ n /m  δ n/log n. 2 3(a> 1 x) 10 H1 ! N (1, 1) ! N (1, m 1) X 1 2 > 2 m minimize (a2i⌘x) yi ⇣ ⌘ x f (x)2 = ⇣ 1 X 2 > 2 1 (x 1) ✓ ◆ (x 1) minimizex f (x) = (ai x) p1yi exp m i=1 p✓ m exp 2 f (x 2 | H1 ) fmX1i=1 (xX | H1 ) > 2⇡ 1 2⇡ X 2 = = exp x == = exp x minimizex L(x) f (x) (a= x)2L(x) yi = 2 x2 p1 exp (x | H0 ) p1 exp fX x fXm (x | H0 ) i 2 2 2 2⇡ 2⇡ 2.3 Which region enjoys i=1 nicers geometry? H1 s H1 H21 H1 2 H1 n region surrounding \ a large diameter that enjoys much nicer \ Interestingly, our theory with 2identifies 2 x >2aR s.t. a> x = ax> x > 1 > > , 1> 1  i  m n > \local i i x 2 R s.t. ai x does = not aimimic x ,an ⇠`2 ball, 1  () i rather, m L(x) xL(x) +⇠log geometry. This but the intersection of ⇠an() `2 ball and x a polytope.+ log ⇠ s region < < 2< < 2 ⇣ For phase X We term it the region X of incoherence and (RIC). retrieval, the RIC includes all ⇣ H0 contraction H0> H 2 H0points 0 > > \ \> 2 2 2 > > minimize f\ (X) => 1 \ eei m XX ej ei X X ej ∈R X xx 2 Rnnobeying s.t. a> =e> i minimize f (X) = XX X \> X j i x i x ej , ei X i a Pe,MAP = (i,j)2⌦ P\⇥ (✓0 )↵ + P⇥ (✓P1 )e,MAP = P⇥ (✓0 )↵ + P⇥ (✓1 ) \ (i,j)2⌦ X ⇣ x−x 2 ≤δ x 2 and (11a) 2 p\
> m ⇣ > \> m ⇣ e> XX↵ minimizeX ↵f (X) = X e X xj − x\eX .X log , (11b) max a> n2 xe\ > 2 j j i i > > \ \> 2 >f (h, x) >= \ \>e XX e 1≤j≤m e X X e minimizeh,x f (h, x) =minimize e> XX e e X X h,x j j i i j j i (i,j)2⌦ i 1,k i 1,k i=1 S1,k =As(W + formalized c1 Fk )ei in i=1numerical constant. S1,k = + cprobability where δ > 0 is some small willk be Lemma 1, (W withkhigh the 1 Fk )e m ⇣ n⇥r > \ \> > Hessian matrix satisfies> X i 2,k > >2 \ \> s.t.find ej XX e (1/2) = eS X X ej(x) ,XX j) 2 ⌦e = (W +(i, c\2O(log F )e= X 2>R s.t. eki\> X = e(W ik> i , k (i, S2,k + cj)2 F2k ⌦ )ei 2,k j ·2,k I> j ∇2efe n) ·jIjnX ne minimizeh,x f (h, x) = ei> XX X X e i i \⇤ ⇤ RIC. ⇤ In words, \ \⇤ n⇤ h\the simultaneously for inbthe Hessian nearly the i=1 find h, x 2 Cn s.t.x find h x a = b x a , i xii⇤i  i i i h,Si x 2 C=i (W bF⇤i1kmatrix h)e aisim = b⇤i hwell-conditioned (with i m condition 1,k i i is.t. i 1,k + c ix i ai ,+ c1 F 1,k k 1 S = (W number bounded by O(log n)), as long as (i) the iterate is not very far from thekglobal (cf. (11a)), 1,k 1 minimizer k )e , 8pixel k i2 2,k 2 > 4 (W> + \c F\> i 2,k , way to8pixel S = )e (r f (x)) (ii) the iterate incoherent with respect to the sensing vectors (cf. (11b)). Another max 2,k k 2 k find X 2and Rn⇥r s.t. remains e> XX e = e X X e , (i, j) 2 ⌦ S = (W + c F )e (r f (x)) i i 2,k k 2 k j j max 2 f (x)) interpret the incoherence (r condition (11b) is that the empirical risk needs to be well-controlled uniformly 2 max f (x))region. max (r ⇤ \ \⇤ See Figure an illustration the above find h,across x 2 all Cnsamples.s.t. b⇤i hi2(a) x>⇤i afori = hi = xi faof(S 1, S i) m i, \bW i a (x x ) find X 2 Rn⇥r a1 a1 a2 aa1>1 (x x\ ) 2. a> 2 (x x\ ) . a1 a a2 1 x\ a2 x p p 2 a kx 1 (x)) max (r f (r>2 f (x)) x\ ) \ max a2 (x x x0 x1 x2 x3 > \ . log n a (x x ) kx x\ ka22 x0\ ) x\1 \ 2 2 3 aa1> (x . log n 0 1 x 2x\ k 3 x0 x10 x\x· 21 x32 x3 1 x xxx x· xkx x x 2 . log n · log n \ x log n a> 1 (x \ xa (a) k 1 1,k > 2,k a1 (x x\ ) Wk = f1 (S1,k , S2,k ) . log\ n 2 ) 3 . log n =xf0 (S1,k12kx , Sx2,k Fk = f2 (S1,k , S2,k ) 0 x21 x 3 x\x k2 x\ ka22Fk > 2 (x kx p x ) .> log n \ x x x x\ k2 x0 a2p(x x\ ) . log n x ) . log n \ (b) kx x k2 \ x1 x2 x3 (c) \ Figure 2: (a) The shaded region is an illustration of the incoherence region, which satisfies a> j (x − x ) . √ 0 1 log n for all points x in the region. (b) When x resides in the desired region, we know that x remains within the `2 ball but might fall out of the incoherence region (the shaded region). Once x1 leaves the incoherence region, we lose control and may overshoot. (c) Our theory reveals that with high probability, all iterates will stay within the incoherence region, enabling fast convergence. The following observation is thus immediate: one can safely adopt a far more aggressive step size (as large as ηt = O(1/ log n)) to achieve acceleration, as long as the iterates stay within the RIC. This, however, fails to be guaranteed by generic gradient descent theory. To be more precise, if the current iterate xt falls within the desired region, then in view of (8), we can ensure `2 error contraction after one iteration, namely, kxt+1 − x\ k2 ≤ kxt − x\ k2 1 and hence xt+1 stays within the local `2 ball and hence satisfies (11a). However, it is 1 not immediately obvious that xt+1 would still stay incoherent with the sensing vectors and satisfy (11b). If xt+1 leaves the RIC, it no longer enjoys the benign local geometry of the loss function, and the algorithm has to slow down in order to avoid overshooting. See Figure 2(b) for a visual illustration. In fact, in almost all regularized gradient descent algorithms mentioned in Section 1.2, one of the main purposes of the proposed regularization procedures is to enforce such incoherence constraints.
\ x is aligned with (and hence very coherent with) one vector aj , then with high probability one has a> j x−x | & √ √ a> nkxk2 , which is significantly larger than log nkxk2 . j x|  4 If 1 11 1 k 1 2 2.4 Implicit regularization However, is regularization really necessary for the iterates to stay within the RIC? To answer this question, t maxj |a> (xt −x\ )| maxj |a> j x | (resp. √logjnkx\ k ) vs. the we plot in Figure 3(a) (resp. Figure 3(b)) the incoherence measure √log nkx \k 2 2 iteration count in a typical Monte Carlo trial, generated in the same way as for Figure 1(a). Interestingly, the incoherence measure remains bounded by 2 for all iterations t > 1. This important observation suggests that one may adopt a substantially more aggressive step size throughout the whole algorithm. 4 2.5 3.5 2 3 1.5 2.5 1 2 0.5 1.5 1 0 0 5 10 15 20 25 30 0 5 10 (a) 15 20 25 30 (b) t \ max1≤j≤m |a> max1≤j≤m |a> xt | j (x −x )| √ (in (a)) and (in (b)) of the gradient Figure 3: The incoherence measure √log nkx\ kj \k log nkx 2 2 iterates vs. iteration count for the phase retrieval problem. The results are shown for n ∈ {20, 100, 200, 1000} and m = 10n, with the step size taken to be ηt = 0.1. The problem instances are generated in the same way as in Figure 1(a). The main objective of this paper is thus to provide a theoretical validation of the above empirical observation. As we will demonstrate shortly, with high probability all iterates along the execution of the algorithm (as well as the spectral initialization) are provably constrained within the RIC, implying fast convergence of vanilla gradient descent (cf. Figure 2(c)). The fact that the iterates stay incoherent with the measurement mechanism automatically, without explicit enforcement, is termed “implicit regularization”. 2.5 A glimpse of the analysis: a leave-one-out trick In order to rigorously establish (11b) for all iterates, the current paper develops a powerful mechanism based on the leave-one-out perturbation argument, a trick rooted and widely used in probability and random matrix theory. Note that the iterate xt is statistically dependent with the design vectors {aj }. Under such circumstances, one often resorts to generic bounds like the Cauchy-Schwarz inequality, which would not yield a desirable estimate. To address this issue, we introduce a sequence of auxiliary iterates {xt,(l) } for each 1 ≤ l ≤ m (for analytical purposes only), obtained by running vanilla gradient descent using all but the lth sample. As one can expect, such auxiliary trajectories serve as extremely good surrogates of {xt } in the sense that xt ≈ xt,(l) , 1 ≤ l ≤ m, t ≥ 0, (12) since their constructions only differ by a single sample. Most importantly, since xt,(l) is independent with the lth design vector, it is much easier to control its incoherence w.r.t. al to the desired level: p
t,(l) (13) a> − x\ . log n x\ 2 . l x Combining (12) and (13) then leads to (11b). See Figure 4 for a graphical illustration of this argument. Notably, this technique is very general and applicable to many other problems. We invite the readers to Section 5 for more details. 12 max (r minimizeX 2 f (x)) 2 f (x)) (r max a1 \ a2 x A Ax x Ax 1 = |Ax| ay1Ax a2 2y = |Ax| x\ 2 1 -3 incoherence region w.r.t. a1 9 2 4 -1 1 {x 4 16 -2 4 -1 1 3 9 4 16 t,(1) minimizeh,x \ kx a> 2 (x p \ x) . log n \ x k2 t,(l) kx ei X (i,j)2⌦ p x) . log n x\ k2 find a> 1 (x f (X) = {x {xt,(l) } } al } {x } X2R s.t. n s.t. b⇤i hi x⇤i ai = a1 a2 x\ a1 a2 x\ · {xt,(l) } (a) al e> i X > e> j XX ei w.r.t. al w.r.t. al t m ⇣ X i=1 n⇥r find h, x 2 C al f (h, x) = incoherence region w.r.t. a1 w.r.t. al max (r 2 max (r 2 f (x a> 1 (x kx a> 2 (x kx (b) Figure 4: Illustration of the leave-one-out sequence w.r.t. al . (a) The sequence {xt,(l) }t≥0 is constructed without using the lth sample. (b) Since the auxiliary sequence {xt,(l) } is constructed without using al , the leave-one-out iterates stay within the incoherence region w.r.t. al with high probability. Meanwhile, {xt } and {xt,(l) } are expected to remain close as their construction differ only in a single sample. 3 Main results This section formalizes the implicit regularization phenomenon underlying unregularized gradient descent, and presents its consequences, namely near-optimal statistical and computational guarantees for phase retrieval, matrix completion, and blind deconvolution. Note that the discrepancy measure dist (·, ·) may vary from problem to problem. 3.1 Phase retrieval Suppose the m quadratic equations 1
\ 2 yj = a> , j x (14) j = 1, 2, . . . , m i.i.d. are collected using random design vectors, namely, aj ∼ N (0, In ), and the nonconvex problem to solve is i2 1 X h >
2 aj x − yj . 4m j=1 m minimizex∈Rn f (x) := (15) The Wirtinger flow (WF) algorithm, first introduced in [CLS15], is a combination of spectral initialization and vanilla gradient descent; see Algorithm 1. Algorithm 1 Wirtinger flow for phase retrieval Input: {aj }1≤j≤m and {yj }1≤j≤m . e0 be the leading eigenvalue and eigenvector of Spectral initialization: Let λ1 (Y ) and x m Y = 1 X yj aj a> j , m j=1 p e0 . respectively, and set x0 = λ1 (Y ) /3 x Gradient updates: for t = 0, 1, 2, . . . , T − 1 do 1
xt+1 = xt − ηt ∇f xt . 1 1 1 (16) (17) 1 Recognizing that the global phase/sign is unrecoverable from quadratic measurements, 1 we introduce the `2 distance modulo the global phase as follows  dist(x, x\ ) := min kx − x\ k2 , kx + x\ k2 . (18) 13 f (x 1 Our finding is summarized in the following theorem. i.i.d. Theorem 1. Let x\ ∈ Rn be a fixed vector. Suppose aj ∼ N (0, In ) for each 1 ≤ j ≤ m and m ≥
c0 n log n for some sufficiently large constant c0 > 0. Assume the step size obeys ηt ≡ η = c1 / log n · kx0 k22 for any sufficiently small constant c1 > 0. Then
there exist some absolute constants 0 < ε < 1 and c2 > 0 such that with probability at least 1 − O mn−5 , the Wirtinger flow iterates (Algorithm 1) satisfy that for all t ≥ 0, dist(xt , x\ ) ≤ ε(1 − ηkx\ k22 /2)t kx\ k2 , p
t \ ≤ c2 log nkx\ k2 . max a> j x −x 1≤j≤m (19a) (19b) Theorem 1 reveals a few intriguing properties of WF. • Implicit regularization: Theorem 1 asserts that the incoherence properties are satisfied throughout the execution of the algorithm (see (19b)), which formally justifies the implicit regularization feature we hypothesized. • Near-constant step size: Consider the case where kx\ k2 = 1. Theorem 1 establishes near-linear convergence of WF with a substantially more aggressive step size η  1/ log n. Compared with the choice η . 1/n admissible in [CLS15, Theorem 3.3], Theorem 1 allows WF to attain -accuracy within O(log n log(1/)) iterations. The resulting computational complexity of WF is   1 , O mn log n log 
which significantly improves upon the result O mn2 log (1/) derived in [CLS15]. As a side note, if the sample size further increases to m  n log2 n, then a constant step size η  1 is also feasible, resulting in an iteration complexity log(1/). This follows since with
high probability, the entire trajectory resides t \ . kx\ k2 . We omit the details here. within a more refined incoherence region maxj a> j x −x • Incoherence of spectral initialization: We have also demonstrated in Theorem 1 that the initial guess x0 falls within the RIC and is hence nearly orthogonal to all design vectors. This provides a finer characterization of spectral initialization, in comparison to prior theory that focuses primarily on the `2 accuracy [NJS13,CLS15]. We expect our leave-one-out analysis to accommodate other variants of spectral initialization studied in the literature [CC17, CLM+ 16, WGE17, LL17, MM17]. 3.2 Low-rank matrix completion Let M ∈ Rn×n be a positive semidefinite matrix5 with rank r, and suppose its eigendecomposition is \ M \ = U \ Σ\ U \> , (20) where U \ ∈ Rn×r consists of orthonormal columns, and Σ\ is an r × r diagonal matrix with eigenvalues in a descending order, i.e. σmax = σ1 ≥ · · · ≥ σr = σmin > 0. Throughout this paper, we assume the condition number κ := σmax /σmin is bounded by a fixed constant, independent of the problem size (i.e. n and r). Denoting X \ = U \ (Σ\ )1/2 allows us to factorize M \ as M \ = X \ X \> . (21) Consider a random sampling model such that each entry of M \ is observed independently with probability 0 < p ≤ 1, i.e. for 1 ≤ j ≤ k ≤ n, ( \ Mj,k + Ej,k with probability p, (22) Yj,k = 0, else, 5 Here, we assume M \ to be positive semidefinite to simplify the presentation, but note that our analysis easily extends to asymmetric low-rank matrices. 14 where the entries of E = [Ej,k ]1≤j≤k≤n are independent sub-Gaussian noise with sub-Gaussian norm σ (see [Ver12, Definition 5.7]). We denote by Ω the set of locations being sampled, and PΩ (Y ) represents the projection of Y onto the set of matrices supported in Ω. We note here that the sampling rate p, if not known, can be faithfully estimated by the sample proportion |Ω|/n2 . To fix ideas, we consider the following nonconvex optimization problem minimizeX∈Rn×r f (X) := 1 X 4p (j,k)∈Ω
2 > e> j XX ek − Yj,k . (23) The vanilla gradient descent algorithm (with spectral initialization) is summarized in Algorithm 2. Algorithm 2 Vanilla gradient descent for matrix completion (with spectral initialization) Input: Y = [Yj,k ]1≤j,k≤n , r, p. Spectral initialization: Let U 0 Σ0 U 0> be the rank-r eigendecomposition of M 0 :=
1 1 PΩ (Y ) = PΩ M \ + E , p p
1/2 and set X 0 = U 0 Σ0 . Gradient updates: for t = 0, 1, 2, . . . , T − 1 do
X t+1 = X t − ηt ∇f X t . (24) Before proceeding to the main theorem, we first introduce a standard incoherence parameter required for matrix completion [CR09]. Definition 3 (Incoherence for matrix completion). A rank-r matrix M \ with eigendecomposition M \ = U \ Σ\ U \> is said to be µ-incoherent if r r µ µr \ \ U 2,∞ ≤ U F= . (25) n n In addition, recognizing that X \ is identifiable only up to orthogonal transformation, we define the optimal transform from the tth iterate X t to X \ as ct := argmin X t R − X \ H R∈O r×r F , (26) where Or×r is the set of r × r orthonormal matrices. With these definitions in place, we have the following theorem. Theorem 2. Let M \ be a rank r, µ-incoherent PSD matrix, and its condition number κ is a fixed constant. Suppose the sample size satisfies n2 p ≥ Cµ3 r3 n log3 n for some sufficiently large constant C > 0, and the noise satisfies r n σmin σ p . (27) 3 p κ µr log3 n
With probability at least 1 − O n−3 , the iterates of Algorithm 2 satisfy  r  n ct − X \ ≤ C4 ρt µr √1 + C1 σ X tH X \ F, (28a) F np σmin p s s ! log n σ n log n t ct \ t + C8 X \ 2,∞ , (28b) X H − X 2,∞ ≤ C5 ρ µr np σmin p  r  n ct − X \ ≤ C9 ρt µr √1 + C10 σ X tH X\ (28c) np σmin p 15 for all 0 ≤ t ≤ T = O(n5 ), where C1 , C4 , C5 , C8 , C9 and C10 are some absolute positive constants and 1 − (σmin /5) · η ≤ ρ < 1, provided that 0 < ηt ≡ η ≤ 2/ (25κσmax ). Theorem 2 provides the first theoretical guarantee of unregularized gradient descent for matrix completion, demonstrating near-optimal statistical accuracy and computational complexity. • Implicit regularization: In Theorem 2, we bound the
`2 /`∞ error of the iterates in a uniform manner \ , which implies the iterates remain incoherent via (28b). Note that X − X \ 2,∞ = maxj e> j X −X 2 with the sensing vectors throughout and have small incoherence parameters (cf. (25)). In comparison, prior works either include a penalty term on {ke> j Xk2 }1≤j≤n [KMO10a,SL16] and/or kXkF [SL16] to encourage an incoherent and/or low-norm solution, or add an extra projection operation to enforce incoherence [CW15, ZL16]. Our results demonstrate that such explicit regularization is unnecessary. • Constant step size: Without loss of generality we may assume that σmax = kM \ k = O(1), which can be done by choosing proper scaling of M \ . Hence we have a constant step size ηt  1. Actually it is more convenient to consider the scale invariant parameter ρ: Theorem 2 guarantees linear convergence of the vanilla gradient descent at a constant rate ρ. Remarkably, the convergence occurs with respect to three different unitarily invariant norms: the Frobenius norm k · kF , the `2 /`∞ norm k · k2,∞ , and the spectral norm k · k. As far as we know, the latter two are established for the first time. Note that our result even improves upon that for regularized gradient descent; see Table 1. • Near-optimal sample complexity: When the rank r = O(1), vanilla gradient descent succeeds under a near-optimal sample complexity n2 p & npoly log n, which is statistically optimal up to some logarithmic factor. • Near-minimal Euclidean error: In view of (28a), as t increases, the Euclidean error of vanilla GD converges to r n σ t ct \ X \ F, (29) X H −X F . σmin p which coincides with the theoretical guarantee in [CW15, Corollary 1] and matches the minimax lower bound established in [NW12, KLT11]. • Near-optimal entrywise error: The `2 /`∞ error bound (28b) immediately yields entrywise control of the empirical risk. Specifically, as soon as t is sufficiently large (so that the first term in (28b) is negligible), we have

ct X t H ct − X \ > ct − X \ X \> X t X t> − M \ ∞ ≤ X t H + X tH ∞ ∞ ct ct − X \ ≤ X tH X tH 2,∞ s σ n log n M\ ∞ , . σmin p 2,∞ ct − X \ + X tH 2,∞ X\ 2,∞ ct −X \ k2,∞ ≤ kX \ k2,∞ and kM \ k∞ = where the last line follows from (28b) as well as the facts that kX t H \ 2 kX k2,∞ . Compared with the Euclidean loss (29), this implies that when r = O(1), the entrywise error of X t X t> is uniformly spread out across all entries. As far as we know, this is the first result that reveals near-optimal entrywise error control for noisy matrix completion using nonconvex optimization, without resorting to sample splitting. Remark 3. Theorem 2 remains valid if the total number T of iterations obeys T = nO(1) . In the noiseless case where σ = 0, the theory allows arbitrarily large T . Finally, we report the empirical statistical accuracy of vanilla gradient descent in the presence of noise. Figure 5 displays the squared relative error of vanilla gradient descent as a function of the signal-to-noise ratio (SNR), where the SNR is defined to be P \
2 kM \ k2F (j,k)∈Ω Mj,k SNR := P ≈ , (30) n2 σ 2 (j,k)∈Ω Var (Ej,k ) 16 -10 -20 -30 -40 -50 -60 -70 -80 -90 10 20 30 40 50 60 70 80 c (measured by k·k , k·k , k·k Figure 5: Squared relative error of the estimate X F 2,∞ modulo global transfor> c c c mation) and M = X X (measured by k·k∞ ) vs. SNR for noisy matrix completion, where n = 500, r = 10, c denotes the estimate returned by Algorithm 2 after convergence. p = 0.1, and ηt = 0.2. Here X and the relative error is measured in terms of the square of the metrics as in (28) as well as the squared entrywise prediction error. Both the relative error and the SNR are shown on a dB scale (i.e. 10 log10 (SNR) and 10 log10 (squared relative error) are plotted). As one can see from the plot, the squared relative error scales inversely proportional to the SNR, which is consistent with our theory.6 3.3 Blind deconvolution Suppose we have collected m bilinear measurements yj = b∗j h\ x\∗ aj , (31) 1 ≤ j ≤ m,

i.i.d. where aj follows a complex Gaussian distribution, i.e. aj ∼ N 0, 21 IK + iN 0, 21 IK for 1 ≤ j ≤ m, and ∗ B := [b1 , · · · , bm ] ∈ Cm×K is formed by the first K columns of a unitary discrete Fourier transform (DFT) matrix F ∈ Cm×m obeying F F ∗ = Im (see Appendix D.3.2 for a brief introduction to DFT matrices). This setup models blind deconvolution, where the two signals under convolution belong to known low-dimensional subspaces of dimension K [ARR14]7 . In particular, the partial DFT matrix B plays an important role in image blind deblurring. In this subsection, we consider solving the following nonconvex optimization problem minimizeh,x∈CK f (h, x) = m X j=1 b∗j hx∗ aj − yj 2 . (32) The (Wirtinger) gradient descent algorithm (with spectral initialization) is summarized in Algorithm 3; here, ∇h f (h, x) and ∇x f (h, x) stand for the Wirtinger gradient and are given in (77) and (78), respectively; see [CLS15, Section 6] for a brief introduction to Wirtinger calculus. It is self-evident that h\ and x\ are only identifiable up to global scaling, that is, for any nonzero α ∈ C, h\ x\∗ =
∗ 1 \ h αx\ . α In light of this, we will measure the discrepancy between   h z := ∈ C2K and x 6 Note z \ :=  \ h ∈ C2K x\ (33)

that when M \ is well-conditioned and when r = O(1), one can easily check that SNR ≈ kM \ k2F / n2 σ 2  2 2 theory says that the squared relative error bound is proportional to σ /σmin . have set the dimensions of the two subspaces equal, and it is straightforward to extend our results to the case of unequal subspace dimensions. 2 /(n2 σ 2 ), and our σmin 7 For simplicity, we 17 via the following function dist z, z
\ := min α∈C s 2 1 h − h\ α 2 2 + kαx − x\ k2 . (34) Algorithm 3 Vanilla gradient descent for blind deconvolution (with spectral initialization) Input: {aj }1≤j≤m , {bj }1≤j≤m and {yj }1≤j≤m . Spectral initialization: Let σ1 (M ), ȟ0 and x̌0 be the leading singular value, left and right singular vectors of m X M := yj bj a∗j , j=1 p p respectively. Set h0 = σ1 (M ) ȟ0 and x0 = σ1 (M ) x̌0 . Gradient updates: for t = 0, 1, 2, . . . , T − 1 do " 1
#  t+1   t  t t ∇ f h , x 2 h t h h
. = − η kx1k2 xt+1 xt ∇ f ht , xt kht k2 x (35) 2 Before proceeding, we need to introduce the incoherence parameter [ARR14, LLSW16], which is crucial for blind deconvolution, whose role is similar to the incoherence parameter (cf. Definition 3) in matrix completion. Definition 4 (Incoherence for blind deconvolution). Let the incoherence parameter µ of h\ be the smallest number such that µ h\ 2 . (36) max b∗j h\ ≤ √ 1≤j≤m m The incoherence parameter describes the spectral flatness of the signal h\ . With this definition in place, we have the following theorem, where for identifiability we assume that h\ 2 = x\ 2 . Theorem 3. Suppose the number of measurements obeys m ≥ Cµ2 K log9 m for some sufficiently large constant C > 0, and suppose the step size η > 0 is taken to be some sufficiently small constant. Then there exist constants c1 , c2 , C1 , C3 , C4 > 0 such that with probability exceeding 1 − c1 m−5 − c1 me−c2 K , the iterates in Algorithm 3 satisfy 
η t 1 dist z t , z \ ≤ C1 1 − (37a) x\ 2 , 16 log2 m
1 max a∗l αt xt − x\ ≤ C3 1.5 x\ 2 , (37b) 1≤l≤m log m µ 1 max b∗l ht ≤ C4 √ log2 m h\ 2 (37c) t 1≤l≤m m α for all t ≥ 0. Here, we denote αt as the alignment parameter, αt := arg min α∈C 1 t h − h\ α 2 2 + αxt − x\ 2 2 . (38) Theorem 3 provides the first theoretical guarantee of unregularized gradient descent for blind deconvolution at a near-optimal statistical and computational complexity. A few remarks are in order. • Implicit regularization: Theorem 3 reveals that the unregularized gradient descent iterates remain incoherent with the sampling mechanism (see (37b) and (37c)). Recall that prior works operate upon a regularized cost function with an additional penalty term that regularizes the global scaling {khk2 , kxk2 } and the incoherence {|b∗j h|}1≤j≤m [LLSW16, HH17, LS17]. In comparison, our theorem implies that it is unnecessary to regularize either the incoherence or the scaling ambiguity, which is somewhat surprising. This justifies the use of regularization-free (Wirtinger) gradient descent for blind deconvolution. 18 • Constant step size: Compared to the step size ηt . 1/m suggested in [LLSW16] for regularized gradient descent, our theory admits a substantially more aggressive step size (i.e. ηt  1) even without regularization. Similar to phase retrieval, the computational efficiency is boosted by a factor of m, attaining -accuracy within O (log(1/)) iterations (vs. O (m log(1/)) iterations in prior theory). • Near-optimal sample complexity: It is demonstrated that vanilla gradient descent succeeds at a near-optimal sample complexity up to logarithmic factors, although our requirement is slightly worse than [LLSW16] which uses explicit regularization. Notably, even under the sample complexity herein, the iteration complexity given in [LLSW16] is still O (m/poly log(m)). • Incoherence of spectral initialization: As in phase retrieval, Theorem 3 demonstrates that the estimates returned by the spectral method are incoherent with respect to both {aj } and {bj }. In contrast, [LLSW16] recommends a projection operation (via a linear program) to enforce incoherence of the initial estimates, which is dispensable according to our theory. • Contraction in k·kF : It is easy to check that the Frobenius norm error satisfies ht xt∗ − h\ x\∗
dist z t , z \ , and therefore Theorem 3 corroborates the empirical results shown in Figure 1(c). 4 F . Related work Solving nonlinear systems of equations has received much attention in the past decade. Rather than directly attacking the nonconvex formulation, convex relaxation lifts the object of interest into a higher dimensional space and then attempts recovery via semidefinite programming (e.g. [RFP10, CSV13, CR09, ARR14]). This has enjoyed great success in both theory and practice. Despite appealing statistical guarantees, semidefinite programming is in general prohibitively expensive when processing large-scale datasets. Nonconvex approaches, on the other end, have been under extensive study in the last few years, due to their computational advantages. There is a growing list of statistical estimation problems for which nonconvex approaches are guaranteed to find global optimal solutions, including but not limited to phase retrieval [NJS13, CLS15, CC17], low-rank matrix sensing and completion [TBS+ 16, BNS16, PKCS16, CW15, ZL15, GLM16], blind deconvolution and self-calibration [LLSW16, LS17, CJ16, LLB17, LLJB17], dictionary learning [SQW17], tensor decomposition [GM17], joint alignment [CC16], learning shallow neural networks [SJL17, ZSJ+ 17], robust subspace learning [NNS+ 14, MZL17, LM14, CJN17]. In several problems [SQW16,SQW17,GM17,GLM16,LWL+ 16,LT16,MBM16,MZL17], it is further suggested that the optimization landscape is benign under sufficiently large sample complexity, in the sense that all local minima are globally optimal, and hence nonconvex iterative algorithms become promising in solving such problems. Below we review the three problems studied in this paper in more details. Some state-of-the-art results are summarized in Table 1. • Phase retrieval. Candès et al. proposed PhaseLift [CSV13] to solve the quadratic systems of equations based on convex programming. Specifically, it lifts the decision variable x\ into a rank-one matrix X \ = x\ x\> and translates the quadratic constraints of x\ in (14) into linear constraints of X \ . By dropping the rank constraint, the problem becomes convex [CSV13, SESS11, CL14, CCG15, CZ15, Tro15a]. Another convex program PhaseMax [GS16,BR17,HV16,DTL17] operates in the natural parameter space via linear programming, provided that an anchor vector is available. On the other hand, alternating minimization [NJS13] with sample splitting has been shown to enjoy much better computational guarantee. In contrast, Wirtinger Flow [CLS15] provides the first global convergence result for nonconvex methods without sample splitting, whose statistical and computational guarantees are later improved by [CC17] via an adaptive truncation strategy. Several other variants of WF are also proposed [CLM+ 16, KÖ16, Sol17], among which an amplitude-based loss function has been investigated [WGE17, ZZLC17, WZG+ 16, WGSC17]. In particular, [ZZLC17] demonstrates that the amplitude-based loss function has a better curvature, and vanilla gradient descent can indeed converge with a constant step size at the order-wise optimal sample complexity. A small sample of other nonconvex phase retrieval methods include [SBE14, SR15, CL16, CFL15, DR17, GX16, Wei15, BEB17, TV17, CLW17, QZEW17], which are beyond the scope of this paper. • Matrix completion. Nuclear norm minimization was studied in [CR09] as a convex relaxation paradigm to solve the matrix completion problem. Under certain incoherence conditions imposed upon the ground truth 19 matrix, exact recovery is guaranteed under near-optimal sample complexity [CT10, Gro11, Rec11, Che15, DR16]. Concurrently, several works [KMO10a, KMO10b, LB10, JNS13, HW14, HMLZ15, ZWL15, JN15, TW16, JKN16, WCCL16, ZWL15] tackled the matrix completion problem via nonconvex approaches. In particular, the seminal work by Keshavan et al. [KMO10a, KMO10b] pioneered the two-stage approach that is widely adopted by later works. Sun and Luo [SL16] demonstrated the convergence of gradient descent type methods for noiseless matrix completion with a regularized nonconvex loss function. Instead of penalizing the loss function, [CW15, ZL16] employed projection to enforce the incoherence condition throughout the execution of the algorithm. To the best of our knowledge, no rigorous guarantees have been established for matrix completion without explicit regularization. A notable exception is [JKN16], which uses unregularized stochastic gradient descent for matrix completion in the online setting. However, the analysis is performed with fresh samples in each iteration. Our work closes the gap and makes the first contribution towards understanding implicit regularization in gradient descent without sample splitting. In addition, entrywise eigenvector perturbation has been studied by [JN15] and [AFWZ17] in order to analyze SVD-based algorithms for matrix completion, which helps us establish theoretical guarantees for the spectral initialization step. • Blind deconvolution. In [ARR14], Ahmed et al. first proposed to invoke similar lifting ideas for blind deconvolution, which translates the bilinear measurements (31) into a system of linear measurements of a rank-one matrix X \ = h\ x\∗ . Near-optimal performance guarantees have been established for convex relaxation [ARR14]. Under the same model, Li et al. [LLSW16] proposed a regularized gradient descent algorithm that directly optimizes the nonconvex loss function (32) with a few regularization terms that account for scaling ambiguity and incoherence. See [CJ16] for a related formulation. In [HH17], a Riemannian steepest descent method is developed that removes the regularization for scaling ambiguity, although they still need to regularize for incoherence. In [AAH17], a linear program is proposed but requires exact knowledge of the signs of the signals. Blind deconvolution has also been studied for other models – interested readers may refer to [Chi16, LS17, LLJB17, LS15, LTR16, ZLK+ 17, WC16]. On the other hand, our analysis framework is based on a leave-one-out perturbation argument. This technique has been widely used to analyze high-dimensional problems with random designs, including but not limited to robust M-estimation [EKBB+ 13,EK15], statistical inference for sparse regression [JM15], likelihood ratio test in logistic regression [SCC17], phase synchronization [ZB17, AFWZ17], ranking from pairwise comparisons [CFMW17], community recovery [AFWZ17], and recovering structured probability matrices [Che17]. In particular, this technique results in tight performance guarantees for the generalized power method [ZB17], the spectral method [AFWZ17, CFMW17], and convex programming approaches [EK15, ZB17, SCC17, CFMW17], however it has not been applied to analyze nonconvex optimization algorithms. Finally, we note that the notion of implicit regularization — broadly defined — arises in settings far beyond the models and algorithms considered herein. For instance, it has been conjectured that in matrix factorization, over-parameterized stochastic gradient descent effectively enforces certain norm constraints, allowing it to converge to a minimal-norm solution as long as it starts from the origin [GWB+ 17]. The stochastic gradient methods have also been shown to implicitly enforce Tikhonov regularization in several statistical learning settings [LCR16]. More broadly, this phenomenon seems crucial in enabling efficient training of deep neural networks [NTS14, NTSS17, ZBH+ 16, SHS17, KMN+ 16]. 5 A general recipe for trajectory analysis In this section, we sketch a general recipe for establishing performance guarantees of gradient descent, which conveys the key idea for proving the main results of this paper. The main challenge is to demonstrate that appropriate incoherence conditions are preserved throughout the trajectory of the algorithm. This requires exploiting statistical independence of the samples in a careful manner, in conjunction with generic optimization theory. Central to our approach is a leave-one-out perturbation argument, which allows to decouple the statistical dependency while controlling the component-wise incoherence measures. 20 General Recipe (a leave-one-out analysis) Step 1: characterize restricted strong convexity and smoothness of f , and identify the region of incoherence and contraction (RIC). Step 2: introduce leave-one-out sequences {X t,(l) } and {H t,(l) } for each l, where {X t,(l) } (resp. {H t,(l) }) is independent of any sample involving φl (resp. ψl ); Step 3: establish the incoherence condition for {X t } and {H t } via induction. Suppose the iterates satisfy the claimed conditions in the tth iteration: (a) show, via restricted strong convexity, that the true iterates (X t+1 , H t+1 ) and the leave-one-out version (X t+1,(l) , H t+1,(l) ) are exceedingly close; (b) use statistical independence to show that X t+1,(l) − X \ (resp. H t+1,(l) − H \ ) is incoherent w.r.t. φl (resp. ψl ), namely, kφ∗l (X t+1,(l) − X \ )k2 and kψl∗ (H t+1,(l) − H \ )k2 are both well-controlled; (c) combine the bounds to establish the desired incoherence condition concerning max kφ∗l (X t+1 − X \ )k2 and max kψl∗ (H t+1 − H \ )k2 . l 5.1 l General model Consider the following problem where the samples are collected in a bilinear/quadratic form as yj = ψj∗ H \ X \∗ φj , 1 ≤ j ≤ m, (39) where the objects of interest H \ , X \ ∈ Cn×r or Rn×r might be vectors or tall matrices taking either real or complex values. The design vectors {ψj } and {φj } are in either Cn or Rn , and can be either random or deterministic. This model is quite general and entails all three examples in this paper as special cases: • Phase retrieval : H \ = X \ = x\ ∈ Rn , and ψj = φj = aj ; • Matrix completion: H \ = X \ ∈ Rn×r and ψj , φj ∈ {e1 , · · · , en }; • Blind deconvolution: H \ = h\ ∈ CK , X \ = x\ ∈ CK , φj = aj , and ψj = bj . For this setting, the empirical loss function is given by m f (Z) := f (H, X) = 2 1 X ∗ ψj HX ∗ φj − yj , m j=1 where we denote Z = (H, X). To minimize f (Z), we proceed with vanilla gradient descent
Z t+1 = Z t − η∇f Z t , ∀t ≥ 0 following a standard spectral initialization, where η is the step size. As a remark, for complex-valued problems, the gradient (resp. Hessian) should be understood as the Wirtinger gradient (resp. Hessian). It is clear from (39) that Z \ = (H \ , X \ ) can only be recovered up to certain global ambiguity. For clarity of presentation, we assume in this section that such ambiguity has already been taken care of via proper global transformation. 5.2 Outline of the recipe We are now positioned to outline the general recipe, which entails the following steps. • Step 1: characterizing local geometry in the RIC. Our first step is to characterize a region R — which we term as the region of incoherence and contraction (RIC) — such that the Hessian matrix ∇2 f (Z) obeys strong convexity and smoothness, 0 ≺ αI  ∇2 f (Z)  βI, 21 ∀Z ∈ R, (40) or at least along certain directions (i.e. restricted strong convexity and smoothness), where β/α scales slowly (or even remains bounded) with the problem size. As revealed by optimization theory, this geometric property (40) immediately implies linear convergence with the contraction rate 1 − O(α/β) for a properly chosen step size η, as long as all iterates stay within the RIC. A natural question then arises: what does the RIC R look like? As it turns out, the RIC typically contains all points such that the `2 error kZ − Z \ kF is not too large and (incoherence) max φ∗j (X − X \ ) j 2 and max ψj∗ (H − H \ ) j 2 are well-controlled. (41) In the three examples, the above incoherence condition translates to: \ – Phase retrieval : maxj a> j (x − x ) is well-controlled; – Matrix completion: X − X \ – Blind deconvolution: is well-controlled; 2,∞ maxj a> (x − j \ x\ ) and maxj b> j (h − h ) are well-controlled. • Step 2: introducing the leave-one-out sequences. To justify that no iterates leave the RIC, we rely on the construction of auxiliary sequences. Specifically, for each l, produce an auxiliary sequence {Z t,(l) = (X t,(l) , H t,(l) )} such that X t,(l) (resp. H t,(l) ) is independent of any sample involving φl (resp. ψl ). As an example, suppose that the φl ’s and the ψl ’s are independently and randomly generated. Then for each l, one can consider a leave-one-out loss function 2 1 X ∗ f (l) (Z) := ψj HX ∗ φj − yj m j:j6=l that discards the lth sample. One further generates {Z t,(l) } by running vanilla gradient descent w.r.t. this auxiliary loss function, with a spectral initialization that similarly discards the lth sample. Note that this procedure is only introduced to facilitate analysis and is never implemented in practice. • Step 3: establishing the incoherence condition. We are now ready to establish the incoherence condition with the assistance of the auxiliary sequences. Usually the proof proceeds by induction, where our goal is to show that the next iterate remains within the RIC, given that the current one does. – Step 3(a): proximity between the original and the leave-one-out iterates. As one can anticipate, {Z t } and {Z t,(l) } remain “glued” to each other along the whole trajectory, since their constructions differ by only a single sample. In fact, as long as the initial estimates stay sufficiently close, their gaps will never explode. To intuitively see why, use the fact ∇f (Z t ) ≈ ∇f (l) (Z t ) to discover that

Z t+1 − Z t+1,(l) = Z t − η∇f (Z t ) − Z t,(l) − η∇f (l) Z t,(l)
≈ Z t − Z t,(l) − η∇2 f (Z t ) Z t − Z t,(l) , which together with the strong convexity condition implies `2 contraction

Z t+1 − Z t+1,(l) F ≈ I − η∇2 f (Z t ) Z t − Z t,(l) ≤ Z t − Z t,(l) F 2 . Indeed, (restricted) strong convexity is crucial in controlling the size of leave-one-out perturbations. – Step 3(b): incoherence condition of the leave-one-out iterates. The fact that Z t+1 and Z t+1,(l) are exceedingly close motivates us to control the incoherence of Z t+1,(l) − Z \ instead, for 1 ≤ l ≤ m. By construction, X t+1,(l) (resp. H t+1,(l) ) is statistically independent of any sample involving the design vector leads to a more friendly analysis for controlling
φl (resp. ψl ), a fact that typically
φ∗l X t+1,(l) − X \ 2 and ψl∗ H t+1,(l) − H \ 2 . – Step 3(c): combining the bounds. With these results in place, apply the triangle inequality to obtain

φ∗l X t+1 − X \ 2 ≤ φl k2 X t+1 − X t+1,(l) F + φ∗l X t+1,(l) − X \ 2 , where the first term
is controlled in Step 3(a) and the second term is controlled in Step 3(b). The term ψl∗ H t+1 − H \ 2 can be bounded similarly. By choosing the bounds properly, this establishes the incoherence condition for all 1 ≤ l ≤ m as desired. 22 6 Analysis for phase retrieval In this section, we instantiate the general recipe presented in Section 5 to phase retrieval and prove Theorem 1. Similar to the Section 7.1 in [CLS15], we are going to use ηt = c1 /(log n · kx\ k22 ) instead of c1 /(log n · kx0 k22 ) as the step size for analysis. This is because with high probability, kx0 k2 and kx\ k2 are rather close in the relative sense. Without loss of generality, we assume throughout this section that x\ 2 = 1 and dist(x0 , x\ ) = kx0 − x\ k2 ≤ kx0 + x\ k2 . (42) In addition, the gradient and the Hessian of f (·) for this problem (see (15)) are given respectively by i
1 X h >
2 aj x − yj a> j x aj , m j=1 m ∇f (x) = i
2 1 Xh 3 a> x − y aj a> j j j , m j=1 (43) m ∇2 f (x) = (44) which are useful throughout the proof. 6.1 6.1.1 Step 1: characterizing local geometry in the RIC Local geometry We start by characterizing the region that enjoys both strong convexity and the desired level of smoothness. This is supplied in the following lemma, which plays a crucial role in the subsequent analysis. Lemma 1 (Restricted strong convexity and smoothness for phase retrieval). Fix any sufficiently small constant C1 > 0 and any sufficiently large constant C2 > 0, and suppose the sample complexity obeys m ≥ c0 n log n for some sufficiently large constant c0 > 0. With probability at least 1 − O(mn−10 ), ∇2 f (x)  (1/2) · In holds simultaneously for all x ∈ Rn satisfying x − x\ 2 ≤ 2C1 ; and ∇2 f (x)  (5C2 (10 + C2 ) log n) · In holds simultaneously for all x ∈ Rn obeying max 1≤j≤m a> j x − x\ \ x−x Proof. See Appendix A.1. 2
(45a) ≤ 2C1 , p ≤ C2 log n. (45b) In words, Lemma 1 reveals that the Hessian matrix is positive definite and (almost) well-conditioned, if one restricts attention to the set of points that are (i) not far away from the truth (cf. (45a)) and (ii) incoherent with respect to the measurement vectors {aj }1≤j≤m (cf. (45b)). 6.1.2 Error contraction As we point out before, the nice local geometry enables `2 contraction, which we formalize below. Lemma 2. With probability exceeding 1 − O(mn−10 ), one has xt+1 − x\ 2 ≤ (1 − η/2) xt − x\ 2 (46) for any xt obeying the conditions (45), provided that the step size satisfies 0 < η ≤ 1/ [5C2 (10 + C2 ) log n]. 23 Proof. This proof applies the standard argument when establishing the `2 error contraction of gradient descent for strongly convex and smooth functions. See Appendix A.2. With the help of Lemma 2, we can turn the proof of Theorem 1 into ensuring that the trajectory {xt }0≤t≤n lies in the RIC specified by (47).8 This is formally stated in the next lemma. Lemma 3. Suppose for all 0 ≤ t ≤ T0 := n, the trajectory {xt } falls within the region of incoherence and contraction (termed the RIC), namely, xt − x\ t max a> l x −x 1≤l≤m 2
\ ≤ C1 , p ≤ C2 log n, (47a) (47b) then the claims in Theorem 1 hold true. Here and throughout this section, C1 , C2 > 0 are two absolute constants as specified in Lemma 1. Proof. See Appendix A.3. 6.2 Step 2: introducing the leave-one-out sequences In comparison to the `2 error bound (47a) that captures the overall loss, the incoherence hypothesis (47b) — which concerns sample-wise control of the empirical risk — is more complicated to establish. This is partly due to the statistical dependence between xt and the sampling vectors {al }. As described in the general recipe, the key idea is the introduction of a leave-one-out version of the WF iterates, which removes a single measurement from consideration. To be precise, for each 1 ≤ l ≤ m, we define the leave-one-out empirical loss function as i2 1 X h >
2 f (l) (x) := aj x − yj , (48) 4m j:j6=l  and the auxiliary trajectory xt,(l) initialization x 0,(l) t≥0 is constructed by running WF w.r.t. f (l) (x). In addition, the spectral is computed based on the rescaled leading eigenvector of the leave-one-out data matrix 1 X Y (l) := yj aj a> (49) j . m j:j6=l  Clearly, the entire sequence xt,(l) t≥0 is independent of the lth sampling vector al . This auxiliary procedure is formally described in Algorithm 4. 6.3 Step 3: establishing the incoherence condition by induction As revealed by Lemma 3, it suffices to prove that the iterates {xt }0≤t≤T0 satisfies (47) with high probability. Our proof will be inductive in nature. For the sake of clarity, we list all the induction hypotheses: max 1≤l≤m xt − x\ 2 xt − xt,(l) 2 t max a> j x −x 1≤j≤m
\ ≤ C1 , r ≤ C3 ≤ C2 p (51a) log n n (51b) log n. (51c) Here C3 > 0 is some universal constant. The induction on (51a), that is, xt+1 − x\ 2 ≤ C1 , (52) has already been established in Lemma 2. This subsection is devoted to establishing (51b) and (51c) for the (t + 1)th iteration, assuming that (51) holds true up to the tth iteration. We defer the justification of the base case (i.e. initialization at t = 0) to Section 6.4. 8 Here, we deliberately change 2C1 in (45a) to C1 in the definition of the RIC (47a) to ensure the correctness of the analysis. 24 Algorithm 4 The lth leave-one-out sequence for phase retrieval Input: {aj }1≤j≤m,j6=l and {yj }1≤j≤m,j6
=l . 0,(l) e Spectral initialization: let λ1 Y (l) and x be the leading eigenvalue and eigenvector of Y (l) = 1 X yj aj a> j , m j:j6=l respectively, and set x0,(l) q
 λ1 Y (l) /3 x e0,(l) − x\ e0,(l) , if x q =
− λ1 Y (l) /3 x e0,(l) , else. 2 e0,(l) + x\ ≤ x 2 , Gradient updates: for t = 0, 1, 2, . . . , T − 1 do
xt+1,(l) = xt,(l) − ηt ∇f (l) xt,(l) . (50) • Step 3(a): proximity between the original and the leave-one-out iterates. The leave-one-out sequence {xt,(l) } behaves similarly to the true WF iterates {xt } while maintaining statistical independence with al , a key fact that allows us to control the incoherence of lth leave-one-out sequence w.r.t. al . We will formally quantify the gap between xt+1 and xt+1,(l) in the following lemma, which establishes the induction in (51b). Lemma 4. Under the hypotheses (51), with probability at least 1 − O(mn−10 ), r log n t+1 t+1,(l) max x −x ≤ C3 , 2 1≤l≤m n (53) as long as the sample size obeys m  n log n and the stepsize 0 < η ≤ 1/ [5C2 (10 + C2 ) log n]. Proof. The proof relies heavily on the restricted strong convexity (see Lemma 1) and is deferred to Appendix A.4. • Step 3(b): incoherence of the leave-one-out iterates. By construction, xt+1,(l) is statistically independent of the sampling vector al . One can thus invoke the standard Gaussian concentration results
and the union bound to derive that with probability at least 1 − O mn−10 , t+1,(l) max a> − x\ l x 1≤l≤m
p ≤ 5 log n xt+1,(l) − x\ 2  p ≤ 5 log n xt+1,(l) − xt+1 2 + xt+1 − x\ ! r (ii) p log n + C1 ≤ 5 log n C3 n p ≤ C4 log n (i) 2  (54) holds for some constant C4 ≥ 6C1 > 0 and n sufficiently large. Here, (i) comes from the triangle inequality, and (ii) arises from the proximity bound (53) and the condition (52). • Step 3(c): combining the bounds. We are now prepared to establish (51c) for the (t + 1)th iteration. Specifically, t+1 max a> − x\ l x 1≤l≤m
t+1 ≤ max a> − xt+1,(l) l x 1≤l≤m 25
t+1,(l) + max a> − x\ l x 1≤l≤m
p ≤ max kal k2 xt+1 − xt+1,(l) 2 + C4 log n 1≤l≤m r (ii) √ p p log n + C4 log n ≤ C2 log n, ≤ 6n · C3 n (i) (55) where (i) follows from the Cauchy-Schwarz inequality and (54), the inequality (ii) is a consequence of (53) and (98), and the last inequality holds as long as C2 /(C3 + C4 ) is sufficiently large. Using mathematical induction and the union bound, we establish (51) for all t ≤ T0 = n with high probability. This in turn concludes the proof of Theorem 1, as long as the hypotheses are valid for the base case. 6.4 The base case: spectral initialization In the end, we return to verify the induction hypotheses for the base case (t = 0), i.e. the spectral initialization obeys (51). The following lemma justifies (51a) by choosing δ sufficiently small. Lemma 5. Fix any small constant δ > 0, and suppose m > c0 n log n for some large constant c0 > 0. e0 as defined in Algorithm 1, and suppose without loss of generality that Consider the two vectors x0 and x (42) holds. Then with probability exceeding 1 − O(n−10 ), one has kY − E [Y ]k ≤ δ, kx0 − x\ k2 ≤ 2δ and e0 − x\ x 2 ≤ √ (56) 2δ. (57) Proof. This result follows directly from the Davis-Kahan sinΘ theorem. See Appendix A.5. We then move on to justifying (51b), the proximity between the original and leave-one-out iterates for t = 0. Lemma 6. Suppose m > c0 n log n for some large constant c0 > 0. Then with probability at least 1 − O(mn−10 ), one has r log n 0 0,(l) ≤ C3 max x − x . (58) 2 1≤l≤m n Proof. This is also a consequence of the Davis-Kahan sinΘ theorem. See Appendix A.6. The final claim (51c) can be proved using the same argument as in deriving (55), and hence is omitted. 7 Analysis for matrix completion In this section, we instantiate the general recipe presented in Section 5 to matrix completion and prove Theorem 2. Before continuing, we first gather a few useful facts regarding the loss function in (23). The gradient of it is given by 
 1 ∇f (X) = PΩ XX > − M \ + E X. (59) p We define the expected gradient (with respect to the sampling set Ω) to be 
 ∇F (X) = XX > − M \ + E X and also the (expected) gradient without noise to be ∇fclean (X) =
1 PΩ XX > − M \ X p and
∇Fclean (X) = XX > − M \ X. (60) In addition, we need the Hessian ∇2 fclean (X), which is represented by an nr×nr matrix. Simple calculations reveal that for any V ∈ Rn×r ,
2
1 1 > vec (V ) ∇2 fclean (X) vec (V ) = PΩ V X > + XV > F + PΩ XX > − M \ , V V > , (61) 2p p where vec(V ) ∈ Rnr denotes the vectorization of V . 26 7.1 7.1.1 Step 1: characterizing local geometry in the RIC Local geometry The first step is to characterize the region where the empirical loss function enjoys restricted strong convexity and smoothness in an appropriate sense. This is formally stated in the following lemma. Lemma 7 (Restricted strong convexity and smoothness for matrix completion). Suppose that the sample size obeys n
2 p ≥ Cκ2 µrn log n for some sufficiently large constant C > 0. Then with probability at least 1 − O n−10 , the Hessian ∇2 fclean (X) as defined in (61) obeys > vec (V ) ∇2 fclean (X) vec (V ) ≥ σmin 2 kV kF 2 ∇2 fclean (X) ≤ and 5 σmax 2 (62) for all X and V = Y HY − Z, with HY := arg minR∈Or×r kY R − ZkF , satisfying: X − X\ where   1/ p κ3 µr log2 n and δ  1/κ. 2,∞ \ ≤  X\ 2,∞ , \ kZ − X k ≤ δkX k, (63a) (63b) Proof. See Appendix B.1. Lemma 7 reveals that the Hessian matrix is well-conditioned in a neighborhood close to X \ that remains incoherent measured in the `2 /`∞ norm (cf. (63a)), and along directions that point towards points which are not far away from the truth in the spectral norm (cf. (63b)). Remark 4. The second condition (63b) is characterized using the spectral norm k·k, while in previous works this is typically presented in the Frobenius norm k · kF . It is also worth noting that the Hessian matrix — even in the infinite-sample and noiseless case — is rank-deficient and cannot be positive definite. As a result, we resort to the form of strong convexity by restricting attention to certain directions (see the conditions on V ). 7.1.2 Error contraction Our goal is to demonstrate the error bounds (28) measured in three different norms. Notably, as long as ct − X \ k2,∞ is sufficiently small. Under our sample the iterates satisfy (28) at the tth iteration, then kX t H ct satisfies the `2 /`∞ condition (63a) required in Lemma 7. Consequently, we complexity assumption, X t H can invoke Lemma 7 to arrive at the following error contraction result. Lemma 8 (Contraction w.r.t. the Frobenius norm). Suppose n2 p ≥ Cκ3 µ3 r3 n log3 n and the noise satisfies (27). If the iterates satisfy (28a) and (28b) at the tth iteration, then with probability at least 1 − O(n−10 ), r σ n \ ct+1 − X \ ≤ C4 ρt+1 µr √1 X t+1 H X + C X\ F 1 F F np σmin p holds as long as 0 < η ≤ 2/(25κσmax ), 1 − (σmin /4) · η ≤ ρ < 1, and C1 is sufficiently large. Proof. The proof is built upon Lemma 7. See Appendix B.2. Further, if the current iterate satisfies all three conditions in (28), then we can derive a stronger sense of error contraction, namely, contraction in terms of the spectral norm. Lemma 9 (Contraction w.r.t. the spectral norm). Suppose n2 p ≥ Cκ3 µ3 r3 n log3 n and the noise satisfies (27). If the iterates satisfy (28) at the tth iteration, then r σ n ct+1 − X \ ≤ C9 ρt+1 µr √1 X t+1 H X \ + C10 X\ (64) np σmin p holds with probability at least 1 − O(n−10 ), provided that 0 < η ≤ 1/ (2σmax ) and 1 − (σmin /3) · η ≤ ρ < 1. 27 Proof. The key observation is this: the iterate that proceeds according to the population-level gradient reduces the error w.r.t. k · k, namely,
ct − η∇Fclean X t H ct − X \ < X t H ct − X \ , X tH ct is sufficiently close to the truth. Notably, the orthonormal matrix H ct is still chosen as long as X t H to be the one that minimizes the k · kF distance (as opposed to k · k), which yields a symmetry property
ct = X t H ct > X \ , crucial for our analysis. See Appendix B.3 for details. X \> X t H 7.2 Step 2: introducing the leave-one-out sequences In order to establish the incoherence properties (28b) for the entire trajectory, which is difficult to deal with directly due to the complicated statistical dependence, we introduce a collection of leave-one-out versions  of {X t }t≥0 , denoted by X t,(l) t≥0 for each 1 ≤ l ≤ n. Specifically, X t,(l) t≥0 is the iterates of gradient descent operating on the auxiliary loss function f (l) (X) := 
 1 PΩ−l XX > − M \ + E 4p 2 F +
1 Pl XX > − M \ 4 2 F . (65) Here, PΩl (resp. PΩ−l and Pl ) represents the orthogonal projection onto the subspace of matrices which vanish outside of the index set Ωl := {(i, j) ∈ Ω | i = l or j = l} (resp. Ω−l := {(i, j) ∈ Ω | i 6= l, j 6= l} and {(i, j) | i = l or j = l}); that is, for any matrix M , ( Mi,j , if (i = l or j = l) and (i, j) ∈ Ω, [PΩl (M )]i,j = (66) 0, else, [PΩ−l (M )]i,j = ( Mi,j , if i 6= l and j 6= l and (i, j) ∈ Ω 0, else and [Pl (M )]i,j = The gradient of the leave-one-out loss function (65) is given by ∇f (l) (X) = ( 0, if i 6= l and j 6= l, Mi,j , if i = l or j = l. (67) 

1 PΩ−l XX > − M \ + E X + Pl XX > − M \ X. p (68) The full algorithm to obtain the leave-one-out sequence {X t,(l) }t≥0 (including spectral initialization) is summarized in Algorithm 5. Algorithm 5 The lth leave-one-out sequence for matrix completion \ \ Input: Y = [Yi,j ]1≤i,j≤n , M·,l , Ml,· , r, p. 0,(l) (l) 0,(l)> Spectral initialization: Let U Σ U be the top-r eigendecomposition of M (l) :=
1

1 P −l (Y ) + Pl M \ = PΩ−l M \ + E + Pl M \ p Ω p with PΩ−l and Pl defined in (67), and set X 0,(l) = U 0,(l) Σ(l) Gradient updates: for t = 0, 1, 2, . . . , T − 1 do
1/2 .
X t+1,(l) = X t,(l) − ηt ∇f (l) X t,(l) . (69) Remark 5. Rather than simply dropping all samples in the lth row/column, we replace the lth row/column with their respective population means. In other words, the leave-one-out gradient forms an unbiased surrogate for the true gradient, which is particularly important in ensuring high estimation accuracy. 28 7.3 Step 3: establishing the incoherence condition by induction We will continue the proof of Theorem 2 in an inductive manner. As seen in Section 7.1.2, the induction hypotheses (28a) and (28c) hold for the (t+1)th iteration as long as (28) holds at the tth iteration. Therefore, we are left with proving the incoherence hypothesis (28b) for all 0 ≤ t ≤ T = O(n5 ). For clarity of analysis, it is crucial to maintain a list of induction hypotheses, which includes a few more hypotheses that complement (28), and is given below.  r  n ct − X \ ≤ C4 ρt µr √1 + C1 σ (70a) X\ F , X tH F np σmin p s s ! log n σ n log n ct − X \ X tH ≤ C5 ρt µr X \ 2,∞ , (70b) + C8 2,∞ np σmin p  r  n ct − X \ ≤ C9 ρt µr √1 + C10 σ (70c) X tH X\ , np σmin p s s ! log n n log n σ ct − X t,(l) Rt,(l) ≤ C3 ρt µr max X t H X \ 2,∞ , (70d) + C7 F 1≤l≤n np σmin p s !
1 n log n σ t,(l) ct,(l) \ t H − X l,· 2 ≤ C2 ρ µr √ + C6 max X X \ 2,∞ (70e) 1≤l≤n np σmin p ct,(l) and Rt,(l) are orthonormal hold for some absolute constants 0 < ρ < 1 and C1 , · · · , C10 > 0. Here, H matrices defined by ct,(l) := arg min X t,(l) R − X \ H r×r R∈O Rt,(l) := arg min R∈O r×r F (71) , ct X t,(l) R − X t H F . (72) Clearly, the first three hypotheses (70a)-(70c) constitute the conclusion of Theorem 2, i.e. (28). The last two hypotheses (70d) and (70e) are auxiliary properties connecting the true iterates and the auxiliary leave-oneout sequences. Moreover, we summarize below several immediate consequences of (70), which will be useful throughout. Lemma 10. Suppose n2 p  κ3 µ2 r2 n log n and the noise satisfies (27). Under the hypotheses (70), one has ct − X t,(l) H ct,(l) X tH ct,(l) − X \ X t,(l) H X t,(l) Rt,(l) − X \ F F 2,∞ ct,(l) − X \ X t,(l) H ct − X t,(l) Rt,(l) , ≤ 5κ X t H F r  1 σ n t,(l) t,(l) \ t ≤ X R −X ≤ 2C4 ρ µr √ + 2C1 X\ F , np σmin p F s s ( ) log n σ n log n t ≤ (C3 + C5 ) ρ µr + (C8 + C7 ) X \ 2,∞ , np σmin p  r  σ 1 n t ≤ 2C9 ρ µr √ + 2C10 X\ . np σmin p (73a) (73b) (73c) (73d) In particular, (73a) follows from hypotheses (70c) and (70d). Proof. See Appendix B.4. In the sequel, we follow the general recipe outlined in Section 5 to establish the induction hypotheses. We only need to establish (70b), (70d) and (70e) for the (t + 1)th iteration, since (70a) and (70c) have been established in Section 7.1.2. Specifically, we resort to the leave-one-out iterates by showing that: first, the true and the auxiliary iterates remain exceedingly close throughout; second, the lth leave-one-out sequence stays incoherent with el due to statistical independence. 29 • Step 3(a): proximity between the original and the leave-one-out iterates. We demonstrate that X t+1 is well approximated by X t+1,(l) , up to proper orthonormal transforms. This is precisely the induction hypothesis (70d) for the (t + 1)th iteration. Lemma 11. Suppose the sample complexity satisfies n2 p  κ4 µ3 r3 n log3 n and the noise satisfies (27). Under the hypotheses (70) for the tth iteration, we have s s σ log n n log n ct+1 − X t+1,(l) Rt+1,(l) ≤ C3 ρt+1 µr X \ 2,∞ + C7 X \ 2,∞ (74) X t+1 H np σmin p F with probability at least 1 − O(n−10 ), provided that 0 < η ≤ 2/(25κσmax ), 1 − (σmin /5) · η ≤ ρ < 1 and C7 > 0 is sufficiently large. Proof. The fact that this difference is well-controlled relies heavily on the benign geometric property of the ct and X t,(l) Rt,(l) Hessian revealed by Lemma 7. Two important remarks are in order: (1) both points X t H ct − X t,(l) Rt,(l) forms a valid direction for restricted strong convexity. satisfy (63a); (2) the difference X t H These two properties together allow us to invoke Lemma 7. See Appendix B.5. • Step 3(b): incoherence of the leave-one-out iterates. Given that X t+1,(l) is sufficiently close to X t+1 , we turn our attention to establishing the incoherence of this surrogate X t+1,(l) w.r.t. el . This amounts to proving the induction hypothesis (70e) for the (t + 1)th iteration. Lemma 12. Suppose the sample complexity meets n2 p  κ3 µ3 r3 n log3 n and the noise satisfies (27). Under the hypotheses (70) for the tth iteration, one has s
1 n log n σ ct+1,(l) − X \ X t+1,(l) H ≤ C2 ρt+1 µr √ X \ 2,∞ + C6 X \ 2,∞ (75) l,· 2 np σmin p −10 with probability √ at least 1 − O(n ), as long as 0 < η ≤ 1/σmax , 1 − (σmin /3) · η ≤ ρ < 1, C2  κC9 and C6  κC10 / log n. Proof. The key observation is that X t+1,(l) is statistically independent from any sample in the lth row/column of the matrix. Since there are an order of np samples in each row/column, we obtain enough information that helps establish the desired incoherence property. See Appendix B.6. • Step 3(c): combining the bounds. The inequalities (70d) and (70e) taken collectively allow us to establish the induction hypothesis (70b). Specifically, for every 1 ≤ l ≤ n, write ct+1 − X \ X t+1 H
l,· ct+1 − X t+1,(l) H ct+1,(l) = X t+1 H and the triangle inequality gives
ct+1 − X \ ct+1 − X t+1,(l) H ct+1,(l) X t+1 H ≤ X t+1 H l,· 2
F l,· + ct+1,(l) − X \ + X t+1,(l) H ct+1,(l) − X \ X t+1,(l) H

l,· , l,· 2 . (76) The second term has already been bounded by (75). Since we have established the induction hypotheses (70c) and (70d) for the (t+1)th iteration, the first term can be bounded by (73a) for the (t+1)th iteration, i.e. ct+1 − X t+1,(l) H ct+1,(l) ≤ 5κ X t+1 H ct+1 − X t+1,(l) Rt+1,(l) . X t+1 H F F Plugging the above inequality, (74) and (75) into (76), we have s s C log n n log n 7 ct+1 − X \ X \ 2,∞ + σ X\ X t+1 H ≤ 5κ C3 ρt+1 µr np σmin p 2,∞ 30 2,∞ ! 1 + C2 ρ µr √ X\ np s log n X\ ≤ C5 ρt+1 µr np t+1 C6 + σ 2,∞ σmin s n log n X\ p s C8 n log n + X\ σ 2,∞ σmin p 2,∞ 2,∞ as long as C5 /(κC3 +C2 ) and C8 /(κC7 +C6 ) are sufficiently large. This establishes the induction hypothesis (70b) and finishes the proof. 7.4 The base case: spectral initialization Finally, we return to check the base case, namely, we aim to show that the spectral initialization satisfies the induction hypotheses (70a)-(70e) for t = 0. This is accomplished via the following lemma. Lemma 13. Suppose the sample size obeys n2
p  µ2 r2 n log n, the noise satisfies (27), and κ = σmax /σmin  1. Then with probability at least 1 − O n−10 , the claims in (70a)-(70e) hold simultaneously for t = 0. Proof. This follows by invoking the Davis-Kahan sinΘ theorem [DK70] as well as the entrywise eigenvector perturbation analysis in [AFWZ17]. We defer the proof to Appendix B.7. 8 Analysis for blind deconvolution In this section, we instantiate the general recipe presented in Section 5 to blind deconvolution and prove Theorem 3. Without loss of generality, we assume throughout that h\ 2 = x\ 2 = 1. Before presenting the analysis, we first gather some simple facts about the empirical loss function in (32). Recall the definition of z in (33), and for notational simplicity, we write f (z) = f (h, x). Since z is complex-valued, we need to resort to Wirtinger calculus; see [CLS15, Section 6] for a brief introduction. The Wirtinger gradient of (32) with respect to h and x are given respectively by ∇h f (z) = ∇h f (h, x) = ∇x f (z) = ∇x f (h, x) = m X j=1 m X j=1
b∗j hx∗ aj − yj bj a∗j x; (b∗j hx∗ aj − yj )aj b∗j h. (77) (78) It is worth noting that the formal Wirtinger gradient contains ∇h f (h, x) and ∇x f (h, x) as well. Nevertheless, since f (h, x) is a real-valued function, the following identities always hold and ∇h f (h, x) = ∇h f (h, x) ∇x f (h, x) = ∇x f (h, x). In light of these observations, one often omits the gradient with respect to the conjugates; correspondingly, the gradient update rule (35) can be written as ht+1 = ht − xt+1 = xt − m X η 2 kxt k2 j=1 m X η 2 kht k2 j=1
b∗j ht xt∗ aj − yj bj a∗j xt , (b∗j ht xt∗ aj − yj )aj b∗j ht . We can also compute the Wirtinger Hessian of f (z) as follows,   A B 2 ∇ f (z) = , B∗ A where A= " Pm ∗ 2 ∗ j=1 aj x bj bj
∗ Pm  ∗ ∗ bj hx aj − yj bj a∗j j=1 #
∗ ∗ ∗ b hx a − y b a j j j j j=1 ∈ C2K×2K ; Pmj 2 ∗ ∗ j=1 bj h aj aj Pm 31 (79a) (79b) (80) B= " Pm 0 j=1
> ∗ ∗ j=1 aj aj x bj bj h Pm bj b∗j h aj a∗j x 0
> # ∈ C2K×2K . Last but not least, we say (h1 , x1 ) is aligned with (h2 , x2 ), if the following holds, ( ) 2 1 2 2 2 kh1 − h2 k2 + kx1 − x2 k2 = min h1 − h2 + kαx1 − x2 k2 . α∈C α 2 To simplify notations, define zet as    1 t et h h ze = t := αtt t e αx x t (81) with the alignment parameter αt given in (38). Then we can see that zet is aligned with z \ and

dist z t , z \ = dist zet , z \ = zet − z \ 2 . 8.1 8.1.1 Step 1: characterizing local geometry in the RIC Local geometry The first step is to characterize the region of incoherence and contraction (RIC), where the empirical loss function enjoys restricted strong convexity and smoothness properties. To this end, we have the following lemma. Lemma 14 (Restricted strong convexity and smoothness for blind deconvolution). Let c > 0 be a sufficiently small constant and δ = c/ log2 m. Suppose the sample size satisfies m ≥ c0 µ2 K
log9 m for some sufficiently large constant c0 > 0. Then with probability at least 1 − O m−10 + e−K log m , the Wirtinger Hessian ∇2 f (z) obeys   2 u∗ D∇2 f (z) + ∇2 f (z) D u ≥ (1/4) · kuk2 and ∇2 f (z) ≤ 3 simultaneously for all z=  h x  where z satisfies and   h1 − h2  x1 − x2   u=  h1 − h2  x1 − x2 max  h − h\ 2 , x − x\  D=  γ 2 IK γ 1 IK γ 2 IK 2 2 , x1 − x\ 2  ,  (82a) ≤ δ;
x − x\ ≤ 2C3 (h1 , x1 ) is aligned with (h2 , x2 ), and they satisfy  max h1 − h\ 2 , h2 − h\  γ 1 IK 1 ; log3/2 m µ max b∗j h ≤ 2C4 √ log2 m; 1≤j≤m m max a∗j 1≤j≤m and  , x2 − x\ 2 (82b) (82c) ≤ δ; (83) and finally, D satisfies for γ1 , γ2 ∈ R, max {|γ1 − 1| , |γ2 − 1|} ≤ δ. Here, C3 , C4 > 0 are numerical constants. 32 (84) Proof. See Appendix C.1. Lemma 14 characterizes the restricted strong convexity and smoothness of the loss function used in blind deconvolution. To the best of our knowledge, this provides the first characterization regarding geometric properties of the Hessian matrix for blind deconvolution. A few interpretations are in order. • The conditions (82) specify the region of incoherence and contraction (RIC). In particular, (82a) specifies a neighborhood that is close to the ground truth in `2 norm, and (82b) and (82c) specify the incoherence region with respect to the sensing vectors {aj } and {bj }, respectively. • Similar to matrix completion, the Hessian matrix is rank-deficient even at the population level. Consequently, we resort to a restricted form of strong convexity by focusing on certain directions. More specifically, these directions can be viewed as the difference between two pre-aligned points that are not far from the truth, which is characterized by (83). • Finally, the diagonal matrix D accounts for scaling factors that are not too far from 1 (see (84)), which allows us to account for different step sizes employed for h and x. 8.1.2 Error contraction The restricted strong convexity and smoothness allow us to establish the contraction of the error measured in terms of dist(·, z \ ) as defined in (34) as long as the iterates stay in the RIC. Lemma 15. Suppose the number of measurements satisfies m  µ2 K log9 m, and the step size η > 0 is some sufficiently small constant. Then

dist z t+1 , z \ ≤ (1 − η/16) dist z t , z \ , provided that
dist z t , z \ ≤ ξ,
e t − x\ ≤ C3 max a∗j x 1≤j≤m 1 , log1.5 m e t ≤ C4 √µ log2 m max b∗j h 1≤j≤m m (85a) (85b) (85c) e t and x et are defined in (81), and ξ  1/ log2 m. for some constants C3 , C4 > 0. Here, h Proof. See Appendix C.2. As a result, if z t satisfies the condition (85) for all 0 ≤ t ≤ T , then the union bound gives


dist z t , z \ ≤ ρ dist z t−1 , z \ ≤ ρt dist z 0 , z \ ≤ ρt c1 , 0 < t ≤ T, where ρ := 1 − η/16. Furthermore, similar to the case of phase retrieval (i.e. Lemma 3), as soon as we demonstrate that the conditions (85) hold for all 0 ≤ t ≤ m, then Theorem 3 holds true. The proof of this claim is exactly the same as for Lemma 3, and is thus omitted for conciseness. In what follows, we focus on establishing (85) for all 0 ≤ t ≤ m. Before concluding this subsection, we make note of another important result that concerns the alignment parameter αt , which will be useful in the subsequent analysis. Specifically, the alignment parameter sequence {αt } converges linearly to a constant whose magnitude is fairly close to 1, as long as the two initial vectors h0 and x0 have similar `2 norms and are close to the truth. Given that αt determines the global scaling of the iterates, this reveals rapid convergence of both kht k2 and kxt k2 , which explains why there is no need to impose extra terms to regularize the `2 norm as employed in [LLSW16, HH17]. Lemma 16. Suppose that m  1. The following two claims hold true. 33 • If |αt | − 1 ≤ 1/2 and dist(z t , z \ ) ≤ C1 / log2 m, then αt+1 cC1 − 1 ≤ c dist(z t , z \ ) ≤ αt log2 m for some absolute constant c > 0; • If |α0 | − 1 ≤ 1/4 and dist(z s , z \ ) ≤ C1 (1 − η/16)s / log2 m for all 0 ≤ s ≤ t, then one has |αs+1 | − 1 ≤ 1/2, 0 ≤ s ≤ t. Proof. See Appendix C.2. The initial condition |α0 | − 1 < 1/4 will be guaranteed to hold with high probability by Lemma 19. 8.2 Step 2: introducing the leave-one-out sequences As demonstrated by the assumptions in Lemma 15, the key is to show that the whole trajectory lies in the region specified by (85a)-(85c). Once again, the difficulty lies in the statistical dependency between the iterates {z t } and the measurement vectors {aj }. We follow the general recipe  and introduce the leave-one out sequences, denoted by ht,(l) , xt,(l) t≥0 for each 1 ≤ l ≤ m. Specifically, ht,(l) , xt,(l) t≥0 is the gradient sequence operating on the loss function X
2 b∗j hx∗ − h\ x\∗ aj . (86) f (l) (h, x) := j:j6=l The whole sequence is constructed by running gradient descent with spectral initialization on the leave-oneout loss (86). The precise description is supplied Algorithm 6.  int,(l)  h t,(l) For notational simplicity, we denote z = and use f (z t,(l) ) = f (ht,(l) , xt,(l) ) interchangeably. xt,(l) Define similarly the alignment parameters αt,(l) := arg min α∈C and denote zet,(l) =  e t,(l) h et,(l) x  1 t,(l) h − h\ α 2 2 + αxt,(l) − x\ 2 , 2 (87) where e t,(l) = h 1 αt,(l) ht,(l) and et,(l) = αt,(l) xt,(l) . x (88) Algorithm 6 The lth leave-one-out sequence for blind deconvolution Input: {aj }1≤j≤m,j6=l , {bj }1≤j≤m,j6=l and {yj }1≤j≤m,j6=l . Spectral initialization: Let σ1 (M (l) ), ȟ0,(l) and x̌0,(l) be the leading singular value, left and right singular vectors of X M (l) := yj bj a∗j , j:j6=l p p respectively. Set h0,(l) = σ1 (M (l) ) ȟ0,(l) and x0,(l) = σ1 (M (l) ) x̌0,(l) . Gradient updates: for t = 0, 1, 2, . . . , T − 1 do "
#  t+1,(l)   t,(l)  1 ∇ f (l) ht,(l) , xt,(l) h h kxt,(l) k22 h
. = −η 1 xt+1,(l) xt,(l) ∇ f (l) ht,(l) , xt,(l) kht,(l) k2 x 2 34 (89) 8.3 Step 3: establishing the incoherence condition by induction As usual, we continue the proof in an inductive manner. For clarity of presentation, we list below the set of induction hypotheses underlying our analysis:
dist z t , z \ ≤ C1 1 , log2 m s
µ µ2 K log9 m max dist z t,(l) , zet ≤ C2 √ , 1≤l≤m m m
1 et − x\ ≤ C3 1.5 , max a∗l x 1≤l≤m log m e t ≤ C4 √µ log2 m, max b∗l h 1≤l≤m m (90a) (90b) (90c) (90d) e t, x et and zet are defined in (81). Here, C1 , C3 > 0 are some sufficiently small constants, while where h C2 , C4 > 0 are some sufficiently large constants. We aim to show that if these hypotheses (90) hold up to the tth iteration, then the same would hold for the (t + 1)th iteration with exceedingly high probability (e.g. 1 − O(m−10 )). The first hypothesis (90a) has already been established in Lemma 15, and hence the rest of this section focuses on establishing the remaining three. To justify the incoherence hypotheses (90c) and (90d) for the (t + 1)th iteration, we need to leverage the nice properties of the leave-one-out sequences, and establish (90b) first. In the sequel, we follow the steps suggested in the general recipe. • Step 3(a): proximity between the original and the leave-one-out iterates. We first justify the hypothesis (90b) for the (t + 1)th iteration via the following lemma. Lemma 17. Suppose the sample complexity obeys m  µ2 K log9 m. There exists a constant c > 0 such that under the hypotheses (90a)-(90d) for the tth iteration, with probability at least 1 − O(m−10 + me−cK ) one has s
µ µ2 K log9 m max dist z t+1,(l) , zet+1 ≤ C2 √ 1≤l≤m m m s µ µ2 K log9 m and max zet+1,(l) − zet+1 2 . C2 √ , 1≤l≤m m m provided that the step size η > 0 is some sufficiently small constant. Proof. As usual, this result follows from the restricted strong convexity, which forces the distance between the two sequences of interest to be contractive. See Appendix C.3. • Step 3(b): incoherence of the leave-one-out iterate xt+1,(l) w.r.t. al . Next, we show that the et+1,(l) — which is independent of al — is incoherent w.r.t. al in the sense that leave-one-out iterate x et+1,(l) − x\ a∗l x
≤ 10C1 1 log 3/2 (91) m
with probability exceeding 1 − O m−10 + e−K log m . To see why, use the statistical independence and the standard Gaussian concentration inequality to show that et+1,(l) − x\ max a∗l x 1≤l≤m
p ≤ 5 log m max 1≤l≤m et+1,(l) − x\ x 2 with probability exceeding 1 − O(m−10 ). It then follows from the triangle inequality that et+1,(l) − x\ x 2 et+1,(l) − x et+1 ≤ x 35 2 et+1 − x\ + x 2 (i) µ ≤ CC2 √ m (ii) ≤ 2C1 s µ2 K log9 m 1 + C1 2 m log m 1 , log2 m √ where (i) follows from Lemmas 15 and 17, and (ii) holds as soon as m  µ2 K log13/2 m. Combining the preceding two bounds establishes (91). • Step 3(c): combining the bounds to show incoherence of xt+1 w.r.t. {al }. The above bounds immediately allow us to conclude that et+1 − x\ max a∗l x
≤ C3 1 3/2 log m
with probability at least 1 − O m−10 + e−K log m , which is exactly the hypothesis (90c) for the (t + 1)th iteration. Specifically, for each 1 ≤ l ≤ m, the triangle inequality yields 1≤l≤m et+1 − x\ a∗l x
et+1 − x et+1,(l) ≤ a∗l x
et+1,(l) − x\ + a∗l x

et+1 − x et+1,(l) + a∗l x et+1,(l) − x\ ≤ kal k2 x 2 s (ii) √ µ 1 µ2 K log9 m ≤ 3 K · CC2 √ + 10C1 3/2 m m log m (iii) 1 ≤ C3 3/2 . log m (i) Here (i) follows from Cauchy-Schwarz, (ii) is a consequence of (190), Lemma 17 and the bound (91), and the last inequality holds as long as m  µ2 K log6 m and C3 ≥ 11C1 . • Step 3(d): incoherence of ht+1 w.r.t. {bl }. It remains to justify that ht+1 is also incoherent w.r.t. its associated design vectors {bl }. This proof of this step, however, is much more involved and challenging, due to the deterministic nature of the bl ’s. As a result, we would need to “propagate” the randomness brought about by {al } to ht+1 in order to facilitate the analysis. The result is summarized as follows. Lemma 18. Suppose that the sample complexity obeys m  µ2 K log9 m. Under the inductive hypotheses (90a)-(90d) for the tth iteration, with probability exceeding 1 − O(m−10 ) we have e t+1 ≤ C4 √µ log2 m max b∗l h m 1≤l≤m as long as C4 is sufficiently large, and η > 0 is taken to be some sufficiently small constant. Proof. The key idea is to divide {1, · · · , m} into consecutive bins each of size poly log(m), and to exploit the randomness (namely, the randomness from al ) within each bin. This binning idea is crucial in ensuring that the incoherence measure of interest does not blow up as t increases. See Appendix C.4. With these steps in place, we conclude the proof of Theorem 3 via induction and the union bound. 8.4 The base case: spectral initialization In order to finish the induction steps, we still need to justify the  induction hypotheses for the base cases, namely, we need to show that the spectral initializations z 0 and z 0,(l) 1≤l≤m satisfy the induction hypotheses (90) at t = 0. To start with, the initializations are sufficiently close to the truth when measured by the `2 norm, as summarized by the following lemma. 36 Lemma 19. Fix any small constant ξ > 0. Suppose the sample size obeys m ≥ Cµ2 K log2 m/ξ 2 for some sufficiently large constant C > 0. Then with probability at least 1 − O(m−10 ), we have  and (92) αh0 − h\ 2 + αx0 − x\ 2 ≤ ξ min α∈C,|α|=1 n o min αh0,(l) − h\ 2 + αx0,(l) − x\ 2 ≤ ξ, 1 ≤ l ≤ m, (93) α∈C,|α|=1 and ||α0 | − 1| ≤ 1/4. Proof. This follows from Wedin’s sinΘ theorem [Wed72] and [LLSW16, Lemma 5.20]. See Appendix C.5. From the definition of dist(·, ·) (cf. (34)), we immediately have s  2 (i)
1 1 2 0 \ \ \ dist z , z = min + kαx − x k2 ≤ min h−h h − h\ α∈C α∈C α α 2 (ii) ≤ min α∈C,|α|=1  αh0 − h\ 2 + αx0 − x\ (iii) ≤ C1 2 2 + αx − x \ 2  1 , log2 m (94) as long as m ≥ Cµ2 K log6 m for some sufficiently large constant C > 0. Here (i) follows from the elementary 2 inequality that a2 + b2 ≤ (a + b) for positive a and b, (ii) holds since the feasible set of the latter one is strictly smaller, and (iii) follows directly from Lemma 19. This finishes the proof of (90a) for t = 0. Similarly, with high probability we have
dist z 0,(l) , z \ ≤ min α∈C,|α|=1 n αh0,(l) − h\ + αx0,(l) − x\ 2 2 o . 1 , log2 m 1 ≤ l ≤ m. (95) Next, when properly aligned, the true initial estimate z 0 and the leave-one-out estimate z 0,(l) are expected to be sufficiently close, as claimed by the following lemma. Along the way, we show that h0 is incoherent w.r.t. the sampling vectors {bl }. This establishes (90b) and (90d) for t = 0. Lemma 20. Suppose that m  µ2 K log3 m. Then with probability at least 1 − O(m−10 ), one has s
µ2 K log5 m µ max dist z 0,(l) , ze0 ≤ C2 √ 1≤l≤m m m (96) and e 0 ≤ C4 max b∗l h 1≤l≤m µ log2 m √ . m (97)
e0 , Proof. The key is to establish that dist z 0,(l) , ze0 can be upper bounded by some linear scaling of b∗l h and vice versa. This allows us to derive bounds simultaneously for both quantities. See Appendix C.6. Finally, we establish (90c) regarding the incoherence of x0 with respect to the design vectors {al }. Lemma 21. Suppose that m  µ2 K log6 m. Then with probability exceeding 1 − O(m−10 ), we have e 0 − x\ max a∗l x 1≤l≤m Proof. See Appendix C.7.
37 ≤ C3 1 log 1.5 m . 9 Discussions This paper showcases an important phenomenon in nonconvex optimization: even without explicit enforcement of regularization, the vanilla form of gradient descent effectively achieves implicit regularization for a large family of statistical estimation problems. We believe this phenomenon arises in problems far beyond the three cases studied herein, and our results are initial steps towards understanding this fundamental phenomenon. That being said, there are numerous avenues open for future investigation, and we conclude the paper with a few of them.
• Improving sample complexity. In the current paper, the required sample complexity O µ3 r3 n log3 n for matrix completion is sub-optimal when the rank r of the underlying matrix is large. While this allows us to achieve a dimension-free iteration complexity, it is slightly higher than the sample complexity derived for regularized gradient descent in [CW15]. We expect our results continue to hold under lower sample
complexity O µ2 r2 n log n , but it calls for a more refined analysis (e.g. a generic chaining argument). • Leave-one-out tricks for more general designs. So far our focus is on independent designs, including the i.i.d. Gaussian design adopted in phase retrieval and partially in blind deconvolution, as well as the independent sampling mechanism in matrix completion. Such independence property creates some sort of “statistical homogeneity”, for which the leave-one-out argument works beautifully. It remains unclear how to generalize such leave-one-out tricks for more general designs (e.g. more general sampling patterns in matrix completion and more structured Fourier designs in phase retrieval and blind deconvolution). In fact, the readers can already get a flavor of this issue in the analysis of blind deconvolution, where the Fourier design vectors require much more delicate treatments than purely Gaussian designs. • Uniform stability. The leave-one-out perturbation argument is established upon a basic fact: when we exclude one sample from consideration, the resulting estimates/predictions do not deviate much from the original ones. This leave-one-out stability bears similarity to the notion of uniform stability studied in statistical learning theory [BE02, LLNT17]. We expect our analysis framework to be helpful for analyzing other learning algorithms that are uniformly stable. • Constrained optimization. We restrict ourselves to study empirical risk minimization problems in an unconstrained setting. It will be interesting to explore if such implicit regularization still holds for constrained nonconvex problems. • Other iterative methods. Iterative methods other than gradient descent have been extensively studied in the nonconvex optimization literature, including alternating minimization, proximal methods, etc. Identifying the implicit regularization feature for a broader class of iterative algorithms is another direction worth exploring. • Connections to deep learning? We have focused on nonlinear systems that are bilinear or quadratic in this paper. Deep learning formulations/architectures, highly nonlinear, are notorious for its daunting nonconvex geometry. However, iterative methods including stochastic gradient descent have enjoyed enormous practical success in learning neural networks (e.g. [ZSJ+ 17, SJL17]), even when the architecture is significantly over-parameterized without explicit regularization. We hope the message conveyed in this paper for several simple statistical models can shed light on why simple forms of gradient descent and variants work so well in learning complicated neural networks. Acknowledgements The work of Y. Chi is supported in part by the grants AFOSR FA9550-15-1-0205, ONR N00014-15-1-2387, NSF CCF-1527456, ECCS-1650449 and CCF-1704245. Y. Chen would like to thank Yudong Chen for inspiring discussions about matrix completion. 38 References [AAH17] A. Aghasi, A. Ahmed, and P. Hand. Branchhull: Convex bilinear inversion from the entrywise product of signals with known signs. arXiv preprint arXiv:1702.04342, 2017. [AFWZ17] E. Abbe, J. Fan, K. Wang, and Y. Zhong. Entrywise eigenvector analysis of random matrices with low expected rank. arXiv preprint arXiv:1709.09565, 2017. [ARR14] A. Ahmed, B. Recht, and J. Romberg. Blind deconvolution using convex programming. IEEE Transactions on Information Theory, 60(3):1711–1732, 2014. [AS08] N. Alon and J. H. Spencer. The Probabilistic Method (3rd Edition). Wiley, 2008. [BE02] O. Bousquet and A. Elisseeff. Stability and generalization. Journal of Machine Learning Research, 2(Mar):499–526, 2002. [BEB17] T. Bendory, Y. C. Eldar, and N. Boumal. Non-convex phase retrieval from STFT measurements. IEEE Transactions on Information Theory, 2017. [BNS16] S. Bhojanapalli, B. Neyshabur, and N. Srebro. Global optimality of local search for low rank matrix recovery. In Advances in Neural Information Processing Systems, pages 3873–3881, 2016. [BR17] S. Bahmani and J. Romberg. Phase retrieval meets statistical learning theory: A flexible convex relaxation. In Artificial Intelligence and Statistics, pages 252–260, 2017. [Bub15] S. Bubeck. Convex optimization: Algorithms and complexity. Foundations and Trends R in Machine Learning, 8(3-4):231–357, 2015. [CC16] Y. Chen and E. Candès. The projected power method: An efficient algorithm for joint alignment from pairwise differences. arXiv preprint arXiv:1609.05820, accepted to Communications on Pure and Applied Mathematics, 2016. [CC17] Y. Chen and E. J. Candès. Solving random quadratic systems of equations is nearly as easy as solving linear systems. Comm. Pure Appl. Math., 70(5):822–883, 2017. [CCG15] Y. Chen, Y. Chi, and A. J. Goldsmith. Exact and stable covariance estimation from quadratic sampling via convex programming. IEEE Transactions on Information Theory, 61(7):4034– 4059, 2015. [CESV13] E. J. Candès, Y. C. Eldar, T. Strohmer, and V. Voroninski. Phase retrieval via matrix completion. SIAM Journal on Imaging Sciences, 6(1):199–225, 2013. [CFL15] P. Chen, A. Fannjiang, and G.-R. Liu. Phase retrieval with one or two diffraction patterns by alternating projections with the null initialization. Journal of Fourier Analysis and Applications, pages 1–40, 2015. [CFMW17] Y. Chen, J. Fan, C. Ma, and K. Wang. Spectral method and regularized MLE are both optimal for top-k ranking. arXiv preprint arXiv:1707.09971, 2017. [Che15] Y. Chen. Incoherence-optimal matrix completion. IEEE Transactions on Information Theory, 61(5):2909–2923, 2015. [Che17] Y. Chen. Regularized mirror descent: A nonconvex approach for learning mixed probability distributions. 2017. [Chi16] Y. Chi. Guaranteed blind sparse spikes deconvolution via lifting and convex optimization. IEEE Journal of Selected Topics in Signal Processing, 10(4):782–794, 2016. [CJ16] V. Cambareri and L. Jacques. A non-convex blind calibration method for randomised sensing strategies. arXiv preprint arXiv:1605.02615, 2016. 39 [CJN17] Y. Cherapanamjeri, P. Jain, and P. Netrapalli. Thresholding based outlier robust PCA. In Conference on Learning Theory, pages 593–628, 2017. [CL14] E. J. Candès and X. Li. Solving quadratic equations via PhaseLift when there are about as many equations as unknowns. Foundations of Computational Mathematics, 14(5):1017–1026, 2014. [CL16] Y. Chi and Y. M. Lu. Kaczmarz method for solving quadratic equations. IEEE Signal Processing Letters, 23(9):1183–1187, 2016. [CLM+ 16] T. T. Cai, X. Li, Z. Ma, et al. Optimal rates of convergence for noisy sparse phase retrieval via thresholded Wirtinger flow. The Annals of Statistics, 44(5):2221–2251, 2016. [CLMW11] E. J. Candès, X. Li, Y. Ma, and J. Wright. Robust principal component analysis? Journal of ACM, 58(3):11:1–11:37, Jun 2011. [CLS15] E. J. Candès, X. Li, and M. Soltanolkotabi. Phase retrieval via Wirtinger flow: Theory and algorithms. IEEE Transactions on Information Theory, 61(4):1985–2007, April 2015. [CLW17] J.-F. Cai, H. Liu, and Y. Wang. Fast rank one alternating minimization algorithm for phase retrieval. arXiv preprint arXiv:1708.08751, 2017. [CR09] E. J. Candès and B. Recht. Exact matrix completion via convex optimization. Foundations of Computational Mathematics, 9(6):717–772, April 2009. [CSPW11] V. Chandrasekaran, S. Sanghavi, P. A. Parrilo, and A. S. Willsky. Rank-sparsity incoherence for matrix decomposition. SIAM Journal on Optimization, 21(2):572–596, 2011. [CSV13] E. J. Candès, T. Strohmer, and V. Voroninski. Phaselift: Exact and stable signal recovery from magnitude measurements via convex programming. Communications on Pure and Applied Mathematics, 66(8):1017–1026, 2013. [CT10] E. Candès and T. Tao. The power of convex relaxation: Near-optimal matrix completion. IEEE Transactions on Information Theory, 56(5):2053 –2080, May 2010. [CW15] Y. Chen and M. J. Wainwright. Fast low-rank estimation by projected gradient descent: General statistical and algorithmic guarantees. arXiv preprint arXiv:1509.03025, 2015. [CYC14] Y. Chen, X. Yi, and C. Caramanis. A convex formulation for mixed regression with two components: Minimax optimal rates. In Conference on Learning Theory, pages 560–604, 2014. [CZ15] T. Cai and A. Zhang. ROP: Matrix recovery via rank-one projections. The Annals of Statistics, 43(1):102–138, 2015. [DK70] C. Davis and W. M. Kahan. The rotation of eigenvectors by a perturbation. iii. SIAM Journal on Numerical Analysis, 7(1):1–46, 1970. [Dop00] F. M. Dopico. A note on sin Θ theorems for singular subspace variations. BIT, 40(2):395–403, 2000. [DR16] M. A. Davenport and J. Romberg. An overview of low-rank matrix recovery from incomplete observations. IEEE Journal of Selected Topics in Signal Processing, 10(4):608–622, 2016. [DR17] J. C. Duchi and F. Ruan. Solving (most) of a set of quadratic equalities: Composite optimization for robust phase retrieval. arXiv preprint arXiv:1705.02356, 2017. [DTL17] O. Dhifallah, C. Thrampoulidis, and Y. M. Lu. Phase retrieval via linear programming: Fundamental limits and algorithmic improvements. arXiv preprint arXiv:1710.05234, 2017. 40 [EK15] N. El Karoui. On the impact of predictor geometry on the performance on high-dimensional ridge-regularized generalized robust regression estimators. Probability Theory and Related Fields, pages 1–81, 2015. [EKBB+ 13] N. El Karoui, D. Bean, P. J. Bickel, C. Lim, and B. Yu. On robust regression with highdimensional predictors. Proceedings of the National Academy of Sciences, 110(36):14557–14562, 2013. [GLM16] R. Ge, J. D. Lee, and T. Ma. Matrix completion has no spurious local minimum. In Advances in Neural Information Processing Systems, pages 2973–2981, 2016. [GM17] R. Ge and T. Ma. On the optimization landscape of tensor decompositions. arXiv preprint arXiv:1706.05598, 2017. [Gro11] D. Gross. Recovering low-rank matrices from few coefficients in any basis. IEEE Transactions on Information Theory, 57(3):1548–1566, March 2011. [GS16] T. Goldstein and C. Studer. Phasemax: Convex phase retrieval via basis pursuit. arXiv preprint arXiv:1610.07531, 2016. [GWB+ 17] S. Gunasekar, B. Woodworth, S. Bhojanapalli, B. Neyshabur, and N. Srebro. Implicit regularization in matrix factorization. arXiv preprint arXiv:1705.09280, 2017. [GX16] B. Gao and Z. Xu. arXiv:1606.08135, 2016. Phase retrieval using Gauss-Newton method. [HH17] W. Huang and P. Hand. Blind deconvolution by a steepest descent algorithm on a quotient manifold. arXiv preprint arXiv:1710.03309, 2017. [Hig92] N. J. Higham. Estimating the matrix p-norm. Numerische Mathematik, 62(1):539–555, 1992. [HKZ12] D. Hsu, S. M. Kakade, and T. Zhang. A tail inequality for quadratic forms of subgaussian random vectors. Electron. Commun. Probab., 17:no. 52, 6, 2012. [HMLZ15] T. Hastie, R. Mazumder, J. D. Lee, and R. Zadeh. Matrix completion and low-rank SVD via fast alternating least squares. Journal of Machine Learning Research, 16:3367–3402, 2015. [HV16] P. Hand and V. Voroninski. An elementary proof of convex phase retrieval in the natural parameter space via the linear program PhaseMax. arXiv preprint arXiv:1611.03935, 2016. [HW14] M. Hardt and M. Wootters. Fast matrix completion without the condition number. Conference on Learning Theory, pages 638 – 678, 2014. [JEH15] K. Jaganathan, Y. C. Eldar, and B. Hassibi. Phase retrieval: An overview of recent developments. arXiv preprint arXiv:1510.07713, 2015. [JKN16] C. Jin, S. M. Kakade, and P. Netrapalli. Provable efficient online matrix completion via nonconvex stochastic gradient descent. In Advances in Neural Information Processing Systems, pages 4520–4528, 2016. [JM15] A. Javanmard and A. Montanari. De-biasing the lasso: Optimal sample size for Gaussian designs. arXiv preprint arXiv:1508.02757, 2015. [JN15] P. Jain and P. Netrapalli. Fast exact matrix completion with finite samples. In Conference on Learning Theory, pages 1007–1034, 2015. [JNS13] P. Jain, P. Netrapalli, and S. Sanghavi. Low-rank matrix completion using alternating minimization. In ACM symposium on Theory of computing, pages 665–674, 2013. [KD09] K. Kreutz-Delgado. The complex gradient operator and the CR-calculus. arXiv preprint arXiv:0906.4835, 2009. 41 arXiv preprint [KLT11] V. Koltchinskii, K. Lounici, and A. B. Tsybakov. Nuclear-norm penalization and optimal rates for noisy low-rank matrix completion. Ann. Statist., 39(5):2302–2329, 2011. [KMN+ 16] N. S. Keskar, D. Mudigere, J. Nocedal, M. Smelyanskiy, and P. T. P. Tang. On large-batch training for deep learning: Generalization gap and sharp minima. arXiv preprint arXiv:1609.04836, 2016. [KMO10a] R. H. Keshavan, A. Montanari, and S. Oh. Matrix completion from a few entries. IEEE Transactions on Information Theory, 56(6):2980 –2998, June 2010. [KMO10b] R. H. Keshavan, A. Montanari, and S. Oh. Matrix completion from noisy entries. J. Mach. Learn. Res., 11:2057–2078, 2010. [KÖ16] R. Kolte and A. Özgür. Phase retrieval via incremental truncated Wirtinger flow. arXiv preprint arXiv:1606.03196, 2016. [Kol11] V. Koltchinskii. Oracle inequalities in empirical risk minimization and sparse recovery problems, volume 2033 of Lecture Notes in Mathematics. Springer, Heidelberg, 2011. [Lan93] S. Lang. Real and functional analysis. Springer-Verlag, New York,, 10:11–13, 1993. [LB10] K. Lee and Y. Bresler. Admira: Atomic decomposition for minimum rank approximation. IEEE Transactions on Information Theory, 56(9):4402–4416, 2010. [LCR16] J. Lin, R. Camoriano, and L. Rosasco. Generalization properties and implicit regularization for multiple passes SGM. In International Conference on Machine Learning, pages 2340–2348, 2016. [LL17] Y. M. Lu and G. Li. Phase transitions of spectral initialization for high-dimensional nonconvex estimation. arXiv preprint arXiv:1702.06435, 2017. [LLB17] Y. Li, K. Lee, and Y. Bresler. Blind gain and phase calibration for low-dimensional or sparse signal sensing via power iteration. In Sampling Theory and Applications (SampTA), 2017 International Conference on, pages 119–123. IEEE, 2017. [LLJB17] K. Lee, Y. Li, M. Junge, and Y. Bresler. Blind recovery of sparse signals from subsampled convolution. IEEE Transactions on Information Theory, 63(2):802–821, 2017. [LLNT17] T. Liu, G. Lugosi, G. Neu, and D. Tao. Algorithmic stability and hypothesis complexity. arXiv preprint arXiv:1702.08712, 2017. [LLSW16] X. Li, S. Ling, T. Strohmer, and K. Wei. Rapid, robust, and reliable blind deconvolution via nonconvex optimization. CoRR, abs/1606.04933, 2016. [LM14] G. Lerman and T. Maunu. Fast, robust and non-convex subspace recovery. arXiv preprint arXiv:1406.6145, 2014. [LS15] S. Ling and T. Strohmer. Self-calibration and biconvex compressive sensing. Inverse Problems, 31(11):115002, 2015. [LS17] S. Ling and T. Strohmer. Regularized gradient descent: A nonconvex recipe for fast joint blind deconvolution and demixing. arXiv preprint arXiv:1703.08642, 2017. [LT16] Q. Li and G. Tang. The nonconvex geometry of low-rank matrix optimizations with general objective functions. arXiv preprint arXiv:1611.03060, 2016. [LTR16] K. Lee, N. Tian, and J. Romberg. Fast and guaranteed blind multichannel deconvolution under a bilinear system model. arXiv preprint arXiv:1610.06469, 2016. [LWL+ 16] X. Li, Z. Wang, J. Lu, R. Arora, J. Haupt, H. Liu, and T. Zhao. Symmetry, saddle points, and global geometry of nonconvex matrix factorization. arXiv preprint arXiv:1612.09296, 2016. 42 [Mat90] R. Mathias. The spectral norm of a nonnegative matrix. Linear Algebra Appl., 139:269–284, 1990. [Mat93] R. Mathias. Perturbation bounds for the polar decomposition. SIAM Journal on Matrix Analysis and Applications, 14(2):588–597, 1993. [MBM16] S. Mei, Y. Bai, and A. Montanari. The landscape of empirical risk for non-convex losses. arXiv preprint arXiv:1607.06534, 2016. [MM17] M. Mondelli and A. Montanari. Fundamental limits of weak recovery with applications to phase retrieval. arXiv preprint arXiv:1708.05932, 2017. [MZL17] T. Maunu, T. Zhang, and G. Lerman. A well-tempered landscape for non-convex robust subspace recovery. arXiv preprint arXiv:1706.03896, 2017. [NJS13] P. Netrapalli, P. Jain, and S. Sanghavi. Phase retrieval using alternating minimization. Advances in Neural Information Processing Systems (NIPS), 2013. [NNS+ 14] P. Netrapalli, U. Niranjan, S. Sanghavi, A. Anandkumar, and P. Jain. Non-convex robust PCA. In Advances in Neural Information Processing Systems, pages 1107–1115, 2014. [NTS14] B. Neyshabur, R. Tomioka, and N. Srebro. In search of the real inductive bias: On the role of implicit regularization in deep learning. arXiv preprint arXiv:1412.6614, 2014. [NTSS17] B. Neyshabur, R. Tomioka, R. Salakhutdinov, and N. Srebro. Geometry of optimization and implicit regularization in deep learning. arXiv preprint arXiv:1705.03071, 2017. [NW12] S. Negahban and M. J. Wainwright. Restricted strong convexity and weighted matrix completion: optimal bounds with noise. J. Mach. Learn. Res., 13:1665–1697, 2012. [PKCS16] D. Park, A. Kyrillidis, C. Caramanis, and S. Sanghavi. Non-square matrix sensing without spurious local minima via the Burer-Monteiro approach. arXiv preprint arXiv:1609.03240, 2016. [QZEW17] Q. Qing, Y. Zhang, Y. Eldar, and J. Wright. Convolutional phase retrieval via gradient descent. Neural Information Processing Systems, 2017. [Rec11] B. Recht. A simpler approach to matrix completion. Journal of Machine Learning Research, 12(Dec):3413–3430, 2011. [RFP10] B. Recht, M. Fazel, and P. A. Parrilo. Guaranteed minimum-rank solutions of linear matrix equations via nuclear norm minimization. SIAM Review, 52(3):471–501, 2010. [RV+ 13] M. Rudelson, R. Vershynin, et al. Hanson-Wright inequality and sub-Gaussian concentration. Electronic Communications in Probability, 18, 2013. [SBE14] Y. Shechtman, A. Beck, and Y. C. Eldar. GESPAR: Efficient phase retrieval of sparse signals. IEEE Transactions on Signal Processing, 62(4):928–938, 2014. [SCC17] P. Sur, Y. Chen, and E. J. Candès. The likelihood ratio test in high-dimensional logistic regression is asymptotically a rescaled chi-square. arXiv preprint arXiv:1706.01191, 2017. [Sch92] B. A. Schmitt. Perturbation bounds for matrix square roots and Pythagorean sums. Linear Algebra Appl., 174:215–227, 1992. [SESS11] Y. Shechtman, Y. C. Eldar, A. Szameit, and M. Segev. Sparsity based sub-wavelength imaging with partially incoherent light via quadratic compressed sensing. Optics express, 19(16), 2011. [SHS17] D. Soudry, E. Hoffer, and N. Srebro. The implicit bias of gradient descent on separable data. arXiv preprint arXiv:1710.10345, 2017. 43 [SJL17] M. Soltanolkotabi, A. Javanmard, and J. D. Lee. Theoretical insights into the optimization landscape of over-parameterized shallow neural networks. arXiv preprint arXiv:1707.04926, 2017. [SL16] R. Sun and Z.-Q. Luo. Guaranteed matrix completion via non-convex factorization. IEEE Transactions on Information Theory, 62(11):6535–6579, 2016. [Sol14] M. Soltanolkotabi. Algorithms and Theory for Clustering and Nonconvex Quadratic Programming. PhD thesis, Stanford University, 2014. [Sol17] M. Soltanolkotabi. Structured signal recovery from quadratic measurements: Breaking sample complexity barriers via nonconvex optimization. arXiv preprint arXiv:1702.06175, 2017. [SQW16] J. Sun, Q. Qu, and J. Wright. A geometric analysis of phase retrieval. In Information Theory (ISIT), 2016 IEEE International Symposium on, pages 2379–2383. IEEE, 2016. [SQW17] J. Sun, Q. Qu, and J. Wright. Complete dictionary recovery over the sphere i: Overview and the geometric picture. IEEE Transactions on Information Theory, 63(2):853–884, 2017. [SR15] P. Schniter and S. Rangan. Compressive phase retrieval via generalized approximate message passing. IEEE Transactions on Signal Processing, 63(4):1043–1055, 2015. [SS12] W. Schudy and M. Sviridenko. Concentration and moment inequalities for polynomials of independent random variables. In Proceedings of the Twenty-Third Annual ACM-SIAM Symposium on Discrete Algorithms, pages 437–446. ACM, New York, 2012. [Tao12] T. Tao. Topics in Random Matrix Theory. Graduate Studies in Mathematics. American Mathematical Society, Providence, Rhode Island, 2012. [tB77] J. M. F. ten Berge. Orthogonal Procrustes rotation for two or more matrices. Psychometrika, 42(2):267–276, 1977. [TBS+ 16] S. Tu, R. Boczar, M. Simchowitz, M. Soltanolkotabi, and B. Recht. Low-rank solutions of linear matrix equations via procrustes flow. In Proceedings of the 33rd International Conference on International Conference on Machine Learning-Volume 48, pages 964–973. JMLR. org, 2016. [Tro15a] J. A. Tropp. Convex recovery of a structured signal from independent random linear measurements. In Sampling Theory, a Renaissance, pages 67–101. Springer, 2015. [Tro15b] J. A. Tropp. An introduction to matrix concentration inequalities. Found. Trends Mach. Learn., 8(1-2):1–230, May 2015. [TV17] Y. S. Tan and R. Vershynin. Phase retrieval via randomized kaczmarz: Theoretical guarantees. arXiv preprint arXiv:1706.09993, 2017. [TW16] J. Tanner and K. Wei. Low rank matrix completion by alternating steepest descent methods. Applied and Computational Harmonic Analysis, 40(2):417–429, 2016. [Ver12] R. Vershynin. Introduction to the non-asymptotic analysis of random matrices. Compressed Sensing, Theory and Applications, pages 210 – 268, 2012. [WC16] L. Wang and Y. Chi. Blind deconvolution from multiple sparse inputs. IEEE Signal Processing Letters, 23(10):1384–1388, 2016. [WCCL16] K. Wei, J.-F. Cai, T. F. Chan, and S. Leung. Guarantees of riemannian optimization for low rank matrix recovery. SIAM Journal on Matrix Analysis and Applications, 37(3):1198–1222, 2016. [Wed72] P.-Å. Wedin. Perturbation bounds in connection with singular value decomposition. BIT Numerical Mathematics, 12(1):99–111, 1972. 44 [Wei15] K. Wei. Solving systems of phaseless equations via Kaczmarz methods: A proof of concept study. Inverse Problems, 31(12):125008, 2015. [WGE17] G. Wang, G. B. Giannakis, and Y. C. Eldar. Solving systems of random quadratic equations via truncated amplitude flow. IEEE Transactions on Information Theory, 2017. [WGSC17] G. Wang, G. B. Giannakis, Y. Saad, and J. Chen. Solving almost all systems of random quadratic equations. arXiv preprint arXiv:1705.10407, 2017. [WWS15] C. D. White, R. Ward, and S. Sanghavi. The local convexity of solving quadratic equations. arXiv preprint arXiv:1506.07868, 2015. [WZG+ 16] G. Wang, L. Zhang, G. B. Giannakis, M. Akçakaya, and J. Chen. Sparse phase retrieval via truncated amplitude flow. arXiv preprint arXiv:1611.07641, 2016. [YWS15] Y. Yu, T. Wang, and R. J. Samworth. A useful variant of the Davis-Kahan theorem for statisticians. Biometrika, 102(2):315–323, 2015. [ZB17] Y. Zhong and N. Boumal. Near-optimal bounds for phase synchronization. arXiv preprint arXiv:1703.06605, 2017. [ZBH+ 16] C. Zhang, S. Bengio, M. Hardt, B. Recht, and O. Vinyals. Understanding deep learning requires rethinking generalization. arXiv preprint arXiv:1611.03530, 2016. [ZCL16] H. Zhang, Y. Chi, and Y. Liang. Provable non-convex phase retrieval with outliers: Median truncated Wirtinger flow. In International conference on machine learning, pages 1022–1031, 2016. [ZL15] Q. Zheng and J. Lafferty. A convergent gradient descent algorithm for rank minimization and semidefinite programming from random linear measurements. In Advances in Neural Information Processing Systems, pages 109–117, 2015. [ZL16] Q. Zheng and J. Lafferty. Convergence analysis for rectangular matrix completion using BurerMonteiro factorization and gradient descent. arXiv preprint arXiv:1605.07051, 2016. [ZLK+ 17] Y. Zhang, Y. Lau, H.-w. Kuo, S. Cheung, A. Pasupathy, and J. Wright. On the global geometry of sphere-constrained sparse blind deconvolution. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4894–4902, 2017. [ZSJ+ 17] K. Zhong, Z. Song, P. Jain, P. L. Bartlett, and I. S. Dhillon. Recovery guarantees for onehidden-layer neural networks. arXiv preprint arXiv:1706.03175, 2017. [ZWL15] T. Zhao, Z. Wang, and H. Liu. A nonconvex optimization framework for low rank matrix estimation. In Advances in Neural Information Processing Systems, pages 559–567, 2015. [ZZLC17] H. Zhang, Y. Zhou, Y. Liang, and Y. Chi. A nonconvex approach for phase retrieval: Reshaped wirtinger flow and incremental algorithms. Journal of Machine Learning Research, 2017. 45 A Proofs for phase retrieval Before proceeding, we gather a few simple facts. The standard concentration inequality for χ2 random variables together with the union bound reveals that the sampling vectors {aj } obey √ (98) max kaj k2 ≤ 6n 1≤j≤m with probability at least 1 − O(me−1.5n ). In addition, standard Gaussian concentration inequalities give p \ (99) max a> j x ≤ 5 log n 1≤j≤m with probability exceeding 1 − O(mn−10 ). A.1 Proof of Lemma 1 We start with the smoothness bound, namely, ∇2 f (x)  O(log n) · In . It suffices to prove the upper bound ∇2 f (x) . log n. To this end, we first decompose the Hessian (cf. (44)) into three components as follows: ∇2 f (x) = m m
i

3 X h >
2 2 X > \
2 \ 2 > \ \> aj x − a> x a a + aj x aj a> + 2 In + 2x\ x\> , j j j j − 2 In + 2x x m j=1 m j=1 {z } | {z } | {z } | :=Λ1 :=Λ2 :=Λ3 \ 2 where we have used yj = (a> j x ) . In the sequel, we control the three terms Λ1 , Λ2 and Λ3 in reverse order. • The third term Λ3 can be easily bounded by kΛ3 k ≤ 2 kIn k + 2 x\ x\> • The second term Λ2 can be controlled by means of Lemma 32:
= 6. kΛ2 k ≤ 2δ for an arbitrarily small constant δ > 0, as long as m ≥ c0 n log n for c0 sufficiently large. • It thus remains to control Λ1 . Towards this we discover that m

3 X > \ aj a> . kΛ1 k ≤ aj x − x\ a> j j x+x m j=1 \ Under the assumption max1≤j≤m a> j x−x \ max a> j x+x 1≤j≤m

(100) √ ≤ C2 log n and the fact (99), we can also obtain \ > \ ≤ 2 max a> j x + max aj x − x 1≤j≤m 1≤j≤m Substitution into (100) leads to
≤ (10 + C2 ) p log n. m kΛ1 k ≤ 3C2 (10 + C2 ) log n · 1 X aj a> ≤ 4C2 (10 + C2 ) log n, j m j=1 where the last inequality is a direct consequence of Lemma 31. Combining the above bounds on Λ1 , Λ2 and Λ3 yields ∇2 f (x) ≤ kΛ1 k + kΛ2 k + kΛ3 k ≤ 4C2 (10 + C2 ) log n + 2δ + 6 ≤ 5C2 (10 + C2 ) log n, as long as n is sufficiently large. This establishes the claimed smoothness property. 46 Next we move on to the strong convexity lower bound. Picking a constant C > 0 and enforcing proper truncation, we get ∇2 f (x) = m m m i
2 1 X > \
2 1 Xh 3 X >
2 > > − 3 a> x − y a a  a a a x 1 a x aj a> j j j j j j j . {|a> j x|≤C } m j=1 m j=1 j m j=1 j | {z } | {z } :=Λ4 :=Λ5 We begin with the simpler term Λ5 . Lemma 32 implies that with probability at least 1 − O(n−10 ),
Λ5 − In + 2x\ x\> ≤ δ holds for any small constant δ > 0, as long as m/(n log n) is sufficiently large. This reveals that Λ5  (1 + δ) · In + 2x\ x\> . To bound Λ4 , invoke Lemma 33 to conclude that with probability at least 1 − c3 e−c2 m (for some constants c2 , c3 > 0),
Λ4 − 3 β1 xx> + β2 kxk22 In ≤ δkxk22 for any small constant δ > 0, provided that m/n is sufficiently large. Here,       and β2 := E ξ 2 1|ξ|≤C , β1 := E ξ 4 1{|ξ|≤C} − E ξ 2 1|ξ|≤C where the expectation is taken with respect to ξ ∼ N (0, 1). By the assumption x − x\ kxk2 ≤ 1 + 2C1 , 2 kxk2 − kx\ k22 ≤ 2C1 (4C1 + 1) , 2 ≤ 2C1 , one has x\ x\> − xx> ≤ 6C1 (4C1 + 1) , which leads to Λ4 − 3 β1 x\ x\> + β2 In
≤ Λ4 − 3 β1 xx> + β2 kxk22 In
+3

β1 x\ x\> + β2 In − β1 xx> + β2 kxk22 In ≤ δkxk22 + 3β1 x\ x\> − xx> + 3β2 In − kxk22 In 2 ≤ δ (1 + 2C1 ) + 18β1 C1 (4C1 + 1) + 6β2 C1 (4C1 + 1) . This further implies i
h 2 Λ4  3 β1 x\ x\> + β2 In − δ (1 + 2C1 ) + 18β1 C1 (4C1 + 1) + 6β2 C1 (4C1 + 1) In . Recognizing that β1 (resp. β2 ) approaches 2 (resp. 1) as C grows, we can thus take C1 small enough and C large enough to guarantee that Λ4  5x\ x\> + 2In . Putting the preceding two bounds on Λ4 and Λ5 together yields   ∇2 f (x)  5x\ x\> + 2In − (1 + δ) · In + 2x\ x\>  (1/2) · In as claimed. A.2 Proof of Lemma 2 Using the update rule (cf. (17)) as well as the fundamental theorem of calculus [Lan93, Chapter XIII, Theorem 4.2], we get x t+1 \ t − x = x − η∇f x t
 Z  \
 \ − x − η∇f x = In − η 0 47 1 2 ∇ f (x (τ )) dτ 
xt − x\ , where we denote x (τ ) = x\ + τ (xt − x\ ), 0 ≤ τ ≤ 1. Here, the first equality makes use of the fact that ∇f (x\ ) = 0. Under the condition (45), it is self-evident that for all 0 ≤ τ ≤ 1, x (τ ) − x\ = kτ (xt − x\ )k2 ≤ 2C1 2 \ max a> l x(τ ) − x 1≤l≤m This means that for all 0 ≤ τ ≤ 1,
and t \ ≤ max a> l τ x −x 1≤l≤m
≤ C2 p log n. (1/2) · In  ∇2 f (x(τ ))  [5C2 (10 + C2 ) log n] · In in view of Lemma 1. Picking η ≤ 1/ [5C2 (10 + C2 ) log n] (and hence kη∇2 f (x(τ ))k ≤ 1), one sees that 0  In − η Z 1 ∇2 f (x (τ )) dτ  (1 − η/2) · In , 0 which immediately yields xt+1 − x\ A.3 2 ≤ In − η Z 1 0 ∇2 f (x (τ )) dτ · xt − x\ ≤ (1 − η/2) xt − x\ 2 2 . Proof of Lemma 3 We start with proving (19a). For all 0 ≤ t ≤ T0 , invoke Lemma 2 recursively with the conditions (47) to reach xt − x\ 2 ≤ (1 − η/2)t x0 − x\ 2 ≤ C1 (1 − η/2)t x\ 2 . (101) This finishes the proof of (19a) for 0 ≤ t ≤ T0 and also reveals that xT0 − x\ ≤ C1 (1 − η/2)T0 x\ 2 2 1 x\ n  2 (102) , provided that η  1/ log n. Applying the Cauchy-Schwarz inequality and the fact (98) indicate that T0 max a> − x\ l x 1≤l≤m
≤ max kal k2 kxT0 − x\ k2 ≤ √ 1≤l≤m 6n · leading to the satisfaction of (45). Therefore, invoking Lemma 2 yields xT0 +1 − x\ 2 ≤ (1 − η/2) xT0 − x\ 2  p 1 \ kx k2  C2 log n, n 1 \ kx k2 . n One can then repeat this argument to arrive at for all t > T0 xt − x\ 2 ≤ (1 − η/2) t x0 − x\ t 2 ≤ C1 (1 − η/2) x\ 2  1 \ kx k2 . n (103) We are left with (19b). It is self-evident that the iterates from 0 ≤ t ≤ T0 satisfy (19b) by assumptions. For t > T0 , we can use the Cauchy-Schhwarz inequality to obtain t \ max a> j x −x 1≤j≤m
≤ max kaj k2 xt − x\ 1≤j≤m 2  where the penultimate relation uses the conditions (98) and (103). 48 √ n· p 1 ≤ C2 log n, n A.4 Proof of Lemma 4 First, going through the same derivation as in (54) and (55) will result in p
t,(l) max a> − x\ ≤ C4 log n l x (104) 1≤l≤m for some C4 < C2 , which will be helpful for our analysis. We use the gradient update rules once again to decompose
h
i xt+1 − xt+1,(l) = xt − η∇f xt − xt,(l) − η∇f (l) xt,(l) h

i
h
i = xt − η∇f xt − xt,(l) − η∇f xt,(l) − η ∇f xt,(l) − ∇f (l) xt,(l) h

i
i > t,(l)
1 h > t,(l)
2 \ 2 = xt − xt,(l) − η ∇f xt − ∇f xt,(l) − η − a> al x al , al x l x | {z } |m {z } (l) (l) :=ν1 :=ν2 where the last line comes from the definition of ∇f (·) and ∇f (l) (·). (l) 1. We first control the term ν2 , which is easier to deal with. Specifically, (l) kν2 k2 ≤ η kal k2 m t,(l) a> l x (i)
2 \ − a> l x
2 n log n . C4 (C4 + 5)(C4 + 10)η m r t,(l) a> l x log n (ii) ≤ cη n r log n , n for any small constant c > 0. Here (i) follows since (98) and, in view of (99) and (104), 



 > t,(l) \ t,(l) 2 \ 2 t,(l) ≤ C4 (C4 + 10) log n, al x − x\ + 2 a> a> − a> ≤ a> − x\ l x l x l x l x p
\ t,(l) t,(l) and a> ≤ a> − x\ + a> l x ≤ (C4 + 5) log n. l x l x And (ii) holds as long as m  n log n. (l) 2. For the term ν1 , the fundamental theorem of calculus [Lan93, Chapter XIII, Theorem 4.2] tells us that   Z 1
(l) ν 1 = In − η ∇2 f (x (τ )) dτ xt − xt,(l) , 0 where we abuse the notation and denote x (τ ) = xt,(l) + τ (xt − xt,(l) ). By the induction hypotheses (51) and the condition (104), one can verify that max a> l 1≤l≤m x (τ ) − x\ 2 ≤ τ xt − x\ 2 + (1 − τ ) xt,(l) − x\ 2 ≤ 2C1 and (105) p


t \ t,(l) x (τ ) − x\ ≤ τ max a> + (1 − τ ) max a> − x\ ≤ C2 log n l x −x l x 1≤l≤m 1≤l≤m for all 0 ≤ τ ≤ 1, as long as C4 ≤ C2 . The second line follows directly from (104). To see why (105) holds, we note that r log n t,(l) \ t,(l) t t \ x −x 2 ≤ x − x 2 + x − x 2 ≤ C3 + C1 , n where the second inequality follows from the induction hypotheses (51b) and (51a). This combined with (51a) gives ! r log n \ x (τ ) − x 2 ≤ τ C1 + (1 − τ ) C3 + C1 ≤ 2C1 n as long as n is large enough, thus justifying (105). Hence by Lemma 1, ∇2 f (x (τ )) is positive definite and almost well-conditioned. By choosing 0 < η ≤ 1/ [5C2 (10 + C2 ) log n], we get (l) ν1 2 ≤ (1 − η/2) xt − xt,(l) 49 2 . (l) (l) 3. Combine the preceding bounds on ν1 and ν2 as well as the induction bound (51b) to arrive at r r log n log n t+1 t+1,(l) t t,(l) x ≤ (1 − η/2) x − x + cη ≤ C3 . −x 2 2 n n (106) This establishes (53) for the (t + 1)th iteration. A.5 Proof of Lemma 5 In view of the assumption (42) that x0 − x\ 2 ≤ x0 + x\ 2 and the fact that x0 = some λ1 (Y ) > 0 (which we will verify below), it is straightforward to see that e0 − x\ x 2 e 0 + x\ ≤ x 2 . p e0 for λ1 (Y ) /3 x One can then invoke the Davis-Kahan sinΘ theorem [YWS15, Corollary 1] to obtain e 0 − x\ x 2 √ ≤2 2 kY − E [Y ]k . λ1 (E [Y ]) − λ2 (E [Y ]) Note that (56) — kY − E[Y ]k ≤ δ — is a direct consequence of Lemma 32. Additionally, the fact that E [Y ] = I + 2x\ x\> gives λ1 (E [Y ]) = 3, λ2 (E [Y ]) = 1, and λ1 (E [Y ]) − λ2 (E [Y ]) = 2. Combining this spectral gap and the inequality kY − E[Y ]k ≤ δ, we arrive at √ e0 − x\ 2 ≤ 2δ. x To connect this bound with x0 , we need to take into account the scaling factor it follows from Weyl’s inequality and (56) that p λ1 (Y ) /3. To this end, |λ1 (Y ) − 3| = |λ1 (Y ) − λ1 (E [Y ])| ≤ kY − E [Y ]k ≤ δ and, as a consequence, λ1 (Y ) ≥ 3 − δ > 0 when δ ≤ 1. This further implies that r λ1 (Y ) −1 1 λ1 (Y ) λ1 (Y ) −1 = q 3 − 1 ≤ δ, ≤ 3 3 3 λ1 (Y ) +1 3 (107) √ √ √ √ where we have used the elementary identity a − b = (a − b) /( a + b). With these bounds in place, we can use the triangle inequality to get r r λ1 (Y ) 0 λ1 (Y ) 0 \ 0 \ e −x e −x e0 + x e0 − x\ x = x x −x 2 = 3 3 2 2 r λ1 (Y ) e0 − x\ 2 ≤ −1 + x 3 √ 1 ≤ δ + 2δ ≤ 2δ. 3 A.6 Proof of Lemma 6 To begin with, repeating the same argument as in Lemma 5 (which we omit here for conciseness), we see that for any fixed constant δ > 0, h i √ e0,(l) − x\ 2 ≤ 2δ, Y (l) − E Y (l) ≤ δ, kx0,(l) − x\ k2 ≤ 2δ, x 1≤l≤m (108) holds with probability at least 1 − O(mn−10 ) as long as m  n log n. The `2 bound on kx0 − x0,(l) k2 is derived as follows. 50 e0 − x e0,(l) 1. We start by controlling x 2 e0 − x e0,(l) x . Combining (57) and (108) yields 2 e0 − x\ ≤ x 2 e0,(l) − x\ + x √ ≤ 2 2δ. 2 e0 − x e0,(l) 2 ≤ x e0 + x e0,(l) 2 , and hence the Davis-Kahan sinΘ For δ sufficiently small, this implies that x theorem [DK70] gives
0,(l)
0,(l) e Y − Y (l) x 0 0,(l)
2 ≤ Y − Y (l) x e e −x e . (109) ≤ x 2 2 (l) λ1 (Y ) − λ2 Y Here, the second inequality uses Weyl’s inequality:

λ1 Y − λ2 Y (l) ≥ λ1 (E[Y ]) − Y − E[Y ] − λ2 (E[Y (l) ]) − Y (l) − E[Y (l) ] ≥ 3 − δ − 1 − δ ≥ 1, with the proviso that δ ≤ 1/2. e0,(l) k2 . Applying the Weyl’s inequality and (56) yields 2. We now connect kx0 − x0,(l) k2 with ke x0 − x |λ1 (Y ) − 3| ≤ kY − E[Y ]k ≤ δ λ1 (Y ) ∈ [3 − δ, 3 + δ] ⊆ [2, 4] =⇒ and, similarly, λ1 (Y (l) ), kY k, kY (l) k ∈ [2, 4]. Invoke Lemma 34 to arrive at
0,(l)   e Y − Y (l) x 1 4 0 0,(l) 2 √ √ x −x √ e0 − x e0,(l) x ≤ + 2 + 2 3 2 2 2
0,(l) e , ≤ 6 Y − Y (l) x 2 (110) 2 (111) where the last inequality comes from (109). 3. Everything then boils down to controlling max 1≤l≤m
0,(l) e Y − Y (l) x
0,(l) e Y − Y (l) x 2 = max 1≤l≤m ≤ max 1≤l≤m 1 m 2 . Towards this we observe that \ a> l x
\ 2 a> l x √
2 e0,(l) al a> l x e0,(l) a> l x al 2 2 m √ log n · n . m r log n n log n  · . (112) n m √ The inequality (i) makes use of the fact maxl a> x\ ≤ 5 log n (cf. (99)), the bound maxl kal k2 ≤ l √ √ e0,(l) ≤ 5 log n (due to statistical independence and standard Gaussian 6 n (cf. (98)), and maxl a> l x concentration). As long as m/(n log n) is sufficiently large, substituting the above bound (112) into (111) leads us to conclude that r log n 0 0,(l) max x − x (113) ≤ C3 2 1≤l≤m n (i) log n · for any constant C3 > 0. B Proofs for matrix completion Before proceeding to the proofs, let us record an immediate consequence of the incoherence property (25): r r κµ κµr X \ 2,∞ ≤ X\ F ≤ X\ . (114) n n 51 where κ = σmax /σmin is the condition number of M \ . This follows since
1/2
1/2 ≤ U \ 2,∞ Σ\ X \ 2,∞ = U \ Σ\ 2,∞ r r
1/2 √ µ µ ≤ U \ F Σ\ ≤ U \ F κσmin n n r r κµ κµr \ X F≤ X\ . ≤ n n Unless otherwise specified, we use the indicator variable δj,k to denote whether the entry in the location (j, k) is included in Ω. Under our model, δj,k is a Bernoulli random variable with mean p. B.1 Proof of Lemma 7 By the expression of the Hessian in (61), one can decompose >
2
1 1 PΩ V X > + XV > F + PΩ XX > − M \ , V V > 2p p
2
2 1
1 > > \> \ > − + V X + XV PΩ V X + X V PΩ XX > − M \ , V V > F F 2p p {z } | {z } vec (V ) ∇2 fclean (X) vec (V ) = = 1 PΩ 2p | :=α1 :=α2
2 1 1 1 2 2 + PΩ V X \> + X \ V > F − V X \> + X \ V > F + V X \> + X \ V > F . 2p 2 2 {z } | {z } | :=α4 :=α3 The basic idea is to demonstrate that: (1) α4 is bounded both from above and from below, and (2) the first three terms are sufficiently small in size compared to α4 . 1. We start by controlling α4 . It is immediate to derive the following upper bound α4 ≤ V X \> 2 F + X \V > 2 F 2 2 ≤ 2kX \ k2 kV kF = 2σmax kV kF . When it comes to the lower bound, one discovers that
o 1n 2 2 α4 = V X \> F + X \ V > F + 2Tr X \> V X \> V 2 h
>
> i 2 ≥ σmin kV kF + Tr Z + X \ − Z V Z + X \ − Z V
2 2 ≥ σmin kV kF + Tr Z > V Z > V − 2 Z − X \ kZk kV kF − Z − X \
2 ≥ (σmin − 5δσmax ) kV kF + Tr Z > V Z > V , 2 2 kV kF (115) where the last line comes from the assumptions that Z − X\ ≤ δ X\ ≤ X\ and kZk ≤ Z − X \ + X \ ≤ 2 X \ . With our assumption V = Y HY − Z in mind, it comes down to controlling
  Tr Z > V Z > V = Tr Z > (Y HY − Z) Z > (Y HY − Z) . From the definition of HY , we see from Lemma 35 that Z > Y HY (and hence Z > (Y HY − Z)) is a symmetric matrix, which implies that   Tr Z > (Y HY − Z) Z > (Y HY − Z) ≥ 0. Substitution into (115) gives 2 α4 ≥ (σmin − 5δσmax ) kV kF ≥ provided that κδ ≤ 1/50. 52 9 2 σmin kV kF , 10 2. For α1 , we consider the following quantity PΩ V X > + XV >
2 F



= PΩ V X > , PΩ V X > + PΩ V X > , PΩ XV >



+ PΩ XV > , PΩ V X > + PΩ XV > , PΩ XV >



= 2 PΩ V X > , PΩ V X > + 2 PΩ V X > , PΩ XV > . Similar decomposition can be performed on PΩ V X \> + X \ V > α1 =
2 F as well. These identities yield



 1 PΩ V X > , PΩ V X > − PΩ V X \> , PΩ V X \> p {z } | :=β1



 1 + . PΩ V X > , PΩ XV > − PΩ V X \> , PΩ X \ V > p {z } | :=β2 For β2 , one has β2 =
> 

E 1D  PΩ V X − X \ , PΩ X − X \ V > p

> 
E 1 1D  PΩ V X \> , PΩ PΩ V X − X \ + , PΩ X \ V > + p p which together with the inequality |hA, Bi| ≤ kAkF kBkF gives |β2 | ≤ 
>  1 PΩ V X − X \ p 2 F + 
>  2 PΩ V X − X \ p F

X − X\ V > PΩ X \ V > This then calls for upper bounds on the following two terms 
>  1 √ PΩ V X − X \ p and F
1 √ PΩ X \ V > p F
F . (116) . The injectivity of PΩ (cf. [CR09, Section 4.2] or Lemma 38)—when restricted to the tangent space of M \ —gives: for any fixed constant γ > 0,
1 √ PΩ X \ V > F ≤ (1 + γ) X \ V > F ≤ (1 + γ) X \ kV kF p
with probability at least 1 − O n−10 , provided that n2 p/(µnr log n) is sufficiently large. In addition, 
>  1 PΩ V X − X \ p 2 F = = 1 p X 1≤j,k≤n X 1≤j≤n   > 2 \ δj,k Vj,· Xk,· − Xk,·    >   X 1 \ \ >  Vj,· Vj,·  δj,k Xk,· − Xk,· Xk,· − Xk,· p 1≤k≤n  >   1 X 2 \ \ δj,k Xk,· − Xk,· Xk,· − Xk,· kV kF 1≤j≤n p 1≤k≤n    1  X 2 2 \ ≤ max δj,k max Xk,· − Xk,· kV kF  p 1≤j≤n  1≤k≤n 2 ≤ max 1≤k≤n ≤ (1 + γ) n X − X \ 53 2 2,∞ 2 kV kF , 
with probability exceeding 1 − O n−10 , which holds as long as np/ log n is sufficiently large. Taken collectively, the above bounds yield that for any small constant γ > 0, q 2 2 2 2 2 2 2 |β2 | ≤ (1 + γ) n X − X \ 2,∞ kV kF + 2 (1 + γ) n kX − X \ k2,∞ kV kF · (1 + γ) kX \ k kV kF   √ 2 2 . 2 n X \ 2,∞ +  n X \ 2,∞ X \ kV kF , where the last inequality makes use of the assumption kX − X \ k2,∞ ≤ kX \ k2,∞ . The same analysis can be repeated to control β1 . Altogether, we obtain   √ 2 2 |α1 | ≤ |β1 | + |β2 | . n2 X \ 2,∞ + n X \ 2,∞ X \ kV kF r   (i) (ii) 1 √ κµr 2 2 2 κµr ≤ n + n σmax kV kF ≤ σmin kV kF , n n 10 p where (i) utilizes the incoherence condition (114) and (ii) holds with the proviso that  κ3 µr  1. 3. To bound α2 , apply the Cauchy-Schwarz inequality to get  

1 1 2 ≤ |α2 | = V , PΩ XX > − M \ V PΩ XX > − M \ kV kF . p p
In view of Lemma 43, with probability at least 1 − O n−10 , as soon as  p
√ 1 2 PΩ XX > − M \ ≤ 2n2 X \ 2,∞ + 4 n log n X \ 2,∞ X \ p r   √ 1 κµr 2 κµr + 4 n log n σmax ≤ σmin ≤ 2n n n 10 κ3 µr log n  1, where we utilize the incoherence condition (114). This in turn implies that 1 2 σmin kV kF . 10 Notably, this bound holds uniformly over all X satisfying the condition in Lemma 7, regardless of the statistical dependence between X and the sampling set Ω. |α2 | ≤ 4. The last term α3 can also be controlled using the injectivity of PΩ when restricted to the tangent space of M \ . Specifically, it follows from the bounds in [CR09, Section 4.2] or Lemma 38 that |α3 | ≤ γ V X \> + X \ V > 2 2 F ≤ 4γσmax kV kF ≤ 1 2 σmin kV kF 10 for any γ > 0 such that κγ is a small constant, as soon as n2 p  κ2 µrn log n. 5. Taking all the preceding bounds collectively yields > vec (V ) ∇2 fclean (X) vec (V ) ≥ α4 − |α1 | − |α2 | − |α3 |   9 3 1 2 2 ≥ − σmin kV kF ≥ σmin kV kF 10 10 2 for all V satisfying our assumptions, and > vec (V ) ∇2 fclean (X) vec (V ) ≤ α4 + |α1 | + |α2 | + |α3 |   3 5 2 2 ≤ 2σmax + σmin kV kF ≤ σmax kV kF 10 2 for all V . Since this upper bound holds uniformly over all V , we conclude that ∇2 fclean (X) ≤ as claimed. 54 5 σmax 2 B.2 Proof of Lemma 8 ct+1 is chosen to minimize the error in terms of the Frobenius norm (cf. (26)), we have Given that H ct+1 − X \ X t+1 H F ct − X \ ≤ X t+1 H = F  t
 t c − X\ X − η∇f X t H F
ct − η∇f X t H ct − X \ = X tH F  
1 (ii) t ct t ct t ct = X H − η ∇fclean X H − PΩ (E) X H − X \ p F


1 ct − η∇fclean X t H ct − X \ − η∇fclean X \ ct , +η ≤ X tH PΩ (E) X t H p F F {z } | | {z } (i) :=α1 (117) :=α2 where (i) follows from the identity ∇f (X t R) = ∇f (X t ) R for any orthonormal matrix R ∈ Or×r , (ii) arises from the definitions of ∇f (X) and ∇fclean (X) (see (59) and (60), respectively), and the last inequality (117) utilizes the triangle inequality and the fact that ∇fclean (X \ ) = 0. It thus suffices to control α1 and α2 . 1. For the second term α2 in (117), it is easy to see that α2 ≤ η 1 PΩ (E) p ct X tH F ≤ 2η 1 PΩ (E) p X\ F ≤ 2ηCσ r n kX \ kF p ct ct − for some absolute constant C > 0. Here, the second inequality holds because X t H ≤ X tH F \ \ \ X F + X F ≤ 2 X F , following the hypothesis (28a) together with our assumptions on the noise and the sample complexity. The last inequality makes use of Lemma 40. 2. For the first term α1 in (117), the fundamental theorem of calculus [Lan93, Chapter XIII, Theorem 4.2] reveals h


i ct − η∇fclean X t H ct − X \ − η∇fclean X \ vec X t H h i h

i ct − X \ − η · vec ∇fclean X t H ct − ∇fclean X \ = vec X t H ! Z 1   2 ct − X \ , = Inr − η ∇ fclean (X(τ )) dτ vec X t H (118) |0 {z } :=A ct − X \ ). Taking the squared Euclidean norm of both sides of the where we denote X(τ ) := X \ + τ (X t H equality (118) leads to 2
2 ct − X \ (Inr − ηA) vec X t H


ct − X \ > Inr − 2ηA + η 2 A2 vec X t H ct − X \ = vec X t H ct − X \ (α1 ) = vec X t H ct − X \ ≤ X tH 2 F
> 2 + η 2 kAk ct − X \ X tH 2 F ct − X \ − 2η vec X t H
>
ct − X \ , A vec X t H (119) where in (119) we have used the fact that ct − X \ vec X t H
>

ct − X \ ≤ kAk2 vec X t H ct − X \ A2 vec X t H 2 2 Based on the condition (28b), it is easily seen that ∀τ ∈ [0, 1], s s ! log n C n log n 8 + σ X\ X (τ ) − X \ 2,∞ ≤ C5 µr np σmin p 55 2 = kAk 2,∞ ct − X \ X tH . 2 F . Taking X = X (τ ) , Y = X t and Z = X \ in Lemma 7, one can easily verify the assumptions therein given our sample size condition n2 p  κ3 µ3 r3 n log3 n and the noise condition (27). As a result, ct − X \ vec X t H
>
ct − X \ ct − X \ ≥ σmin X t H A vec X t H 2 Substituting these two inequalities into (119) yields   25 2 2 ct − X \ − σmin η X t H (α1 ) ≤ 1 + η 2 σmax 4 2 F 2 F and kAk ≤ 5 σmax . 2  σmin  ct − X \ ≤ 1− η X tH 2 2 as long as 0 < η ≤ (2σmin )/(25σmax ), which further implies that  σmin  ct − X \ α1 ≤ 1 − η X tH 4 F 2 F . 3. Combining the preceding bounds on both α1 and α2 and making use of the hypothesis (28a), we have r   ct+1 − X \ ≤ 1 − σmin η X t H ct − X \ + 2ηCσ n X \ X t+1 H F 4 p F F   r r  σmin  σ n n 1 ≤ 1− η X \ F + C1 X \ F + 2ηCσ X\ F C4 ρt µr √ 4 np σmin p p   r  σmin  C1 n σmin  1 \ t X F+ 1− X\ F ≤ 1− η C4 ρ µr √ η + 2ηC σ 4 np 4 σmin p r 1 σ n t+1 \ ≤ C4 ρ µr √ X F + C1 X\ F np σmin p 2 as long as 0 < η ≤ (2σmin )/(25σmax ), 1 − (σmin /4) · η ≤ ρ < 1 and C1 is sufficiently large. This completes the proof of the contraction with respect to the Frobenius norm. B.3 Proof of Lemma 9 To facilitate analysis, we construct an auxiliary matrix defined as follows 
 ft+1 := X t H ct − η 1 PΩ X t X t> − M \ + E X \ . X p (120) With this auxiliary matrix in place, we invoke the triangle inequality to bound ft+1 − X \ . ct+1 − X \ ≤ X t+1 H ct+1 − X ft+1 + X X t+1 H | {z } | {z } :=α1 (121) :=α2 ft+1 is also not far from the truth. 1. We start with the second term α2 and show that the auxiliary matrix X t+1 f The definition of X allows one to express 
 ct − η 1 PΩ X t X t> − M \ + E X \ − X \ α2 = X t H p
1 ct − η 1 PΩ X t X t> − X \ X \> X \ − X \ ≤η PΩ (E) X \ + X t H p p
1 ct − η X t X t> − X \ X \> X \ − X \ ≤η PΩ (E) X \ + X t H p | {z } (122) :=β1 1 +η PΩ X t X t> − X \ X p |
\>
X \ − X t X t> − X \ X \> X \ , {z } :=β2 56 (123) where we have used the triangle inequality to separate the population-level component (i.e. β1 ), the perturbation (i.e. β2 ), and the noise component. In what follows, we will denote ct − X \ ∆t := X t H which, by Lemma 35, satisfies the following symmetry property ct> X t> X \ = X \> X t H ct H ∆t> X \ = X \> ∆t . =⇒ (124) (a) The population-level component β1 is easier to control. Specifically, we first simplify its expression as
β1 = ∆t − η ∆t ∆t> + ∆t X \> + X \ ∆t> X \
≤ ∆t − η ∆t X \> + X \ ∆t> X \ + η ∆t ∆t> X \ . {z } | {z } | :=γ1 :=γ2 The leading term γ1 can be upper bounded by γ1 = ∆t − η∆t Σ\ − ηX \ ∆t> X \ = ∆t − η∆t Σ\ − ηX \ X \> ∆t
1

1 t 1 = ∆ Ir − 2ηΣ\ + Ir − 2ηM \ ∆t ≤ Ir − 2ηΣ\ + Ir − 2ηM \ ∆t 2 2 2 where the second identity follows from the symmetry property (124). By choosing η ≤ 1/(2σmax ), one has 0  Ir − 2ηΣ\  (1 − 2ησmin ) Ir and 0  Ir − 2ηM \  Ir , and further one can ensure γ1 ≤ 1 [(1 − 2ησmin ) + 1] ∆t = (1 − ησmin ) ∆t . 2 (125) Next, regarding the higher order term γ2 , we can easily obtain γ2 ≤ η ∆t 2 X\ . (126) The bounds (125) and (126) taken collectively give β1 ≤ (1 − ησmin ) ∆t + η ∆t 2 X\ . (127) (b) We now turn to the perturbation part β2 by showing that
  1 1 β2 = PΩ ∆t ∆t> + ∆t X \> + X \ ∆t> X \ − ∆t ∆t> + ∆t X \> + X \ ∆t> X \ η p



1 1 ≤ PΩ ∆t X \> X \ − ∆t X \> X \ + PΩ X \ ∆t> X \ − X \ ∆t> X \ p p F F {z } | {z } | :=θ1 :=θ2

1 + PΩ ∆t ∆t> X \ − ∆t ∆t> X \ , p F | {z } (128) :=θ3 where the last inequality holds due to the triangle inequality as well as the fact that kAk ≤ kAkF . In the sequel, we shall bound the three terms separately.

• For the first term θ1 in (128), the lth row of p1 PΩ ∆t X \> X \ − ∆t X \> X \ is given by   n n X 1X 1 \ \  \> \> (δl,j − p) ∆tl,· Xj,· Xj,· = ∆tl,·  (δl,j − p) Xj,· Xj,· p j=1 p j=1 57 where, as usual, δl,j = 1{(l,j)∈Ω} . Lemma 41 together with the union bound reveals that  q n 1 1X 2 2 \ \> \ 2 \ \ (δl,j − p) Xj,· Xj,· . p kX k2,∞ kX k log n + X 2,∞ log n p j=1 p s kX \ k22,∞ σmax log n kX \ k22,∞ log n  + p p for all 1 ≤ l ≤ n with high probability. This gives   n X 1X 1 \> \  \> \ ∆tl,·  ≤ ∆tl,· 2 (δl,j − p) Xj,· Xj,· (δl,j − p) Xj,· Xj,· p j=1 p j 2 s   kX \ k2 σ  \ 2 log n kX k log n max 2,∞ 2,∞ . ∆tl,· 2 + ,   p p which further reveals that v u uX n u 1X \> \ (δl,j − p) ∆tl,· Xj,· Xj,· θ1 = t p j l=1 2 . ∆t 2 F s  kX \ k2  s  kX \ k2  \ 2 σ log n kX k2,∞ log n  2,∞ max +  p p   \ 2 rkX k log n 2,∞ + . ∆t   p p (s ) (ii) κµr2 log n κµr3/2 log n t . ∆ + σmax np np (i) 2,∞ rσmax log n √ (iii) ≤ γσmin ∆t , √ for arbitrarily small γ > 0. Here, (i) follows from k∆t kF ≤ r k∆t k, (ii) holds owing to the incoherence condition (114), and (iii) follows as long as n2 p  κ3 µr2 n log n. • For the second term θ2 in (128), denote

A = PΩ X \ ∆t> X \ − p X \ ∆t> X \ , whose lth row is given by Al,· = \ Xl,· n X j=1 \ (δl,j − p) ∆t> j,· Xj,· . Recalling the induction hypotheses (28b) and (28c), we define s s log n σ n log n X \ 2,∞ + C8 X\ ∆t 2,∞ ≤ C5 ρt µr np σmin p r n σ 1 ∆t ≤ C9 ρt µr √ X \ + C10 X \ := ψ. np σmin p (129) 2,∞ := ξ (130) (131) With these two definitions in place, we now introduce a “truncation level” ω := 2pξσmax 58 (132) that allows us to bound θ2 in terms of the following two terms v v v u n u n u n X X u uX 1u 1 1 2 2 2 θ2 = t kAl,· k2 ≤ t kAl,· k2 1{kAl,· k ≤ω} + t kAl,· k2 1{kAl,· k ≥ω} . 2 2 p p p l=1 l=1 l=1 {z } {z } | | :=φ1 :=φ2 We will apply different strategies when upper bounding the terms φ1 and φ2 , with their bounds given in the following two lemmas under the induction hypotheses (28b) and (28c). Lemma 22. Under the conditions in Lemma 9, there exist some constants c, C > 0 such that with probability exceeding 1 − c exp(−Cnr log n), q (133) φ1 . ξ pσmax kX \ k22,∞ nr log2 n holds simultaneously for all ∆t obeying (130) and (131). Here, ξ is defined in (130).
Lemma 23. Under the conditions in Lemma 9, with probability at least 1 − O n−10 , q 2 φ2 . ξ κµr2 p log2 n X \ (134) holds simultaneously for all ∆t obeying (130) and (131). Here, ξ is defined in (130). The bounds (133) and (134) together with the incoherence condition (114) yield s q q κµr2 log2 n 1 1 2 ξσmax . θ2 . ξ pσmax kX \ k22,∞ nr log2 n + ξ κµr2 p log2 n X \ . p p p • Next, we assert that the third term θ3 in (128) has the same upper bound as θ2 . The proof follows by repeating the same argument used in bounding θ2 , and is hence omitted. Take the previous three bounds on θ1 , θ2 and θ3 together to arrive at s κµr2 log2 n t e β2 ≤ η (|θ1 | + |θ2 | + |θ3 |) ≤ ηγσmin ∆ + Cη ξσmax p e > 0. for some constant C (c) Substituting the preceding bounds on β1 and β2 into (123), we reach (i)
1 α2 ≤ 1 − ησmin + ηγσmin + η ∆t X \ PΩ (E) X \ ∆t + η p s s s 2 log2 n κµr log n σ n log n e + Cη σmax C5 ρt µr X \ 2,∞ + C8 X\ p np σmin p (ii)  σmin  1 ≤ 1− η ∆t + η PΩ (E) X \ 2 p s s s 2 log2 n κµr log n σ n log n t \ e + Cη σmax C5 ρ µr X 2,∞ + C8 X\ p np σmin p r (iii)  σmin  n ≤ 1− η ∆t + Cησ X\ 2 p s r  r  κ2 µ2 r3 log3 n 1 σ n t e σmax C5 ρ µr + C8 X\ + Cη np np σmin p 2,∞ 2,∞ ! ! (135) for some constant C > 0. Here, (i) uses the definition of ξ (cf. (130)), (ii) holds if γ is small enough and k∆t k X \  σmin , and (iii) follows from Lemma 40 as well as the incoherence condition (114). An 59 immediate consequence of (135) is that under the sample size condition and the noise condition of this lemma, one has ft+1 − X \ X \ ≤ σmin /2 X (136) if 0 < η ≤ 1/σmax . 2. We then move on to the first term α1 in (121), which can be rewritten as ct R1 − X ft+1 , α1 = X t+1 H with ct R1 = H
−1 ft+1 satisfies (a) First, we claim that X ct+1 := arg min H R∈O r×r ct R − X \ X t+1 H ft+1 R − X \ Ir = arg min X r×r R∈O F F . (137) (138) , ft+1 is already rotated to the direction that is most “aligned” with X \ . This important meaning that X ft+1 is property eases the analysis. In fact, in view of Lemma 35, (138) follows if one can show that X \> X ct is symmetric symmetric and positive semidefinite. First of all, it follows from Lemma 35 that X \> X t H and, hence, by definition, 
 ft+1 = X \> X t H ct − η X \> PΩ X t X t> − M \ + E X \ X \> X p is also symmetric. Additionally, ft+1 − M \ ≤ X ft+1 − X \ X \> X X \ ≤ σmin /2, where the second inequality holds according to (136). Weyl’s inequality guarantees that ft+1  X \> X thus justifying (138) via Lemma 35. 1 σmin Ir , 2 (b) With (137) and (138) in place, we resort to Lemma 37 to establish the bound. Specifically, take ft+1 and X2 = X t+1 H ct , and it comes from (136) that X1 = X X1 − X \ Moreover, we have in which ct − X ft+1 kX1 − X2 k X \ = X t+1 H X\ ,  
 1 ct X t − η PΩ X t X t> − M \ + E X t H p   
 ct − η 1 PΩ X t X t> − M \ + E X \ − X tH p   t t>
  t t 1 c − X\ . X H = −η PΩ X X − M \ + E p ct − X ft+1 = X t+1 H This allows one to derive X \ ≤ σmin /2.       ct − X \ + η 1 PΩ (E) X t H ct − X ft+1 ≤ η 1 PΩ X t X t> − M \ X t H ct − X \ X t+1 H p p 60  ≤ η 2n ∆t 2 2,∞ √ + 4 n log n ∆t 2,∞ X \ + Cσ r  n ∆t p (139) for some absolute constant C > 0. Here the last inequality follows from Lemma 40 and Lemma 43. As a consequence,  r  √ n \ t 2 t \ kX1 − X2 k X ≤ η 2n ∆ 2,∞ + 4 n log n ∆ 2,∞ X + Cσ ∆t X \ . p Under our sample size condition and the noise condition (27) and the induction hypotheses (28), one can show kX1 − X2 k X \ ≤ σmin /4. Apply Lemma 37 and (139) to reach ct − X ft+1 α1 ≤ 5κ X t+1 H  √ 2 ≤ 5κη 2n ∆t 2,∞ + 2 n log n ∆t 2,∞ X \ + Cσ r  n ∆t . p 3. Combining the above bounds on α1 and α2 , we arrive at r   ct+1 − X \ ≤ 1 − σmin η ∆t + ηCσ n X \ X t+1 H 2 p s r  r  C8 κ2 µ2 r3 log3 n 1 n t e σmax C5 ρ µr + X\ + Cη σ np np σmin p  r  √ n t 2 t \ + 5ηκ 2n ∆ 2,∞ + 2 n log n ∆ 2,∞ X + Cσ ∆t p r σ 1 n \ t+1 X + C10 X\ , ≤ C9 ρ µr √ np σmin p with the proviso that ρ ≥ 1 − (σmin /3) · η, κ is a constant, and n2 p  µ3 r3 n log3 n. B.3.1 Proof of Lemma 22 In what follows, we first assume that the δj,k ’s are independent, and then use the standard decoupling trick to extend the result to symmetric sampling case (i.e. δj,k = δk,j ). To begin with, we justify the concentration bound for any ∆t independent of Ω, followed by the standard covering argument that extends the bound to all ∆t . For any ∆t independent of Ω, one has 2 and \ \ B := max Xl,· (δl,j − p) ∆t> ≤ X \ 2,∞ ξ j,· Xj,· 1≤j≤n 2   n >  X 2 \ \ \ \ t>  V := E  (δl,j − p) Xl,· ∆t> j,· Xj,· Xl,· ∆j,· Xj,· j=1 ≤p \ Xl,· \ ≤ p Xl,· ≤ 2p 2 2 2 n X 2 X \ 2,∞ t ∆t> j,· ∆j,· j=1 X\ 2 \ 2 X 2,∞ 2 2,∞ ψ2 ξ 2 σmax , where ξ and ψ are defined respectively in (130) and (131). Here, the last line makes use of the fact that √ X \ 2,∞ ψ  ξ X \ = ξ σmax , (140) 61 as long as n is sufficiently large. Apply the matrix Bernstein inequality [Tro15b, Theorem 6.1.1] to get !  ct2 P kAl,· k2 ≥ t ≤ 2r exp − 2 2 2pξ 2 σmax kX \ k2,∞ + t · kX \ k2,∞ ξ ! ct2 ≤ 2r exp − 2 4pξ 2 σmax kX \ k2,∞ for some constant c > 0, provided that t ≤ 2pσmax ξ. This upper bound on t is exactly the truncation level ω we introduce in (132). With this in mind, we can easily verify that kAl,· k2 1{kAl,· k ≤ω} 2   2 is a sub-Gaussian random variable with variance proxy not exceeding O pξ 2 σmax X \ 2,∞ log r . Therefore, invoking the concentration bounds for quadratic functions [HKZ12, Theorem 2.1] yields that for some constants C0 , C > 0, with probability at least 1 − C0 e−Cnr log n , φ21 = n X l=1 2 kAl,· k2 1{kAl,· k 2 ≤ω } . pξ 2 σmax kX \ k22,∞ nr log2 n. Now that we have established an upper bound on any fixed matrix ∆t (which holds with exponentially high probability), we can proceed to invoke the standard epsilon-net argument to establish a uniform bound over all feasible ∆t . This argument is fairly standard, and is thus omitted; see [Tao12, Section 2.3.1] or the 1 proof of Lemma 42. In conclusion, we have that with probability exceeding 1 − C0 e− 2 Cnr log n , v u n q uX 2 kAl,· k2 1{kAl,· k ≤ω} . pξ 2 σmax kX \ k22,∞ nr log2 n φ1 = t (141) 2 l=1 holds simultaneously for all ∆t ∈ Rn×r obeying the conditions of the lemma. In the end, we comment on how to extend the bound to the symmetric sampling pattern where δj,k = δk,j . 2 Recall from (129) that the diagonal element δl,l cannot change the `2 norm of Al,· by more than X \ 2,∞ ξ. As a result, changing all the diagonals {δl,l } cannot change the quantity of interest (i.e. φ1 ) by more than √ 2 n X \ 2,∞ ξ. This is smaller than the right hand side of (141) under our incoherence and sample size conditions. Hence from now on we ignore the effect of {δl,l } and focus on off-diagonal terms. The proof then follows from the same argument as in [GLM16, Theorem D.2]. More specifically, we can employ the construction of Bernoulli random variables introduced therein to demonstrate that the upper bound in (141) 0 0 still holds if the indicator δi,j is replaced by (τi,j + τi,j )/2, where τi,j and τi,j are independent copies of the symmetric Bernoulli random variables. Recognizing that sup∆t φ1 is a norm of the Bernoulli random variables τi,j , one can repeat the decoupling argument in [GLM16, Claim D.3] to finish the proof. We omit the details here for brevity. B.3.2 Proof of Lemma 23 Observe from (129) that kAl,· k2 ≤ X \ ≤ X\ 2,∞ n X j=1 2,∞   \ (δl,j − p) ∆t> j,· Xj,· n X \ t δl,j ∆t> j,· Xj,· + p ∆ j=1 62 (142)  X\    \ δl,1 X1,·      .. \  t> ≤ X \ 2,∞  δl,1 ∆t>  + pψ X   1,· , · · · , δl,n ∆n,· . \ δl,n Xn,·

√ ≤ X \ 2,∞ Gl ∆t · 1.2 p X \ + pψ X \ ,   (143) where ψ is as defined in (131) and Gl (·) is as defined in Lemma 41. Here, the last inequality follows from Lemma 41, namely, for some constant C > 0, the following holds with probability at least 1 − O(n−10 )   \ δl,1 X1,·   12 q   2 .. \ 2 \ 2 2 \ \ + C pkX k2,∞ kX k log n + CkX k2,∞ log n  ≤ p X  . \ δl,n Xn,· r  1 κµr κµr log n 2 √ (144) ≤ p+C p log n + C X \ ≤ 1.2 p X \ , n n where we also use the incoherence condition (114) and the sample complexity condition n2 p  κµrn log n. Hence, the event kAl,· k2 ≥ ω = 2pσmax ξ together with (142) and (143) necessarily implies that n X j=1 t Gl ∆
≥ \ (δl,j − p) ∆t> j,· Xj,· ≥ 2pσmax 2pσmax ξ kX \ kkX \ k2,∞ √ 1.2 p − pψ ≥ √ 2 pkX \ kξ kX \ k2,∞ − ξ kX \ k2,∞ √ pψ 1.2 and √ ≥ 1.5 p ξ X\ , kX \ k2,∞ where the last inequality follows from the bound (140). As a result, with probability at least 1 − O(n−10 ) (i.e. when (144) holds for all l’s) we can upper bound φ2 by v v u n uX uX u n 2 2 , √ √ kAl,· k2 1{kAl,· k ≥ω} ≤ t kAl,· k2 1 φ2 = t 1.5 pξ σmax t 2 l=1 kGl (∆ )k≥ l=1 kX \ k2,∞ where the indicator functions are now specified with respect to kGl (∆t )k. Next, we divide into multiple cases based on the size of kGl (∆t )k. By Lemma 42, for some constants c1 , c2 > 0, with probability at least 1 − c1 exp (−c2 nr log n), n X l=1 αn 1{kGl (∆t )k≥4√pψ+√2k rξ} ≤ k−3 2 (145) for any k ≥ 0 and any α & log n. We claim that it suffices to consider the set of sufficiently large k obeying √ √ 2k rξ ≥ 4 pψ or equivalently k ≥ log 16pψ 2 ; rξ 2 (146) otherwise we can use (140) to obtain √ √ √ √ 4 pψ + 2k rξ ≤ 8 pψ  1.5 p ξ kX \ k2,∞ X\ , which contradicts the event kAl,· k2 ≥ ω. Consequently, we divide all indices into the following sets n √ √
o Sk = 1 ≤ l ≤ n : Gl ∆t ∈ 2k rξ, 2k+1 rξ 63 (147) defined for each integer k obeying (146). Under the condition (146), it follows from (145) that n X l=1 1{kGl (∆t )k≥√2k+2 rξ} ≤ n X l=1 αn 1{kGl (∆t )k≥4√pψ+√2k rξ} ≤ k−3 , 2 meaning that the cardinality of Sk satisfies |Sk+2 | ≤ αn 2k−3 or |Sk | ≤ αn 2k−5 which decays exponentially fast as k increases. Therefore, when restricting attention to the set of indices within Sk , we can obtain r sX  √ 2 (i) √ 2 2 kAl,· k2 ≤ |Sk | · kX \ k2,∞ 1.2 2k+1 rξ p kX \ k + pψ kX \ k l∈Sk r   √ αn √ \ k+1 rξ p X \ + pψ X \ 2 X 2 2,∞ 2k−5 r √ (ii) αn √ ≤ 4 X \ 2,∞ 2k+1 rξ p X \ k−5 2 p (iii) 2 ≤ 32 ακµr2 pξ X \ , ≤ where (i) follows from the bound (143) and the constraint (147) in Sk , (ii) is a consequence of (146) and (iii) uses the incoherence condition (114). Now that we have developed an upper bound with respect to each Sk , we can add them up to yield the final upper bound. Note that there are in total no more than O (log n) different sets, i.e. Sk = ∅ if k ≥ c1 log n for c1 sufficiently large. This arises since √ √ √ √ kGl (∆t )k ≤ k∆t kF ≤ nk∆t k2,∞ ≤ nξ ≤ n rξ and hence 1{kGl (∆t )k≥4√pψ+√2k rξ} = 0 and Sk = ∅ if k/ log n is sufficiently large. One can thus conclude that φ22 ≤ c1X log n k=log 16pψ 2 rξ2 p leading to φ2 . ξ ακµr2 p log n X \ large constant c > 0. B.4 2 X l∈Sk 2 kAl,· k2 . p ακµr2 pξ X \  2 2 · log n, . The proof is finished by taking α = c log n for some sufficiently Proof of Lemma 10 ct and X2 = X t,(l) Rt,(l) , we get 1. To obtain (73a), we invoke Lemma 37. Setting X1 = X t H s (i) (ii) 1 1 C10 n log n \ \ t X1 − X X ≤ C9 ρ µr √ σmax + σ σmax ≤ σmin , np σmin p 2 where (i) follows from (70c) and (ii) holds as long as n2 p  κ2 µ2 r2 n and the noise satisfies (27). In addition, kX1 − X2 k X \ ≤ kX1 − X2 kF X \ s (i) log n ≤ C3 ρt µr X\ np 64 C7 + σ 2,∞ σmin s n log n X\ p 2,∞ ! X\ (ii) t ≤ C3 ρ µr (iii) ≤ s log n C7 σmax + σ np σmin s n log n σmax p 1 σmin , 2 where (i) utilizes (70d), (ii) follows since X \ 2,∞ ≤ X \ , and (iii) holds if n2 p  κ2 µ2 r2 n log n and the noise satisfies (27). With these in place, Lemma 37 immediately yields (73a). ct,(l) . The second inequality is con2. The first inequality in (73b) follows directly from the definition of H t,(l) t,(l) cerned with the estimation error of X R with respect to the Frobenius norm. Combining (70a), (70d) and the triangle inequality yields X t,(l) Rt,(l) − X \ F ct − X \ ≤ X tH 1 ≤ C4 ρt µr √ X\ np + F 1 ≤ C4 ρ µr √ X\ np C1 σ + F σmin t 1 X\ ≤ 2C4 ρt µr √ np F C1 σ σmin + r r 2C1 σ σmin n p n p r ct − X t,(l) Rt,(l) + X tH F s s C7 σ n log n log n t \ \ X F + C3 ρ µr X 2,∞ + X \ 2,∞ np σmin p s s r r log n κµ C σ n log n κµ 7 \ t \ X F + C3 ρ µr X F+ X\ np n σmin p n F n X\ p F F (148) , where the last step holds true as long as n  κµ log n. 3. To obtain (73c), we use (70d) and (70b) to get ct − X \ ct − X t,(l) Rt,(l) X t,(l) Rt,(l) − X \ ≤ X tH + X tH 2,∞ 2,∞ F s s s C8 σ n log n log n log n ≤ C5 ρt µr X \ 2,∞ + X \ 2,∞ + C3 ρt µr X\ np σmin p np s s C8 + C7 log n n log n t \ ≤ (C3 + C5 ) ρ µr X 2,∞ + X \ 2,∞ . σ np σmin p C7 σ + 2,∞ σmin s n log n X\ p 4. Finally, to obtain (73d), one can take the triangle inequality ct,(l) − X \ ≤ X t,(l) H ct,(l) − X t H ct X t,(l) H F ct − X t,(l) Rt,(l) ≤ 5κ X t H ct − X \ + X tH F ct − X \ , + X tH where the second line follows from (73a). Combine (70d) and (70c) to yield ct,(l) − X \ X t,(l) H s s ! r C C10 n log n 1 n 7 \ t \ t + σ X X + σ X\ + C ρ µr ≤ 5κ C3 ρ µr √ 9 2,∞ 2,∞ σmin p np σmin p s s ! r r κµr log n C n log n 1 C10 σ n 7 \ t t \ ≤ 5κ X C3 ρ µr + σ + C9 ρ µr √ X + X\ n np σmin p np σmin p r 1 2C10 σ n ≤ 2C9 ρt µr √ X\ + X\ , np σmin p log n X\ np where the second inequality uses the incoherence of X \ (cf. (114)) and the last inequality holds as long as n  κ3 µr log n. 65 2,∞ B.5 Proof of Lemma 11 From the definition of Rt+1,(l) (see (72)), we must have ct+1 − X t+1,(l) Rt+1,(l) X t+1 H F ct − X t+1,(l) Rt,(l) ≤ X t+1 H F . The gradient update rules in (24) and (69) allow one to express
 t h t,(l)
i  ct − X t+1,(l) Rt,(l) = X t − η∇f X t H c − X X t+1 H − η∇f (l) X t,(l) Rt,(l)
h
i ct − η∇f X t H ct − X t,(l) Rt,(l) − η∇f (l) X t,(l) Rt,(l) = X tH h i
ct − X t,(l) Rt,(l) − η ∇f (X t H ct ) − ∇f X t,(l) Rt,(l) ) = X tH h

i − η ∇f X t,(l) Rt,(l) − ∇f (l) X t,(l) Rt,(l) , where we have again used
the fact that ∇f (X t ) R = ∇f (X t R) for any orthonormal matrix R ∈ Or×r (similarly for ∇f (l) X t,(l) ). Relate the right-hand side of the above equation with ∇fclean (X) to reach h


i ct − X t+1,(l) Rt,(l) = X t H ct − X t,(l) Rt,(l) − η ∇fclean X t H ct − ∇fclean X t,(l) Rt,(l) X t+1 H | {z } (l) :=B1      1 t,(l) t,(l)> \ t,(l) t,(l)> \ − η PΩl X X − M − Pl X X −M X t,(l) Rt,(l) p {z } | (l) :=B2   1 ct − X t,(l) Rt,(l) + η 1 PΩ (E) X t,(l) Rt,(l) , + η PΩ (E) X t H p p l | {z } | {z } (l) (149) (l) :=B3 :=B4 where we have used the following relationship between ∇f (l) (X) and ∇f (X): 

1 ∇f (l) (X) = ∇f (X) − PΩl XX > − M \ + E X + Pl XX > − M \ X p (150) for all X ∈ Rn×r with PΩl and Pl defined respectively in (66) and (67). In the sequel, we control the four terms in reverse order. (l) 1. The last term B4 is controlled via the following lemma. Lemma 24. Suppose that the sample size obeys n2 p > Cµ2 r2 n log2 n for some sufficiently large constant
(l) C > 0. Then with probability at least 1 − O n−10 , the matrix B4 as defined in (149) satisfies (l) B4 F . ησ s n log n X\ p 2,∞ . (l) 2. The third term B3 can be bounded as follows (l) B3 F 1 ≤η PΩ (E) p t ct X H −X t,(l) R t,(l) where the second inequality comes from Lemma 40. (l) 3. For the second term B2 , we have the following lemma. 66 F . ησ r n ct − X t,(l) Rt,(l) X tH p F , Lemma 25. Suppose that the sample size obeys n2 p  µ2 r2 n log n. Then with probability exceeding
(l) 1 − O n−10 , the matrix B2 as defined in (149) satisfies (l) B2 F .η s κ2 µ2 r2 log n X t,(l) Rt,(l) − X \ np 2,∞ (151) σmax . (l) 4. Regarding the first term B1 , apply the fundamental theorem of calculus [Lan93, Chapter XIII, Theorem 4.2] to get   Z 1   (l)
2 ct − X t,(l) Rt,(l) , (152) vec B1 = Inr − η ∇ fclean (X(τ )) dτ vec X t H 0   ct − X t,(l) Rt,(l) . Going through where we abuse the notation and denote X(τ ) := X t,(l) Rt,(l) + τ X t H the same derivations as in the proof of Lemma 8 (see Appendix B.2), we get  σmin  (l) ct − X t,(l) Rt,(l) (153) η X tH B1 F ≤ 1 − 4 F 2 ). with the proviso that 0 < η ≤ (2σmin )/(25σmax Applying the triangle inequality to (149) and invoking the preceding four bounds, we arrive at ct+1 − X t+1,(l) Rt+1,(l) X t+1 H F s   σmin κ2 µ2 r2 log n ct − X t,(l) Rt,(l) + Cη e ≤ 1− η X tH X t,(l) Rt,(l) − X \ σmax 4 np F 2,∞ s r n n log n t t t,(l) t,(l) c −X e e X \ 2,∞ X H R + Cησ + Cησ p p F s  r  σmin n κ2 µ2 r2 log n ct − X t,(l) Rt,(l) + Cη e e = 1− η + Cησ X tH X t,(l) Rt,(l) − X \ 4 p np F s n log n e X \ 2,∞ + Cησ p s   2σmin κ2 µ2 r2 log n ct − X t,(l) Rt,(l) + Cη e ≤ 1− η X tH X t,(l) Rt,(l) − X \ σmax 9 np F 2,∞ s n log n e X \ 2,∞ + Cησ p 2,∞ σmax p e > 0. Here the last inequality holds as long as σ n/p  σmin , which is satisfied for some absolute constant C under our noise condition (27). This taken collectively with the hypotheses (70d) and (73c) leads to ct+1 − X t+1,(l) Rt+1,(l) X t+1 H sF s !   log n n log n 2σmin σ t \ \ η C3 ρ µr ≤ 1− X 2,∞ + C7 X 2,∞ 9 np σmin p s s s " # κ2 µ2 r2 log n log n σ n log n t e + Cη (C3 + C5 ) ρ µr + (C8 + C7 ) X\ np np σmin p s n log n e + Cησ X \ 2,∞ p 67 2,∞ σmax  σmin  ≤ 1− η C3 ρt µr 5 s log n X\ np 2,∞ + C7 σ σmin s n log n X\ p 2,∞ as long as C7 > 0 is sufficiently large, where we have used the sample complexity assumption n2 p  κ4 µ2 r2 n log n and the step size 0 < η ≤ 1/(2σmax ) ≤ 1/(2σmin ). This finishes the proof. B.5.1 Proof of Lemma 24 By the unitary invariance of the Frobenius norm, one has (l) B4 F = η PΩl (E) X t,(l) p F , where all nonzero entries of the matrix PΩl (E) reside in the lth row/column. Decouple the effects of the lth row and the lth column of PΩl (E) to reach p (l) B4 η F ≤ n X t,(l) δl,j El,j Xj,· {z } j=1 | :=uj + | 2 X t,(l) δl,j El,j Xl,· j:j6=l (154) , }2 {z :=α where δl,j := 1{(l,j)∈Ω} indicates whether the (l, j)-th entry is observed. Since X t,(l) is independent of {δl,j }1≤j≤n and {El,j }1≤j≤n , we can treat the first term as a sum of independent vectors {uj }. It is easy to verify that , ≤ X t,(l) kδl,j El,j kψ1 . σ X t,(l) kuj k2 2,∞ 2,∞ ψ1 where k · kψ1 denotes the sub-exponential norm [KLT11, Section 6]. Further, one can calculate     n n X X t,(l) t,(l)> t,(l) t,(l)> 2 Xj,· Xj,·  = pσ 2 X t,(l) V := E  (δl,j El,j ) Xj,· Xj,·  . pσ 2 E  j=1 j=1 2 F . Invoke the matrix Bernstein inequality [KLT11, Proposition 2] to discover that with probability at least
1 − O n−10 , n X j=1 . uj 2 . p V log n + kuj k q ψ1 log2 n 2 pσ 2 X t,(l) F log n + σ X t,(l) log2 n 2,∞ p + σ X t,(l) log2 n . σ np log n X t,(l) 2,∞ 2,∞ p . σ np log n X t,(l) , 2,∞ where the third inequality follows from X t,(l) 2 2 F ≤ n X t,(l) 2 , 2,∞ and the last inequality holds as long as np  log n. Additionally, the remaining term α in (154) can be controlled using the same argument, giving rise to p α . σ np log n X t,(l) 2,∞ . We then complete the proof by observing that X t,(l) 2,∞ = X t,(l) Rt,(l) 2,∞ ≤ X t,(l) Rt,(l) − X \ 2,∞ + X\ 2,∞ ≤ 2 X\ 2,∞ , (155) where the last inequality follows by combining (73c), the sample complexity condition n2 p  µ2 r2 n log n, and the noise condition (27). 68 B.5.2 Proof of Lemma 25 For notational simplicity, we denote C := X t,(l) X t,(l)> − M \ = X t,(l) X t,(l)> − X \ X \> . (156) Since the Frobenius norm is unitarily invariant, we have (l) B2 F   1 PΩl (C) − Pl (C) X t,(l) p | {z } =η . F :=W Again, all nonzero entries of the matrix W reside in its lth row/column. We can deal with the lth row and the lth column of W separately as follows p (l) B2 η ≤ F n X (δl,j − p) Cl,j Xj,· n X (δl,j − p) Cl,j Xj,· j=1 . j=1 t,(l) + + √ (δl,j − p) kCk∞ Xl,· t,(l) np kCk∞ Xl,· t,(l) (δl,j − p) Cl,j Xj,· 1≤j≤n V := 2 2 , 2 where δl,j := 1{(l,j)∈Ω} and the second line relies on the fact that L := max t,(l) 2 j:j6=l 2 t,(l) sX 2 ≤ kCk∞ X t,(l) P 2 j:j6=l (δl,j − p)  np. It follows that (i) 2,∞ ≤ 2 kCk∞ X \ 2,∞ , n n X X  t,(l) t,(l)> t,(l) t,(l)> 2 2 E (δl,j − p) Cl,j Xj,· Xj,· ≤ pkCk2∞ Xj,· Xj,· j=1 j=1 2 (ii) 2 = p kCk∞ X t,(l) F 2 ≤ 4p kCk∞ X \ 2 F . Here, (i) is a consequence of (155). In addition, (ii) follows from X t,(l) F = X t,(l) Rt,(l) F ≤ X t,(l) Rt,(l) − X \ F + X\ F ≤ 2 X\ F , where the last inequality comes from (73b), the sample complexity condition n2 p  µ2 r2 n log n, and the noise condition (27). The matrix Bernstein inequality [Tro15b, Theorem 6.1.1] reveals that n X j=1 t,(l) . (δl,j − p) Cl,j Xj,· 2 p V log n + L log n . q 2 2 p kCk∞ kX \ kF log n + kCk∞ X \
with probability exceeding 1 − O n−10 , and as a result, p (l) B2 η F . p p log n kCk∞ X \ F + √ np kCk∞ X \ 2,∞ log n (157) 2,∞ as soon as np  log n. To finish up, we make the observation that  > kCk∞ = X t,(l) Rt,(l) X t,(l) Rt,(l) − X \ X \> ≤  X t,(l) R t,(l)  > −X X t,(l) Rt,(l) \ 69 ∞ ∞  > + X \ X t,(l) Rt,(l) − X \ − X \ X \> ∞ ≤ X t,(l) Rt,(l) − X \ ≤3 X t,(l) R t,(l) −X X t,(l) Rt,(l) 2,∞ \ 2,∞ X \ 2,∞ 2,∞ + X\ 2,∞ X t,(l) Rt,(l) − X \ 2,∞ (158) , where the last line arises from (155). This combined with (157) gives s r log n n (l) \ B2 .η kCk∞ X F + η kCk∞ X \ 2,∞ p p F s r (i) log n n 2 X \ 2,∞ X \ F + η X \ 2,∞ X t,(l) Rt,(l) − X \ X t,(l) Rt,(l) − X \ .η p p 2,∞ 2,∞ s r r (ii) κµr log n κµr2 n t,(l) t,(l) \ . η X X t,(l) Rt,(l) − X \ R −X σmax + η σmax p n p 2,∞ 2,∞ n s κ2 µ2 r2 log n σmax , X t,(l) Rt,(l) − X \ .η np 2,∞ where (i) comes from (158), and (ii) makes use of the incoherence condition (114). B.6 Proof of Lemma 12 We first introduce an auxiliary matrix  h  
i ft+1,(l) := X t,(l) H ct,(l) − η 1 PΩ−l X t,(l) X t,(l)> − M \ + E + Pl X t,(l) X t,(l)> − M \ X \ . X p (159) With this in place, we can use the triangle inequality to obtain  ct+1,(l) − X \ X t+1,(l) H  l,· 2 ≤     ft+1,(l) − X \ ct+1,(l) − X ft+1,(l) + X . X t+1,(l) H l,· 2 l,· 2 {z } | {z } | :=α1 (160) :=α2 In what follows, we bound the two terms α1 and α2 separately. ft+1,(l) (see (159)) that 1. Regarding the second term α2 of (160), we see from the definition of X   h   i ft+1,(l) − X \ ct,(l) − η X t,(l) X t,(l)> − X \ X \> X \ − X \ , X = X t,(l) H l,· (161) l,· where we also utilize the definitions of PΩ−l and Pl in (67). For notational convenience, we denote ct,(l) − X \ . ∆t,(l) := X t,(l) H (162) This allows us to rewrite (161) as   h  i t,(l) ft+1,(l) − X \ X = ∆l,· − η ∆t,(l) X \> + X \ ∆t,(l)> X \ l,· t,(l) l,· t,(l) h i − η ∆t,(l) ∆t,(l)> X \ l,· t,(l) \ = ∆l,· − η∆l,· Σ\ − ηXl,· ∆t,(l)> X \ − η∆l,· ∆t,(l)> X \ , which further implies that t,(l) t,(l) α2 ≤ ∆l,· − η∆l,· Σ\ t,(l) ≤ ∆l,· \ + η Xl,· ∆t,(l)> X \ 2 Ir − ηΣ\ + η X \ 2 Ir − ηΣ\ + 2η X \ t,(l) ≤ ∆l,· 2 2,∞ 2,∞ 70 t,(l) 2 ∆t,(l) ∆t,(l) + η ∆l,· ∆t,(l)> X \ t,(l) X \ + η ∆l,· X\ . 2 2 ∆t,(l) X\ t,(l) Here, the last line follows from the fact that ∆l,· 2 hypothesis (70e) to get t,(l) ∆l,· 1 ≤ C2 ρ µr √ X\ np t 2 2,∞ + C6 ≤ X\ σ σmin 2,∞ s . To see this, one can use the induction n log n X\ p 2,∞  X\ 2,∞ (163) p as long as np  µ2 r2 and σ (n log n) /p  σmin . By taking 0 < η ≤ 1/σmax , we have 0  Ir − ηΣ\  (1 − ησmin ) Ir , and hence can obtain t,(l) α2 ≤ (1 − ησmin ) ∆l,· 2 + 2η X \ 2,∞ X\ . ∆t,(l) (164) An immediate consequence of the above two inequalities and (73d) is α2 ≤ kX \ k2,∞ . (165) 2. The first term α1 of (160) can be equivalently written as α1 = where ct,(l) R1 = H Simple algebra yields α1 ≤ ≤  
−1  ct,(l) R1 − X ft+1,(l) X t+1,(l) H ct+1,(l) := arg min H R∈O r×r ct,(l) − X ft+1,(l) X t+1,(l) H ct,(l) − X ft+1,(l) X t+1,(l) H | {z :=β1   Here, to bound the the second term we have used ft+1,(l) X l,· 2 ft+1,(l) − X \ ≤ X l,· l,· 2 l,· 2 ft+1,(l) + X l,· +2 X \ } \ + Xl,· l,· 2 , ct,(l) R − X \ X t+1,(l) H R1 l,· 2  2,∞ 2 F , kR1 − Ir k kR1 − Ir k . | {z } :=β2 2 \ = α2 + Xl,· 2 ≤ 2 X\ 2,∞ , where the last inequality follows from (165). It remains to upper bound β1 and β2 . For both β1 and β2 , ct,(l) − X ft+1,(l) . By the definition of X ft+1,(l) in (159) and the a central quantity to control is X t+1,(l) H gradient update rule for X t+1,(l) (see (69)), one has ct,(l) − X ft+1,(l) X t+1,(l) H    h  
i 1 t,(l) ct,(l) t,(l) t,(l)> \ t,(l) t,(l)> \ t,(l) ct,(l) = X H − η PΩ−l X X − M + E + Pl X X −M X H p     h  
i ct,(l) − η 1 PΩ−l X t,(l) X t,(l)> − M \ + E + Pl X t,(l) X t,(l)> − M \ X \ − X t,(l) H p      1 η t,(l) t,(l)> \ \> t,(l) t,(l)> \ \> = −η PΩ−l X X −X X + Pl X X −X X ∆t,(l) + PΩ−l (E) ∆t,(l) . p p (166) It is easy to verify that (ii) (i) 1 1 PΩ−l (E) ≤ PΩ (E) . σ p p 71 r n (iii) δ ≤ σmin p 2 for δ > 0 sufficiently small. Here, (i) uses the elementary fact that the spectral norm of a submatrix is no more than that of the matrix itself, (ii) arises from Lemma 40 and (iii) is a consequence of the noise condition (27). Therefore, in order to control (166), we need to upper bound the following quantity γ :=     1 PΩ−l X t,(l) X t,(l)> − X \ X \> + Pl X t,(l) X t,(l)> − X \ X \> . p (167) To this end, we make the observation that γ≤   1 PΩ X t,(l) X t,(l)> − X \ X \> p | {z } :=γ1     1 + PΩl X t,(l) X t,(l)> − X \ X \> − Pl X t,(l) X t,(l)> − X \ X \> , p | {z } (168) :=γ2 where PΩl is defined in (66). An application of Lemma 43 reveals that γ1 ≤ 2n X t,(l) Rt,(l) − X \ 2 2,∞ √ + 4 n log n X t,(l) Rt,(l) − X \ 2,∞ X\ , where Rt,(l) ∈ Or×r is defined in (72). Let C = X t,(l) X t,(l)> − X \ X \> as in (156), and one can bound the other term γ2 by taking advantage of the triangle inequality and the symmetry property: v uX r r n (i) (ii) 2u n n 2 2 t γ2 ≤ (δl,j − p) Cl,j . X \ 2,∞ , kCk∞ . X t,(l) Rt,(l) − X \ p j=1 p p 2,∞ Pn 2 where (i) comes from the standard Chernoff bound j=1 (δl,j − p)  np, and in (ii) we utilize the bound established in (158). The previous two bounds taken collectively give 2 √ γ ≤ 2n X t,(l) Rt,(l) − X \ + 4 n log n X t,(l) Rt,(l) − X \ 2,∞ r n δ e +C X t,(l) Rt,(l) − X \ X \ 2,∞ ≤ σmin p 2 2,∞ 2,∞ X\ (169) e > 0 and δ > 0 sufficiently small. The last inequality follows from (73c), the incoherence for some constant C condition (114) and our sample size condition. In summary, we obtain   ct,(l) − X ft+1,(l) ≤ η γ + 1 PΩ−l (E) X t+1,(l) H ∆t,(l) ≤ ηδσmin ∆t,(l) , (170) p for δ > 0 sufficiently small. With the estimate (170) in place, we can continue our derivation on β1 and β2 . (a) With regard to β1 , in view of (166) we can obtain  (i) β1 = η  ≤η (ii) = η ≤η  X t,(l) X t,(l)> − X \ X \> X t,(l) X t,(l)> − X \ X \>  t,(l) ∆ t,(l) ∆l,·  2 X t,(l) ct,(l) H >   ∆t,(l) 2 l,· 2 \ +X ∆ \ X t,(l) + Xl,· 72 l,· ∆t,(l) t,(l)>  ∆t,(l) l,· 2 2 ∆t,(l)  ∆t,(l) t,(l) ≤ η ∆l,· 2 \ ∆t,(l) + η Xl,· X t,(l) 2 ∆t,(l) 2 (171) , where (i) follows from the definitions of PΩ−l and Pl (see (67) and note that all entries in the lth row of PΩ−l (·) are identically zero), and the identity (ii) is due to the definition of ∆t,(l) in (162). (b) For β2 , we first claim that ft+1,(l) R − X \ X Ir := arg min R∈O r×r F (172) , whose justification follows similar reasonings as that of (138), and is therefore omitted. In particular, ft+1,(l) is symmetric and it gives rise to the facts that X \> X ft+1,(l) X
> X\  1 σmin Ir . 2 (173) We are now ready to invoke Lemma 36 to bound β2 . We abuse the notation and denote C :=

ft+1,(l) > X \ and E := X t+1,(l) H ct,(l) − X ft+1,(l) > X \ . We have X kEk ≤ 1 σmin ≤ σr (C) . 2 The first inequality arises from (170), namely, ct,(l) − X ft+1,(l) kEk ≤ X t+1,(l) H (i) ≤ ηδσmin X \ 2 (ii) ≤ X\ X \ ≤ ηδσmin ∆t,(l) 1 σmin , 2 where (i) holds since ∆t,(l) ≤ X \ and (ii) holds true for δ sufficiently small and η ≤ 1/σmax . Invoke Lemma 36 to obtain 2 kEk σr−1 (C) + σr (C) 2 ct,(l) − X ft+1,(l) ≤ X t+1,(l) H σmin β2 = kR1 − Ir k ≤ ≤ 2δη ∆t,(l) X\ (174) X\ , (175) where (174) follows since σr−1 (C) ≥ σr (C) ≥ σmin /2 from (173), and the last line comes from (170). (c) Putting the previous bounds (171) and (175) together yields t,(l) α1 ≤ η ∆l,· 2 \ ∆t,(l) + η Xl,· X t,(l) 2 ∆t,(l) 2 + 4δη X \ 2,∞ ∆t,(l) X\ . 3. Combine (160), (164) and (176) to reach  ct+1,(l) − X \ X t+1,(l) H t,(l) + η ∆l,· 2  t,(l) l,· 2 X t,(l)  ≤ 1 − ησmin + η X t,(l) (i) ≤ (1 − ησmin ) ∆l,· \ ∆t,(l) + η Xl,· ∆t,(l)  t,(l) 2 2 + 2η X \ ∆t,(l) 2 2,∞ + 4δη X \ + 4η X \ 2,∞ ∆t,(l) ! (ii)  1 C6 n log n σmin  t η C2 ρ µr √ + X \ 2,∞ σ ≤ 1− 2 np σmin p   r 1 2C10 n + 4η X \ X \ 2,∞ 2C9 ρt µr √ X\ + σ X\ np σmin p ∆l,· s 73 2 ∆t,(l) 2,∞ X\ X\ ∆t,(l) X\ (176) (iii) t+1 ≤ C2 ρ 1 X\ µr √ np C6 + σ 2,∞ σmin s n log n X\ p 2,∞ . Here, (i) follows since ∆t,(l) ≤ X \ and δ is sufficiently small, (ii) invokes the hypotheses (70e) and (73d) and recognizes that s ! 1 σmin 2C n log n 10 2C9 µr √ ≤ X t,(l) ∆t,(l) ≤ 2 X \ X\ + X\ σ np σmin np 2 holds under the sample size√and noise condition, while (iii) is valid as long as 1 − (σmin /3) · η ≤ ρ < 1, C2  κC9 and C6  κC10 / log n. B.7 Proof of Lemma 13 For notational convenience, we define the following two orthonormal matrices Q := arg min U 0R − U \ r×r R∈O and F Q(l) := arg min U 0,(l) R − U \ r×r R∈O F . ct (see (26)) is called the orthogonal Procrustes problem [tB77]. It is well-known The problem of finding H t c always exists and is given by that the minimizer H
ct = sgn X t> X \ . H Here, the sign matrix sgn(B) is defined as sgn(B) := U V > (177) for any matrix B with singular value decomposition B = U ΣV > , where the columns of U and V are left and right singular vectors, respectively. Before proceeding, we make note of the following perturbation bounds on M 0 and M (l) (as defined in Algorithm 2 and Algorithm 5, respectively):
1 1 PΩ M \ − M \ + PΩ (E) p p r r r r (ii) n n n σ n√ 2 ≤ C σmin M \ 2,∞ + Cσ =C X \ 2,∞ + C √ p p p σmin p  r r  (iii) (iv) 1 √ σ n ≤ C µr σmax + √ X \  σmin , np σmin p (i) M0 − M\ ≤ (178) for some universal constant C > 0. Here, (i) arises from the triangle inequality, (ii) utilizes Lemma 39 and Lemma 40, (iii) follows from the incoherence condition (114) and (iv) holds under our sample complexity assumption that n2 p  µ2 r2 n and the noise condition (27). Similarly, we have  r r  1 √ σ n (l) \ σmax + √ X \  σmin . (179) M − M . µr np σmin p Combine Weyl’s inequality, (178) and (179) to obtain Σ0 − Σ\ ≤ M 0 − M \  σmin and Σ(l) − Σ\ ≤ M (l) − M \  σmin , (180) and     1 σmin ≤ σr Σ(l) ≤ σ1 Σ(l) ≤ 2σmax . 2 (181) which further implies

1 σmin ≤ σr Σ0 ≤ σ1 Σ0 ≤ 2σmax 2 We start by proving (70a), (70b) and (70c). The key decomposition we need is the following h
i




c0 − X \ = U 0 Σ0 1/2 H c0 − Q + U 0 Σ0 1/2 Q − Q Σ\ 1/2 + U 0 Q − U \ Σ\ 1/2 . X 0H 74 (182) 1. For the spectral norm error bound in (70c), the triangle inequality together with (182) yields c0 − X \ ≤ X 0H Σ0
1/2 c0 − Q + H Σ0
1/2 Q − Q Σ\
1/2 + √ σmax U 0 Q − U \ , where we have also used the fact that kU 0 k = 1. Recognizing that M 0 − M \  σmin (see (178)) and the assumption σmax /σmin . 1, we can apply Lemma 47, Lemma 46 and Lemma 45 to obtain Σ0
1/2 c0 − Q . H Q − Q Σ\ 1 M0 − M\ , σmin
1/2 U 0Q − U \ . .√ 1 σmin (183a) 1 M0 − M\ , σmin (183b) M0 − M\ . (183c) These taken collectively imply the advertised upper bound √ 1 1 1 σmax M0 − M\ + √ M0 − M\ . √ M0 − M\ σmin σmin σmin  r r r  1 σmax σ n . µr X\ , + np σmin σmin p c0 − X \ . X 0H
1/2 √ ≤ 2σmax (see (181)) and the bounded condition number Σ0 where we also utilize the fact that assumption, i.e. σmax /σmin . 1. This finishes the proof of (70c). 2. With regard to the Frobenius norm bound in (70a), one has c0 − X \ X 0H F √ c0 − X \ r X 0H  r  r r  r  √ (i) 1 σ n √ 1 σ n √ σmax √ + r X \ = µr + r√ σmin . µr np σmin p np σmin p σmin  r r  (ii) 1 σ n √ . µr + r X\ F . np σmin p ≤ Here (i) arises from (70c) and (ii) holds true since σmax /σmin  1 and the proof of (70a). √ √ r σmin ≤ X \ F , thus completing 3. The proof of (70b) follows from similar arguments as used in proving (70c). Combine (182) and the triangle inequality to reach n o
1/2

c0 − X \ c0 − Q + Σ0 1/2 Q − Q Σ\ 1/2 X 0H ≤ U 0 2,∞ Σ0 H 2,∞ √ + σmax U 0 Q − U \ 2,∞ . Plugging in the estimates (178), (181), (183a) and (183b) results in  r r  √ 1 σ n 0 c0 \ X H −X + X \ U 0 2,∞ + σmax U 0 Q − U \ . µr np σmin p 2,∞ 2,∞ . It remains to study the component-wise error of U 0 . To this end, it has already been shown in [AFWZ17, Lemma 14] that  r r  1 σ n U 0 2,∞ . U \ 2,∞ (184) U 0 Q − U \ 2,∞ . µr + U \ 2,∞ and np σmin p 75 under our assumptions. These combined with the previous inequality give  r  r r  r  1 n √ 1 n σ σ 0 c0 \ \ . µr X H −X σmax U 2,∞ . µr X\ + + np σmin p np σmin p 2,∞ 2,∞ , where the last relation is due to the observation that √ σmax U \ . 2,∞ √ σmin U \ 2,∞ ≤ X\ 2,∞ . 4. We now move on to proving (70e). Recall that Q(l) = arg minR∈Or×r U 0,(l) R − U \ inequality, c0,(l) − X \ X 0,(l) H \ \ Note that Xl,· = Ml,· U \ Σ\
0,(l) l,· 2 0,(l) 0,(l) Xl,· ≤
−1/2 Xl,· ≤ Xl,· c0,(l) − Q(l) H c0,(l) H 2
−Q 2 + (l) X 0,(l) Q(l) − X \ + X 0,(l) Q (l) −X and, by construction of M (l) , (l) = Ml,· U 0,(l) Σ(l)
−1/2 \ = Ml,· U 0,(l) Σ(l)
−1/2 F
\ . By the triangle l,· 2
l,· 2 (185) . . We can thus decompose   n h 
−1/2 (l)
−1/2 i  0,(l) (l)
−1/2 o \ X 0,(l) Q(l) − X \ = Ml,· U 0,(l) Σ(l) Q − Q(l) Σ\ + U Q − U \ Σ\ , l,· which further implies that X 0,(l) Q(l) − X \
l,· 2 ≤ M\ 2,∞  Σ(l) In order to control this, we first see that Σ(l)
−1/2 Q(l) − Q(l) Σ\
−1/2
−1/2 = Σ(l) . 1 . Q(l) − Q(l) Σ\
−1/2 +√  1 U 0,(l) Q(l) − U \ . σmin (186)
−1/2 h (l)
1/2
1/2 (l) i \
−1/2 Q Σ\ − Σ(l) Q Σ Q(l) Σ\ σmin 1
1/2 − Σ(l) M (l) − M \ , 3/2 σmin
−1/2 Q(l) where the penultimate inequality uses (181) and the last inequality arises from Lemma 46. Additionally, Lemma 45 gives 1 U 0,(l) Q(l) − U \ . M (l) − M \ . σmin Plugging the previous two bounds into (186), we reach X 0,(l) (l) Q −X \
l,· 2 . 1 M 3/2 σmin where the last relation follows from M \ Note that this also implies that (·)l,· 2 0,(l) Xl,· (l) −M 2,∞ 2 \ M \ = X \ X \> ≤2 X \ 2,∞ 2,∞ 2,∞ . ≤  µr √ r 1 σ + np σmin σmax X \ 2,∞ r  n X\ p 0,(l) 0,(l) 2 = Xl,· Q(l) 2 ≤ . To see this, one has by the unitary invariance of X 0,(l) Q(l) − X \ 76 . and the estimate (179). , Xl,· 2,∞
l,· 2 \ + Xl,· 2 ≤ 2 X\ 2,∞ . Substituting the above bounds back to (185) yields in X 0,(l) c0,(l) H −X \
 r r  1 σ n 0,(l) (l) c + µr . X 2,∞ H −Q X\ + np σmin p  r r  1 n σ . µr X \ 2,∞ , + np σmin p \ l,· 2 2,∞ where the second line relies on Lemma 47, the bound (179), and the condition σmax /σmin  1. This establishes (70e). 5. Our final step is to justify (70d). Define B := arg minR∈Or×r U 0,(l) R − U 0 R0,(l) (cf. (72)), one has c0 − X 0,(l) R0,(l) ≤ X 0,(l) B − X 0 . X 0H F . From the definition of F F Recognizing that X 0,(l) B − X 0 = U 0,(l) h Σ(l)
1/2 B − B Σ0
1/2 i we can use the triangle inequality to bound
1/2
1/2 X 0,(l) B − X 0 ≤ Σ(l) B − B Σ0 F F  
1/2 , + U 0,(l) B − U 0 Σ0 + U 0,(l) B − U 0 Σ0 F In view of Lemma 46 and the bounds (178) and (179), one has Σ(l)
−1/2 B − BΣ1/2 F .√ 1 σmin From Davis-Kahan’s sinΘ theorem [DK70] we see that U 0,(l) B − U 0 F . 1 σmin These estimates taken together with (181) give X 0,(l) B − X 0 F .√ 1 σmin
M 0 − M (l) U 0,(l)
M 0 − M (l) U 0,(l) F
M 0 − M (l) U 0,(l) F
1/2 . . . F .
It then boils down to controlling M 0 − M (l) U 0,(l) F . Quantities of this type have showed up multiple times already, and hence we omit the proof details for conciseness (see Appendix B.5). With probability
at least 1 − O n−10 , s ( s )
0,(l) log n n log n 0 (l) M −M U . µr σmax + σ U 0,(l) . np p 2,∞ F If one further has U 0,(l) 2,∞ . U\ 2,∞ .√ 1 X\ σmin 2,∞ (187) , then taking the previous bounds collectively establishes the desired bound s ) ( s log n σ n log n 0 c0 0,(l) 0,(l) + X\ X H −X R . µr np σmin p F 2,∞ . Proof of Claim (187). Denote by M (l),zero the matrix derived by zeroing out the lth row/column of M (l) , and U (l),zero ∈ Rn×r containing the leading r eigenvectors of M (l),zero . On the one hand, [AFWZ17, Lemma 4 and Lemma 14] demonstrate that max kU (l),zero k2,∞ . kU \ k2,∞ . 1≤l≤n 77 On the other hand, by the Davis-Kahan sin Θ theorem [DK70] we obtain     1 U 0,(l) sgn U 0,(l)> U (l),zero − U (l),zero . M (l) − M (l),zero U (l),zero σmin F where sgn(A) denotes the sign matrix of A. For any j 6= l, one has     U (l),zero = M (l) − M (l),zero M (l) − M (l),zero j,· (l),zero since the lth row of Ul,· (l),zero j,l Ul,· F (188) , = 01×r , is identically zero by construction. In addition,   U (l),zero M (l) − M (l),zero l,· As a consequence, one has   M (l) − M (l),zero U (l),zero F 2 \ U (l),zero = Ml,· =  2 M (l) − M (l),zero ≤ M\  l,· 2,∞ ≤ σmax U \ U (l),zero which combined with (188) and the assumption σmax /σmin  1 yields   U 0,(l) sgn U 0,(l)> U (l),zero − U (l),zero . U \ F 2 2,∞ ≤ σmax U \ . 2,∞ , 2,∞ The claim (187) then follows by combining the above estimates:   U 0,(l) = U 0,(l) sgn U 0,(l)> U (l),zero 2,∞ 2,∞   (l),zero 0,(l) ≤ kU k2,∞ + U sgn U 0,(l)> U (l),zero − U (l),zero F . kU \ k2,∞ , where we have utilized the unitary invariance of k·k2,∞ . C Proofs for blind deconvolution Before proceeding to the proofs, we make note of the following concentration results. The standard Gaussian concentration inequality and the union bound give p (189) max a∗l x\ ≤ 5 log m 1≤l≤m with probability at least 1 − O(m−10 ). In addition, with probability exceeding 1 − Cm exp(−cK) for some constants c, C > 0, √ max kal k2 ≤ 3 K. (190) 1≤l≤m In addition, the population/expected Wirtinger Hessian at the truth z \ is given by   IK 0 0 h\ x\> 
 0 IK
x\ h\> 0 . ∗ ∇2 F z \ =  \   IK 0 0
x h\> \ \> ∗ hx 0 0 IK C.1 (191) Proof of Lemma 14 First, we find it convenient to decompose the Wirtinger Hessian (cf. (80)) into the expected Wirtinger Hessian at the truth (cf. (191)) and the perturbation part as follows:


∇2 f (z) = ∇2 F z \ + ∇2 f (z) − ∇2 F z \ . (192)
The proof then proceeds by showing that (i) the population Hessian ∇2 F z \ satisfies the restricted
strong 2 2 \ convexity and smoothness properties as advertised, and (ii) the perturbation ∇ f (z) − ∇ F z is wellcontrolled under our assumptions. We start by controlling the population Hessian in the following lemma. 78 Lemma 26. Instate the notation and the conditions of Lemma 14. We have


 2 ∇2 F z \ = 2 and u∗ D∇2 F z \ + ∇2 F z \ D u ≥ kuk2 . The next step is to bound the perturbation. To this end, we define the set S := {z : z satisfies (82)} , and derive the following lemma. Lemma 27. Suppose the sample complexity satisfies m  µ2 K log9 m, c >
0 is a sufficiently small constant, and δ = c/ log2 m. Then with probability at least 1 − O m−10 + e−K log m , one has sup ∇2 f (z) − ∇2 F z \ z∈S
≤ 1/4. Combining the two lemmas, we can easily see that for z ∈ S,

∇2 f (z) ≤ ∇2 F z \ + ∇2 f (z) − ∇2 F z \ ≤ 2 + 1/4 ≤ 3, which verifies the smoothness upper bound. In addition,   u∗ D∇2 f (z) + ∇2 f (z) D u 

 
 
 = u∗ D∇2 F z \ + ∇2 F z \ D u + u∗ D ∇2 f (z) − ∇2 F z \ u + u∗ ∇2 f (z) − ∇2 F z \ Du (i)


 2 ≥ u∗ D∇2 F z \ + ∇2 F z \ D u − 2 kDk ∇2 f (z) − ∇2 F z \ kuk2 (ii) 2 ≥ kuk2 − 2 (1 + δ) · (iii) ≥ 1 2 kuk2 4 1 2 kuk2 , 4 where (i) uses the triangle inequality, (ii) holds because of Lemma 27 and the fact that kDk ≤ 1 + δ, and (iii) follows if δ ≤ 1/2. This establishes the claim on the restricted strong convexity. C.1.1 Proof of Lemma 26 We start by proving the identity ∇2 F z \   h\ 1  0  , u1 = √  2 0  x\ 
= 2. Let  0 \ 1  x  , u2 = √  2  h\  0   h\ 1  0  , u3 = √  2 0  −x\   0 \ 1  x  . u4 = √  2  −h\  0 Recalling that kh\ k2 = kx\ k2 = 1, we can easily check that these four vectors form an orthonormal set. A little algebra reveals that
∇2 F z \ = I4K + u1 u∗1 + u2 u∗2 − u3 u∗3 − u4 u∗4 , which immediately implies ∇2 F z \
= 2.
We now turn attention to the restricted strong convexity. Since u∗ D∇2 F z \ u is the complex conjugate

of u∗ ∇2 F z \ Du as both ∇2 F (z \ ) and D are Hermitian, we will focus on the first term u∗ D∇2 F z \ u. This term can be rewritten as
u∗ D∇2 F z \ u 79   h\ x\> h1 − h2   x1 − x2 0    h1 − h2 0 x1 − x2 IK  \ \> h1 − h2 + h x (x1 − x2 ) h i 
(ii) ∗ x1 − x2 + x\ h\> (h1 − h2 ) ∗ ∗ ∗  = γ1 (h1 − h2 ) , γ2 (x1 − x2 ) , γ1 h1 − h2 , γ2 (x1 − x2 )   x\ h\>
∗ (x1 − x2 ) + (h1 − h2 ) 
∗ h\ x\> (h1 − h2 ) + (x1 − x2 )   ∗ h1 − h2 + h\ (x1 − x2 ) x\ ∗ \ \ h i
∗ ∗  x1 − x2 + x (h1 − h2 ) h  ∗ ∗  = γ1 (h1 − h2 ) , γ2 (x1 − x2 ) , γ1 h1 − h2 , γ2 (x1 − x2 )   h1 − h2 + h\ (x1 − x2 )∗ x\  ∗ x1 − x2 + x\ (h1 − h2 ) h\ IK h i 
∗ 0 (i) ∗ ∗ ∗ = (h1 − h2 ) , (x1 − x2 ) , h1 − h2 , (x1 − x2 ) D   0
∗ h\ x\>  2 0 IK
∗ x\ h\> 0 0 x\ h\> IK 0 2 = 2γ1 kh1 − h2 k2 + 2γ2 kx1 − x2 k2 ∗ ∗ ∗ ∗ + (γ1 + γ2 ) (h1 − h2 ) h\ (x1 − x2 ) x\ + (γ1 + γ2 ) (h1 − h2 ) h\ (x1 − x2 ) x\ , | | {z } {z } :=β (193) =β
where (i) uses the definitions of u and ∇2 F z \ , and (ii) follows from the definition of D. In view of the assumption (84), we can obtain   2 2 2 2 2 2γ1 kh1 − h2 k2 + 2γ2 kx1 − x2 k2 ≥ 2 min {γ1 , γ2 } kh1 − h2 k2 + kx1 − x2 k2 ≥ (1 − δ) kuk2 , where the last inequality utilizes the identity 2 2 2 (194) 2 kh1 − h2 k2 + 2 kx1 − x2 k2 = kuk2 . It then boils down to controlling β. Toward this goal, we decompose β into the following four terms

∗ ∗ ∗ ∗ β = (h1 − h2 ) h2 (x1 − x2 ) x2 + (h1 − h2 ) h\ − h2 (x1 − x2 ) x\ − x2 | {z } | {z } :=β1 + (h1 − h2 ) | ∗ \ |β2 | ≤ h\ − h2 2 x\ − x2 2 ∗ h − h2 (x1 − x2 ) x2 + (h1 − {z } | :=β3 Since h2 − h\ 2 and x2 − x\ regarding β2 , we discover that
2 :=β2 ∗ h2 ) h2 (x1 − x2 ) {z ∗ :=β4
x\ − x2 . } are both small by (83), β2 , β3 and β4 are well-bounded. Specifically, kh1 − h2 k2 kx1 − x2 k2 ≤ δ 2 kh1 − h2 k2 kx1 − x2 k2 ≤ δ kh1 − h2 k2 kx1 − x2 k2 , where the second inequality is due to (83) and the last one holds since δ < 1. Similarly, we can obtain |β3 | ≤ δ kx2 k2 kh1 − h2 k2 kx1 − x2 k2 ≤ 2δ kh1 − h2 k2 kx1 − x2 k2 , and |β4 | ≤ δ kh2 k2 kh1 − h2 k2 kx1 − x2 k2 ≤ 2δ kh1 − h2 k2 kx1 − x2 k2 , where both lines make use of the facts that kx2 k2 ≤ x2 − x\ 2 + x\ 2 ≤1+δ ≤2 and kh2 k2 ≤ h2 − h\ 2 + h\ 2 ≤ 1 + δ ≤ 2. Combine the previous three bounds to reach 2 |β2 | + |β3 | + |β4 | ≤ 5δ kh1 − h2 k2 kx1 − x2 k2 ≤ 5δ where we utilize the elementary inequality ab ≤ (a2 + b2 )/2 and the identity (194). 80 2 kh1 − h2 k2 + kx1 − x2 k2 5 2 = δ kuk2 , 2 4 (195)     The only remaining term is thus β1 . Recalling that (h1 , x1 ) and (h2 , x2 ) are aligned by our assumption, we can invoke Lemma 56 to obtain ∗ 2 2 (h1 − h2 ) h2 = kx1 − x2 k2 + x∗2 (x1 − x2 ) − kh1 − h2 k2 , which allows one to rewrite β1 as o n ∗ 2 2 β1 = kx1 − x2 k2 + x∗2 (x1 − x2 ) − kh1 − h2 k2 · (x1 − x2 ) x2   2 ∗ ∗ 2 2 = (x1 − x2 ) x2 kx1 − x2 k2 − kh1 − h2 k2 + (x1 − x2 ) x2 . Consequently,    2 2 ∗ ∗ 2 β1 + β1 = 2 (x1 − x2 ) x2 2 + 2Re (x1 − x2 ) x2 kx1 − x2 k2 − kh1 − h2 k2    2 ∗ 2 ≥ 2Re (x1 − x2 ) x2 kx1 − x2 k2 − kh1 − h2 k2 (i) ∗ 2 ≥ − (x1 − x2 ) x2 kuk2 (ii) 2 ≥ −4δ kuk2 . Here, (i) arises from the triangle inequality that 2 2 2 1 2 kuk2 , 2 2 kx1 − x2 k2 − kh1 − h2 k2 ≤ kx1 − x2 k2 + kh1 − h2 k2 = and (ii) occurs since kx1 − x2 k2 ≤ kx1 − x\ k2 + kx2 − x\ k2 ≤ 2δ and kx2 k2 ≤ 2 (see (195)). To finish up, note that γ1 + γ2 ≤ 2(1 + δ) ≤ 3 for δ < 1/2. Substitute these bounds into (193) to obtain

2 u∗ D∇2 F z \ u ≥ (1 − δ) kuk2 + (γ1 + γ2 ) β + β
2 ≥ (1 − δ) kuk2 + (γ1 + γ2 ) β1 + β1 − 2 (γ1 + γ2 ) (|β2 | + |β3 | + |β4 |) 5 2 2 2 ≥ (1 − δ) kuk2 − 12δ kuk2 − 6 · δ kuk2 4 2 ≥ (1 − 20.5δ) kuk2 1 2 ≥ kuk2 2 as long as δ is small enough. C.1.2 Proof of Lemma 27
In view of the expressions of ∇2 f (z) and ∇2 F z \ (cf. (80) and (191)) and the triangle inequality, we get ∇2 f (z) − ∇2 F z \
(196) ≤ 2α1 + 2α2 + 4α3 + 4α4 , where the four terms on the right-hand side are defined as follows α1 = m X a∗j x bj b∗j − IK , m X
b∗j hx∗ aj − yj bj a∗j , j=1 α3 = j=1 2 α2 = m X j=1 α4 = 2 b∗j h aj a∗j − IK , m X bj b∗j h aj a∗j x j=1 In what follows, we shall control supz∈S αj for j = 1, 2, 3, 4 separately. 81
> − h\ x\> . 1. Regarding the first term α1 , the triangle inequality gives α1 ≤ | m X m X 2 a∗j x bj b∗j − j=1 {z 2 a∗j x\ bj b∗j + j=1 | } :=β1 m X j=1 2 a∗j x\ bj b∗j − IK . {z } :=β2 • To control β1 , the key observation is that a∗j x and a∗j x\ are extremely close. We can rewrite β1 as β1 = m  X a∗j x 2 j=1 − a∗j x\ 2  m X bj b∗j ≤ a∗j x 2 j=1 − a∗j x\ 2 bj b∗j , (197) where a∗j x 2 − a∗j x\ 2 (i) =  a∗j x − x\
∗
+ 2 a∗j x − x\ a∗j x\ (iii) p 1 1 ≤ 4C32 3 + 4C3 3/2 · 5 log m log m log m 1 . C3 . log m (ii) ≤ a∗j x − x\


∗ ∗ \
∗
a∗j x − x\ + a∗j x − x\ aj x + a∗j x\ a∗j x − x\ 2 Here, the first line (i) uses the identity for u, v ∈ C, |u|2 − |v|2 = u∗ u − v ∗ v = (u − v)∗ (u − v) + (u − v)∗ v + v ∗ (u − v), the second relation (ii) comes from the triangle inequality, and the third line (iii) follows from (189) and the assumption (82b). Substitution into (197) gives β1 ≤ max 1≤j≤m 2 a∗j x − 2 a∗j x\ m X bj b∗j . C3 j=1 where the last inequality comes from the fact that Pm j=1 1 , log m bj b∗j = IK .
• The other term β2 can be bounded through Lemma 59, which reveals that with probability 1−O m−10 , β2 . r K log m. m Taken collectively, the preceding two bounds give r K 1 sup α1 . log m + C3 . m log m z∈S Hence P(supz∈S α1 ≤ 1/32) = 1 − O(m−10 ). 2. We are going to prove that P(supz∈S α2 ≤ 1/32) = 1 − O(m−10 ). The triangle inequality allows us to bound α2 as α2 ≤ | m X j=1 2 2 2 b∗j h aj a∗j − khk2 IK + khk2 IK − IK . {z } :=θ1 (h) 82 | {z :=θ2 (h) } The second term θ2 (h) is easy to control. To see this, we have 2 θ2 (h) = khk2 − 1 = khk2 − 1 (khk2 + 1) ≤ 3δ < 1/64, where the penultimate relation uses the assumption that h − h\ khk2 − 1 ≤ δ, 2 ≤ δ and hence khk2 ≤ 1 + δ ≤ 2. For the first term θ1 (h), we define a new set  H := h ∈ CK : kh − h\ k2 ≤ δ and max 1≤j≤m 2C4 µ log2 m √ ≤ m b∗j h  . It is easily seen that supz∈S θ1 ≤ suph∈H θ1 . We plan to use the standard covering argument to show that   P sup θ1 (h) ≤ 1/64 = 1 − O(m−10 ). (198) h∈H To this end, we define cj (h) = |b∗j h|2 for every 1 ≤ j ≤ m. It is straightforward to check that θ1 (h) = m X j=1 m X j=1 c2j = m X j=1 |b∗j h|4 ≤  cj (h) aj a∗j − IK max 1≤j≤m |b∗j h|2 X m j=1
max |cj | ≤ , |b∗j h|2 = 1≤j≤m  max 1≤j≤m |b∗j h|2 for h ∈ H. In the above argument, we have used the facts that m X j=1  |b∗j h|2 = h∗  m X j=1  Pm   j=1 2C4 µ log2 m √ m khk22 ≤4  2 (199) , 2C4 µ log2 m √ m 2 (200) bj b∗j = IK and bj b∗j  h = khk22 ≤ (1 + δ)2 ≤ 4, together with the definition of H. Lemma 57 combined with (199) and (200) readily yields that for any fixed h ∈ H and any t ≥ 0, ( )! 2 t t e1 K − C e2 min , Pm 2 P(θ1 (h) ≥ t) ≤ 2 exp C max1≤j≤m |cj | j=1 cj   e1 K − C e2 mt min {1, t/4} , ≤ 2 exp C (201) 4C42 µ2 log4 m e1 , C e2 > 0 are some universal constants. where C Now we are in a position to strengthen this bound to obtain uniform control of θ1 over H. Note that for any h1 , h2 ∈ H, |θ1 (h1 ) − θ1 (h2 )| ≤ m X j=1
|b∗j h1 |2 − |b∗j h2 |2 aj a∗j + kh1 k22 − kh2 k22 = max |b∗j h1 |2 − |b∗j h2 |2 1≤j≤m m X j=1 aj a∗j + kh1 k22 − kh2 k22 , where |b∗j h2 |2 − |b∗j h1 |2 = (h2 − h1 )∗ bj b∗j h2 + h∗1 bj b∗j (h2 − h1 ) 83 ≤ 2 max{kh1 k2 , kh2 k2 }kh2 − h1 k2 kbj k22 4K ≤ 4kh2 − h1 k2 kbj k22 ≤ kh2 − h1 k2 m and kh1 k22 − kh2 k22 = |h∗1 (h1 − h2 ) − (h1 − h2 )∗ h2 | ≤ 2 max{kh1 k2 , kh2 k2 }kh2 − h1 k2 ≤ 4kh1 − h2 k2 . o n P m ∗ Define an event E0 = j=1 aj aj ≤ 2m . When E0 happens, the previous estimates give |θ1 (h1 ) − θ1 (h2 )| ≤ (8K + 4)kh1 − h2 k2 ≤ 10Kkh1 − h2 k2 , ∀h1 , h2 ∈ H. e be an ε-net covering H (see [Ver12, Definition 5.1]). We have Let ε = 1/(1280K), and H ( ) !   1 1 sup θ1 (h) ≤ ∩ E0 ⊆ sup θ1 ≤ 128 64 h∈H e h∈H and as a result,  1 P sup θ1 (h) ≥ 64 h∈H  1 ≤ P sup θ1 (h) ≥ 128 e h∈H !   e · max P θ1 (h) ≥ 1 + P(E0c ) ≤ |H| + P(E0c ). e 128 h∈H e ≤C e3 K log K for some absolute Lemma 57 forces that P(E0c ) = O(m−10 ). Additionally, we have log |H| e constant C3 > 0 according to [Ver12, Lemma 5.2]. Hence (201) leads to     m(1/128) min {1, (1/128)/4} 1 e e e e |H| · max P θ1 (h) ≥ ≤ 2 exp C3 K log K + C1 K − C2 e 128 4C42 µ2 log4 m h∈H ! e e3 K log m − C4 m ≤ 2 exp 2C µ2 log4 m e4 > 0. Under the sample complexity m  µ2 K log5 m, the right-hand side of the for some constant C
above display is at most O m−10 . Combine the estimates above to establish the desired high-probability bound for supz∈S α2 . 3. Next, we will demonstrate that
P( sup α3 ≤ 1/96) = 1 − O m−10 + e−K log m . z∈S To this end, we let   a∗1   A =  ...  ∈ Cm×K , a∗m   b∗1   B =  ...  ∈ Cm×K , b∗m where for each 1 ≤ j ≤ m,   C=   c1 (z) c2 (z) ··· cm (z)   ∈ Cm×m ,  cj (z) := b∗j hx∗ aj − yj = b∗j (hx∗ − h\ x\∗ )aj . As a consequence, we can write α3 = kB ∗ CAk. The key observation is that both the `∞ norm and the Frobenius norm
of C are well-controlled. Specifically, we claim for the moment that with probability at least 1 − O m−10 , kCk∞ = max |cj | ≤ C 1≤j≤m 84 µ log5/2 m √ ; m (202a) 2 kCkF = m X j=1 2 |cj | ≤ 12δ 2 , (202b) where C > 0 is some absolute constant. This motivates us to divide the entries in C into multiple groups based on their magnitudes. To be precise, introduce R := 1 + dlog2 (Cµ log7/2 m)e sets {Ir }1≤r≤R , where ) ( Cµ log5/2 m Cµ log5/2 m √ √ < |cj | ≤ , 1≤r ≤R−1 Ir = j ∈ [m] : 2r m 2r−1 m SR−1
and IR = {1, · · · , m} \ r=1 Ir . An immediate consequence of the definition of Ir and the norm constraints in (202) is the following cardinality bound 2 |Ir | ≤ kCkF minj∈Ir |cj | 2 ≤ 12δ 2 Cµ log5/2 2 √ r m m 2 = 12δ 2 4r m log5 m | {z } C 2 µ2 (203) δr for 1 ≤ r ≤ R − 1. Since {Ir }1≤r≤R form a partition of the index set {1, · · · , m}, it is easy to see that B ∗ CA = R X r=1 (BIr ,· )∗ CIr ,Ir AIr ,· , where DI,J denotes the submatrix of D induced by the rows and columns of D having indices from I and J , respectively, and DI,· refers to the submatrix formed by the rows from the index set I. As a result, one can invoke the triangle inequality to derive α3 ≤ R−1 X r=1 kBIr ,· k · kCIr ,Ir k · kAIr ,· k + kBIR ,· k · kCIR ,IR k · kAIR ,· k . (204) Recognizing that B ∗ B = IK , we obtain kBIr ,· k ≤ kBk = 1 for every 1 ≤ r ≤ R. In addition, by construction of Ir , we have kCIr ,Ir k = max |cj | ≤ j∈Ir Cµ log5/2 m √ 2r−1 m for 1 ≤ r ≤ R, and specifically for R, one has kCIR ,IR k = max |cj | ≤ j∈IR Cµ log5/2 m 1 √ ≤√ , 2R−1 m m log m which follows from the definition of R, i.e. R = 1 + dlog2 (Cµ log7/2 m)e. Regarding kAIr ,· k, we discover that kAIR ,· k ≤ kAk and in view of (203), kAIr ,· k ≤ sup I:|I|≤δr m kAI,· k , 1 ≤ r ≤ R − 1. Substitute the above estimates into (204) to get α3 ≤ R−1 X r=1 Cµ log5/2 m kAk √ sup kAI,· k + √ . r−1 2 m I:|I|≤δr m m log m 85 (205) √ It remains to upper bound kAk and supI:|I|≤δr m kAI,· k. Lemma 57 tells us that kAk ≤ 2 m with
probability at least 1 − O m−10 . Furthermore, we can invoke Lemma 58 to bound supI:|I|≤δr m kAI,· k for each 1 ≤ r ≤ R − 1. It is easily seen from our assumptions m  µ2 K log9 m and δ = c/ log2 m that δr  K/m. In addition, 7/2 12δ 2 41+log2 (Cµ log 12δ 2 4R−1 ≤ δr ≤ 2 2 5 C µ log m C 2 µ2 log5 m m) = 48δ 2 log2 m = 48c  1. log2 m e2 , C e3 > 0 By Lemma 58 we obtain that for some constants C ! ! q e2 C e3 C e3 δr m log(e/δr ) ≤ 2 exp − δr m log(e/δr ) P sup kAI,· k ≥ 4C 3 I:|I|≤δr m ! e2 C e3 C ≤ 2 exp − δr m ≤ 2e−K . 3 Taking the union
bound and substituting the estimates above into (205), we see that with probability at
least 1 − O m−10 − O (R − 1)e−K , √ q Cµ log5/2 m e3 δr m log(e/δr ) + √ 2 m √ · 4 C 2r−1 m m log m r=1 R−1 X q e3 log(e/δr ) + 2 ≤ 4δ 12C log m r=1 p 1 . . (R − 1)δ log(e/δ1 ) + log m α3 ≤ Note that µ ≤ √ R−1 X m, R − 1 = dlog2 (Cµ log7/2 m)e . log m, and s   r e eC 2 µ2 log5 m log . log m. = log δ1 48δ 2

Therefore, with probability exceeding 1 − O m−10 − O e−K log m , sup α3 . δ log2 m + z∈S 1 . log m By taking c to be small enough in δ = c/ log2 m, we get  

P sup α3 ≥ 1/96 ≤ O m−10 + O e−K log m z∈S as claimed. Finally, it remains to justify (202). For all z ∈ S, the triangle inequality tells us that |cj | ≤ b∗j h(x − x\ )∗ aj + b∗j (h − h\ )x\∗ aj
b∗j h + b∗j h\ · a∗j x\   p 2C3 2C4 µ log2 m µ 2C4 µ log2 m √ √ √ · + + 5 log m ≤ 3/2 m m m log m ≤ b∗j h · a∗j (x − x\ ) + ≤C µ log5/2 m √ , m 86 for some large constant C > 0, where we have used the definition of S and the fact (189). The claim (202b) follows directly from [LLSW16, Lemma 5.14]. To avoid confusion, we use µ1 to refer to the parameter µ therein. Let L = m, N = K, d0 = 1, µ1 = C4 µ log2 m/2, and ε = 1/15. Then S ⊆ Nd0 ∩ Nµ1 ∩ Nε , and the sample complexity condition L  µ21 (K + N ) log2 L is satisfied
because we have assumed m  µ2 K log6 m. Therefore with probability exceeding 1 − O m−10 + e−K , we obtain that for all z ∈ S, 2 kCkF ≤ 5 hx∗ − h\ x\∗ 4 2 F . The claim (202b) can then be justified by observing that hx∗ − h\ x\∗ F = h x − x\
∗
+ h − h\ x\∗ F ≤ khk2 x − x\ 2 + h − h\ 2 x\ 2 ≤ 3δ. 4. It remains to control α4 , for which we make note of the following inequality α4 ≤ | m X j=1 bj b∗j (hx> − h\ x\> )aj aj ∗ + {z } θ3 | m X j=1 bj b∗j h\ x\> (aj aj ∗ − IK ) {z } θ4 with aj denoting the entrywise conjugate of aj . Since {aj } has the same joint distribution as {aj }, by the same argument used for bounding α3 we obtain control of the first term, namely,   P sup θ3 ≥ 1/96 = O(m−10 + e−K log m). z∈S Note that m  µ2 K log m/δ 2 and δ  1. According to [LLSW16, Lemma 5.20],     P sup θ4 ≥ 1/96 ≤ P sup θ4 ≥ δ = O(m−10 ). z∈S z∈S Putting together the above bounds, we reach P(supz∈S α4 ≤ 1/48) = 1 − O(m−10 + e−K log m). 5. Combining all the previous bounds for supz∈S αj and (196), we deduce that with probability 1−O(m−10 + e−K log m),
1 1 1 1 1 ∇2 f (z) − ∇2 F z \ ≤ 2 · +2· +4· +4· = . 32 32 96 48 4 C.2 Proofs of Lemma 15 and Lemma 16 Proof of Lemma 15. In view of the definition of αt+1 (see (38)), one has dist z t+1 , z \
2 = 1 αt+1 2 ht+1 − h\ 2 + αt+1 xt+1 − x\ 2 2 ≤ 1 t+1 h − h\ αt 2 2 + αt xt+1 − x\ The gradient update rules (79) imply that  

1 η 1 t+1 t e t − η ∇h f zet , h = ht − ∇ f z =h h 2 2 t kx k2 αt αt ke xt k2  

η η et − αt xt+1 = αt xt − ∇x f z t =x ∇x f zet , 2 t 2 t e kh k2 kh k2 87 2 2 . b t+1 = 1 ht+1 and x e t = 1 ht and x et = αt xt as in (81). Let h bt+1 = αt xt+1 . We further where we denote h αt αt get       −2  b t+1 − h\ e t − h\ ke xt k2 IK h h ∇h f (e zt) −2   ∇ f (e  x  et − x\   t+1 t et − x\  h IK  x z )    b   x  2 . = − η   b t+1     −2 e t − h\  zt)    ∇h f (e  h  ke xt k2 IK − h\   h e t −2 IK ∇x f (e zt) bt+1 − x\ et − x\ x x h 2 | {z } :=D (206)
\ The fundamental theorem of calculus (see Appendix D.3.1) together with the fact that ∇f z = 0 tells us  
    e t − h\ h ∇h f (e z t ) − ∇h f z \
∇h f (e zt) Z 1  x t \    ∇x f (e z t ) − ∇x f z \  zt)   e −x  2  =  ∇x f (e =  , (207) ∇ f (z (τ )) dτ   ∇h f (e e t − h\  z t )   ∇h f (e z t ) − ∇h f (z \ )   h  0 | {z } ∇x f (e zt) ∇x f (e z t ) − ∇x f (z \ ) e t − x\ x :=A
where we denote z (τ ) := z \ + τ zet − z \ and ∇2 f is the Wirtinger Hessian. To further simplify notation,   b t+1 h denote zbt+1 = t+1 . The identity (207) allows us to rewrite (206) as b x  t+1   t  zb − z\ ze − z \ = (I − ηDA) . (208) zbt+1 − z \ zet − z \ Take the squared Euclidean norm of both sides of (208) to reach   t ∗  1 zet − z \ ze − z \ ∗ t+1 \ 2 zb −z 2 = (I − ηDA) (I − ηDA) zet − z \ 2 zet − z \     t ∗
zet − z \ 1 ze − z \ 2 2 I + η AD A − η (DA + AD) = zet − z \ 2 zet − z \  t  t   ∗ η ze − z \ ze − z \ t \ 2 2 2 2 ≤ (1 + η kAk kDk ) ze − z 2 − (DA + AD) . zet − z \ 2 zet − z \ Since z (τ ) lies between zet and z \ , we conclude from the assumptions (85) that for all 0 ≤ τ ≤ 1,
 max h (τ ) − h\ 2 , x (τ ) − x\ 2 ≤ dist z t , z \ ≤ ξ ≤ δ;
1 max a∗j x (τ ) − x\ ≤ C3 3/2 ; 1≤j≤m log m µ max b∗j h (τ ) ≤ C4 √ log2 m 1≤j≤m m for ξ > 0 sufficiently small. Moreover, it is straightforward to see that et γ1 := x satisfy −2 2 et γ2 := h and max {|γ1 − 1| , |γ2 − 1|} . max n e t − h\ h −2 2 et − x\ , x 2 2 o ≤δ as long as ξ > 0 is sufficiently small. We can now readily invoke Lemma 14 to arrive at  zet − z \ zet − z \ ∗ kAk kDk ≤ 3(1 + δ) ≤ 4 (DA + AD)  zet − z \ zet − z \  88 ≥ 1 4  and zet − z \ zet − z \  2 = 2 1 t ze − z \ 2 2 2 . (209) Substitution into (209) indicates that 2 zbt+1 − z \ 2 When 0 < η ≤ 1/128, this implies that zbt − z \ and hence zet+1 − z \ 2 ≤ zbt+1 − z \
≤ 1 + 16η 2 − η/4 zet − z \ 2 2 1/2 2 2 ≤ (1 − η/8) zet − z \ ≤ (1 − η/8) This completes the proof of Lemma 15. zet − z \ 2 2 2 2 . , ≤ (1 − η/16) dist(z t , z \ ). Proof of Lemma 16. Reuse the notation in this subsection, namely, zbt+1 = bt+1 = αt xt+1 . From (210), one can tell that and x zet+1 − z \ Invoke Lemma 52 with β = αt to get 2 ≤ zbt+1 − z \ αt+1 − αt . zbt+1 − z \ ≤ dist(z t , z \ ). 2 2  b t+1 h bt+1 x  b t+1 = with h (210) 1 t+1 h αt ≤ dist(z t , z \ ). This combined with the assumption ||αt | − 1| ≤ 1/2 implies that αt ≥ 1 2 αt+1 − αt αt+1 1 = − 1 . dist(z t , z \ ) . C1 2 . αt αt log m and This finishes the proof of the first claim. The second claim can be proved by induction. Suppose that |αs | − 1 ≤ 1/2 and dist(z s , z \ ) ≤ C1 (1 − η/16)s / log2 m hold for all 0 ≤ s ≤ τ ≤ t , then using our result in the first part gives |α τ +1 0 | − 1 ≤ |α | − 1 + ≤ 1 + 4 η 16 τ X α s+1 s=0 −α s cC1 1 ≤ 2 2 log m τ X 1 ≤ +c dist(z s , z \ ) 4 s=0 for m sufficiently large. The proof is then complete by induction. C.3 Proof of Lemma 17 Define the alignment parameter between z t,(l) and zet as 2 1 t,(l) 1 h − ht + αxt,(l) − αt xt t α α α∈C 2  t,(l)  b h Further denote, for simplicity of presentation, zbt,(l) = t,(l) with b x t,(l) αmutual := argmin b t,(l) := h 1 t,(l) αmutual ht,(l) 89 2 . t,(l) bt,(l) := αmutual xt,(l) . x and Clearly, zbt,(l) is aligned with zet . Armed with the above notation, we have s
1 t+1,(l) 1 dist z t+1,(l) , zet+1 = min h − ht+1 t+1 α α α 2 2 2 + αxt+1,(l) − αt+1 xt+1 2 2 v u u = min t αt αt+1 α v u u u ≤t ≤ max αt αt+1  ! ! αt+1 αt  1 1 αt+1 t+1,(l) h − ht+1 t α α αt 1 ht+1,(l) t,(l)  ! 2 + 2 2 1 − ht+1  αt  +  αt+1 αt αt+1 αt αmutual 2  " 1 1 t+1 # t+1,(l) t h − h α t,(l) αt α mutual , t+1 t,(l) α αmutual xt+1,(l) − αt xt+1 t+1   αt α t+1 xt+1,(l) − αt xt+1 α  t,(l) αmutual xt+1,(l) − αt xt+1  2 2 2 (211) 2 (212) , 2 t,(l) where (211) follows by taking α = ααt αmutual . The latter bound is more convenient to work with when controlling the gap between z t,(l) and z t . We can then apply the gradient update rules (79) and (89) to get # " 1 ht+1,(l) − α1t ht+1 t,(l) αmutual t,(l) αmutual xt+1,(l) − αt xt+1     
η 1 (l) t,(l) t,(l) t,(l) h ,x − α1t ht − kxηt k2 ∇h f (ht , xt ) h − t,(l) 2 ∇h f  αt,(l) kx k2 2 mutual   =    t,(l)
η η αmutual xt,(l) − t,(l) 2 ∇x f (l) ht,(l) , xt,(l) − αt xt − kht k2 ∇x f (ht , xt ) 2 kh k2    

η (l) b t,(l) bt,(l) e t − ηt 2 ∇h f h b t,(l) − e t, x et h ,x − h h 2∇ f ke x k2 kxbt,(l) k2 h  =

  . η η t (l) b t,(l) bt,(l) t e t, x e bt,(l) − b t,(l) e x − x ∇ f h , x − ∇ f h x x 2 et 2 kh k2     kh k2 By construction, we can write the leave-one-out gradients as ∇h f (l) (h, x) = ∇h f (h, x) − (b∗l hx∗ al − yl ) bl a∗l x ∇x f (l) (h, x) = ∇h f (h, x) − (b∗l hx∗ al − and yl )al b∗l h, which allow us to continue the derivation and obtain 
 t " η t,(l) 1 t+1 # 1 t+1,(l) b t,(l) − b t,(l) , x e − ηt 2 ∇h f b h ∇ f h − h h − h 2 h ke x k2 t,(l) bt,(l) k x αt k αmutual  2  =
t,(l) η t,(l) t,(l) t t,(l) t+1,(l) t t+1 b b e − eηt 2 ∇x f b − x x − b t,(l) 2 ∇x f h , x αmutual x −α x kh k2 kh k2     1 ∗ b t,(l) bt,(l)∗ ∗ bt,(l) x a − y b h b a x 2 l l l l l  kxbt,(l) k2    −η . 1 b t,(l) x b t,(l) bt,(l)∗ al − yl al b∗l h b∗l h 2 t,(l) b kh k2 {z } |
  e t, x et h
  e t, x et h :=J3 This further gives " 1 t,(l) ht+1,(l) − αmutual t,(l) αmutual xt+1,(l) 1 t+1 h αt − αt xt+1 #      η η t,(l) t,(l) bt,(l) t t et b b e e − t,(l) 2 ∇h f h , x − h − t,(l) 2 ∇h f h , x  h kxb k2 kxb k2  =      t,(l) η η t,(l) t,(l) t t t b e b b e − b t,(l) 2 ∇x f h , x e x − b t,(l) 2 ∇x f h , x − x kh k2 kh k2 | {z     +η  |   e t, x et − t,(l) 2 ∇h f h kxb k2     1 1 t e t, x e − ∇ f h 2 2 x etk b t,(l) k kh kh 2 {z2 1 ke xt k22 1  :=ν1 :=ν2 In what follows, we bound the three terms ν1 , ν2 , and ν3 separately. 90    −ην3 .  }     } (213) 1. Regarding the first term ν1 , one can adopt the same strategy as in Appendix C.2. Specifically, write   
 t η η t t,(l) b t,(l) − e − b h ∇ f ∇ f (e z ) z − h 2 2 h h   kxbt,(l) k2 kxbt,(l) k2  b t,(l) − h et   h
 t   t,(l) η η t,(l) t b e − b t,(l) 2 ∇x f (e  x − b t,(l) 2 ∇x f zb z)   x − x t,(l) et  −x   b   kh k2 kh k2 =     
b t,(l) − h et    b t,(l)   η η h t,(l) t t e − kb ∇ f zb − h − kb ∇ f (e z)   h xt,(l) k22 h xt,(l) k22 h t,(l) − x t     b e x
η η bt,(l) − b t,(l) et − b t,(l) x ∇x f zbt,(l) − x ∇x f (e zt) kh k22 kh k22   −2  
bt,(l) 2 IK x ∇h f zbt,(l)
− ∇h f (e zt) −2   b t,(l)   ∇ f zbt,(l) − ∇ f (e  IK h zt)  x  x   2
−η .   t,(l) − ∇ f (e t)  t,(l) −2 b ∇ f z z     h h b x I K
2   −2 ∇x f zbt,(l) − ∇x f (e zt) b t,(l) IK h 2 | {z } :=D The fundamental theorem of calculus (see Appendix D.3.1) reveals that   
b t,(l) − h et h ∇h f zbt,(l)
− ∇h f (e zt) Z 1  ∇ f zbt,(l) − ∇ f (e  t,(l) t  b et −x z)  x  x  x
∇2 f (z (τ )) dτ  t,(l)  = t,(l) t b et − ∇h f (e z)   ∇h f zb  h −h
{z } |0 t,(l) t t,(l) ∇x f zb − ∇x f (e z) b et x −x :=A    , 
where we abuse the notation and denote z (τ ) = zet + τ zbt,(l) − zet . In order to invoke Lemma 14, we need to verify the conditions required therein. Recall the induction hypothesis (90b) that s
µ µ2 K log9 m dist z t,(l) , zet = zbt,(l) − zet 2 ≤ C2 √ , m m and the fact that z (τ ) lies between zbt,(l) and zet . For all 0 ≤ τ ≤ 1: √ (a) If m  µ2 K log13/2 m, then o n z (τ ) − z \ 2 ≤ max zbt,(l) − z \ 2 , zet − z \ 2 ≤ zet − z \ 2 + zbt,(l) − zet s 1 µ µ2 K log9 m 1 ≤ C1 2 + C2 √ ≤ 2C1 2 , m m log m log m 2 where we have used the induction hypotheses (90a) and (90b); (b) If m  µ2 K log6 m, then max a∗j x (τ ) − x\ 1≤j≤m


bt,(l) − x et + a∗j x e t − x\ = max τ a∗j x 1≤j≤m bt,(l) − x et ≤ max a∗j x 1≤j≤m
e t − x\ + max a∗j x ≤ max kaj k2 zbt,(l) − zet 1≤j≤m √ µ ≤ 3 K · C2 √ m s 1≤j≤m 2 + C3 1 log 3/2
m µ2 K log9 m 1 1 + C3 3/2 ≤ 2C3 3/2 , m log m log m which follows from the bound (190) and the induction hypotheses (90b) and (90c); 91 (214) (c) If m  µK log5/2 m, then
b t,(l) − h e t + b∗ h et max b∗j h (τ ) = max τ b∗j h j 1≤j≤m 1≤j≤m b t,(l) − h et ≤ max b∗j h 1≤j≤m b t,(l)
et + max b∗j h 1≤j≤m et et ≤ max kbj k2 h − h 2 + max b∗j h 1≤j≤m 1≤j≤m s r K µ2 K log9 m µ µ µ · C2 √ + C4 √ log2 m ≤ 2C4 √ log2 m, (215) ≤ m m m m m p which makes use of the fact kbj k2 = K/m as well as the induction hypotheses (90b) and (90d). These properties satisfy the condition (82) required in Lemma 14. The other two conditions (83) and (84) are also straightforward to check and hence we omit it. Thus, we can repeat the argument used in Appendix C.2 to obtain kν1 k2 ≤ (1 − η/16) · zbt,(l) − zet 2 . 2. In terms of the second term ν2 , it is easily seen that ( 1 1 1 kν2 k2 ≤ max − , 2 2 et et 2 bt,(l) 2 x x h 2 2 − 1 2 2 b t,(l) h )  ∇h f (e zt) ∇x f (e zt)  . 2 We first note that the upper bound on k∇2 f (·) k (which essentially provides a Lipschitz constant on the gradient) in Lemma 14 forces
    1 ∇h f (e z t ) − ∇h f z \
∇h f (e zt) = . zet − z \ 2 . C1 2 , t \ ∇x f (e zt) ∇ f (e z ) − ∇ f z x x log m 2 2
where the first identity follows since ∇h f z \ = 0, and the last inequality comes from the induction bt,(l) 2  1, one can easily verify that hypothesis (90a). Additionally, recognizing that ke xt k2  x 1 2 ke xt k2 − 1 bt,(l) x 2 2 = bt,(l) x 2 2 2 − ke xt k2 2 2 2 bt,(l) ke xt k2 · x . bt,(l) x 2 et − x 2 bt,(l) − x et . x 2 . A similar bound holds for the other term involving h. Combining the estimates above thus yields kν2 k2 . C1 1 zbt,(l) − zet log2 m 2 . 3. When it comes to the last term ν3 , one first sees that   b t,(l) x b t,(l) x bt,(l)∗ al − yl bl a∗l x bt,(l) ≤ b∗l h bt,(l)∗ al − yl kbl k2 a∗l x bt,(l) . b∗l h 2 The bounds (189) and (214) taken collectively yield bt,(l) ≤ a∗l x\ + a∗l x bt,(l) − x\ a∗l x
. p log m + C3 In addition, the same argument as in obtaining (215) tells us that 1 log 3/2 m  p log m. b t,(l) − h\ ) . C4 √µ log2 m. b∗l (h m Combine the previous two bounds to obtain b t,(l) x b t,(l) (b b t,(l) − h\ )x\∗ al bt,(l)∗ al − yl ≤ b∗l h b∗l h xt,(l) − x\ )∗ al + b∗l (h 92 (216) t,(l) b t,(l) − h\ ) · a∗ x\ b t,(l) · a∗ (b − x\ ) + b∗l (h ≤ b∗l h l l x   ∗ b t,(l) \ ∗ \ ∗ t,(l) \ b t,(l) − h\ ) · a∗ x\ ≤ bl (h · al (b − h ) + bl h x − x ) + b∗l (h l   log2 m µ 1 log2 m p log5/2 m . C4 µ √ +√ · C3 3/2 + C4 µ √ · log m . C4 µ √ . m m m m log m Substitution into (216) gives  b t,(l) x bt,(l)∗ al b∗l h Similarly, we can also derive  − yl  bt,(l) bl a∗l x 2 log5/2 m . C4 µ √ · m r K p · log m. m (217)  b t,(l) b t,(l) x b t,(l) x b t,(l) ≤ b∗ h bt,(l)∗ al − yl kal k2 b∗l h bt,(l)∗ al − yl al b∗l h b∗l h l . C4 µ log5/2 m √ µ √ · K · C4 √ log2 m m m Putting these bounds together indicates that µ kν3 k2 . (C4 ) √ m 2 s µ2 K log9 m . m The above bounds taken together with (212) and (213) ensure the existence of a constant C > 0 such that   s  t+1    t 2 K log9 m 
α α 1 µ η µ 2 zbt,(l) − zet 2 + C (C4 ) η √ dist z t+1,(l) , zet+1 ≤ max 1− , t+1 + CC1 η 2   αt α 16 m m log m   s 2 K log9 m  (i) 1 − η/21  µ η  t,(l) µ 2 ≤ 1− zb − zet 2 + C (C4 ) η √  1 − η/20  20 m m s  µ µ2 K log9 m η  t,(l) 2 zb − zet 2 + 2C (C4 ) η √ ≤ 1− 21 m m s 
η µ µ2 K log9 m 2 = 1− dist z t,(l) , zet + 2C (C4 ) η √ 21 m m s (ii) µ µ2 K log9 m ≤ C2 √ . m m Here, (i) holds as long as m is sufficiently large such that CC1 1/log2 m  1 and  t+1  α αt 1 − η/21 max , < , αt αt+1 1 − η/20 (218) which is guaranteed by Lemma 16. The inequality (ii) arises from the induction hypothesis (90b) and taking C2 > 0 is sufficiently large. e t+1 , x et+1 ) and Finally we establish the second inequality claimed in the lemma. Take (h1 , x1 ) = (h t+1,(l) t+1,(l) b b (h2 , x2 ) = (h ,x ) in Lemma 55. Since both (h1 , x1 ) and (h2 , x2 ) are close enough to (h\ , x\ ), we deduce that s µ µ2 K log9 m zet+1,(l) − zet+1 2 . zbt+1,(l) − zet+1 2 . C2 √ m m as claimed. 93 C.4 Proof of Lemma 18 Before going forward, we make note of the following inequality max b∗l 1≤l≤m 1 αt+1 ht+1 ≤ αt αt+1 1 t+1 1 h ≤ (1 + δ) max b∗l ht+1 t 1≤l≤m α αt max b∗l 1≤l≤m for some small δ  log−2 m, where the last relation follows from Lemma 16 that 1 αt+1 −1 . ≤δ αt log2 m for m sufficiently large. In view of the above inequality, the focus of our subsequent analysis will be to control maxl b∗l α1t ht+1 . The gradient update rule for ht+1 (cf. (79a)) gives m X
1 t+1 e t e tx et∗ − h\ x\∗ aj a∗j x et , h = h − ηξ bj b∗j h t α j=1 e t = 1 ht and x et = αt xt . Here and below, we denote ξ = 1/ke xt k22 for notational convenience. The where h αt above formula can be further decomposed into the following terms m X 1 t+1 e t t e t a∗ x h = h − ηξ bj b∗j h je αt j=1  = 1 − ηξ x\ − ηξ where we use the fact that | Pm j=1 m X j=1 2 2  e t − ηξ h | e t a∗ x\ bj b∗j h j 2 + ηξ j=1 m X j=1 2 m X et bj b∗j h\ x\∗ aj a∗j x t e t a∗ x bj b∗j h je {z 2 − a∗j x\ 2
} :=v1 − x\ {z 2
+ ηξ 2 } :=v2 | m X j=1 et , bj b∗j h\ x\∗ aj a∗j x {z } :=v3 bj b∗j = IK . In the sequel, we shall control each term separately. 1. We start with |b∗l v1 | by making the observation that m h X
∗ t
∗

i 1 ∗ t \ ∗ \ ∗ t \ ∗ e t a∗ x e e e b∗l bj b∗j h |bl v1 | = − x a x + a x a x − x j j j j ηξ j=1 ≤ m X j=1 |b∗l bj |  et max b∗j h 1≤j≤m  e t − x\ max a∗j x 1≤j≤m
Combining the induction hypothesis (90c) and the condition (189) yields et ≤ max a∗j x e t − x\ max a∗j x 1≤j≤m 1≤j≤m
+ max a∗j x\ ≤ C3 1≤j≤m as long as m is sufficiently large. This further implies et − x\ max a∗j x 1≤j≤m
et + a∗j x\ a∗j x
≤ C3 1 log 3/2 m 1 log3/2 m et + a∗j x\ a∗j x +5
 . p p log m ≤ 6 log m p · 11 log m ≤ 11C3 1 . log m Substituting it into (219) and taking Lemma 48, we arrive at   1 ∗ e t · C3 1 . C3 max b∗ h e t ≤ 0.1 max b∗ h et , |bl v1 | . log m · max b∗j h j j 1≤j≤m 1≤j≤m 1≤j≤m ηξ log m with the proviso that C3 is sufficiently small. 94 (219) 2. We then move on to |b∗l v3 |, which obeys m m X X
1 ∗ ∗ ∗ \ \∗ ∗ \ et − x\ . b∗l bj b∗j h\ x\∗ aj a∗j x bl bj bj h x aj aj x + |b v3 | ≤ ηξ l j=1 j=1 (220) Regarding the first term, we have the following lemma, whose proof is given in Appendix C.4.1. 2 Lemma 28. Suppose
m ≥ CK log m for some sufficiently large constant C > 0. Then with probability −10 at least 1 − O m , one has m X µ b∗l bj b∗j h\ x\∗ aj a∗j x\ − b∗l h\ . √ . m j=1 For the remaining term, we apply the same strategy as in bounding |b∗l v1 | to get m X b∗l bj b∗j h\ x\∗ aj a∗j t e −x x j=1 \
≤ m X j=1 |b∗l bj |  max 1≤j≤m b∗j h\  max 1≤j≤m a∗j p µ 1 ≤ 4 log m · √ · C3 3/2 · 5 log m m log m µ . C3 √ , m t e −x x \
 max 1≤j≤m a∗j x\  where the second line follows from the incoherence (36), the induction hypothesis (90c), the condition (189) and Lemma 48. Combining the above three inequalities and the incoherence (36) yields µ 1 ∗ µ µ |b v3 | . b∗l h\ + √ + C3 √ . (1 + C3 ) √ . ηξ l m m m 3. Finally, we need to control |b∗l v2 |. For convenience of presentation, we will only bound |b∗1 v2 | in the sequel, but the argument easily extends to all other bl ’s. The idea is to group {bj }1≤j≤m into bins each containing τ adjacent vectors, and to look at each bin separately. Here, τ  poly log(m) is some integer to be specified later. For notational simplicity, we assume m/τ to be an integer, although all arguments continue to hold when m/τ is not an integer. For each 0 ≤ l ≤ m − τ , the following summation over τ adjacent data obeys b∗1 τ X j=1 = b∗1 et bl+j b∗l+j h τ X j=1 = a∗l+j x\ et bl+1 b∗l+1 h  τ  X   a∗l+j x\ j=1 + b∗1 τ X j=1 2  2 − x\ a∗l+j x\ − x\ 2    2 2  2 2  − x\ 2 2  + b∗1 τ X j=1 e t + b∗ b∗1 bl+1 b∗l+1 h 1 ∗ et bl+1 (bl+j − bl+1 ) h  a∗l+j x\ 2
t ∗ e bl+j b∗l+j − bl+1 b∗l+1 h al+j x\ τ X j=1 − x\ et (bl+j − bl+1 ) b∗l+j h 2 2  .  a∗l+j x\ 2 2 − x\ − x\ 2 2 2 2   (221) We will now bound each term in (221) separately.
Pτ ∗ \ 2 \ 2 • Before bounding the first term in (221), we first bound the pre-factor j=1 |al+j x | − kx k2 . Notably, the fluctuation of this quantity does not grow fast as it is the sum of i.i.d. random variables 95 2 over a group of relatively large size, i.e. τ . Since 2 a∗j x\ follows the χ22 distribution, by standard
concentration results (e.g. [RV+ 13, Theorem 1.1]), with probability exceeding 1 − O m−10 , τ X a∗l+j x\ 2 j=1 − kx\ k22
. p τ log m. With this result in place, we can bound the first term in (221) as   τ  X p
2 e t . τ log m |b∗ bl+1 | max b∗ h et . b∗1 bl+1 b∗l+1 h a∗l+j x\ − kx\ k22 l 1   1≤l≤m j=1 Taking the summation over all bins gives   m m τ τ −1 X τ −1  X X p
2 ∗ ∗ t \ 2 ∗ \ e et . b bkτ +1 bkτ +1 h . τ log m akτ +j x − kx k2 |b∗1 bkτ +1 | max b∗l h  1  1≤l≤m k=0 j=1 (222) k=0 It is straightforward to see from the proof of Lemma 48 that m τ −1 X k=0 |b∗1 bkτ +1 | = kb1 k22 + m τ −1 X k=1 |b∗1 bkτ +1 | ≤ K +O m  log m τ  (223) . Substituting (223) into the previous inequality (222) gives     s m √ τ τ −1 X  3 X
2 e t .  K τ log m + log m  max b∗ h et a∗kτ +j x\ − kx\ k22 b∗ b b∗ h l   1 kτ +1 kτ +1 1≤l≤m m τ k=0 j=1 et , ≤ 0.1 max b∗l h 1≤l≤m √ 3 as long as m  K τ log m and τ  log m. • The second term of (221) obeys b∗1 τ X j=1 et (bl+j − bl+1 ) b∗l+j h  a∗l+j x\ 2 − x\ 2 2  v v uX uX τ τ  u 2u et t |b∗1 (bl+j − bl+1 )| t a∗l+j x\ ≤ max b∗l h 1≤l≤m j=1 2 j=1 2 − kx\ k2 v uX u τ √ 2 et t . τ max b∗l h |b∗1 (bl+j − bl+1 )| , 1≤l≤m 2 j=1 where the first inequality is due to Cauchy-Schwarz, and the second one holds because of the following lemma, whose proof can be found in Appendix C.4.2. Lemma 29. Suppose
τ ≥ C log4 m for some sufficiently large constant C > 0. Then with probability exceeding 1 − O m−10 , τ   X 2 2 2 a∗j x\ − x\ 2 . τ. j=1 With the above bound in mind, we can sum over all bins of size τ to obtain b∗1 m τ −1 τ XX k=0 j=1 et (bkτ +j − bkτ +1 ) b∗kτ +j h 96 n a∗l+j x\ 2 − x\ 2 2 o   v m uX τ −1 √ X  u τ 2 et t . τ |b∗1 (bkτ +j − bkτ +1 )| max b∗ h   1≤l≤m l j=1 k=0 ≤ 0.1 max 1≤l≤m et b∗l h . Here, the last line arises from Lemma 51, which says that for any small constant c > 0, as long as m  τ K log m v m uX τ −1 X u τ 1 2 t |b∗1 (bkτ +j − bkτ +1 )| ≤ c √ . τ j=1 k=0 • The third term of (221) obeys b∗1 τ X j=1 ∗ et bl+1 (bl+j − bl+1 ) h ≤ |b∗1 bl+1 | .τ  τ X  |b∗1 bl+1 | 2 a∗l+j x\ j=1 max 0≤l≤m−τ, 1≤j≤τ where the last line relies on the inequality τ X n a∗l+j x \ 2 j=1 − 2 x\ 2 2 a∗l+j x\ − x\ 2 − x\ 2   2 o ∗ et max (b − bl+1 ) h  0≤l≤m−τ, 1≤j≤τ l+j 2 ∗ et (bl+j − bl+1 ) h , v u τ  √ uX ≤ τt a∗l+j x\ 2 j=1 − 2 kx\ k2 2 .τ owing to Lemma 29 and the Cauchy-Schwarz inequality. Summing over all bins gives m τ −1 X b∗1 k=0 .τ τ X j=1 m τ −1 X k=0 . log m ∗ et bkτ +1 (bkτ +j − bkτ +1 ) h |b∗1 bkτ +1 | max 0≤l≤m−τ, 1≤j≤τ max 0≤l≤m−τ, 1≤j≤τ n a∗kτ +j x\ 2 − x\ 2 2 o ∗ et (bl+j − bl+1 ) h ∗ et (bl+j − bl+1 ) h , where the last relation makes use of (223) with the proviso that m  Kτ . It then boils down to bounding ∗ et et . Without loss of generality, it suffices to look at (bj − b1 )∗ h max0≤l≤m−τ, 1≤j≤τ (bl+j − bl+1 ) h for all 1 ≤ j ≤ τ . Specifically, we claim for the moment that µ ∗ et max (bj − b1 ) h ≤ cC4 √ log m m (224) 1≤j≤τ for some sufficiently small constant c > 0, provided that m  τ K log4 m. As a result, m τ −1 X k=0 b∗1 τ X j=1 ∗ et bkτ +1 (bkτ +j − bkτ +1 ) h n a∗kτ +j x\ 2 − x\ 2 2 o µ . cC4 √ log2 m. m • Putting the above results together, we get m τ τ −1 n X X 1 ∗ \ e t a∗ |b1 v2 | ≤ b∗1 bkτ +j b∗kτ +j h kτ +j x ηξ j=1 2 k=0 97 − 2 x\ 2 o ≤ 0.2 max 1≤l≤m et b∗l h   µ 2 +O cC4 √ log m . m 4. Combining the preceding bounds guarantees the existence of some constant C8 > 0 such that   e t + C8 (1 + C3 )ηξ √µ + C8 ηξcC4 √µ log2 m e t + 0.3ηξ max b∗ h e t+1 ≤ (1 + δ) (1 − ηξ) b∗ h b∗l h l l 1≤l≤m m m      (i) µ 1 µ µ 2 2 √ √ √ (1 − 0.7ηξ) C4 ≤ 1+O log m + C8 (1 + C3 )ηξ + C8 ηξcC4 log m m m m log2 m (ii) µ ≤ C4 √ log2 m. m Here, (i) uses the induction hypothesis (90d), and (ii) holds as long as c > 0 is sufficiently small (so that (1 + δ)C8 ηξc  1) and η > 0 is some sufficiently small constant. In order for the proof to go through, it suffices to pick τ = c10 log4 m for some sufficiently large constant c10 > 0. Accordingly, we need the sample size to exceed m  µ2 τ K log4 m  µ2 K log8 m. Finally, it remains to verify the claim (224), which we accomplish in Appendix C.4.3. C.4.1 Proof of Lemma 28 Denote wj = b∗l bj b∗j h\ x\∗ aj a∗j x\ . P m Recognizing that E[aj a∗j ] = IK and j=1 bj b∗j = IK , we can write the quantity of interest as the sum of independent random variables, namely, m X j=1 b∗l bj b∗j h\ x\∗ aj a∗j x\ − b∗l h\ = m X j=1 (wj − E [wj ]) . Further, the sub-exponential norm (see definition in [Ver12]) of wj − E [wj ] obeys (i) (ii) kwj − E [wj ]kψ1 ≤ 2 kwj kψ1 ≤ 4 |b∗l bj | b∗j h\ 2 a∗j x\ ψ 2 (iii) . |b∗l bj | √ µ (iv) µ K √ ≤ , m m where (i) arises from the centering property of the sub-exponential norm (see [Ver12, Remark 5.18]), (ii) utilizes the relationship between the sub-exponential norm and the sub-Gaussian norm [Ver12, Lemma 5.14] and (iii) is a consequence of the incoherence condition (36) and the fact that a∗j x\ ψ . 1, and (iv) follows 2 p from kbj k2 = K/m. Let M = maxj∈[m] kwj − E [wj ]kψ1 and V2 = m X j=1 which follows since Pm 2 kwj − E [wj ]kψ1 . 2 ∗ ∗ j=1 |bl bj | = bl P m  X j=1  m ∗ j=1 bj bj Pm 2 1, j=1 aj µ |b∗l bj | √ m 2 = µ2 µ2 K 2 kbl k2 = , m m2 bl = kbl k22 = K/m. Let aj = kwj − E [wj ]kψ1 and Xj = (wj − E[wj ])/aj . Since kXj kψ1 = = V 2 and maxj∈[m] |aj | = M , we can invoke [Ver12, Proposition 5.16] to obtain that      m X t t2   P , , aj Xj ≥ t ≤ 2 exp −c min M V2 j=1 √ where c > 0 is some universal constant. By taking t = µ/ m, we see there exists some constant c0 such that     √  m X µ µ/ m µ2 /m ∗ ∗ \ \∗ ∗ \ ∗ \   P bl bj bj h x aj aj x − bl h ≥ √ ≤ 2 exp −c min , M V2 m j=1 98    √ µ2 /m µ/ m √ ≤ 2 exp −c0 min , 2 µ K/m µ K/m2  o np = 2 exp −c0 min m/K, m/K . We conclude the proof by observing that m  K log2 m as stated in the assumption. C.4.2 Proof of Lemma 29
2 From the elementary inequality (a − b) ≤ 2 a2 + b2 , we see that τ  X a∗j x\ 2 j=1  2 2 − x\ 2 ≤2 τ  X a∗j x\ 4 + x\ j=1 4 2  =2 τ X a∗j x\ 4 (225) + 2τ, j=1 Pτ 4 where the last identity holds true since x\ 2 = 1. It thus suffices to control j=1 a∗j x\ . Let ξi = a∗j x\ , which is a standard complex Gaussian random variable. Since the ξi ’s are statistically independent, one has ! τ X |ξi |4 ≤ C4 τ Var i=1 for some constant C4 > 0. It then follows from the hypercontractivity concentration result for Gaussian polynomials that [SS12, Theorem 1.9] ( τ ) 1/4 !  X  4 
c2 τ 2 4 Pτ P |ξi | − E |ξi | ≥ cτ ≤ C exp −c2 Var ( i=1 |ξi |4 ) i=1 !  2 2 1/4 !  2 1/4 c τ c ≤ C exp −c2 = C exp −c2 τ 1/4 C4 τ C4 ≤ O(m−10 ), for some constants c, c2 , C > 0, with the proviso that τ  log4 m. As a consequence, with probability at least 1 − O(m−10 ), τ τ h i X X 4 4 a∗j x\ . τ + E a∗j x\  τ, j=1 j=1 which together with (225) concludes the proof. C.4.3 Proof of Claim (224) We will prove the claim by induction. Again, observe that αt−1 ∗ et ∗ 1 t h = (bj − b1 ) h = (bj − b1 ) αt αt ∗ (bj − b1 ) 1 αt−1 ht ≤ (1 + δ) (bj − b1 ) ∗ ∗ 1 for some δ  log−2 m, which allows us to look at (bj − b1 ) t−1 ht instead. α Use the gradient update rule for ht (cf. (79a)) once again to get 1 αt−1 t 1 t−1 m X bl b∗l t−1 t−1∗ \ \∗ h x −h x 2 kxt−1 k2 l=1 m   X e t−1 − ηθ e t−1 x et−1∗ − h\ x\∗ al a∗l x et−1 , =h bl b∗l h h = αt−1 h − η l=1 99
al a∗l xt−1 ! 1 αt−1 ht 2 . 2 et−1 where we denote θ := 1/ x ∗ (bj − b1 ) 1 αt−1 This further gives rise to ∗ e t−1 ∗ ht = (bj − b1 ) h − ηθ (bj − b1 ) m X l=1   e t−1 x et−1 et−1∗ − h\ x\∗ al a∗l x bl b∗l h m   X ∗ e t−1 ∗ e t−1 x et−1 et−1∗ − h\ x\∗ x = (bj − b1 ) h − ηθ (bj − b1 ) bl b∗l h l=1 m   X ∗ e t−1 x et−1 et−1∗ − h\ x\∗ (al a∗l − IK ) x − ηθ (bj − b1 ) bl b∗l h = 1− ηθke xt−1 k22

∗ e t−1 ∗ et−1 (bj − b1 ) h + ηθ (bj − b1 ) h\ x\∗ x | {z } ∗ − ηθ (bj − b1 ) | l=1 :=β1 m X l=1  e t−1 x et−1∗ − h\ x\∗ (al a∗l − IK ) x et−1 , bl b∗l h  {z :=β2 where the last identity makes use of the fact that Pm l=1 bl b∗l = IK . For β1 , one can get } µ 1 ∗ xt−1 k2 ≤ 4 √ , |β1 | ≤ (bj − b1 ) h\ kx\ k2 ke ηθ m et−1 and x\ are extremely close, i.e. where we utilize the incoherence condition (36) and the fact that x
et−1 − x\ 2 ≤ dist z t−1 , z \  1 x =⇒ ke xt−1 k2 ≤ 2. Regarding the second term β2 , we have ) (m   X 1 ∗ e t−1 x et−1∗ − h\ x\∗ (al a∗l − IK ) x et−1 . (bj − b1 ) bl max b∗l h |β2 | ≤ 1≤l≤m ηθ l=1 | {z } :=ψ The term ψ can be bounded as follows e t−1 x et−1∗ (al a∗l − I) x et−1 + max b∗l h\ x\∗ (al a∗l − IK ) x et−1 ψ ≤ max b∗l h 1≤l≤m 1≤l≤m e t−1 max x et−1∗ (al a∗l − IK ) x et−1 + max b∗l h\ max x\∗ (al a∗l − IK ) x et−1 ≤ max b∗l h 1≤l≤m 1≤l≤m 1≤l≤m 1≤l≤m   µ ∗ e t−1 √ . . log m max bl h + 1≤l≤m m Here, we have used the incoherence condition (36) and the facts that et−1 ≤ a∗l x et−1 (e xt−1 )∗ (al a∗l − I) x x \∗ (al a∗l t−1 e − I) x ≤ 2 2 et−1 2 a∗l x et−1 + x a∗l x\ 2 2 2 . log m, et−1 + x 2 kx\ k2 . log m, which are immediate consequences of (90c) and (189). Combining this with Lemma 50, we see that for any small constant c > 0   1 µ 1 ∗ e t−1 √ |β2 | ≤ c max b h + ηθ log m 1≤l≤m l m holds as long as m  τ K log4 m. To summarize, we arrive at  ∗ et et−1 (bj − b1 ) h ≤ (1 + δ) 1 − ηθ x 2 2  µ 1 ∗ e t−1 (bj − b1 ) h + 4ηθ √ + cηθ log m m 100  e t−1 + √µ max b∗l h 1≤l≤m m  . et−1 Making use of the induction hypothesis (85c) and the fact that x 2 2 ≥ 0.9, we reach   µ cµηθ ∗ e t−1 t e + cC4 ηθ √ log m + √ (bj − b1 ) h ≤ (1 + δ) (1 − 0.9ηθ) (bj − b1 ) h . m m log m ∗ Recall that δ  1/ log2 m. As a result, if η > 0 is some sufficiently small constant and if   µ µ µ ∗ e t−1 ≤ 10c C4 √ log m + √ (bj − b1 ) h ≤ 20cC4 √ log m m ηθ m log m m holds, then one has µ ∗ et ≤ 20cC4 √ log m. (bj − b1 ) h m Therefore, this concludes the proof of the claim (224) by induction, provided that the base case is true, i.e. for some c > 0 sufficiently small µ ∗ e0 (bj − b1 ) h ≤ 20cC4 √ log m. m (226) The claim (226) is proved in Appendix C.6 (see Lemma 30). C.5 Proof of Lemma 19 Recall that ȟ0 and x̌0 are the leading left and right singular vectors of M , respectively. Applying a variant of Wedin’s sinΘ theorem [Dop00, Theorem 2.1], we derive that min α∈C,|α|=1  αȟ0 − h\ 2 + αx̌0 − x\ 2 ≤ c1 kM − E [M ]k , σ1 (E [M ]) − σ2 (M ) (227) for some universal constant c1 > 0. Regarding the numerator of (227), it has been shown in [LLSW16, Lemma 5.20] that for any ξ > 0, kM − E [M ]k ≤ ξ (228) with probability exceeding 1 − O(m−10 ), provided that m≥ c2 µ2 K log2 m ξ2 for some universal constant c2 > 0. For the denominator of (227), we can take (228) together with Weyl’s inequality to demonstrate that σ1 (E [M ]) − σ2 (M ) ≥ σ1 (E [M ]) − σ2 (E [M ]) − kM − E [M ]k ≥ 1 − ξ, where the last inequality utilizes the facts that σ1 (E [M ]) = 1 and σ2 (E[M ]) = 0. These together with (227) reveal that min α∈C,|α|=1  αȟ0 − h\ 2 + αx̌0 − x\ 2 ≤ c1 ξ ≤ 2c1 ξ 1−ξ (229) as long as ξ ≤ 1/2. p Now we connect the preceding bound (229) with the scaled singular vectors h0 = σ1 (M ) ȟ0 and p x0 = σ1 (M ) x̌0 . For any α ∈ C with |α| = 1, from the definition of h0 and x0 we have αh0 − h\ + αx0 − x\ 2 2 = p
σ1 (M ) αȟ0 − h\ 101 2 + p
σ1 (M ) αx̌0 − x\ 2 . Since αȟ0 , αx̌0 are also the leading left and right singular vectors of M , we can invoke Lemma 60 to get p
2 |σ1 (M ) − σ1 (E [M ])| p αh0 − h\ 2 + αx0 − x\ 2 ≤ σ1 (E[M ]) αȟ0 − h\ 2 + αx̌0 − x\ 2 + p σ1 (M ) + σ1 (E [M ]) 2 |σ1 (M ) − σ1 (E [M ])| p . (230) = αȟ0 − h\ 2 + αx̌0 − x\ 2 + σ1 (M ) + 1 In addition, we can apply Weyl’s inequality once again to deduce that (231) |σ1 (M ) − σ1 (E[M ])| ≤ kM − E[M ]k ≤ ξ, where the last inequality comes from (228). Substitute (231) into (230) to obtain αh0 − h\ 2 + αx0 − x\ 2 ≤ αȟ0 − h\ 2 + αx̌0 − x\ Taking the minimum over α, one can thus conclude that   αȟ0 − h\ αh0 − h\ 2 + αx0 − x\ 2 ≤ min min α∈C,|α|=1 α∈C,|α|=1 2 2 (232) + 2ξ. + αx̌0 − x\ 2 + 2ξ ≤ 2c1 ξ + 2ξ, where the last inequality comes from (229). Since ξ is arbitrary, by taking m/(µ2 K log2 m) to be large enough, we finish the proof for (92). Carrying out similar arguments (which we omit here), we can also establish (93). The last claim in Lemma 19 that ||α0 | − 1| ≤ 1/4 is a direct corollary of (92) and Lemma 52. C.6 Proof of Lemma 20 The proof is composed of three steps: • In the first step, we show that the normalized singular vectors of M and M (l) are close enough; see (240). • We then proceed by passing this proximity result to the scaled singular vectors; see (243). • Finally, we translate the usual `2 distance metric to the distance function we defined in (34); see (245). Along the way, we also prove the incoherence of h0 with respect to {bl }. Here comes the formal proof. Recall that ȟ0 and x̌0 are respectively the leading left and right singular vectors of M , and ȟ0,(l) and x̌0,(l) are respectively the leading left and right singular vectors of M (l) . Invoke Wedin’s sinΘ theorem [Dop00, Theorem 2.1] to obtain

n o M − M (l) x̌0,(l) 2 + ȟ0,(l)∗ M − M (l) 2 0 0,(l) 0 0,(l)
min αȟ − ȟ + αx̌ − x̌ ≤ c1 2 2 α∈C,|α|=1 σ1 M (l) − σ2 (M ) for some universal constant c1 > 0. Using the Weyl’s inequality we get

σ1 M (l) − σ2 (M ) ≥ σ1 E[M (l) ] − kM (l) − E[M (l) ]k − σ2 (E[M ]) − kM − E[M ]k ≥ 3/4 − kM (l) − E[M (l) ]k − kM − E[M ]k ≥ 1/2, where the penultimate inequality follows from
σ1 E[M (l) ] ≥ 3/4 for m sufficiently large, and the last inequality comes from [LLSW16, Lemma 5.20], provided that m ≥ c2 µ2 K log2 m for some sufficiently large constant c2 > 0. As a result, denoting n o β 0,(l) := argmin αȟ0 − ȟ0,(l) 2 + αx̌0 − x̌0,(l) 2 (233) α∈C,|α|=1 allows us to obtain β 0,(l) ȟ0 − ȟ0,(l) + β 0,(l) x̌0 − x̌0,(l) 2 2 ≤ 2c1 n
M − M (l) x̌0,(l) + ȟ0,(l)∗ M − M (l) 2
2 o . (234) It then boils down to controlling the two terms on the right-hand side of (234). By construction, M − M (l) = bl b∗l h\ x\∗ al a∗l . 102 • To bound the first term, observe that
M − M (l) x̌0,(l) = bl b∗l h\ x\∗ al a∗l x̌0,(l) = kbl k2 b∗l h\ a∗l x\ · a∗l x̌0,(l) 2 s 2 µ K log m ≤ 30 √ · , (235) m m p where we use the fact that kbl k2 = K/m, the
incoherence condition (36), the bound (189) and the fact that with probability exceeding 1 − O m−10 , p max a∗l x̌0,(l) ≤ 5 log m, 2 1≤l≤m due to the independence between x̌0,(l) and al . • To bound the second term, for any α e obeying |e α| = 1 one has ȟ0,(l)∗ M − M (l)
2 = ȟ0,(l)∗ bl b∗l h\ x\∗ al a∗l 2 = kal k2 b∗l h\ a∗l x\ · b∗l ȟ0,(l) p √ µ ≤ 3 K · √ · 5 log m · b∗l ȟ0,(l) m r r 2 (ii)
µ K log m µ2 K log m ∗ ∗ 0 ≤ 15 α ebl ȟ + 15 bl α eȟ0 − ȟ0,(l) m m r r r 2 2 (iii) µ K log m ∗ 0 µ K log m K bl ȟ + 15 ≤ 15 · α eȟ0 − ȟ0,(l) m m m (i) 2 . Here, (i) arises from the incoherence condition (36) together with the bounds (189) p and (190), the inequality (ii) comes from the triangle inequality, and the last line (iii) holds since kbl k2 = K/m and |e α| = 1. Substitution of the above bounds into (234) yields β 0,(l) ȟ0 − ȟ0,(l) + β 0,(l) x̌0 − x̌0,(l) 2 2  s r r r  2 2 K log m µ K log m ∗ 0 µ2 K log m K µ + 15 bl ȟ + 15 · α eȟ0 − ȟ0,(l) ≤ 2c1 30 √ ·  m m m m m   2 . Since the previous inequality holds for all |e α| = 1, we can choose α e = β 0,(l) and rearrange terms to get r r !  µ2 K log m K  0,(l) 0 β ȟ − ȟ0,(l) + β 0,(l) x̌0 − x̌0,(l) 1 − 30c1 m m 2 2 s r µ K log2 m µ2 K log m ∗ 0 ≤ 60c1 √ · + 30c1 bl ȟ . m m m p p Under the condition that m  µK log1/2 m, one has 1 − 30c1 µ2 K log m/m · K/m ≥ 12 , and therefore s r K log2 m µ2 K log m ∗ 0 µ 0,(l) 0 0,(l) 0,(l) 0 0,(l) β ȟ − ȟ + 60c1 + β x̌ − x̌ ≤ 120c1 √ · bl ȟ , m m m 2 2 which immediately implies that o n max β 0,(l) ȟ0 − ȟ0,(l) + β 0,(l) x̌0 − x̌0,(l) 1≤l≤m 2 2 s r µ K log2 m µ2 K log m ≤ 120c1 √ · + 60c1 max b∗l ȟ0 . 1≤l≤m m m m 103 (236) We then move on to b∗l ȟ0 . The aim is to show that max1≤l≤m b∗l ȟ0 can also be upper bounded by the left-hand side of (236). By construction, we have M x̌0 = σ1 (M ) ȟ0 , which further leads to b∗l ȟ0 = 1 b∗ M x̌0 σ1 (M ) l (i) ≤2 m X (b∗l bj ) b∗j h\ x\∗ aj a∗j x̌0 j=1   m X  ∗ \ ≤ 2 |b∗l bj | max bj h a∗j x\ a∗j x̌0 j=1 1≤j≤m  n µ  p ≤ 8 log m · √ · 5 log m max a∗j x̌0,(j) + kaj k2 β 0,(j) x̌0 − x̌0,(j) 1≤j≤m m s 2 µ log m µ2 K log3 m + 120 max β 0,(j) x̌0 − x̌0,(j) , ≤ 200 √ 1≤j≤m m m 2 (ii) 2 o (237) where β 0,(j) is as defined in (233). Here, (i) comes from the lower bound σ1 (M ) ≥ 1/2. The bound (ii) follows by the incoherence condition (36), the bound (189), the triangle inequality, as well as the estimate √ Pmcombining ∗ ∗ 0,(j) ≤ 5 log m j=1 |bl bj | ≤ 4 log m from Lemma 48. The last line uses the upper estimate max1≤j≤m aj x̌ and (190). Our bound (237) further implies s 2 µ log m µ2 K log3 m max b∗l ȟ0 ≤ 200 √ (238) + 120 max β 0,(j) x̌0 − x̌0,(j) . 1≤j≤m 1≤l≤m m m 2 The above bound (238) taken together with (236) gives n o s K log2 m + β 0,(l) x̌0 − x̌0,(l) 1≤l≤m m 2 2  s r µ log2 m µ2 K log m  µ2 K log3 m + 60c1 200 √ + 120 max β 0,(j) x̌0 − x̌0,(j) 1≤j≤m m m m max β 0,(l) ȟ0 − ȟ0,(l) As long as m  µ2 K log2 m we have 60c1 we are left with max 1≤l≤m n β 0,(l) ȟ0 − ȟ0,(l) µ ≤ 120c1 √ · m 2  . (239) q p µ2 K log m/m · 120 µ2 K log3 m/m ≤ 1/2. Rearranging terms, 2 + β 0,(l) x̌0 − x̌0,(l) 2 o µ ≤ c3 √ m s µ2 K log5 m m (240) for some constant c3 > 0. Further, this bound combined with (238) yields s s 2 3 2 µ log m µ K log m µ µ2 K log5 m µ log2 m max b∗l ȟ0 ≤ 200 √ + 120 · c3 √ ≤ c2 √ 1≤l≤m m m m m m (241) for some constant c2 > 0, with the proviso that m  µ2 K log2 m. We now translate the preceding bounds to the scaled version. Recall from the bound (231) that (242) 1/2 ≤ 1 − ξ ≤ kM k = σ1 (M ) ≤ 1 + ξ ≤ 2, as long as ξ ≤ 1/2. For any α ∈ C with |α| = 1, αȟ0 , αx̌0 are still the leading left and right singular vectors of M . Hence, we can use Lemma 60 to derive that n o


σ1 M − σ1 M (l) ≤ M − M (l) x̌0,(l) + αȟ0 − ȟ0,(l) + αx̌0 − x̌0,(l) kM k 2 104 2 2 
M − M (l) x̌0,(l) ≤ and 2 +2 + αx0 − x0,(l) 2 p
q
0 = σ1 (M ) αȟ − σ1 M (l) ȟ0,(l) αh0 − h0,(l) ≤ p σ1 (M ) 2 n αȟ0 − ȟ0,(l) √ n ≤ 2 αȟ0 − ȟ0,(l) 2 2 2 + αx0 − x0,(l) + αx̌0 − x̌0,(l) 2 ≤ √ 2 which together with (235) and (240) implies min α∈C,|α|=1 n αh0 − h0,(l) + 2 + αx̌0 − x̌0,(l) Taking the previous two bounds collectively yields αh0 − h0,(l) n 2 o + αȟ0 − ȟ0,(l) 2 o q p
σ1 (M )αx̌0 − σ1 M (l) x̌0,(l) 2 o √
M − M (l) x̌0,(l) 2 2 + αx̌0 − x̌0,(l) + αx0 − x0,(l) 2 2 σ1 (M ) − σ1 (M (l) ) p +p σ1 (M ) + σ1 (M (l) ) 2 σ1 (M ) − σ1 (M (l) ) . 2 2 +6 o n αȟ0 − ȟ0,(l) µ ≤ c5 √ m s 2 + αx̌0 − x̌0,(l) µ2 K log5 m m 2 o , (243) for some constant c5 > 0, as long as ξ is sufficiently small. Moreover, we have n 1 0 α 0,(l) h − h + α0 x0 − αα0 x0,(l) ≤ 2 h0 − αh0,(l) + x0 − αx0,(l) 2 2 2 α0 α0 for any |α| = 1, where α0 is defined in (38) and, according to Lemma 19, satisfies 1/2 ≤ |α0 | ≤ 2. 2 o (244) Therefore, r 2 α 0,(l) 2 h0 − h + α0 x0 − αα0 x0,(l) 0 α∈C,|α|=1 2 2 α  α 0,(l) 1 0 + α0 x0 − αα0 x0,(l) ≤ min h − h α∈C,|α|=1 α0 α0 2 n o ≤ 2 min h0 − αh0,(l) + x0 − αx0,(l) α∈C,|α|=1 2 2 s µ µ2 K log5 m ≤ 2c5 √ . m m min 1 α0 2  Furthermore, we have dist z 0,(l) 0 , ze
r 1 0,(l) 1 0 2 2 h − h 2 + αx0,(l) − α0 x0 2 0 α∈C α α r 1 0 α 0,(l) 2 ≤ min h − h + α0 x0 − αα0 x0,(l) 0 α∈C,|α|=1 2 α α0 s µ2 K log5 m µ , ≤2c5 √ m m = min 2 2 (245) where the second line follows since the latter is minimizing over a smaller feasible set. This completes the proof for the claim (96). 105 e 0 , one first sees that Regarding b∗l h p b∗l h0 = σ1 (M )b∗l ȟ0 ≤ √ 2c2 µ log2 m √ , m where the last relation holds due to (241) and (242). Hence, using the property (244), we have e 0 = b∗ b∗l h l 1 α0 h0 ≤ 1 α0 √ µ log2 m b∗l h0 ≤ 2 2c2 √ , m which finishes the proof of the claim (97). Before concluding this section, we note a byproduct of the proof. Specifically, we can establish the claim required in (226) using many results derived in this section. This is formally stated in the following lemma. Lemma 30. Fix any small constant
c > 0. Suppose the number of samples obeys m  τ K log4 m. Then with probability at least 1 − O m−10 , we have µ ∗ e0 ≤ c √ log m. max (bj − b1 ) h m 1≤j≤τ Proof. Instate the notation and hypotheses in Appendix C.6. Recognize that p ∗ 1 ∗ 1 ∗ e0 h0 = (bj − b1 ) σ1 (M )ȟ0 (bj − b1 ) h = (bj − b1 ) 0 0 α α 1 p ∗ σ1 (M ) (bj − b1 ) ȟ0 ≤ α0 ∗ ≤ 4 (bj − b1 ) ȟ0 , ∗ where the√ last inequality comes from (242) and (244). It thus suffices to prove that (bj − b1 ) ȟ0 ≤ cµ log m/ m for some c > 0 small enough. To this end, it can be seen that ∗ 1 ∗ (bj − b1 ) M x̌0 σ1 (M ) m X ∗ ≤2 (bj − b1 ) bk b∗k h\ x\∗ ak a∗k x̌0 (bj − b1 ) ȟ0 = ≤2 (i) ≤c (ii) k=1 m X k=1 ∗ (bj − b1 ) bk ! max 1≤k≤m  b∗k h\ a∗k x\ a∗k x̌0 n  1 µ  p √ · · 5 log m max a∗j x̌0,(j) + kaj k2 α0,(j) x̌0 − x̌0,(j) 1≤j≤m m log2 m µ 1 µ . c√ ≤ c √ log m, m log m m 2 o (246) where (i) comes from Lemma 50, the incoherence condition (36), and the estimate (189). The last line (ii) holds since we have already established (see (237) and (240)) n o p max a∗j x̌0,(j) + kaj k2 α0,(j) x̌0 − x̌0,(j) . log m. 1≤j≤m 2 The proof is then complete. C.7 Proof of Lemma 21 Recall that α0 and α0,(l) are the alignment parameters between z 0 and z \ , and between z 0,(l) and z \ , respectively, that is,   1 0 2 h − h\ 22 + αx0 − x\ 2 , α0 := argmin α α∈C 106 α0,(l) := argmin α∈C Also, we let 0,(l) αmutual := argmin α∈C   1 0,(l) h − h\ α 2 2 1 0,(l) 1 0 h h − α α0 + αx0,(l) − x\ 2 2 2 2  . + αx0,(l) − α0 x0 2 2  . The triangle inequality together with (94) and (245) then tells us that s 2 1 2 0,(l) h0,(l) − h\ + αmutual x0,(l) − x\ 2 0,(l) 2 αmutual s s 2 2 2 1 0 1 0 1 2 0,(l) ≤ h − h0,(l) + α0 x0 − αmutual x0,(l) + h − h\ + kα0 x0 − x\ k2 0 0,(l) 2 2 α0 α 2 αmutual s µ µ2 K log5 m 1 ≤ 2c5 √ + C1 2 m m log m 1 ≤ 2C1 2 , log m √ where the last relation holds as long as m  µ2 K log9/2 m. Let 1 1 0,(l) x1 = α0 x0 , h1 = h0 and x2 = αmutual x0,(l) , h2 = h0,(l) . 0 0,(l) α αmutual It is easy to see that x1 , h1 , x2 , h2 satisfy the assumptions in Lemma 55, which implies s s 2 2 1 1 0 2 1 0 1 2 0,(l) 0,(l) 0,(l) 0,(l) 0 0 h − h h − h0,(l) + α0 x0 − αmutual x0,(l) + α x −α x 2 . 0 0 0,(l) 0,(l) 2 2 2 α α α αmutual s µ2 K log5 m µ .√ , (247) m m where the last line comes from (245). With this upper estimate at hand, we are now ready to show that with high probability, a∗l α0 x0 − x\
(i) ≤ a∗l α0,(l) x0,(l) − x\
+ a∗l α0 x0 − α0,(l) x0,(l) p ≤ 5 log m α0,(l) x0,(l) − x\ (ii)
+ kal k2 α0 x0 − α0,(l) x0,(l) s (iii) p √ 1 µ µ2 K log5 m √ . log m · + K m m log2 m (iv) . 1 log 3/2 m 2 2 , where (i) follows from the triangle inequality, (ii) uses Cauchy-Schwarz and the independence between x0,(l) and al , (iii) holds because of (95) and (247) under the condition m  µ2 K log6 m, and (iv) holds true as long as m  µ2 K log4 m. 107 D D.1 D.1.1 Technical lemmas Technical lemmas for phase retrieval Matrix concentration inequalities i.i.d. Lemma 31. Suppose that aj ∼ N (0, In ) for every 1 ≤ j ≤ m. Fix any small constant δ > 0. With probability at least 1 − C2 e−c2 m , one has m 1 X aj a> j − In ≤ δ, m j=1 as long as m ≥ c0 n for some sufficiently large constant c0 > 0. Here, C2 , c2 > 0 are some universal constants. Proof. This is an immediate consequence of [Ver12, Corollary 5.35]. i.i.d. Lemma 32. Suppose that aj ∼ N (0, In ), for every 1 ≤ j ≤ m. Fix any small constant δ > 0. With probability at least 1 − O(n−10 ), we have m
1 X > \
2 \ 2 \ \> a x aj a> ≤ δkx\ k22 , j − kx k2 In + 2x x m j=1 j provided that m ≥ c0 n log n for some sufficiently large constant c0 > 0. Proof. This is adapted from [CLS15, Lemma 7.4]. i.i.d. Lemma 33. Suppose that aj ∼ N (0, In ), for every 1 ≤ j ≤ m. Fix any small constant δ > 0 and any constant C > 0. Suppose m ≥ c0 n for some sufficiently large constant c0 > 0. Then with probability at least 1 − C2 e−c2 m , m
1 X >
2 > > 2 aj x 1{|a> ≤ δkxk22 , x|≤C} aj aj − β1 xx + β2 kxk2 In j m j=1 holds for some absolute constants c2 , C2 > 0, where     β1 := E ξ 4 1{|ξ|≤C} − E ξ 2 1|ξ|≤C and with ξ being a standard Gaussian random variable. ∀x ∈ Rn   β2 = E ξ 2 1|ξ|≤C Proof. This is supplied in [CC17, supplementary material]. D.1.2 Matrix perturbation bounds Lemma 34. Let λ1 (A), u be the leading eigenvalue and eigenvector of a symmetric matrix A, respectively, e u e respectively. Suppose that e be the leading eigenvalue and eigenvector of a symmetric matrix A, and λ1 (A), e e λ1 (A), λ1 (A), kAk, kAk ∈ [C1 , C2 ] for some C1 , C2 > 0. Then, p λ1 (A) u − Proof. Observe that q e u e λ1 (A) q p e u e λ1 (A) u − λ1 (A) 2 ≤ 2 p ≤
e u A−A √ 2 C1 λ1 (A) u − q 108 2 p  C2 e k2 . + ku − u C2 + √ C1 e u λ1 (A) + 2 q q e u − λ1 (A) e u e λ1 (A) 2 p ≤ λ1 (A) − q e + λ1 (A) q e ku − u e k2 , λ1 (A) where the last inequality follows since kuk2 = 1. Using the identity p λ1 (A) − q √ a− √ (248) √ √ b = (a − b)/( a + b), we have

e e λ1 A − λ1 (A) λ1 A − λ1 (A) e = √ ≤ λ1 (A) , q p 2 C1 e λ1 (A) + λ1 (A) e This combined with (248) yields where the last inequality comes from our assumptions on λ1 (A) and λ1 (A). q p e u e λ1 (A) u − λ1 (A) 2
e λ1 A − λ1 (A) p √ e k2 . ≤ + C2 ku − u 2 C1 (249)
e , use the relationship between the eigenvalue and the eigenvector to obtain To control λ1 A − λ1 (A) e = u> Au − u eu e > Ae λ1 (A) − λ1 (A)
e u + u> Au e −u e + u e −u eu e > Au e > Au e > Ae ≤ u> A − A
e , e u + 2 ku − u e k2 A ≤ A−A 2 which together with (249) gives q p e u e λ1 (A) u − λ1 (A) 2 ≤ ≤ as claimed. D.2 D.2.1
e u A−A e p e k2 A + 2 ku − u √ e k2 + C2 ku − u 2 C1
 e u p  A−A C2 2 √ e k2 + √ + C2 ku − u 2 C1 C1 2 Technical lemmas for matrix completion Orthogonal Procrustes problem The orthogonal Procrustes problem is a matrix approximation problem which seeks an orthogonal matrix R b to be the minimizer of to best “align” two matrices A and B. Specifically, for A, B ∈ Rn×r , define R minimizeR∈Or×r kAR − BkF . (250) b of (250). The first lemma is concerned with the characterization of the minimizer R b is the minimizer of (250) if and only if R b > A> B is symmetric and Lemma 35. For A, B ∈ Rn×r , R positive semidefinite. Proof. This is an immediate consequence of [tB77, Theorem 2]. Let A> B = U ΣV > be the singular value decomposition of A> B ∈ Rr×r . It is easy to check that b := U V > satisfies the conditions that R b > A> B is both symmetric and positive semidefinite. In view of R b = U V > is the minimizer of (250). In the special case when C := A> B is invertible, R b enjoys Lemma 35, R the following equivalent form:
b=H c (C) := C C > C −1/2 , R (251) c (·) is an Rr×r -valued function on Rr×r . This motivates us to look at the perturbation bounds for where H c (·), which is formulated in the following lemma. the matrix-valued function H 109 Lemma 36. Let C ∈ Rr×r be a nonsingular matrix. Then for any matrix E ∈ Rr×r with kEk ≤ σmin (C) and any unitarily invariant norm |||·|||, one has c (·) is defined above. where H c (C + E) − H c (C) H 2 |||E|||, σr−1 (C) + σr (C) ≤ Proof. This is an immediate consequence of [Mat93, Theorem 2.3]. With Lemma 36 in place, we are ready to present the following bounds on two matrices after “aligning” them with X \ . Lemma 37. Instate the notation in Section 3.2. Suppose X1 , X2 ∈ Rn×r are two matrices such that X1 − X \ X \ ≤ σmin /2, kX1 − X2 k X \ (252a) (252b) ≤ σmin /4. Denote R1 := argmin X1 R − X \ R2 := argmin X2 R − X \ and F R∈O r×r R∈O r×r F . Then the following two inequalities hold true: kX1 R1 − X2 R2 k ≤ 5κ kX1 − X2 k kX1 R1 − X2 R2 kF ≤ 5κ kX1 − X2 kF . and Proof. Before proving the claims, we first gather some immediate consequences of the assumptions (252). > Denote C = X1> X \ and E = (X2 − X1 ) X \ . It is easily seen that C is invertible since C − X \> X \ ≤ X1 − X \ (i) X \ ≤ σmin /2 (ii) =⇒ σr (C) ≥ σmin /2, (253) where (i) follows from the assumption (252a) and (ii) is a direct application of Weyl’s inequality. In addition, C + E = X2> X \ is also invertible since (i) (ii) kEk ≤ kX1 − X2 k X \ ≤ σmin /4 < σr (C) , where (i) arises from the assumption (252b) and (ii) holds because of (253). When both C and C + E are invertible, the orthonormal matrices R1 and R2 admit closed-form expressions as follows R1 = C C > C
−1/2 and Moreover, we have the following bound on kX1 k: (i) kX1 k ≤ X1 − X \ + X \ (ii) ≤ h i−1/2 > R2 = (C + E) (C + E) (C + E) . σmin σmax + X\ ≤ + X\ 2 kX \ k 2 kX \ k (iii) ≤ 2 X\ , (254) where (i) is the triangle inequality, (ii) uses the assumption (252a) and (iii) arises from the fact that X \ = √ σmax . With these in place, we turn to establishing the claimed bounds. We will focus on the upper bound on kX1 R1 − X2 R2 kF , as the bound on kX1 R1 − X2 R2 k can be easily obtained using the same argument. Simple algebra reveals that kX1 R1 − X2 R2 kF = k(X1 − X2 ) R2 + X1 (R1 − R2 )kF ≤ kX1 − X2 kF + kX1 k kR1 − R2 kF ≤ kX1 − X2 kF + 2 X \ kR1 − R2 kF , 110 (255) where the first inequality uses the fact that kR2 k = 1 and the last inequality comes from (254). An application of Lemma 36 leads us to conclude that 2 kEkF σr (C) + σr−1 (C) 2 > (X2 − X1 ) X \ ≤ σmin F 2 \ ≤ kX2 − X1 kF X , σmin kR1 − R2 kF ≤ (256) (257) where (256) utilizes (253). Combine (255) and (257) to reach kX1 R1 − X2 R2 kF ≤ kX1 − X2 kF + 4 kX2 − X1 kF X \ 2 σmin ≤ (1 + 4κ) kX1 − X2 kF , which finishes the proof by noting that κ ≥ 1. D.2.2 Matrix concentration inequalities This section collects various measure concentration results regarding the Bernoulli random variables {δj,k }1≤j,k≤n , which is ubiquitous in the analysis for matrix completion. Lemma 38. Fix any
small constant δ > 0, and suppose that m  δ −2 µnr log n. Then with probability exceeding 1 − O n−10 , one has 1 (1 − δ)kBkF ≤ √ kPΩ (B)kF ≤ (1 + δ)kBkF p holds simultaneously for all B ∈ Rn×n lying within the tangent space of M \ . Proof. This result has been established in [CR09, Section 4.2] for asymmetric sampling patterns (where each (i, j), i 6= j is included in Ω independently). It is straightforward to extend the proof and the result to symmetric sampling patterns (where each (i, j), i ≥ j is included in Ω independently). We omit the proof for conciseness. 2 Lemma 39. Fix a matrix M ∈ Rn×n
. Suppose n p ≥ c0 n log n for some sufficiently large constant c0 > 0. −10 With probability at least 1 − O n , one has 1 PΩ (M ) − M ≤ C p r n kM k∞ , p where C > 0 is some absolute constant. Proof. See [KMO10a, Lemma 3.2]. Similar to Lemma 38, the result therein was provided for the asymmetric sampling patterns but can be easily extended to the symmetric case. Lemma 40. Recall from Section 3.2 that E ∈ Rn×n is the symmetric noise matrix. Suppose the sample size
obeys n2 p ≥ c0 n log2 n for some sufficiently large constant c0 > 0. With probability at least 1 − O n−10 , one has r 1 n PΩ (E) ≤ Cσ , p p where C > 0 is some universal constant. Proof. See [CW15, Lemma 11]. 111 Lemma 41. Fix some matrix A ∈ Rn×r with n ≥ 2r and some 1 ≤ l ≤ n. Suppose {δl,j }1≤j≤n are independent Bernoulli random variables with means {pj }1≤j≤n no more than p. Define   > > r×n Gl (A) := δl,1 A> . 1,· , δl,2 A2,· , · · · , δl,n An,· ∈ R Then one has Median [kGl (A)k] ≤ s q 2 p kAk + 2 2 2 2p kAk2,∞ kAk log (4r) + 2 kAk2,∞ 3 log (4r) and for any constant C ≥ 3, with probability exceeding 1 − n−(1.5C−1) n X j=1 (δl,j − p)A> j,· Aj,· ≤ C and kGl (A)k ≤ s 2 p kAk + C q  2 2 2 p kAk2,∞ kAk log n + kAk2,∞ log n , q  2 2 2 p kAk2,∞ kAk log n + kAk2,∞ log n . Proof. By the definition of Gl (A) and the triangle inequality, one has > 2 kGl (A)k = Gl (A) Gl (A) = n X j=1 δl,j A> j,· Aj,· ≤ n X j=1 Therefore, it suffices to control the first term. It can be seen that zero-mean random matrices. Letting  (δl,j − pj ) A> j,· Aj,· 1≤j≤n are i.i.d. 2 (δl,j − pj ) A> j,· Aj,· ≤ kAk2,∞ L := max 1≤j≤n and V := 2 (δl,j − pj ) A> j,· Aj,· + p kAk . n n h i h i X X 2 2 2 2 2 > E (δl,j − pj ) A> A A A ≤ E (δ − p ) kAk A> l,j j j,· j,· j,· j,· j,· Aj,· ≤ p kAk2,∞ kAk 2,∞ j=1 j=1 and invoking matrix Bernstein’s inequality [Tro15b, Theorem 6.1.1], one has for all t ≥ 0,   ! n   X −t2 /2 > (δl,j − pj ) Aj,· Aj,· ≥ t ≤ 2r · exp P . 2 2 2   p kAk2,∞ kAk + kAk2,∞ · t/3 (258) j=1 We can thus find an upper bound on Median h P n j=1 (δl,j − pj ) A> j,· Aj,· i by finding a value t that ensures the right-hand side of (258) is smaller than 1/2. Using this strategy and some simple calculations, we get   2 n q X 2 kAk2,∞ 2 2  Median  (δl,j − pj ) A> A ≤ kAk 2p kAk log (4r) + log (4r) j,· j,· 2,∞ 3 j=1 and for any C ≥ 3, n X j=1 (δl,j − p j ) A> j,· Aj,·  q 2 2 2 p kAk2,∞ kAk log n + kAk2,∞ log n ≤C holds with probability at least 1 − n−(1.5C−1) . As a consequence, we have s 2 q 2 kAk2,∞ 2 2 2 Median [kGl (A)k] ≤ p kAk + 2p kAk2,∞ kAk log (4r) + log (4r), 3 112 and with probability exceeding 1 − n−(1.5C−1) , q  2 2 2 2 2 kGl (A)k ≤ p kAk + C p kAk2,∞ kAk log n + kAk2,∞ log n . This completes the proof. Lemma 42. Let {δl,j }1≤l≤j≤n be i.i.d. Bernoulli random variables with mean p and δl,j = δj,l . For any ∆ ∈ Rn×r , define   > > r×n Gl (∆) := δl,1 ∆> . 1,· , δl,2 ∆2,· , · · · , δl,n ∆n,· ∈ R Suppose the sample size obeys n2 p  κµrn log2 n. Then for any k > 0 and α > 0 large enough, with probability at least 1 − c1 e−αCnr log n/2 , n X l=1 1{kGl (∆)k≥4√pψ+2√krξ} ≤ holds simultaneously for all ∆ ∈ Rn×r obeying s log n X\ k∆k2,∞ ≤ C5 ρt µr np and 2,∞ 2αn log n k + C8 σ 1 k∆k ≤ C9 ρt µr √ X \ + C10 σ np r s n log n X\ p 2,∞ := ξ n X \ := ψ, p where c1 , C5 , C8 , C9 , C10 > 0 are some absolute constants. Proof. For simplicity of presentation, we will prove the claim for the asymmetric case where {δl,j }1≤l,j≤n are independent. The results immediately carry over to the symmetric case as claimed in this lemma. To see this, note that we can always divide Gl (∆) into Gl (∆) = Gupper (∆) + Glower (∆), l l where all nonzero components of Gupper (∆) come from the upper triangular part (those blocks with l ≤ j l ), while all nonzero components of Glower (∆) are from the lower triangular part (those blocks with l > j). l We can then look at {Gupper (∆) | 1 ≤ l ≤ n} and {Gupper (∆) | 1 ≤ l ≤ n} separately using the argument l l we develop for the asymmetric case. From now on, we assume that {δl,j }1≤l,j≤n are independent. e ∈ Rn×r , Suppose for the moment that ∆ is statistically independent of {δl,j }. Clearly, for any ∆, ∆ e Gl (∆) − Gl (∆)

e ≤ Gl (∆) − Gl ∆ e ≤ Gl (∆) − Gl ∆ v uX 2 u n e j,· ≤t ∆j,· − ∆ F 2 j=1
e , := d ∆, ∆ which implies that kGl (∆)k is 1-Lipschitz with respect to the metric d (·, ·). Moreover, max kδl,j ∆j,· k2 ≤ k∆k2,∞ ≤ ξ 1≤j≤n according to our assumption. Hence, Talagrand’s inequality [CC16, Proposition 1] reveals the existence of some absolute constants C, c > 0 such that for all λ > 0
P {kGl (∆)k − Median [kGl (∆)k] ≥ λξ} ≤ C exp −cλ2 . (259) We then proceed to control Median [kGl (∆)k]. A direct application of Lemma 41 yields r p 2ξ 2 √ Median [kGl (∆)k] ≤ 2pψ 2 + p log (4r)ξψ + log (4r) ≤ 2 pψ, 3 113 where the last relation holds since pψ 2  ξ 2 log r, which follows by combining the definitions of ψ and ξ, the sample size condition np  κµr log2 n, and the incoherence condition (114). Thus, substitution into (259) √ and taking λ = kr give n o √ √ P kGl (∆)k ≥ 2 pψ + krξ ≤ C exp (−ckr) (260) for any k ≥ 0. Furthermore, invoking [AS08, Corollary A.1.14] and using the bound (260), one has !   n X t log t P 1{kGl (∆)k≥2√pψ+√krξ} ≥ tnC exp (−ckr) ≤ 2 exp − nC exp (−ckr) 2 l=1 for any t ≥ 6. Choose t = α log n/ [kC exp (−ckr)] ≥ 6 to obtain !   n X αC αn log n √ √ ≤ 2 exp − P nr log n . 1{kGl (∆)k≥2 pψ+ krξ} ≥ k 2 (261) l=1 Pn So far we have demonstrated that for any fixed ∆ obeying our assumptions, l=1 1{kGl (∆)k≥2√pψ+√krξ} is well controlled with exponentially high probability. In order to extend the results to all feasible ∆, we resort to the standard -net argument. Clearly, due to the homogeneity property of kGl (∆)k, it suffices to restrict attention to the following set: where ψ/ξ . kX k/kX k2,∞ . \ \ √ n. We then proceed with the following steps. 1. Introduce the auxiliary function    1, χl (∆) = (262) S = {∆ | min {ξ, ψ} ≤ k∆k ≤ ψ} , √ kGl (∆)k−2 pψ− √ √ 2 pψ+ krξ  √  0, √ √ if kGl (∆)k ≥ 4 pψ + 2 krξ, √ √ √ √ krξ , if kGl (∆)k ∈ [2 pψ + krξ, 4 pψ + 2 krξ], else. Clearly, this function is sandwiched between two indicator functions 1{kGl (∆)k≥4√pψ+2√krξ} ≤ χl (∆) ≤ 1{kGl (∆)k≥2√pψ+√krξ} . Note that χl is more convenient to work with due to continuity. 2. Consider an -net N [Tao12, Section 2.3.1] of the set S as defined in (262). For any  = 1/nO(1) , one can find such a net with cardinality log |N | . nr log n. Apply the union bound and (261) to yield ! ! n n X X αn log n αn log n P χl (∆) ≥ , ∀∆ ∈ N ≤ P 1{kGl (∆)k≥2√pψ+√krξ} ≥ , ∀∆ ∈ N k k l=1 l=1     αC αC ≤ 2|N | exp − nr log n ≤ 2 exp − nr log n , 2 4 as long as α is chosen to be sufficiently large. 3. One can then use the continuity argument to extend the bound to all ∆ outside the -net, i.e. with exponentially high probability, n X 2αn log n χl (∆) ≤ , ∀∆ ∈ S k l=1 =⇒ n X l=1 1{kGl (∆)k≥4√pψ+2√krξ} ≤ n X l=1 χl (∆) ≤ 2αn log n , k This is fairly standard (see, e.g. [Tao12, Section 2.3.1]) and is thus omitted here. 114 ∀∆ ∈ S We have thus concluded the proof. 2 Lemma 43. Suppose the sample size obeys
n p ≥ Cκµrn log n for some sufficiently large constant C > 0. −10 Then with probability at least 1 − O n ,
1 PΩ XX > − X \ X \> ≤ 2n2 X \ p 2 2,∞ √ + 4 n log n X \ 2,∞ X\ holds simultaneously for all X ∈ Rn×r satisfying X − X\ 2,∞ ≤  X\ 2,∞ , (263) where  > 0 is any fixed constant. Proof. To simplify the notations hereafter, we denote ∆ := X − X \ . With this notation in place, one can decompose XX > − X \ X \> = ∆X \> + X \ ∆> + ∆∆> , which together with the triangle inequality implies that



1 1 1 1 PΩ XX > − X \ X \> ≤ PΩ ∆X \> + PΩ X \ ∆> + PΩ ∆∆> p p p p

1 1 = PΩ ∆∆> +2 PΩ ∆X \> . p p | {z } | {z } :=α1 (264) :=α2 In the sequel, we bound α1 and α2 separately. 1. Recall from [Mat90, Theorem 2.5] the elementary inequality that (265) kCk ≤ |C| , where |C| := [|ci,j |]1≤i,j≤n for any matrix C = [ci,j ]1≤i,j≤n . In addition, for any matrix D := [di,j ]1≤i,j≤n such that |di,j | ≥ |ci,j | for all i and j, one has |C| ≤ |D| . Therefore α1 ≤ 1 PΩ p ∆∆>
2 ≤ k∆k2,∞
1 PΩ 11> . p Lemma 39 then tells us that with probability at least 1 − O(n−10 ), r
1 n PΩ 11> − 11> ≤ C p p (266) for some universal constant C > 0, as long as p  log n/n. This together with the triangle inequality yields r

1 1 n PΩ 11> ≤ PΩ 11> − 11> + 11> ≤ C + n ≤ 2n, (267) p p p provided that p  1/n. Putting together the previous bounds, we arrive at 2 α1 ≤ 2n k∆k2,∞ . (268) 2. Regarding the second term α2 , apply the elementary inequality (265) once again to get

PΩ ∆X \> ≤ PΩ ∆X \> ,
which motivates us to look at PΩ ∆X
\> instead. A key step of this part is to take advantage of the `2,∞ norm constraint of PΩ ∆X \> . Specifically, we claim for the moment that with probability exceeding 1 − O(n−10 ),
2 2 (269) PΩ ∆X \> 2,∞ ≤ 2pσmax k∆k2,∞ := θ 115 holds under our sample size condition. In addition, we also have the following trivial `∞ norm bound
PΩ ∆X \> ∞ ≤ k∆k2,∞ X \ 2,∞ := γ. (270) In what follows, for simplicity of presentation, we will denote
A := PΩ ∆X \> . (271) (a) To facilitate the analysis of kAk, we first introduce k0 + 1 = 12 log (κµr) auxiliary matrices9 Bs ∈ Rn×n that satisfy kX 0 −1 kAk ≤ kBk0 k + kBs k . (272) s=0 To be precise, each Bs is defined such that ( 1 1 if Aj,k ∈ ( 2s+1 γ, 21s γ], s γ, [Bs ]j,k = 2 0, else, ( 1 if Aj,k ≤ 2k10 γ, k γ, [Bk0 ]j,k = 2 0 0, else, for 0 ≤ s ≤ k0 − 1 and which clearly satisfy (272); in words, Bs is constructed by rounding up those entries of A within a prescribed magnitude interval. Thus, it suffices to bound kBs k for every s. To this end, we start with s = k0 and use the definition of Bk0 to get q (ii) (iii) √ (i) 1 2 k∆k2,∞ X \ 2,∞ ≤ 4 np k∆k2,∞ X \ , kBk0 k ≤ kBk0 k∞ (2np) ≤ 4np √ κµr where (i) arises from Lemma 44, with 2np being a crude upper bound on the number of nonzero entries in each row and each column. This can be derived by applying the standard Chernoff bound on Ω. The second inequality (ii) relies on the definitions of γ and k0 . The last one (iii) follows from the incoherence condition (114). Besides, for any 0 ≤ s ≤ k0 − 1, by construction one has 2 2 kBs k2,∞ ≤ 4θ = 8pσmax k∆k2,∞ and kBs k∞ = 1 γ, 2s where θ is as defined in (269). Here, we have used the fact that the magnitude of each entry of Bs is at most 2 times that of A. An immediate implication is that there are at most 2 kBs k2,∞ 2 kBs k∞ 2 ≤ nonzero entries in each row of Bs and at most 8pσmax k∆k2,∞ := kr
2 1 2s γ kc = 2np nonzero entries in each column of Bs , where kc is derived from the standard Chernoff bound on Ω. Utilizing Lemma 44 once more, we discover that q p √ 1 p 2 kBs k ≤ kBs k∞ kr kc = s γ kr kc = 16np2 σmax k∆k2,∞ = 4 np k∆k2,∞ X \ 2 for each 0 ≤ s ≤ k0 − 1. Combining all, we arrive at kAk ≤ 9 For simplicity, we assume is not an integer. 1 2 kX 0 −1 s=0 √ kBs k + kBk0 k ≤ (k0 + 1) 4 np k∆k2,∞ X \ log (κµr) is an integer. The argument here can be easily adapted to the case when 116 1 2 log (κµr) √ ≤ 2 np log (κµr) k∆k2,∞ X \ √ ≤ 2 np log n k∆k2,∞ X \ , where the last relation holds under the condition n ≥ κµr. This further gives √ 1 α2 ≤ kAk ≤ 2 n log n k∆k2,∞ X \ . p (b) In order to finish the proof of this part, we need to justify the claim (269). Observe that 2 
 2 Xn  \> ∆l,· Xj,· δl,j PΩ ∆X \> l,· = j=1 2  Xn  \> \ = ∆l,· δl,j Xj,· Xj,· ∆> l,· j=1 Xn 2 \> \ δl,j Xj,· Xj,· ≤ k∆k2,∞ j=1 (273) (274) for every 1 ≤ l ≤ n, where δl,j indicates whether the entry with the index (l, j) is observed or not. Invoke Lemma 41 to yield h i 2 Xn \> \ \> \> \> δl,j Xj,· Xj,· = δl,1 X1,· , δl,2 X2,· , · · · , δl,n Xn,· j=1 q  2 2 2 ≤ pσmax + C p kX \ k2,∞ kX \ k log n + X \ 2,∞ log n ! r pκµr log n κµr log n ≤ p+C +C σmax n n (275) ≤ 2pσmax , with high probability, as soon as np  κµr log n. Combining (274) and (275) yields as claimed in (269).  PΩ ∆X \>
 2 l,· 2 2 ≤ 2pσmax k∆k2,∞ , 1≤l≤n 3. Taken together, the preceding bounds (264), (268) and (273) yield
√ 1 2 PΩ XX > − X \ X \> ≤ α1 + 2α2 ≤ 2n k∆k2,∞ + 4 n log n k∆k2,∞ X \ . p The proof is completed by substituting the assumption k∆k2,∞ ≤  X \ 2,∞ . In the end of this subsection, we record a useful lemma to bound the spectral norm of a sparse Bernoulli matrix. n ×n Lemma 44. Let A ∈ {0, 1} 1 2 be a binary matrix, and suppose that there √ are at most kr and kc nonzero entries in each row and column of A, respectively. Then one has kAk ≤ kc kr . Proof. This immediately follows from the elementary inequality kAk2 ≤ kAk1→1 kAk∞→∞ (see [Hig92, equation (1.11)]), where kAk1→1 and kAk∞→∞ are the induced 1-norm (or maximum absolute column sum norm) and the induced ∞-norm (or maximum absolute row sum norm), respectively. D.2.3 Matrix perturbation bounds Lemma 45. Let M ∈ Rn×n be a symmetric matrix with the top-r eigendecomposition U ΣU > . Assume M − M \ ≤ σmin /2 and denote b := argmin U R − U \ . Q F R∈O r×r Then there is some numerical constant c3 > 0 such that b − U \ ≤ c3 M − M \ . UQ σmin 117 Proof. Define Q = U > U \ . The triangle inequality gives
b − Q + U U >U \ − U \ . b − Q + UQ − U\ ≤ Q b − U\ ≤ U Q UQ (276) [AFWZ17, Lemma 3] asserts that b−Q ≤4 Q M − M \ /σmin
2 as long as M − M \ ≤ σmin /2. For the remaining term in (276), one can use U \> U \ = Ir to obtain U U > U \ − U \ = U U > U \ − U \ U \> U \ ≤ U U > − U \ U \> , which together with the Davis-Kahan sinΘ theorem [DK70] reveals that U U >U \ − U \ ≤ c2 M − M\ σmin b − Q , U U > U \ − U \ and (276) to reach for some constant c2 > 0. Combine the estimates on Q b − U\ ≤ UQ  4 σmin M −M \ 2 + c2 c3 M − M\ ≤ M − M\ σmin σmin for some numerical constant c3 > 0, where we have utilized the fact that M − M \ /σmin ≤ 1/2. f ∈ Rn×n be two symmetric matrices with top-r eigendecompositions U ΣU > and Lemma 46. Let M , M > e e e f − M \ ≤ σmin /4, and suppose σmax /σmin is U ΣU , respectively. Assume M − M \ ≤ σmin /4 and M bounded by some constant c1 > 0, with σmax and σmin the largest and the smallest singular values of M \ , respectively. If we denote e , Q := argmin U R − U F R∈O r×r then there exists some numerical constant c3 > 0 such that e 1/2 ≤ √ c3 f−M Σ1/2 Q − QΣ M σmin e 1/2 Σ1/2 Q − QΣ and F ≤√ c3 σmin
f−M U M F . Proof. Here, we focus on the Frobenius norm; the bound on the operator norm follows from the same argument, and hence we omit the proof. Since k·kF is unitarily invariant, we have e 1/2 Σ1/2 Q − QΣ F e 1/2 = Q> Σ1/2 Q − Σ F , e 1/2 are the matrix square roots of Q> ΣQ and Σ, e respectively. In view of the matrix where Q> Σ1/2 Q and Σ square root perturbation bound [Sch92, Lemma 2.1], e 1/2 Σ1/2 Q − QΣ F ≤  1/2 σmin (Σ)  1 > e   Q ΣQ − Σ e 1/2 + σmin (Σ) F ≤√ where the last inequality follows from the lower estimates
σmin (Σ) ≥ σmin Σ\ − kM − M \ k ≥ σmin /4 1 e Q> ΣQ − Σ σmin e ≥ σmin /4. Recognizing that Σ = U > M U and Σ e =U e >M fU e , one gets and, similarly, σmin (Σ)
>
e e >M fU e Q> ΣQ − Σ = UQ M UQ − U F F
>

>

>

f UQ f UQ − U e >M f UQ ≤ UQ M UQ − UQ M + UQ M F
>f >fe e e + U M UQ − U MU F 118 F F , (277) ≤
f−M U M F e + 2 UQ − U
f−M U M f ≤ M F where the last relation holds due to the upper estimate F e + 4σmax U Q − U F (278) , f ≤ M\ + M f − M \ ≤ σmax + σmin /4 ≤ 2σmax . M Invoke the Davis-Kahan sinΘ theorem [DK70] to obtain e UQ − U F ≤
f−M U M c2 f) σr (M ) − σr+1 (M F ≤ 2c2 σmin for some constant c2 > 0, where the last inequality follows from the bounds
σr (M ) ≥ σr M \ − kM − M \ k ≥ 3σmin /4,
f ) ≤ σr+1 M \ + kM f − M \ k ≤ σmin /4. σr+1 (M
f−M U M F (279) , Combine (277), (278), (279) and the fact σmax /σmin ≤ c1 to reach for some constant c3 > 0. e 1/2 Σ1/2 Q − QΣ F ≤√ c3 σmin
f−M U M F Lemma 47. Let M ∈ Rn×n be a symmetric matrix with the top-r eigendecomposition U ΣU > . Denote X = U Σ1/2 and X \ = U \ (Σ\ )1/2 , and define b := argmin U R − U \ Q R∈O r×r c := argmin XR − X \ H and F R∈O r×r F . Assume M − M \ ≤ σmin /2, and suppose σmax /σmin is bounded by some constant c1 > 0. Then there exists a numerical constant c3 > 0 such that b−H c ≤ Q c3 M − M\ . σmin Proof. We first collect several useful facts about the spectrum of Σ. Weyl’s inequality tells us that Σ − Σ\ ≤ M − M \ ≤ σmin /2, which further implies that
σr (Σ) ≥ σr Σ\ − Σ − Σ\ ≥ σmin /2 Denote kΣk ≤ Σ\ + Σ − Σ\ ≤ 2σmax . and Q = U >U \ H = X >X \. and Simple algebra yields H = Σ1/2 Q Σ\
1/2 b = Σ1/2 Q − Q |
Σ\
1/2 b − QΣ b 1/2 + Σ1/2 Q {z :=E
Σ\ It can be easily seen that σr−1 (A) ≥ σr (A) ≥ σmin /2, and b · kEk ≤ Σ1/2 · Q − Q Σ\
1/2 b − QΣ b 1/2 · + Σ1/2 Q b − QΣ b 1/2 , b +√σmax Σ1/2 Q ≤ 2σmax Q − Q | {z } | {z } :=α :=β which can be controlled as follows. 119
1/2 b ΣΣ\ +Q } | {z Σ\ :=A
1/2
1/2 . } • Regarding α, use [AFWZ17, Lemma 3] to reach b ≤ 4 M − M\ α= Q−Q • For β, one has (ii) (i) b > Σ1/2 Q b − Σ1/2 β= Q ≤ 2σr 1 b > ΣQ b−Σ
Q Σ1/2 2 2 /σmin . (iii) = 2σr 1 b − QΣ b
ΣQ , Σ1/2 where (i) and (iii) come from the unitary invariance of k·k, and (ii) follows from the matrix square root perturbation bound [Sch92, Lemma 2.1]. We can further take the triangle inequality to obtain b − Q) − (Q b − Q)Σ b − QΣ b = ΣQ − QΣ + Σ(Q ΣQ b ≤ kΣQ − QΣk + 2 kΣk Q − Q

b = U M − M \ U \> + Q Σ\ − Σ + 2 kΣk Q − Q

b ≤ U M − M \ U \> + Q Σ\ − Σ + 2 kΣk Q − Q ≤ 2 M − M \ + 4σmax α, where the last inequality uses the Weyl’s inequality kΣ\ −Σk ≤ kM −M \ k and the fact that kΣk ≤ 2σmax . • Rearrange the previous bounds to arrive at kEk ≤ 2σmax α + √ σmax √
1 2 M − M \ + 4σmax α ≤ c2 M − M \ σmin for some numerical constant c2 > 0, where we have used the assumption that σmax /σmin is bounded. b = sgn (A) (see definition in (177)), we are ready to invoke Lemma 36 to deduce that Recognizing that Q for some constant c3 > 0. D.3 D.3.1 b−H c ≤ Q 2 c3 kEk ≤ M − M\ σr−1 (A) + σr (A) σmin Technical lemmas for blind deconvolution Wirtinger calculus In this section, we formally prove the fundamental theorem of calculus and the mean-value form of Taylor’s theorem under the Wirtinger calculus; see (283) and (284), respectively. Let f : Cn → R be a real-valued function. Denote z = x + iy ∈ Cn , then f (·) can alternatively be viewed as a function R2n → R. There is a one-to-one mapping connecting the Wirtinger derivatives and the conventional derivatives [KD09]:     x z = J −1 , (280a) z y     x z ∇R f = J ∗ ∇C f , (280b) y z     x z J, (280c) ∇2R f = J ∗ ∇2C f y z where the subscripts R and C represent calculus in the real (conventional) sense and in the complex (Wirtinger) sense, respectively, and   In iIn J= . In −iIn 120 With these relationships in place, we are ready to verify the fundamental theorem of calculus using the Wirtinger derivatives. Recall from [Lan93, Chapter XIII, Theorem 4.2] that  ∇R f x1 y1   x2 y2  x (τ ) y (τ )   − ∇R f where  := = Z 1 ∇2R f 0 x2 y2  +τ Substitute the identities (280) into (281) to arrive at J ∗ ∇C f  z1 z1  − J ∗ ∇C f  z2 z2    x (τ ) y (τ ) x1 y1  −   dτ x2 y2    x1 y1  −  x2 y2  , (281) .      z1 z2 dτ J J −1 − z1 z2 0    Z 1     z (τ ) z2 z1 ∇2C f − , (282) = J∗ dτ z z2 z (τ ) 1 0 = J∗ Z 1 ∇2C f  z (τ ) z (τ )  where z1 = x1 + iy1 , z2 = x2 + iy2 and         z (τ ) z2 z1 z2 := +τ − . z2 z1 z2 z (τ ) Simplification of (282) gives ∇C f  z1 z1  − ∇C f  z2 z2  = Z 0 1 ∇2C f  z (τ ) z (τ )  dτ   z1 z1  −  z2 z2  . Repeating the above arguments, one can also show that    ∗   1 z1 − z2 z1 − z2 z1 − z2 ∗ 2 f (z1 ) − f (z2 ) = ∇C f (z2 ) + ∇C f (e z) , z1 − z2 z1 − z2 2 z1 − z2 (283) (284) where ze is some point lying on the vector connecting z1 and z2 . This is the mean-value form of Taylor’s theorem under the Wirtinger calculus. D.3.2 Discrete Fourier transform matrices Let B ∈ Cm×K be the first K columns of a discrete Fourier transform (DFT) matrix F ∈ Cm×m , and denote by bl the lth column of the matrix B ∗ . By definition, ∗ 1  bl = √ 1, ω (l−1) , ω 2(l−1) , · · · , ω (K−1)(l−1) , m 2π where ω := e−i m with i representing the imaginary unit. It is seen that for any j 6= l, b∗l bj K−1 K−1 K−1 1 X k(l−1) k(j−1) (i) 1 X k(l−1) k(1−j) 1 X l−j
k (ii) 1 1 − ω K(l−j) = ω ·ω ω ·ω = ω = . = m m m m 1 − ω l−j k=0 k=0 (285) k=0 Here, (i) uses ω α = ω −α for all α ∈ R, while the last identity (ii) follows from the formula for the sum of a finite geometric series when ω l−j 6= 1. This leads to the following lemma. Lemma 48. For any m ≥ 3 and any 1 ≤ l ≤ m, we have m X j=1 |b∗l bj | ≤ 4 log m. 121 Proof. We first make use of the identity (285) to obtain m X j=1 |b∗l bj | = 2 kbl k2   m m π K 1 X sin K (l − j) m 1 X 1 − ω K(l−j)   , = + + π m 1 − ω l−j m m sin (l − j) m j:j6=l j:j6=l 2 where the last identity follows since kbl k2 = K/m and, for all α ∈ R, 2π π π π |1 − ω α | = 1 − e−i m α = e−i m α ei m α − e−i m α
 π . = 2 sin α m (286) Without loss of generality, we focus on the case when l = 1 in the sequel. Recall that for c > 0, we denote by bcc the largest integer that does not exceed c. We can continue the derivation to get m X j=1 |b∗1 bj |   m m π (i) 1 X K 1 1 X sin K (1 − j) m K     + ≤ = + π π m m j=2 sin (1 − j) m m j=2 sin (j − 1) m m   b m2 c+1 m X 1 K 1 1  X      + = +  π π m m sin (j − 1) m sin (j − 1) m j=2 j=b m 2 c+2   bm m 2 c+1 X X 1 1 K (ii) 1    +    =  + , π π m m sin (j − 1) sin (m + 1 − j) m m j=2 j=b m 2 c+2
π where (i) follows from sin K (1 − j) m ≤ 1 and |sin (x)| = |sin (−x)|, and (ii) relies on the fact that sin (x) = sin (π − x). The property that sin (x) ≥ x/2 for any x ∈ [0, π/2] allows one to further derive     m+1 m m b b b m m 2 c+1 2 c 2 c−1 X X X 1 K 2 X 1 2m 1  X 2m  K + =  + |b∗1 bj | ≤  + + m (j − 1) π (m + 1 − j) π m π k k m m j=1 j=2 k=1 k=1 j=b 2 c+2 m (i) 4 X (iii) 1 K (ii) 4 ≤ + ≤ (1 + log m) + 1 ≤ 4 log m, π k m π k=1 where in (i) we extend the range of the summation, (ii) uses the elementary inequality and (iii) holds true as long as m ≥ 3. ∗ Pm k=1 k −1 ≤ 1 + log m The next lemma considers the difference of two inner products, namely, (bl − b1 ) bj . m , we have Lemma 49. For all 0 ≤ l − 1 ≤ τ ≤ 10 ∗ (bl − b1 ) bj ≤ ( 8τ /π 4τ K (j−l) m + (j−l)2 8τ /π K 4τ m−(j−l) m + [m−(j−1)]2 for for   l+τ ≤j ≤ m 2 + 1, m 2 + l ≤ j ≤ m − τ. In addition, for any j and l, the following uniform upper bound holds ∗ (bl − b1 ) bj ≤ 2 K . m Proof. Given (285), we can obtain for j 6= l and j 6= 1, ∗ (bl − b1 ) bj = = 1 1 − ω K(l−j) 1 − ω K(1−j) − m 1 − ω l−j 1 − ω 1−j 1 1 − ω K(l−j) 1 − ω K(1−j) 1 − ω K(1−j) 1 − ω K(1−j) − + − l−j l−j l−j m 1−ω 1−ω 1−ω 1 − ω 1−j 122 
1 − ω K(1−j) 1 ω K(1−j) − ω K(l−j) + ω l−j − ω 1−j l−j m 1−ω (1 − ω l−j ) (1 − ω 1−j )
1 1 − ω K(l−1) 2 1 ≤ + 1 − ω 1−l , m 1 − ω l−j m (1 − ω l−j ) (1 − ω 1−j ) = where the last line is due to the triangle inequality and |ω α | = 1 for all α ∈ R. The identity (286) allows us to rewrite this bound as (   ) π h sin (1 − l) m πi 1 1 ∗     . sin K (l − 1) + (bl − b1 ) bj ≤ (287) π π m sin (l − j) m m sin (1 − j) m Combined with the fact that |sin x| ≤ 2 |x| for all x ∈ R, we can upper bound (287) as ( ) π 2τ m 1 1 π ∗  2Kτ +  ,   (bl − b1 ) bj ≤ π π m sin (l − j) m m sin (1 − j) m where we also utilize the assumption 0 ≤ l − 1 ≤ τ . Then for l + τ ≤ j ≤ bm/2c + 1, one has (l − j) π π ≤ m 2 and (1 − j) π π ≤ . m 2 Therefore, utilizing the property sin (x) ≥ x/2 for any x ∈ [0, π/2], we arrive at   2 π 4τ 4τ K 8τ /π ∗ (bl − b1 ) bj ≤ 2Kτ + ≤ + , (j − l) π m j−1 (j − l) m (j − l)2 where the last inequality holds since j − 1 > j − l. Similarly we can obtain the upper bound for bm/2c + l ≤ j ≤ m − τ using nearly identical argument (which is omitted for brevity). The uniform upper bound can be justified as follows ∗ (bl − b1 ) bj ≤ (kbl k2 + kb1 k2 ) kbj k2 ≤ 2K/m. 2 The last relation holds since kbl k2 = K/m for all 1 ≤ l ≤ m. Next, we list two consequences of the above estimates in Lemma 50 and Lemma 51. Lemma 50. Fix any constant c > 0 that is independent of m and K. Suppose m ≥ Cτ K log4 m for some sufficiently large constant C > 0, which solely depends on c. If 0 ≤ l − 1 ≤ τ , then one has m X j=1 ∗ (bl − b1 ) bj ≤ c . log2 m Proof. For some constant c0 > 0, we can split the index set [m] into the following three disjoint sets n j m ko A1 = j : l + c0 τ log2 m ≤ j ≤ , 2 n jmk o A2 = j : + l ≤ j ≤ m − c0 τ log2 m , 2 and A3 = [m] \ (A1 ∪ A2 ) . With this decomposition in place, we can write m X j=1 ∗ (bl − b1 ) bj = X j∈A1 ∗ (bl − b1 ) bj + X j∈A2 ∗ (bl − b1 ) bj + We first look at A1 . By Lemma 49, one has for any j ∈ A1 , ∗ (bl − b1 ) bj ≤ 8τ /π 4τ K + , j − l m (j − l)2 123 X j∈A3 ∗ (bl − b1 ) bj . and hence X j∈A1 bm 2 c+1 X ∗ (bl − b1 ) bj ≤ 8τ /π 4τ K + j − l m (j − l)2 j=l+c0 τ log2 m ! ≤ m 4τ K X 1 8τ + m k π k=1 m X k=c0 τ log2 m 1 k2 K 16τ 1 , log m + m π c0 τ log2 m Pm Pm where the last inequality arises from k=1 k −1 ≤ 1 + log m ≤ 2 log m and k=c k −2 ≤ 2/c. Similarly, for j ∈ A2 , we have ≤ 8τ ∗ (bl − b1 ) bj ≤ 4τ K 8τ /π + , m − (j − l) m [m − (j − 1)]2 which in turn implies X j∈A2 ∗ (bl − b1 ) bj ≤ 8τ K 16τ 1 . log m + m π c0 τ log2 m Regarding j ∈ A3 , we observe that

|A3 | ≤ 2 c0 τ log2 m + l ≤ 2 c0 τ log2 m + τ + 1 ≤ 4c0 τ log2 m. ∗ This together with the simple bound (bl − b1 ) bj ≤ 2K/m gives X ∗ (bl − b1 ) bj ≤ 2 j∈A3 K 8c0 τ K log2 m |A3 | ≤ . m m The previous three estimates taken collectively yield m X j=1 ∗ (bl − b1 ) bj ≤ 1 1 16τ K log m 32τ 8c0 τ K log2 m + ≤c 2 + 2 m π c0 τ log m m log m as long as c0 ≥ (32/π) · (1/c) and m ≥ 8c0 τ K log4 m/c. Lemma 51. Fix any constant c > 0 that is independent of m and K. Consider an integer τ > 0, and suppose that m ≥ Cτ K log m for some large constant C > 0, which depends solely on c. Then we have v bm/τ c u τ X uX c 2 t |b∗1 (bkτ +j − bkτ +1 )| ≤ √ . τ j=1 k=0 Proof. The proof strategy is similar to the one used in Lemma 50. First notice that ∗ |b∗1 (bkτ +j − bkτ +1 )| = (bm − bm+1−j ) bkτ . As before, for some c1 > 0, we can split the index set {1, · · · , bm/τ c} into three disjoint sets  ko n jj m k + 1 − j /τ , B1 = k : c1 ≤ k ≤ 2 k n jj m k o B2 = k : + 1 − j /τ + 1 ≤ k ≤ b(m + 1 − j) /τ c − c1 , n j2m ko and B3 = 1, · · · , \ (B1 ∪ B2 ) , τ where 1 ≤ j ≤ τ . 124 By Lemma 49, one has ∗ (bm − bm+1−j ) bkτ ≤ Hence for any k ∈ B1 , v uX u τ √ 2 t |b∗1 (bkτ +j − bkτ +1 )| ≤ τ j=1 4τ K 8τ /π , + kτ m (kτ )2 4τ K 8τ /π + kτ m (kτ )2 ! k ∈ B1 . = √ τ  4K 8/π + 2 km k τ  , which further implies that v  τ m  Xu uX √ X √ K log m 16 1 1 8/π 4K 2 t √ |b∗1 (bkτ +j − bkτ +1 )| ≤ τ + 2 ≤8 τ + , km k τ m π τ c1 j=1 k∈B1 k=c1 Pm Pm where the last inequality follows since k=1 k −1 ≤ 2 log m and k=c1 k −2 ≤ 2/c1 . A similar bound can be obtained for k ∈ B2 . For the remaining set B3 , observe that |B3 | ≤ 2c1 . ∗ This together with the crude upper bound (bl − b1 ) bj ≤ 2K/m gives v √ r τ Xu uX √ 2K 4c1 τ K 2 2 ∗ ∗ t |b1 (bkτ +j − bkτ +1 )| ≤ |B3 | τ max |b1 (bkτ +j − bkτ +1 )| ≤ |B3 | τ · ≤ . j m m j=1 k∈B3 The previous estimates taken collectively yield v   √ bm/τ c u τ X uX √ K log m 16 1 1 1 4c1 τ K 2 ∗ t √ + ≤ c√ , |b1 (bkτ +j − bkτ +1 )| ≤ 2 8 τ + m π c m τ τ 1 j=1 k=0 as long as c1  1/c and m/(c1 τ K log m)  1/c. D.3.3 Complex-valued alignment Let gh,x (·) : C → R be a real-valued function defined as gh,x (α) := 1 h − h\ α 2 2 + αx − x\ 2 2 , which is the key function in the definition (34). Therefore, the alignment parameter of (h, x) to (h\ , x\ ) is the minimizer of gh,x (α). This section is devoted to studying various properties of gh,x (·). To begin with, the Wirtinger gradient and Hessian of gh,x (·) can be calculated as " #   ∂gh,x (α,α) 2 −2 2 −2 α kxk2 − x∗ x\ − α−1 (α) khk2 + (α) h\∗ h ∂α ∇gh,x (α) = ∂gh,x (α,α) = ; (288) 2 −1 2 α kxk2 − x\∗ x − (α) α−2 khk2 + α−2 h∗ h\ ∂α  2 −4 2 −3 2 −3 kxk2 + |α| khk2 2α−1 (α) khk2 − 2 (α) h\∗ h ∇ gh,x (α) = . (289) −1 2 2 −4 2 2 (α) α−3 khk2 − 2α−3 h∗ h\ kxk2 + |α| khk2   The first lemma reveals that, as long as β1 h, βx is sufficiently close to (h\ , x\ ), the minimizer of gh,x (α) cannot be far away from β. 2  125 Lemma 52. Assume theres exists β ∈ C with 1/2 ≤ |β| ≤ 3/2 such that max δ ≤ 1/4. Denote by α b the minimizer of gh,x (α), then we necessarily have n 1 h β − h\ 2 , βx − x\ 2 o ≤ α − β| ≤ 18δ. |b α| − |β| ≤ |b Proof. The first inequality is a direct consequence of the triangle inequality. Hence we concentrate on the second one. Notice that by assumption, gh,x (β) = 1 h − h\ β 2 2 + βx − x\ 2 2 ≤ 2δ 2 , (290) which immediately implies that gh,x (b α) ≤ 2δ 2 . It thus suffices to show that for any α obeying |α − β| > 18δ, 2 one has gh,x (α) > 2δ , and hence it cannot be the minimizer. To this end, we lower bound gh,x (α) as follows:
2 = (α − β) x + βx − x\ 2 h
∗ i 2 2 2 = |α − β| kxk2 + βx − x\ 2 + 2Re (α − β) βx − x\ x
∗ 2 2 ≥ |α − β| kxk2 − 2 |α − β| βx − x\ x . gh,x (α) ≥ αx − x\ Given that βx − x\ 2 2 2 ≤ δ ≤ 1/4 and x\ 2 = 1, we have kβxk2 ≥ kx\ k2 − βx − x\ 2 ≥ 1 − δ ≥ 3/4, which together with the fact that 1/2 ≤ |β| ≤ 3/2 implies kxk2 ≥ 1/2 and βx − x\
∗ and kxk2 ≤ 2 x ≤ βx − x\ Taking the previous estimates collectively yields gh,x (α) ≥ 2 kxk2 ≤ 2δ. 1 2 |α − β| − 4δ |α − β| . 4 It is self-evident that once |α − β| > 18δ, one gets gh,x (α) > 2δ 2 , and hence α cannot be the minimizer as gh,x (α) > gh,x (β) according to (290). This concludes the proof. The next lemma reveals the local strong convexity of gh,x (α) when α is close to one.  Lemma 53. Assume that max h − h\ 2 , x − x\ 2 ≤ δ for some sufficiently small constant δ > 0. Then, for any α satisfying |α − 1| ≤ 18δ and any u, v ∈ C, one has    1 2 u 2 ∗ ∗ 2 [u , v ] ∇ gh,x (α) ≥ |u| + |v| , v 2 where ∇2 gh,x (·) stands for the Wirtinger Hessian of gh,x (·). Proof. For simplicity of presentation, we use gh,x (α, α) and gh,x (α) interchangeably. By (289), for any u, v ∈ C , one has      h  i u 2 −4 2 2 2 −3 2 −3 ∗ ∗ 2 [u , v ] ∇ gh,x (α) = kxk2 + |α| khk2 |u| + |v| +2 Re u∗ v 2α−1 (α) khk2 − 2 (α) h\∗ h . v | {z } | {z } :=β1 :=β2 2 2 We would like to demonstrate that this is  at least on the order of |u| + |v| . We first develop a lower bound on β1 . Given the assumption that max h − h\ 2 , x − x\ 2 ≤ δ, one necessarily has 1 − δ ≤ kxk2 ≤ 1 + δ and 126 1 − δ ≤ khk2 ≤ 1 + δ. Thus, for any α obeying |α − 1| ≤ 18δ, one has     −4 2 −4 2 β1 ≥ 1 + |α| (1 − δ) ≥ 1 + (1 + 18δ) (1 − δ) ≥ 1 as long as δ > 0 is sufficiently small. Regarding the second term β2 , we utilizes the conditions |α − 1| ≤ 18δ, kxk2 ≤ 1 + δ and khk2 ≤ 1 + δ to get −3 2 α−1 khk2 − h\∗ h
−3 2 = 2 |u| |v| |α| α−1 − 1 khk2 − (h\ − h)∗ h   −3 2 ≤ 2 |u| |v| |α| α−1 − 1 khk2 + h − h\ 2 khk2   18δ −3 2 ≤ 2 |u| |v| (1 − 18δ) (1 + δ) + δ (1 + δ) 1 − 18δ
2 2 . δ |u| + |v| , |β2 | ≤ 2 |u| |v| |α| 2 2 where the last relation holds since 2 |u| |v| ≤ |u| + |v| and δ > 0 is sufficiently small. Combining the previous bounds on β1 and β2 , we arrive at     1  u 2 2 2 2 [u∗ , v ∗ ] ∇2 gh,x (α) ≥ (1 − O(δ)) |u| + |v| ≥ |u| + |v| v 2 as long as δ is sufficiently small. This completes the proof. Additionally, in a local region surrounding the optimizer, the alignment parameter is Lipschitz continuous, namely, the difference of the alignment parameters associated with two distinct vector pairs is at most proportional to the `2 distance between the two vector pairs involved, as demonstrated below. Lemma 54. Suppose that the vectors x1 , x2 , h1 , h2 ∈ CK satisfy  max x1 − x\ 2 , h1 − h\ 2 , x2 − x\ 2 , h2 − h\ 2 (291) ≤ δ ≤ 1/4 for some sufficiently small constant δ > 0. Denote by α1 and α2 the minimizers of gh1 ,x1 (α) and gh2 ,x2 (α), respectively. Then we have |α1 − α2 | . kx1 − x2 k2 + kh1 − h2 k2 . Proof. Since α1 minimizes gh1 ,x1 (α), the mean-value form of Taylor’s theorem (see Appendix D.3.1) gives gh1 ,x1 (α2 ) ≥ gh1 ,x1 (α1 ) = gh1 ,x1 (α2 ) + ∇gh1 ,x1 (α2 ) ∗  α1 − α2 α1 − α2  + 1 (α1 − α2 , α1 − α2 ) ∇2 gh1 ,x1 (e α) 2  α1 − α2 α1 − α2  , where α e is some complex number lying between α1 and α2 , and ∇gh1 ,x1 and ∇2 gh1 ,x1 are the Wirtinger gradient and Hessian of gh1 ,x1 (·), respectively. Rearrange the previous inequality to obtain |α1 − α2 | . k∇gh1 ,x1 (α2 )k2 λmin (∇2 gh1 ,x1 (e α)) (292)
as long as λmin ∇2 gh1 ,x1 (e α) > 0. This calls for evaluation of the Wirtinger gradient and Hessian of gh1 ,x1 (·). Regarding the Wirtinger Hessian, by the assumption (291), we can invoke Lemma 52 with β = 1 to reach max {|α1 − 1| , |α2 − 1|} ≤ 18δ. This together with Lemma 53 implies
λmin ∇2 gh1 ,x1 (e α) ≥ 1/2, since α e lies between α1 and α2 . 127 For the Wirtinger gradient, since α2 is the minimizer of gh2 ,x2 (α), the first-order optimality condition [KD09, equation (38)] requires ∇gh2 ,x2 (α2 ) = 0 , which gives k∇gh1 ,x1 (α2 )k2 = k∇gh1 ,x1 (α2 ) − ∇gh2 ,x2 (α2 )k2 . Plug in the gradient expression (288) to reach k∇gh1 ,x1 (α2 ) − ∇gh2 ,x2 (α2 )k2 i √ h 2 −2 2 −2 = 2 α2 kx1 k2 − x∗1 x\ − α2−1 (α2 ) kh1 k2 + (α2 ) h\∗ h1 h i 2 −2 2 −2 − α2 kx2 k2 − x∗2 x\ − α2−1 (α2 ) kh2 k2 + (α2 ) h\∗ h2 2 . |α2 | 2 kx1 k2 1 2 . |α2 | kx1 k2 − kx2 k2 + x∗1 x\ − x∗2 x\ + − 2 kx2 k2 + kx1 − x2 k2 + 1 2 3 |α2 | |α2 | 1 2 kh1 k2 − kh2 k2 + 2 3 kh1 k2 − 2 kh2 k2 + 2 1 |α2 | 2 |α2 | h\∗ h1 − h\∗ h2 kh1 − h2 k2 , where the last line follows from the triangle inequality. It is straightforward to see that 1/2 ≤ |α2 | ≤ 2, 2 2 kx1 k2 − kx2 k2 . kx1 − x2 k2 , 2 2 kh1 k2 − kh2 k2 . kh1 − h2 k2 under the condition (291) and the assumption kx\ k2 = kh\ k2 = 1, where the first inequality follows from Lemma 52. Taking these estimates together reveals that k∇gh1 ,x1 (α2 ) − ∇gh2 ,x2 (α2 )k2 . kx1 − x2 k2 + kh1 − h2 k2 . The proof is accomplished by substituting the two bounds on the gradient and the Hessian into (292). Further, if two vector pairs are both close to the optimizer, then their distance after alignement (w.r.t. the optimizer) cannot be much larger than their distance without alignment, as revealed by the following lemma. Lemma 55. Suppose that the vectors x1 , x2 , h1 , h2 ∈ CK satisfy  max x1 − x\ 2 , h1 − h\ 2 , x2 − x\ 2 , h2 − h\ 2 ≤ δ ≤ 1/4 (293) for some sufficiently small constant δ > 0. Denote by α1 and α2 the minimizers of gh1 ,x1 (α) and gh2 ,x2 (α), respectively. Then we have 2 kα1 x1 − α2 x2 k2 + 1 1 h1 − h2 α1 α2 2 2 2 2 . kx1 − x2 k2 + kh1 − h2 k2 . Proof. To start with, we control the magnitudes of α1 and α2 . Lemma 52 together with the assumption (293) guarantees that 1/2 ≤ |α1 | ≤ 2 and 1/2 ≤ |α2 | ≤ 2. Now we can prove the lemma. The triangle inequality gives kα1 x1 − α2 x2 k2 = kα1 (x1 − x2 ) + (α1 − α2 ) x2 k2 ≤ |α1 | kx1 − x2 k2 + |α1 − α2 | kx2 k2 (i) ≤ 2 kx1 − x2 k2 + 2 |α1 − α2 | (ii) . kx1 − x2 k2 + kh1 − h2 k2 , where (i) holds since |α1 | ≤ 2 and kx2 k2 ≤ 1 + δ ≤ 2, and (ii) arises from Lemma 54 that |α1 − α2 | . kx1 − x2 k2 + kh1 − h2 k2 . Similarly,   1 1 1 1 1 h1 − h2 = (h1 − h2 ) + − h2 α1 α2 α1 α1 α2 2 2 128 1 1 1 kh1 − h2 k2 + kh2 k2 − α1 α1 α2 |α1 − α2 | ≤ 2 kh1 − h2 k2 + 2 |α1 α2 | . kx1 − x2 k2 + kh1 − h2 k2 , ≤ where the last inequality comes from Lemma 54 as well as the facts that |α1 | ≥ 1/2 and |α2 | ≥ 1/2 as q shown above. Combining all of the above bounds and recognizing that kx1 − x2 k2 + kh1 − h2 k2 ≤ 2 2 2 kx1 − x2 k2 + 2 kh1 − h2 k2 , we conclude the proof. Finally, there is a useful identity associated with the minimizer of ge(α) as defined below. Lemma 56. For any h1 , h2 , x1 , x2 ∈ CK , denote α] := arg min ge(α), α e1 = e1 = α] x1 and h Let x 1 h1 , α] where then we have e 1 − x2 x 2 2 1 h1 − h2 α ge (α) := e 1 − h2 + x∗2 (e x1 − x2 ) = h Proof. We can rewrite the function ge (α) as 2 2 2 2 2 + kαx1 − x2 k2 .
e 1 − h2 ∗ h2 . + h  ∗   1 1 h1 h2 − h∗2 h1 α α 1 1 1 2 2 2 2 = αα kx1 k2 + kx2 k2 − αx∗1 x2 − αx∗2 x1 + kh1 k2 + kh2 k2 − h∗1 h2 − h∗2 h1 . αα α α 2 2 2 ∗ ge (α) = |α| kx1 k2 + kx2 k2 − (αx1 ) x2 − x∗2 (αx1 ) + 1 α 2 2 2 kh1 k2 + kh2 k2 − The first-order optimality condition [KD09, equation (38)] requires     ∂e g 1 1 1 2 2 = α] kx1 k2 − x∗1 x2 + ] − 2 kh1 k2 − − 2 h∗2 h1 = 0, ∂α α=α] α α] α] which further simplifies to 2 e1 e∗1 x2 = h ke x1 k2 − x 2 2 e1 − h∗2 h e 1 = 1 h1 , and α] 6= 0 (otherwise ge(α] ) = ∞ and cannot be the minimizer). Furthermore, e1 = α] x1 , h since x α] this condition is equivalent to
e 1 − h2 ∗ h e 1. e∗1 (e x x1 − x2 ) = h Recognizing that
∗ e∗1 (e e1 − x2 (e x x1 − x2 ) = x∗2 (e x1 − x2 ) + x x1 − x2 ) = x∗2 (e x1 − x2 ) + ke x1 − x2 k22 ,




∗ e∗ h e 1 − h2 = h∗ h e 1 − h2 + h e 1 − h2 h e 1 − h2 = h∗ h e 1 − h2 + kh e 1 − h2 k2 , h 1 2 2 2 we arrive at the desired identity. D.3.4 Matrix concentration inequalities The proof for  blind deconvolution is largely built upon the concentration of random matrices that are functions of aj a∗j . In this subsection, we collect the measure concentration results for various forms of random matrices that we encounter in the analysis. 129 

i.i.d. Lemma 57. Suppose aj ∼ N 0, 12 IK + iN 0, 21 IK for every 1 ≤ j ≤ m, and {cj }1≤j≤m are a set of e1 , C e2 > 0 such that for all t ≥ 0 fixed numbers. Then there exist some universal constants C   ( )! m X t t2 ∗ e e   cj (aj aj − IK ) ≥ t ≤ 2 exp C1 K − C2 min . P , Pm 2 maxj |cj | j=1 cj j=1 Proof. This is a simple variant of [Ver12, Theorem 5.39], which uses the Bernstein inequality and the standard covering argument. Hence we omit its proof.

i.i.d. Lemma 58. Suppose aj ∼ N 0, 21 IK +iN 0, 21 IK for every 1 ≤ j ≤ m. Then there exist some absolute e1 , C e2 , C e3 > 0 such that for all max{1, 3C e1 K/C e2 }/m ≤ ε ≤ 1, one has constants C   ! X e2 C e3 e C e ∗ e   aj aj ≥ 4C3 εm log ≤ 2 exp − P sup εm log , ε 3 ε |J|≤εm j∈J where J ⊆ [m] and |J| denotes its cardinality. Proof. The proof relies on Lemma 57 and the union bound. First, invoke Lemma 57 to see that for any fixed J ⊆ [m] and for all t ≥ 0, we have     X  e1 K − C e2 |J| min t, t2 , P (aj a∗j − IK ) ≥ |J| t ≤ 2 exp C (294) j∈J e1 , C e2 > 0, and as a result, for some constants C    X (i) P  sup aj a∗j ≥ dεme(1 + t) ≤ P  sup |J|≤εm j∈J  |J|=dεme ≤ P  sup (ii) ≤  |J|=dεme  X j∈J X j∈J  aj a∗j ≥ dεme(1 + t)  (aj a∗j − IK ) ≥ dεmet    m e1 K − C e2 dεme min t, t2 , · 2 exp C dεme where dce denotes the smallest integer that is no smaller than c. Here, (i) holds since we take the supremum
over a larger set and (ii) results from (294) and the union bound. Apply the elementary inequality nk ≤ (en/k)k for any 0 ≤ k ≤ n to obtain   dεme    X  em e1 K − C e2 dεme min t, t2 exp C P  sup aj a∗j ≥ dεme(1 + t) ≤ 2 dεme |J|≤εm j∈J    e1 K − C e2 εm min t, t2 exp C ε h  i  e1 K − εm C e2 min t, t2 − 2 log(e/ε) . = 2 exp C ≤2  e 2εm (295) where the second inequality uses εm ≤ dεme ≤ 2εm whenever 1/m ≤ ε ≤ 1. e3 ≥ max{1, 6/C e2 } and t = C e3 log(e/ε). To see this, it is The proof is then completed by taking C 2 e e2 εm/3 ≤ C e2 εmt/3, and easy to check that min{t, t } = t since t ≥ 1. In addition, one has C1 K ≤ C e 2 log(e/ε) ≤ C2 t/3. Combine the estimates above with (295) to arrive at     X X (i) e3 εm log(e/ε) ≤ P  sup P  sup aj a∗j ≥ 4C aj a∗j ≥ dεme(1 + t) |J|≤εm j∈J |J|≤εm 130 j∈J h  i  e1 K − εm C e2 min t, t2 − 2 log(e/ε) ≤ 2 exp C ! e2 C e3 C e2 t/3 = 2 exp − εm log(e/ε) ≤ 2 exp −εmC 3  (ii)  e3 log(e/ε). The inequality as claimed. Here (i) holds due to the facts that dεme ≤ 2εm and 1 + t ≤ 2t ≤ 2C (ii) arises from the estimates listed above.
Lemma 59. Suppose m  K log3 m. With probability exceeding 1 − O m−10 , we have m X 2 a∗j x\ bj b∗j j=1 Proof. The identity Pm j=1 − IK . r K log m. m bj b∗j = IK allows us to rewrite the quantity on the left-hand side as m X j=1 m  X 2 a∗j x\ bj b∗j − IK = j=1 |  − 1 bj b∗j , {z } 2 a∗j x\ :=Zj where the Zj ’s are independent zero-mean random matrices. To control the above spectral norm, we resort to the matrix Bernstein inequality [Kol11, Theorem 2.7]. To this end, we first need to upper bound the sub-exponential norm k · kψ1 (see definition in [Ver12]) of each summand Zj , i.e. kZj k 2 ψ1 = kbj k2 a∗j x\ 2 −1 2 ψ1 . kbj k2 a∗j x\ 2 ψ1 . K , m where we make use of the facts that 2 kbj k2 = K/m a∗j x\ and We further need to bound the variance parameter, that is,   " m m X X 2 ∗ σ0 := E  Zj Zj  = E a∗j x\ j=1 . m X j=1 bj b∗j bj b∗j = j=1 2 2 ψ1 . 1. 2 − 1 bj b∗j bj b∗j # m K K X bj b∗j = , m j=1 m 
2  Pm where the second line arises since E |a∗j x\ |2 − 1  1, kbj k22 = K/m, and j=1 bj b∗j = IK . A direct application of the matrix
Bernstein inequality [Kol11, Theorem 2.7] leads us to conclude that with probability exceeding 1 − O m−10 , m X Zj . max j=1 (r ) r K K K 2 log m, log m  log m, m m m where the last relation holds under the assumption that m  K log3 m. D.3.5 Matrix perturbation bounds We also need the following perturbation bound on the top singular vectors of a given matrix. The following lemma is parallel to Lemma 34. 131 Lemma 60. Let σ1 (A), u and v be the leading singular value, left and right singular vectors of A, respece u e respectively. e and v e be the leading singular value, left and right singular vectors of A, tively, and let σ1 (A), e Suppose σ1 (A) and σ1 (A) are not identically zero, then one has
e v A−A e ≤ σ1 (A) − σ1 (A) q p e u e σ1 (A) u − σ1 (A) + 2 q p e v e σ1 (A) v − σ1 (A) 2 2 e ; e k2 + kv − v ek2 ) A + (ku − u ≤ Proof. The first claim follows since p e 2 σ1 (A) − σ1 (A) q e k2 + kv − v ek2 ) + p σ1 (A) (ku − u . e σ1 (A) + σ1 (A) ev e = u∗ Av − u e ∗ Ae σ1 (A) − σ1 (A)
e −u e + u e −u ev e v + u∗ Av e ∗ Av e ∗ Av e ∗ Ae ≤ u∗ A − A
e v + ku − u e + A e kv − v e k2 A ek2 . ≤ A−A 2 With regards to the second claim, we see that q p p p e u e ≤ e σ1 (A) u − σ1 (A) σ1 (A) u − σ1 (A) u 2 = = Similarly, one can obtain p σ1 (A) v − q p p e k2 + σ1 (A) ku − u p e k2 + p σ1 (A) ku − u e v e σ1 (A) 2 ≤ p Add these two inequalities to complete the proof. 2 + σ1 (A) − e− σ1 (A) u q e σ1 (A) q e u e σ1 (A) e σ1 (A) − σ1 (A) q . e σ1 (A) + σ1 (A) ek2 + p σ1 (A) kv − v 132 p e σ1 (A) − σ1 (A) q . e σ1 (A) + σ1 (A) 2 

Expansion of polynomial Lie group integrals in terms of certain maps on surfaces, and factorizations of permutations Marcel Novaes arXiv:1601.08206v2 [math-ph] 7 Feb 2016 Instituto de Fı́sica, Universidade Federal de Uberlândia, Uberlândia, MG, 38408-100, Brazil Abstract Using the diagrammatic approach to integrals over Gaussian random matrices, we find a representation for polynomial Lie group integrals as infinite sums over certain maps on surfaces. The maps involved satisfy a specific condition: they have some marked vertices, and no closed walks that avoid these vertices. We also formulate our results in terms of permutations, arriving at new kinds of factorization problems. 1 1.1 Introduction Background We are interested in integrals over the orthogonal group O(N ) of N -dimensional real matrices O satisfying OOT = 1, where T means transpose, and over the unitary group U (N ) of N -dimensional complex matrices U satisfying U U † = 1, where † means transpose conjugate. These groups have a unique invariant probability measure, known as Haar measure, and integrals over them may be seen as averages over ensembles of random matrices. We will considerQaverages of functions that are polynomial Q in the matrix elements, † n i.e. quantities like k=1 Uik ,jk Upk ,qk for the unitary group and nk=1 Oik ,jk Oikb ,jkb for the orthogonal one (results for the unitary symplectic group Sp(N ) are very close to those for O(N ), so we do not consider this group in detail). From the statistical point of view, these are joint moments of the matrix elements, considered as correlated random variables. Their study started in physics [1, 2], with applications to quantum chaos [3, 4, 5], and found its way into mathematics, initially for the unitary group [6] and soon after for the orthogonal and symplectic ones [7, 8]. Since then, they have been extensively explored [9, 10], related to Jucys-Murphy elements [11, 12] and generalized to symmetric spaces [13]. Unsurprisingly, after these developments some new applications have been found in physics [14, 15, 16, 17, 18, 19]. For the unitary group average, the result is different from zero only if the q-labels are a permutation of the i-labels, and the p-labels are a permutation of the j-labels. The basic building blocks of the calculation, usually called Weingarten functions, are of the kind WgU N (π) = Z dU U (N ) n Y † , Uk,k Uk,π(k) (1) k=1 where π is an element of the permutation group, π ∈ Sn . In general, if there is more than one permutation relating the sets of labels (due to repeated indices, e.g. h|U1,2 |4 i), the result is a sum of Weingarten functions. The cycletype of a permutation in Sn is a partition of n, α = (α1 , α2 , ...) ⊢ n = |α|, whose parts are the lengths of its cycles; the function WgU N (π) depends only on the cycletype of π [6, 7]. 1 The result of the orthogonal group average is different from zero only if the i-labels satisfy some matching (see Section 2.1 for matchings) and the j-labels also satisfy some matching [7, 9, 8]. For concreteness, we may choose the i’s to satisfy only the trivial matching, and the j’s to satisfy only some matching m of cosettype β. The basic building blocks of polynomial integrals over the orthogonal group are Z O dOO1,j1 O1,jb1 O2,j2 O2,jb2 · · · On,jn On,jnb . (2) WgN (β) = Ø(N ) As examples, let us mention Z U WgN ((2)) = U (N ) † † dU U1,1 U1,2 U2,2 U2,1 = −1 , (N − 1)N (N + 1) corresponding to the permutation π = (12), which has cycletype (2), and Z −1 O dOO1,1 O1,2 O2,2 O2,1 = WgN ((2)) = , (N − 1)N (N + 2) Ø(N ) (3) (4) corresponding to the matching {{1, b 2}, {2, b 1}} for the j’s, which has cosettype (2). Our subject is the combinatorics associated with the large N asymptotics of these integrals. For the unitary case, this has been addressed in previous works [4, 6, 11, 20, 21], where connections with maps and factorizations of permutations have already appeared. However, our approach is different, and the maps/factorizations we encounter are new. Our treatment of the orthogonal case is also new. We proceed by first considering Weingarten functions in the context of ensembles of random truncated matrices; then relating those to Gaussian ensembles, and finally using the rich combinatorial structure of the latter. In what follows we briefly present our results. A review of some basic concepts can be found in Section 2. Results for the unitary group are obtained in Section 3, while the orthogonal group is considered in Section 4. In an Appendix we discuss several different factorization problems that are related to 1/N expansions of Weingarten functions. 1.2 1.2.1 Results and discussion for the unitary group Maps For the unitary group, we represent the Weingarten function as an infinite sum over orientable maps. Theorem 1 If α is a partition with ℓ(α) parts, then WgU N (α) = (−1)ℓ(α) X χ N N 2|α|+ℓ(α) χ X (−1)V (w) , (5) w∈B(α,χ) where the first sum is over Euler characteristic, and V (w) is the number of vertices in the map w. As we will discuss with more detail in Section 3.2 the (finite) set B(α, χ) contains all maps, not necessarily connected, with the following properties: i) they are orientable with Euler characteristic χ; ii) they have ℓ(α) marked vertices with valencies (2α1 , 2α2 , ...); iii) all other vertices have even valence larger than 2; iv) all closed walks along the boundaries of the edges (which we see as ribbons) visit the marked vertices in exactly one corner; v) they are face-bicolored and have |α| faces of each color. As an example, the 1/N expansion for the function in (3), for which α = (2), starts − 1 1 1 − 5 − 7 − ··· , 3 N N N 2 (6) Figure 1: A few of the maps that contribute to the Weingarten function (3), according to the expansion (5). They are face-bicolored, with different colors being indicated by solid and dashed lines, respectively. All boundary walks visit the marked vertex, which is the black disk. Their contributions are discussed in the text. The first two orders come from maps like the ones shown in Figure 1. The map in panel a) is the only leading order contribution. It has χ = 2 and V = 2, so its value is indeed −1/N 3 . The next order comes from elements of B((2), 0). There are in total 21 different maps in that set having four vertices of valence 4; one of them is shown in panel b). There are 28 different maps in that set having two vertices of valence 4 and one vertex of valence 6; one of them is shown in panel c). Finally, there are 8 different maps in that set having one vertex of valence 4 and one vertex of valence 8; one of them is shown in panel d). They all have χ = 0, and their combined contribution is (−21 + 28 − 8)/N 5 = −1/N 5 . 1.2.2 Factorizations By a factorization of Π we mean an ordered pair, f ≡ (τ1 , τ2 ), such that Π = τ1 τ2 . We call Π the ‘target’ of the factorization f . If Π ∈ Sk , then the Euler characteristic of f is given by χ = ℓ(Π) − k + ℓ(τ1 ) + ℓ(τ2 ). The number of factorizations of a permutation depends only on its cycletype, so it makes sense to restrict attention to specific representatives. Call a permutation ‘standard’ if each of its cycles contains only adjacent numbers, and the cycles have weakly decreasing length when ordered with respect to least element. For example, (123)(45) is standard. Since WgU N (π) depends only on the cycletype of π, we may take π to be standard. As we will discuss with more detail in Section 3.3, the relevant factorizations for WgU N (π) are those whose target is of the kind Π = πρ, where the ‘complement’ ρ is a standard permutation acting on the set {n + 1, ..., n + m}, for some m ≥ 0. They satisfy the following properties: i) they have Euler characteristic χ; ii) the complement ρ has no fixed points; iii) every cycle of the factors τ1 , τ2 must have exactly one element in {1, ..., n}. Notice that the last condition implies ℓ(τ1 ) = ℓ(τ2 ) = n. Let the (finite) set of all such factorizations be denoted F(π, χ). Then, we have 3 Theorem 2 Let π ∈ Sn be a standard permutation, then WgU N (π) = Q (−1)ℓ(π) X χ N N 2n+ℓ(π) χ X f ∈F (π,χ) (−1)ℓ(Π) , zρ (7) where zρ = j j vj vj !, with vj the number of times part j occurs in the cycletype of the complement ρ. Theorem 2 follows from Theorem 1 by a simple procedure for associating factorizations to maps in B(α, χ), discussed in Section 3.3. Associations of this kind are well known [22, 23, 24]. For example, the leading order in Eq.(6), for which π = (12), has a contribution from the factorization (12)(34) = (14)(23) · (13)(24), which has ρ = (34) and is one of two factorizations that can be associated with the map in Figure 1a. Several factorizations can be associated with the other maps in Figure 1. We mention one possibility for each of them: (12)(34)(56)(78) = (148)(25763) · (13)(285746) for the map in Figure 1b; (12)(345)(67) = (17)(23546) · (16)(27435) for the map in Figure 1c; (12)(3456) = (164)(253) · (1365)(24) for the map in Figure 1d. Notice how they have different complements. Other factorization problems are also related to the coefficients in the 1/N expansion for WgU N (π) (see the Appendix). Collins initially showed [6] that they can be expressed in terms of the number of ‘Proper’ factorizations of π. Matsumoto and Novak later showed [11] that the coefficients count Monotone factorizations. On the other hand, Berkolaiko and Irving recently defined [25] Inequivalent-Cycle factorizations and showed that X X (−1)r Iα,χ (r) = (−1)r Pα,χ (r) = (−1)n+ℓ(α) Mα,χ , (8) r≥0 r≥0 where Iα,χ (r) and Pα,χ (r) are, respectively, the numbers of Inequivalent-Cycle and Proper factorizations of π, with cycletype α ⊢ n, into r factors having Euler characteristic χ, while Mα,χ is the number of Monotone factorizations of π with Euler characteristic χ. Interestingly, our factorizations satisfy a very similar sum rule, namely X f ∈F (π;χ) (−1)ℓ(Π) = (−1)n+ℓ(α) Mα,χ . zρ (9) An important difference between (9) and (8) is that the factorizations in (9) take place in Sn+m for some m ≥ 0, while all those in (8) take place in Sn . Notice that our factorizations must satisfy condition iii), which is related to the distribution of the elements from the set {1, ..., n} among the cycles of the factors τ1 , τ2 . This is close in spirit to the kind of questions studied by Bóna, Stanley and others [26, 27], which count factorizations satisfying some placement conditions on the elements of the target. 1.3 1.3.1 Results and discussion for the orthogonal group Maps For the orthogonal group, we represent the Weingarten function as an infinite sum over maps (orientable and non-orientable). Theorem 3 Let β be a partition with ℓ(β) parts, then WgO N +1 (β) (−2)ℓ(β) X χ N = 2|β|+ℓ(β) N χ 4 X w∈N B(β,χ)  1 − 2 V (w) , (10) Figure 2: Two of the maps that contribute to the Weingarten function (4), according to the expansion (10). They are not orientable: some of the ribbons have ‘twists’. where the first sum is over Euler characteristic, and V (w) is the number of vertices in the map w. As we will discuss with more detail in Section 4.2 the (finite) set N B(β, χ) contains all maps, not necessarily connected or orientable, with the following properties: i) they have Euler characteristic χ; ii) they have ℓ(β) marked vertices with valencies (2β1 , 2β2 , ...); iii) all other vertices have even valence larger than 2; iv) all closed walks along the boundaries of the edges visit the marked vertices in exactly one corner; v) they are face-bicolored and have |β| faces of each color. Notice that the expansion is in powers of N −1 for the group O(N + 1) and not for O(N ). For example, the first terms in the expansion of the function in Eq.(4) at dimension N + 1 are 1 4 13 −1 = − 3 + 4 − 5 ... (11) N (N + 1)(N + 3) N N N The leading order comes from the map shown in Figure 1a, which is orientable. The next order comes from non-orientable maps in the set N B((2), 1). There are in total 8 different maps having two vertices of valence 2; one of them is shown in Figure 2a. There are in total 4 different maps having one vertex of valence 2 and one vertex of valence 3; one of them is shown in Figure 2b. Their combined contribution is (8 − 4)/N 4 = 4/N 4 . Remark It is known [13] that the Weingarten function of the unitary symplectic group Sp(N ) is proportional to WgO −2N . Therefore the appropriate dimension for map expansion in Sp(N ) is also different from N : it has to be Sp(N − 1/2) (a non-integer dimension is to be understood by keeping in mind that Weingarten functions are rational functions of N ). Interestingly, if we assign a parameter α = 2, 1, 1/2 for orthogonal, unitary and symplectic groups, respectively (sometimes called Jack parameter), then the appropriate dimensions for the map expansion in powers of N −1 can be written as N + α − 1. 1.3.2 Factorizations c = {b c Define the Let [n] = {1, ..., n} and [n] 1, ..., n b}. We consider the action of S2n on [n]∪ [n]. ‘hat’ involution on permutations as π b(a) = π −1 (b a) (assuming b b a = a). We call permutations that are invariant under this transformation ‘palindromic’, e.g. (12b 2b 1) and (12)(b 2b 1) are palindromic. Given a partition β, define π ∈ Sn to be the standard permutation that has cycletype β and define Π = πρ, where the ‘complement’ ρ is a standard permutation acting on the set {n + 1, ..., n + m} for some m ≥ 0. Define the fixed-point free involutions p1 , whose cycles are (a b a) for 1 ≤ a ≤ n + m, and p2 whose cycles are of the type (b a a + 1), but with the additions computed modulo the cycle lengths of Π, i.e. \ p2 = (b 1 2)(b 2 3) · · · (βb1 1)(β\ 1 + 1 β1 + 2) · · · (β1 + β2 β1 + 1) · · · . 5 (12) b They provide a factorization of the palindromic version of the target, p2 p1 = Π Π. b The problem we need to solve is to find all factorizations Π Π = f2 f1 , that satisfy the following properties: i) their Euler characteristic, defined as ℓ(Π) − m − n + ℓ(f1 ) + ℓ(f2 ), is χ; ii) the complement ρ has no fixed points; iii) the factors may be written as f1 = θp1 and f2 = p2 θ for some fixed-point free involution θ; iv) f1 is palindromic; v) every cycle c Clearly, the crucial quantity of the factors f1 , f2 contains exactly one element in [n] ∪ [n]. is actually θ. Let the (finite) set of all pairs (Π, θ) satisfying these conditions be denoted N F(β, χ). Then, we have Theorem 4 For a given partition β of length ℓ(β), WgO N +1 (β) Q (−2)ℓ(β) X χ N = 2|β|+ℓ(β) N χ X (Π,θ)∈N F (β,χ) 1 zρ  1 − 2 ℓ(Π) , (13) where zρ = j j vj vj !, with vj the number of times part j occurs in the cycletype of the complement ρ. Theorem 4 follows from Theorem 3 by a simple procedure for describing combinatorially the maps in N B(β, χ), discussed in Section 4.3. Such kind of descriptions are well known [28, 29, 30]. For example, the leading order in Eq.(11) has a contribution from the pair Π = (12)(34), θ = (1b 3)(2b 4)(4b 2)(3b 1). The next order comes from factorizations associated with the maps in Figure 2. We mention one for each of them: Π = (12)(34)(56), θ = (1b 3)(b 13)(25)(b 24)(b 46)(b 5b 6) b b b b b for a) and Π = (12)(345), θ = (13)(13)(25)(24)(45) for b). Other factorizations are related to the coefficients in the 1/N expansion of orthogonal Weingarten functions, as we discuss in the Appendix. Matsumoto has shown [31] that these coefficients count certain factorizations that are matching analogues of the Monotone family. Following the steps in [6] we show that the coefficients can be expressed in terms of analogues of Proper factorizations. These relations hold for O(N ), however, not O(N + 1). Therefore the relation to our results is less direct. On the other hand, Berkolaiko and Kuipers have shown [32] that certain ‘Palindromic Monotone’ factorizations are related to the 1/N expansion for O(N + 1). The appropriate analogue of Inequivalent-Cycle factorizations is currently missing. 1.4 Connection with physics This work was originally motivated by applications in physics, in the semiclassical approximation to the quantum mechanics of chaotic systems. Without going into too much detail, the problem in question was to construct correlated sets of chaotic trajectories that are responsible for the relevant quantum effects in the semiclassical limit [33]. Connections between this topic and factorizations of permutations had already been noted [34, 32, 35, 36, 37]. In [18] we suggested that such sets of trajectories could be obtained from the diagrammatic expansion of a certain matrix integral, with the proviso that the dimension of the matrices had to be set to zero after the calculation. When we realized the connection to truncated random unitary matrices, this became our Theorem 1. The suggestion from [18], initially restricted to systems without time-reversal symmetry, was later extended to systems that have this symmetry [19]; the connection with truncated random orthogonal matrices then gave rise to our Theorem 3. 1.5 Acknowledgments The connection between our previous work [18] and truncated unitary matrices was first suggested by Yan Fyodorov. 6 Figure 3: Three examples of graphs associated with matchings. Edges coming from the trivial matching, t, are drawn in dashed line. In a) we have m = {{1, b 2}, {2, b 1}, {3, b 3}}, which can be produced as (12)(t) or (b 1b 2)(t), among others ways; permutations (12) and (b 1b 2) thus belong to the same coset in S2n /Hn . In b) m = {{2, b 3}, {3, b 2}}, which can be produced as (23)(t) or (b 2b 3)(t), among others. The matchings in both a) and b) have the same cosettype, namely (2, 1); the permutations producing them therefore belong to the same double coset in Hn \S2n /Hn . In c) m = {{1, 2}, {b 1, b 2}, {3, b 4}, {4, b 5}, {5, b 3}}, and its cosettype is (3, 2). This work had financial support from CNPq (PQ-303318/2012-0) and FAPEMIG (APQ00393-14). 2 2.1 Basic concepts Partitions, Permutations and Matchings By α ⊢ n we mean Pα is a partition of n, i.e. a weakly decreasing sequence of positive integers such that i αi = |α| = n. The number of non-zero parts is called the length of the partition and denoted by ℓ(α). The group of permutations of n elements is Sn . The cycletype of π ∈ Sn is the partition α ⊢ n whose parts are the lengths of the cycles of π. The length of a permutation is the number of cycles it has, ℓ(π) = ℓ(α), while its rank is r(π) = n − ℓ(π) = r(α). We multiply permutations from right to left, e.g. (13)(12) = (123). The conjugacyQ class Cλ is the set of all permutations with cycletype λ. Its size is |Cλ | = n!/zλ , with zλ = j j vj vj !, where vj is the number of times part j occurs in λ. c = {b c be a set with 2n elements. A matching Let [n] = {1, ..., n}, [n] 1, ..., n b} and let [n]∪ [n] on this set is a collection of n disjoint subsets with two elements each. The set of all such matchings is Mn . The trivial matching is defined as t = {{1, b 1}, {2, b 2}, ..., {n, n b}}. A permutation π acts on matchings by replacing blocks like {a, b} by {π(a), π(b)}. If π(t) = m we say that π produces the matching m. c two Given a matching m, let Gm be a graph with 2n vertices having labels in [n] ∪ [n], vertices being connected by an edge if they belong to the same block in either m or t. Since each vertex belongs to two edges, all connected components of Gm are cycles of even length. The cosettype of m is the partition of n whose parts are half the number of edges in the connected components of Gm . See examples in Figure 3. A permutation has the same cosettype as the matching it produces. c It We realize the group S2n as the group of all permutations acting on the set [n] ∪ [n]. n has a subgroup called the hyperoctahedral, Hn , with |Hn | = 2 n! elements, which leaves invariant the trivial matching and is generated by permutations of the form (a b a) or of b the form (a b)(b a b). The cosets S2n /Hn ∼ Mn can therefore be represented by matchings. The trivial matching identifies the coset of the identity permutation. We may inject Mn into S2n by using fixed-point free involutions. This is done by the simple identification m = {{a1 , a2 }, {a3 , a4 }, ...} 7→ σm = (a1 a2 )(a3 a4 ) · · · . 7 3 3 5 3 5 4 4 5 3 3 5 4 4 5 3 5 4 4 Figure 4: Wick’s rule for complex Gaussian random matrices. Starting from hTr(ZZ † )3 i = P P † † † b1 ,b2 ,b3 Za1 b1 Zb1 a2 Za2 b2 Zb2 a3 Za3 b3 Zb3 a1 , we arrange the elements around a vertex a1 ,a2 ,a3 (left panel). Then, we produce all possible connections between the marked ends of the arrows, respecting orientation. Two of them are shown. The map in the middle panel has only one face of each color, and χ = 0. The map in the right panel has three faces of one color and a single face of the other, with χ = 2. The double cosets Hn \S2n /Hn , on the other hand, are indexed by partitions of n: two permutations belong to the same double coset if and only if they have the same cosettype [38] (hence this terminology). We denote by Kλ the double coset of all permutations with cosettype λ. Its size is |Kλ | = |Hn ||Cλ |2r(λ) . Given a sequence of 2n numbers, (i1 , ..., i2n ) we say that it satisfies the matching m if the elements coincide when paired according to m. This is quantified by the function Q ∆m (i) = b∈m δib1 ,ib2 , where the product runs over the blocks of m and b1 , b2 are the elements of block b. 2.2 Maps Maps are graphs drawn on a surface, with a well defined sense of rotation around each vertex. We represent the edges of a map by ribbons. These ribbons meet at the vertices, which are represented by disks, and the region where two ribbons meet is called a corner. It is possible to go from one vertex to another by walking along a boundary of a ribbon. Upon arriving at a vertex, a walker may move around the boundary of the disk to a boundary of the next ribbon, and then depart again. Such a walk we call a boundary walk. A boundary walk that eventually retraces itself is called closed and delimits a face of the map. Let V , E and F be respectively the numbers of vertices, edges and faces of a map. The Euler characteristic of the map is χ = V − E + F , and it is additive in the connected components (we do not require maps to be connected, and each connected component is understood to be drawn on a different surface). In all of the maps used in this work, ribbons have boundaries of two different colors. For convenience, we represent those colors by simply drawing these boundaries with two types os lines: dashed lines and solid lines. Ribbons are attached to vertices in such a way that all corners and all faces have a well defined color, i.e. our maps are face-bicolored. Examples are shown in Figures 1 and 2. 2.3 Non-hermitian Gaussian random matrices We consider N -dimensional matrices for which the elements are independent and identicallydistributed Gaussian random variables, the so-called Ginibre ensembles [39]. For real matrices we will use the notation M , for complex ones we use Z. 8 3 3 5 3 3 5 4 5 5 4 4 4 Figure 5: Wick’s rule for real Gaussian random matrices. Starting from hTr(M M T )3 i, we arrange the elements around a vertex (left panel). Then, we produce all possible connections between the marked ends of the arrows. One of them is shown, which has two faces of one color and a single face of the other, with χ = 1. Notice that the boundaries of the edges are not oriented, and that two of them have twists. Normalization constants for these ensembles are defined as Z Z Ω T † ZR = dM e− 2 Tr(M M ) , ZC = dZe−ΩTr(ZZ ) . (14) They can be computed (as we do later) using singular value decomposition. Average values are denoted by Z Ω 1 T dM e− 2 Tr(M M ) f (M ), (15) hf (M )i = ZR and Z 1 † dZe−ΩTr(ZZ ) f (Z), (16) hf (Z)i = ZC the meaning of h·i being clear from context. We have the simple covariances hMab Mcd i = 1 δac δbd , Ω (17) and 1 δac δbd . (18) Ω Polynomial integrals may be computed using Wick’s rule, which is a combinatorial prescription for combining covariances. It simply states that, since the elements are independent, the average of a product can be decomposed in terms of products of covariances. In the complex case, we may consider the elements of Z fixed and then permute the elements of Z † in all possible ways, * n + n Y 1 X Y † δak ,dπ(k) δbk ,cπ(k) . (19) Zak bk Zck dk = n Ω ∗ ∗ Zcd i = 0, hZab Zcd i = hZab ∗ i= hZab Zcd π∈Sn k=1 k=1 For example, hZa1 b1 Zd∗1 c1 Za2 b2 Zd∗2 c2 i = 1 1 δa1 ,d1 δb1 ,c1 δa2 ,d2 δb2 ,c2 + 2 δa1 ,d2 δb1 ,c2 δa2 ,d1 δb2 ,c1 . 2 Ω Ω In the real case, we must consider all possible matchings among the elements, * 2n + Y 1 X ∆m (a)∆m (b). Mak bk = n Ω k=1 m∈Mn 9 (20) (21) For example, 1 1 δa1 ,a2 δa3 ,a4 δb1 ,b2 δb3 ,b4 + 2 δa1 ,a3 δa2 ,a4 δb1 ,b3 δb2 ,b4 2 Ω Ω 1 + 2 δa1 ,a4 δa2 ,a3 δb1 ,b4 δb2 ,b3 . Ω hMa1 b1 Ma2 b2 Ma3 b3 Ma4 b4 i = (22) These Wick’s rules have a well known diagrammatic interpretation (see, e.g., [40, 41, 42, 43, 44]). In the complex case, matrix elements are represented by ribbons having borders oriented in the same direction but with different colors. Ribbons from elements of Z have a marked head, while ribbons from elements of Z † have a marked tail. Ribbons coming from traces are arranged around vertices, so that all marked ends are on the outside and all corners have a well defined color. Wick’s rule consists in making all possible connections between ribbons, using marked ends, respecting orientation. This produces a map (not necessarily connected). According to Eq.(18), each edge leads to a factor Ω−1 . In the real cases, the boundaries of the ribbons are not oriented and the maps need not be orientable: the edges may contain a ‘twist’. We show an example for the complex case in Figure 4, and an example for the real case in Figure 5. 3 Unitary Group 3.1 Truncations Let U be a random matrix uniformly distributed in U (N ) with the appropriate normalized Haar measure. Let A be the M1 × M2 upper left corner of U , with N ≥ M1 + M2 and M1 ≤ M2 . It is known [45, 46, 47, 48, 49] that A, which satisfies AA† < 1M1 , becomes distributed with probability density given by P (A) = 1 det(1 − AA† )N0 , Y1 (23) where N0 = N − M1 − M2 (24) and Y1 is a normalization constant. The value of Y1 can be computed using the singular value decomposition A = W DV , where W and V are matrices from U (M1 ) and U (M2 ), respectively. Matrix D is real, diagonal and non-negative. Let T = D 2 = diag(t1 , t2 , ...). Then [40, 50, 51], Y1 = Z dW U (M1 ) Z dV U (M2 ) where ∆(t) = Z M1 1Y 0 i=1 M1 M1 Y Y 2 −M1 dti (1 − ti )N0 tM |∆(t)|2 , i (25) (tj − ti ). (26) i=1 j=i+1 If we denote the angular integrals by Z U (M1 ) dW Z U (M2 ) dV = V1 , (27) then Selberg’s integral tells us that [52, 53] Y1 = V 1 M1 Y Γ(j + 1)Γ(M2 + 1 − j)Γ(N − M2 − j + 1) . Γ(N − j + 1) j=1 10 (28) e Consider now an even smaller subblock of U , which is contained in A. Namely, let U be the N1 × N2 upper left corner of U , with N1 ≤ M1 and N2 ≤ M2 . We shall make use of e can be computed either by the obvious fact that integrals involving matrix elements of U integrating over U or over A. In particular, the quantity WgU N (π) = Z dU U (N ) n Y k=1 with n ≤ N1 , N2 and π ∈ Sn , can also be written as WgU N (π) Z 1 = Y1 ek,k U e† U k,π(k) , n Y † N0 dA det(1 − AA ) AA† <1M1 (29) Ak,k A†k,π(k). (30) k=1 Notice that, although this may not be evident at first sight, the right-hand-side of equation (30) is actually independent of M1 and M2 . 3.2 Sum over maps The key to the diagrammatic formulation of our integral is the identity −N0 † det(1 − AA† )N0 = eN0 Trlog(1−AA ) = e P∞ 1 † q q=1 q Tr(AA ) . (31) We shall consider the first term in the series separately from the rest, and incorporate it into the measure, i.e. we will write −N0 dAe P∞ 1 † q q=1 q Tr(AA ) −N0 = dG (A)e P∞ 1 † q q=2 q Tr(AA ) , (32) where dG (A) is a Gaussian measure, † dG (A) = dAe−N0 Tr(AA ) . (33) We have WgU N (π) 1 = Y1 Z −N0 AA† <1M1 dG (A)e P∞ 1 † q q=2 q Tr(AA ) n Y Ak,k A†k,π(k) . (34) k=1 Taking into account that the series in the exponent diverges for AA† ≥ 1M1 and that e−∞ = 0, we extend the integration to general matrices A, WgU N (π) 1 = Y1 Z −N0 dG (A)e P∞ 1 † q q=2 q Tr(AA ) n Y Ak,k A†k,π(k). (35) k=1 Now we are in the realm of Gaussian integrals, and may apply Wick’s rule. For each cycle of π, the elements of A and A† in the last product can be arranged in counterclockwise order around vertices. This produces what we call ‘marked’ vertices, in number of ℓ(π). Formally expanding the exponential in Eq.(35) as a Taylor series in N0 will produce other vertices, let us call them ‘internal’, all of them of even valence larger than 2. This leads to an infinite sum over maps with arbitrary numbers of internal vertices and edges. The contribution of a map will be proportional to (−1)v N0v−E , if it has v internal vertices and E edges, which can be written as (−1)v N0v−E = (−1)ℓ(π) F +ℓ(π) N0 11 N0χ (−1)V . (36) However, the application of Wick’s rule may lead to closed boundary walks that visit only the internal vertices and avoid the marked ones. If a map has r1 closed boundary walks of one color and r2 closed boundary walks of the other color that avoid the marked vertices, its contribution will be proportional to M1r1 M2r2 . Crucially, since we know that WgU N (π) is actually independent of M1 , M2 , we are free to consider only the cases r1 = r2 = 0, or equivalently to take the simplifying limits M1 → 0 and M2 → 0. We will therefore be left only with maps in which all closed boundary walks visit the marked vertices. Another point we must address is that R the normalization constant Y1 is not equal to the corresponding Gaussian one, ZC = dG (A), so that the diagrammatic expression for WgU N (π) will be multiplied by ZC /Y1 . Using again the singular value decomposition of A, and  M 1 M 2 Y Z ∞Y M1 M1 1 −N0 ti M2 −M1 2 Γ(j + 1)Γ(M2 + 1 − j), (37) dti e ti |∆(t)| = N0 0 j=1 i=1 which can be obtained as a limiting case of the Selberg integral, we arrive at ZC = V 1 1 N0M1 M2 M1 Y j!(M2 − j)! (38) j=1 and, therefore, M1 Y 1 ZC (N − j)! . = M1 M2 Y1 (N − M2 − j)! N0 j=1 (39) In the end, we have the simplification ZC = 1. M2 →0 Y1 lim (40) The limit M1 → 0 is trivial. Notice that the limit of N0 is simply N . Because of the way we arranged the elements around the marked edges, our maps will always have 2n faces, being n of each color. We have therefore produced the maps in B(α, χ) and proved our Theorem 1. 3.3 Sum over factorizations Suppose a standard permutation π ∈ Sn with cycletype α ⊢ n. We shall associate factorizations of π with the maps in B(α, χ). In order to do that, we must label the corners of the maps. We proceed as follows. Consider first the marked vertex with largest valency, and pick a corner delimited by solid line. Label this corner and the corner following it in counterclockwise order with number 1. Then label the next two corners with the number 2, and so on until the label α1 is used twice. Proceed to another marked vertex in weakly decreasing order of valency, and repeat the above with integers from α1 + 1 to α1 + α2 . Repeat until all marked vertices have been labelled, producing thus the cycles of π. The same procedure is then applied to the internal vertices, producing the cycles of another standard permutation ρ, acting on the set {n + 1, ..., E}, where E = n + m is the number of edges in the map. Notice that ρ has no fixed points, since all internal vertices of maps in B(α, χ) have even valence larger than 2. Let Π = πρ. See Figure 6, where Π = (12)(34) for panel a), Π = (12)(34)(56)(78) for panel b), Π = (12)(345)(67) for panel c), and Π = (12)(3456) for panel d). Define the permutation ω1 to be such that its cycles have the integers in the order they are visited by the arrows in solid line. In the example of Figure 6, this would be 12 Figure 6: Labeling the maps from Figure 1, we can associate with them certain factorizations of permutations. ω1 = (14)(23) for panel a), ω1 = (184)(23675) for panel b), ω1 = (17)(26453) for panel c) and ω1 = (146)(235) for panel d). The cycles of ω1 correspond to the closed boundary walks in solid line. Permutation τ2 is defined analogously in terms of the arrows in dashed lines. In Figure 6, this would be τ2 = (13)(24) for panel a), τ2 = (13)(285746) for panel b), τ2 = (16)(27435) for panel c) and τ2 = (1365)(24) for panel d). The cycles of τ2 correspond to the closed boundary walks in dashed line. Suppose an initial integer, i. The arrow in dashed line which departs from the corner labelled by i arrives at the corner labelled by τ2 (i). On the other hand, the image of i under the permutation Π corresponds to the label of an outgoing arrow in solid line which, following ω1 , also arrives at τ2 (i). Therefore, we have, by construction, ω1 Π = τ2 or, equivalently, writing τ1 = ω1−1 , we have the factorization Π = τ1 τ2 . (41) For the maps in B(α, χ) all boundary walks visit the marked vertices, which means that all cycles of τ1 ad of τ2 must have exactly one element in the set {1, ..., n}. Therefore, the permutations satisfy the conditions we listed in Section 1.2.2. When we label the vertices of the map to produce a factorization, there are two kinds of ambiguities. First, for a vertex of valency 2j there are j possible choices for the first corner to be labelled. Second, if there are mj vertices of valency j, there are mj ! ways Q to order them. Hence, to a map for which the complement is ρ there correspond zρ = j j mj mj ! factorizations, where mj is the multiplicity of part j in the cycletype of ρ. The sum in (5) can indeed be written as WgU N (π) = 1 N 2n+ℓ(π) X χ Nχ X f ∈F (π,χ) (−1)ℓ(ρ) , zρ (42) where F(π, χ) is the set of factorizations of the kind we have described for given α and χ. This proves our Theorem 2. 13 4 Orthogonal group 4.1 Truncations Let O be a random matrix uniformly distributed in O(N + 1) with the appropriate normalized Haar measure. Let A be the M1 × M2 upper left corner of O, with N ≥ M1 + M2 and M1 ≤ M2 . It is known [48, 54] that A, which satisfies AAT < 1M1 , becomes distributed with probability density given by P (A) = 1 det(1 − AAT )N0 /2 , Y2 (43) where N0 = N − M1 − M2 (44) and Y2 is a normalization constant. Notice that we start with O(N + 1) and not O(N ). The value of Y2 can be computed using the singular value decomposition A = W DV , where W and V are matrices from O(M1 ) and O(M2 ), respectively. Matrix D is real, diagonal and non-negative. Let T = D 2 = diag(t1 , t2 , ...). Then [50, 51], Y2 = Z dW O(M1 ) Z dV O(M2 ) M1 1Y 0 i=1 If we denote the angular integrals by Selberg’s integral that Y2 = V 2 Z R (M2 −M1 −1)/2 dti (1 − ti )N0 /2 ti O(M1 ) dW R O(M2 ) dV |∆(t)|. (45) = V2 , then we have again from M1 Y Γ(j/2 + 1)Γ((M2 + 1 − j)/2)Γ((N − M2 − j)/2 + 1) . Γ((N − j)/2 + 1)Γ(3/2) (46) j=1 e Consider now an even smaller subblock of O, which is contained in A. Namely, let O be the N1 × N2 upper left corner of O, with N1 ≤ M1 and N2 ≤ M2 . The average value of e can be computed either by integrating over O or over any function of matrix elements of O A. In particular, the quantity WgO N +1 (β) = Z dO O(N +1) n Y k=1 ek,j O ek,j , O k b k (47) where k ≤ N1 , N2 and the j’s only satisfy some matching of cosettype β, can also be written as Z n Y 1 T N0 /2 (β) = WgO Ak,jk Ak,jkb . dA det(1 − AA ) (48) N +1 Y2 AAT <1M k=1 1 Notice that the right-hand-side of equation (48) is actually independent of M1 and M2 . 4.2 Sum over maps Analogously to the unitary case, we have WgO N +1 (β) N0 1 = Y2 Z − dG (A)e N0 2 P∞ 1 T q q=2 q Tr(AA ) 2n Y Ak,jk Ak,jkb , (49) k=1 T where now dG (A) = e− 2 Tr(AA ) . The diagrammatical considerations proceed as previously, except that we use the Wick’s rule of the real case and the resulting maps need not be 14 orientable. Also, a map now contributes (−N0 /2) for each internal vertex and 1/N0 for each edge. This gives a total contribution which is proportional to  V  v 1 N0χ 1 v−E = F +ℓ(β) − (−2)ℓ(β) , (50) N0 − 2 2 N where v is the number of internal vertices, V = v + ℓ is the total number of vertices, E is the number of edges and χ = F − E + V is the Euler characteristic, where F is the number of faces. When we take M2 → 0, and then M1 → 0, we arrive at maps with no closed boundary walks that avoid the marked vertices, having 2n faces, n of each color. We thus arrive at the maps in the set N B(β, χ). The Gaussian normalization constant is Z ∞Y Z M1 M1 Y N − 20 ti (M2 −M1 −1)/2 ti |tj − ti | (51) dti e ZR = dG (A) = V2 0 = V2  j=i+1 i=1  M1 M2 Y M1 2 2 Γ(1 + j/2)Γ((M2 + 1 − j)/2) , N0 Γ(3/2) (52) j=1 and we have ZR = lim lim M2 →0 M2 →0 Y2  2 N0  M1 M2 Y M1 2 j=1 Γ((N + 2 − j)/2) = 1. Γ((N − M2 + 2 − j)/2) (53) Once again, the limit M1 → 0 is trivial. This reduces N0 to N . Taking into account the contribution of the maps, already discussed, we arrive at our Theorem 3. 4.3 Sum over factorizations We now label the maps in N B(β, χ) in order to relate them to permutations. We only need to change slightly the labelling procedure we used for the maps in B(α, χ) in Section 3.3. First, we replace the labels of the corners in dashed line by ‘hatted’ versions. Second, instead of labelling corners, we now label half-edges, by rotating the previous labels counterclockwise. This is shown in Figure 7 (where the hatted labels are enclosed by boxes, while the normal ones are enclosed by circles). The unhatted labels, read in anti-clockwise order around vertices, produce a permutation Π which is standard. This can be written as Π = πρ, where π ∈ Sn has cycletype β and the complement ρ acts on the set {n + 1, ..., E} where E = n + m is the number of edges. As before, ρ has no fixed points, since all internal vertices of maps in N B(β, χ) have even valence larger than 2. A fixed-point free involution θ can be constructed from the labels that appear at the ends of each edge. Namely, in the examples shown in Figure 7 it is given by θ = (1b 3)(b 13)(2b 4)(b 24) b b b b b b b b b b b for a), θ = (13)(13)(25)(24)(46)(56) for b) and θ = (13)(13)(25)(24)(45) for c). We also define the hatted version of any permutation π by the equation π b(a) = π −1 (b a), b assuming b a = a. This is clearly an involution. Permutations that are invariant under this operation are called ‘palindromic’, such as (12b 2b 1) or (12)(b 2b 1). Any permutation that can be written as πb π where π is another permutation is automatically palindromic. Define two special fixed-point free involutions, and p1 = (1 b 1)(2 b 2) · · · , \ p2 = (b 1 2)(b 2 3) · · · (βb1 1)(β\ 1 + 1 β1 + 2) · · · (β1 + β2 β1 + 1) · · · . 15 (54) (55) Figure 7: Labelling some of the maps from Figures 1 and 2, in the way appropriate for the orthogonal case. Notice that the cycles of p1 contain labels which delimit corners of dashed line, while that the cycles of p2 contain labels which delimit corners of solid line. Notice also that they b factor the palindromic version of the vertex permutation, p2 p1 = Π Π. By construction, the permutation f1 = θp1 contains every other label encountered along boundary walks around the faces delimited by boundaries in dashed line. For example, in Figure 7 it would be f1 = (13)(24)(b 4b 2)(b 3b 1) for a), f1 = (13)(246b 5)(5b 6b 4b 2)(31) for b) and b b b f1 = (13)(245)(542)(31) for c). In particular, f1 is always palindromic. Conversely, permutation f2 = p2 θ contains every other label encountered along boundary walks around the faces delimited by boundaries in solid line. In Figure 7 it would be f2 = (14)(23)(b 3b 2)(b 4b 1) for a), f2 = (14)(2b 663)(b 3b 2)(b 4b 55b 1) for b) and f2 = (14)(2b 43)(b 3b 2)(b 1b 55) for c). Permutation f2 needs not be palindromic. Notice that f1 and f2 are factors for the b = f2 f1 . palindromic version of the vertex permutation, Π Π For the maps in N B(β, χ) all boundary walks visit the marked vertices, which means that all cycles of f1 and of f2 must have exactly one element in the set {1, ..., n, b 1, ..., n b}. Therefore, the permutations satisfy the conditions we listed in Section 1.3.2. When we label the vertices of the map to produce a factorization, the same ambiguities arise as for the unitary group, which are accounted for by division by the factor zρ . We have therefore arrived at factorizations in N F(β, χ) and proved our Theorem 4. Appendix - Other factorizations 4.4 Unitary case A monotone factorization of π is a sequence (τ1 , ..., τk ) of transpositions τi = (si ti ), ti > si and ti ≥ ti−1 , such that π = τ1 · · · τk . The number of transpositions, k, is the length of the factorization. Let Mαk be the number of length k monotone factorizations of π, with cycletype α. Using the theory of Jucys-Murphy elements, Matsumoto and Novak showed that ∞ X n+ℓ(α) WgU (α) = (−1) Mαk N −n−k . (56) N k=0 A proper factorization of π is a sequence of permutations (τ1 , ..., τk ), in which no one is the identity, such that P π = τ1 · · · τk . The depth of a proper factorization is te sum of the ranks of the factors, kj=1 r(τj ). Let Pαk,d be the number of proper factorizations of π, with cycletype α, having length k 16 and depth d. It is known that [55] Pαk,d   k  X Y χ (µ ) 1 X λ j |Cµj | χλ (1n )χλ (α) δPj r(µj ),d , = n)   n! χ (1 λ µ ···µ λ⊢n 1 k (57) j=1 where all partitions µ are different from 1n . Starting from the character expansion of the Weingarten function, 1 X χλ (1n )2 χλ (α), (58) WgU (α) = N n!2 sλ (1N ) λ⊢n Collins used [6] the Schur function expansion  n n X 1 χλ (1 )N  sλ (1N ) = |Cµ |χλ (µ)N ℓ(µ) = 1+ n! n! µ⊢n X |Cµ | µ⊢n,µ6=1n to arrive at the expression WgU N (α) =  χλ (µ) −r(µ)  N , χλ (1n ) ∞ k X X Y χλ (µj ) −r(µj ) 1 X n k N . |Cµj | χ (1 )χ (α) (−1) λ λ n) n!N n χ (1 λ µ ···µ λ⊢n k=1 1 k (59) (60) j=1 Comparing with (57) ones concludes that WgU N (α) ! ∞ d X X k k,d = N −n−d . (−1) Pα d=0 (61) k=1 A cycle factorization π is a sequence of permutations (τ1 , ..., τk ), in which all factors have only one cycle besides singletons, i.e. their cycletypes are hook partitions. Inequivalent cycle factorizations are equivalence classes of cycle factorizations, two factorizations being equivalent if they differ by the swapping of adjacent commuting factors. Berkolaiko and Irving show [25] that the number of such factorizations of π, with cycletype α, having length k and depth d, denoted by us Iαk,d , satisfy X X (−1)k Pαk,d = (−1)n+ℓ(α) Mαd . (62) (−1)k Iαk,d = k k These results are indexed by depth, but one can use Euler characteristic instead, by resorting to the equality χ = n + ℓ(α) − d. 4.5 Orthogonal case c Let h be the operation of ‘forgetting’ Consider again permutations acting on the set [n]∪ [n]. the hat, i.e. h(a) = h(b a) = a for all a ∈ [n]. Matsumoto defined the following analogue of monotone factorizations [31]. Let m be a matching and let (τ1 , ..., τk ) be a sequence of transpositions τi = (si ti ), in which all ti ∈ [n] with ti ≥ ti−1 and ti > h(si ), such that m = τ1 · · · τk (t), where t is the trivial matching. Let fk be the number of length k such factorizations of some m with cycletype β. Then, M α WgO N (β) = ∞ X k=0 fk N −n−k . (−1)k M β (63) The analogue of proper factorizations in this context is, for a permutation σ with cosettype β, a sequence of permutations (τ1 , ..., τk ), no one having the same cosettype as the 17 identity, such that σ = τ1 · · · τk . Let Peβk,d be the number of such factorizations having length k and depth d. We know from [28, 29, 30] that (actually these works only consider k = 2, but the extension to higher values of k is straightforward)   k   X Y X 1 |Kµj |ωλ (µj ) δPj r(µj ),d , (64) χ2λ (12n )ωλ (β) Peβk,d =   (2n)! µ ···µ λ⊢n 1 where ωλ (τ ) = k j=1 1 X χ2λ (τ ξ) 2n n! (65) ξ∈Hn are the zonal spherical function of the Gelfand pair (S2n , Hn ) (they depend only on the cosettype of τ ; see [38]). The relation between the above factorizations and orthogonal Weingarten functions comes as follows. The character-theoretic expression for the orthogonal Weingarten function is [8] 2n n! X χ2λ (1n ) ωλ (τ ), (66) WgO N (τ ) = (2n)! Zλ (1N ) λ⊢n where Zλ are zonal polynomials. Following the same procedure used for the unitary group, P we expand Zλ (1N ) = 2n1n! µ⊢n |Kµ |ωλ (µ)N ℓ(µ) to arrive at WgO N (τ ) k ∞ X Y |Kµj | 2n n! X χ2λ (1n )ωλ (τ ) X k (−1) ω (µ )N −r(µj ) . = n n n! λ j (2n)! N 2 µ ···µ λ⊢n k=1 1 k (67) j=1 Comparing with (64), we see that WgO N (τ ) = ∞ d X X d=0 k=1 ! (−1)k ek,d P N −n−d . (2n n!)k−1 β (68) Berkolaiko and Kuipers have provided a combinatorial description of the coefficients in the 1/N expansion of the function WgO N +1 [32] (they actually worked with the so-called Circular Orthogonal Ensemble of unitary symmetric matrices, but the Weingarten function of that ensemble coincides [13] with WgO N +1 ). A palindromic monotone factorization is a sequence (τ1 , ..., τk ) of transpositions τi = (si ti ), with ti > si and ti ≥ ti−1 , such that πb π= k c τ1 · · · τk τbk · · · τb1 . Let Mβ be the number of length k palindromic monotone factorizations of πb π , with π a permutation of cycletype β. Then, WgO N +1 (β) = ∞ X k=0 ck N −n−k . (−1)k M β (69) An appropriate analogue of inequivalent cycle factorizations is currently missing, but we conjecture that, whatever they are, their counting function will be related to coefficients in 1/N expansions of orthogonal Weingarten functions. References [1] D. Weingarten, Asymptotic behavior of group integrals in the limit of infinite rank. J. Math. Phys. 19, 999 (1978). [2] S. Samuel, U (N ) integrals, 1/N , and de Wit–’t Hooft anomalies. J. Math. Phys. 21, 2695 (1980). 18 [3] P. A. Mello, Averages on the unitary group and applications to the problem of disordered conductors. J. Phys. A: Math. Gen. 23 4061 (1990). [4] P. W. Brouwer and C.W.J. Beenakker, Diagrammatic method of integration over the unitary group, with applications to quantum transport in mesoscopic systems. J. Math. Phys. 37, 4904 (1996). [5] M. Degli Esposti and A. Knauf, On the form factor for the unitary group. J. Math. Phys. 45, 4957 (2004). [6] B. Collins, Moments and cumulants of polynomial random variables on unitary groups, the Itzykson-Zuber integral, and free probability. Int. Math. Res. Not. 17, 953 (2003). [7] B. Collins and P. Śniady, Integration with respect to the Haar measure on unitary, orthogonal and symplectic group. Commun. Math. Phys. 264, 773 (2006). [8] B. Collins and S. Matsumoto, On some properties of orthogonal Weingarten functions. J. Math. Phys. 50, 113516 (2009). [9] J.-B. Zuber, The large-N limit of matrix integrals over the orthogonal group, J. Phys. A: Math. Theor. 41, 382001 (2008). [10] T. Banica, B. Collins and J.-M. Schlenker, On polynomial integrals over the orthogonal group. J. Combinat. Theory A 118, 78 (2011). [11] S. Matsumoto and J. Novak, Jucys-Murphy elements and unitary matrix integrals. Int. Math. Research Not. 2, 362 (2013). [12] P. Zinn-Justin, Jucys-Murphy elements and Weingarten matrices. Lett. Math. Phys. 91, 119 (2010). [13] S. Matsumoto, Weingarten calculus for matrix ensembles associated with compact symmetric spaces. Random Matrices: Theory and Applications 2, 1350001 (2013). [14] A.J. Scott, Optimizing quantum process tomography with unitary 2-designs. J. Phys. A: Math. Theor. 41, 055308 (2008). [15] M. Žnidarič, C. Pineda and I. Garcı́a-Mata, Non-Markovian behavior of small and large complex quantum systems. Phys. Rev. Lett. 107, 080404 (2011). [16] M. Cramer, Thermalization under randomized local Hamiltonians. New J. Phys. 14, 053051 (2012). [17] Vinayak and M. Žnidarič, Subsystem’s dynamics under random Hamiltonian evolution. J. Phys. A: Math. Theor. 45, 125204 (2012). [18] M. Novaes, A semiclassical matrix model for quantum chaotic transport. J. Phys. A: Math. Theor. 46, 502002 (2013). [19] M. Novaes, Semiclassical matrix model for quantum chaotic transport with timereversal symmetry. Annals of Physics 361, 51 (2015). [20] P. Zinn-Justin and J.-B. Zuber, On some integrals over the U (N ) unitary group and their large N limit, J. Phys. A: Math. Gen. 36, 3173 (2003). [21] B. Collins, A. Guionnet and E. Maurel-Segala, Asymptotics of unitary and orthogonal matrix integrals, Advances in Mathematics 222, 172 (2009). 19 [22] M. Bousquet-Mélou, G. Schaeffer, Enumeration of planar constellations. Adv. in Appl. Math. 24, 337 (2000). [23] J. Irving, Minimal transitive factorizations of permutations into cycles. Canad. J. Math. 61, 1092 (2009). [24] J. Bouttier, Matrix integrals and enumeration of maps. Chapter 26 in The Oxford Handbook of Random Matrix Theory (Oxford, 2011), G. Akemann, J. Baik and P. Di Francesco (Editors). [25] G. Berkolaiko and arXiv:1405.5255v2. J. Irving, Inequivalent factorizations of permutations. [26] M. Bóna and B. Pittel, On the cycle structure of the product of random maximal cycles, arXiv:1601.00319v1. [27] O. Bernardi, A. Morales, R. Stanley and R. Du, Separation probabilities for products of permutations, Combinatorics, Probability and Computing 23, 201 (2014). [28] A.H. Morales and E.A. Vassilieva, Bijective evaluation of the connection coefficients of the double coset algebra, arXiv:1011.5001v1. [29] P.J. Hanlon, R.P. Stanley and J.R. Stembridge, Some combinatorial aspects of the spectra of normally distributed random matrices. Contemporary Mathematics 138, 151 (1992). [30] I.P. Goulden and D.M. Jackson, Maps in locally orientable surfaces, the double coset algebra, and zonal polynomials. Can. J. Math. 48, 569 (1996). [31] S. Matsumoto, Jucys–Murphy elements, orthogonal matrix integrals, and Jack measures. Ramanujan J. 26, 69 (2011). [32] G. Berkolaiko and J. Kuipers, Combinatorial theory of the semiclassical evaluation of transport moments I: Equivalence with the random matrix approach. J. Math. Phys. 54, 112103 (2013). [33] S. Müller, S. Heusler, P. Braun and F. Haake, Semiclassical approach to chaotic quantum transport. New Journal of Physics 9, 12 (2007). [34] S. Müller, S. Heusler, P. Braun, F. Haake and A. Altland, Periodic-orbit theory of universality in quantum chaos. Phys. Rev. E 72, 046207 (2005). [35] G. Berkolaiko and J. Kuipers, Combinatorial theory of the semiclassical evaluation of transport moments II: Algorithmic approach for moment generating functions. J. Math. Phys. 54, 123505 (2013). [36] M. Novaes, Semiclassical approach to universality in quantum chaotic transport. Europhys. Lett. 98, 20006 (2012). [37] M. Novaes, Combinatorial problems in the semiclassical approach to quantum chaotic transport. J. Phys. A: Math. Theor. 46, 095101 (2013). [38] I.G. Macdonald, Symmetric Functions and Hall Polynomials, 2nd edn. (Oxford University Press, Oxford, 1995). [39] J. Ginibre, Statistical Ensembles of Complex, Quaternion, and Real Matrices. J. Math. Phys. 6, 440 (1965). 20 [40] T.R. Morris, Chequered surfaces and complex matrices. Nuclear Physics B 356 (1991) 703. [41] G. ’t Hooft, A planar diagram theory for strong interactions. Nucl. Phys. B 72, 461 (1974). [42] D. Bessis, C. Itzykson and J.B. Zuber, Quantum field theory techniques in graphical enumeration. Adv. Appl. Math. 1, 109 (1980). [43] P. Di Francesco, Rectangular matrix models and combinatorics of colored graphs. Nuclear Physics B 648, 461 (2003). [44] M. Novaes, Statistics of time delay and scattering correlation functions in chaotic systems. II. Semiclassical Approximation. J. Math. Phys. 56, 062109 (2015). [45] W.A. Friedman and P.A. Mello, Marginal distribution of an arbitrary square submatrix of the S-matrix for Dyson’s measure. J. Phys. A: Math. Gen. 18, 425 (1985). [46] K. Życzkowski and H.-J. Sommers, Truncations of random unitary matrices. J. Phys. A: Math. Gen. 33, 2045 (2000). [47] Y.A. Neretin, Hua-type integrals over unitary groups and over projective limits of unitary groups. Duke Math. J. 114, 239 (2002). [48] P.J. Forrester, Quantum conductance problems and the Jacobi ensemble. J. Phys. A: Math. Gen. 39, 6861 (2006). [49] Y.V. Fyodorov and B.A. Khoruzhenko, A few remarks on colour-flavour transformations, truncations of random unitary matrices, Berezin reproducing kernels and Selbergtype integrals. J. Phys. A: Math. Theor. 40, 669 (2007). [50] J. Shen, On the singular values of Gaussian random matrices. Linear Algebra Appl. 326, 1 (2001). [51] A. Edelman and N.R. Rao, Random matrix theory. Acta Numerica 14, 233 (2005). [52] M.L. Mehta, Random Matrices, Chapter 17 (Academic Press, 2004). [53] P.J. Forrester and S.O. Warnaar, The importance of the Selberg integral. Bull. Amer. Math. Soc. 45, 489 (2008). [54] B.A. Khoruzhenko, H.-J. Sommers and K. Życzkowski, Truncations of random orthogonal matrices. Phys. Rev. E 82, 040106(R) (2010). [55] R.P. Stanley, Enumerative Combinatorics, vol. 2, Cambridge Univ. Press, Cambridge, 1999. 21 

A REMARK ON TORSION GROWTH IN HOMOLOGY AND VOLUME OF 3-MANIFOLDS arXiv:1802.09244v1 [math.GR] 26 Feb 2018 HOLGER KAMMEYER Abstract. We show that Lück’s conjecture on torsion growth in homology implies that two 3-manifolds have equal volume if the fundamental groups have the same set of finite quotients. The purpose of this note is to relate two well-known open problems which both deal with a residually finite fundamental group Γ of an odd-dimensional aspherical manifold. The first one [11, Conjecture 1.12(2)] predicts that the ℓ2 -torsion ρ(2) (Γ) determines the exponential rate at which torsion in middledegree homology grows along a chain of finite index normal subgroups. Conjecture A. Let M be an aspherical closed manifold of dimension 2d+1. Suppose that Γ = π1 M is residually finite and letTΓ = Γ0 ≥ Γ1 ≥ · · · be any chain of finite index normal subgroups of Γ with ∞ n=0 Γn = {1}. Then lim n→∞ log |Hd (Γn )tors | = (−1)d ρ(2) (Γ). [Γ : Γn ] The term |Hd (Γn )tors | denotes the order of the torsion subgroup of Hd (Γn ). The ℓ2 -torsion ρ(2) (Γ) is the ℓ2 -counterpart to Reidemeister torsion as surveyed in [12] and [7]. The second conjecture says that volume of 3-manifolds can be recovered from the finite quotients of the fundamental group. Conjecture B. Let Γ and Λ be infinite fundamental groups of connected, b∼ b Then closed, orientable, irreducible 3-manifolds and suppose that Γ = Λ. vol(Γ) = vol(Λ). b of Γ is the projective limit over all finite Here the profinite completion Γ quotients of Γ. Two groups have isomorphic profinite completions if and only if they have the same set of finite quotients [18, Corollary 3.2.8]. If Γ = π1 M for a 3-manifold M with the stated properties, then Thurston geometrization applies to M : there is a minimal choice of finitely many disjointly embedded incompressible tori in M , unique up to isotopy, which cut M into pieces such that each piece carries one out of eight geometries. The sum of the volumes of the hyperbolic pieces gives the well-defined quantity vol(Γ). Conjecture B is often stated as a question [2, Question 3.18]. But we dare to promote it to a conjecture in view of the following result. Theorem 1. Conjecture A implies Conjecture B. The theorem seems to be folklore among the experts in the field but I could not find a proof in the literature so that this note is meant as a service to the community. 2010 Mathematics Subject Classification. 20E18, 57M27. 1 2 HOLGER KAMMEYER The contrapositive of Theorem 1 says that constructing two profinitely isomorphic 3-manifold groups with differing covolume would disprove Conjecture A. Funar [4] and Hempel [6] constructed examples of closed 3-manifolds with non-isomorphic but profinitely isomorphic fundamental groups. These examples carry Sol and H2 ×R geometry, respectively, and thus all have zero volume by definition. Wilkes [19] showed that Hempel’s examples are the only ones among Seifert-fiber spaces. It seems to be open whether there exist such examples with H3 -geometry. As a first step in the negative direction, Bridson and Reid [3] showed that the figure eight knot group is determined among 3-manifold groups by the profinite completion. The paper at hand is divided into two sections. Section 1 presents the proof of Theorem 1. As a complement, Section 2 discusses how the related asymptotic volume conjecture and the Bergeron–Venkatesh conjecture fit into the picture. I wish to thank S. Kionke and J. Raimbault for helpful discussions during the junior trimester program “Topology” at the Hausdorff Research Institute for Mathematics in Bonn. 1. Proof of Theorem 1 For the moment, let Γ and Λ be any two finitely generated, residually finite groups. To prepare the proof of Theorem 1, we collect a couple of propositions from the survey article [17] and include more detailed proofs for the sake of a self-contained treatment. We first recall that the open b are precisely the subgroups of finite index. One direction is subgroups of Γ b easy: Γ is compact and the cosets of an open subgroup form a disjoint open cover. The converse is a deep theorem due to Nikolov and Segal [15] that crucially relies on the assumption that Γ is finitely generated. The proof moreover invokes the classification of finite simple groups. The assumption that Γ is residually finite says precisely that the canonical b is an embedding. If Q is a finite group, the universal property of map Γ → Γ b b Q) → Hom(Γ, Q) is a surjection. By Γ says that the restriction map Hom(Γ, ϕ b− the above, the kernel of any homomorphism Γ → Q is open which implies that ϕ is continuous and is thus determined by the values on the dense subset b Thus Hom(Γ, b Q) → Hom(Γ, Q) is in fact a bijection which clearly Γ ⊂ Γ. b Q) → Epi(Γ, Q) of surjective homomorphisms. restricts to a bijection Epi(Γ, This has the following consequence. b then there is an epimorphism Proposition 2. If Λ embeds densely into Γ, H1 (Λ) → H1 (Γ). Proof. Let p be a prime number which does not divide the group order |H1 (Λ)tors | and let us set r = dimQ H1 (Γ; Q). It is apparent that we have an epimorphism Γ → (Z/pZ)r ⊕ H1 (Γ)tors . By the above remarks, this b → (Z/pZ)r ⊕ H1 (Γ)tors . epimorphism extends uniquely to an epimorphism Γ b the latter map restricts to an epimorphism Since Λ embeds densely into Γ, r Λ → (Z/pZ) ⊕ H1 (Γ)tors . This epimorphism must lift to an epimorphism Λ → Zr ⊕ H1 (Γ)tors ∼ = H1 (Γ) because p is coprime to |H1 (Λ)tors |. Of course this last epimorphism factors through the abelianization H1 (Λ).  A REMARK ON TORSION GROWTH IN HOMOLOGY 3 b ∼ b then Corollary 3. The abelianization is a profinite invariant: if Γ = Λ, ∼ H1 (Γ) = H1 (Λ). Proof. Since we have surjections in both directions the groups H1 (Γ) and H1 (Λ) have the same free abelian rank. Thus either surjection restricts to an isomorphism of the free parts and thus induces a surjection of the finite torsion quotients—which then must be a bijection.  b called the profinite Let us now endow Γ with the subspace topology of Γ, topology of Γ. For the open subgroups of Γ we have the same situation as b we observed for Γ. Proposition 4. A subgroup H ≤ Γ is open in the profinite topology if and only if H has finite index in Γ. b carries the coarsest topology under which the projecProof. Recall that Γ b tions Γ → Γ/Γi for finite index normal subgroups Γi E Γ are continuous. b → Γ/Γi are the canonical projections, it Since the compositions Γ → Γ b is given by the follows that a subbase for the subspace topology of Γ ⊂ Γ cosets of finite index normal subgroups of Γ. T If H has finite index in Γ, then so does the normal core N = g∈Γ gHg−1 because N is precisely the kernel of the permutation representation S of Γ on the homogeneous set Γ/H defined by left translation. Thus H = h∈H hN is open. Conversely, let H ≤ Γ be open. Then H is a union of finite intersections of finite index normal subgroups of Γ. In particular H contains a finite index subgroup, whence has finite index itself.  b defines a 1-1–correspondence Proposition 5. Taking closure H 7→ H in Γ from the open (or finite index) subgroups of Γ to the open (or finite index) b The inverse is given by intersection H 7→ H ∩ Γ with Γ. subgroups of Γ. This correspondence preserves the index, sends a normal subgroup N E Γ to ∼ b and in the latter case we have Γ/N b a normal subgroup N E Γ, = Γ/N . The proof is given in [18, Prop. 3.2.2, p. 84]. Here is an easy consequence. Corollary 6. For H1 , H2 ≤ Γ of finite index we have H1 ∩ H2 = H1 ∩ H2 . b and we get Proof. By the proposition H1 ∩ H2 has finite index in Γ (H1 ∩ H2 ) ∩ Γ = (H1 ∩ Γ) ∩ (H2 ∩ Γ) = H1 ∩ H2 . Applying the proposition again yields H1 ∩ H2 = H1 ∩ H2 .  Note that for a finitely generated, residually finite group Γ there is a canonical choice of a chain Γ = M1 ≥ M2 ≥ M3 ≥ · · · T of finite index normal subgroups Mn E Γ satisfying ∞ n=1 Mn = {1}. Simply define Mn to be the intersection of the (finitely many!) normal subgroups of index at most n. By the last two results, Mn is the intersection of all b with index at most n. normal subgroups of Γ T Proposition 7. The intersection ∞ n=1 Mn is trivial. 4 HOLGER KAMMEYER T b Proof. Let g ∈ ∞ n=1 Mn ⊂ Γ. Since Γ is finitely generated, it has only countably many subgroups of finite index. Therefore the description of the topolb given above shows that Γ b is second and thus first countable. Hence ogy of Γ b with limi→∞ gi = g. we can pick a sequence (gi ) from the dense subset Γ ⊂ Γ b b Let pn : Γ → Γ/Mn and pbn : Γ → Γ/Mn denote the canonical projections. Since pbn is continuous, we have Mn = pbn (g) = lim pbn (gi ) i→∞ ∼ b and hence limi→∞ pn (gi ) = Mn ∈ Γ/Mn because Γ/M n = Γ/Mn by Proposition 5. As Γ/Mn is discrete, the sequence pn (gi ) is eventually constant. This means that for all n ≥ 1 there is N ≥ 1 such that for all i ≥ N we have pn (gi ) = Mn , or equivalently gi ∈ Mn . But the open sets Mn form a neighborhood basis of 1 ∈ Γ as follows from the description of the profinite topology of Γ given in the proof of Proposition 4. So the last statement gives b (and hence Γ) is Hausdorff, we conclude g = 1.  limi→∞ gi = 1. Since Γ b is residually finite as an abstract group. Before we give It follows that Γ the proof of Theorem 1, we put down one more observation. If H ≤ Γ is any b is a profinite group so that the universal subgroup, then the closure H in Γ b b → H which restricts property of H gives a canonical homomorphism η : H to the identity on H. This is always an epimorphism because the image is dense, as it contains H, and closed because it is compact and H is Hausdorff. However, in general we cannot expect that η is injective, not even if H is finitely generated. Nevertheless: b → Proposition 8. If H ≤ Γ has finite index, then the canonical map η : H H is an isomorphism. Proof. Let h ∈ ker η. The group H is finitely generated because it is a finite b is second and hence index subgroup of Γ. As above we conclude that H b we can thus pick a sequence of first countable. Since H lies densely in H, elements hi ∈ H such that limi→∞ hi = h. By continuity of η, we obtain limi→∞ η(hi ) = η(h) = 1 and thus limi→∞ hi = 1 in the topology of H. A neighborhood basis of 1 ∈ H is given by the sets Mn ∩ H where Mn are the b from above. It follows that for all n ≥ 1 finite index normal subgroups of Γ there exists N ≥ 1 such that for all i ≥ N we have hi ∈ Mn ∩H. Since H has finite index in Γ, it follows that any finite index normal subgroup K E H has also finite index as a subgroup of Γ. Thus there exists n ≥ 1 such that Mn lies in the normal core of K as a subgroup of Γ. Hence for all K E H of finite index there exists N ≥ 1 such that for all i ≥ N we have hi ∈ K. But the finite index normal subgroups K E H form a neighborhood basis of 1 ∈ H in the profinite topology of H. Hence we have limi→∞ hi = 1 in b Since H b is Hausdorff, we conclude h = 1. the topology of H.  Proof of Theorem 1. Note that Γ and Λ are finitely generated and residually b∼ b finite, as a consequence of geometrization [5]. We fix an isomorphism Γ = Λ. Again, let Mn ≤ Γ be the intersection of all normal subgroups of Γ of index at most n. By Proposition 5 it follows that Ln = Λ ∩ Mn is the intersection of all normal subgroups of Λ of index at most n and [Γ : Mn ] = A REMARK ON TORSION GROWTH IN HOMOLOGY 5 T T [Λ : Ln ]. By Proposition 7 we have n Mn = {1} so that n Ln = {1}. From dn ∼ cn so that Corollary 3 implies |H1 (Mn )tors | = Proposition 8 we get M =L |H1 (Ln )tors |. A theorem of Lück and Schick [13, Theorem 0.7] conjectured in Lott and Lück [9, Conjecture 7.7] shows that ρ(2) (Γ) = − vol(Γ)/6π and similarly for Λ, see also [12, Theorem 4.3, p. 216]. If Conjecture A holds true, this implies log |H1 (Ln )tors | log |H1 (Mn )tors | = 6π lim = vol(Λ).  n→∞ n→∞ [Γ : Mn ] [Λ : Ln ] vol(Γ) = 6π lim 2. Related conjectures One can find companion conjectures to Conjecture A in the literature which likewise predict an exponential rate of torsion growth in homology proportional to volume. However, these conjectures restrict the aspherical manifolds under consideration in one way or another. Specifically dealing with 3-manifolds is Lê’s asymptotic volume conjecture. Conjecture C. Let Γ be the fundamental group of a connected, orientable, irreducible, compact 3-manifold whose boundary is either empty or a collection of tori. Then lim sup Γn →{1} vol(Γ) log |H1 (Γn )tors | = . [Γ : Γn ] 6π The conjecture appears in [8, Conjecture 1 (a)]. The volume vol(Γ) is defined by a geometric decomposition as before which also exists for toroidal boundary. The lim sup on the left hand side is defined as the lowest upper bound of all lim sups along sequences (Γn ) of (not necessarily nested!) finite index normal subgroups of Γ with lim supn Γn = {1}. Recall that by definition \ [ lim supn Γn = Γn N ≥0 n≥N so that the condition lim supn Γn = {1} is actually equivalent to requiring ( 1 if g = e, lim trC[Γ/Γn ] (gΓn ) = trC[Γ] (g) = n→∞ 0 otherwise, for all g ∈ Γ where the traces are the usual traces of group algebras given by the unit coefficient. Question 9. Does Conjecture C imply Conjecture B? The proof of Theorem 1 does not immediately carry over to Question 9 as lim supn Γn = {1} for some sequence (Γn ) does not imply lim sup Λn = {1} for the groups Λn = Λ ∩ Γn . Here is an example. Example 10. Let Γ = Z×Z with (nested) chain of subgroups Γn = 2n Z×3nZ. Q b=Z b × Z. b From the description Z b∼ Clearly, we have Γ = p Zp it is apparent b : N Z] b = N for any N ≥ 1. Since N Z is the only subgroup of index N that [Z 6 HOLGER KAMMEYER b Thus we have Γn = 2n Z b × 3n Z. b in Z, Proposition 5 implies that N Z = N Z. It follows that     ∞ \ Y Y b × Z. b Γn ∼ Zp  × Z2 × {0} × Zq  ≤ Z = {0} × n=1 p>2 q>3 b be the subgroup generated by the two elements So if we let Λ ≤ Γ ((0, 1, 1, . . .), (1, 0, 0, 0, . . .)) and ((1, 0, 0, 0 . . .), (0, 1, 1, 1, . . .)) Q b × Z, b then clearly Λ ∼ b so that in p Zp × p Zp ∼ =Z = Z × Z is dense in Γ b→Λ=Γ b is a surjective homomorphism of isomorphic the canonical map Λ finitely generated profinite groups. Hence it T must be an isomorphism [18, Proposition 2.5.2, p. 46]. However, we have ∞ n=1 Λn 6= {0} even though T ∞ Γ = {0}. n n=1 Q We remark that Lê has proven the inequality “≤” of Conjecture C, even if the subgroups are not required to be normal. Another conjecture, which leaves both the realm of 3-manifolds and of normal subgroups, is due to Bergeron and Venkatesh [1, Conjecture 1.3]. It does however assume a somewhat rigorous arithmetic setting. This is what we want to present next. Let G be a semisimple algebraic group, defined and anisotropic over Q. Let Γ ≤ G(Q) be a congruence subgroup. This means that for some (and then for any) Q-embedding ρ : G → GLn there is k ≥ 1 such that the group ρ(Γ) contains the kernel of ρ(G) ∩ GLn (Z) → GLn (Z/kZ) as a subgroup of finite index. Fix an algebraic representation of G on a finite-dimensional Q-vector space W and let M ⊂ W be a Γ-invariant Z-lattice, which always exists according to [16, Remark, p. T 173]. Let Γ = Γ0 ≥ Γ1 ≥ · · · be a chain of congruence subgroups with n Γn = {1}. For a maximal compact subgroup K of G = G(R), we denote by X = G/K the symmetric space associated with G. Let g and k be the Lie algebras of G and K and let δ(G) = rankC g ⊗ C − rankC k ⊗ C be the deficiency of G, sometimes also known as the fundamental rank δ(X) of X. Conjecture D. For each d ≥ 1 there is a constant cG,M,d ≥ 0 such that log |Hd (Γn ; M )tors | = cG,M,d vol(Γ) n→∞ [Γ : Γn ] lim and cG,M,d > 0 if and only if δ(G) = 1 and dim X = 2d + 1. In this case the volume vol(Γ) is the volume of the closed locally symmetric space Γ\X which is defined by means of a Haar measure on G and as such only unique up to scaling. But any rescaling of this measure would also rescale the constant cG,M,d by the reciprocal value so that the product is well-defined. To make sure that cG,M,d really only depends on G, M , and d, we agree upon the following normalization of the Haar measure. The Killing form on g restricts to a positive definite form on the subspace p in the orthogonal Cartan decomposition g = k ⊕ p. Identifying p with the tangent space TK X, we obtain a G-invariant metric on X by translation. We require that the volume of Γ\X determined by Haar measure be equal to the volume of Γ\X as Riemannian manifold. A REMARK ON TORSION GROWTH IN HOMOLOGY 7 To relate Conjecture D to Conjecture B, we need to restrict our attention to arithmetic hyperbolic 3-manifolds. These are quotients of hyperbolic 3-space H3 by arithmetic Kleinian groups. A Kleinian group is a discrete subgroup Γ ≤ PSL(2, C) ∼ = Isom+ (H3 ) such that vol(Γ) = vol(Γ\H3 ) < ∞. A Kleinian group Γ ≤ PSL(2, C) is called arithmetic if there exists a semisimple linear algebraic Q-group H ≤ GLn and an epimorphism of Lie groups φ : H(R)0 → PSL(2, C) with compact kernel such that Γ is commensurable with φ(H(Z) ∩ H(R)0 ). Here H(R)0 denotes the unit component and two subgroups of a third group are called commensurable if their intersection has finite index in both subgroups. Note that we consider PSL(2, C) as a real Lie group so that the complexified lie algebra is sl(2, C) ⊕ sl(2, C) and hence δ(PSL(2, C)) = 1. There is an alternative and equivalent approach to the definition of arithmetic Kleinian groups via orders in quaternion algebras over number fields [14]. b = Λ. b Question 11. Let Γ and Λ be arithmetic Kleinian groups such that Γ Suppose Conjecture D holds true. Can we conclude that vol(Γ) = vol(Λ)? Again, various problems arise when trying to adapt the proof of Theorem 1 to settle this question in the affirmative. To be more concrete, a direct translation fails for the following reason. Let Mn be the intersection of all normal subgroups of index at most n in the arithmetic group H(Z) corresponding to Γ as above. Then Mn will not consist of congruence subgroups. In fact, H(Z) has the congruence subgroup property if and only if all the groups Mn are congruence subgroups. But the congruence subgroup property is well known to fail for all arithmetic Kleinian groups [10]. Instead, one could try to start with a chain of congruence subgroups Γn of Γ but then it seems unclear if or under what circumstances the chain Λn = Γn ∩ Λ consists of congruence subgroups in Λ. We remark that for the trivial coefficient system Z ⊂ Q, Conjecture D is wide open. However, in our relevant case of δ(G) = 1, Bergeron and Venkatesh construct strongly acyclic coefficient modules M with the property that the spectrum of the Laplacian acting on M ⊗Z C-valued p-forms on Γn \X is bounded away from zero for all p and n. In the special case G = SL(2, C), they show that Conjecture D holds true for any strongly acyclic M . References [1] N. Bergeron and A. Venkatesh, The asymptotic growth of torsion homology for arithmetic groups, J. Inst. Math. Jussieu 12 (2013), no. 2, 391–447. MR 3028790 ↑6 [2] M. Boileau and S. Friedl, Profinite completions and 3-manifold groups, RIMS Kokyuroku 1991 (2016), 54–68. http://hdl.handle.net/2433/224629. ↑1 [3] M. Bridson and A. Reid, Profinite rigidity, fibering, and the figure-eight knot (2015). e-print. arXiv:1505.07886 ↑2 [4] L. Funar, Torus bundles not distinguished by TQFT invariants, Geom. Topol. 17 (2013), no. 4, 2289–2344. With an appendix by Funar and Andrei Rapinchuk. MR 3109869 ↑2 [5] J. Hempel, Residual finiteness for 3-manifolds, Combinatorial group theory and topology (Alta, Utah, 1984), Ann. of Math. Stud., vol. 111, Princeton Univ. Press, Princeton, NJ, 1987, pp. 379–396. MR 895623 ↑4 [6] , Some 3-manifold groups with the same finite quotients (2014). e-print. arXiv:1409.3509 ↑2 8 HOLGER KAMMEYER [7] H. Kammeyer, Introduction to ℓ2 -invariants, lecture notes, 2018. Available for download at http://topology.math.kit.edu/21_679.php. ↑1 [8] T. T. Q. Lê, Growth of homology torsion in finite coverings and hyperbolic volume (2014). e-print. arXiv:1412.7758 ↑5 [9] J. Lott and W. Lück, L2 -topological invariants of 3-manifolds, Invent. Math. 120 (1995), no. 1, 15–60. MR 1323981 ↑5 [10] A. Lubotzky, Group presentation, p-adic analytic groups and lattices in SL2 (C), Ann. of Math. (2) 118 (1983), no. 1, 115–130. MR 707163 ↑7 [11] W. Lück, Approximating L2 -invariants and homology growth, Geom. Funct. Anal. 23 (2013), no. 2, 622–663. MR 3053758 ↑1 [12] , L2 -invariants: theory and applications to geometry and K-theory, Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics, vol. 44, Springer-Verlag, Berlin, 2002. MR 1926649 ↑1, 5 [13] W. Lück and T. Schick, L2 -torsion of hyperbolic manifolds of finite volume, Geom. Funct. Anal. 9 (1999), no. 3, 518–567. MR 1708444 ↑5 [14] C. Maclachlan and A. W. Reid, The arithmetic of hyperbolic 3-manifolds, Graduate Texts in Mathematics, vol. 219, Springer-Verlag, New York, 2003. MR 1937957 ↑7 [15] N. Nikolov and D. Segal, On finitely generated profinite groups. I. Strong completeness and uniform bounds, Ann. of Math. (2) 165 (2007), no. 1, 171–238. MR 2276769 ↑2 [16] V. Platonov and A. Rapinchuk, Algebraic groups and number theory, Pure and Applied Mathematics, vol. 139, Academic Press, Inc., Boston, MA, 1994. Translated from the 1991 Russian original by Rachel Rowen. MR 1278263 ↑6 [17] A. W. Reid, Profinite properties of discrete groups, Groups St Andrews 2013, London Math. Soc. Lecture Note Ser., vol. 422, Cambridge Univ. Press, Cambridge, 2015, pp. 73–104. MR 3445488 ↑2 [18] L. Ribes and P. Zalesskii, Profinite groups, Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics, vol. 40, SpringerVerlag, Berlin, 2000. MR 1775104 ↑1, 3, 6 [19] G. Wilkes, Profinite rigidity for Seifert fibre spaces, Geom. Dedicata 188 (2017), 141– 163. MR 3639628 ↑2 Karlsruhe Institute of Technology, Institute for Algebra and Geometry, Germany E-mail address: holger.kammeyer@kit.edu URL: www.math.kit.edu/iag7/~kammeyer/ 

On automorphisms of finite p-groups Hemant Kalra and Deepak Gumber School of Mathematics arXiv:1803.07853v1 [math.GR] 21 Mar 2018 Thapar Institute of Engineering and Technology, Patiala - 147 004, India emails: happykalra26@gmail.com, dkgumber@gmail.com It is proved in [J. Group Theory, 10 (2007), 859-866] that if G is a finite p-group such that (G, Z(G)) is a Camina pair, then |G| divides | Aut(G)|. We give a very short and elementary proof of this result. 2010 Mathematics Subject Classification: 20D15, 20D45. Keywords: Camina pair, class-preserving automorphism. 1 Introduction Let G be a finite non-abelian p-group. The problem “Does the order, if it is greater than p2 , of a finite non-cyclic p-group divide the order of its automorphism group?” is a well-known problem [6, Problem 12.77] in finite group theory. Gaschütz [4] proved that any finite p-group of order at least p2 admits a non-inner automorphism of order a power of p. It follows that the problem has an affirmative answer for finite p-groups with center of order p. This immediately answers the problem positively for finite p-groups of maximal class. Otto [7] also gave an independent proof of this result. Fouladi et al. [3] gave a supportive answer to the problem for finite p-groups of co-class 2. For more details on this problem, one can see the introduction in the paper of Yadav [8]. In [8, Theorem A], Yadav proved that if G is a finite p-group such that (G, Z(G)) is a Camina pair, then |G| divides | Aut(G)|. He also proved the important result [8, Corollary 4.4] that the group of all class-preserving outer automorphisms is non-trivial for finite p-groups G with (G, Z(G)) a Camina pair. In this paper, we give different and very short proofs of these results of Yadav using elementary arguments. Let G be a finite p-group. Then (G, Z(G)) is called a Camina pair if xZ(G) ⊆ xG for all x ∈ G − Z(G), where xG denotes the conjugacy class of x in G. In particular, if (G, G′ ) is a Camina pair, then G is called a Camina p-group. 2 Proofs We shall need the following lemma which is a simple modification of a lemma of Alperin [1, Lemma 3]. Lemma 2.1. Let G be any group and B be a central subgroup of G contained in a normal subgroup A of G. Then the group AutB A (G)of all automorphisms of G that induce the identity on both A and G/B is isomorphic onto Hom(G/A, B). Theorem 2.2. Let G be a finite p-group such that (G, Z(G)) is a Camina pair. Then |G| divides | Aut(G)|. Proof. Observe that Z(G) ≤ G′ ≤ Φ(G) and, therefore, Z(G) ≤ Z(M ) for every maximal subgroup M of G. Suppose that Z(G) < Z(M1 ) for some maximal subgroup M1 of G. Let G = M1 hg1 i, where g1 ∈ G − M1 and g1p ∈ M1 . Let g ∈ Z(M1 ) − Z(G). Then |Z(G)| ≤ |[g, G]| = |[g, M1 hg1 i]| = |[g, hg1 i]| ≤ p 1 implies that |Z(G)| = p. The result therefore follows by Gaschütz [4]. We therefore suppose that Z(G) = Z(M ) for every maximal subgroup M of G. We prove that CG (M ) ≤ M . Assume that there exists g0 ∈ CG (M0 ) − M0 for some maximal subgroup M0 of G. Then G = M0 hg0 i and thus g0 ∈ Z(G), because g0 commutes with M0 . This is a contradiction because Z(G) ≤ Φ(G). Therefore CG (M ) ≤ M for Z(G) every maximal subgroup M of G. Consider the group AutM (G) which is isomorZ(G) phic to Hom(G/M, Z(G)) by Lemma 2.1. It follows that AutM (G) is non-trivial. Z(G) Let α ∈ AutM (G) ∩ (Inn(G)). Then α is an inner automorphism induced by some g ∈ CG (M ) = Z(M ). Since Z(G) = Z(M ), α is trivial. It follows that Z(G) |(AutM Z(G) (G))(Inn(G))| = |(AutM (G))||(Inn(G))| = |Z(G)||G/Z(G)| = |G|, because Z(G) is elementary abelian by Theorem 2.2 of [5]. This completes the proof. Corollary 2.3. Let G be a finite Camina p-group. Then |G| divides | Aut(G)|. Proof. It is a well known result [2] that nilpotence class of G is at most 3. Also, it follows from [5, Lemma 2.1, Theorem 5.2, Corollary 5.3] that (G, Z(G)) is a Camina pair. The result therefore follows from Theorem 2.2. An automorphism α of G is called a class-preserving automorphism of G if α(x) ∈ xG for each x ∈ G. The group of all class-preserving automorphisms of G is denoted by Autc (G). An automophism β of G is called a central automorphism if x−1 β(x) ∈ Z(G) for each x ∈ G. It is easy to see that if (G, Z(G)) is a Camina pair, then the group of all central automorphisms fixing Z(G) element-wise is contained in Autc (G). Remark 2.4. It follows from the proof of Theorem 2.2 that if G is a finite p-group such that (G, Z(G)) is a Camina pair and |Z(G)| ≥ p2 , then Z(G) | Autc (G)| ≥ |(AutM (G))(Inn(G))| = |G|. Thus, in particular, we obtain the following result of Yadav [8]. Corollary 2.5 ([8, Corollary 4.4]). Let G be a finite p-group such that (G, Z(G)) is a Camina pair and |Z(G)| ≥ p2 . Then Autc (G)/ Inn(G) is non-trivial. The following example shows that Remark 2.4 is not true if |Z(G)| = p. Example 2.6. Consider a finite p-group G of nilpotence class 2 such that (G, Z(G)) is a Camina pair and |Z(G)| = p. Since cl(G) = 2, exp(G/Z(G)) = exp(G′ ) and hence G′ = Z(G) = Φ(G). Let |G| = pn , where n ≥ 3, and let {x1 , x2 , . . . , xn−1 } be the minimal generating set of G. Then | Autc (G)| ≤ n−1 Y |xi G | = pn−1 = |G/Z(G)|. i=1 Acknowledgment: Research of first author is supported by Thapar Institute of Engineering and Technology and also by SERB, DST grant no. MTR/2017/000581. Research of second author is supported by SERB, DST grant no. EMR/2016/000019. References [1] J. L. Alperin, Groups with finitely many automorphisms, Pacific J. Math., 12 (1962), 1-5. 2 [2] R. Dark and C. M. Scoppola, On Camina groups of prime power order, J. Algebra, 181 (1996), 787-802. [3] S. Fouladi, A. R. Jamali and R. Orfi, Automorphism groups of finite p-groups of co-class 2, J. Group Theory, 10 (2007), 437-440. [4] W. Gaschütz, Nichtabelsche p-Gruppen besitzen aussere p-Automorphismen, J. Algebra, 4 (1966), 1-2. [5] I. D. Macdonald, Some p-groups of Frobenius and extra-special type, Israel J. Math., 40 (1981), 350-364. [6] V. D. Mazurov and E. I. Khukhro, The Kourovka notebook, Unsolved problems in group theory, 18th augmented edn. (Russian Academy of Sciences Siberian Division, Institute of Mathematics, 2014). Also available at ArXiv. [7] A. D. Otto, Central automorphisms of a finite p-group, Trans. Amer. Math. Soc., 125 (1966), 280-287. [8] M. K. Yadav, On automorphisms of finite p-groups, J. Group Theory, 10 (2007), 859-866. 3 

Faster Information Gathering in Ad-Hoc Radio Tree Networks Marek Chrobak∗ Kevin Costello† arXiv:1512.02179v1 [cs.DS] 7 Dec 2015 December 8, 2015 Abstract We study information gathering in ad-hoc radio networks. Initially, each node of the network has a piece of information called a rumor, and the overall objective is to gather all these rumors in the designated target node. The ad-hoc property refers to the fact that the topology of the network is unknown when the computation starts. Aggregation of rumors is not allowed, which means that each node may transmit at most one rumor in one step. We focus on networks with tree topologies, that is we assume that the network is a tree with all edges directed towards the root, but, being ad-hoc, its actual topology is not known. We provide two deterministic algorithms for this problem. For the model that does not assume any collision detection nor acknowledgement mechanisms, we give an O(n log log n)-time algorithm, improving the previous upper bound of O(n log n). We also show that this running time can be further reduced to O(n) if the model allows for acknowledgements of successful transmissions. 1 Introduction We study the problem of information gathering in ad-hoc radio networks. Initially, each node of the network has a piece of information called a rumor, and the objective is to gather all these rumors, as quickly as possible, in the designated target node. The nodes communicate by sending messages via radio transmissions. At any time step, several nodes in the network may transmit. When a node transmits a message, this message is sent immediately to all nodes within its range. When two nodes send their messages to the same node at the same time, a collision occurs and neither message is received. Aggregation of rumors is not allowed, which means that each node may transmit at most one rumor in one step. The network can be naturally modeled by a directed graph, where an edge (u, v) indicates that v is in the range of u. The ad-hoc property refers to the fact that the actual topology of the network is unknown when the computation starts. We assume that nodes are labeled by integers 0, 1, ..., n − 1. An information gathering protocol determines a sequence of transmissions of a node, based on its label and on the previously received messages. Our results. In this paper, we focus on ad-hoc networks with tree topologies, that is the underlying ad-hoc network is assumed to be a tree with all edges directed towards the root, although the actual topology of this tree is unknown. ∗ Department of Computer Science, University of California at Riverside, USA. Research supported by NSF grants CCF-1217314 and CCF-1536026. † Department of Mathematics, University of California at Riverside, USA. Research supported by NSA grant H98230-13-1-0228 1 We consider two variants of the problem. In the first one, we do not assume any collision detection or acknowledgment mechanisms, so none of the nodes (in particular neither the sender nor the intended recipient) are notified about a collision after it occurred. In this model, we give a deterministic algorithm that completes information gathering in time O(n log log n). Our result significantly improves the previous upper bound of O(n log n) from [5]. To our knowledge, no lower bound for this problem is known, apart from the trivial bound of Ω(n) (since each rumor must be received by the root in a different time step). In the second part of the paper, we also consider a variant where acknowledgments of successful transmissions are provided to the sender. All the remaining nodes, though, including the intended recipient, cannot distinguish between collisions and absence of transmissions. Under this assumption, we show that the running time can be improved to O(n), which is again optimal for trivial reasons, up to the implicit constant. While we assume that all nodes are labelled 0, 1, ..., n − 1 (where n is the number of vertices), our algorithms’ asymptotic running times remain the same if the labels are chosen from a larger range 0, 1, ..., N − 1, as long as N = O(n). Related work. The problem of information gathering for trees was introduced in [5], where the model without any collision detection was studied. In addition to the O(n log n)-time algorithm without aggregation – that we improve in this paper – [5] develops an O(n)-time algorithm for the model with aggregation, where a message can include any number of rumors. Another model studied in [5], called fire-and-forward, requires that a node cannot store any rumors; a rumor received by a node has to be either discarded or immediately forwarded. For fire-and-forward protocols, a tight bound of Θ(n1.5 ) is given in [5]. The information gathering problem is closely related to two other information dissemination primitives that have been well studied in the literature on ad-hoc radio networks: broadcasting and gossiping. All the work discussed below is for ad-hoc radio networks modeled by arbitrary directed graphs, and without any collision detection capability. In broadcasting, a single rumor from a specified source node has to be delivered to all other nodes in the network. The naïve RoundRobin algorithm (see the next section) completes broadcasting in time O(n2 ). Following a sequence of papers [6, 18, 2, 3, 21, 11] where this naïve bound was gradually improved, it is now known that broadcasting can be solved in time O(n log D log log(D∆/n)) [10], where D is the diameter of G and ∆ is its maximum in-degree. This nearly matches the lower bound of Ω(n log D) from [9]. Randomized algorithms for broadcasting have also been well studied [1, 19, 11]. The gossiping problem is an extension of broadcasting, where each node starts with its own rumor, and all rumors need to be delivered to all nodes in the network. The time complexity of deterministic algorithms for gossiping is a major open problem in the theory of ad-hoc radio networks. Obviously, the lower bound of Ω(n log D) for broadcasting [9] applies to gossiping as well, but no better lower bound is known. It is also not known whether gossiping can be solved in time O(n polylog(n)) with a deterministic algorithm, even if message aggregation is allowed. The best currently known upper bound is O(n4/3 log4 n) [16] (see [6, 25] for some earlier work). The case when no aggregation is allowed (or with limited aggregation) was studied in [4]. Randomized algorithms for gossiping have also been well studied [11, 20, 7]. Interested readers can find more information about gossiping in the survey paper [15]. Connections to other problems. This research, as well as the earlier work in [5], was motivated by the connections between information gathering in trees and other problems in distributed 2 computing involving shared channels, including gossiping in radio networks and MAC contention resolution. For arbitrary graphs, assuming aggregation, one can solve the gossiping problem by running an algorithm for information gathering and then broadcasting all rumors (as one message) to all nodes in the network. Thus an O(n polylog(n))-time algorithm for information gathering would resolve in positive the earlier-discussed open question about the complexity of gossiping. Due to this connection, developing an O(n polylog(n))-time algorithm for information gathering on arbitrary graphs is likely to be very difficult – if possible at all. We hope that developing efficient algorithms for trees, or for some other natural special cases, will ultimately lead to some insights helpful in resolving the complexity of the gossiping problem in arbitrary graphs. Some algorithms for ad-hoc radio networks (see [4, 17], for example) involve constructing a spanning subtree of the network and disseminating information along this subtree. Better algorithms for information gathering on trees may thus be useful in addressing problems for arbitrary graphs. The problem of contention resolution for multiple-access channels (MAC) has been widely studied in the literature. (See, for example, [12, 22, 14] and the references therein.) There are in fact myriad of variants of this problem, depending on the characteristics of the communication model. Generally, the instance of the MAC contention resolution problem involves a collection of transmitters connected to a shared channel (e.g. ethernet). Some of these transmitters need to send their messages across the channel, and the objective is to design a distributed protocol that will allow them to do that. The information gathering problem for trees is in essence an extension of MAC contention resolution to multi-level hierarchies of channels, where transmitters have unique identifiers, and the structure of this hierarchy is not known. 2 Preliminaries We now provide a formal definition of our model and introduce notation, terminology, and some basic properties used throughout the paper. Radio networks with tree topology. In the paper we focus exclusively on radio networks with tree topologies. Such a network will be represented by a tree T with root r and with n = |T | nodes. The edges in T are directed towards the root, representing the direction of information flow: a node can send messages to its parent, but not to its children. We assume that each node v ∈ T is assigned a unique label from [n] = {0, 1, ..., n − 1}, and we denote this label by label(v). For a node v, by deg(v) we denote the degree of v, which is the number of v’s children. For any subtree X of T and a node v ∈ X, we denote by Xv the subtree of X rooted at v that consists of all descendants of v in X. For any integer γ = 1, 2, ..., n − 1 and any node v of T define the γ-height of v as follows. If v is a leaf then the γ-height of v is 0. If v is an internal node then let g be the maximum γ-height of a child of v. If v has fewer than γ children of γ-height equal g then the γ-height of v is g. Otherwise, the γ-height of v is g + 1. The γ-height of v will be denoted by heightγ (v). In case when more than one tree are under consideration, to resolve potential ambiguity we will write heightγ (v, T ) for the γ-height of v in T . The γ-height of a tree T , denoted heightγ (T ), is defined as heightγ (r), that is the γ-height of its root. Its name notwithstanding, the definition of γ-height is meant to capture the “bushiness” of a tree. For example, if T is just a path then its γ-height is equal 0 for each γ ≥ 2. The concept of γ-height generalizes Strahler numbers [23, 24], introduced in hydrology to measure the size of 3 2 1 0 0 2 1 0 1 0 1 0 0 0 0 0 1 0 1 0 1 0 0 0 0 0 0 0 0 Figure 1: An example showing a tree and the values of 3-heights for all its nodes. streams in terms of the complexity of their tributaries. Figure 1 gives an example of a tree and values of 3-heights for all its nodes. The lemma below is a slight refinement of an analogous lemma in [5], and it will play a critical role in our algorithms. Lemma 1. Suppose that T has q leaves, and let 2 ≤ γ ≤ q. Then heightγ (T ) ≤ logγ q. Equivalently, any tree having height j must have at least γ j leaves. This can be seen by induction on j – if v is a vertex which is furthest from the room among all vertices of height j, then v by definition has γ descendants of height j − 1, each of which has γ j−1 leaf descendants by inductive hypothesis. Information gathering protocols. Each node v of T has a label (or an identifier) associated with it, and denoted label(v). When the computation is about to start, each node v has also a piece of information, ρv , that we call a rumor. The computation proceeds in discrete, synchronized time steps, numbered 0, 1, 2, .... At any step, v can either be in the receiving state, when it listens to radio transmissions from other nodes, or in the transmitting state, when it is allowed to transmit. When v transmits at a time t, the message from v is sent immediately to its parent in T . As we do not allow rumor aggregation, this message may contain at most one rumor, plus possibly O(log n) bits of other information. If w is v’s parent, w will receive v’s message if and only if w is in the receiving state and no collision occurred, that is if no other child of w transmitted at time t. In Sections 3 and 4 we do not assume any collision detection nor acknowledgement mechanisms, so if v’s message collides with a message from one of its siblings, neither v nor w receive any notification. (In other words, w cannot distinguish collisions between its children’s transmissions from background noise.) We relax this requirement in Section 5, by assuming that v (and only v) will obtain an acknowledgment from w after each successful transmission. The objective of an information gathering protocol is to deliver all rumors from T to its root r, as quickly as possible. Such a protocol needs to achieve its goal even without the knowledge of the topology of T . More formally, a gathering protocol A can be defined as a function that, at each time t, and for each given node v, determines the action of v at time t based only on v’s label and the information received by v up to time t. The action of v at each time step t involves choosing its state (either receiving or transmitting) and, if it is in the transmitting state, choosing which rumor to transmit. We will say that A runs in time T (n) if, for any tree T and any assignment of labels to its nodes, after at most T (n) steps all rumors are delivered to r. A simple example of an information gathering protocol is called RoundRobin. In RoundRobin nodes transmit one at a time, in n rounds, where in each round they transmit in the order 0, 1, ..., n−1 4 of their labels. For any node v, when it is its turn to transmit, v transmits any rumor from the set of rumors that have been received so far (including its own rumor) but not yet transmitted. In each round, each rumor that is still not in r will get closer to r, so after n2 steps all rumors will reach r. Strong k-selectors. Let S̄ = (S1 , S2 , ..., Sm ) be a family of subsets of {0, 1, ..., n − 1}. S̄ is called a strong k-selector if, for each k-element set A ⊆ {0, 1, ..., n − 1} and each a ∈ A, there is a set Si such that Si ∩ A = {a}. As shown in [13, 9], for each k there exists a strong k-selector S̄ = (S0 , S1 , ..., Sm−1 ) with m = O(k 2 log n). We will make extensive use of strong k-selectors in our algorithm. At a certain time in the computation our protocols will “run” S̄, for an appropriate choice of k, by which we mean that it will execute a sequence of m consecutive steps, such that in the jth step the nodes from Sj will transmit, while those not in Sj will stay quiet. This will guarantee that, for any node v with at most k − 1 siblings, there will be at least one step in the execution of S̄ where v will transmit but none of its siblings will. Therefore at least one of v’s transmissions will be successful. 3 √ An O(n log n)-Time Protocol √ We first give a gathering protocol SimpleGather for trees with running time O(n log n). Our faster protocol will be presented in the next section. We fix three parameters: √ p p K = 2b log nc , D = dlogK ne = O( log n), D0 = dlog K 3 e = O( log n). We also fix a strong K-selector S̄ = (S0 , S1 , ..., Sm−1 ), where m ≤ CK 2 log n, for some integer constant C. By Lemma 1, we have that heightK (T ) ≤ D. We call a node v of T light if |Tv | ≤ n/K 3 ; otherwise we say that v is heavy. Let T 0 be the subtree of T induced by the heavy nodes. By the definition of heavy nodes, T 0 has at most K 3 leaves, so height2 (T 0 ) ≤ D0 . Also, obviously, r ∈ T 0 . To streamline the description of our algorithm we will allow each node to receive and transmit messages at the same time. We will also assume a preprocessing step allowing each v to know both the size of its subtree Tv (in particular, whether it is in T 0 or not), its K-height, and, if it is in T 0 , its 2-height in the subtree T 0 . We later explain both the preprocessing and how to modify the algorithm to remove the receive/transmit assumption. The algorithm consists of two epochs. Epoch 1 consists of D + 1 stages, each lasting O(n) steps. In this epoch only light vertices participate. The purpose of this epoch is to gather all rumors from T in T 0 . Epoch 2 has D0 + 1 stages and only heavy vertices participate in the computation during this epoch. The purpose of epoch 2 is to deliver all rumors from T 0 to r. We describe the computation in the two epochs separately. A detailed description of Algorithm SimpleGather is given in Pseudocode 1. To distinguish between computation steps (which do not consume time) and communication steps, we use command “at time t”. When the algorithm reaches this command it waits until time step t to continue processing. Each message transmission takes one time step. For each node v we maintain a set Bv of rumors received by v, including its own rumor ρv . Epoch 1: light vertices. Let 0 ≤ h ≤ D, and let v be a light vertex whose K-height equals h. Then v will be active only during stage h which starts at time αh = (C + 1)hn. This stage is divided into two parts. In the first part of stage h, v will transmit according to the strong K-selector S̄. Specifically, this part has n/K 3 iterations, where each iteration corresponds to a complete execution of S̄. At 5 any time, some of the rumors in Bv may be marked; the marking on a rumor indicates that the algorithm has already attempted to transmit it using S̄. At the beginning of each iteration, v chooses any rumor ρz ∈ Bv it has not yet marked, then transmits ρz in the steps that use sets Si containing the label of v. This ρz is then marked. If the parent u of v has degree at most K, the definition of strong K-selectors guarantees that ρz will be received by u, but if u’s degree is larger it may not have received ρz . Note that the total number of steps required for this part of stage h is (n/K 3 ) · m ≤ Cn, so these steps will be completed before the second part of stage h starts. In the second part, that starts at time αh + Cn, we simply run a variant of the RoundRobin protocol, but cycling through rumors instead of nodes: in the l-th step of this part, all nodes holding the rumor of the node with label l transmit that rumor (note that due to the tree topology it is impossible for two siblings to both be holding rumor `). We claim that the following invariant holds for all h = 0, 1, ..., D: (Ih ) Let w ∈ T and let u be a light child of w with heightK (u) ≤ h − 1. Then at time αh node w has received all rumors from Tu . To prove this invariant we proceed by induction on h. If h = 0 the invariant (I0 ) holds vacuously. So suppose that invariant (Ih ) holds for some value of h. We want to prove that (Ih+1 ) is true when stage h + 1 starts. We thus need to prove the following claim: if u is a light child of w with heightK (u) ≤ h then at time αh+1 all rumors from Tu will arrive in w. If heightK (u) ≤ h − 1 then the claim holds, immediately from the inductive assumption (Ih ). So assume that heightK (u) = h. Consider the subtree H rooted at u and containing all descendants of u whose K-height is equal to h. By the inductive assumption, at time αh any w0 ∈ H has all rumors from the subtrees rooted at its descendants of K-height smaller than h, in addition to its own rumor ρw0 . Therefore all rumors from Tu are already in H and each of them has exactly one copy in H, because all nodes in H were idle before time αh . When the algorithm executes the first part of stage h on H, then each v node in H whose parent is also in H will successfully transmit an unmarked rumor during each pass through the strong Kselector – indeed, our definition of H guarantees that v has at most K − 1 siblings in H, so by the definition of strong selector it must succeed at least once. We make the following additional claim: Claim 1. At all times during stage h, the collection of nodes in H still holding unmarked rumors forms an induced tree of H The claim follows from induction: At the beginning of the stage the nodes in H still hold their own original rumor, and it is unmarked since those nodes were idle so far. As the stage progresses, each parent of a transmitting child will receive a new (and therefore not yet marked) rumor during each run through the strong selector, so no holes can ever form. In particular, node u will receive a new rumor during every run through the strong selector until it has received all rumors from its subtree. Since the tree originally held at most |Tu | ≤ n/K 3 rumors originally, u must have received all rumors from its subtree after at most n/K 3 runs through the selector. Note that, as heightK (u) = h, u will also attempt to transmit its rumors to w during this part, but, since we are not making any assumptions about the degree of w, there is no guarantee that w will receive them. This is where the second part of this stage is needed. Since in the second part each rumor is transmitted without collisions, all rumors from u will reach w before time αh+1 , completing the inductive step and the proof that (Ih+1 ) holds. In particular, using Invariant (Ih ) for h = D, we obtain that after epoch 1 each heavy node w will have received rumors from the subtrees rooted at all its light children. Therefore at that time all rumors from T will be already in T 0 , with each rumor having exactly one copy in T 0 . 6 Pseudocode 1 SimpleGather(v) √ 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: K = 2b log nc , D = dlogK ne Bv ← {ρv } . Initially v has only ρv Throughout: all rumors received by v are automatically added to Bv if |Tv | ≤ n/K 3 then . v is light (epoch 1) h ← heightK (v, T ) ; αh ← (C + 1)nh . v participates in stage h for i = 0, 1, ..., n/K 3 − 1 do . iteration i at time αh + im if Bv contains an unmarked rumor then . Part 1: strong K-selector choose any unmarked ρz ∈ Bv and mark it for j = 0, 1, ..., m − 1 do at time αh + im + j if label(v) ∈ Sj then Transmit(ρz ) for l = 0, 1, ..., n − 1 do at time αh + Cn + l z ← node with label(z) = l if ρz ∈ Bv then Transmit(ρz ) else g ← height2 (v, T 0 ) ; αg0 ← αD+1 + 2ng for i = 0, 1, ..., n − 1 do at time αg0 + i if Bv contains an unmarked rumor then choose any unmarked ρz ∈ Bv and mark it Transmit(ρz ) for l = 0, 1, ..., n − 1 do at time αg0 + 2n + l z ← node with label(z) = l if ρz ∈ Bv then Transmit(ρz ) 7 . Part 2: RoundRobin . v is heavy (epoch 2) . v participates in stage g . Part 1: all nodes transmit . Part 2: RoundRobin w u H Figure 2: Proving Invariant (Ih ). Dark-shaded subtrees of Tu consist of light nodes with K-height at most h − 1. H consists of the descendants of u with K-height equal h. Epoch 2: heavy vertices. In this epoch we have at most D0 + 1 stages, and only heavy nodes in T 0 participate in the computation. When the epoch starts, all rumors are already in T 0 . In stage D + 1 + g the nodes in T 0 whose 2-height is equal g will participate. Similar to the stages of epoch 1, this stage has two parts and the second part executes RoundRobin, as before. The difference is that now, in the first part, instead of using the strong K-selector, each heavy node will transmit at each of the n steps. We need to show that r will receive all rumors at the end. The argument is similar as for light vertices, but with a twist, since we do not use selectors now; instead we have steps when all nodes transmit. In essence, we show that each stage reduces by at least one the 2-depth of the minimum subtree of T 0 that contains all rumors. Specifically, we show that the following invariant holds for all g = 0, 1, ..., D0 : (Jg ) Let w ∈ T 0 and let u ∈ T 0 be a child of w with height2 (u, T 0 ) ≤ g − 1. Then at time αg0 node w has received all rumors from Tu . We prove invariant (Jg ) by induction on g. For g = 0, (J0 ) holds vacuously. Assume that (Jg ) holds for some g. We claim that (Jg+1 ) holds right after stage g. Choose any child u of w with height2 (u, T 0 ) ≤ g. If height2 (u, T 0 ) ≤ g − 1, we are done, by the inductive assumption. So we can assume that height2 (u, T 0 ) = g. Let P be the subtree of T 0 rooted at u and consisting of all descendants of u whose 2-height in T 0 is equal g. Then P is simply a path. By the inductive assumption, for each w0 ∈ P , all rumors from the subtrees of w0 rooted at its children of 2-height at most g − 1 are in w0 . Thus all rumors from Tu are already in P . All nodes in P participate in stage g, but their children outside P do not transmit. Therefore each transmission from any node x ∈ P − {u} during stage g will be successful. Due to pipelining, all rumors from P will reach u after the first part of stage g. (This conclusion can also be derived from treating this computation on P as an instance of the token collection game in Section 2, with each step of the transmissions being one step of the game.) In the second part, all rumors from u will be successfully sent to w. So after stage g all rumors from Tu will be in w, completing the proof that (Jg+1 ) holds. Removing simplifying assumptions. At the beginning of this section we made some simplifying assumptions. It still remains to explain how to modify our algorithm so that it works even if these 8 assumptions do not hold. These modification are similar to those described in [5], but we include them here for the sake of completeness. First, we assumed a preprocessing step whereby each v knows certain parameters of its subtree Tv , including the size, its K-height, etc. The justification for this lies in the algorithm from [5] for information gathering in trees with aggregation. Such an algorithm can be modified to compute in linear time any function f such that f (v) is uniquely determined by the values of f on the children of v. The modification is that each node u, when its sends its message (which, in the algorithm from [5] contains all rumors from Tu ), it will instead send the value of f (u). A node v, after it receives all values of f (u) from each child u, will then compute f (v)1 . We also assumed that each node can receive and transmit messages at the same time. We now need to modify the algorithm so that it receives messages only in the receiving state and transmits only in the transmitting state. For the RoundRobin steps this is trivial: a node v is in the transmitting state only if it is scheduled to transmit, otherwise it is in the receiving state. For other steps, we will explain the modification for light and heavy nodes separately. Consider the computation of the light nodes during the steps when they transmit according to the strong selector. Instead of the strong K-selector, we can use the strong (K + 1)-selector, which will not affect the asymptotic running time. When a node v is scheduled to transmit, it enters the transmitting state, otherwise it is in the receiving state. In the proof, where we argue that the message from v will reach its parent, instead of applying the selector argument to v and its siblings, we apply it to the set of nodes consisting of v, its siblings, and its parent, arguing that there will be a step when v is the only node transmitting among its siblings and its parent is in the receiving state. Finally, consider the computation of the heavy nodes, at steps when all of them transmit. We modify the algorithm so that, in any stage g, the iteration (in Line 18) of these steps is preceded by O(n)-time preprocessing. Recall that the nodes whose 2-height in T 0 is equal g form disjoint paths. We can run a one round of RoundRobin where each node transmits an arbitrary message. This way, each node will know whether it is the first node on one of these paths or not. If a node x is first on some path, say P , x sends a message along this path, so that each node y ∈ P can compute its distance from x. Then, in the part where all nodes transmit, we replace each step by two consecutive steps (even and odd), and we use parity to synchronize the computation along these paths: the nodes at even positions are in the receiving state at even steps and in the transmitting state at odd steps, and the nodes at odd positions do the opposite. Summarizing this section, we have presented Algorithm SimpleGather √ that completes information gathering in ad-hoc radio networks with tree topologies in time O(n log n). In the next section, we will show how to improve this bound to O(n log log n). A Protocol with Running Time O(n log log n) 4 In this section we consider the same basic model of information gathering in trees as in Section 3, that is, the model does not provide √ any collision detection and rumor aggregation is not allowed. We show that how to refine our O(n log n) protocol SimpleGather to improve the running time to O(n log log n). This is the first main result of our paper, as summarized in the theorem below. Theorem 1. The problem of information gathering on trees, without rumor aggregation, can be solved in time O(n log log n). 1 It needs to be emphasized here that in our model only communication steps contribute to the running time; all calculations are assumed to be instantaneous. 9 Our protocol that achieves running time O(n log log n) will be called FastGather. This protocol can be thought of as an iterative application of the idea behind Algorithm SimpleGather from Section 3. We assume that the reader is familiar with Algorithm SimpleGather and its analysis, and in our presentation we will focus on the high level ideas behind Algorithm FastGather, referring the reader to Section 3 for the implementation of some details. As before, we use notation T for the input tree and n = |T | denotes the number of vertices in T . We assume that n is sufficiently large, and we will establish some needed lower bounds for n as we work through the proof. We fix some arbitrary integer constant β ≥ 2. For ` = 1, 2, ..., let −` K` = dnβ e. So K1 = dn1/β e, the sequence (K` )` is non-increasing, and lim`→∞ K` = 2. Let L be −` the largest value of ` for which nβ ≥ log n. (Note that L is well defined for sufficiently large n, since β is fixed). We thus have the following exact and asymptotic bounds: L ≤ logβ (log n/ log log n) KL ≥ log n L = Θ(log log n) KL = Θ(log n). ` ) we denote a strong K -selector of size m ≤ CK 2 log n, For ` = 1, 2, ..., L, by S̄ ` = (S1` , S2` , ..., Sm ` ` ` ` for some integer constant C. As discussed in Section 2, such selectors S̄ ` exist. Let T (0) = T , and for each ` = 1, 2, ..., L, let T (`) be the subtree of T induced by the nodes v with |Tv | ≥ n/K`3 . Each tree T (`) is rooted at r, and T (`) ⊆ T (`−1) for ` ≥ 1. For ` 6= 0, the definition of T (`) implies also that it has at most K`3 leaves, so, by Lemma 1, its K`+1 -height is at −` −(`+1) , we have most logK`+1 (K`3 ). Since K` ≤ 2nβ and K`+1 ≥ nβ logK`+1 (K`3 ) = 3 logK`+1 (K` ) −` ≤ 3 lognβ−(`+1) (2nβ ) ≤ 3 lognβ−(`+1) nβ = 3β + −` + 3 lognβ−(`+1) 2 3β `+1 3β L+1 ≤ 3β + ≤ 3β + 1, log n log n where the last inequality holds as long as 3β L+1 ≤ log n, which is true if n is large enough (since by the first of the "exact and asymptotic bounds" above we have β L ≤ log n/ log log n). We thus obtain that the K`+1 -height of T (`) is at most D = 3β + 1 = O(1). As in the previous section, we will make some simplifying assumptions. First, we will assume that all nodes can receive and transmit messages at the same time. Second, we will also assume that each node v knows the size of its subtree |Tv | and its K` -heights, for each ` ≤ L. After completing the description of the algorithm we will explain how to modify it so that these assumptions will not be needed. Algorithm FastGather consists of L + 1 epochs, numbered 1, 2, ..., L + 1. In each epoch ` ≤ L only the nodes in T (`−1) − T (`) participate. At the beginning of epoch ` all rumors will be already collected in T (`−1) , and the purpose of epoch ` is to move all these rumors into T (`) . Each of these L epochs will run in time O(n), so their total running time will be O(nL) = O(n log log n). In epoch L + 1, only the nodes in T (L) participate. At the beginning of this epoch all rumors will be already in T (L) , and when this epoch is complete, all rumors will be collected in r. This epoch will take time O(n log log n). Thus the total running time will be also O(n log log n). We now provide the details. Epochs ` = 1, 2, ..., L. In epoch `, only the nodes in T (`−1) − T (`) are active, all other nodes will be quiet. The computation in this epoch is very similar to the computation of light nodes (in 10 epoch 1) in Algorithm SimpleGather. Epoch ` starts at time γ` = (D + 1)(C + 1)(` − 1)n and lasts (D + 1)(C + 1)n steps. Let v ∈ T (`−1) − T (`) . The computation of v in epoch ` consists of D + 1 identical stages. Each stage h = 0, 1, ..., D starts at time step α`,h = γ` + (C + 1)hn and lasts (C + 1)n steps. Stage h has two parts. The first part starts at time α`,h and lasts time Cn. During this part we execute dn/K`3 e iterations, each iteration consisting of running the strong K`+1 -selector S̄ ` . The time needed to execute these iterations is at most d 2 log n 2K`+1 n 2 e · CK log n ≤ Cn · `+1 K`3 K`3 ≤ Cn · 8n2β −(`+1) log n n3β −` 8 log n n(3−2/β)β −` 8 log n ≤ Cn · (3−2/β)β −L ≤ Cn, n = Cn · where the last inequality holds as long as n3−2/β β −L ≥ 8 log n, which is again true for sufficiently large n (recall that β ≥ 2 is constant, and L = Θ(log log n)). Thus all iterations executing the strong selector will complete before time α`,h + Cn. Then v stays idle until time α`,h + Cn, which is when the second part starts. In the second part we run the RoundRobin protocol, which takes n steps. So stage h will complete right before step α`,h + (C + 1)n = α`,h+1 . Note that the whole epoch will last (D + 1)(C + 1)n steps, as needed. We claim that the algorithm preserves the following invariant for ` = 1, 2, ..., L: (I` ) Let w ∈ T and let u ∈ T − T (`−1) be a child of w. Then w will receive all rumors from Tu before time γ` , that is before epoch ` starts. For ` = 1, invariant (I1 ) holds vacuously, because T (0) = T . In the inductive step, assume that (I` ) holds for some epoch `. We want to show that (I`+1 ) holds right after epoch ` ends. In other words, we will show that if w has a child u ∈ T − T (`) then w will receive all rumors from Tu before time γ`+1 . So let u ∈ T −T (`) . If u 6∈ T (`−1) then (I`+1 ) holds for u, directly from the inductive assumption. We can thus assume that u ∈ T (`−1) − T (`) . The argument now is very similar to that for Algorithm SimpleGather in Section 3, when we analyzed epoch 1. For each h = 0, 1, ..., D we prove a refined version of condition (I`+1 ): (`−1) − T (`) be child of w with height (u, T (`−1) ) ≤ h − 1. Then (I`+1 K` h ) Let w ∈ T and let u ∈ T w will receive all rumors from Tu before time α`,h , that is before stage h. The proof is the same as for Invariant (Ih ) in Section 3, proceeding by induction on h. For each fixed h we consider a subtree H rooted at u and consisting of all descendants of u in T (`−1) whose K` -height is at least h. By the inductive assumption, at the beginning of stage h all rumors from Tu are already in H. Then, the executions of S̄ ` , followed by the execution of RoundRobin, will move all rumors from H to w. We omit the details here. By applying condition (I`+1 h ) with h = D, we obtain that after all stages of epoch ` are complete, that is at right before time γ`+1 , w will receive all rumors from Tu . Thus ` invariant (I`+1 ) will hold. Epoch L + 1. Due to the definition of L, we have that T (L) contains at most KL3 = O(log3 n) leaves, so its 2-depth is at most D0 = log(KL3 ) = O(log log n), by Lemma 1. The computation in this epoch 11 is similar to epoch 2 from Algorithm SimpleGather. As before, this epoch consists of D0 + 1 stages, where each stage g = 0, 1, ..., D0 has two parts. In the first part, we have n steps in which each node transmits. In the second part, also of length n, we run one iteration of RoundRobin. Let αg0 = γL + 2gn. To prove correctness, we show that the following invariant holds for all stages g = 0, 1, ..., D0 : (Jg ) Let w ∈ T (L) and let u ∈ T (L) be a child of w with height2 (u, T (L) ) ≤ g − 1. Then at time αg0 node w has received all rumors from Tu . The proof is identical to the proof of the analogous Invariant (Jg ) in Section 3, so we omit it here. Applying Invariant (Jg ) with g = D0 , we conclude that after stage D0 , the root r will receive all rumors. As for the running time, we recall that L = O(log log n). Each epoch ` = 1, 2, ..., L has D + 1 = O(1) stages, where each stage takes time O(n), so the execution of the first L epochs will take time O(n log log n). Epoch L + 1 has D0 + 1 = O(log log n) stages, each stage consisting of O(n) steps, so this epoch will complete in time O(n log log n). We thus have that the overall running time of our protocol is O(n log log n). It remains to explain that the simplifying assumptions we made at the beginning of this section are not needed. Computing all subtree sizes and all K` -heights can be done recursively bottom-up, using the linear-time information gathering algorithm from [5] that uses aggregation. This was explained in Section 3. The difference now is that each node has to compute L + 1 = O(log log n) values K` , and, since we limit bookkeeping information in each message up to O(log n) bits, these values need to be computed separately. Nevertheless, the total pre-computation time will still be O(n log log n). Removing the assumption that nodes can receive and transmit at the same time can be done in the same way as in Section 3. Roughly, in each epoch ` = 1, 2, ..., L, any node v ∈ T (`−1) − T (`) uses a strong (K` + 1)-selector (instead of a strong K` -selector) to determine whether to be in the receiving or transmitting state. In epoch L the computation (in the steps when all nodes transmit) is synchronized by transmitting a control message along induced paths, and then choosing the receiving or transmitting state according to node parity. Summarizing, we have proved that the running time of Algorithm SimpleGather is O(n log log n), thus completing the proof of Theorem 1. 5 An O(n)-time Protocol with Acknowledgments In this section we consider a slightly different communication model from that in Sections 3 and 4. We now assume that acknowledgments of successful transmissions are provided to the sender. All the remaining nodes, including the intended recipient, cannot distinguish between collisions and absence of transmissions. The main result of this section, as summarized in the theorem below, is that with this feature it is possible to achieve the optimal time O(n). Theorem 2. The problem of information gathering on trees without rumor aggregation can be solved in time O(n) if acknowledgments are provided. The overall structure of our O(n) time protocol, called Algorithm LinGather, is similar to Algorithm SimpleGather. It consists of two epochs. The first epoch does not use the acknowledgement feature, and it is in fact identical to Epoch 1 in Algorithm SimpleGather, except for a 12 different choice of the parameters. After this epoch, lasting time O(n), all rumors will be collected in the subtree T 0 consisting of the heavy nodes. In the second epoch, only the heavy nodes in T 0 will participate in the computation, and the objective of this epoch is to move all rumors already collected in T 0 to the root r. The key obstacle to be overcome in this epoch is congestion stemming from the fact that, although T 0 may be small, its nodes have many rumors to transmit. This congestion means that simply repeatedly applying k-selectors is no longer enough. For example, if the root has ` children, each with n/` rumors, then repeating an `-selector n/` times would get all the rumors to the root. However, we know from the selector size bounds in [8] that this would take total time Ω(n` log n/ log `), which is far too slow. While in this particular scenario RoundRobin will collect all rumors in the root in time O(n), this example can be enhanced to show that simple combinations of k-selectors and RoundRobin do not seem to be sufficient to gather all rumors from T 0 in the root in linear time. To overcome this obstacle, we introduce two novel tools that will play a critical role in our algorithm. The first tool is a so-called amortizing selector family. Since a parent, say with ` children, receives at most one rumor per round, it clearly cannot simultaneously be receiving rumors at an average rate greater than 1` from each child individually. With the amortizing family, we will be able to achieve this bound within a constant fraction over long time intervals, so long as each child knows (approximately) how many siblings it is competing with. Of course, such a family will not be useful unless a node can obtain an accurate estimate of its parent’s degree, which will be the focus of our second tool, k-distinguishers. Using a k-distinguisher a node can determine whether its number of active siblings ` is at least k or at most 2k. While this information is not sufficient to determine the exact relation between ` and k, we show how to combine different k-distinguishers to obtain another structure, called a cardinality estimator, that will determine the value of ` within a factor of 4. Using this estimate, and the earlier described amortizing selector, a node can quickly transmit its rumors to its parent. This will allow us to gather all rumors from T 0 in the root in time O(n). This section is divided into three parts. In Sections 5.1 and 5.2 we give precise definitions and constructions of our combinatorial structures. Our O(n)-time protocol LinGather is described and analyzed in Section 5.3. 5.1 Construction of Amortizing Selectors We now define the concept of an amortizing selector family S̄. Similarly to a strong selector, this amortizing family will be a collection of subsets of the underlying label set [n], though now it will be doubly indexed. Specifically, S̄ = {Sij }, where 1 ≤ i ≤ s and each j ∈ {1, 2, 4, 8, . . . , k}, for some parameters s and k. We say that S̄ succeeds at cumulative rate q if the following statement is true: For each j ∈ {1, 2, 4, . . . , k2 }, each subset A ⊆ {1, . . . , n} satisfying j/2 ≤ |A| ≤ 2j, and each q element v ∈ A there are at least |A| s distinct i for which v ∈ Sij and A ∩ (Si(j/2) ∪ Sij ∪ Si(2j) ) = {v}. In the case j = 1 the set Si(j/2) is defined to be empty. Here s can be thought of as the total running time of the selector, j as a node’s estimate of its parent’s degree, and k as some bound on the maximum degree handled by the selector. A node fires at time step i if and only if its index is contained in the set Sij . What the above statement is then saying that for any subset A of siblings, if |A| is at most k/2 and each child estimates |A| within a factor of 2 then each child will transmit q . at rate at least |A| 13 Theorem 3. There are fixed constants c, C > 0 such that the following is true: For any k and n and any s ≥ Ck 2 log n, there is an amortizing selector with parameters n, k, s succeeding with cumulative rate c. Let k, n, s be given such that s ≥ Ck 2 log n, where k is a constant to be determined later. We form our selector probabilistically: For each v, i, and j, we independently include v in Sij with probability 2−j . Observe that by monotonicity it suffices to check the selector property for the case |A| = 2j: If we replace A with a larger set A0 containing A and satisfying |A0 | ≤ 4|A|, then for any v ∈ A and any i satisfying A0 ∩ (Si(j/2) ∪ Sij ∪ Si(2j) ) = {v}, we also have A ∩ (Si(j/2) ∪ Sij ∪ Si(2j) ) = {v}. (c/4)s cs So if there are at least |A 0 | distinct i satisfying the first equality, there are at least |A| satisfying the second equality. Now fix j ∈ {1, 2, 4, . . . , k}, a set A ⊆ [n] with |A| = 2j and some v ∈ A, and let the random variables X and Y be defined by  X = | i : v ∈ Sij and A ∩ (Si(j/2) ∪ Sij ∪ Si(2j) ) = {v} | The expected value of X is µX        1 2 2j−1 1 2j−1 1 2j−1 =s 1− 1− 1− . j j j 2j Utilizing the bound       1 2j−1 1 2j−1 1 2 2j−1 1− 1− ≥ 7, 1− j j 2j e we have µX ≥ e−7 Now let c = 1 . 4e7 s j Applying the Chernoff bound, we get Pr[X ≤ c s 1 ] ≤ Pr[X ≤ µX ] ≤ e−µX /8 |A| 2 7 ≤ e−s/8e j . We now use the union bound over all choices
of j, v, and S. We have at most n choices of n v, at most log n choices for j, and at most 2j−1 ≤ n2k−1 choices of S given j and v. Thus the probability that our family is not an amortizing selector is at most   7 7 n log n · n2k−1 · e−s/8e j ≤ 2n2k log n · e−Ck log n/8e which is smaller than 1 for sufficiently large C. This implies the existence of the amortized selector family. 14 5.2 Construction of k-Distinguishers and Cardinality Estimators In this section, we define k-distinguishers and show how to construct them. Let S̄ = (S1 , S2 , ..., Sm ), where Sj ⊆ [n] for each j. For A ⊆ [n] and a ∈ A, define Hitsa,A (S̄) = {j : Sj ∩ A = {a}}, that is Hitsa,A (S̄) is the collection of indices j for which Sj intersects A exactly on a. Note that, using this terminology, S̄ is a strong k-selector if and only if Hitsa,A (S̄) 6= ∅ for all sets A ⊆ [n] of cardinality at most k and all a ∈ A. We say that S̄ is a k-distinguisher if there is a threshold value ξ (which may depend on k) such that, for any A ⊆ [n] and a ∈ A, the following conditions hold: if |A| ≤ k then |Hitsa,A (S̄)| > ξ, and if |A| ≥ 2k then |Hitsa,A (S̄)| < ξ. We make no assumptions on what happens for |A| ∈ {k + 1, k + 2, ..., 2k − 1}. The idea is this: consider a fixed a, and imagine that we have some set A that contains a, but its other elements are not known. Suppose that we also have an access to a hit oracle that for any set S will tell us whether S ∩ A = {a} or not. With this oracle, we can then use a k-distinguisher S̄ to extract some information about the cardinality of A by calculating the cardinality of Hitsa,A (S̄). If |Hitsa,A (S̄)| ≤ ξ then we know that |A| > k, and if |Hitsa,A (S̄)| ≥ ξ then we know that |A| < 2k. What we will soon show, again by a probabilistic construction, is that not-too-large k-distinguishers exist: Theorem 4. For any n ≥ 2 and 1 ≤ k ≤ n/2 there exists a k-distinguisher of size m = O(k 2 log n). In our framework, the acknowledgement of a message received from a parent corresponds exactly to such a hit oracle. So if all nodes fire according to such a k-distinguisher, each node can determine in time O(k 2 log n) either that its parent has at least k children or that it has at most 2k children. Now let λ be a fixed parameter between 0 and 1. For each i = 0, 1, ..., dλ log ne, let S̄ i be a i 2 -distinguisher of size O(22i log n) and with threshold value ξi . We can then concatenate these Pdλ log ne k-distinguishers to obtain a sequence S̃ of size i=0 O(22i log n) = O(n2λ log n). We will refer to S̃ as a cardinality estimator, because applying our hit oracle to S̄ we can estimate a cardinality of an unknown set within a factor of 4, making O(n2λ log n) hit queries. More specifically, consider again a scenario where we have a fixed a and some unknown set A containing a, where |A| ≤ nλ . Using the hit oracle, compute the values hi = |Hitsa,A (S̄ i )|, for all i. If i0 is the smallest i for which hi > ξi , then by definition of our distinguisher we must have 2i0 −1 < |A| < 2(2i0 ). In our gathering framework, this corresponds to each node in the tree being able to determine in time O(n2λ log n) a value of j (specifically, i0 − 1) such that the number of children of its parent is between 2j and 2j+2 , which is exactly what we need to be able to run the amortizing selector. It remains to show the existence of k-distinguishers, i.e. to prove Theorem 4. Let m = Ck 2 log n, where C is some sufficiently large constant whose value we will determine later. We choose the collection of random sets S̄ = (S1 , S2 , ..., Sm ), by letting each Sj be formed by independently including each x ∈ [n] in Sj with probability 1/2k. Thus, for any set A and a ∈ A, the probability that Sj ∩ A = {a} is (1/2k)(1 − 1/2k)|A|−1 , and the expected value of |Hitsa,A (S̄)| is E[|Hitsa,A (S̄)|] = m · 1 2k 1−
1 |A|−1 . 2k (1) Recall that to be a k-distinguisher our set needs to satisfy (for suitable ξ) the following two properties: (d1) if |A| ≤ k then |Hitsa,A (S̄)| > ξ (d2) if |A| ≥ 2k then |Hitsa,A (S̄)| < ξ. 15 We claim that, for a suitable value of ξ, the probability that there exists a set A ⊆ [n] and some a ∈ A for which S̄ does not satisfy both conditions is smaller than 1 (and in fact tends to 0) This will be sufficient to show an existence of a k-distinguisher with threshold value ξ. Observe that in order to be a k-distinguisher it is sufficient that S̄ satisfies (d1) for sets A with |A| = k and satisfies (d2) for sets A with |A| = 2k. This is true because the value of |Hitsa,A (S̄)| is monotone with respect to the inclusion: if a ∈ A ⊆ B then Hitsa,A (S̄) ⊇ Hitsa,B (S̄). Now consider some fixed a ∈ [n] and two sets A1 , A2 ⊆ [n] such that |A1 | = k, |A2 | = 2k and a ∈ A1 ∩ A2 . For i = 1, 2, we consider two corresponding random variables Xi = |Hitsa,Ai (S̄)| and their expected values µi = Exp[Xi ]. For any integer k ≥ 1 we have 1 e1/2 1 e ≤ (1 − ≤ (1 − 1 k−1 2k ) 1 2k−1 2k ) ≤ 12 . From (1), substituting m = Ck 2 log n, this gives us the corresponding estimates for µ1 and µ2 : 1 Ck log n 2e1/2 1 2e Ck log n ≤ µ1 ≤ µ2 ≤ 14 Ck log n Since e−1/2 > 21 , we can choose an  ∈ (0, 1) and ξ for which (1 + )µ2 < ξ < (1 − )µ1 . Thus the probability that S̄ violates (d1) for A = A1 is 2 µ /2 1 Pr[X1 ≤ ξ] ≤ Pr[X1 ≤ (1 − )µ1 ] ≤ e− 2 e−1/2 Ck log n/4 ≤ e− , where in the second inequality we use the Chernoff bound for deviations below the mean. Similarly, using the Chernoff bound for deviations above the mean, we can bound the probability of S̄ violating (d2) for A = A2 as follows: 2µ Pr[X2 ≥ ξ] ≤ Pr[X2 ≥ (1 + )µ2 ] ≤ e− 2 e−1 Ck log n/6 ≤ e− 2 /3 . To finish off the proof, we apply the union bound. We
have at most n choices for a, at most
n n k−1 ≤ n2k−1 choices of A , and at most 2k−1 choices of A . Note also that 1 2 k−1 ≤ n 2k−1 ≤ n −1/2 −1 e /4 > e /6. Putting it all together, the probability that S̄ is not a k-distinguisher is at most

n n n · k−1 · Pr[X1 ≤ ξ] + n · 2k−1 · Pr[X2 ≥ ξ] ≤ n2k · (Pr[X1 ≤ ξ] + Pr[X2 ≥ ξ]) ≤ 2n2k · e− = 2nk(2− for C large enough. 16 2 e−1 Ck log n/6 2 e−1 C/6) < 1, 5.3 Linear-time Protocol As before, T is the input tree with n nodes. We will recycle the notions of light and heavy nodes from Section 3, although now we will use slightly different parameters. Let δ > 0 be a small constant, and let K = dnδ e. We say that v ∈ T is light if |Tv | ≤ n/K 3 and we call v heavy otherwise. By T 0 we denote the subtree of T induced by the heavy nodes. Algorithm LinGather. Our algorithm will consist of two epochs. The first epoch is essentially identical to Epoch 1 in Algorithm SimpleGather, except for a different choice of the parameters. The objective of this epoch is to collect all rumors in T 0 in time O(n). In the second epoch, only the heavy nodes in T 0 will participate in the computation, and the objective of this epoch is to gather all rumors from T 0 in the root r. This epoch is quite different from our earlier algorithms and it will use the combinatorial structures obtained in the previous sections to move all rumors from T 0 to r in time O(n). Epoch 1: In this epoch only light nodes will participate, and the objective of Epoch 1 is to move all rumors into T 0 . In this epoch we will not be taking advantage of the acknowledgement mechanism. As mentioned earlier, except for different choices of parameters, this epoch is essentially identical to Epoch 1 of Algorithm SimpleGather, so we only give a very brief overview here. We use a strong K-selector S̄ of size m ≤ CK 2 log n. Let D = dlogK ne ≤ 1/δ = O(1). By Lemma 1, the K-depth of T is at most D. Epoch 1 consists of D + 1 stages, where in each stage h = 0, 1, ..., D, nodes of K-depth h participate. Stage h consists of n/K 3 executions of S̄, followed by an execution of RoundRobin, taking total time n/K 3 · m + n = O(n). So the entire epoch takes time (D + 1) · O(n) = O(n) as well. The proof of correctness (namely that after this epoch all rumors are in T 0 ) is identical as for Algorithm SimpleGather. Epoch 2: When this epoch starts, all rumors are already gathered in T 0 , and the objective is to push them further to the root. For the second epoch, we restrict our attention to the tree T 0 of heavy nodes. As before, no parent in this tree can have more than K 3 = n3δ children, since each child is itself the ancestor of a subtree of size n/K 3 . We will further assume the existence of a fixed amortizing selector family with parameters k = 2K 3 and s = K 8 ,
as well as a fixed
cardinality estimator with parameter λ = 3δ running in time D1 = O n6δ log n = O K 6 log n . Our protocol will be divided into stages, each consisting of 2(D1 + K 8 ) steps. A node will be active in a given stage if at the beginning of the stage it has already received all of its rumors, but still has at least one rumor left to transmit (it is possible for a node to never become active, if it receives its last rumor and then finishes transmitting before the beginning of the next stage). During each odd-numbered time step of a stage, all nodes (active or not) holding at least one rumor they have not yet successfully passed on transmit such a rumor. The even-numbered time steps are themselves divided into two parts. In the first D1 even steps, all active nodes participate in the aforementioned cardinality estimator. At the conclusion of the estimator, each node knows a j such that their parent has between 2j and 2j+2 active children. Note that active siblings do not necessarily have the same estimate for their family size. For the remainder of the even steps, each active node fires using the corresponding 2j+1 -selector from the amortizing family. The stages repeat until all rumors have reached the root. Our key claim, stated below, is that the rumors aggregate at least at a steady rate over time – each node with subtree size m in the original tree T will have received all m rumors within O(m) steps of the start of the epoch. (I` ) For any heavy node v such that v has subtree size m in T , and any 0 ≤ j ≤ m, that node 17 has received at least j rumors within time C(2m + j) of the beginning of Epoch 2, where C is some sufficiently large absolute constant. In particular, the root has received all of the rumors by time 3Cn. We will show this invariant holds by induction on the node’s height within T 0 . If the node is a leaf the statement follows from our analysis of Epoch 1 (the node has received all rumors from its subtree by the beginning of epoch 2). Now assume that a node u with subtree size m + 1 has k children within T 0 , and that those children have subtree sizes a1 ≥ a2 ≥ · · · ≥ ak ≥ K 3 . Node u may also received some number a0 of messages from non-heavy children (these children, if any, will have already completed their transmissions during the previous epoch). Let v be a child having subtree size a1 (chosen arbitrarily, if there are two children with maximal subtree size). Let t2 be defined by  3Ca2 + 3c (a2 + · · · + ak ) + K 12 if k > 1 t2 = 0 if k = 1 We make the following additional claims. Claim 2: By time t2 , all children except possibly v will have completed all of their transmissions. Proof. By inductive hypothesis, all children except v will have received all of their rumors by time 3Ca2 . During each stage from that point until all non-v nodes complete, one of the following things must be true. • Some active node transmits the final rumor it is holding, and stops transmitting. This can happen during at most K 3 stages, since there are at most K 3 children and each child only becomes active once it already has received all rumors from its subtree. • All active nodes have rumors to transmit throughout the entire stage. If there were j active nodes total during the stage, then by the definition of our amortizing selector family, the 8 parent received at least c 2Kj rumors from each child during the stage. In particular, it must have received at least 2K 8 c(j − 1) ≥ cK 8 j new rumors from children other than v. Combining the two types of stages, the non-v children will have all finished in at most K3 + 1 (a2 + · · · + ak ) cK 8 complete stages after time 3Ca2 . Since each stage takes time 2(D1 + K 8 ) = (2 + o(1))K 8 , the bound follows. Claim 3: Let k, m, and v be as above. By time 2Cm, all children except possibly v have completed their transmissions. Proof. This is trivial for k = 1. For larger k follows from the previous claim, together with the estimate that (for sufficiently large C) 4 5 2Cm ≥ 2C(a1 + · · · + ak ) ≥ 4Ca2 + 2C(a3 + · · · + ak ) ≥ (t2 − K 12 ) ≥ t2 3 4 Here the middle inequality holds for any C > 2/c, while the latter inequality holds since t2 ≥ a2 ≥ n/K 3  K 12 . 18 By the above claim, node v is the only node that could still be transmitting at time 2Cm. In particular, if it has a rumor during an odd numbered time step after this point, it successfully transmits. By assumption, v will have received at least j rumors by time C(2m + j) for each j. This implies it will successfully broadcast j rumors by time C(2m + j) + 2 for each 0 ≤ j ≤ a1 . By time C[2(m + 1) + j], the parent has either received all rumors from all of its children (if j > a1 ), or at least j rumors from a1 alone (if j ≤ a1 ). Either way, it has received at least j rumors total, and the induction is complete. References [1] Noga Alon, Amotz Bar-Noy, Nathan Linial, and David Peleg. A lower bound for radio broadcast. J. Comput. Syst. Sci., 43(2):290–298, 1991. [2] Danilo Bruschi and Massimiliano Del Pinto. Lower bounds for the broadcast problem in mobile radio networks. Distrib. Comput., 10(3):129–135, April 1997. [3] Bogdan S. Chlebus, Leszek Gasieniec, Alan Gibbons, Andrzej Pelc, and Wojciech Rytter. Deterministic broadcasting in ad hoc radio networks. Distributed Computing, 15(1):27–38, 2002. [4] Malin Christersson, Leszek Gasieniec, and Andrzej Lingas. Gossiping with bounded size messages in ad hoc radio networks. In Automata, Languages and Programming, 29th International Colloquium, ICALP’02, pages 377–389, 2002. [5] Marek Chrobak, Kevin Costello, Leszek Gasieniec, and Darek R. Kowalski. Information gathering in ad-hoc radio networks with tree topology. In Proc. 8th International Conference on Combinatorial Optimization and Applications (COCOA’14), pages 129–145, 2014. [6] Marek Chrobak, Leszek Gasieniec, and Wojciech Rytter. Fast broadcasting and gossiping in radio networks. J. Algorithms, 43(2):177–189, 2002. [7] Marek Chrobak, Leszek Gasieniec, and Wojciech Rytter. A randomized algorithm for gossiping in radio networks. Networks, 43(2):119–124, 2004. [8] Andrea E. F. Clementi, Angelo Monti, and Riccardo Silvestri. Selective families, superimposed codes, and broadcasting on unknown radio networks. In Prof. 12th Annual Symposium on Discrete Algorithms (SODA’01), pages 709–718, 2001. [9] Andrea E. F. Clementi, Angelo Monti, and Riccardo Silvestri. Distributed broadcast in radio networks of unknown topology. Theor. Comput. Sci., 302(1-3):337–364, 2003. [10] Artur Czumaj and Peter Davies. Almost optimal deterministic broadcast in radio networks. CoRR, abs/1506.00853, 2015. [11] Artur Czumaj and Wojciech Rytter. Broadcasting algorithms in radio networks with unknown topology. Journal of Algorithms, 60(2):115 – 143, 2006. [12] G. De Marco and D.R. Kowalski. Contention resolution in a non-synchronized multiple access channel. In 2013 IEEE 27th International Symposium on Parallel Distributed Processing (IPDPS), pages 525–533, 2013. 19 [13] P. Erdös, P. Frankl, and Z. Füredi. Families of finite sets in which no set is covered by the union of r others. Israel Journal of Mathematics, 51(1–2):79–89, 1985. [14] Antonio Fernández Anta, Miguel A. Mosteiro, and Jorge Ramón Muñoz. Unbounded contention resolution in multiple-access channels. Algorithmica, 67(3):295–314, 2013. [15] Leszek Gasieniec. On efficient gossiping in radio networks. In 16th Int. Colloquium on Structural Information and Communication Complexity (SIROCCO’09), pages 2–14, 2009. [16] Leszek Gasieniec, Tomasz Radzik, and Qin Xin. Faster deterministic gossiping in directed ad hoc radio networks. In Proc. Scandinavian Workshop on Algorithm Theory (SWAT’04), pages 397–407, 2004. [17] Dariusz R. Kowalski. On selection problem in radio networks. In Proceedings of the Twentyfourth Annual ACM Symposium on Principles of Distributed Computing, PODC ’05, pages 158–166, 2005. [18] Dariusz R. Kowalski and Andrzej Pelc. Faster deterministic broadcasting in ad hoc radio networks. SIAM J. Discrete Math., 18(2):332–346, 2004. [19] Eyal Kushilevitz and Yishay Mansour. An Ω(D log(N/D)) lower bound for broadcast in radio networks. SIAM J. Comput., 27(3):702–712, 1998. [20] Ding Liu and Manoj Prabhakaran. On randomized broadcasting and gossiping in radio networks. In 8th Annual Int. Conference on Computing and Combinatorics (COCOON’02), pages 340–349, 2002. [21] Gianluca De Marco. Distributed broadcast in unknown radio networks. In Proc. 19th Annual ACM-SIAM Symposium on Discrete Algorithms (SODA’08), pages 208–217, 2008. [22] Gianluca De Marco and Dariusz R. Kowalski. Fast nonadaptive deterministic algorithm for conflict resolution in a dynamic multiple-access channel. SIAM Journal on Computing, 44(3):868– 888, 2015. [23] A.N. Strahler. Hypsometric (area-altitude) analysis of erosional topology. Bull. Geol. Soc. Amer., 63:117–1142, 1952. [24] X.G. Viennot. A Strahler bijection between Dyck paths and planar trees. Discrete Mathematics, 246:317–329, 2003. [25] Ying Xu. An O(n1.5 ) deterministic gossiping algorithm for radio networks. Algorithmica, 36(1):93–96, 2003. 20 

Gayaz Khakimzyanov Institute of Computational Technologies, Novosibirsk, Russia Denys Dutykh CNRS–LAMA, Université Savoie Mont Blanc, France arXiv:1707.01301v1 [physics.flu-dyn] 5 Jul 2017 Oleg Gusev Institute of Computational Technologies, Novosibirsk, Russia Nina Shokina Institute of Computational Technologies, Novosibirsk, Russia Dispersive shallow water wave modelling. Part II: Numerical simulation on a globally flat space arXiv.org / hal Last modified: July 7, 2017 Dispersive shallow water wave modelling. Part II: Numerical simulation on a globally flat space Gayaz Khakimzyanov, Denys Dutykh∗, Oleg Gusev, and Nina Yu. Shokina Abstract. In this paper we describe a numerical method to solve numerically the weakly dispersive fully nonlinear Serre–Green–Naghdi (SGN) celebrated model. Namely, our scheme is based on reliable finite volume methods, proven to be very efficient for the hyperbolic part of equations. The particularity of our study is that we develop an adaptive numerical model using moving grids. Moreover, we use a special form of the SGN equations where non-hydrostatic part of pressure is found by solving a nonlinear elliptic equation. Moreover, this form of governing equations allows to determine the natural form of boundary conditions to obtain a well-posed (numerical) problem. Key words and phrases: nonlinear dispersive waves; moving adaptive grids; finite volumes; conservative finite differences MSC: [2010] 76B15 (primary), 76M12, 65N08, 65N06 (secondary) PACS: [2010] 47.35.Bb (primary), 47.35.Fg (secondary) Key words and phrases. nonlinear dispersive waves; non-hydrostatic pressure; moving adaptive grids; finite volumes; conservative finite differences. ∗ Corresponding author. Dispersive shallow water wave modelling. Part II 3 / 66 Contents 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 2 Mathematical model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 2.1 Well-posedness conditions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 Linear waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 2.2 Conservative form of equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 Intermediate conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 2.3 One-dimensional case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 Boundary conditions on the elliptic part . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 Unicity of the elliptic equation solution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12 2.4 Vector short-hand notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 Flat bottom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14 2.5 Linear dispersion relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 3 Numerical method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 3.1 Adaptive mesh construction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16 Initial grid generation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 Grid motion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 3.2 The SGN equations on a moving grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18 3.3 Predictor–corrector scheme on moving grids . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19 Predictor step . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 Corrector step . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24 Well-balanced property . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26 3.4 Numerical scheme for linearized equations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 Linear stability of the scheme . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 Discrete dispersion relation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32 4 Numerical results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 4.1 Solitary wave propagation over the flat bottom . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 Uniform grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 34 Adaptive grid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 4.2 Solitary wave/wall interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 36 Wave action on the wall . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39 4.3 Solitary wave/bottom step interaction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40 4.4 Wave generation by an underwater landslide . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 5.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55 5.2 Perspectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 A Derivation of the non-hydrostatic pressure equation B Acronyms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 References . . . . . . . . . . . . . 57 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 Dispersive shallow water wave modelling. Part II 5 / 66 1. Introduction In 1967 D. Peregrine derived the first two-dimensional Boussinesq-type system of equations [117]. This model described the propagation of long weakly nonlinear waves over a general non-flat bottom. From this landmark study the modern era of long wave modelling started. On one hand researchers focused on the development of new models and in parallel the numerical algorithms have been developed. We refer to [20] for a recent ‘reasoned’ review of this topic. The present manuscript is the continuation of our series of papers devoted to the long wave modelling. In the first part of this series we derived the so-called base model [92], which encompasses a number of previously known models (but, of course, not all of nonlinear dispersive systems). The governing equations of the base model are Ht + ∇ · [ HU ] = 0 , p̌ 1 h ∇P (HU )t + (ū · ∇)(HU ) = ∇h − ūt + (ū · ∇)ū + H H H i + H(U · ∇) ū + HU ∇ · ū , (1.1) (1.2) def where U := ū + U is the modified horizontal velocity and U = U (H, ū) is the closure relation to be specified later. Depending on the choice of this variable various models can be obtained (see [92, Section §2.4]). Variables P and p̌ are related to the fluid pressure. The physical meaning of these variables is reminded below in Section 2. In the present paper we propose an adaptive numerical discretization for a particular, but very popular nowadays model which can be obtained from the base model (1.1), (1.2). Namely, if we choose U ≡ 0 (thus, U becomes the depth-averaged velocity u) then we obtain equations equivalent to the celebrated Serre–Green–Naghdi (SGN) equations [72, 126, 127] (rediscovered later independently by many other researchers). This system will be the main topic of our numerical study. Most often, adaptive techniques for dispersive wave equations involve the so-called Adaptive Mesh Refinement (AMR) [121] (see also [15] for nonlinear shallow water equations). The particularity of our study is that we conserve the total number of grid points and the adaptivity is achieved by judiciously redistributing them in space [83, 84]. The ideas of redistributing grid nodes is stemming from the works of Bakhvalov [7], Il’in [85] and others [1, 134]. The base model (1.1), (1.2) admits an elegant conservative form [92]: Ht + ∇ · [ HU ] = 0 , i (H U )t + ∇ · Hū ⊗ U + P(H, ū) · I + H U ⊗ ū = p̌ ∇h , h (1.3) (1.4) where I ∈ Mat 2 × 2 (R) is the identity matrix and the operator ⊗ denotes the tensorial product. We note that the pressure function P(H, ū) incorporates the familiar hydrostatic g H2 well-known from the Nonlinear Shallow Water Equations (NSWE) pressure part 2 G. Khakimzyanov, D. Dutykh, et al. 6 / 66 [11, 43]. By setting U ≡ 0 we obtain readily from (1.3), (1.4) the conservative form of the SGN equations (one can notice that the mass conservation equation (1.1) was already in conservative form). Nonlinear dispersive wave equations represent certain numerical difficulties since they involve mixed derivatives (usually of the horizontal velocity variable, but sometimes of the total water depth as well) in space and time. These derivatives have to be approximated numerically, thus leaving a lot of room for the creativity. Most often the so-called Method Of Lines (MOL) is employed [97, 120, 123, 128], where the spatial derivatives are discretized first and the resulting system of coupled Ordinary Differential Equations (ODEs) is then approached with more or less standard ODE techniques, see e.g. [76, 77]. The MOL separates the choice of time discretization from the procedure of discretization in space, even if the interplay between two schemes might be important. For example, it would be natural to choose the same order of accuracy for both schemes. Let us review the available spatial discretization techniques employed in recent numerical studies. We focus essentially on fully nonlinear weakly dispersive models, even if some interesting works devoted to Boussinesq-type and unidirectional equations will be mentioned. First of all, dispersive wave equations with the dispersion relation given by a rational function (à la BBM [14, 116]) usually involve the inversion of an elliptic operator. This gives the first idea of employing the splitting technique between the hyperbolic and elliptic operators. This idea was successfully realized in e.g. [8, 9, 18, 82]. Historically, perhaps the finite difference techniques were applied first to dispersive (and more general non-hydrostatic) wave equations [24, 35–37, 108, 109, 143, 147]. Then, naturally we arrive to the development of continuous Galerkin/Finite Element type discretizations [2, 17, 45, 47, 114, 131, 139]. See also a recent review [46] and references therein. Pseudospectral Fourier-type methods can also be successfully applied to the SGN equations [52]. See [62] for a pedagogical review of pseudo-spectral and radial basis function methods for some shallow water equations. More recently, the finite volume type methods were applied to dispersive equations [30, 52, 57, 58, 89, 100]. In the present study we also employ a predictor–corrector finite volume type scheme [129], which is described in details below. The present article is organized as follows. In Section 2 we present the governing equations in 2D and 1D spatial dimensions. The numerical method is described in Section 3. Several numerical illustrations are shown in Section 4 including the solitary wave/wall or bottom interactions and even a realistic underwater landslide simulation. Finally, in Section 5 we outline the main conclusions and perspectives of the present study. In Appendices A we provide some details on analytical derivations used in this manuscript. 2. Mathematical model In this study we consider the following system of the Serre–Green–Naghdi (SGN) equations, which describes the incompressible homogeneous fluid flow in a layer bounded from below by the impermeable bottom y = −h(x, t) and above by the free surface 7 / 66 Dispersive shallow water wave modelling. Part II l y O α η (x,t) x u(x,t) d h(x,t) H(x,t) Figure 1. Sketch of the fluid domain in 2D. y = η (x, t), x = (x1 , x2 ) ∈ R2 :   Ht + ∇ · Hu = 0 , (2.1) p̌ ∇P = ∇h , (2.2) ut + (u · ∇) u + H H where for simplicity we drop the bars over the horizontal velocity variable u(x, t) =
def u1 (x, t), u2(x, t) . Function H(x, t) := h(x, t) + η(x, t) being the total water depth. The sketch of the fluid domain is schematically depicted in Figure 1. For the derivation of equations (2.1), (2.2) we refer to the first part of the present series of papers [92]. The depth-integrated pressure P(u, H) is defined as g H2 − ℘(x, t) , 2 where ℘(x, t) is the non-hydrostatic part of the pressure: def P (u, H) := ℘ (x, t) def := H2 H3 R1 + R2 , 3 2 (2.3) with def def R1 := D(∇ · u) − (∇ · u)2 , R2 := D 2 h , def D := ∂t + u · ∇ . Above, D is the total or material derivative operator. On the right hand side of equation def (2.2) we have the pressure trace at the bottom p̌ := p|y = −h , which can be written as p̌ (x, t) = gH − ̺ (x, t) , where ̺(x, t) is again the non-hydrostatic pressure contribution: def ̺ (x, t) := H2 R1 + H R2 . 2 (2.4) 8 / 66 G. Khakimzyanov, D. Dutykh, et al. Equations above are much more complex comparing to the classical NSWE (or SaintVenant equations) [43], since they contain mixed derivatives up to the third order. From the numerical perspective these derivatives have to be approximated. However, the problem can be simplified if we ‘extract’ a second order sub-problem for the non-hydrostatic component of the pressure. Indeed, it can be shown (see Appendix A) that function ℘(x, t) satisfies the following second order linear elliptic equation with variable coefficients (by analogy with incompressible Navier–Stokes equations, where the pressure is found numerically by solving a Poisson-type problem [33, 79]):     ∇h   2 Υ − 3 (∇℘ · ∇h) ∇h ∇℘ ℘ = F , (2.5) − 6 − · + ∇· ∇· H HΥ H3 Υ H2 Υ {z } | (⋆) def where Υ := 4 + | ∇h |2 and F, R are defined as   R ∇h 6R def u u F := ∇ · g ∇η + − + 2 (∇ · u) 2 − 2 1 x1 1 x2 , Υ HΥ u 2 x1 u 2 x2   def R := −g ∇η · ∇h + (u · ∇)∇h · u + htt + 2 u · ∇ht . (2.6) (2.7) Symbol | · | in (2.6) denotes the determinant of a 2 × 2 matrix. Equation (2.5) is uniformly elliptic and it does not contain time derivatives of the fluid velocity u. If the coefficient (⋆) is positive (for instance, it is the case for a flat bottom h(x, t) ≡ const), we deal with a positive operator and stable robust discretizations can be proposed. Taking into account the fact that equation (2.5) is linear with respect to the variable ℘(x, t), its discrete counterpart can be solved by direct or iterative methods∗. Well-posedness of this equation is discussed below (see Section 2.1). The boundary conditions for equation (2.5) will be discussed below in Section 2.3.1 (in 1D case only, the generalization to 2D is done by projecting on the normal direction to the boundary). Introduction of the variable ℘(x, t) allows to rewrite equation (2.2) in the following equivalent form: ut + (u · ∇)u + g ∇H = g ∇h + ∇℘ − ̺ ∇h . H (2.8) The non-hydrostatic pressure at the bottom ̺(x, t) can be expressed through ℘ in the following way: i 1 h 6℘ ℘ ̺(x, t) = + H R + ∇ · ∇h . (2.9) Υ H The derivation of this equation (2.9) is given in Appendix A as well. So, thanks to this relation (2.9), the usage of equation (2.4) is not necessary anymore. Once we found the function ℘(x, t), we can compute the bottom component from (2.9). ∗In our implementation we use the direct Thomas algorithm, since in 1D the resulting linear system of equations is tridiagonal with the dominant diagonal. Dispersive shallow water wave modelling. Part II 9 / 66 Remark 1. It can be easily seen that taking formally the limit ℘ → 0 and ̺ → 0 of vanishing non-hydrostatic pressures, allows us to recover the classical NSWE (or SaintVenant equations) [43]. Thus, the governing equations verify the Bohr correspondence principle [16]. 2.1. Well-posedness conditions In order to obtain a well-posed elliptic problem (2.5), one has to ensure that coefficient (⋆) is positive. This coefficient involves the bathymetry function h(x, t) and the total water depth H(x, t). In other words, the answer depends on local depth and wave elevation. It is not excluded that for some wave conditions the coefficient (⋆) may become negative. In the most general case the positivity condition is trivial and, thus, not very helpful, i.e.  ∇h  2 Υ − 3 (⋆) ≡ > 0. (2.10) · + ∇· H3 Υ H2 Υ On the flat bottom h(x, t) → d = const we know that the above condition is satisfied 1 > 0. Consequently, by continuity of the coefficient (⋆) since Υ → 4 and (⋆) → 2 H3 we conclude that the same property will hold for some (sufficiently small) variations of the depth h(x, t), i.e. | ∇h | ≪ 1. In practice it can be verified that bathymetry variations can be even finite so that condition (2.10) still holds. Remark 2. It may appear that restrictions on the bathymetry variations are inherent to our formulation only. However, it is the case of all long wave models, even if this assumption does not appear explicitly in the derivation. For instance, bottom irregularities will inevitably generate short waves ( i.e. higher frequencies) during the wave propagation process. A priori, this part of the spectrum is not modeled correctly by approximate equations, unless some special care is taken. 2.1.1 Linear waves Let us take the limit of linear waves η → 0 in expression (⋆). It will become then  ∇h  2 Υ − 3 def ⋆ ) := . · + ∇ · (⋆) → (⋆ h3 Υ h2 Υ ⋆ ) then takes the following form: The positivity∗ condition of (⋆

 2 Υ + h hx1 x1 1 − h2x1 + h2x2 + hx2 x2 1 + h2x1 − h2x2 − 4 hx1 hx2 hx1 x2 > 0 . If we restrict our attention to the one-dimensional bathymetries (i.e. hx2 → 0), then we obtain an even simpler condition: 2 1 + hx2 , hxx > − · h 1 − hx2 ∗Non-negativity, to be more precise. 10 / 66 G. Khakimzyanov, D. Dutykh, et al. where by x we denote x1 for simplicity. The last condition can be easily checked at the problem outset. A further simplification is possible if we additionally assume that | ∇h | ≡ | hx | < 1 , (2.11) 2 hxx > − . h (2.12) then we have the following elegant condition: 2.2. Conservative form of equations Equations (2.1), (2.2) admit an elegant conservative form, which is suitable for numerical simulations:   Ht + ∇ · Hu = 0 , (2.13) (Hu)t + ∇ · F = g H ∇h + ∇℘ − ̺∇h , (2.14) where the flux matrix F (H, u) is the same as in NSWE (or Saint-Venant equations):   g H2 2 H u1 · u2  def H u1 + 2 . F (H, u) :=  g H2 H u1 · u2 H u22 + 2 Notice that it is slightly different from the (fully-)conservative form given in Part I [92]. Conservative equations∗ (2.13), (2.14) can be supplemented by the energy conservation equation which can be used to check the accuracy of simulation (in conservative case, i.e. ht ≡ 0) and/or to estimate the energy of generated waves [54]: h  Pi = −p̌ ht , (2.15) (H E )t + ∇ · H u E + H where the total energy E is defined as g def (H − 2h) . E := 12 | u | 2 + 61 H 2 (∇ · u)2 + 21 H (Dh) (∇ · u) + 21 (Dh)2 + 2 Notice that equation (2.15) is not independent. It is a differential consequence of the mass and momentum conservations (2.13), (2.14) (as it is the case for incompressible flows in general). 2.2.1 Intermediate conclusions As a result, the system of nonlinear dispersive equations (2.1), (2.2) was split in two main parts: (1) Governing equations (2.13), (2.14) in the form of (hyperbolic) balance laws with source terms ∗It is not difficult to see that the mass conservation equation (2.1) is already in a conservative form in the SGN model. Thus, equations (2.1) and (2.13) are obviously identical. 11 / 66 Dispersive shallow water wave modelling. Part II (2) A scalar nonlinear elliptic equation to determine the non-hydrostatic part of the pressure ℘(x, t) (and consequently ̺(x, t) as well) This splitting idea will be exploited below in the numerical algorithm in order to apply the most suitable and robust algorithm for each part of the solution process [9]. 2.3. One-dimensional case In this study for the sake of simplicity we focus on the two-dimensional physical problem, i.e. one horizontal and one vertical dimensions. The vertical flow structure being resolved using the asymptotic expansion (see Part I [92] of this series of papers), thus we deal with def PDEs involving one spatial (horizontal) dimension (x := x1 ) and one temporal variable t ∈ R+ . The horizontal velocity variable u(x, t) becomes a scalar function in this case. Below we provide the full set of governing equations (which follow directly from (2.13), (2.14) and (2.5)): Ht + [ H u ]x = 0 , h g H2 i = g H hx + (H u)t + H u2 + 2 x   h h i  ℘x  2 Υ − 3 x ℘ = F, − 6 4 · + HΥ x H3 Υ H2 Υ x (2.16) ℘x − ̺ hx , (2.17) (2.18) def where Υ := 4 + h2x and def F := def  R hx gηx + Υ  x − 6R + 2 u2x , HΥ R := −g ηx hx + u2 hxx + htt + 2 u hxt . The last equations can be trivially obtained from corresponding two-dimensional versions given in (2.6), (2.7). This set of equations will be solved numerically below (see Section 3). 2.3.1 Boundary conditions on the elliptic part First, we rewrite elliptic equation (2.18) in the following equivalent form:   K ℘x x − K0 ℘ = F , where (2.19) h 2 Υ − 3  h  i 4 def x . , K0 := 6 + 3 2 HΥ H Υ H Υ x We assume that we have to solve an initial-boundary value problem for the system (2.16)– (2.18). If we have a closed numerical wave tank∗ (as it is always the case in laboratory def K := ∗Other possibilities have to be discussed separately. 12 / 66 G. Khakimzyanov, D. Dutykh, et al. experiments), then on vertical walls the horizontal velocity satisfies: ∀t ∈ R+ . u(x, t) |x = 0 = u(x, t) |x = ℓ ≡ 0, For the situation where the same boundary condition holds on both boundaries, we introduce a short-hand notation: =ℓ u(x, t) |xx = 0 ≡ 0, ∀t ∈ R+ . Assuming that equation (2.8) is valid up to the boundaries, we obtain the following boundary conditions for the elliptic equation (2.18):  x=ℓ  ℘ − ̺h x x − g ηx = 0, ∀t ∈ R+ . H x=0 Or in terms of equation (2.19) we equivalently have: x=ℓ   R hx  6 hx ℘  = g η + K℘ x − x H 2Υ Υ x=0 x=ℓ , x=0 ∀t ∈ R+ . (2.20) The boundary conditions for the non-hydrostatic pressure component ℘ are of the 3rd kind (sometimes they are referred to as of Robin-type). For the case where locally at the =ℓ boundaries the bottom is flat (to the first order), i.e. hx |xx = 0 ≡ 0, then we have the nd (non-homogeneous) Neumann boundary condition of the 2 kind: =ℓ K℘x |x = 0 = g ηx |xx = 0 , x=ℓ ∀t ∈ R+ . For a classical Poisson-type equation this condition would not be enough to have a wellposed problem. However, we deal rather with a Helmholtz-type equation (if K0 > 0). So, the flat bottom does not represent any additional difficulty for us and the unicity of the solution can be shown in this case as well. 2.3.2 Unicity of the elliptic equation solution The mathematical structure of equation (2.19) is very advantageous since it allows to show the following Theorem 1. Suppose that the Boundary Value Problem (BVP) (2.20) for equation (2.19) admits a solution and the following conditions are satisfied: K0 > 0, then this solution is unique. hx |x = 0 > 0, hx |x = ℓ 6 0 , (2.21) Proof. Assume that there are two such solutions ℘1 and ℘2 . Then, their difference def ℘ := ℘1 − ℘2 satisfies the following homogeneous BVP:   K ℘x x − K0 ℘ = 0 , (2.22) x=ℓ  6 hx ℘  ℘ = 0. (2.23) K x − H 2Υ x=0 13 / 66 Dispersive shallow water wave modelling. Part II Let us multiply the first equation (2.22) by ℘ and integrate over the computational domain: ˆ ℓ ˆ ℓ   2 K ℘x x ℘ dx − K0 ℘ dx = 0 . 0 |0 {z } () Integration by parts of the first integral () yields: ˆ ℓ ˆ ℓ x=ℓ 2 2 K ℘x ℘| − K ℘x ℘|x = 0 − K ℘x dx − K0 ℘ dx = 0 . 0 0 And using boundary conditions (2.23) we finally obtain: ˆ ℓ ˆ ℓ x=ℓ 6 hx 6 hx 2 2 2 2 ℘ ℘ − − K ℘x dx − K0 ℘ dx = 0 . H 2Υ H 2Υ 0 0 x=0 Taking into account this Theorem assumptions (2.21) and the fact that K > 0, the last identity leads to a contradiction, since the left hand side is strictly negative. Consequently,  the solution to equation (2.19) with boundary condition (2.20) is unique. Remark 3. Conditions in Theorem 1 are quite natural. The non-negativity of coefficient K0 has already been discussed in Section 2.1. Two other conditions mean that the water depth is increasing in the offshore direction (hx |x = 0 > 0) and again it is decreasing ( hx |x = ℓ 6 0) when we approach the opposite shore. 2.4. Vector short-hand notation For the sake of convenience we shall rewrite governing equations (2.16), (2.17) in the following vectorial form:   v t + F (v) x = G (v, ℘, ̺, h) , (2.24) where we introduced the following vector-valued functions:   ! Hu def def H v := , F (v) :=  2 g H2 , Hu Hu + 2 and the source term is defined as ! def 0 G (v, ℘, ̺, h) := . g H hx + ℘x − ̺hx The point of view that we adopt in this study is to view the SGN equations as a system of hyperbolic equations (2.24) with source terms G (v, ℘x , ̺, h). Obviously, one has to solve also the elliptic equation (2.18) in order to compute the source term G . The Jacobian matrix of the advection operator coincides with that of classical NSWE equations: ! def dF (v) 0 1 A (v) := = . dv −u2 + g H 2u 14 / 66 G. Khakimzyanov, D. Dutykh, et al. Eigenvalues of the Jacobian matrix A (v) can be readily computed: def p λ− = u − s , λ+ = u + s , s := gH. (2.25) The Jacobian matrix appears naturally in the non-divergent form of equations (2.24): v t + A (v) · v x = G , (2.26) By multiplying both sides of the last equation by A (v) we obtain the equations for the advection flux function F (v): Ft + A (v) · Fx = A · G . (2.27) In order to study the characteristic form of equations one needs also to know the matrix of left and right eigenvectors correspondingly: ! ! 1 def def s −1 1 −λ+ 1 L := 2 . (2.28) , R := s 2 −λ− λ+ −λ− 1 If we introduce also the diagonal matrix of eigenvalues ! def λ− 0 Λ := , 0 λ+ the following relations can be easily checked: R·Λ·L ≡ A , R·L = L ·R ≡ I, where I is the identity 2 × 2 matrix. 2.4.1 Flat bottom Equations above become particularly simple on the flat bottom. In this case the bathymetry functions is constant, i.e. h(x, t) ≡ d = const > 0 . Substituting it into governing equations above, we straightforwardly obtain: Ht + [ H u ]x = 0 , h g H2 i 2 (H u)t + H u + = ℘x , 2 x  ℘x  3 ℘ − = g ηxx + 2 u2x . H x H3 15 / 66 Dispersive shallow water wave modelling. Part II Solitary wave solution. Equations above admit an elegant analytical solution known as the solitary wave ∗. It is given by the following expressions: √ 
3αg υ · η(x, t) 2 η(x, t) = α · sech u(x, t) = x − x0 − υ t , , (2.29) 2dυ d + η(x, t) where α is the wave amplitude, x0 ∈ R is the initial wave position and υ is the velocity defined as def p υ := g (d + α) . The non-hydrostatic pressure under the solitary wave can be readily computed as well: ℘(x, t) = g  H 2 (x, t) − d 2  − d υ u(x, t) . 2 One can derive also periodic travelling waves known as cnoidal waves. For their expressions we refer to e.g. [52, 55]. 2.5. Linear dispersion relation The governing equations (2.16)–(2.18) after linearizations take the following form: ηt + d ux = 0 , ut + g ηx = ℘xx def where c := form √ − 3 d2 ℘ ℘x d , = c2 ηxx , g d is the linear gravity wave speed. By looking for plane wave solutions of η(x, t) = α ei (kx − ω t) u(x, t) = υ ei (kx − ω t) , ℘(x, t) = ρ ei (kx − ω t) ,  where k is the wave number, ω(k) is the wave frequency and α, υ, ρ ∈ R are some (constant) real amplitudes. The necessary condition for the existence of plane wave solutions reads ck ω(k) = ± r . (2.30) (kd)2 1 + 3 2π into the last formula and dividing both sides by By substituting the definition of k = λ k we obtain the relation between the phase speed cp and the wavelength λ:
c def ω k(λ)
. = r cp (λ) := k λ 4π 2 d 2 1 + 3λ2 ∗Solitary , waves are to be distinguished from the so-called solitons which interact elastically [48]. Since the SGN equations are not integrable (for the notion of integrability we refer to e.g. [145]), the interaction of solitary waves is inelastic [114]. 16 / 66 G. Khakimzyanov, D. Dutykh, et al. This dispersion relation is accurate to 2nd order at the limit of long waves kd → 0. There are many other nonlinear dispersive wave models which share the same linear dispersion relation, see e.g. [61, 64, 117, 149]. However, their nonlinear properties might be very different. 3. Numerical method The construction of numerical schemes for hyperbolic conservation laws on moving grids was described in our previous work [94]. In the present manuscript we make an extension of this technology to dispersive PDEs illustrated on the example of the SGN equations (2.16)–(2.18). The main difficulty which arises in the dispersive case is handling of high order (possibly mixed) derivatives. The SGN system is an archetype of such systems with sufficient degree of nonlinearity and practically important applications in Coastal Engineering [96]. 3.1. Adaptive mesh construction In the present work we employ the method of moving grids initially proposed in early 60’s by Tikhonov & Samarskii [135, 136] and developed later by Bakhvalov (1969) [7] and Il’in (1969) [85]. This technology was recently described by the authors to steady [90] and unsteady [94] problems. For more details we refer to our recent publications [90, 94]. An alternative recent approach can be found in e.g. [3, 4]. In the present Section we just recall briefly the main steps of the method. The main idea consists in assuming that there exists a (time-dependent) diffeomorphism def from the reference domain Q := [0, 1] to the computational domain I = [0, ℓ]: x(q, t) : Q 7→ I . It is natural to assume that boundaries of the domains correspond to each other, i.e. x(0, t) = 0 , x(1, t) = ℓ , ∀t > 0 . We shall additionally assume that the Jacobian of this map is bounded from below and above def ∂x 0 < m 6 J(q, t) := 6 M < +∞ (3.1) ∂q by some real constants m and M. The construction of this diffeomorphism x(q, t) is the heart of the matter in the moving grid method. We employ the so-called equidistribution method. The required non-uniform grid Ih of the computational domain I is then obtained as the image of the uniformly distributed nodes Qh under the mapping x(q, t): xj = x(qj , t) , qj = j ∆q , ∆q = 1 , N 17 / 66 Dispersive shallow water wave modelling. Part II where N is the total number of grid points. Notice, that strictly speaking, we do not even need to know the mapping x(q, t) in other points besides {qj }N j=0 . Under condition (3.1) it easily follows that the maximal discretization step in the physical space vanishes when we refine the mesh in the reference domain Qh : max | xj+1 − xj | 6 M ∆q → 0 , j=0 ... ,N −1 as ∆q → 0 . 3.1.1 Initial grid generation Initially, the desired mapping x(q, 0) is obtained as a solution to the following nonlinear elliptic problem   d dx ̟(x) = 0, x(0) = 0, x(1) = ℓ , (3.2) dq dq where we drop in this Section the 2nd constant argument 0. The function ̟(x) is the so-called monitor function. Its choice will be specified below, but we can say that this functions has to be positive defined and bounded from below, i.e. ̟(x) > C > 0 , ∀x ∈ R . In practice the lower bound C is taken for simplicity to be equal to 1. A popular choice of the monitor function is, for example, ϑ0 ∈ R+ , ̟[η](x) = 1 + ϑ0 | η | , where η is the free surface elevation. Another possibility consists in taking into account the free surface gradient: ϑ1 ∈ R+ , ̟[η](x) = 1 + ϑ1 | ηx | , or even both effects: ̟[η](x) = 1 + ϑ0 | η | + ϑ1 | ηx | , ϑ0,1 ∈ R+ . In some simple cases equation (3.2) can be solved analytically (see e.g. [90]). However, in most cases we have to solve the nonlinear elliptic problem (3.2) numerically. For this purpose we use an iterative scheme, where at every stage we have a linear three-diagonal problem to solve: 1 ∆q  (n+1) (n) ̟(xj+1/2 ) xj+1 (n+1) − xj ∆q (n+1) − (n) ̟(xj−1/2 ) xj (n+1) − xj−1 ∆q  = 0, n ∈ N0 . (3.3) The iterations are continued until the convergence is achieved to the prescribed tolerance parameter (typically ∝ 10−10 ). 18 / 66 G. Khakimzyanov, D. Dutykh, et al. 3.1.2 Grid motion In unsteady computations the grid motion is given by the following nonlinear parabolic equation:   ∂ ∂x ∂x ̟(x, t) = β , β ∈ R+ . (3.4) ∂q ∂q ∂t The parameter β plays the rôle of the diffusion coefficient here. It is used to control the smoothness of nodes trajectories. Equation (3.4) is discretized using an implicit scheme:   n+1 xn+1 − xnj xn+1 xn+1 − xn+1 1 j j+1 − xj j j−1 n n ̟j+1/2 = β − ̟j−1/2 , (3.5) ∆q ∆q ∆q τ with boundary conditions xn+1 = 0, xn+1 = ℓ as above. We would like to reiterate 0 N that at every time step we solve only one additional (tridiagonal) linear system. Nonlinear iterative computations are performed only once when we project the initial condition on the ad-hoc non-uniform grid. So, the additional overhead due to the mesh motion is linear in complexity, i.e. O(N). Similarly to the elliptic case (3.2), equation (3.4) admits smooth solutions provided that the monitor function ̟(x, t) is bounded from below by a positive constant. In numerical examples shown below we always take monitor functions which satisfy the condition ̟(x, t) > 1, ∀x ∈ I , ∀t > 0 . Thus, for any t > 0 equation (3.4) provides us the required diffeomorphism between the reference domain Q and the computational domain I. 3.2. The SGN equations on a moving grid Before discretizing the SGN equations (2.16)–(2.18), we have to pose them on the reference domain Q. The composed functions will be denoted as:
def ů(q, t) := (u ◦ x) (q, t) ≡ u x(q, t), t .
def And we introduce similar notations for all other variables, e.g. H̊(q, t) := H x(q, t), t . The conservative (2.24) and non-conservative (2.26), (2.27) forms of hyperbolic equations read:   ˚ − xt v̊ (Jv̊)t + F = G˚, (3.6) q  1 ˚ 1 ˚ Fq − xt v̊ q = G, v̊ t + J J   ˚t + 1 A˚· F ˚q − xt v̊ q = 1 A˚· G˚, F J J where the terms on the right-hand sides are defined similarly as above: ! 0 def G˚ := . ˚ − ˚ g H̊ h̊q + ℘ ̺ h̊q q (3.7) (3.8) 19 / 66 Dispersive shallow water wave modelling. Part II The non-hydrostatic pressure on the bottom is computed in Q space as:  ˚ ˚ h̊  ℘ h̊q2 6℘ def 1 def q q ˚ ̺ := + H̊ R̊ + , Υ̊ := 4 + J2 J2 Υ̊ H̊ (3.9) Finally, we just have to specify the expression for R̊:  i η̊q h̊q ů 2 h h̊q i xt  2 ů − xt h xt def R̊ := −g 2 + + h̊t − h̊q h̊q . · h̊t − + J J J q J J J t q We have to specify also the equations which allow us to find the non-hydrostatic part of ˚ . Equation (2.19) posed on the reference domain Q reads: the pressure field ℘  ˚  ˚ = F̊ , K̊ ℘q q − K̊0 ℘ (3.10) where the coefficients and the right-hand side are defined as   h̊   2 J Υ̊ − 3 4 def def q · + , , K̊0 := 6 K̊ := Υ̊ H̊ 3 J H̊ Υ̊ J H̊ 2 Υ̊ q h η̊ ůq2 R̊ h̊q i 6 R̊ J def q F̊ := g − + + 2 . J J J Υ̊ q H̊ Υ̊ Finally, the boundary conditions are specified now at q = 0 and q = 1. For the hyperbolic part of the equations they are ů(0, t) = 0 ů(1, t) = 0 ∀t > 0 . For the elliptic part we have the following mixed-type boundary conditions:   q=1  q=1  R̊ 1 4 ℘ ˚ ˚ − 6 h̊q ℘ . gη̊q + h̊q = q J Υ̊ J H̊ Υ̊ J H̊ 2 Υ̊ q=0 q=0 (3.11) 3.3. Predictor–corrector scheme on moving grids In this Section we describe the numerical finite volume discretization of the SGN equations on a moving grid. We assume that the reference domain Q is discretized with a def  N uniform grid Qh := qj = j ∆q j = 0 , with the uniform spacing ∆q = N1 . Then, the grid Inh in the physical domain I at every time instance t = t n > 0 is given by the image of the uniform grid Qh under the mapping x(q, t) , i.e. xnj = x(qj , t n ) , j = 0, 1, . . . , N or def  N simply Inh = x(Qh , t n ). We assume that we know the discrete solution∗ v̊ ♯n := v̊ jn j = 0 , def  ˚ n N ˚ n := ℘ ℘j j = 0 at the current time t = tn and we already constructed the non-uniform ♯ def  N grid xn+1 := xn+1 at the following time layer t n+1 using the equidistribution method j ♯ j =0 described above. We remind that the non-uniform grid at the following layer is constructed based only on the knowledge of v̊ ♯n . ∗With symbol ♯ we denote the set of solution values at discrete spatial grid nodes. 20 / 66 G. Khakimzyanov, D. Dutykh, et al. 3.3.1 Predictor step In the nonlinear case, during the predictor step the hyperbolic part of equations is solved two times: def  N −1 • First, using equation (3.7) we compute the discrete solution values v̊ ∗♯, c := v̊ ∗j+1/2 j = 0 def  N −1 in the cell centers Qh, c := qj+1/2 = qj + ∆q . 2 j =0 • Then, using equation (3.8) we compute the values of the flux vector equally in the def  N −1 ˚∗ ˚∗ := . F cell centers F ♯, c j+1/2 j = 0 We rewrite equations (3.7), (3.8) in the characteristic form by multiplying them on the left by the matrix L̊ (of left eigenvectors of the Jacobian A˚):   1 ˚q − xt v̊ q = 1 L̊ · G˚, L̊ · F J J   1 ˚t + ˚q − xt v̊ q = 1 Λ̊ · L̊ · G˚, L̊ · F Λ̊ · L̊ · F J J L̊ · v̊ t + The discretization of last equations reads: 1 n 1  n v̊ ∗j+1/2 − v̊ nj+1/2 ˚q − xt v̊ q , = L̊ · F L̊ · G˚ + τ /2 J J j+1/2 j+1/2 (3.12) ˚∗ ˚n n 1 1 F  n
n j+1/2 − Fj+1/2 −1 ˚ ˚ Λ̊ · L̊ · Fq − xt v̊ q Λ̊ · L̊ · G + , = D · L̊ j+1/2 · τ /2 J J j+1/2 j+1/2 (3.13) D−1 · L̊
n · j+1/2 where τ is the time step, L̊nj+1/2 is an approximation of matrix L̊ in the cell centers Qh, c (it will be specified below). The matrix D is composed of cell parameters for each equation: ! ! 1, n −, n 1 + θ 0 1 + λ 0 def def j+1/2 j+1/2 n n Dj+1/2 := , Λ̊j+1/2 := , 2, n n 0 1 + θj+1/2 0 1 + λ+, j+1/2 n with λ±, j+1/2 being the approximations of eigenvalues (2.25) in the cell centers Qh, c (it will be specified below). On the right-hand side the source term is ! 0 def n
n G˚j+1/2 := , ˚ − ˚ g H̊ h̊q + ℘ ̺ h̊ q q j+1/2 where derivatives with respect to q are computed using central differences: n ˚ ℘ q, j+1/2 def := ˚n ℘ j+1 ˚n − ℘ j , ∆q def h̊nq, j+1/2 := h̊nj+1 − h̊nj . ∆q 21 / 66 Dispersive shallow water wave modelling. Part II n The value of the non-hydrostatic pressure trace at the bottom ˚ ̺j+1/2 is computed according ˚n in cell centers are computed as: to formula (3.9). Solution vector v̊ n and the fluxes F ♯, c ♯, c ˚n ˚n + def Fj+1 + Fj ˚n := , F . j+1/2 2 2 The derivatives of these quantities are estimated using simple finite differences: n n ˚n ˚n def Fj+1 − Fj def v̊ j+1 − v̊ j ˚n := , F . v̊ nq, j+1/2 := q, j+1/2 ∆q ∆q Finally, we have to specify the computation of some mesh-related quantities: def v̊ nj+1/2 := xnt, j xn+1 − xnj j , := τ def v̊ nj+1 v̊ nj def xnt, j+1/2 := xnt, j + xnt, j , 2 def Jnj+1/2 ≡ xnq, j+1/2 := xnj+1 − xnj . ∆q n The approximation of the matrix of left eigenvectors L̊nj+1/2 and eigenvalues λ±, j+1/2 depends on the specification of the Jacobian matrix A˚n . Our approach consists in choosing the j+1/2 discrete approximation in order to have at discrete level
n ˚n F A˚· v̊ q j+1/2 , q, j+1/2 ≡ (3.14) which is the discrete analogue of the continuous identity F̊q ≡ A˚ · v̊ q . Basically, our philosophy consists in preserving as many as possible continuous properties at the discrete level. For example, the following matrix satisfies the condition (3.14): !
n 0 1 n A˚j+1/2 = = R̊ · Λ̊ · L̊ j+1/2 n n n n −uj uj+1 + g Hj+1/2 2 uj+1/2 The matrices Lnj+1/2 and Rnj+1/2 = (Lnj+1/2 )−1 are computed by formulas (2.28). The n Jacobian matrix A˚j+1/2 eigenvalues can be readily computed: q q def def n n n n n n 2 − un un := λ±, := (u ± s) , s (u ) + g H > g Hj+1/2 > 0. j+1/2 j+1/2 j j+1 j+1/2 j+1/2 j+1/2 Thanks to the discrete differentiation rule (3.14), we can derive elegant formulas for the ˚∗ by drastically simplifying the scheme (3.12), (3.13): predicted values v̊ ∗♯, c , F ♯, c h
in τ ¯ v̊ ∗j+1/2 = v̊ − R̊ · D · Λ̊ · P̊ − L̊ · G˚ , (3.15) 2J j+1/2 h
in τ ¯ ∗ ˚ ˚ ˚ Fj+1/2 = F − , (3.16) R̊ · D · Λ̊ · Λ̊ · P̊ − L̊ · G 2J j+1/2 where we introduced two matrices:
n def def ¯ n := L̊ · v̊ q j+1/2 . P̊j+1/2 Λ̊nj+1/2 := Λ̊nj+1/2 − xt,nj+1/2 · I , 1,2 Finally, the scheme parameters θj+1/2 are chosen as it was explained in our previous works [94, 129] for the case of Nonlinear Shallow Water Equations. This choice guarantees the TVD property of the resulting scheme. 22 / 66 G. Khakimzyanov, D. Dutykh, et al. ˚∗ , Non-hydrostatic pressure computation. Once we determined the predicted values v̊ ∗♯, c , F ♯, c we have to determine also the predicted value for the non-hydrostatic pressure components ˚ ∗ located in cell centers Q . In order to discretize the elliptic equation (3.10) we apply ℘ h, c ♯, c the same finite volume philosophy. Namely, we integrate equation (3.10) over one cell [qj , qj+1 ]. Right now for simplicity we consider an interior element. The approximation near boundaries will be discussed below. The integral form of equation (3.10) reads ˆ qj+1  qj ˚ ∗  dq K̊ ℘ q q − ˆ qj+1 ˚ ∗ dq K̊0 ℘ ˆ = qj qj+1 (3.17) F̊ dq . qj The coefficients K̊, K̊0 are evaluated using the predicted value of the total water depth n H̊♯,∗ c . If the scheme parameter θj+1/2 ≡ 0 , ∀j = 0, . . . , N − 1, then the predictor value would lie completely on the middle layer t = t n + τ2 . However, this simple choice of n {θj+1/2 }jN=−1 0 does not ensure the desired TVD property [10, 129]. The solution of this integral equation will give us the predictor value for the non˚ ∗ . The finite difference scheme for equation (3.10) is obtained hydrostatic pressure ℘ ♯, c by applying the following quadrature formulas to all the terms in integral equation (3.17): ˆ qj+1  qj ˆ qj+1 qj ˆ qj+1 qj ˚∗ ˚∗ ℘ ℘  − K̊ + K̊ j+3/2 j+1/2 j+3/2 j+1/2 ˚ · K̊ ℘ q q dq ≃ 2 ∆q ∗ K̊j+1/2 + K̊j−1/2 − · 2  ˚ ∗ dq ≃ K̊0 ℘ ∆q · h 12 Jn (H̊ ∗ )3 h + · Υ̊ − 3 i Υ̊ ˚∗ ℘ j+1/2 ˚∗ − ℘ j−1/2 , ∆q j+1/2 3 h̊nq Υ̊ Jn (H̊ ∗ )2 i j+3/2 − h 3 h̊nq Υ̊ Jn (H̊ ∗ )2 i j−1/2  ˚∗ ℘ j+1/2 ,  η̊ ∗  (ů∗ )2  η̊ ∗ R̊ h̊nq  R̊ h̊nq  6 R̊ Jn  q q q − g n + . F̊ dq ≃ ∆q · 2 n − + g n + J J J Υ̊ Jn j+1 Υ̊ Jn j Υ̊ H̊ ∗ j+1/2 In approximation formulas above we introduced the following notations: def K̊j+1/2 := h 4 Υ̊ Jn H̊ ∗ i j+1/2 , def Υ̊j+1/2 := 4 +  h̊n 2 q n J j+1/2 , def Jnj := Jnj+1/2 + Jnj−1/2 2 . 23 / 66 Dispersive shallow water wave modelling. Part II In this way we obtain a three-point finite difference approximation of the elliptic equation (3.10) in interior of the domain, i.e. j = 1, . . . , N − 2 : ˚∗ ˚∗ ˚∗ ˚∗ K̊j+1/2 + K̊j−1/2 ℘j+1/2 − ℘j−1/2 K̊j+3/2 + K̊j+1/2 ℘j+3/2 − ℘j+1/2 · − · 2 ∆q 2 ∆q   n h 12 Jn Υ̊ − 3 i i i h h 3 h̊nq 3 h̊q − ∆q · · = − + j+1/2 Υ̊ (H̊ ∗ )3 Υ̊ Jn (H̊ ∗ )2 j+3/2 Υ̊ Jn (H̊ ∗ )2 j−1/2  (ů∗ )2  η̊ ∗  η̊ ∗ R̊ h̊nq  R̊ h̊nq  6 R̊ Jn  q q q ∆q · 2 n − + g n + − g n + . (3.18) J J J Υ̊ Jn j+1 Υ̊ Jn j Υ̊ H̊ ∗ j+1/2 Two missing equations are obtained by approximating the integral equation (3.17) in intervals adjacent to the boundaries. As a result, we obtain a linear system of equations N −1 ˚∗ where unknowns are {℘ j+1/2 }j = 0 . The approximation in boundary cells will be illustrated on the left boundary [q0 ≡ 0, q1 ]. The right-most cell [qN −1 , qN ≡ 1] can be treated similarly. Let us write down one-sided quadrature formulas for the first cell: K̊3/2 + K̊1/2 · 2 + ˚∗ ℘ 3/2 h ˚∗ ˚∗ − ℘ 4℘ 1/2 q − ∗ ∆q J H̊ Υ̊ {z | 1 3 h̊nq Υ̊ Jn (H̊ ∗ )2 i 3/2 + h q=0 }  h 12 Jn Υ̊ − 3 i ∗ ˚ ℘ − 1/2 ∆q · · 1/2 Υ̊ (H̊ ∗ )3 3 h̊nq Υ̊ Jn (H̊ ∗ )2  (ů∗ )2  η̊ ∗ 6 R̊ Jn  q q = ∆q · 2 n − + g n ∗ J J Υ̊ H̊ 1/2 i  + ˚∗ 6 h̊nq ℘ J (H̊ ∗ )2 Υ̊ q=0 {z } | 2  η̊ ∗ R̊ h̊nq  R̊ h̊nq  q + − g n + J Υ̊ Jn 1 Υ̊ Jn {z | 3 1/2 . q =0 } It can be readily noticed that terms 1 + 2 + 3 vanish thanks to the boundary condition (3.11) (the part at q = 0). The same trick applies to the right-most cell [qN −1 , qN ≡ 1]. We reiterate on the fact that in our scheme the boundary conditions are taken into account exactly. Consequently, in two boundary cells we obtain a two-point finite difference approximation to equation (3.10). The resulting linear system of equations can be solved using e.g. the direct Thomas algorithm with linear complexity O(N). Under 6 0 the numerical solution exists, it is > 0 , h̊q the conditions K̊0 > 0 , h̊q q =0 unique and stable [122]. q=1 24 / 66 G. Khakimzyanov, D. Dutykh, et al. 3.3.2 Corrector step During the corrector step we solve again separately the hyperbolic and elliptic parts of the SGN equations. In order to determine the vector of conservative variables v̊ n+1 we use ♯ an explicit finite volume scheme based on the conservative equation (3.6): ˚∗ − xt · v̊ ∗ F (Jv̊)n+1 − (Jv̊)nj j + τ
j+1/2 ˚∗ − xt · v̊ ∗ F − ∆q
j−1/2 = G˚j∗ , (3.19) where def G˚j∗ :=    0 ,
˚∗ − ˚ ̺∗ h̊nq + ♭ + ℘ q g H̊ n + ♭ ∗ ˚ ℘ q, j def := ˚∗ ℘ j+1/2 j ˚∗ − ℘ j−1/2 , ∆q and n+1 n+1 n n H̊j+1 + H̊j−1 + 2 H̊jn+1 + 2 H̊jn + H̊j+1 + H̊j−1 , 8 n+1 n+1 n n def h̊j+1 − h̊j−1 + h̊j+1 − h̊j−1 := . 4 ∆q def H̊jn + ♭ := h̊nq + ♭ (3.20) (3.21) The algorithm of the corrector scheme can be summarized as follows: (1) From the mass conservation equations (the first component in (3.19)) we find the total water depth H̊♯n+1 in interior nodes of the grid (2) Using the method of characteristics and the boundary conditions ůn+1 = ůn+1 ≡ 0 0 N n+1 n+1 we determine the total water depth H̊0 , H̊N in boundary points q0 ≡ 0 and qN ≡ 1 (3) Then, using the momentum conservation equation (the second component in (3.19)) we find the momentum values (H̊ ů)n+1 on the next time layer. ♯ In this way, we obtain an explicit scheme despite the fact that the right hand side G˚♯∗ depends on the water depth H̊♯n+1 at the new time layer t = tn+1 . ˚ n+1 is comNon-hydrostatic pressure correction. The non-hydrostatic pressure correction ℘ ♯ puted by integrating locally the elliptic equation (3.10) around each grid point: ˆ qj+1/2  qj−1/2 ˚ n+1  dq − K̊ ℘ q q ˆ qj+1/2 qj−1/2 ˚ n+1 dq = K̊0 ℘ ˆ qj+1/2 qj−1/2 F̊ n+1 dq , j = 1, . . . , N −1 , 25 / 66 Dispersive shallow water wave modelling. Part II The details of integrals approximations are similar to the predictor step described above. Consequently, we provide directly the difference scheme in interior nodes: ˚ n+1 ℘ j+1 ˚ n+1 ˚ n+1 − ℘ ˚ n+1 ℘ − ℘ j j j−1 Kj+1/2 − Kj−1/2 = ∆q ∆q   n+1 n+1  n+1  ( Υ̊ − 3) J h̊ ( Υ̊ − 3) J h̊ q q ˚ − 6 ℘j ∆q = + ∆q − + Υ̊ H̊ 3 Υ̊ J H̊ 2 j−1/2 Υ̊ H̊ 3 Υ̊ J H̊ 2 j+1/2  η̊  ů2  η̊ 6 R̊ J n+1 R̊ h̊q n+1 R̊ h̊q n+1 q q q − + + − g , (3.22) ∆q 2 + g J J J Υ̊ J j+1/2 Υ̊ J j−1/2 Υ̊ H̊ j where 4 def Kj+1/2 := Υ̊ J H̊
n+1 , def Υ̊n+1 j+1/2 := 4+ j+1/2 h h̊n+1 − h̊n+1 i2 j+1 j n+1 xn+1 j+1 − xj , def Jn+1 j+1/2 := n+1 xn+1 j+1 − xj . ∆q In order to complete the scheme description, we have to specify the discretization of the elliptic equation (3.10) in boundary cells. To be specific we take again the left-most cell [q0 ≡ 0, q1/2 ]. The integral equation in this cell reads: ˆ q1/2 ˆ q1/2 ˆ q1/2  ˚ n+1  n+1 ˚ ℘ ℘ K̊ q dq − K̊0 dq = F̊ n+1 dq . q q0 q0 q0 And the corresponding difference equation is K1/2 ˚ n+1 ℘ 1  n+1 ˚ n+1 ˚ n+1 ˚ n+1 ℘ n+1 4℘ − ℘ ( Υ̊ − 3) J h̊ 6 h̊ q q q 0 ˚ − 6 ℘0 = − ∆q + + ∆q Υ̊ H̊ 3 J H̊ Υ̊ q = 0 J H̊ 2 Υ̊ 1/2 J H̊ 2 Υ̊ q = 0 | | {z } {z } 31 32 n+1  n+1  ů2  η̊ R̊ h̊q R̊ h̊q  3 R̊ J  η̊q q q . + − + + ∆q g − g J J J Υ̊ J Υ̊ J q = 0 Υ̊ H̊ 1/2 | {z } 33 By taking into account the boundary condition (3.11) we obtain that three under-braced terms vanish: 31 + 32 + 33 ≡ 0 . A similar two-point approximation can be obtained by integrating over the right-most cell h   ∆q i qN −1/2 , qN ≡ 1 − , 1 . In this way we obtain again a three-diagonal system of 2 linear equations which can be efficiently solved with the Thomas algorithm [81]. 26 / 66 G. Khakimzyanov, D. Dutykh, et al. Stability of the scheme. In order to ensure the stability of (nonlinear) computations, we impose a slightly stricter restriction on the time step τ than the linear analysis given below predicts (see Section 3.4.1). Namely, at every time layer we apply the same restriction as for hyperbolic (non-dispersive) Nonlinear Shallow Water Equations [94]: n, ± max{ Cj+1/2 } 6 1, j n, ± where Cj+1/2 are local Courant numbers [40] which are defined as follows τ h | λ± − xt | in def n, ± := Cj+1/2 . ∆q J j+1/2 3.3.3 Well-balanced property It can be easily established that the predictor–corrector scheme presented above preserves exactly the so-called states ‘lake-at-rest’: Lemma 1. Assume that the bottom is stationary ( i.e. ht ≡ 0 , but not necessary flat) and initially the fluid is at the ‘lake-at-rest’ state, i.e. η̊j0 ≡ 0 , ůj0 ≡ 0 j = 0, 1, 2, . . . , N . (3.23) Then, the predictor–corrector scheme will preserve this state at all time layers. Proof. In order to prove this Lemma, we employ the mathematical induction [80]. First, we have to discuss the generation of the initial grid and how it will be transformed to the next time layer along with the discrete numerical solution: x0♯ ֒→ x1♯ , v̊ 0♯ ֒→ v̊ ∗c, ♯ ֒→ v̊ 1♯ . Then, by assuming that our statement is true at the nth time layer, we will have to show that it is true on the upcoming (n + 1)th layer. This will complete the proof [80]. If the monitoring function ̟(x, t) depends only on the free surface elevation η(x, t) and fluid velocity u(x, t), then the monitoring function ̟(x, t) ≡ 1 thanks to Lemma assumption (3.23). And the equidistribution principle (3.2) will give us the uniform mesh. However, in most general situations one can envisage the grid adaptation upon the bathymetry profile∗ h(x, t). Consequently, in general we can expect that the mesh will be non-uniform even under condition (3.23), since hx 6= 0 . However, we know that the initial grid satisfies the fully converged discrete equidistribution principle (3.3). From now on we assume that the initial grid is generated and it is not necessarily uniform. In order to construct the grid at the next layer, we solve just one linear equation (3.5). Since, system (3.5) is diagonally dominant, its solution exists and it is unique [122]. It is not difficult to check that the set of values {xj1 ≡ xj0 }N j = 0 solves the system (3.5). It follows from two observations: • The right-hand side of (3.5) vanishes when x1j ≡ x0j , ∀j = 0, . . . , N . 0 • The monitor function {̟j+1/2 }jN=−10 is evaluated on the previous time layer t = 0 . ∗In the present study we do not consider such example. However, the idea of grid adaptation upon the bathymetry function certainly deserves to be studied more carefully. 27 / 66 Dispersive shallow water wave modelling. Part II Thus, we obtain that x♯1 ≡ x♯0 . Consequently, we have xt,0 j ∀j = 0, . . . , N . ≡ 0 and Jj1 = Jj0 , ˚ 0 and ˚ In order to complete the predictor step we need to determine the quantities ℘ ̺♯0 ♯ 0 on which depends the source term G˚j+1/2 . These quantities are uniquely determined by ˚ 0 are obtained by solving linear equations prescribed initial conditions. For instance, ℘ ♯ (3.22). We showed above also that the solution to this equation is unique. We notice also that the right-hand side in equation (3.22) vanishes under conditions of this Lemma. ˚ 0 ≡ 0 . By applying a finite difference analogue of equation Consequently, we obtain ℘ ♯ (3.9) we obtain also that ̺♯0 ≡ 0 . As the result, at the ‘lake-at-rest’ state the right-hand side of predictor equations (3.12), (3.13) reads ! 0 0 G˚j+1/2 = . 0 (g h̊ h̊q )j+1/2 Taking into account the fact that the mesh does not evolve x0♯ ֒→ x1♯ ≡ x0♯ , we obtain q ¯0 0 0 g h̊j+1/2 , xt,0 j ≡ 0 and thus Λ̊j+1/2 ≡ Λ̊j+1/2 , sj+1/2 ≡ ! !

h̊q, j+1/2 h̊q, j+1/2 0 0 ¯ . , L̊ · G˚ j+1/2 ≡ Λ̊ · P̊ j+1/2 ≡ h̊q, j+1/2 h̊q, j+1/2 Consequently, the predictor step (3.15), (3.16) gives us the following values: 0 v̊ ∗j+1/2 ≡ v̊ j+1/2 , ˚∗ ˚0 F j+1/2 ≡ Fj+1/2 . For the sake of clarity, we rewrite the last predictions in component-wise form: ! ! 0 h̊ j+1/2 ˚∗ 2 v̊ ∗j+1/2 ≡ , F . g h̊j+1/2 j+1/2 ≡ 0 2 ∗ Thus, H̊j+1/2 ≡ h̊j+1/2 . As an intermediate conclusion of the predictor step we have: ∗ ηj+1/2 ≡ 0, ů∗j+1/2 ≡ 0 , ˚ ∗, ˚ ∗ and all dispersive corrections ℘ ♯ ̺♯ vanish as well by applying similar arguments to equation (3.18). The corrector step (3.19), written component-wise reads: (J ů H̊)j1 − (J ů H̊)j0 τ (J H̊)j1 − (J H̊)j0 = 0, τ 2 2 g h̊j+1/2 − g h̊j−1/2
♭ + = g H̊ h̊q j 2 ∆q From the first equation above taking into account that Jj1 ≡ Jj0 and H̊j0 = h̊j we obtain H̊j1 = h̊j . And thus, by the definition of the total water depth we obtain η̊j1 ≡ 0 . In 28 / 66 G. Khakimzyanov, D. Dutykh, et al. the second equation above by condition (3.23) we have that ůj0 ≡ 0 . Moreover, in the left-hand side: 2 2 g h̊j+1/2 − g h̊j−1/2 2 ∆q = g (h̊j+1 − h̊j−1 ) · (h̊j+1 + 2 h̊j + h̊j−1 ) . 8 ∆q (3.24) The right-hand side of the same corrector equation can be rewritten using definitions (3.20), (3.21) as
♭ 2 h̊j+1 + 4 h̊j + 2 h̊j−1 2 h̊j+1 − 2 h̊j−1 g H̊ h̊q j = g · . (3.25) 8 4 ∆q Comparing equation (3.24) with (3.25) yields the desired well-balanced property of the predictor–corrector scheme and thus ůj1 ≡ 0 . By assuming that (3.23) is verified at the time layer t = t n and repeating precisely the same reasoning as above (by substituting superscripts 0 ← n and 1 ← n + 1) we obtain that (3.23) is verified at the next time layer t = t n+1 . It completes the proof of this Lemma.  We would like to mention that the well-balanced property of the proposed scheme was checked also in numerical experiments on various configurations of general uneven bottoms (not reported here for the sake of manuscript compactness) — in all cases we witnessed the preservation of the ‘lake-at-rest’ state up to the machine precision. This validates our numerical implementation of the proposed algorithm. 3.4. Numerical scheme for linearized equations In order to study the numerical scheme stability and its dispersive properties, we consider the discretization of the linearized SGN equations on a uniform unbounded grid (for simplicity we consider an IVP without boundary conditions). The governing equations after linearization can be written as (we already gave these equations in Section 2.5) ηt + d ux = 0 , ut + g ηx = ℘xx − 3 d2 ℘ 1 d ℘x , = c2 ηxx , √ where c = g d is the speed of linear gravity waves. We shall apply to these PDEs precisely the same scheme as described above. Since the grid is uniform, we can return to the original notation, i.e. v̊ ≡ v, etc. Let ∆x be the discretization step in the computational domain Ih and τ is the local time step. We introduce the following finite difference operators (illustrated on the free surface elevation η ♯n ): ηt,nj ηjn+1 − ηjn := , τ def ηx,n j n ηj+1 − ηjn := , ∆x def n η(x), j n n ηj+1 − ηj−1 := , 2 ∆x def 29 / 66 Dispersive shallow water wave modelling. Part II n n n n n n ηj+1 − 2 ηjn + ηj−1 def ηxx, j + ηxx, j+1 def ηxx, j+1 − ηxx, j n n , η , η . := := xx, j+1/2 xxx, j ∆x2 2 ∆x +∞ ∗ ∗ Then, at the predictor step we compute auxiliary quantities {ηj+1/2 }+∞ j = −∞ , {uj+1/2 }j = −∞ ∗ and {℘j+1/2 }+∞ j = −∞ . First, we solve the hyperbolic part of the linearized SGN equations:
∗ n ηj+1/2 − 21 ηj+1 + ηjn + d unx, j = 0 , ∗ τj+1/2
u∗j+1/2 − 12 unj+1 + unj 1 ℘n + g ηx,n j = , ∗ τj+1/2 d x, j def n ηxx, j := and then we solve the elliptic equation to find {℘j+1/2 }+∞ j = −∞ : ∗ ℘∗j+3/2 ∗ ∗ ∗ ∗ η∗ − 2 ηj+1/2 + ηj−1/2 − 2 ℘j+1/2 + ℘j−1/2 3 ℘∗ 2 j+3/2 − 2 j+1/2 = c , ∆x2 d ∆x2 def τ ∗ n n where τj+1/2 := (1 + θj+1/2 ) and θj+1/2 is the numerical scheme parameter [94], 2 whose choice guarantees the TVD property (strictly speaking the proof was done for scalar hyperbolic equations only). Then, the predicted values are used on the second — corrector step, to compute all n+1 n+1 +∞ physical quantities {ηjn+1 }+∞ }j = −∞ and {℘j }+∞ j = −∞ , {uj j = −∞ on the next time layer n+1 t = t : u∗j+1/2 − u∗j−1/2 = 0, (3.26) ηt,nj + d ∆x ∗ ∗ ∗ ∗ ηj+1/2 − ηj−1/2 1 ℘j+1/2 − ℘j−1/2 n ut, j + g = , (3.27) ∆x d ∆x 3 ℘n+1 n+1 ℘n+1 = c2 ηxx, (3.28) xx, j − j . d2 j It can be easily checked that the scheme presented above has the first order accuracy if n n θj+1/2 = const , ∀j and the second order if θj+1/2 ≡ 0 , ∀j . However, the last condition can be somehow relaxed. There is an interesting case of quasi-constant values of the scheme parameter: n θj+1/2 = O (τ + ∆x) . In this case the scheme is second order accurate as well. In the present Section we perform n a theoretical analysis of the scheme and we shall assume for simplicity that θj+1/2 ≡ const. Consequently, from now on we shall drop the index j + 1/2 in the intermediate time step ∗ τj+1/2 . 3.4.1 Linear stability of the scheme In this Section we apply the so-called von Neumann stability analysis to the predictor– corrector scheme described above [28]. In order to study the scheme stability, first we 30 / 66 G. Khakimzyanov, D. Dutykh, et al. +∞ ∗ ∗ exclude algebraically the predicted values {ηj+1/2 }+∞ j = −∞ , {uj+1/2 }j = −∞ from difference equations. The resulting system reads: n ∗℘ ηt,nj + d un(x), j = τ ∗ c2 ηxx, j − τ xx, j , n unt, j ℘∗j+3/2 − 2 ℘∗j+1/2 + ℘∗j−1/2 + g n η(x), j ∗ 2 = τ c unxx, j − ℘j−1/2 , (3.30) ∆x
− τ ∗ d unxxx, j . (3.31) 1 + d ℘∗j+1/2 (3.29) ∗ 3 ℘∗ n = c2 ηxx, j+1/2 d2 j+1/2 We substitute in all difference relations above the following elementary harmonics ∆x 2 ηjn = Λ0 ρn e i j ξ , − unj = Ψ0 ρn e i j ξ , ℘nj = Φ0 ρn ei j ξ , ℘∗j+1/2 = Φ∗0 (ρ) e i (j+1/2) ξ , (3.32) def where ξ := k · ∆x ∈ [0, π] is the scaled wavenumber and ρ is the transmission factor between the time layers tn and tn+1 . As a result, from equations (3.28) and (3.31) we obtain the following expressions for Φ0 and Φ∗0 : h i 2 2 4 c2 d 2 2 ∗ n 2c d ∗ 4d 2 Φ0 = ‫ג‬ Λ , Φ (ρ) = ρ ‫ג‬ Λ sin(ξ) − i τ ‫ג‬ Ψ , 0 0 0 0 3 ℏ ∆x 2 3 ℏ ∆x 2 ∆x where we introduced some short-hand notations: ξ  4 d2 2 τ def def def def def , k := 4 c2 ℵ 2 ‫ ג‬2 , i := c ℵ sin(ξ) , ℏ := 1 + , ‫ ג‬:= sin ‫ ג‬. ℵ := ∆x 2 3 ∆x 2 By substituting just obtained expressions for Φ0 and Φ∗0 into equations (3.29), (3.30) we obtain two linear equations with respect to amplitudes Λ0 and Ψ0 : h 2 c2 ℵ2 (1 + θ) 2 i ρ − 1 + ‫ ג‬Λ0 + i ℵ d sin(ξ) Ψ0 = 0 , ℏ h 2 c2 ℵ2 (1 + θ) 2 i g ℵ sin(ξ) Λ0 + ρ − 1 + ‫ ג‬Ψ0 = 0 . i ℏ ℏ The necessary condition to have non-trivial solutions gives us an algebraic equation for the transmission factor ρ: k (1 + θ) k2 (1 + θ)2 i2 (ρ − 1) + + = 0. ℏ 4 ℏ2 ℏ This quadratic equation admits two distinct roots: (ρ − 1)2 + ρ± = 1 − k (1 + θ) i ± i√ . 2ℏ ℏ (3.33) The necessary stability condition | ρ | 6 1 is equivalent to the following condition on quadratic equation coefficients: h 4 ς2 i d def 2 2 2 c ℵ (1 + θ) ζ − 1 + ζ (ζ + θ) 6 0 , ς := , (3.34) 3 ∆x def which has to be fulfilled for all ζ := ‫ ג‬2 ∈ [0, 1]. The parameter ς characterizes the grid resolution relative to the mean water depth. This parameter appears in stability condition Dispersive shallow water wave modelling. Part II 31 / 66 along with the Courant ratio ℵ. It is one of the differences with non-dispersive equations whose discretization stability depends only on ℵ. Further thoughts about stability. When long waves travel towards the shoreline, their shoal- ing process is often accompanied with the formation of undular bores [75, 116]. Undular bores have dispersive nature and cannot be correctly described by dispersionless models. In [75] it was shown that satisfactory description of dispersive effects in shallow water d def = 2 ∼ 4 . In another study [70] it was shown environments is obtained for ς := ∆x that for satisfactory modeling of trans-oceanic wave propagation it is sufficient to choose ς ≈ 4 in deep ocean and ς ≈ 1 in shallow coastal areas. In other words, it is sufficient to choose the grid size equal to water depth in shallow waters and in deep areas — four times smaller than the water depth. On coarser grids the numerical dispersion may dominate √ 3 over the physical one [70]. In the present study we shall assume that parameter ς > . 2 Substituting into equation (3.34) the value ζ ≡ 0 we obtain that for stability reasons necessarily the scheme parameter θ > 0 . Since the predictor layer should be in between time layers t = tn and t = tn+1 we have θ 6 1 . Then, for fixed values of parameters ς and θ the stability condition (3.34) takes the following form: r 4 1 + ς2 θ 3 . cℵ 6 1 + θ For θ ≡ 0 the last condition simply becomes: cℵ 6 1, and it does not depend on parameter ς. However, when θ > 0, then the scheme stability depends on the mesh refinement ς relative to the mean water depth. Surprisingly, more we refine the grid, less stringent becomes the stability barrier. In the asymptotic limit ς ≫ 1 we obtain the following restriction on the time step τ : √ 2 θ 1 τ 6 √ τ0 < √ τ0 ≈ 0.58 τ0 , 3 (1 + θ) 3 def where τ0 := dc ≡ √dgd is the characteristic time scale of gravity waves. Above we used the following obvious inequality: √ 1 + θ > 2 θ, ∀ θ ∈ R+ . So, in practice for sufficiently refined grids the stability condition de facto does not involve the grid spacing ∆x anymore. This property is very desirable for numerical simulations. For the sake of comparison we give here (without underlying computations) the stability restriction of the same predictor–corrector scheme for NSWE equations: cℵ 6 √ 1 . 1 + θ 32 / 66 G. Khakimzyanov, D. Dutykh, et al. So, another surprising conclusion obtained from this linear stability analysis is that the SGN equations require in fine a less stringent condition on the time step than corresponding dispersionless NSWE. Most probably, this conclusion can be explained by the regularization effect of the dispersion. Indeed, the NSWE bores are replaced by smooth undular bores whose regularity is certainly higher. The smoothness of solutions allows to use a larger time step τ to propagate the numerical solution. This conclusion was checked in (fully nonlinear) numerical experiments (not reported here) where the time step τ was artificially pushed towards the stability limits. In general, the omission of dispersive effects yields a stricter stability condition. The authors of [71] came experimentally to similar conclusions about the time step limit in dispersive and hydrostatic simulations. Our theoretical analysis reported above may serve as a basis of rational explanation of this empirical fact. This result is to be compared with a numerical scheme proposed in [41] for a weakly nonlinear weakly dispersive water wave model. They used splitting technique and solved an elliptic equation to determine the non-hydrostatic pressure correction. The main drawback of the scheme proposed in [41] is the stability condition: ∆x > 1.5 d . One can easily see that a numerical computation with a sufficiently refined grid is simply impossible with that scheme. Our method is free of such drawbacks. 3.4.2 Discrete dispersion relation The dispersion relation properties are crucial to understand and explain the behaviour of the numerical solution [101]. In this Section we perform the dispersion relation analysis of the proposed above predictor–corrector scheme. This analysis is based on the study of elementary plane-wave solutions (3.32). The continuous case was already analyzed in Section 2.5. Dispersive properties of the scheme can be completely characterized by the def phase error ∆ϕ := φ − ϕ committed during solution transfer from time layer t = tn to t = tn+1 = tn + τ . Here we denote by φ the phase shift due to the SGN equations dynamics and ϕ is its discrete counterpart. From equations (2.30) and (3.33) we obtain correspondingly: φ = arg(e−i ω τ ) ≡ −ω τ = ± r cℵξ , ξ ∈ [0, π] , ς 2 ξ2 1 + 3 i h k (1 + θ)  /| ρ | , ϕ = arg ρ = ± arccos 1 − 2ℏ (3.35) (3.36) In other words, the phase change φ is predicted by the ‘exact’ SGN equations properties, while ϕ comes from the approximate dynamics as predicted by the predictor–corrector Dispersive shallow water wave modelling. Part II 33 / 66 scheme. Since we are interested in long wave modelling, we can consider Taylor expansions of the phase shifts in the limit ξ → 0 (assuming that ς and ℵ are kept constant): h i cℵ 2 3 φ = ± cℵξ − ς ξ + O(ξ 4 ) , 6 i h
cℵ (c ℵ)2 (3 θ + 1) − 1 − ς 2 ξ 3 + O(ξ 4 ) . ϕ = ± cℵξ + 6 The asymptotic expression for the phase error is obtained by subtracting above expressions: ∆ϕ = ∓  cℵ  (c ℵ)2 (3 θ + 1) − 1 ξ 3 + O(ξ 4 ) . 6 From the last relation one can see that the leading part of the phase error has the same asymptotic order as the ‘physical’ dispersion of the SGN equations. In general, this result is not satisfactory. However, this situation can be improved if for the given scheme parameter θ > 0, the Courant ratio ℵ is chosen according to the following formula: cℵ = √ 1 . 1 + 3θ In this case the numerical phase error will be one order lower than the physical dispersion of the SGN system. In Figure 2 we represent graphically phase shifts predicted by various models. The dashed line (1) is the phase shift of the predictor–corrector scheme given by equation (3.36) (taken with + sign) for the parameters values θ = 0 , c ℵ = 1 , ς = 2 . The continuous dispersion relation are shown with the dotted line (3) (the SGN equations, formula (3.35)) and the solid line (4) (full Euler equations): s tanh(ς ξ) φEuler = ± c ℵ ξ . ςξ It can be seen that our predictor–corrector scheme provides a better approximation to the dispersion relation than the scheme proposed by Peregrine [116] (dash-dotted line (2) in Figure 2). The analysis of the discrete dispersion relation of Peregrine’s scheme is not given here, but we provide only the final result for the phase change:  i2 . φPeregrine = ± arccos 1 − 2ℏ In Figure 2 one can see that the predictor–corrector scheme (curve (1)) approximates well the dispersion relation of the SGN equations (curve (3)) up to ξ = k · ∆x > π4 . In terms of the wave length λ we obtain that λ ? 8 ∆x and for ς = 2 we obtain the inequality λ ? 4 d. So, as the main result of the present analysis we conclude that our scheme is able to propagate accurately water waves whose length is four times longer than the mean water depth d. 34 / 66 G. Khakimzyanov, D. Dutykh, et al. Φ 1.2 4 3 0.8 1 0.4 2 0.0 0 1 2 ξ 3 Figure 2. Phase shifts in different models: (1) predictor–corrector scheme; (2) Peregrine’s numerical scheme [116]; (3) the SGN equations; (4) full Euler equations. 4. Numerical results Below we present a certain number of test cases which aim to validate and illustrate the performance of the numerical scheme described above along with our implementation of this method. 4.1. Solitary wave propagation over the flat bottom As we saw above in Section 2.4.1, in a special case of constant water depth h(x, t) = d the SGN equations admit solitary wave solutions (given by explicit simple analytical formulas) which propagate with constant speed without changing their shapes. 4.1.1 Uniform grid These analytical solutions can be used to estimate the accuracy of the fully discrete numerical scheme. Consequently, we take a sufficiently large domain [0, ℓ] with ℓ p = 80. In this Section all lengths are relative to the water depth d, and time is scaled with g/d. 35 / 66 Dispersive shallow water wave modelling. Part II 0.4 0.4 η /d η /d 1 2 3 4 0.3 0.2 0.2 0.1 0.1 0 0 55 60 65 70 x/d 1 2 3 4 0.3 75 10 (a) 15 20 t(g/d) 1/2 25 (b) Figure 3. Propagation of a solitary wave over the flat bottom: (a) free surface profile at t = 20; (b) wave gauge data at x = 60. Various lines denote: (1) — N = 80, (2) — N = 160, (3) — N = 320, (4) — the exact analytical solution given by formula (2.29). For instance, if the solitary wave amplitude α = 0.7, then α d = 0.7 d in dimensional variables. So, the solitary wave is initially located at x0 = 40. In computations below we take a solitary wave of amplitude α = 0.4. In general, the SGN travelling wave solutions approximate fairly well those of the full Euler model up to amplitudes α > 21 (see [49] for comparisons). In Figure 3 we show a zoom on free surface profile (a) at t = 20 and wave gauge data (b) in a fixed location x = 60 for various spatial (and uniform) resolutions. By this time, the solitary wave propagated the horizontal distance of 20 mean water depths. It can be seen that the numerical solution converges to the analytical one. In order to quantify the accuracy of the numerical solution we measure the relative l∞ discrete error: def k εh k∞ := α−1 k ηh − η k∞ , where ηh stands for the numerical and η – for the exact free surface profiles. The factor α−1 is used to obtain the dimensionless error. Then, the order of convergence k can be estimated as   k ε h k∞ k ≃ log2 . k εh/2 k∞ The numerical results in Table 1 indicate that k → 2, when N → +∞. This validates the proposed scheme and the numerical solver. 36 / 66 G. Khakimzyanov, D. Dutykh, et al. N ς 80 160 320 640 1280 2560 1 2 4 8 16 32 k ε h k∞ 0.2442 0.1277 0.3344 × 10−1 0.8639 × 10−2 0.2208 × 10−2 0.5547 × 10−3 k — 0.94 1.93 1.95 1.97 1.99 Table 1. Numerical estimation of the convergence order for the analytical d solitary wave propagation test case. The parameter ς = ∆x characterizes the mesh resolution relative to the mean water depth d. 4.1.2 Adaptive grid In order to show the performance of the adaptive algorithm, we adopt two monitor functions in our computations: ̟0 [ η ] (x, t) = 1 + ϑ0 | η(x, t) | , ̟1 [ η ](x, t) = 1 + ϑ0 | η(x, t) | + ϑ1 | ηx (x, t) | , (4.1) (4.2) where ϑ 0, 1 > 0 are some positive constants. In numerical simulations we use ϑ 0 = ϑ 1 = 10 and only N = 80 grid points. Above we showed that numerical results are rather catastrophic when these 80 grid points are distributed uniformly (see Figure 3). Numerical results on adaptive moving grids obtained with monitor functions ̟ 0, 1 (x, t) are shown in Figure 4. The monitor function ̟ 0 (x, t) ensures that points concentrate around the wave crest, leaving the areas in front and behind relatively rarefied. The visual comparison of panels 4(b) and 4(c) shows that the inclusion of the spatial derivative ηx into the monitor function ̟1 (x, t) yields the increase of dense zones around the wave crest. With an adaptive grid involving only N = 80 points we obtain a numerical solution of quality similar to the uniform grid with N = 320 points. 4.2. Solitary wave/wall interaction For numerous practical purposes in Coastal Engineering it is important to model correctly wave/structure interaction processes [118]. In this Section we apply the above proposed numerical algorithm to the simulation of a simple solitary wave/vertical wall interaction. The reason is two-fold: (1) Many coastal structures involve vertical walls as building elements, (2) This problem is well studied by previous investigators and, consequently, there is enough available data/results for comparisons 37 / 66 Dispersive shallow water wave modelling. Part II 20 t(g/d) η /d t(g/d) 1/2 1/2 20 0.4 1 2 3 4 0.3 0.2 0.1 15 15 10 10 5 5 0 55 60 65 (a) 70 x/d 75 0 30 40 50 (b) 60 x/d 70 0 30 40 50 60 x/d 70 (c) Figure 4. Propagation of a solitary wave over the flat bottom simulated with moving adapted grids: (a) free surface profile at t = 20; (b) trajectory of some grid points predicted with monitor function ̟0 (x, t) ; (c) the same but with monitor function ̟1 (x, t) . On panel (a) the lines are defined as: (1) — numerical solution on a uniform fixed grid; (2) — numerical solution predicted with monitor function ̟0 (x, t) ; (3) — the same with ̟1 (x, t) ; (4) — exact analytical solution. We would like to underline that this problem is equivalent to the head-on collision of two equal solitary waves due to simple symmetry considerations. This ‘generalized’ problem was studied in the past using experimental [111, 125], numerical [26, 67] and analytical techniques [21, 113, 132]. More recently this problem gained again some interest of researchers [22, 25, 39, 52, 57, 58, 107, 138]. Despite the simple form of the obstacle, the interaction process of sufficiently large solitary waves with it takes a highly non-trivial character as it will be highlighted below. Figure 5(a) shows the free surface dynamics as it is predicted by the SGN equations solved numerically using the moving grid with N = 320 nodes. The initial condition consists of an exact analytical solitary wave (2.29) of amplitude α = 0.4 moving rightwards to the vertical wall (where the wall boundary condition u = 0 is imposed∗ on the velocity, for the pressure see Section 2.3.1). The computational domain is chosen to be sufficiently large [0, ℓ] = [0, 80], so there is no interaction with the boundaries at t = 0. Initially the solitary wave is located at x0 = 40 (right in the middle). The bottom is flat h(x, t) = d = const in this test case. From Figure 5(a) it can be clearly seen that the reflection process generates a train of weakly nonlinear waves which propagate with different speeds in agreement with the dispersion relation. The moving grid was constructed using the monitor function ̟1 (x, t) from the previous Section (see the definition in equation (4.2)). with ϑ0 = ϑ1 = 10. The resulting trajectories of mesh nodes are shown in Figure 5(b). The grid is clearly refined around the solitary wave and nodes follow it. Moreover, we ∗The same condition is imposed on the left boundary as well, even if during our simulation time there are no visible interactions with the left wall boundary. 38 / 66 G. Khakimzyanov, D. Dutykh, et al. (a) (b) Figure 5. Solitary wave (amplitude α = 0.4)/vertical wall interaction in the framework of the SGN equations: (a) space-time plot of the free surface elevation; (b) nodes trajectories. For the sake of clarity every 5th node is shown only, the total number of nodes N = 320. would like to note also a slight mesh refinement even in the dispersive tail behind the reflected wave (it is not clearly seen in Figure 5(b) since we show only every 5th node). One of the main interesting characteristics that we can compute from these numerical experiments is the maximal wave run-up R on the vertical wall: def R := sup 0 6 t 6 T {η(ℓ, t)} . The sup is taken in some time window when the wave/wall interaction takes place. For the class of incident solitary wave solutions it is clear that maximal run-up R will depend on the (dimensionless) solitary wave amplitude α. In [132] the following asymptotic formula was derived in the limit α → 0:   R (α) = 2 α 1 + 14 α + 83 α2 + O(α4 ) . (4.3) The last approximation was already checked against full the Euler simulations [39, 67] and even laboratory experiments [111]. Figure 6 shows the dependence of the maximal run-up R on the incident solitary wave amplitude α as it is predicted by our numerical model, by formula (4.3) and several other experimental [42, 110, 111, 144] and numerical [26, 39, 67] studies. In particular, one can see that almost all models agree fairly well up to the amplitudes α > 0.4. Then, there is an apparent ‘separation’ of data in two branches. Again, our numerical model gives a very good agreement with experimental data from [110, 111, 144] up to the amplitudes α > 0.7. 39 / 66 Dispersive shallow water wave modelling. Part II 2.4 R/d 1 2 3 4 5 6 7 8 9 2.0 1.6 1.2 0.8 0.4 0.0 0 0.1 0.2 0.3 0.4 0.5 0.6 α /d 0.7 Figure 6. Dependence of the maximal run-up R on the amplitude α of the incident solitary wave. Experimental data: (1) — [144], (2) — [111], (3) — [42], (4) — [110]. Numerical data: (5) — [67], (6) — [26], (7) — [39]. The solid line (8) — our numerical results, the dashed line (9) — the analytical prediction (4.3). 4.2.1 Wave action on the wall The nonlinear dispersive SGN model can be used to estimate also the wave force exerted on the vertical wall. Moreover, we shall show below that this model is able to capture the non-monotonic behaviour of the force when the incident wave amplitude is increased. This effect was first observed experimentally [144] and then numerically [148]. For the 2D case with flat bottom the fluid pressure p(x, y, t) can be expressed: h H2
(y + d)2 i p(x, y, t) R1 , = g H − (y + d) − − ρ 2 2 def − d 6 y 6 η(x, t) , (4.4) with R1 := uxt + u uxx − u2x . The horizontal wave loading exerted on the vertical wall located at x = ℓ is given by the following integral: ˆ η(ℓ,t) F0 (t) H3 g H2 = − R̄1 , p(ℓ, y, t) dy = ρ 2 3 −d 40 / 66 η /d F/ρgd 2 G. Khakimzyanov, D. Dutykh, et al. 1.5 1.5 1 2 3 4 1.0 1.0 0.5 0.5 0.0 0.0 25 30 35 40 45 1 2 3 4 t(g/d) 1/2 25 (a) 30 35 40 45 t(g/d) 1/2 (b) Figure 7. Solitary wave/vertical wall interaction: (a) time series of wave run-up on the wall; (b) dynamic wave loading on the wall. Different lines correspond to different incident solitary wave amplitudes: (1) — α = 0.1, (2) — α = 0.3, (3) — α = 0.5, (4) — α = 0.7. where due to boundary conditions R̄1 = uxt − u2x . After removing the hydrostatic force, we obtain the dynamic wave loading computed in our simulations: h H2 F(t) H3 d2 i − = g − R̄1 . ρ 2 2 3 The expression for corresponding tilting moment can be found in [52, Remark 3]. Figure 7 shows the wave elevation (a) and the dynamic wave loading (b) on the vertical wall. From Figure 7(b) it can be seen that the force has one maximum for small amplitude solitary waves. However, when we gradually increase the amplitude (i.e. α ? 0.4), the second (local) maximum appears. For such large solitary waves a slight run-down phenomenon can be noticed in Figure 7(a). We reiterate that this behaviour is qualitatively and quantitatively correct comparing to the full Euler equations [25, 39]. However, the complexity of the nonlinear dispersive SGN model and, consequently, the numerical algorithm to solve it, is much lower. 4.3. Solitary wave/bottom step interaction Water waves undergo continuous changes while propagating over general uneven bottoms. Namely, the wave length and wave amplitude are modified while propagating over bottom irregularities. Such transformations have been studied in the literature [44, 98]. In the present Section we focus on the process of a Solitary Wave (SW) transformation over a Dispersive shallow water wave modelling. Part II 41 / 66 bottom step. In the early work by Madsen & Mei (1969) [106] it was shown using long wave approximation that a solitary wave can be disintegrated into a finite number of SWs with decreasing amplitudes while passing over an underwater step. This conclusion was supported in [106] by laboratory data as well. This test case was used later in many works, see e.g. [29, 53, 99]. We illustrate the behaviour of the adaptive numerical algorithm as well as the SGN model on the solitary wave/bottom interaction problem. The bottom bathymetry is given by the following discontinuous function: ( −h0 , 0 6 x 6 xs , y = −h(x) = −hs , xs < x 6 ℓ , where ℓ is the numerical wave tank length, h0 (respectively hs ) are the still water depths on the left (right) of the step located at x = xs . We assume also that 0 < hs < h0 . The initial condition is a solitary wave located at x = x0 and propagating rightwards. For the experiment cleanliness we assume that initially the solitary wave does not ‘feel’ the step. In other words it is located sufficiently far from the abrupt change in bathymetry. In our experiment we choose x0 so that η(xs ) > 0.01 α, where α is the SW amplitude. The main parameters in this problem are the incident wave amplitude α and the bottom step jump ∆bs = h0 − hs . Various theoretical and experimental studies show that a solitary wave undergoes a splitting into a reflected wave and a finite number of solitary waves after passing over an underwater step. See [98] for a recent review on this topic. Amplitudes and the number of solitary waves over the step were determined in [88] in the framework of the shallow water theory. These expressions were reported later in [124] and this result was improved recently in [115]. However, in the vicinity of the step, one may expect important vertical accelerations of fluid particles, which are simplified (or even neglected) in shallow water type theories. Nevertheless, in [115] a good agreement of this theory with numerical and experimental data was reported. There is also another difficulty inherent to the bottom step modelling. In various derivations of shallow water models there is an implicit assumption that the bathymetry gradient ∇h is bounded (or even small | ∇h | ≪ 1, e.g. in the Boussinesq-type equations [19]). On the other hand, numerical tests and comparisons with the full (Euler and even Navier–Stokes) equations for finite values of | ∇h | ∼ O(1) show that resulting approximate models have a larger applicability domain than it was supposed at the outset [19]. In the case of a bottom step, the bathymetry function is even discontinuous which is an extreme case we study in this Section. There are two main approaches to cope with this problem. One consists in running the approximate model directly on discontinuous bathymetry, and the resulting eventual numerical instabilities are damped out by ad-hoc dissipative terms (see e.g. references in [115]). The magnitude of these terms allows to increase the scheme dissipation, and overall computation appears to be stable. The difficulty of this approach consists in the fine tuning of dissipation, since • Insufficient dissipation will make the computation unstable, G. Khakimzyanov, D. Dutykh, et al. 42 / 66 • Excessive dissipation will yield unphysical damping of the solution. An alternative approach consists in replacing the discontinuous bathymetry by a smoothed  ℓs ℓs  version over certain length xs − , where ℓs is the smoothing length on , xs + 2 2 which the jump from h0 to hs is replaced by a smooth variation. For instance, in all numerical computations reported in [124] the smoothing length was chosen to be ℓs = 60 cm independently of the water depths before h0 and after hs the step. In another work [66] the smoothing length was arbitrarily set to ℓs = 20 cm independently of other parameters. Unfortunately, in a recent work [146] the smoothing procedure was not described at all. Of course, this method is not perfect since the bathymetry is slightly modified. However, one can expect that sufficiently long waves will not ‘notice’ this modification. This assumption was confirmed by the numerical simulations reported in [35, 66, 124]. In the present work we also employ the bottom smoothing procedure. However, the smoothing length ℓs is chosen in order to have a well-posed problem for the elliptic operator (2.5). For simplicity, we use the sufficient condition (2.12) (obtained under restriction (2.11)), which is not necessarily optimal, but it allows us to invert stably the nonlinear elliptic operator (2.18). Namely, the smoothed step has the following analytical expression:  ℓs  , −h0 , 0 6 x 6 xs −    2  
ℓs ℓs ∆bs y = −h(x) = −h0 + (4.5) · 1 + sin ζ , xs − 6 x 6 xs + ,  2 2 2     ℓs −h , xs + 6 x 6 ℓ, s 2 def π(x − xs ) where ζ := . For this bottom profile, the inequalities (2.11), (2.12) take the ℓs form:  π π π ∆bs , cos ζ < 1 , ∀ ζ ∈ − , 2 ℓs 2 2 2 π 2 ∆bs sin ζ > − . 2 h + h ∆bs 2 ℓs 0 s − sin ζ 2 2 These inequalities have corresponding solutions: π p π ∆bs ℓs > , ℓs > h0 ∆bs . 2 2 The last inequalities are verified simultaneously if the second inequality is true. If we assume that the bottom step height ∆bs is equal to the half of the water depth before it, then we obtain the following condition: π ℓs > √ h0 ≈ 1.11 h0 . 2 2 We underline that the last condition is only sufficient and stable numerical computations can most probably be performed even for shorter smoothing lengths ℓs . For instance, we tested the value ℓs = h0 and everything went smoothly. 43 / 66 Dispersive shallow water wave modelling. Part II Parameter Value Wave tank length, ℓ 35 m 3.65 cm Solitary wave amplitude, α 11 m Solitary wave initial position, x0 20 cm Water depth before the step, h0 Water depth after the step, hs 10 cm h0 Water depth used in scaling, d 10 cm Bottom step jump, ∆bs Bottom step location, xs 14 m 350 Number of grid points, N 17.6 s Simulation time, T Table 2. Values of various numerical parameters used in the solitary wave/bottom step interaction test case. In [124] the results of 80 experiments are reported for various values of α and h0 (for fixed values of the bottom jump ∆bs = 10 cm). In our work we repeated all experiments from [124] using the SGN equations solved numerically with the adaptive predictor–corrector algorithm described above. In general, we obtained a very good agreement with experimental data from [124] in terms of the following control parameters: • number of solitary waves moving over the step, • amplitudes of solitary waves over the step, • amplitude of the (main) reflected wave. We notice that the amplitude of the largest solitary wave over the step corresponds perfectly to the measurements. However, the variation in the amplitude of subsequent solitary waves over the step could reach in certain cases 20%. Remark 4. The conduction of laboratory experiments on the solitary wave/bottom step interaction encounters a certain number of technical difficulties [27, 124] that we would like to mention. First of all, the wave maker generates a solitary wave with some dispersive components. Moreover, one has to take the step sufficiently long so that the transmitted wave has enough time to develop into a finite number of visible well-separated solitary waves. Finally, the reflections of the opposite wave flume’s wall are to be avoided as well in order not to pollute the measurements. Consequently, the successful conduction of experiments and accurate measurement of wave characteristics requires a certain level of technique. We would like to mention the exemplary experimental work [78] on the head-on collision of solitary waves. Below we focus on one particular case of α = 3.65 cm. All other parameters are given in Table 2. It corresponds to the experiment N◦ 24 from [124]. The free surface dynamics is depicted in Figure 8(a) and the trajectories of every second grid node are 44 / 66 G. Khakimzyanov, D. Dutykh, et al. (a) (b) Figure 8. Interaction of a solitary wave with an underwater step: (a) space-time plot of the free surface elevation y = η(x, t) in the dimensional time interval [0 s, 17.6 s]; (b) trajectories of every second grid node. Numerical parameters are provided in Table 2. shown in Figure 8(b). For the mesh adaptation we use the monitor function (4.2) with ϑ1 = ϑ2 = 10 . In particular, one can see that three solitary waves are generated over the step. This fact agrees well with the theoretical predictions [88, 115]. Moreover, one can see that the distribution of grid points follows perfectly all generated waves (over the step and the reflected wave). Figure 9(a) shows the free surface dynamics in the vicinity of the bottom step. In particular, one can see that the wave becomes notoriously steep by the time instance t = 3 s and during later times it splits into one reflected and three transmitted waves. The free surface profile at the final simulation time y = η(x, T ) is depicted in Figure 9(b). On the same panel the experimental measurements are shown with empty circles ◦ , which show a very good agreement with our numerical simulations. In our numerical experiments we go even further since a nonlinear dispersive wave model (such as the SGN equations employed in this study) can provide also information about the internal structure of the flow (i.e. beneath the free surface). For instance, the nonhydrostatic component of the pressure field can be easily reconstructed∗: h η + y + 2h i pd (x, y, t) = −(η − y) · R1 + R2 , −h(x) 6 y 6 η(x, t) . (4.6) ρ 2 where the quantities R1,2 are defined in (2.3) as (see also the complete derivation in [92]): R1 = uxt + u uxx − ux2 , R2 = ut hx + u [ u hx ]x . ∗Please, notice that formula (4.4) is not applicable here, since the bottom is not flat anymore. 45 / 66 Dispersive shallow water wave modelling. Part II η /d η /d 0.20 0.25 1 2 3 4 5 6 0.15 0.10 0.20 1 2 0.15 0.05 0.10 0.00 0.05 0.00 -0.05 60 70 (a) 80 x/d 140 150 160 170 x/d (b) Figure 9. Free surface profiles y = η(x, t) during the interaction process of a solitary wave with an underwater step: (a) initial condition (1), t = 1.5 s (2), t = 2.0 s (3), t = 2.5 s (4), t = 3.0 s (5), smoothed bottom profile given by formula (4.5) (6); (b) free surface profile y = η(x, T ) (1) at the final simulation time t = T . The experimental points (2) are taken from [124], experiment N◦ 24. Numerical parameters are provided in Table 2. We do not consider the hydrostatic pressure component since its variation is linear with water depth y: ph = ρ g (η − y) . Even if the dispersive pressure component pd might be negligible comparing to the hydrostatic one ph , its presence is crucial to balance the effects of nonlinearity, which results in the existence of solitary waves, as one of the most widely known effects in dispersive wave propagation [48]. The dynamic pressure field and several other physical quantities under a solitary wave were computed and represented graphically in the framework of the full Euler equations in [51]. A good qualitative agreement with our results can be reported. The balance of dispersive and nonlinear effects results also in the symmetry of the non-hydrostatic pressure distribution with respect to the wave crest. It can be seen in Figure 10(a,d ) before and after the interaction process. On the other hand, during the interaction process the symmetry is momentaneously broken (see Figure 10(b,c)). However, with the time going on, the system relaxes again to a symmetric∗ pressure distribution shown in Figure 10(d ). Knowledge of the solution
to the SGN equations allows to reconstruct also the velocity † field ũ(x, y, t), ṽ(x, y, t) in the fluid bulk. Under the additional assumption that the ∗The †This symmetry here is understood with respect to the vertical axis passing by the wave crest. information can be used later to compute fluid particle trajectories [69], for example. 46 / 66 G. Khakimzyanov, D. Dutykh, et al. pd /ρ gd y/d 0.005 0.002 -0.000 -0.003 -0.005 -0.008 -0.010 -0.013 -0.015 -0.018 0.2 0.0 0.0 -0.2 -0.4 -0.4 56 0.005 0.003 0.001 -0.001 -0.003 -0.006 -0.008 -0.010 -0.012 -0.014 0.2 -0.2 54 pd /ρgd y/d x/d 58 68 70 (a) (b) pd /ρ gd y/d 0.007 0.004 0.001 -0.002 -0.005 -0.008 -0.011 -0.014 -0.017 -0.021 0.2 0.0 0.0 -0.4 -0.4 (c) 80 x/d 0.011 0.004 -0.004 -0.011 -0.018 -0.025 -0.032 -0.040 -0.047 -0.054 0.2 -0.2 78 pd /ρgd y/d -0.2 76 x/d 72 164 166 168 x/d (d) Figure 10. Non-hydrostatic pressure distribution during a solitary wave/underwater step interaction process at different instances of time: (a) t = 0.1 s, (b) t = 2.0 s, (c) t = 3.0 s, (d) t = 17.5 s. Numerical parameters are provided in Table 2. flow is potential,
one can derive the following asymptotic (neglecting the terms of the order 4 2
O(µ4 ) ≡ O λd4 in the horizontal velocity ũ(x, y, t) and of the order O(µ2 ) ≡ O λd2 for the vertical one ṽ(x, y, t)) representation formula [64] (see also the derivation in [92] for 47 / 66 Dispersive shallow water wave modelling. Part II y/d y/d 0.2 0.2 0.0 0.0 -0.2 -0.2 -0.4 -0.4 -0.6 65 70 x/d 75 (a) -0.6 60 70 x/d 80 (b) Figure 11. Reconstructed velocity field in the fluid during the solitary wave interaction process with an underwater step: (a) t = 2.0 s , (b) t = 3.0 s . Solid blue lines show a few streamlines. Numerical parameters are provided in Table 2. the 3D case with moving bottom): ũ(x, y, t) = u +   H2
(y + h)2  uxx , − y − h · [ u hx ]x + ux hx + − 2 6 2 (4.7) H ṽ(x, y, t) = −u hx − (y + h) ux . (4.8) The formulas above allow to compute the
velocity vector field in the fluid domain at any time (when the solution H(x, t), u(x, t) is available) and in any point (x, y) above the bottom y = −h(x) and beneath the free surface y = η(x, t) . Figure 11 shows a numerical application of this reconstruction technique at two different moments of time t = 2 and 3 s during the interaction process with the bathymetry change. In particular, in Figure 11(a) one can see that important vertical particle velocities emerge during the interaction with the bottom step. In subsequent time moments one can see the division of the flow in two structures (see Figure 11(b)): the left one corresponds to the reflected wave, while the right structure corresponds to the transmitted wave motion. The reconstructed velocity fields via the SGN model compare fairly well with the 2D Navier–Stokes predictions [115]. However, the computational complexity of our approach is significantly lower than the simulation of the full Navier–Stokes equations. This is probably the main advantage of the proposed modelling methodology. G. Khakimzyanov, D. Dutykh, et al. 48 / 66 4.4. Wave generation by an underwater landslide As the last illustration of the proposed above numerical scheme, we model wave generation by the motion of an underwater landslide over uneven bottom. This test-case is very challenging since it involves rapid bottom motion (at least of its part). We recall that all previous tests were performed on a static bottom (i.e. ht ≡ 0 ). The numerical simulation of underwater landslides is an important application where the inclusion of non-hydrostatic effects is absolutely crucial [105]. Moreover, the accurate prediction of generated waves allows to assess more accurately the natural hazard induced by unstable sliding masses (rockfalls, debris, ground movements) [140]. Usually, the precise location of unstable underwater masses is unknown and the numerical simulation is a preferred tool to study these processes. The landslide can be modelled as a solid undeformable body moving down the slope [34, 60, 74, 142]. Another possibility consists in representing the landslide as another fluid layer of higher density (possibly also viscosity) located near the bottom [23, 68]. In some works the landslide motion was not simulated (e.g. [86]) and the initial wave field generated by the landslide motion was determined using empirical formulas [73]. Then, this initial condition was propagated using an appropriate water wave model [86]. However, strictly speaking the employed empirical models are valid only for an absolutely rigid landslide sliding down a constant slope. Subsequent numerical simulations showed that the bottom shape influences quite substantially the generated wave field [13]. Consequently, for realistic modelling of real world cases one needs to take into account the actual bathymetry [103] and even possible deformations of the landslide during its motion [102]. In a recent experimental work [102] the deformability of the landslide was achieved by composing it with four solid parts interconnected by springs. The idea to represent a landslide as a finite number of blocks was used in numerical [119] and theoretical [137] investigations. In the present study we use the quasideformable∗ landslide model [12, 56, 59]. In this model the landslide deforms according to encountered bathymetry changes, however, at every time instance, all components of the velocity vector are the same in all particles which constitute the landslide (as in a solid rigid body). We shall use two long wave models: • The SGN equations (fully nonlinear non-hydrostatic weakly dispersive model) • NSWE equations† (standard hydrostatic dispersionless model) The advantage of the SGN equations over other frequently used long wave models [86, 105, 141] are: • The Galilean invariance • The energy balance equation (consistent with the full Euler [65]) NSWE were employed in [5] to model the real world 16th October 1979 Nice event. It looks like the consensus on the importance of dispersive effects in landslide modeling is far from ∗This †The [94]. model can be visualized if you imagine a landslide composed of infinitely many solid blocks. numerical algorithm to solve NSW equations on a moving grid was presented and validated in Dispersive shallow water wave modelling. Part II 49 / 66 being achieved. For example, in [86] the authors affirm that the inclusion of dispersion gives results very similar to NSWE. In other works [71, 104, 133] the authors state that dispersive effects significantly influence the resulting wave field, especially during long time propagation. Consequently, in the present study we use both the SGN and NSWE equations to shed some light on the rôle of dispersive effects. Consider a 1D fluid domain bounded from below by the solid (static) impermeable bottom given by the following function:   h+ + h− h+ − h− h0 (x) = + tanh ̥(x − ξ̥ ) , (4.9) 2 2 where h+ and h− are water depths at ± ∞ correspondingly (the domain we take is finite, of course). We assume for definiteness that h+ < h− < 0 . We have also by definition hh − h i 1 2 tan θ0 def def 0 + > 0, ξ̥ := > 0, ln ̥ := h− − h+ 2̥ h− − h+ where h0 ≡ h0 (0) is water depth in x = 0 and θ0 is the maximal slope angle, which is reached at the inflection point ξ̥ . It can be easily checked that h+ + h− < h0 < h− . 2 Equation (4.9) gives us the static part of the bottom shape. The following equation prescribes the shape of the bathymetry including the unsteady component: y = −h(x, t) = h0 (x) + ζ(x, t) , where function ζ(x, t) prescribes the landslide shape. In the present study we assume that the landslide initial shape is given by the following analytical formula:   h 2 π x − x (0)
i   ~ ν c   1 + cos , | x − xc (0) | 6 , 2 ν 2 ζ(x, 0) =  ν  0 , , | x − xc (0) | > 2 where xc (0), ~ and ν are initial landslide position, height and width (along the axis Ox) correspondingly. Initially we put the landslide at the unique∗ point where the water depth is equal to h0 = 100 m, i.e. hh − h i 1 0 + ≈ 8 323.5 m . ln xc (0) = ξ̥ − 2̥ h− − h+ For t > 0 the landslide position xc (t) and its velocity v(t) are determined by solving a second order ordinary differential equation which describes the balance of all the forces acting on the sliding mass [12]. This model is well described in the literature [56, 59] and we do not reproduce the details here. ∗This point is unique since the static bathymetry h0 (x) is a monotonically increasing function of its argument x . 50 / 66 G. Khakimzyanov, D. Dutykh, et al. Parameter Value Fluid domain length, ℓ 80 000 m −5.1 m Water depth, h0 (0) −500 m Rightmost water depth, h+ −5 m Leftmost water depth, h− Maximal bottom slope, θ0 6◦ 20 m Landslide height, ~ 5000 m Landslide length, ν Initial landslide position, xc (0) 8 323.5 m 1.0 Added mass coefficient, Cw 1.0 Hydrodynamic resistance coefficient, Cd 1.5 Landslide density, ρsl /ρw ∗ Friction angle, θ 1◦ 1000 s Final simulation time, T 400 Number of grid points, N Monitor function parameter, ϑ0 200 Table 3. Numerical and physical parameters used in landslide simulation. In Figure 12(a) we show the dynamics of the moving bottom from the initial condition at t = 0 to the final simulation time t = T . All parameters are given in Table 3. It can be clearly seen that landslide’s motion significantly depends on the underlying static bottom shape. In Figure 12(b) we show landslide’s barycenter trajectory x = xc (t) (line 1), its velocity v = v(t) (line 2) and finally the static bottom profile y = h0 (x) (line 3). From the landslide speed plot in Figure 12(b) (line 2), one can see that the mass is accelerating during the first 284.2 s and slows down during 613.4 s. The distances traveled by the landslide during these periods have approximatively the same ratio ≈ 2 . It is also interesting to notice that the landslide stops abruptly its motion with a negative (i.e. nonzero) acceleration. In order to simulate water waves generated by the landslide, we take the fluid domain I = [0, ℓ]. For simplicity, we prescribe wall boundary conditions∗ at x = 0 and x = ℓ. Undisturbed water depth at both ends is h0 and ≈ h+ respectively. The computational domain length ℓ is chosen to be sufficiently large to avoid any kind of reflections from the right boundary. Initially the fluid is at rest with undisturbed free surface, i.e. η(x, 0) ≡ 0 , ∗It u(x, 0) ≡ 0 . would be better to prescribe transparent boundary conditions here, but this question is totally open for the SGN equations. 51 / 66 Dispersive shallow water wave modelling. Part II 0 900 10 v, m/s 20 30 t, s 3 0 y, m 1 600 -200 300 2 -400 0 (a) 0 10000 20000 x, m (b) Figure 12. Generation of surface waves by an underwater landslide motion: (a) dynamics of the moving bottom; (b) graphics of functions (1) x = xc (t) , (2) v = v(t) , (3) y = h0 (x) . Two outer red circles denote landslide initial t = 0 and terminal t = 897.6 s positions. Middle red circle denotes landslide position at the moment of time t = 284.2 s where landslide’s speed is maximal vmax ≈ 26.3 m/s . The black square shows the inflection point ξ̥ position. The maximal speed is achieved well below the inflection point ξ̥ . Numerical parameters are given in Table 3. Segment I is discretized using N = 400 points. In order to redistribute optimally mesh nodes, we employ the monitor function defined in equation (4.1), which refines the grid where the waves are large (regardless whether they are of elevation or depression type). In Figure 13 we show the surface y = η(x, t) in space-time, which shows the main features of the generated wave field. The left panel (a) is the dispersive SGN prediction, while (b) is the computation with NSWE that we include into this study for the sake of comparison. For instance, one can see that the dispersive wave system is much more complex even if NSWE seem to reproduce the principal wave components. The dispersive components follow the main wave travelling rightwards. There is also at least one depression wave moving towards the shore. The motion of grid points is shown in Figure 14. The initial grid was chosen to be uniform, since the free surface was initially flat. However, during the wave generation process the grid adapts to the solution. The numerical method redistributes the nodes according to the chosen monitor function ̟0 [ η ] (x, t) , i.e. where the waves are large (regardless whether they are of elevation or depression type). We would like to underline the fact that in order to achieve a similar accuracy on a uniform grid, one would need about 4 N points. In Figure 15 we show two snapshots of the free surface elevation at two moments of time (a) and wave gauge records collected at two different spatial locations (b). In particular, 52 / 66 G. Khakimzyanov, D. Dutykh, et al. (a) (b) Figure 13. Generation of surface waves y = η(x, t) by an underwater landslide motion: (a) the SGN model (dispersive); (b) NSWE equations (dispersionless). Numerical parameters are given in Table 3. we observe that there is a better agreement between NSWE and the SGN model in shallow regions (i.e. towards x = 0), while a certain divergence between two models becomes more apparent in deeper regions (towards the right end x = ℓ). In the previous Section 4.3 we showed the internal flow structure during nonlinear transformations of a solitary wave over a static step. In this Section we show that SGN equations can be used to reconstruct and to study the physical fields in situations where the bottom moves abruptly. In order to reconstruct the non-hydrostatic field between moving bottom and the free surface, one can use formula (4.6), but the quantity R2 has some extra terms due to the bottom motion: R2 = ut hx + u [ u hx ]x + htt + 2 u hxt . In Figure 16 we show the non-hydrostatic pressure field at two different moments of time. More precisely, we show a zoom on the area of interest around the landslide only. In panel (a) t = t1 = 150 s and the landslide barycenter is located at xc (t1 ) = 9456 m . Landslide moves downhill with the speed v(t1 ) = 15.72 m/s and it continues to accelerate. In particular, one can see that there is a zone of positive pressure in front of the landslide and a zone of negative pressure just behind. This fact has implications on the fluid particle trajectories around the landslide. In right panel (b) we show the moment of time t = t2 = 400 s . At this moment xc (t2 ) = 15 264 m and v(t2 ) = 21.4 m/s . The non-hydrostatic pressure distribution qualitatively changed. Zones of positive and negative pressure switched their respective positions. Moreover, in Figure 15 we showed that dispersive effects start to be noticeable at the free surface only after t > 400 s and by t = 800 s they are flagrant. In Figure 17 we show the velocity fields in the fluid bulk at corresponding moments of 53 / 66 t, s Dispersive shallow water wave modelling. Part II 800 600 400 200 0 0 20000 40000 60000 x, m Figure 14. Trajectories of every second grid node during the underwater landslide simulation in the framework of the SGN equations. Numerical parameters are given in Table 3. time t1 and t2 . We notice some similarities between the fluid flow around a landslide with an air flow around an airfoil. To our knowledge the internal hydrodynamics of landslide generated waves on a general non-constant sloping bottom and in the framework of SGN equations has not been shown before. We remind that in the presence of moving bottom one should use the following reconstruction formulas for the velocity field (which are slightly different from (4.7), (4.8)):   H2
(y + h)2  uxx , − y − h · [ ht + u hx ]x + ux hx + − 2 6 2 ṽ(x, y, t) = −ht − u hx − (y + h) ux . ũ(x, y, t) = u + H 54 / 66 G. Khakimzyanov, D. Dutykh, et al. y, m y, m 1 2 3 4 2 3 0 0 -3 -2 1 2 3 4 -6 -4 -9 0 20000 40000 60000 0 x, m 200 (a) 400 600 800 t, s (b) Figure 15. Generation of surface waves by an underwater landslide: (a) free surface elevation profiles y = η(x, t1, 2 ) at t1 = 300 s (1,3) and t2 = 800 s (2,4); (b) free surface elevation y = η(x1, 2 , t) as a function of time in two spatial locations x1 = 20000 m (1,3) and x2 = 40000 m (2,4). The SGN predictions are represented with solid lines (1,2) and NSWE with dashed lines (3,4). Numerical parameters are given in Table 3. y, m pd , kPa y, m pd , kPa 0.0 41.780 35.960 30.140 24.320 18.500 12.681 6.861 1.041 -4.779 -10.599 0.0 41.780 35.960 30.140 24.320 18.500 12.681 6.861 1.041 -4.779 -10.599 -100.0 -200.0 -100.0 -200.0 -300.0 -300.0 -400.0 -400.0 -500.0 -500.0 5000 7500 10000 (a) 12500 x, m 12500 15000 17500 x, m 20000 (b) Figure 16. Generation of surface waves by an underwater landslide. Isolines of the non-hydrostatic pressure at two moments of time: t = 150 s (a); t = 400 s (b). Numerical parameters are given in Table 3. 55 / 66 Dispersive shallow water wave modelling. Part II y, m y, m 0.0 0.0 -100.0 -100.0 -200.0 -200.0 -300.0 -300.0 -400.0 -400.0 -500.0 -500.0 5000 7500 10000 (a) 12500 x, m 12500 15000 17500 x, m 20000 (b) Figure 17. Generation of surface waves by an underwater landslide. The reconstructed velocity field at two moments of time: t = 150 s (a); t = 400 s (b). Numerical parameters are given in Table 3. These formulas naturally become (4.7), (4.8) if the bottom is static, i.e. ht ≡ 0 . 5. Discussion Above we presented a detailed description of the numerical algorithm and a number of numerical tests which illustrate its performance. The main conclusions and perspectives of this study are outlined below. 5.1. Conclusions In the second part of our series of papers we focused on the development of numerical algorithms for shallow water propagation over globally flat spaces (i.e. we allow some variations of the bathymetry in the limits discussed in Section 2.1). The main distinction of our work is that the proposed algorithm allows for local mesh adaptivity by moving the grid points where they are needed. The performance of our method was illustrated on several test cases ranging from purely academic ones (e.g. propagation of a solitary waves, which allowed us to estimate the overall accuracy of the scheme) to more realistic applications with landslide-generated waves [12]. The mathematical model chosen in this study allows us to have a look into the distribution of various physical fields in the fluid bulk. In particular, in some interesting cases we reconstructed the velocity field and the hydrostatic pressure distribution beneath the free surface. G. Khakimzyanov, D. Dutykh, et al. 56 / 66 We studied the linear stability of the proposed finite volume discretization. It was shown that the resulting scheme possesses an interesting and possible counter-intuitive property: smaller we take the spatial discretization step ∆x, less restrictive is becoming the stability CFL-type condition on the time step τ . This result was obtained using the classical von Neumann analysis [28]. However, we show (and we compute it) that there exists the upper limit of allowed time steps. Numerical schemes with such properties seem to be new. We considered also in great detail the question of wall boundary conditions∗ for the SGN system. It seems that this issue was not properly addressed before. The wall boundary condition for the elliptic part of equations follows naturally from the form of the momentum equation we chose in this study. Finally, in numerical experiments we showed how depth-integrated SGN equations can be used to study nonlinear transformations of water waves over some bathymetric features (such as an underwater step or a sliding mass). Moreover, we illustrated clearly that SGN equations (and several other approximate dispersive wave models) can be successfully used to reconstruct the flow field under the wave. The accuracy of this reconstruction will be studied in future works by direct comparisons with the full Euler equations where these quantities are resolved. 5.2. Perspectives The main focus of our study was set on the adaptive spatial discretization. The first natural continuation of our study is the generalization to 3D physical problems (i.e. involving two horizontal dimensions). The main difficulty is to generalize the mesh motion algorithm to this case, even if some ideas have been proposed in the literature [6]. In the present computations the time step was chosen to ensure the linear CFL condition. In other words, it was chosen in order to satisfy the numerical solution stability. In future works we would like to incorporate an adaptive time stepping procedure along the lines of e.g. [130] aimed to meet the prescribed error tolerance. Of course, the extension of the numerical method presented in this study to three-dimensional flows (i.e. two horizontal dimensions) represents the main important extension of our work. Further improvement of the numerical algorithm can be expected if we include also some bathymetric features (such as ∇h) into the monitor function ̟[η, h](x, t) . Physically this improvement is fully justified since water waves undergo constant transformations over bottom irregularities (as illustrated in Sections 4.3 & 4.4). A priori, everything is ready to perform these further numerical experiments. Ideally, we would like to generalize the algorithm presented in this study for the Serre– Green–Naghdi (SGN) equations to the base model in its most general form (1.3), (1.4). In this way we would be able to incorporate several fully nonlinear shallow water models (discussed in Part I [91]) in the same numerical framework. It would allow the great ∗The x=ℓ wall boundary condition for the velocity component u(x, t) is straightforward, i.e. u(x, t)|x = 0 = 0. However, there was an open question of how to prescribe the boundary conditions for the elliptic part of the equations. Dispersive shallow water wave modelling. Part II 57 / 66 flexibility in applications to choose and to assess the performance of various approximate models. Moreover, in the present study we raised the question of boundary conditions for SGN equations. However, non-reflecting (or transparent) boundary conditions would allow to take much smaller domains in many applications. Unfortunately, this question is totally open to our knowledge for the SGN equations (however, it is well understood for NSWE). In future works we plan to fill in this gap as well. Finally, the SGN equations possess a number of variational structures. The Hamiltonian formulation can be found e.g. in [87]. Various Lagrangians can be found in [38, 63, 95, 112]. Recently, a multi-symplectic formulation for SGN equations has been proposed [32]. All these available variational structures raise an important question: after the discretization can we preserve them at the discrete level as well? It opens beautiful perspectives for the development of structure-preserving numerical methods as it was done for the classical Korteweg–de Vries [50] and nonlinear Schrödinger [31] equations. In the following parts of this series of papers we shall discuss the derivation of the SGN equations on a sphere [91] and their numerical simulation using the finite volume method [93]. Acknowledgments This research was supported by RSCF project No 14–17–00219. The authors would like to thank Prof. Emmanuel Audusse (Université Paris 13, France) who brought our attention to the problem of boundary conditions for the SGN equations. A. Derivation of the non-hydrostatic pressure equation In this Appendix we give some hints for the derivation of the non-hydrostatic pressure equation (2.5) and relation (2.9). Let us start with the latter. For this purpose we rewrite equation (2.8) in a more compact form using the total derivative operator: Du = −g ∇η + ∇℘ − ̺∇h , (A.1) H and ̺ (see equations (2.3) and (2.4) corre- By definition of non-hydrostatic quantities ℘ spondingly) we obtain: 3℘ H ̺ = + R2 . 2H 4 We have to substitute into the last relation the expression for R2 :
R2 = (Du) · ∇h + u · (u · ∇)∇h + htt + 2 u · ∇ht , along with the expression (A.1) for the horizontal acceleration Du of fluid particles. After simple algebraic computations one obtains (2.9). 58 / 66 G. Khakimzyanov, D. Dutykh, et al. The derivation of equation (2.5) is somehow similar. First, from definitions (2.3), (2.4) we obtain another relation between non-hydrostatic pressures: H3 H R1 + ̺, 12 2 with R1 rewritten in the following form: ℘ (A.2) = R1 = ∇ · (Du) − 2 (∇ · u)2 + 2 u 1 x1 u 1 x2 . u 2 x1 u 2 x2 Substituting into equation (A.2) the just shown relation (2.9) with the last expression for R1 , yields the required equation (2.5). B. Acronyms In the text above the reader could encounter the following acronyms: SW: Solitary Wave AMR: Adaptive Mesh Refinement BBM: Benjamin–Bona–Mahony BVP: Boundary Value Problem CFL: Courant–Friedrichs–Lewy IVP: Initial Value Problem MOL: Method Of Lines ODE: Ordinary Differential Equation PDE: Partial Differential Equation SGN: Serre–Green–Naghdi TVD: Total Variation Diminishing NSWE: Nonlinear Shallow Water Equations References [1] G. B. Alalykin, S. K. Godunov, L. L. Kireyeva, and L. A. Pliner. Solution of OneDimensional Problems in Gas Dynamics on Moving Grids. Nauka, Moscow, 1970. 5 Dispersive shallow water wave modelling. Part II 59 / 66 [2] J. S. Antunes Do Carmo, F. J. Seabra Santos, and E. Barthélemy. Surface waves propagation in shallow water: A finite element model. Int. J. Num. Meth. Fluids, 16(6):447–459, mar 1993. 6 [3] C. Arvanitis and A. I. Delis. Behavior of Finite Volume Schemes for Hyperbolic Conservation Laws on Adaptive Redistributed Spatial Grids. SIAM J. Sci. Comput., 28(5):1927–1956, jan 2006. 16 [4] C. Arvanitis, T. Katsaounis, and C. Makridakis. Adaptive Finite Element Relaxation Schemes for Hyperbolic Conservation Laws. ESAIM: Mathematical Modelling and Numerical Analysis, 35(1):17–33, 2010. 16 [5] S. Assier-Rzadkieaicz, P. Heinrich, P. C. Sabatier, B. Savoye, and J. F. Bourillet. Numerical Modelling of a Landslide-generated Tsunami: The 1979 Nice Event. Pure Appl. Geophys., 157(10):1707–1727, oct 2000. 48 [6] B. N. Azarenok, S. A. Ivanenko, and T. Tang. Adaptive Mesh Redistibution Method Based on Godunov’s Scheme. Commun. Math. Sci., 1(1):152–179, 2003. 56 [7] N. S. Bakhvalov. The optimization of methods of solving boundary value problems with a boundary layer. USSR Computational Mathematics and Mathematical Physics, 9(4):139– 166, jan 1969. 5, 16 [8] V. B. Barakhnin and G. S. Khakimzyanov. On the algorithm for one nonlinear dispersive shallow-water model. Russ. J. Numer. Anal. Math. Modelling, 12(4):293–317, 1997. 6 [9] V. B. Barakhnin and G. S. Khakimzyanov. The splitting technique as applied to the solution of the nonlinear dispersive shallow-water equations. Doklady Mathematics, 59(1):70–72, 1999. 6, 11 [10] T. J. Barth and M. Ohlberger. Finite Volume Methods: Foundation and Analysis. In E. Stein, R. de Borst, and T. J. R. Hughes, editors, Encyclopedia of Computational Mechanics. John Wiley & Sons, Ltd, Chichester, UK, nov 2004. 22 [11] E. Barthélémy. Nonlinear shallow water theories for coastal waves. Surveys in Geophysics, 25:315–337, 2004. 6 [12] S. A. Beisel, L. B. Chubarov, D. Dutykh, G. S. Khakimzyanov, and N. Y. Shokina. Simulation of surface waves generated by an underwater landslide in a bounded reservoir. Russ. J. Numer. Anal. Math. Modelling, 27(6):539–558, 2012. 48, 49, 55 [13] S. A. Beisel, L. B. Chubarov, and G. S. Khakimzyanov. Simulation of surface waves generated by an underwater landslide moving over an uneven slope. Russ. J. Numer. Anal. Math. Modelling, 26(1):17–38, 2011. 48 [14] T. B. Benjamin, J. L. Bona, and J. J. Mahony. Model equations for long waves in nonlinear dispersive systems. Philos. Trans. Royal Soc. London Ser. A, 272:47–78, 1972. 6 [15] M. J. Berger, D. L. George, R. J. LeVeque, and K. T. Mandli. The GeoClaw software for depth-averaged flows with adaptive refinement. Advances in Water Resources, 34(9):1195– 1206, sep 2011. 5 [16] N. Bohr. Über die Serienspektra der Element. Zeitschrift für Physik, 2(5):423–469, oct 1920. 9 [17] J. L. Bona, V. A. Dougalis, and O. A. Karakashian. Fully discrete Galerkin methods for the Korteweg-de Vries equation. Computers & Mathematics with Applications, 12(7):859–884, jul 1986. 6 [18] P. Bonneton, F. Chazel, D. Lannes, F. Marche, and M. Tissier. A splitting approach for the fully nonlinear and weakly dispersive Green-Naghdi model. J. Comput. Phys., 230:1479– 1498, 2011. 6 G. Khakimzyanov, D. Dutykh, et al. 60 / 66 [19] M.-O. Bristeau, N. Goutal, and J. Sainte-Marie. Numerical simulations of a non-hydrostatic shallow water model. Comput. & Fluids, 47(1):51–64, aug 2011. 41 [20] M. Brocchini. A reasoned overview on Boussinesq-type models: the interplay between physics, mathematics and numerics. Proc. R. Soc. A, 469(2160):20130496, oct 2013. 5 [21] J. G. B. Byatt-Smith. The reflection of a solitary wave by a vertical wall. J. Fluid Mech., 197:503–521, 1988. 37 [22] F. Carbone, D. Dutykh, J. M. Dudley, and F. Dias. Extreme wave run-up on a vertical cliff. Geophys. Res. Lett., 40(12):3138–3143, 2013. 37 [23] M. J. Castro, M. de la Asuncion, J. Macias, C. Parés, E. D. Fernandez-Nieto, J. M. GonzalezVida, and T. Morales de Luna. IFCP Riemann solver: Application to tsunami modelling using CPUs. In M. E. Vazquez-Cendon, A. Hidalgo, P. Garcia-Navarro, and L. Cea, editors, Numerical Methods for Hyperbolic Equations: Theory and Applications, pages 237–244. CRC Press, Boca Raton, London, New York, Leiden, 2013. 48 [24] V. Casulli. A semi-implicit finite difference method for non-hydrostatic, free-surface flows. Int. J. Num. Meth. Fluids, 30(4):425–440, jun 1999. 6 [25] J. Chambarel, C. Kharif, and J. Touboul. Head-on collision of two solitary waves and residual falling jet formation. Nonlin. Processes Geophys., 16:111–122, 2009. 37, 40 [26] R. K.-C. Chan and R. L. Street. A computer study of finite-amplitude water waves. J. Comp. Phys., 6(1):68–94, aug 1970. 37, 38, 39 [27] K.-A. Chang, T.-J. Hsu, and P. L.-F. Liu. Vortex generation and evolution in water waves propagating over a submerged rectangular obstacle. Coastal Engineering, 44(1):13–36, sep 2001. 43 [28] J. G. Charney, R. Fjörtoft, and J. Neumann. Numerical Integration of the Barotropic Vorticity Equation. Tellus, 2(4):237–254, nov 1950. 29, 56 [29] F. Chazel. Influence of bottom topography on long water waves. M2AN, 41:771–799, 2007. 41 [30] F. Chazel, D. Lannes, and F. Marche. Numerical simulation of strongly nonlinear and dispersive waves using a Green-Naghdi model. J. Sci. Comput., 48:105–116, 2011. 6 [31] J.-B. Chen, M.-Z. Qin, and Y.-F. Tang. Symplectic and multi-symplectic methods for the nonlinear Schrödinger equation. Computers & Mathematics with Applications, 43(8-9):1095– 1106, apr 2002. 57 [32] M. Chhay, D. Dutykh, and D. Clamond. On the multi-symplectic structure of the SerreGreen-Naghdi equations. J. Phys. A: Math. Gen, 49(3):03LT01, jan 2016. 57 [33] A. Chorin. Numerical solution of the Navier-Stokes equations. Math. Comp., 22:745–762, 1968. 8 [34] L. B. Chubarov, S. V. Eletsky, Z. I. Fedotova, and G. S. Khakimzyanov. Simulation of surface waves by an underwater landslide. Russ. J. Numer. Anal. Math. Modelling, 20(5):425– 437, 2005. 48 [35] L. B. Chubarov, Z. I. Fedotova, Y. I. Shokin, and B. G. Einarsson. Comparative Analysis of Nonlinear Dispersive Shallow Water Models. Int. J. Comp. Fluid Dyn., 14(1):55–73, jan 2000. 6, 42 [36] L. B. Chubarov and Y. I. Shokin. The numerical modelling of long wave propagation in the framework of non-linear dispersion models. Comput. & Fluids, 15(3):229–249, jan 1987. [37] R. Cienfuegos, E. Barthélemy, and P. Bonneton. A fourth-order compact finite volume scheme for fully nonlinear and weakly dispersive Boussinesq-type equations. Part II: boundary conditions and validation. Int. J. Num. Meth. Fluids, 53(9):1423–1455, mar 2007. 6 Dispersive shallow water wave modelling. Part II 61 / 66 [38] D. Clamond and D. Dutykh. Practical use of variational principles for modeling water waves. Phys. D, 241(1):25–36, 2012. 57 [39] M. J. Cooker, P. D. Weidman, and D. S. Bale. Reflection of a high-amplitude solitary wave at a vertical wall. J. Fluid Mech., 342:141–158, 1997. 37, 38, 39, 40 [40] R. Courant, K. Friedrichs, and H. Lewy. Über die partiellen Differenzengleichungen der mathematischen Physik. Mathematische Annalen, 100(1):32–74, 1928. 26 [41] M. H. Dao and P. Tkalich. Tsunami propagation modelling - a sensitivity study. Nat. Hazards Earth Syst. Sci., 7:741–754, 2007. 32 [42] V. H. Davletshin. Force action of solitary waves on vertical structures. In Tsunami meeting, pages 41–43, Gorky, 1984. Institute of Applied Physics. 38, 39 [43] A. J. C. de Saint-Venant. Théorie du mouvement non-permanent des eaux, avec application aux crues des rivières et à l’introduction des marées dans leur lit. C. R. Acad. Sc. Paris, 73:147–154, 1871. 6, 8, 9 [44] M. W. Dingemans. Water wave propagation over uneven bottom. World Scientific, Singapore, 1997. 40 [45] V. A. Dougalis and O. A. Karakashian. On Some High-Order Accurate Fully Discrete Galerkin Methods for the Korteweg-de Vries Equation. Mathematics of Computation, 45(172):329, oct 1985. 6 [46] V. A. Dougalis and D. E. Mitsotakis. Theory and numerical analysis of Boussinesq systems: A review. In N. A. Kampanis, V. A. Dougalis, and J. A. Ekaterinaris, editors, Effective Computational Methods in Wave Propagation, pages 63–110. CRC Press, 2008. 6 [47] V. A. Dougalis, D. E. Mitsotakis, and J.-C. Saut. On some Boussinesq systems in two space dimensions: Theory and numerical analysis. Math. Model. Num. Anal., 41(5):254–825, 2007. 6 [48] P. G. Drazin and R. S. Johnson. Solitons: An introduction. Cambridge University Press, Cambridge, 1989. 15, 45 [49] A. Duran, D. Dutykh, and D. Mitsotakis. On the Galilean Invariance of Some Nonlinear Dispersive Wave Equations. Stud. Appl. Math., 131(4):359–388, nov 2013. 35 [50] D. Dutykh, M. Chhay, and F. Fedele. Geometric numerical schemes for the KdV equation. Comp. Math. Math. Phys., 53(2):221–236, 2013. 57 [51] D. Dutykh and D. Clamond. Efficient computation of steady solitary gravity waves. Wave Motion, 51(1):86–99, jan 2014. 45 [52] D. Dutykh, D. Clamond, P. Milewski, and D. Mitsotakis. Finite volume and pseudo-spectral schemes for the fully nonlinear 1D Serre equations. Eur. J. Appl. Math., 24(05):761–787, 2013. 6, 15, 37, 40 [53] D. Dutykh and F. Dias. Dissipative Boussinesq equations. C. R. Mecanique, 335:559–583, 2007. 41 [54] D. Dutykh and F. Dias. Energy of tsunami waves generated by bottom motion. Proc. R. Soc. A, 465:725–744, 2009. 10 [55] D. Dutykh and D. Ionescu-Kruse. Travelling wave solutions for some two-component shallow water models. J. Diff. Eqns., 261(2):1099–1114, jul 2016. 15 [56] D. Dutykh and H. Kalisch. Boussinesq modeling of surface waves due to underwater landslides. Nonlin. Processes Geophys., 20(3):267–285, may 2013. 48, 49 [57] D. Dutykh, T. Katsaounis, and D. Mitsotakis. Finite volume schemes for dispersive wave propagation and runup. J. Comput. Phys., 230(8):3035–3061, apr 2011. 6, 37 G. Khakimzyanov, D. Dutykh, et al. 62 / 66 [58] D. Dutykh, T. Katsaounis, and D. Mitsotakis. Finite volume methods for unidirectional dispersive wave models. Int. J. Num. Meth. Fluids, 71:717–736, 2013. 6, 37 [59] D. Dutykh, D. Mitsotakis, S. A. Beisel, and N. Y. Shokina. Dispersive waves generated by an underwater landslide. In E. Vazquez-Cendon, A. Hidalgo, P. Garcia-Navarro, and L. Cea, editors, Numerical Methods for Hyperbolic Equations: Theory and Applications, pages 245–250. CRC Press, Boca Raton, London, New York, Leiden, 2013. 48, 49 [60] F. Enet and S. T. Grilli. Experimental study of tsunami generation by three-dimensional rigid underwater landslides. J. Waterway, Port, Coastal and Ocean Engineering, 133(6):442– 454, 2007. 48 [61] R. C. Ertekin, W. C. Webster, and J. V. Wehausen. Waves caused by a moving disturbance in a shallow channel of finite width. J. Fluid Mech., 169:275–292, aug 1986. 16 [62] M. S. Fabien. Spectral Methods for Partial Dierential Equations that Model Shallow Water Wave Phenomena. Master, University of Washington, 2014. 6 [63] Z. I. Fedotova and E. D. Karepova. Variational principle for approximate models of wave hydrodynamics. Russ. J. Numer. Anal. Math. Modelling, 11(3):183–204, 1996. 57 [64] Z. I. Fedotova and G. S. Khakimzyanov. Shallow water equations on a movable bottom. Russ. J. Numer. Anal. Math. Modelling, 24(1):31–42, 2009. 16, 46 [65] Z. I. Fedotova, G. S. Khakimzyanov, and D. Dutykh. Energy equation for certain approximate models of long-wave hydrodynamics. Russ. J. Numer. Anal. Math. Modelling, 29(3):167–178, jan 2014. 48 [66] Z. I. Fedotova and V. Y. Pashkova. Methods of construction and the analysis of difference schemes for nonlinear dispersive models of wave hydrodynamics. Russ. J. Numer. Anal. Math. Modelling, 12(2), 1997. 42 [67] J. D. Fenton and M. M. Rienecker. A Fourier method for solving nonlinear water-wave problems: application to solitary-wave interactions. J. Fluid Mech., 118:411–443, apr 1982. 37, 38, 39 [68] E. D. Fernandez-Nieto, F. Bouchut, D. Bresch, M. J. Castro-Diaz, and A. Mangeney. A new Savage-Hutter type models for submarine avalanches and generated tsunami. J. Comput. Phys., 227(16):7720–7754, 2008. 48 [69] G. R. Flierl. Particle motions in large-amplitude wave fields. Geophysical & Astrophysical Fluid Dynamics, 18(1-2):39–74, aug 1981. 45 [70] S. Glimsdal, G. K. Pedersen, K. Atakan, C. B. Harbitz, H. P. Langtangen, and F. Lovholt. Propagation of the Dec. 26, 2004, Indian Ocean Tsunami: Effects of Dispersion and Source Characteristics. Int. J. Fluid Mech. Res., 33(1):15–43, 2006. 31 [71] S. Glimsdal, G. K. Pedersen, C. B. Harbitz, and F. Løvholt. Dispersion of tsunamis: does it really matter? Natural Hazards and Earth System Science, 13(6):1507–1526, jun 2013. 32, 49 [72] A. E. Green, N. Laws, and P. M. Naghdi. On the theory of water waves. Proc. R. Soc. Lond. A, 338:43–55, 1974. 5 [73] S. Grilli, S. Vogelmann, and P. Watts. Development of a 3D numerical wave tank for modeling tsunami generation by underwater landslides. Engng Anal. Bound. Elem., 26:301– 313, 2002. 48 [74] S. T. Grilli and P. Watts. Tsunami Generation by Submarine Mass Failure. I: Modeling, Experimental Validation, and Sensitivity Analyses. Journal of Waterway Port Coastal and Ocean Engineering, 131(6):283, 2005. 48 Dispersive shallow water wave modelling. Part II 63 / 66 [75] J. Grue, E. N. Pelinovsky, D. Fructus, T. Talipova, and C. Kharif. Formation of undular bores and solitary waves in the Strait of Malacca caused by the 26 December 2004 Indian Ocean tsunami. J. Geophys. Res., 113(C5):C05008, may 2008. 31 [76] E. Hairer, S. P. Nørsett, and G. Wanner. Solving ordinary differential equations: Nonstiff problems. Springer, 2009. 6 [77] E. Hairer and G. Wanner. Solving Ordinary Differential Equations II. Stiff and DifferentialAlgebraic Problems. Springer Series in Computational Mathematics, Vol. 14, 1996. 6 [78] J. Hammack, D. Henderson, P. Guyenne, and M. Yi. Solitary wave collisions. In Proc. 23rd International Conference on Offshore Mechanics and Arctic Engineering, 2004. 43 [79] F. H. Harlow and J. E. Welch. Numerical Calculation of Time-Dependent Viscous Incompressible Flow of Fluid with Free Surface. Phys. Fluids, 8:2182, 1965. 8 [80] H. Hermes. Introduction to Mathematical Logic. Universitext. Springer Berlin Heidelberg, Berlin, Heidelberg, 1973. 26 [81] N. J. Higham. Accuracy and Stability of Numerical Algorithms. SIAM Philadelphia, 2nd ed. edition, 2002. 25 [82] J. Horrillo, Z. Kowalik, and Y. Shigihara. Wave Dispersion Study in the Indian OceanTsunami of December 26, 2004. Marine Geodesy, 29(3):149–166, dec 2006. 6 [83] W. Huang. Practical Aspects of Formulation and Solution of Moving Mesh Partial Differential Equations. J. Comp. Phys., 171(2):753–775, aug 2001. 5 [84] W. Huang and R. D. Russell. Adaptive mesh movement - the MMPDE approach and its applications. J. Comp. Appl. Math., 128(1-2):383–398, mar 2001. 5 [85] A. M. Il’in. Differencing scheme for a differential equation with a small parameter affecting the highest derivative. Mathematical Notes of the Academy of Sciences of the USSR, 6(2):596–602, aug 1969. 5, 16 [86] M. Ioualalen, S. Migeon, and O. Sardoux. Landslide tsunami vulnerability in the Ligurian Sea: case study of the 1979 October 16 Nice international airport submarine landslide and of identified geological mass failures. Geophys. J. Int., 181(2):724–740, mar 2010. 48, 49 [87] R. S. Johnson. Camassa-Holm, Korteweg-de Vries and related models for water waves. J. Fluid Mech., 455:63–82, 2002. 57 [88] A. Kabbaj. Contribution à l’étude du passage des ondes de gravité et de la génération des ondes internes sur un talus, dans le cadre de la théorie de l’eau peu profonde. Thèse, Université Scientifique et Médicale de Grenoble, 1985. 41, 44 [89] M. Kazolea and A. I. Delis. A well-balanced shock-capturing hybrid finite volume-finite difference numerical scheme for extended 1D Boussinesq models. Appl. Numer. Math., 67:167–186, 2013. 6 [90] G. Khakimzyanov and D. Dutykh. On supraconvergence phenomenon for second order centered finite differences on non-uniform grids. J. Comp. Appl. Math., 326:1–14, dec 2017. 16, 17 [91] G. S. Khakimzyanov, D. Dutykh, and Z. I. Fedotova. Dispersive shallow water wave modelling. Part III: Model derivation on a globally spherical geometry. Submitted, pages 1–40, 2017. 56, 57 [92] G. S. Khakimzyanov, D. Dutykh, Z. I. Fedotova, and D. E. Mitsotakis. Dispersive shallow water wave modelling. Part I: Model derivation on a globally flat space. Submitted, pages 1–40, 2017. 5, 7, 10, 11, 44, 46 [93] G. S. Khakimzyanov, D. Dutykh, and O. Gusev. Dispersive shallow water wave modelling. Part IV: Numerical simulation on a globally spherical geometry. Submitted, pages 1–40, G. Khakimzyanov, D. Dutykh, et al. [94] [95] [96] [97] [98] [99] [100] [101] [102] [103] [104] [105] [106] [107] [108] [109] [110] [111] [112] [113] [114] 64 / 66 2017. 57 G. S. Khakimzyanov, D. Dutykh, D. E. Mitsotakis, and N. Y. Shokina. Numerical solution of conservation laws on moving grids. Submitted, pages 1–28, 2017. 16, 21, 26, 29, 48 J. W. Kim, K. J. Bai, R. C. Ertekin, and W. C. Webster. A derivation of the Green-Naghdi equations for irrotational flows. J. Eng. Math., 40(1):17–42, 2001. 57 J. W. Kim and R. C. Ertekin. A numerical study of nonlinear wave interaction in regular and irregular seas: irrotational Green-Naghdi model. Marine Structures, 13(4-5):331–347, jul 2000. 16 H. O. Kreiss and G. Scherer. Method of lines for hyperbolic equations. SIAM Journal on Numerical Analysis, 29:640–646, 1992. 6 A. A. Kurkin, S. V. Semin, and Y. A. Stepanyants. Transformation of surface waves over a bottom step. Izvestiya, Atmospheric and Oceanic Physics, 51(2):214–223, mar 2015. 40, 41 Z. Lai, C. Chen, G. W. Cowles, and R. C. Beardsley. A nonhydrostatic version of FVCOM: 1. Validation experiments. J. Geophys. Res., 115(C11):C11010, nov 2010. 41 O. Le Métayer, S. Gavrilyuk, and S. Hank. A numerical scheme for the Green-Naghdi model. J. Comp. Phys., 229(6):2034–2045, 2010. 6 D. Y. Le Roux. Spurious inertial oscillations in shallow-water models. J. Comp. Phys., 231(24):7959–7987, oct 2012. 32 E. K. Lindstrøm, G. K. Pedersen, A. Jensen, and S. Glimsdal. Experiments on slide generated waves in a 1:500 scale fjord model. Coastal Engineering, 92:12–23, oct 2014. 48 P. L.-F. Liu, T.-R. Wu, F. Raichlen, C. E. Synolakis, and J. C. Borrero. Runup and rundown generated by three-dimensional sliding masses. J. Fluid Mech., 536(1):107–144, jul 2005. 48 F. Løvholt, G. Pedersen, and G. Gisler. Oceanic propagation of a potential tsunami from the La Palma Island. J. Geophys. Res., 113(C9):C09026, sep 2008. 49 P. Lynett and P. L. F. Liu. A numerical study of submarine-landslide-generated waves and run-up. Proc. R. Soc. A, 458(2028):2885–2910, dec 2002. 48 O. S. Madsen and C. C. Mei. The transformation of a solitary wave over an uneven bottom. J. Fluid Mech., 39(04):781–791, dec 1969. 41 P. A. Madsen, H. B. Bingham, and H. Liu. A new Boussinesq method for fully nonlinear waves from shallow to deep water. J. Fluid Mech., 462:1–30, 2002. 37 P. A. Madsen, R. Murray, and O. R. Sorensen. A new form of the Boussinesq equations with improved linear dispersion characteristics. Coastal Engineering, 15:371–388, 1991. 6 P. A. Madsen and O. R. Sorensen. A new form of the Boussinesq equations with improved linear dispersion characteristics. Part 2. A slowly-varying bathymetry. Coastal Engineering, 18:183–204, 1992. 6 S. V. Manoylin. Some experimental and theoretical methods of estimation of tsunami wave action on hydro-technical structures and seaports. Technical report, Siberian Branch of Computing Center, Krasnoyarsk, 1989. 38, 39 T. Maxworthy. Experiments on collisions between solitary waves. J. Fluid Mech, 76:177–185, 1976. 37, 38, 39 J. W. Miles and R. Salmon. Weakly dispersive nonlinear gravity waves. J. Fluid Mech., 157:519–531, 1985. 57 S. M. Mirie and C. H. Su. Collision between two solitary waves. Part 2. A numerical study. J. Fluid Mech., 115:475–492, 1982. 37 D. Mitsotakis, B. Ilan, and D. Dutykh. On the Galerkin/Finite-Element Method for the Serre Equations. J. Sci. Comput., 61(1):166–195, feb 2014. 6, 15 Dispersive shallow water wave modelling. Part II 65 / 66 [115] E. Pelinovsky, B. H. Choi, T. Talipova, S. B. Wood, and D. C. Kim. Solitary wave transformation on the underwater step: Asymptotic theory and numerical experiments. Applied Mathematics and Computation, 217(4):1704–1718, oct 2010. 41, 44, 47 [116] D. H. Peregrine. Calculations of the development of an undular bore. J. Fluid Mech., 25(02):321–330, mar 1966. 6, 31, 33, 34 [117] D. H. Peregrine. Long waves on a beach. J. Fluid Mech., 27:815–827, 1967. 5, 16 [118] D. H. Peregrine. Water-Wave Impact on Walls. Annu. Rev. Fluid Mech., 35:23–43, 2003. 36 [119] B. Ranguelov, S. Tinti, G. Pagnoni, R. Tonini, F. Zaniboni, and A. Armigliato. The nonseismic tsunami observed in the Bulgarian Black Sea on 7 May 2007: Was it due to a submarine landslide? Geophys. Res. Lett., 35(18):L18613, sep 2008. 48 [120] S. C. Reddy and L. N. Trefethen. Stability of the method of lines. Numerische Mathematik, 62(1):235–267, 1992. 6 [121] G. Sadaka. Solution of 2D Boussinesq systems with FreeFem++: the flat bottom case. Journal of Numerical Mathematics, 20(3-4):303–324, jan 2012. 5 [122] A. A. Samarskii. The Theory of Difference Schemes. CRC Press, New York, 2001. 23, 26 [123] W. E. Schiesser. Method of lines solution of the Korteweg-de vries equation. Computers Mathematics with Applications, 28(10-12):147–154, 1994. 6 [124] F. J. Seabra-Santos, D. P. Renouard, and A. M. Temperville. Numerical and Experimental study of the transformation of a Solitary Wave over a Shelf or Isolated Obstacle. J. Fluid Mech, 176:117–134, 1987. 41, 42, 43, 45 [125] F. J. Seabra-Santos, A. M. Temperville, and D. P. Renouard. On the weak interaction of two solitary waves. Eur. J. Mech. B/Fluids, 8(2):103–115, 1989. 37 [126] F. Serre. Contribution à l’étude des écoulements permanents et variables dans les canaux. La Houille blanche, 8:830–872, 1953. 5 [127] F. Serre. Contribution to the study of long irrotational waves. La Houille blanche, 3:374–388, 1956. 5 [128] L. F. Shampine. ODE solvers and the method of lines. Numerical Methods for Partial Differential Equations, 10(6):739–755, 1994. 6 [129] Y. I. Shokin, Y. V. Sergeeva, and G. S. Khakimzyanov. Predictor-corrector scheme for the solution of shallow water equations. Russ. J. Numer. Anal. Math. Modelling, 21(5):459–479, jan 2006. 6, 21, 22 [130] G. Söderlind and L. Wang. Adaptive time-stepping and computational stability. J. Comp. Appl. Math., 185(2):225–243, 2006. 56 [131] O. R. Sørensen, H. A. Schäffer, and L. S. Sørensen. Boussinesq-type modelling using an unstructured finite element technique. Coastal Engineering, 50(4):181–198, feb 2004. 6 [132] C. H. Su and R. M. Mirie. On head-on collisions between two solitary waves. J. Fluid Mech., 98:509–525, 1980. 37, 38 [133] D. R. Tappin, P. Watts, and S. T. Grilli. The Papua New Guinea tsunami of 17 July 1998: anatomy of a catastrophic event. Nat. Hazards Earth Syst. Sci., 8:243–266, 2008. 49 [134] P. D. Thomas and C. K. Lombart. Geometric conservation law and its application to flow computations on moving grid. AIAA Journal, 17(10):1030–1037, 1979. 5 [135] A. N. Tikhonov and A. A. Samarskii. Homogeneous difference schemes. Zh. vych. mat., 1(1):5–63, 1961. 16 [136] A. N. Tikhonov and A. A. Samarskii. Homogeneous difference schemes on non-uniform nets. Zh. vych. mat., 2(5):812–832, 1962. 16 G. Khakimzyanov, D. Dutykh, et al. 66 / 66 [137] S. Tinti, E. Bortolucci, and C. Vannini. A Block-Based Theoretical Model Suited to Gravitational Sliding. Natural Hazards, 16(1):1–28, 1997. 48 [138] J. Touboul and E. Pelinovsky. Bottom pressure distribution under a solitonic wave reflecting on a vertical wall. Eur. J. Mech. B/Fluids, 48:13–18, nov 2014. 37 [139] M. Walkley and M. Berzins. A finite element method for the two-dimensional extended Boussinesq equations. Int. J. Num. Meth. Fluids, 39(10):865–885, aug 2002. 6 [140] S. N. Ward. Landslide tsunami. J. Geophysical Res., 106:11201–11215, 2001. 48 [141] P. Watts, S. T. Grilli, J. T. Kirby, G. J. Fryer, and D. R. Tappin. Landslide tsunami case studies using a Boussinesq model and a fully nonlinear tsunami generation model. Natural Hazards And Earth System Science, 3(5):391–402, 2003. 48 [142] P. Watts, F. Imamura, and S. T. Grilli. Comparing model simulations of three benchmark tsunami generation cases. Science of Tsunami Hazards, 18(2):107–123, 2000. 48 [143] G. Wei and J. T. Kirby. Time-Dependent Numerical Code for Extended Boussinesq Equations. J. Waterway, Port, Coastal and Ocean Engineering, 121(5):251–261, sep 1995. 6 [144] N. N. Zagryadskaya, S. V. Ivanova, L. S. Nudner, and A. I. Shoshin. Action of long waves on a vertical obstacle. Bulletin of VNIIG, 138:94–101, 1980. 38, 39 [145] V. E. Zakharov. What Is Integrability? Springer Series in Nonlinear Dynamics, 1991. 15 [146] Y. Zhang, A. B. Kennedy, N. Panda, C. Dawson, and J. J. Westerink. Boussinesq-GreenNaghdi rotational water wave theory. Coastal Engineering, 73:13–27, mar 2013. 42 [147] B. B. Zhao, R. C. Ertekin, and W. Y. Duan. A comparative study of diffraction of shallowwater waves by high-level IGN and GN equations. J. Comput. Phys., 283:129–147, feb 2015. 6 [148] M. I. Zheleznyak. Influence of long waves on vertical obstacles. In E. N. Pelinovsky, editor, Tsunami Climbing a Beach, pages 122–139. Applied Physics Institute Press, Gorky, 1985. 39 [149] M. I. Zheleznyak and E. N. Pelinovsky. Physical and mathematical models of the tsunami climbing a beach. In E. N. Pelinovsky, editor, Tsunami Climbing a Beach, pages 8–34. Applied Physics Institute Press, Gorky, 1985. 16 G. Khakimzyanov: Institute of Computational Technologies, Siberian Branch of the Russian Academy of Sciences, Novosibirsk 630090, Russia E-mail address: Khak@ict.nsc.ru D. Dutykh: LAMA, UMR 5127 CNRS, Université Savoie Mont Blanc, Campus Scientifique, F-73376 Le Bourget-du-Lac Cedex, France E-mail address: Denys.Dutykh@univ-savoie.fr URL: http://www.denys-dutykh.com/ O. Gusev: Institute of Computational Technologies, Siberian Branch of the Russian Academy of Sciences, Novosibirsk 630090, Russia E-mail address: gusev_oleg_igor@mail.ru N. Yu. Shokina: Institute of Computational Technologies, Siberian Branch of the Russian Academy of Sciences, Novosibirsk 630090, Russia E-mail address: Nina.Shokina@googlemail.com URL: https://www.researchgate.net/profile/Nina_Shokina/ 

Beat by Beat: Classifying Cardiac Arrhythmias with Recurrent Neural Networks arXiv:1710.06319v2 [cs.LG] 24 Oct 2017 Patrick Schwab, Gaetano C Scebba, Jia Zhang, Marco Delai, Walter Karlen Mobile Health Systems Lab, Department of Health Sciences and Technology ETH Zurich, Switzerland Abstract With tens of thousands of electrocardiogram (ECG) records processed by mobile cardiac event recorders every day, heart rhythm classification algorithms are an important tool for the continuous monitoring of patients at risk. We utilise an annotated dataset of 12,186 single-lead ECG recordings to build a diverse ensemble of recurrent neural networks (RNNs) that is able to distinguish between normal sinus rhythms, atrial fibrillation, other types of arrhythmia and signals that are too noisy to interpret. In order to ease learning over the temporal dimension, we introduce a novel task formulation that harnesses the natural segmentation of ECG signals into heartbeats to drastically reduce the number of time steps per sequence. Additionally, we extend our RNNs with an attention mechanism that enables us to reason about which heartbeats our RNNs focus on to make their decisions. Through the use of attention, our model maintains a high degree of interpretability, while also achieving state-of-the-art classification performance with an average F1 score of 0.79 on an unseen test set (n=3,658). 1. Introduction Cardiac arrhythmias are a heterogenous group of conditions that is characterised by heart rhythms that do not follow a normal sinus pattern. One of the most common arrhythmias is atrial fibrillation (AF) with an agedependant population prevalence of 2.3 - 3.4% [1]. Due to the increased mortality associated with arrhythmias, receiving a timely diagnosis is of paramount importance for patients [1, 2]. To diagnose cardiac arrhythmias, medical professionals typically consider a patient’s electrocardiogram (ECG) as one of the primary factors [2]. In the past, clinicians recorded these ECGs mainly using multi-lead clinical monitors or Holter devices. However, the recent advent of mobile cardiac event recorders has given patients the ability to remotely record short ECGs using devices with a single lead. We propose a machine-learning approach based on recurrent neural networks (RNNs) to differentiate between various types of heart rhythms in this more challenging setting with just a single lead and short ECG record lengths. To ease learning of dependencies over the temporal dimension, we introduce a novel task formulation that harnesses the natural beat-wise segmentation of ECG signals. In addition to utilising several heartbeat features that have been shown to be highly discriminative in previous works, we also use stacked denoising autoencoders (SDAE) [3] to capture differences in morphological structure. Furthermore, we extend our RNNs with a soft attention mechanism [4–7] that enables us to reason about which ECG segments the RNNs prioritise for their decision making. 2. Methodology Our cardiac rhythm classification pipeline consists of multiple stages (figure 1). The core idea of our setup is to extract a diverse set of features from the sequence of heartbeats in an ECG record to be used as input features to an ensemble of RNNs. We blend the individual models’ predictions into a per-class classification score using a multilayer perceptron (MLP) with a softmax output layer. The following paragraphs explain the stages shown in figure 1 in more detail. ECG Dataset. We use the dataset of the PhysioNet Computing in Cardiology (CinC) 2017 challenge [8] which contains 12,186 unique single-lead ECG records of varying length. Experts annotated each of these ECGs as being either a normal sinus rhythm, AF, an other ar- ECG Preprocessing Features Level 1 Models Level 2 Blender Model1 Normalise Segment Extract Features ... Modeln Blend Classification Normal AF Other Noise Figure 1. An overview of our ECG classification pipeline. rhythmia or too noisy to classify. The challenge organisers keep 3,658 (30%) of these ECG records private as a test set. Additionally, we hold out a non-stratified random subset of 20% of the public dataset as a validation set. For some RNN configurations, we further augment the training data with labelled samples extracted from other PhysioNet databases [9–12] in order to even out misbalanced class sizes in the training set. As an additional measure against the imbalanced class distribution of the dataset, we weight each training sample’s contribution to the loss function to be inversely proportional to its class’ prevalence in the overall dataset. Normalisation. Prior to segmentation, we normalise the ECG recording to have a mean value of zero and a standard deviation of one. We do not apply any additional filters as all ECGs were bandpass-filtered by the recording device. Segmentation. Following normalisation, we segment the ECG into a sequence of heartbeats. We decide to reformulate the given task of classifying arrhythmias as a sequence classification task over heartbeats rather than over raw ECG readings. The motivation behind the reformulation is that it significantly reduces the number of time steps through which the error signal of our RNNs has to propagate. On the training set, the reformulation reduces the mean number of time steps per ECG from 9000 to just 33. To perform the segmentation, we use a customised QRS detector based on Pan-Tompkin’s [13] that identifies Rpeaks in the ECG recording. We extend their algorithm by adapting the threshold with a moving average of the ECG signal to be more resilient against the commonly encountered short bursts of noise. For the purpose of this work, we define heartbeats using a symmetric fixed size window with a total length of 0.66 seconds around R-peaks. We pass the extracted heartbeat sequence in its original order to the feature extraction stage. Feature Extraction. We extract a diverse set of features from each heartbeat in an ECG recording. Specifically, we extract the time since the last heartbeat (δRR), the relative wavelet energy (RWE) over five frequency bands, the total wavelet energy (TWE) over those frequency bands, the R amplitude, the Q amplitude, QRS duration and wavelet entropy (WE). Previous works demonstrated the efficacy of all of these features in discriminating cardiac arrhythmias from normal heart rhythms [14–18]. In addition to the aforementioned features, we also train two SDAEs on the heartbeats in an unsupervised manner with the goal of learning more nuanced differences in morphology of individual heartbeats. We train one SDAE on the extracted heartbeats of the training set and the other on their wavelet coefficients. We then use the encoding side of the SDAEs to extract low-dimensional embeddings of each heartbeat and each heartbeat’s wavelet coefficients to be used as additional input features. Finally, we concatenate all ex- tracted features into a single feature vector per heartbeat and pass them to the level 1 models in original heartbeat sequence order. Level 1 Models. We build an ensemble of level 1 models to classify the sequence of per-beat feature vectors. To increase the diversity within our ensemble, we train RNNs in various binary classification settings and with different hyperparameters. We use RNNs with 1 - 5 recurrent layers that consist of either Gated Recurrent Units (GRU) [19] or Bidirectional Long Short-Term Memory (BLSTM) units [20], followed by an optional attention layer, 1 - 2 forward layers and a softmax output layer. Additionally, we infer a nonparametric Hidden Semi-Markov Model (HSMM) [21] with 64 initial states for each class in an unsupervised setting. In total, our ensemble of level 1 models consists of 15 RNNs and 4 HSMMs. We concatenate the ECG’s normalised log-likelihoods under the per-class HSMMs and the RNNs’ softmax outputs into a single prediction vector. We pass the prediction vector of the level 1 models to the level 2 blender model. Level 2 Blender. We use blending [22] to combine the predictions of our level 1 models and a set of ECG-wide features into a final per-class classification score. The additional features are the RWE and WE over the whole ECG and the absolute average deviation (AAD) of the WE and δRR of all beats. We employ a MLP with a softmax output layer as our level 2 blender model. In order to avoid overfitting to the training set, we train the MLP on the validation set. Hyperparameter Selection. To select the hyperparameters of our level 1 RNNs, we performed a grid search on the range of 0 - 75% for the dropout and recurrent dropout percentages, 60 - 512 for the number of units per hidden layer and 1 - 5 for the number of recurrent layers. We found that RNNs trained with 35% dropout, 65% recurrent dropout, 80 units per hidden layer and 5 recurrent layers (plus an additional attention layer) achieve consistently strong results across multiple binary classification settings. For our level 2 blender model, we utilise Bayesian optimisation [23] to select the number of layers, number of hidden units per layer, dropout and number of training epochs. We perform a 5-fold cross validation on the validation set to select the blender model’s hyperparameters. 2.1. Attention over Heartbeats Attention mechanisms have been shown to allow for greater interpretability of neural networks in a variety of tasks in computer vision and natural language processing [4–7]. In this work, we apply soft attention over the heartbeats contained in ECG signals in order to gain a deeper understanding of the decision-making process of our RNNs. Consider the case of an RNN that is processing a sequence of T heartbeats. The topmost recurrent layer Normal: 94 % Normal vs. all Other vs. all Other: 67 % ECG at (s) (s) Figure 2. A visualisation of the attention values at (top) of two different RNNs over two sample ECG recordings (bottom). The graphs on top of the ECG recordings show the attention values at associated with each identified heartbeat (dashed line). The labels in the left and right corners of the attention value graphs show the settings the model was trained for and their classification confidence, respectively. The recording on the left (A02149) represents a normal sinus rhythm. Due to the regular heart rhythm in the ECG, a distinctive pattern of approximately equally weighted attention on each heartbeat emerges from our RNN that was trained to distinguish between normal sinus rhythms and all other types of rhythms. The recording on the right (A04661) is labelled as an other arrhythmia. The RNN trained to identify other arrhythmias focuses primarily on a sudden, elongated pause in the heart rhythm to decide that the record is most likely an other arrhythmia. outputs a hidden state ht at every time step t ∈ [1, T ] of the sequence. We extend some of our RNNs with additive soft attention over the hidden states ht to obtain a context vector c that attenuates the most informative hidden states ht of a heartbeat sequence. Based on the definition in [6], we use the following set of equations: ut = tanh(Wbeat ht + bbeat ) (1) at = sof tmax(uTt ubeat ) X c= at ht (2) [24]. In contrast, we achieve state-of-the-art performance with significantly fewer trainable parameters by harnessing the natural heartbeat segmentation of ECGs and discriminative features from previous works. Additionally, we pay consideration to the fact that interpretability remains a challenge in applying machine learning to the medical domain [25] by extending our models with an attention mechanism that enables medical professionals to reason about which heartbeats contributed most to the decisionmaking process of our RNNs. (3) t Where equation (1) is a single-layer MLP with a weight matrix Wbeat and bias bbeat to obtain ut as a hidden representation of ht [6]. In equation (2), we calculate the attention factors at for each heartbeat by computing a softmax over the dot-product similarities of every heartbeat’s ut to the heartbeat context vector ubeat . ubeat corresponds to a hidden representation of the most informative heartbeat [6]. We jointly optimise Wbeat , bbeat and ubeat with the other RNN parameters during training. In figure 2, we showcase two examples of how qualitative analysis of the attention factors at of equation (2) provides a deeper understanding of our RNNs’ decision making. 3. Related Work Our work builds on a long history of research in detecting cardiac arrhythmias from ECG records by making use of features that have been shown to be highly discriminative in distinguishing certain arrhythmias from normal heart rhythms [14–18]. Recently, Rajpurkar et al. proposed a 34-layer convolutional neural network (CNN) to reach cardiologist-level performance in classifying a large set of arrhythmias from mobile cardiac event recorder data 4. Results and Conclusion We present a machine-learning approach to distinguishing between multiple types of heart rhythms. Our approach utilises an ensemble of RNNs to jointly identify temporal and morphological patterns in segmented ECG recordings of any length. In detail, our approach reaches an average F1 score of 0.79 on the private test set of the PhysioNet CinC Challenge 2017 (n = 3, 658) with class-wise F1 scores of 0.90, 0.79 and 0.68 for normal rhythms, AF and other arrhythmias, respectively. On top of its state-of-theart performance, our approach maintains a high degree of interpretability through the use of a soft attention mechanism over heartbeats. In the spirit of open research, we make an implementation of our cardiac rhythm classification system available through the PhysioNet 2017 Open Source Challenge. Future Work. Based on our discussions with a cardiologist, we hypothesise that the accuracy of our models could be further improved by incorporating contextual information, such as demographic information, data from other clinical assessments and behavioral aspects. Acknowledgements This work was partially funded by the Swiss National Science Foundation (SNSF) project No. 167302 within the National Research Program (NRP) 75 “Big Data” and SNSF project No. 150640. We thank Prof. Dr. med. Firat Duru for providing valuable insights into the decisionmaking process of cardiologists. References [1] Ball J, Carrington MJ, McMurray JJ, Stewart S. Atrial fibrillation: Profile and burden of an evolving epidemic in the 21st century. International Journal of Cardiology 2013; 167(5):1807–1824. [2] Camm AJ, Kirchhof P, Lip GY, Schotten U, Savelieva I, Ernst S, Van Gelder IC, Al-Attar N, Hindricks G, Prendergast B, et al. Guidelines for the management of atrial fibrillation. European Heart Journal 2010;31:2369–2429. [3] Vincent P, Larochelle H, Lajoie I, Bengio Y, Manzagol PA. Stacked denoising autoencoders: Learning useful representations in a deep network with a local denoising criterion. Journal of Machine Learning Research 2010; 11(Dec):3371–3408. [4] Bahdanau D, Cho K, Bengio Y. Neural machine translation by jointly learning to align and translate. In International Conference on Learning Representations, 2015. [5] Xu K, Ba J, Kiros R, Cho K, Courville A, Salakhudinov R, Zemel R, Bengio Y. Show, attend and tell: Neural image caption generation with visual attention. In International Conference on Machine Learning. 2015; 2048–2057. [6] Yang Z, Yang D, Dyer C, He X, Smola AJ, Hovy EH. Hierarchical attention networks for document classification. In Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2016; 1480–1489. [7] Zhang Z, Xie Y, Xing F, McGough M, Yang L. MDNet: A Semantically and Visually Interpretable Medical Image Diagnosis Network. In International Conference on Computer Vision and Pattern Recognition, arXiv preprint arXiv:1707.02485, 2017. [8] Clifford GD, Liu CY, Moody B, Lehman L, Silva I, Li Q, Johnson AEW, Mark RG. AF classification from a short single lead ECG recording: The Physionet Computing in Cardiology Challenge 2017. In Computing in Cardiology, 2017. [9] Goldberger AL, Amaral LAN, Glass L, Hausdorff JM, Ivanov PC, Mark RG, Mietus JE, Moody GB, Peng CK, Stanley HE. PhysioBank, PhysioToolkit, and PhysioNet: Components of a New Research Resource for Complex Physiologic Signals. Circulation 2000;101(23):e215–e220. [10] Moody GB, Mark RG. The impact of the MIT-BIH arrhythmia database. IEEE Engineering in Medicine and Biology Magazine 2001;20(3):45–50. [11] Moody G. A new method for detecting atrial fibrillation using RR intervals. In Computers in Cardiology. IEEE, 1983; 227–230. [12] Greenwald SD, Patil RS, Mark RG. Improved detection and classification of arrhythmias in noise-corrupted electrocardiograms using contextual information. In Computers in Cardiology. IEEE, 1990; 461–464. [13] Pan J, Tompkins WJ. A real-time QRS detection algorithm. IEEE Transactions on Biomedical Engineering 1985; 3:230–236. [14] Sarkar S, Ritscher D, Mehra R. A detector for a chronic implantable atrial tachyarrhythmia monitor. IEEE Transactions on Biomedical Engineering 2008;55(3):1219–1224. [15] Tateno K, Glass L. Automatic detection of atrial fibrillation using the coefficient of variation and density histograms of RR and δRR intervals. Medical and Biological Engineering and Computing 2001;39(6):664–671. [16] Garcı́a M, Ródenas J, Alcaraz R, Rieta JJ. Application of the relative wavelet energy to heart rate independent detection of atrial fibrillation. computer methods and programs in biomedicine 2016;131:157–168. [17] Ródenas J, Garcı́a M, Alcaraz R, Rieta JJ. Wavelet entropy automatically detects episodes of atrial fibrillation from single-lead electrocardiograms. Entropy 2015;17(9):6179– 6199. [18] Alcaraz R, Vayá C, Cervigón R, Sánchez C, Rieta J. Wavelet sample entropy: A new approach to predict termination of atrial fibrillation. In Computers in Cardiology. IEEE, 2006; 597–600. [19] Chung J, Gulcehre C, Cho K, Bengio Y. Empirical evaluation of gated recurrent neural networks on sequence modeling. In Neural Information Processing Systems, Workshop on Deep Learning, arXiv preprint arXiv:1412.3555, 2014. [20] Graves A, Jaitly N, Mohamed Ar. Hybrid speech recognition with deep bidirectional lstm. In Automatic Speech Recognition and Understanding, IEEE Workshop on. IEEE, 2013; 273–278. [21] Johnson MJ, Willsky AS. Bayesian nonparametric hidden semi-markov models. Journal of Machine Learning Research 2013;14(Feb):673–701. [22] Wolpert DH. Stacked generalization. Neural networks 1992;5(2):241–259. [23] Bergstra J, Yamins D, Cox DD. Hyperopt: A python library for optimizing the hyperparameters of machine learning algorithms. In Proceedings of the 12th Python in Science Conference. 2013; 13–20. [24] Rajpurkar P, Hannun AY, Haghpanahi M, Bourn C, Ng AY. Cardiologist-level arrhythmia detection with convolutional neural networks. arXiv preprint, arXiv:1707.01836, 2017. [25] Cabitza F, Rasoini R, Gensini G. Unintended consequences of machine learning in medicine. Journal of the American Medical Association 2017;318(6):517–518. Address for correspondence: Patrick Schwab, ETH Zurich Balgrist Campus, BAA D, Lengghalde 5, 8092 Zurich patrick.schwab@hest.ethz.ch 

arXiv:cs/0512057v1 [cs.PL] 14 Dec 2005 Resource Control for Synchronous Cooperative Threads∗ Roberto M. Amadio Université Paris 7 † Silvano Dal Zilio CNRS Marseille‡ 7th May 2017 Abstract We develop new methods to statically bound the resources needed for the execution of systems of concurrent, interactive threads. Our study is concerned with a synchronous model of interaction based on cooperative threads whose execution proceeds in synchronous rounds called instants. Our contribution is a system of compositional static analyses to guarantee that each instant terminates and to bound the size of the values computed by the system as a function of the size of its parameters at the beginning of the instant. Our method generalises an approach designed for first-order functional languages that relies on a combination of standard termination techniques for term rewriting systems and an analysis of the size of the computed values based on the notion of quasi-interpretation. We show that these two methods can be combined to obtain an explicit polynomial bound on the resources needed for the execution of the system during an instant. As a second contribution, we introduce a virtual machine and a related bytecode thus producing a precise description of the resources needed for the execution of a system. In this context, we present a suitable control flow analysis that allows to formulate the static analyses for resource control at byte code level. 1 Introduction The problem of bounding the usage made by programs of their resources has already attracted considerable attention. Automatic extraction of resource bounds has mainly focused on (first-order) functional languages starting from Cobham’s characterisation [18] of polynomial time functions by bounded recursion on notation. Following work, see e.g. ∗ Work partially supported by ACI Sécurité Informatique CRISS. Laboratoire Preuves, Programmes et Systèmes, UMR-CNRS 7126. ‡ Laboratoire d’Informatique Fondamentale de Marseille, UMR-CNRS 6166 † 1 [8, 19, 21, 23], has developed various inference techniques that allow for efficient analyses while capturing a sufficiently large range of practical algorithms. Previous work [10, 24] has shown that polynomial time or space bounds can be obtained by combining traditional termination techniques for term rewriting systems with an analysis of the size of computed values based on the notion of quasi-interpretation. Thus, in a nutshell, resource control relies on termination and bounds on data size. This approach to resource control should be contrasted with traditional worst case execution time technology (see, e.g., [30]): the bounds are less precise but they apply to a larger class of algorithms and are functional in the size of the input, which seems more appropriate in the context of the applications we have in mind (see below). In another direction, one may compare the approach with the one based on linear logic (see, e.g., [7]): while in principle the linear logic approach supports higher-order functions, it does not offer yet a user-friendly programming language. In [3, 4], we have considered the problem of automatically inferring quasi-interpretations in the space of multi-variate max-plus polynomials. In [1], we have presented a virtual machine and a corresponding bytecode for a first-order functional language and shown how size and termination annotations can be formulated and verified at the level of the bytecode. In particular, we can derive from the verification an explicit polynomial bound on the space required to execute a given bytecode. In this work, we aim at extending and adapting these results to a concurrent framework. As a starting point, we choose a basic model of parallel threads interacting on shared variables. The kind of concurrency we consider is a cooperative one. This means that by default a running thread cannot be preempted unless it explicitly decides to return the control to the scheduler. In preemptive threads, the opposite hypothesis is made: by default a running thread can be preempted at any point unless it explicitly requires that a series of actions is atomic. We refer to, e.g., [28] for an extended comparison of the cooperative and preemptive models. Our viewpoint is pragmatic: the cooperative model is closer to the sequential one and many applications are easier to program in the cooperative model than in the preemptive one. Thus, as a first step, it makes sense to develop a resource control analysis for the cooperative model. The second major design choice is to assume that the computation is regulated by a notion of instant. An instant lasts as long as a thread can make some progress in the current instant. In other terms, an instant ends when the scheduler realizes that all threads are either stopped, or waiting for the next instant, or waiting for a value that no thread can produce in the current instant. Because of this notion of instant, we regard our model as synchronous. Because the model includes a logical notion of time, it is possible for a thread to react to the absence of an event. The reaction to the absence of an event is typical of synchronous languages such as Esterel [9]. Boussinot et al. have proposed a weaker version of this feature where the reaction to the absence happens in the following instant [13] and they have implemented it in various programming environments based on C, Java, and Scheme [31]. Applications suited to this programming style include: event-driven applications, graphical user interfaces, simulations (e.g. N-bodies problem, cellular automata, ad hoc networks), web 2 services, multiplayer online games, . . . Boussinot et al. have also advocated the relevance of this concept for the programming of mobile code and demonstrated that the possibility for a ‘synchronous’ mobile agent to react to the absence of an event is an added factor of flexibility for programs designed for open distributed systems, whose behaviours are inherently difficult to predict. These applications rely on data structure such as lists and trees whose size needs to be controlled. Recently, Boudol [12] has proposed a formalisation of this programming model. Our analysis will essentially focus on a small fragment of this model without higher-order functions, and where the creation of fresh memory cells (registers) and the spawning of new threads is only allowed at the very beginning of an instant. We believe that what is left is still expressive and challenging enough as far as resource control is concerned. Our analysis goes in three main steps. A first step is to guarantee that each instant terminates (Section 3.1). A second step is to bound the size of the computed values as a function of the size of the parameters at the beginning of the instant (Section 3.2). A third step, is to combine the termination and size analyses. Here we show how to obtain polynomial bounds on the space and time needed for the execution of the system during an instant as a function of the size of the parameters at the beginning of the instant (Section 3.3). A characteristic of our static analyses is that to a great extent they make abstraction of the memory and the scheduler. This means that each thread can be analysed separately, that the complexity of the analyses grows linearly in the number of threads, and that an incremental analysis of a dynamically changing system of threads is possible. Preliminary to these analyses, is a control flow analysis (Section 2.1) that guarantees that each thread performs each read instruction (in its body code) at most once in an instant. This condition is instrumental to resource control. In particular, it allows to regard behaviours as functions of their initial parameters and the registers they may read in the instant. Taking this functional viewpoint, we are able to adapt the main techniques developed for proving termination and size bounds in the first-order functional setting. We point out that our static size analyses are not intended to predict the size of the system after arbitrarily many instants. This is a harder problem which in general requires an understanding of the global behaviour of the system and/or stronger restrictions on the programs we can write. For the language studied in this paper, we advocate a combination of our static analyses with a dynamic controller that at the end of each instant checks the size of the parameters of the system and may decide to stop some threads taking too much space. Along the way and in appendix A, we provide a number of programming examples illustrating how certain synchronous and/or concurrent programming paradigms can be represented in our model. These examples suggest that the constraints imposed by the static analyses are not too severe and that their verification can be automated. As a second contribution, we describe a virtual machine and the related bytecode for our programming model (Section 4). This provides a more precise description of the resources needed for the execution of the systems we consider and opens the way to the verification of resource bounds at the bytecode level, following the ‘typed assembly language’ approach adopted in [1] for the purely functional fragment of the language. More precisely, we de3 scribe a control flow analysis that allows to recover the conditions for termination and size bounds at bytecode level and we show that the control flow analysis is sufficiently liberal to accept the code generated by a rather standard compilation function. Proofs are available in appendix B. 2 A Model of Synchronous Cooperative Threads A system of synchronous cooperative threads is described by (1) a list of mutually recursive type and constructor definitions and (2) a list of mutually recursive function and behaviour definitions relying on pattern matching. In this respect, the resulting programming language is reminiscent of Erlang [5], which is a practical language to develop concurrent applications. The set of instructions a behaviour can execute is rather minimal. Indeed, our language can be regarded as an intermediate code where, for instance, general patternmatching has been compiled into a nesting of if then else constructs and complex control structures have been compiled into a simple tail-recursive form. Types We denote type names with t, t′ , . . . and constructors with c, c′ , . . . We will also denote with r, r′ , . . . constructors of arity 0 and of ‘reference’ type (see equation of kind (2) below) and we will refer to them as registers (thus registers are constructors). The values v, v ′, . . . computed by programs are first order terms built out of constructors. Types and constructors are declared via recursive equations that may be of two kinds: (1) (2) t = . . . | c of t1 , . . . , tn | . . . t = Ref (t′ ) with . . . | r = v | . . . In (1) we declare a type t with a constructor c of functional type (t1 , . . . , tn ) → t. In (2) we declare a type t of registers referencing values of type t′ and a register r with initial value v. As usual, type definitions can be mutually recursive (functional and reference types can be intermingled) and it is assumed that all types and constructors are declared exactly once. This means that we can associate a unique type with every constructor and that with respect to this association we can say when a value is well-typed. For instance, we may define the type nat of natural numbers in unary format by the equation nat = z | s of nat and the type llist of linked lists of natural numbers by the equations nlist = nil | cons of (nat, llist) and llist = Ref (nlist) with r = cons(z, r). The last definition declares a register r of type llist with initial value the infinite (cyclic) list containing only z’s. Finally, we have a special behaviour type, beh. Elements of type beh do not return a value but produce side effects. We denote with β either a regular type or beh. Expressions We let x, y, . . . denote variables ranging over values. The size |v| of a value v is defined by |c| = 0 and |c(v1 , . . . , vn )| = 1 + |v1 | + · · · + |vn |. In the following, we will use the vectorial notation a to denote either a vector a1 , . . . , an or a sequence a1 · · · an 4 of elements. We use σ, σ ′ , . . . to denote a substitution [v/x], where v and x have the same length. A pattern p is a well-typed term built out of constructors and variables. In particular, a shallow linear pattern p is a pattern c(x1 , . . . , xn ), where c is a constructor of arity n and the variables x1 , . . . , xn are all distinct. Expressions, e, and expression bodies, eb, are defined as: e ::= x | c(e1 , . . . , ek ) | f (e1 , . . . , en ) eb ::= e | match x with p then eb else eb where f is a functional symbol of type (t1 , . . . , tn ) → t, specified by an equation of the kind f (x1 , . . . , xn ) = eb, and where p is a shallow linear pattern. A closed expression body eb evaluates to a value v according to the following standard rules: (e1 ) (e4 ) r⇓r  (e2 ) e⇓v c(e) ⇓ c(v) (e3 ) [v/x]eb 1 ⇓ v  match c(v) with c(x) ⇓v then eb 1 else eb 2 e ⇓ v, (e5 )  f (x) = eb, [v/x]eb ⇓ v f (e) ⇓ v eb 2 ⇓ v c 6= d  match c(v) with d(x) ⇓v then eb 1 else eb 2 Since registers are constructors, rule (e1 ) is a special case of rule (e2 ); we keep the rule for clarity. Behaviours Some function symbols may return a thread behaviour b, b′ , . . . rather than a value. In contrast to ‘pure’ expressions, a behaviour does not return a result but produces side-effects by reading and writing registers. A behaviour may also affect the scheduling status of the thread executing it. We denote with b, b′ , . . . behaviours defined as follows: b ::= stop | f (e) | yield .b | next.f (e) | ̺ := e.b | read ̺ with p1 ⇒ b1 | · · · | pn ⇒ bn | [ ] ⇒ f (e) | match x with c(x) then b1 else b2 where: (i) f is a functional symbol of type t1 , . . . , tn → beh, defined by an equation f (x) = b, (ii) ̺, ̺′ , . . . range over variables and registers, and (iii) p1 , . . . , pn are either shallow linear patterns or variables. We also denote with [ ] a special symbol that will be used in the default case of read expressions (see the paragraph Scheduler below). Note that if the pattern pi is a variable then the following branches including the default one can never be executed. The effect of the various instructions is informally described as follows: stop, terminates the executing thread for ever; yield .b, halts the execution and hands over the control to the scheduler — the control should return to the thread later in the same instant and execution resumes with b; f (e) and next.f (e) switch to another behaviour immediately or at the beginning of the following instant; r := e.b, evaluates the expression e, assigns its value to r and proceeds with the evaluation of b; read r with p1 ⇒ b1 | · · · | pn ⇒ bn | [ ] ⇒ b, waits until the value of r matches one of the patterns p1 , . . . , pn (there could be no 5 delay) and yields the control otherwise; if at the end of the instant the thread is always stuck waiting for a matching value then it starts the behaviour b in the following instant; match v with p then b1 else b2 filters the value v according to the pattern p, it never blocks the execution. Note that if p is a pattern and v is a value there is at most one matching substitution σ such that v = σp. X Behaviour reduction is described by the 9 rules below. A reduction (b, s)→(b′ , s′) means that the behaviour b with store s runs an atomic sequence of actions till b′ , producing a store s′ , and returning the control to the scheduler with status X. A status is a value in {N, R, S, W } that represents one of the four possible state of a thread — N stands for next (the thread will resume at the beginning of the next instant), R for run, S for stopped, and W for wait (the thread is blocked on a read statement). (b1 ) S (stop, s) → (stop, s) (b2 ) R (yield .b, s) → (b, s) (b3 ) N (next .f (e), s) → (f (e), s) X X (b2 , s) → (b′ , s′ ), c 6= d ([v/x]b1 , s) → (b′ , s′ )    match c(v) match c(v) (b ) (b4 ) 5 X X  with c(x)  with d(x) , s → (b′ , s′ ) , s → (b′ , s′ ) then b1 else b2 then b1 else b2 X no pattern matches s(r) s(r) = σp, (σb, s) → (b′ , s′ ) (b7 ) (b6 ) W X (read r . . . , s) → (read r . . . , s) (read r with · · · | p ⇒ b | . . . , s) → (b′ , s′ )  (b8 ) e ⇓ v, f (x) = b, X ([v/x]b, s) → (b′ , s′ ) X (f (e), s) → (b′ , s′ ) (b9 ) e ⇓ v, X (b, s[v/r]) → (b′ , s′ ) X (r := e.b, s) → (b′ , s′ ) We denote with be either an expression body or a behaviour. All expressions and behaviours are supposed to be well-typed. As usual, all formal parameters are supposed to be distinct. In the match x with c(y) then be 1 else be 2 instruction, be 1 may depend on y but not on x while be 2 may depend on x but not on y. Systems We suppose that the execution environment consists of n threads and we associate with every thread a distinct identity that is an index in Zn = {0, 1, . . . , n − 1}. We let B, B ′ , . . . denote systems of synchronous threads, that is finite mappings from thread indexes to pairs (behaviour, status). Each register has a type and a default value — its value at the beginning of an instant — and we use s, s′ , . . . to denote a store, an association between registers and their values. We suppose that at the beginning of each instant the store is so , such that each register is assigned its default value. If B is a system and i ∈ Zn is a valid thread index then we denote with B1 (i) the behaviour executed by the thread i and with B2 (i) its current status. Initially, all threads have status R, the current thread index is 0, and B1 (i) is a behaviour expression of the shape f (v) for all i ∈ Zn . System reduction is described by a relation (B, s, i) → (B ′ , s′ , i′ ): the system B with store s and 6 current thread (index) i runs an atomic sequence of actions and becomes (B ′ , s′ , i′ ). X ′ ′ ′ ′ (s1 ) (B1 (i), s) → (b , s ), B2 (i) = R,′ B′ = B[(b , X)/i], (B, s, i) → (B [(B1 (k), R)/k], s′ , k) X (s2 ) (B1 (i), s) → (b′ , s′ ), B2 (i) = R, B ′ = B[(b′ , X)/i], B ′′ = U (B ′ , s′ ), N (B ′′ , so , 0) = k (B, s, i) → (B ′′ , so , k) N (B ′ , s′ , i) = k N (B ′ , s′ , i) ↑, Scheduler The scheduler is determined by the functions N and U. To ensure progress of the scheduling, we assume that if N returns an index then it must be possible to run the corresponding thread in the current instant and that if N is undefined (denoted N (. . . ) ↑) then no thread can be run in the current instant. If N (B, s, i) = k then B2 (k) = R or ( B2 (k) = W and B1 (k) = read r with · · · | p ⇒ b | . . . and some pattern matches s(r) i.e., ∃σ σp = s(r) ) If N (B, s, i) ↑ then ∀k ∈ Zn , B2 (k) ∈ {N, S} or ( B2 (k) = W, B1 (k) = read r with . . . , and no pattern matches s(r) ) When no more thread can run, the instant ends and the function U performs the following status transitions: N → R, W → R. We assume here that every thread in status W takes the [ ] ⇒ . . . branch at the beginning of the next instant. Note that the function N is undefined on the updated system if and only if all threads are stopped.  if B(i) = (b, S)  (b, S) U (B, s)(i) = (b, R) if B(i) = (b, N )  (f (e), R) if B(i) = (read r with · · · | [ ] ⇒ f (e), W ) Example 1 (channels and signals) The read instruction allows to read a register subject to certain filter conditions. This is a powerful mechanism which recalls, e.g., Linda communication [15], and that allows to encode various forms of channel and signal communication. (1) We want to represent a one place channel c carrying values of type t. We introduce a new type ch(t) = empty | full of t and a register c of type Ref (ch(t)) with default value empty. A thread should send a message on c only if c is empty and it should receive a message only if c is not empty (a received message is discarded). These operations can be modelled using the following two derived operators: send (c, e).b =def read c with empty ⇒ c := full(e).b receive(c, x).b =def read c with full(x) ⇒ c := empty.b (2) We want to represent a fifo channel c carrying values of type t such that a thread can always emit a value on c but may receive only if there is at least one message in the 7 channel. We introduce a new type fch(t) = nil | cons of t, fch(t) and a register c of type Ref (fch(t)) with default value nil. Hence a fifo channel is modelled by a register holding a list of values. We consider two read operations — freceive to fetch the first message on the channel and freceiveall to fetch the whole queue of messages — and we use the auxiliary function insert to queue messages at the end of the list: fsend (c, e).b =def read c with l ⇒ c := insert(e, l).b freceive(c, x).b =def read c with cons(x, l) ⇒ c := l.b freceiveall (c, x).b =def read c with cons(y, l) ⇒ c := nil.[cons(y, l)/x]b insert(x, l) = match l with cons(y, l′ ) then cons(y, insert(x, l′ )) else cons(x, nil) (3) We want to represent a signal s with the typical associated primitives: emitting a signal and blocking until a signal is present. We define a type sig = abst | prst and a register s of type Ref (sig) with default value abst, meaning that a signal is originally absent: emit(s).b =def s := prst.b wait (s).b =def read s with prst ⇒ b Example 2 (cooperative fragment) The cooperative fragment of the model with no synchrony is obtained by removing the next instruction and assuming that for all read instructions the branch [ ] ⇒ f (e) is such that f (. . . ) = stop. Then all the interesting computation happens in the first instant; threads still running in the second instant can only stop. By using the representation of fifo channels presented in Example 1(2) above, the cooperative fragment is already powerful enough to simulate, e.g., Kahn networks [20]. Next, to make possible a compositional and functional analysis for resource control, we propose to restrict the admissible behaviours and we define a simple preliminary control flow analysis that guarantees that this restriction is met. We then rely on this analysis to define a symbolic representation of the states reachable by a behaviour. Finally, we extract from this symbolic control points suitable order constraints which are instrumental to our analyses for termination and value size limitation within an instant. 2.1 Read Once Condition We require and statically check on the call graph of the program (see below) that threads can perform any given read instruction at most once in an instant. 1. We assign to every read instruction in a system a distinct fresh label, y, and we collect all these labels in an ordered sequence, y1 , . . . , ym . In the following, we will sometimes use the notation read hyi ̺ with . . . in the code of a behaviour to make visible the label of a read instruction. 2. With every function symbol f defined by an equation f (x) = b we associate the set L(f ) of labels of read instructions occurring in b. 8 3. We define a directed call graph G = (N, E) as follows: N is the set of function symbols in the program defined by an equation f (x) = b and (f, g) ∈ E if g ∈ Call(b) where Call (b) is the collection of function symbols in N that may be called in the current instant and which is formally defined as follows: Call (stop) = Call (next .g(e)) = ∅ Call (f (e)) = {f } Call (yield .b) = Call (̺ := e.b) = Call (b) Call (match x with p then b1 else b2 ) = Call (b1 ) S ∪ Call (b2 ) Call (read ̺ with p1 ⇒ b1 | · · · | pn ⇒ bn | [ ] ⇒ b) = i=1,...,n Call (bi ) We write f E ∗ g if the node g is the node f in the graph G. We denote S reachable from ∗ with R(f ) the set of labels {L(g) | f E g} and with yf the ordered sequence of labels in R(f ). The definition of Call is such that for every sequence of calls in the execution of a thread within an instant we can find a corresponding path in the call graph. Definition 3 (read once condition) A system satisfies the read once condition if in the call graph there are no loops that go through a node f such that L(f ) 6= ∅. Example 4 (alarm) We consider the representation of signals as in Example 1(3). We assume two signals sig and ring. The behaviour alarm(n, m) will emit a signal on ring if it detects that no signal is emitted on sig for m consecutive instants. The alarm delay is reset to n if the signal sig is present. alarm(x, y) = match y with s(y ′ ) then read hui sig with prst ⇒ next.alarm(x, x) | [ ] ⇒ alarm(x, y ′ ) else ring := prst.stop Hence u is the label associated with the read instruction and L(alarm) = {u}. Since the call graph has just one node, alarm, and no edges, the read once condition is satisfied. To summarise, the read once condition is a checkable syntactic condition that safely approximates the semantic property we are aiming at. Proposition 5 If a system satisfies the read once condition then in every instant every thread runs every read instruction at most once (but the same read instruction can be run by several threads). The following simple example shows that without the read once restriction, a thread can use a register as an accumulator and produce an exponential growth of the size of the data within an instant. 9 Example 6 (exponentiation) We recall that nat = z | s of nat is the type of tally natural numbers. The function dble defined below doubles the value of its parameter so that |dble(n)| = 2|n|. We assume r is a register of type nat with initial value s(z). Now consider the following recursive behaviour: dble(n) = match n with s(n′ ) then s(s(dble(n′ ))) else z exp(n) = match n with s(n′ ) then read r with m ⇒ r := dble(m).exp(n′ ) else stop The function exp does not satisfy the read once condition since the call graph has a loop on the exp node. The evaluation of exp(n) involves |n| reads to the register r and, after each read operation, the size of the value stored in r doubles. Hence, at end of the instant, the register contains a value of size 2|n| . The read once condition does not appear to be a severe limitation on the expressiveness of a synchronous programming language. Intuitively, in most synchronous algorithms every thread reads some bounded number of variables before performing some action. Note that while the number of variables is bounded by a constant, the amount of information that can be read in each variable is not. Thus, for instance, a ‘server’ thread can just read one variable in which is stored the list of requests produced so far and then it can go on scanning the list and replying to all the requests within the same instant. 2.2 Control Points From a technical point of view, an important consequence of the read once condition is that a behaviour can be described as a function of its parameters and the registers it may read during an instant. This fact is used to associate with a system satisfying the read once condition a finite number of control points. A control point is a triple (f (p), be, i) where, intuitively, f is the currently called function, p represents the patterns crossed so far in the function definition plus possibly the labels of the read instructions that still have to be executed, be is the continuation, and i is an integer flag in {0, 1, 2} that will be used to associate with the control point various kinds of conditions. If the function f returns a value and is defined by the equation f (x) = eb, then we associate with f the set C(f, x, eb) defined as follows: C(f, p, eb) = case eb of  : {(f (p), eb, 0)} e match x with c(y) : {(f (p), eb, 2)} ∪ C(f, [c(y)/x]p, eb 1 ) ∪ C(f, p, eb 2 ) then eb 1 else eb 2 On the other hand, suppose the function f is a behaviour defined by the equation f (x) = b. Then we generate a fresh function symbol f + whose arity is that of f plus the size of R(f ), thus regarding the labels yf (the ordered sequence of labels in R(f )) as part of the formal 10 parameters of f + . The set of control points associated with f + is the set C(f + , (x · yf ), b) defined as follows: C(f + , p, b) = case b of (C1 ) stop : {(f + (p), b, 2)} (C2 ) g(e) : {(f + (p), b, 0)} ′ (C3 ) yield .b : {(f + (p), b, 2)} ∪ C(f + , p, b′ ) (C4 ) next.g(e) : {(f + (p), b, 2), (f + (p), g(e), 2)} (C5 ) ̺ := e.b′ : {(f + (p), b, 2), (f + (p), e, 1)} ∪ C(f +, p, b′ )   match x with c(y) {(f + (p), b, 2)} ∪ C(f + , ([c(y)/x]p), b1 ) (C6 ) : then b1 else b2 ∪ C(f + , p, b2 )   {(f + (p), b, 2), (f + (p), g(e), 2)} read hyi ̺ with p1 ⇒ b1 | · · · | : ∪ C(f + , ([p1 /y]p), b1 ) ∪ . . . (C7 ) pn ⇒ bn | [ ] ⇒ g(e) ∪ C(f +, ([pn /y]p), bn ) By inspecting the definitions, we can check that a control point (f (p), be, i) has the property that Var(be) ⊆ Var (p). Definition 7 An instance of a control point (f (p), be, i) is an expression body or a behaviour be ′ = σ(be), where σ is a substitution mapping the free variables in be to values. The property of being an instance of a control point is preserved by expression body evaluation, behaviour reduction and system reduction. Thus the control points associated with a system do provide a representation of all reachable configurations. Indeed, in Appendix B we show that it is possible to define the evaluation and the reduction on pairs of control points and substitutions. Proposition 8 Suppose (B, s, i) → (B ′ , s′ , i′ ) and that for all thread indexes j ∈ Zn , B1 (j) is an instance of a control point. Then for all j ∈ Zn , we have that B1′ (j) is an instance of a control point. In order to prove the termination of the instant and to obtain a bound on the size of computed value, we associate order constraints with control points: Control point (f (p), e, 0) (f + (p), g(e), 0) (f + (p), e, 1) (f + (p), be, 2) Associated constraint f (p) ≻0 e f + (p) ≻0 g+ (e, yg ) f + (p) ≻1 e no constraints A program will be deemed correct if the set of constraints obtained from all the function definitions can be satisfied in suitable structures. We say that a constraint e ≻i e′ has index i. We rely on the constraints of index 0 to enforce termination of the instant and on those of index 0 or 1 to enforce a bound on the size of the computed values. Note that the constraints are on pure first order terms, a property that allows us to reuse techniques developed in the standard term rewriting framework (cf. Section 3). 11 Example 9 With reference to Example 4, we obtain the following control points: (alarm + (x, y, u), match . . . , 2) (alarm + (x, y, u), prst, 1) (alarm + (x, s(y ′ ), u), read . . . , 2) (alarm + (x, s(y ′ ), prst), next.alarm(x, x), 2) (alarm + (x, y, u), ring := prst.stop, 2) (alarm + (x, z, u), stop, 2) (alarm + (x, s(y ′ ), u), alarm (x, y ′ ), 2) (alarm + (x, s(y ′ ), prst), alarm(x, x), 2) The triple (alarm + (x, y, u), prst, 1) is the only control point with a flag different from 2. It corresponds to the constraint alarm + (x, y, u) ≻1 prst, where u is the label associated with the only read instruction in the body of alarm. We note that no constraints of index 0 are generated and so, in this simple case, the control flow analysis can already establish the termination of the thread and all is left to do is to check that the size of the data is under control, which is also easily verified. In Example 2, we have discussed a possible representation of Kahn networks in the cooperative fragment of our model. In general Kahn networks there is no bound on the number of messages that can be written in a fifo channel nor on the size of the messages. Much effort has been put into the static scheduling of Kahn networks (see, e.g., [22, 16, 17]). This analysis can be regarded as a form of resource control since it guarantees that the number of messages in fifo channels is bounded (but says nothing about their size). The static scheduling of Kahn network is also motivated by performance issues, since it eliminates the need to schedule threads at run time. Let us look in some detail at the programming language Lustre, that can be regarded as a language for programming Kahn networks that can be executed synchronously. Example 10 (read once vs. Lustre) A Lustre network is composed of four types of nodes: the combinatorial node, the delay node, the when node, and the merge node. Each node may have several input streams and one output stream. The functional behaviour of each type of node is defined by a set of recursive definitions. For instance, the node When has one boolean input stream b — with values of type bool = false | true — and one input stream s of values. A When node is used to output values from s whenever b is true. This behaviour may be described by the following recursive definitions: When(false · b, x · s) = When(b, s), When(true · b, x · s) = x · When(b, s), and When(b, s) = ǫ otherwise. Here is a possible representation of the When node in our model, where the input streams correspond to one place channels b, c (cf. Example 1(1)), the output stream to a one place channel c′ and at most one element in each input stream is processed per instant. When() = read hui b with full(true) ⇒ read hvi c with full(x) ⇒ c′ := x.next.When() | [ ] ⇒ When() | full(false) ⇒ next.When() | [ ] ⇒ When() While the function When has no formal parameters, we consider the function When + with two parameters u and v in our size and termination analyses. 12 3 Resource Control Our analysis goes in three main steps: first, we guarantee that each instant terminates (Section 3.1), second we bound the size of the computed values as a function of the size of the parameters at the beginning of the instant (Section 3.2), and third we combine the termination and size analyses to obtain polynomial bounds on space and time (Section 3.3). As we progress in our analysis, we refine the techniques we employ. Termination is reduced to the general problem of finding a suitable well-founded order over first-order terms. Bounding the size of the computed values is reduced to the problem of synthesizing a quasi-interpretation. Finally, the problem of obtaining polynomial bounds is attacked by combining recursive path ordering termination arguments with quasi-interpretations. We selected these techniques because they are well established and they can handle a significant spectrum of the programs we are interested in. It is to be expected that other characterisations of complexity classes available in the literature may lead to similar results. 3.1 Termination of the Instant We recall that a reduction order > over first-order terms is a well-founded order that is closed under context and substitution: t > s implies C[t] > C[s] and σt > σs, where C is any one hole context and σ is any substitution (see, e.g, [6]). Definition 11 (termination condition) We say that a system satisfies the termination condition if there is a reduction order > such that all constraints of index 0 associated with the system hold in the reduction order. In this section, we assume that the system satisfies the termination condition. As expected this entails that the evaluation of closed expressions succeeds. Proposition 12 Let e be a closed expression. Then there is a value v such that e ⇓ v and e ≥ v with respect to the reduction order. Moreover, the following proposition states that a behaviour will always return the control to the scheduler. Proposition 13 (progress) Let b be an instance of a control point. Then for all stores X s, there exist X, b′ and s′ such that (b, s) → (b′ , s′ ). Finally, we can guarantee that at each instant the system will reach a configuration in which the scheduler detects the end of the instant and proceeds to the reinitialisation of the store and the status (as specified by rule (s2 )). Theorem 14 (termination of the instant) All sequences of system reductions involving only rule (s1 ) are finite. 13 Proposition 13 and Theorem 14 are proven by exhibiting a suitable well-founded measure which is based both on the reduction order and the fact that the number of reads a thread may perform in an instant is finite. Example 15 (monitor max value) We consider a recursive behaviour monitoring the register i (acting as a fifo channel) and parameterised on a number x representing the largest value read so far. At each instant, the behaviour reads the list l of natural numbers received on i and assigns to o the greatest number in x and l. f (x) = yield .read hii i with l ⇒ f1 (maxl (l, x)) f1 (x) = o := x.next.f (x) max (x, y) = match x with s(x′ ) then match y with s(y ′ ) then s(max (x′ , y ′ )) else s(x′ ) else y maxl (l, x) = match l with cons(y, l′ ) then maxl (l′ , max (x, y)) else x It is easy to prove the termination of the thread by recursive path ordering, where the function symbols are ordered as f + > f1+ > maxl > max , the arguments of maxl are compared lexicographically from left to right, and the constructor symbols are incomparable and smaller than any function symbol. 3.2 Quasi-interpretations Our next task is to control the size of the values computed by the threads. To this end, we propose a suitable notion of quasi-interpretation (cf. [10, 3, 4]). Definition 16 (assignment) Given a program, an assignment q associates with constructors and function symbols, functions over the non-negative reals R+ such that: (1) If c is a constant then qc is the constant 0. (2) If c is a constructor with arity n ≥ 1 then qc is a function in (R+ )n → R+ such that qc (x1 , . . . , xn ) = d + Σi∈1..n xi , for some d ≥ 1. (3) If f is a function (name) with arity n then qf : (R+ )n → R+ is monotonic and for all i ∈ 1..n we have qf (x1 , . . . , xn ) ≥ xi . An assignment q is extended to all expressions e as follows, giving a function expression qe with variables in Var (e): qx = x , qc(e1 ,...,en) = qc (qe1 , . . . , qen ) , qf (e1 ,...,en) = qf (qe1 , . . . , qen ) . Here qx is the identity function and, e.g., qc (qe1 , . . . , qen ) is the functional composition of the function qc with the functions qe1 , . . . , qen . It is easy to check that there exists a constant δq depending on the assignment q such that for all values v we have |v| ≤ qv ≤ δq · |v|. Thus the quasi-interpretation of a value is always proportional to its size. 14 Definition 17 (quasi-interpretation) An assignment is a quasi-interpretation, if for all constraints associated with the system of the shape f (p) ≻i e, with i ∈ {0, 1}, the inequality qf (p) ≥ qe holds over the non-negative reals. Quasi-interpretations are designed so as to provide a bound on the size of the computed values as a function of the size of the input data. In the following, we assume given a suitable quasi-interpretation, q, for the system under investigation. Example 18 With reference to Examples 6 and 15, the following assignment is a quasiinterpretation (the parameter i corresponds to the label of the read instruction in the body of f ). We give no quasi-interpretations for the function exp because it fails the read once condition: qnil = qz = 0 , qs (x) = x + 1 , qcons (x, l) = x + l + 1 , qdble (x) = 2 · x , qf + (x, i) = x + i , qf + (x) = x , qmaxl (x, y) = qmax (x, y) = max (x, y) . 1 One can show [3, 4] that in the purely functional fragment of our language every value v computed during the evaluation of an expression f (v1 , . . . , vn ) satisfies the following condition: |v| ≤ qv ≤ qf (v1 ,...,vn ) = qf (qv1 , . . . , qvn ) ≤ qf (δq · |v1 |, . . . , δq · |vn |) . (1) We generalise this result to threads as follows. Theorem 19 (bound on the size of the values) Given a system of synchronous threads B, suppose that at the beginning of the instant B1 (i) = f (v) for some thread index i. Then the size of the values computed by the thread i during an instant is bounded by qf + (v,u) where u are the values contained in the registers at the time they are read by the thread (or some constant value, if they are not read at all). Theorem 19 is proven by showing that quasi-interpretations satisfy a suitable invariant. In the following corollary, we note that it is possible to express a bound on the size of the computed values which depends only on the size of the parameters at the beginning of the instant. This is possible because the number of reads a system may perform in an instant is bounded by a constant. Corollary 20 Let B be a system with m distinct read instructions and n threads. Suppose B1 (i) = fi (vi ) for i ∈ Zn . Let c be a bound of the size of the largest parameter of the functions fi and the largest default value of the registers. Suppose h is a function bounding all the quasi-interpretations, that is, for all the functions fi+ we have h(x) ≥ qf + (x, . . . , x) i over the non-negative reals. Then the size of the values computed by the system B during an instant is bounded by hn·m+1 (c). 15 Example 21 The n·m iterations of the function h predicted by Corollary 20 correspond to a tight bound, as shown by the following example. We assume n threads and one register, r, of type nat with default value z. The control of each thread is described as follows: f (x0 ) = read r with x1 ⇒ r := dble(max (x1 , x0 )). read r with x2 ⇒ r := dble(x2 ). ...... read r with xm ⇒ r := dble(xm ).next .f (dble(xm )) . For this system we have c ≥ |x0 | and h(x) = qdble (x) = 2 · x. It is easy to show that, at the end of an instant, there have been n · m assignments to the register r (m for every thread in the system) and that the value stored in r is dble n·m (x0 ) of size 2n·m · |x0 |. 3.3 Combining Termination and Quasi-interpretations To bound the space needed for the execution of a system during an instant we also need to bound the number of nested recursive calls, i.e. the number of frames that can be found on the stack (a precise definition of frame is given in the following Section 4). Unfortunately, quasi-interpretations provide a bound on the size of the frames but not on their number (at least not in a direct implementation that does not rely on memoization). One way to cope with this problem is to combine quasi-interpretations with various families of reduction orders [24, 10]. In the following, we provide an example of this approach based on recursive path orders which is a widely used and fully mechanizable technique to prove termination [6]. Definition 22 We say that a system terminates by LPO, if the reduction order associated with the system is a recursive path order where: (1) symbols are ordered so that function symbols are always bigger than constructor symbols and two distinct constructor symbols are incomparable; (2) the arguments of function symbols are compared with respect to the lexicographic order and those of constructor symbols with respect to the product order. Note that because of the hypotheses on constructors, this is actually a special case of the lexicographic path order. For the sake of brevity, we still refer to it as LPO. Definition 23 We say that a system admits a polynomial quasi-interpretation if it has a quasi-interpretation where all functions are bounded by a polynomial. The following property is a central result of this paper. Theorem 24 If a system B terminates by LPO and admits a polynomial quasi-interpretation then the computation of the system in an instant runs in space polynomial in the size of the parameters of the threads at the beginning of the instant. The proof of Theorem 24 is based on Corollary 20 that provides a polynomial bound on the size of the computed values and on an analysis of nested calls in the LPO order that can be found in [10]. The point is that the depth of such nested calls is polynomial in the size of the values and that this allows to effectively compute a polynomial bounding the space necessary for the execution of the system. 16 Example 25 We can check that the order used in Example 15 for the functions f + , f1+ , max and maxl is indeed a LPO. Moreover, from the quasi-interpretation given in Example 18, we can deduce that the function h(x) has the shape a · x + b (it is affine). In practice, many useful functions admit quasi-interpretations bound by an affine function such as the max-plus polynomials considered in [3, 4]. The combination of LPO and polynomial quasi-interpretation actually provides a characterisation of PSPACE. In order to get to PTIME a further restriction has to be imposed. Among several possibilities, we select one proposed in [11]. We say that the system terminates by linear LPO if it terminates by LPO as in definition 22 and moreover if in all the constraints f (p) ≻0 e or f + (p) ≻0 g + (e) of index 0 there is at most one function symbol on the right hand side which has the same priority as the (unique) function symbol on the left-hand side. For instance, the Example 15 falls in this case. In op. cit., it is shown by a simple counting argument that the number of calls a function may generate is polynomial in the size of its arguments. One can then restate theorem 24 by replacing LPO with linear LPO and PSPACE with PTIME. We stress that these results are of a constructive nature, thus beyond proving that a system ‘runs in PSPACE (or PTIME)’, we can extract a definite polynomial that bounds the size needed to run a system during an instant. In general, the bounds are rather rough and should be regarded as providing a qualitative rather than quantitative information. In the purely functional framework, M. Hofmann [19] has explored the situation where a program is non-size increasing which means that the size of all intermediate results is bounded by the size of the input. Transferring this concept to a system of threads is attractive because it would allow to predict the behaviour of the system for arbitrarily many instants. However, this is problematic. For instance, consider again example 25. By Theorem 24, we can prove that the computation of a system running the behaviour f (x0 ) in an instant requires a space polynomial in the size of x0 . Note that the parameter of f is the largest value received so far in the register i. Clearly, bounding the value of this parameter for arbitrarily many instants requires a global analysis of the system which goes against our wish to produce a compositional analysis in the sense explained in the Introduction. An alternative approach which remains to be explored could be to develop linguistic tools and a programming discipline that allow each thread to control locally the size of its parameters. 4 A Virtual Machine We describe a simple virtual machine for our language thus providing a concrete intuition for the data structures required for the execution of the programs and the scheduler. Our motivations for introducing a low-level model of execution for synchronous threads are twofold: (i) it offers a simple formal definition for the space needed for the execution of an instant (just take the maximal size of a machine configuration), and (ii) it explains some of the elaborate mechanisms occurring during the execution, like the synchronisation with 17 the read instruction and the detection of the end of an instant. A further motivation which is elaborated in Section 4.5 is the possibility to carry on the static analyses for resource control at bytecode level. The interest of bytecode verification is now well understood, and we refer the reader to [25, 26]. 4.1 Data Structures We suppose given the code for all the threads running in a system together with a set of types and constructor names and a disjoint set of function names. A function name f will also denote the sequence of instructions of the associated code: f [i] stands for the ith instruction in the (compiled) code of f and |f | stands for the number of instructions. The configuration of the machine is composed of a store s, that maps registers to their current values, a sequence of records describing the state of each thread in the system, and three local registers owned by the scheduler whose role will become clear in Section 4.3. A thread identifier, t, is simply an index in Zn . The state of a thread t is a pair (st t , Mt ) where st t is a status and Mt is the memory of the thread. A memory M is a sequence of frames, and a frame is a triple (f, pc, ℓ) composed of a function name, the value of the program counter (a natural number in 1..|f |), and a stack of values ℓ = v1 · · · vk . We denote with |ℓ| the number of values in the stack. The status of a thread is defined as in the source language, except for the status W which is refined into W (j, n) where: j is the index where to jump at the next instant if the thread does not resume in the current instant, and n is the (logical) time at which the thread is suspended (cf. Section 4.3). 4.2 Instructions The set of instructions of the virtual machine together with their operational meaning is described in Table 1. All instructions operate on the frame of the current thread t and the memory Mt — the only instructions that depend on or affect the store are read and write. For every segment of bytecode, we require that the last instruction is either return, stop or tcall and that the jump index j in the instructions branch c j and wait j is within the segment. 4.3 Scheduler In Table 2 we describe a simple implementation of the scheduler. The scheduler owns three registers: (1) tid that stores the identity of the current thread, (2) time for the current time, and (3) wtime for the last time the store was modified. The notion of time here is of a logical nature: time passes whenever the scheduler transfers control to a new thread. Like in the source language, so denotes the store at the beginning of each instant. The scheduler triggers the execution of the current instruction of the current thread, whose index is stored in tid, with a call to run(tid). The call returns the label X associated with the instruction in Table 1. By convention, take X = ǫ when no label is displayed. If X 6= ǫ then the scheduler must take some action. Assume tid stores the thread index t. We denote 18 Table 1: Bytecode instructions f [pc] load k branch c j branch c j build c n call g n tcall g n return read r read k write r write k Current memory Following memory ′ M · (f, pc, ℓ · v · ℓ ) → M · (f, pc + 1, ℓ · v · ℓ′ · v), |ℓ| = k − 1 M · (f, pc, ℓ · c(v1 , . . . , vn )) → M · (f, pc + 1, ℓ · v1 · · · vn ) M · (f, pc, ℓ · d(. . .)) → M · (f, j, ℓ · d(. . .)) c 6= d M · (f, pc, ℓ · v1 · · · vn ) → M · (f, pc + 1, ℓ · c(v1 , . . . , vn )) M · (f, pc, ℓ · v1 · · · vn ) → M · (f, pc, ℓ · v1 · · · vn ) · (g, 1, v1 · · · vn ) M · (f, pc, ℓ · v1 · · · vn ) → M · (g, 1, v1 · · · vn ) M · (g, pc ′ , ℓ′ · v′ ) · (f, pc, ℓ · v) → M · (g, pc ′ + 1, ℓ′ · v), ar (f ) = |v′ | (M · (f, pc, ℓ), s) → (M · (f, pc + 1, ℓ · s(r)), s) (M · (f, pc, ℓ · r · ℓ′ ), s) → (M · (f, pc + 1, ℓ · r · ℓ′ · s(r)), s), |ℓ| = k − 1 (M · (f, pc, ℓ · v), s) → (M · (f, pc + 1, ℓ), s[v/r]) (M · (f, pc, ℓ · r · ℓ′ · v), s) → (M · (f, pc + 1, ℓ · r · ℓ′ ), s[v/r]), |ℓ| = k − 1 stop M · (f, pc, ℓ) → ǫ yield M · (f, pc, ℓ) → M · (f, pc + 1, ℓ) next M · (f, pc, ℓ) → M · (f, pc + 1, ℓ) wait j M · (f, pc, ℓ · v) → M · (f, j, ℓ) S R N W pc tid the program counter of the top frame (f, pc t , ℓ) in Mt , if any, Itid the instruction f [pc t ] (the current instruction in the thread) and st tid the state st t of the thread. Let us explain the role of the status W (j, n) and of the registers time and wtime. We assume that a thread waiting for a condition to hold can check the condition without modifying the store. Then a thread waiting since time m may pass the condition only if the store has been modified at a time n with m < n. Otherwise, there is no point in passing the control to it1 . With this data structure we also have a simple method to detect the end of an instant, it arises when no thread is in the running status and all waiting threads were interrupted after the last store modification occurred. In models based on preemptive threads, it is difficult to foresee the behaviour of the scheduler which might depend on timing information not available in the model. For this reason and in spite of the fact that most schedulers are deterministic, the scheduler is often modelled as a non-deterministic process. In cooperative threads, as illustrated here, the interrupt points are explicit in the program and it is possible to think of the scheduler as a deterministic process. Then the resulting model is deterministic and this fact considerably simplifies its programming, debugging, and analysis. 1 Of course, this condition can be refined by recording the register on which the thread is waiting, the shape of the expected value,. . . 19 Table 2: An implementation of the scheduler for t in Zn do { st t := R; } s := so ; tid := time := wtime := 0; while (tid ∈ Zn ) { if Itid = (write ) then wtime := time; if Itid = (wait j ) then st tid := W (pc tid + 1, time); X := run(tid); if X 6= ǫ then { if X 6= W then st tid := X; tid := N (tid, st ); if tid ∈ Zn then { st tid := R; time := time + 1; } else { s := so ; wtime := time; tid := N (0, st ); forall i in Zn do { if st i = W (j, ) then pc i := j; if st i 6= S then st i := R; } } } If N (tid, st) = k ∈ Zn If N (tid, st) ∈ / Zn (initialisation) (the initial thread is of index 0) (loop until all threads are blocked) (record store modified) (save continuation for next instant) (run current thread) (update thread status) (compute index of next active thread) (test whether all threads are blocked) (if not, prepare next thread to run) (else, initialisation of the new instant) (select thread to run, starting from 0) Conditions on N : then st k = R or (st k = W (j, n) and n < wtime) then ∀k ∈ Zn (st k 6= R and (st k = W (j, n) implies n ≥ wtime)) 20 Table 3: Compilation of source code to bytecode Compilation of expression bodies: =  C ′ (e, η) · return (branch c j) · C(eb 1 , η ′ · y) · if η = η ′ · x      match x with c(y) (j : C(eb 2 , η)) ,η = C then eb 1 else eb 2 (load i(x, η)) · (branch c j) · o.w.    C(eb 1 , η · y) · (j : C(eb 2 , η · x)) C(e, η) Auxiliary compilation of expressions: C ′ (x, η) = (load i(x, η)) C ′ (c(e1 , . . . , en ), η) = C ′ (e1 , η) · . . . · C ′ (en , η) · (build c n) C ′ (f (e1 , . . . , en ), η) = C ′ (e1 , η) · . . . · C ′ (en , η) · (call f n) Compilation of behaviours: C(stop, η) C(f (e1 , . . . , en ), η) C(yield .b, η) C(next.f (e), η) C(̺ := e.b, η) = = = = = stop C ′ (e1 , η) · · · C ′ (en , η) · (tcall f n) yield · C(b, η) next · C(f (e), η) ′ C  (e, η) · (write i(̺, η))′ · C(b, η) (branch c j) · C(b1 , η · y) · if η = η ′ · x      match x with c(y) (j : C(b2 , η)) C ,η = then b1 else b2 (load i(x, η)) · (branch c j) · o.w.    C(b1 , η · y) · (j : C(b2 , η · x))    j0 : (read i(̺, η)) · . . . · read ̺ with · · · | cℓ (yℓ ) ⇒ bℓ | C ,η = jℓ : (branch cℓ jℓ+1 ) · C(bℓ , η · yℓ )·  · · · y k ⇒ bk · · · j : · · · jk : C(bk , η · yk )  ℓ+1    j0 : (read i(̺, η)) · . . . · read ̺ with · · · | cℓ (yℓ ) ⇒ bℓ | ,η = jℓ : (branch cℓ jℓ+1 ) · C(bℓ , η · yℓ )·  C · · · | [ ] ⇒ g(e) jℓ+1 : · · · jn : (wait j0 ) · C(g(e), η) 21 4.4 Compilation In Table 3, we describe a possible compilation of the intermediate language into bytecode. We denote with η a sequence of variables. If x is a variable and η a sequence then i(x, η) is the index of the rightmost occurrence of x in η. For instance, i(x, x · y · x) = 3. By convention, i(r, η) = r if r is a register constant. We also use the notation j : C(be, η) to indicate that j is the position of the first instruction of C(be, η). This is just a convenient notation since, in practice, the position can be computed explicitly. With every function definition f (x1 , . . . , xn ) = be we associate the bytecode C(be, x1 · · · xn ). Example 26 (compiled code) We show below the result of the compilation of the function alarm in Example 4: 1 2 3 4 5 4.5 : : : : : branch s 12 read sig branch prst 8 next load 1 6 7 8 9 10 : : : : : load 1 tcall alarm 2 wait 2 load 1 load 2 11 12 13 14 : : : : tcall alarm 2 build prst 0 write ring stop Control Flow Analysis Revisited As a first step towards control flow analysis, we analyse the flow graph of the bytecode generated. Definition 27 (flow graph) The flow graph of a system is a directed graph whose nodes are pairs (f, i) where f is a function name in the program and i is an instruction index, 1 ≤ i ≤ |f |, and whose edges are classified as follows: Successor: An edge ((f, i), (f, i+ 1)) if f [i] is a load, branch, build, call, read, write, or yield instruction. Branch: An edge ((f, i), (f, j)) if f [i] = branch c j. Wait: An edge ((f, i), (f, j)) if f [i] = wait j. Next: An edge ((f, i), (f, i + 1)) if f [i] is a wait or next instruction. Call: An edge ((f, i), (g, 1)) if f [i] = call g n or f [i] = tcall g n. The following is easily checked by inspecting the compilation function. Properties Tree and Read-Wait entail that the only cycles in the flow graph of a function correspond to the compilation of a read instruction. Property Next follows from the fact that, in a behaviour, an instruction next is always followed by a function call f (e). Property ReadOnce is a transposition of the read once condition (Section 2.1) at the level of the bytecode. 22 Proposition 28 The flow graph associated with the compilation of a well-formed system satisfies the following properties: Tree: Let G′ be the flow graph without wait and call edges. Let G′f be the full subgraph of G′ whose nodes have the shape (f, i). Then G′f is a tree with root (f, 1). Read-Wait: If f [i] = wait j then f [j] = read r and there is a unique path from (f, j) to (f, i) and in this path, every node corresponds to a branch instruction. Next: Let G′ be the flow graph without call edges. If ((f, i), (f, i + 1)) is a next edge then for all nodes (f, j) accessible from (f, i + 1), f [j] is not a read instruction. Read-Once: Let G′ be the flow graph without wait edges and next edges. If the source code satisfies the read once condition then there is no loop in G′ that goes through a node (f, i) such that f [i] is a read instruction. In [1], we have presented a method to perform resource control verifications at bytecode level. This work is just concerned with the functional fragment of our model. Here, we outline its generalisation to the full model. The main problem is to reconstruct a symbolic representation of the values allocated on the stack. Once this is done, it is rather straightforward to formulate the constraints for the resource control. We give first an informal description of the method. 1. For every segment f of bytecode instructions with, say, formal parameters x1 , . . . , xn and for every instruction i in the segment, we compute a sequence of expressions e1 · · · em and a substitution σ. 2. The expressions (ei )i∈1..m are related to the formal parameters via the substitution σ. More precisely, the variables in the expressions are contained in σx1 , . . . , σxn and the latter forms a linear pattern. 3. Next, let us look at the intended usage of the formal expressions. Suppose at run time the function f is called with actual parameters u1 , . . . , un and suppose that following this call, the control reaches instruction i with a stack ℓ. Then we would like that: • The values u1 , . . . , un match the pattern σx1 , . . . , σxn via some substitution ρ. • The stack ℓ contains exactly m values v1 , . . . , vm whose types are the ones of e1 , . . . , em , respectively. • Moreover ρ(ei ) is an over-approximation (w.r.t. size and/or termination) of the value vi , for i = 1, . . . , m. In particular, if ei is a pattern, we want that ρ(ei ) = vi . We now describe precisely the generation of the expressions and the substitutions. This computation is called shape analysis in [1]. For every function f and index i such that f [i] is a read instruction we assume a fresh variable xf,i . Given a total order on the function 23 symbols, such variables can be totally ordered with respect to the index (f, i). Moreover, for every index i in the code of f , we assume a countable set xi,j of distinct variables. We assume that the bytecode comes with annotations assigning a suitable type to every constructor, register, and function symbol. With every function symbol f of type t → beh, comes a fresh function symbol f + of type t, t′ → beh so that |t′| is the number of read instructions accessible from f within an instant. Then, as in the definition of control points (Section 2.2), the extra arguments in f + corresponds to the values read in the registers within an instant. The order is chosen according to the order of the variables associated with the read instructions. In the shape analysis, we will consider well-typed expressions obtained by composition of such fresh variables with function symbols, constructors, and registers. In order to make explicit the type of a variable x we will write xt . For every function f , the shape analysis computes a vector σ = σ1 , . . . , σ|f | of substitutions and a vector E = E1 , . . . , E|f | of sequences of well-typed expressions. We let Ei and σi denote the sequence Ei and the substitution σi respectively (the ith element in the vector), and Ei [k] the k th element in Ei . We also let hi = |Ei | be the length of the ith sequence. 1 n We assume σ1 = id and E1 = xt1,1 · · · xt1,n , if f : t1 , . . . , tn → β is a function of arity n. The main case is the branch instruction: f [i] = branch c j Conditions c : t → t, Ei = E · e, e : t, and either e = c(e), σi+1 = σi , Ei+1 = E · e or e = d(e), c 6= d, σj = σi , Ej = Ei n 1 )/x], or e = xt , σj = σi , Ej = Ei , σ ′ = [c(xti+1,h , . . . , xti+1,h i+1 i ′ ′ σi+1 = σ ◦ σi , Ei+1 = σ (E) · xi+1,hi · · · xi+1,hi+1 . The constraints for the remaining instructions are given in Table 4, where it is assumed that σi+1 = σi except for the instructions tcall and return (that have no direct successors in the code of the function). Example 29 We give the shape of the values on the stack (a side result of the shape analysis) for the bytecode obtained from the compilation of the function f defined in Example 15: Instruction 1 : yield 2 : read i 3 : load 1 Shape x x x·l Instruction 4 : call maxl 2 5 : call f1 1 6 : return Shape x·l·x x · maxl (l, x) x · f1 (maxl (l, x)) Note that the code has no branch instruction, hence the substitution σ is always the identity. Once the shapes are generated it is rather straightforward to determine a set of constraints that entails the termination of the code and a bound on the size of the computed values. For instance, assuming the reduction order is a simplification order, it is enough to require that f + (x, l) > f1 (maxl (l, x)), i.e. the shape of the returned value, f1 (maxl(l, x)), is less than the shape of the call, f + (x, l). 24 Table 4: Shape analysis at bytecode level f [i] = load k build c n call g n tcall g n return read r read k write r write k yield next wait j Conditions k ∈ 1..hi , Ei+1 = Ei · Ei [k] c : t → t, Ei = E · e, |e| = n, e : t, Ei+1 = E · c(e) g : t → t, Ei = E · e, |e| = n, e : t, Ei+1 = E · g(e) g : t → β, Ei = E · e, |e| = n, e : t f : t → t, Ei = E · e, e : t r : Ref (t), Ei+1 = Ei · xtf,i k ∈ 1..hi , Ei [k] : Ref (t), Ei+1 = Ei · xtf,i r : Ref (t), Ei = E · e, e : t, Ei+1 = E k ∈ 1..hi , Ei [k] : Ref (t), Ei = E · e, e : t, Ei+1 = E Ei+1 = Ei Ei+1 = Ei Ei = Ej · xtf,j , Ei+1 = Ej , σi = σj If one can find a reduction order and an assignment satisfying the constraints generated from the shape analysis then one can show the termination of the instant and provide bounds on the size of the computed values. We refrain from developing this part which is essentially an adaptation of Section 3 at bytecode level. Moreover, a detailed treatment of the functional fragment is available in [1]. Instead, we state that the shape analysis is always successful on the bytecode generated by the compilation function described in Table 3 (see Appendix B.8). This should suggest that the control flow analysis is not overly constraining though it can certainly be enriched in order to take into account some code optimisations. Theorem 30 The shape analysis succeeds on the compilation of a well-formed program. 5 Conclusion The execution of a thread in a cooperative synchronous model can be regarded as a sequence of instants. One can make each instant simple enough so that it can be described as a function — our experiments with writing sample programs show that the restrictions we impose do not hinder the expressivity of the language. Then well-known static analyses used to bound the resources needed for the execution of first-order functional programs can be extended to handle systems of synchronous cooperative threads. We believe this provides some evidence for the relevance of these techniques in concurrent/embedded programming. We also expect that our approach can be extended to a richer programming model including more complicated control structures. The static analyses we have considered do not try to analyse the whole system. On the contrary, they focus on each thread separately and can be carried out incrementally. Moreover, it is quite possible to perform them at bytecode level. These characteristics are 25 particularly interesting in the framework of ‘mobile code’ where threads can enter or leave the system at the end of each instant as described in [12]. Acknowledgements and Publication History We would like to thank the referees for their valuable comments. Thanks to G. Boudol and F. Dabrowski for comments and discussions on a preliminary version of this article that was presented at the 2004 International Conference on Concurrency Theory. In the present paper, we consider a more general model which includes references as first class values and requires a reformulation of the control flow analysis. Moreover, we present a new virtual machine, a number of examples, and complete proofs not available in the conference paper. References [1] R. Amadio, S. Coupet-Grimal, S. Dal-Zilio, and L. Jakubiec. A functional scenario for bytecode verification of resource bounds. In Proceedings of CSL – International Conference on Computer Science Logic, Lecture Notes in Computer Science 3210, Springer, 2004. [2] R. Amadio, S. Dal-Zilio. Resource control for synchronous cooperative threads. In Proceedings CONCUR – 15th International Conference on Concurrency Theory, Lecture Notes in Computer Science 3170, Springer, 2004. [3] R. Amadio. Max-plus quasi-interpretations. In Proceedings of TLCA – 6th International Conference on Typed Lambda Calculi and Applications, Lecture Notes in Computer Science 2701, Springer, 2003. [4] R. Amadio. Synthesis of max-plus quasi-interpretations. In Fundamenta Informaticae, 65(1-2):29-60, 2005. [5] J. Armstrong, R. Virding, C. Wikström, M. Williams. Concurrent Programming in Erlang. Prentice-Hall 1996. [6] F. Baader and T. Nipkow. Term rewriting and all that. Cambridge University Press, 1998. [7] P. Baillot and V. Mogbil, Soft lambda calculus: a language for polynomial time computation. In Proceedings of FOSSACS – 7th International Conference on Foundations of Software Science and Computation Structures, Lecture Notes in Computer Science 2987, Springer, 2004. [8] S. Bellantoni and S. Cook. A new recursion-theoretic characterization of the poly-time functions. Computational Complexity, 2:97–110, 1992. [9] G. Berry and G. Gonthier, The Esterel synchronous programming language. Science of computer programming, 19(2):87–152, 1992. 26 [10] G. Bonfante, J.-Y. Marion, and J.-Y. Moyen. On termination methods with space bound certifications. In Proceedings Perspectives of System Informatics, Lecture Notes in Computer Science 2244, Springer, 2001. [11] G. Bonfante, J.-Y. Marion, J.-Y. Moyen. Quasi-interpretations. Internal report LORIA, November 2004, available from the authors. [12] G. Boudol, ULM, a core programming model for global computing. In Proceedings of ESOP – 13th European Symposium on Programming, Lecture Notes in Computer Science 2986, Springer, 2004. [13] F. Boussinot and R. De Simone, The SL Synchronous Language. IEEE Trans. on Software Engineering, 22(4):256–266, 1996. [14] J. Buck. Scheduling dynamic dataflow graphs with bounded memory using the token flow model. PhD thesis, University of California, Berkeley, 1993. [15] N. Carriero and D. Gelernter. Linda in Context. Communication of the ACM, 32(4): 444-458, 1989. [16] P. Caspi. Clocks in data flow languages. Theoretical Computer Science, 94:125–140, 1992. [17] P. Caspi and M. Pouzet. Synchronous Kahn networks. In Proceedings of ICFP – ACM SIGPLAN International Conference on Functional Programming, SIGPLAN Notices 31(6), ACM Press, 1996. [18] A. Cobham. The intrinsic computational difficulty of functions. In Proceedings Logic, Methodology, and Philosophy of Science II, North Holland, 1965. [19] M. Hofmann. The strength of non size-increasing computation. In Proceedings of POPL – 29th SIGPLAN-SIGACT Symposium on Principles of Programming Languages, ACM, 2002. [20] G. Kahn. The semantics of a simple language for parallel programming. In Proceedings IFIP Congress, North-Holland, 1974. [21] N. Jones. Computability and complexity, from a programming perspective. MIT-Press, 1997. [22] E. Lee and D. Messerschmitt. Static scheduling of synchronous data flow programs for digital signal processing. IEEE Transactions on Computers, 1:24–35, 1987. [23] D. Leivant. Predicative recurrence and computational complexity i: word recurrence and poly-time. Feasible mathematics II, Clote and Remmel (eds.), Birkhäuser:320– 343, 1994. 27 [24] J.-Y. Marion. Complexité implicite des calculs, de la théorie à la pratique. Université de Nancy. Habilitation à diriger des recherches, 2000. [25] G. Morriset, D. Walker, K. Crary and N. Glew. From system F to typed assembly language. In ACM Transactions on Programming Languages and Systems, 21(3):528569, 1999. [26] G. Necula. Proof carrying code. In Proceedings of POPL – 24th SIGPLAN-SIGACT Symposium on Principles of Programming Languages, ACM, 1997. [27] M. Odersky. Functional nets. In Proceedings of ESOP – 9th European Symposium on Programming, Lecture Notes in Computer Science 1782, Springer, 2000. [28] J. Ousterhout. Why threads are a bad idea (for most purposes). Invited talk at the USENIX Technical Conference, 1996. [29] Th. Park. Bounded scheduling of process networks. PhD thesis, University of California, Berkeley, 1995. [30] P. Puschner and A. Burns (eds.), Real time systems 18(2/3), Special issue on Worstcase execution time analysis, 2000. [31] Reactive Programming, INRIA http://www-sop.inria.fr/mimosa/rp. A Sophia-Antipolis, Mimosa Project. Readers-Writers and Other Synchronisation Patterns A simple, maybe the simplest, example of synchronisation and resource protection is the single place buffer. The buffer (initially empty) is implemented by a thread listening to two signals. The first on the register put to fill the buffer with a value if it is empty, the second on the register get to emit the value stored in the buffer by writing it in the special register result and flush the buffer. In this encoding, the register put is a one place channel and get is a signal as in Example 1. Moreover, owing to the read once condition, we are not able to react to several put/get requests during the same instant — only if the buffer is full can we process one get and one put request in the same instant. Note that the value of the buffer is stored on the function call to full (v), hence we use function parameters as a kind of private memory (to compare with registers that model shared memory). empty() = read put with full(x) ⇒ next.full (x) | [ ] ⇒ empty() full(x) = read get with prst ⇒ result := x.yield .empty() | [ ] ⇒ full(x) Another common example of synchronisation pattern is a situation where we need to protect a resource that may be accessed both by ‘readers’ (which access the resource without modifying it) and ‘writers’ (which can access and modify the resource). This form of 28 access control is common in databases and can be implemented using traditional synchronisation mechanisms such as semaphores, but this implementation is far from trivial [27]. In our encoding, a control thread secures the access to the protected resource. The other threads, which may be distinguished by their identity id (a natural number), may initiate a request to access / release the resource by sending a special value on the dedicated register req. The thread regulating the resource may acknowledge at most one request per instant and allows the sender of a request to proceed by writing its id on the register allow at the next instant. The synchronisation constraints are as follows: there can be multiple concurrent readers, there can be only one writer at any one time, pending write requests have priority over pending read requests (but do not preempt ongoing read operations). We define a new algebraic datatype for assigning requests: request = startRead(nat) | startWrite(nat) | endRead | endWrite | none The value startRead(id ) indicates a read request from the thread id , the other constructors correspond to requests for starting to write, ending to read or ending to write — the value none stands for no requests. A startRead operation requires that there are no pending writes to proceed. In that case we increment the number of ongoing readers and allow the caller to proceed. By contrast, a startWrite puts the monitor thread in a state waiting to process the pending write request (function pwrite), which waits for the number of readers to be null and then allows the thread that made the pending write request to proceed. An endRead and endWrite request is always immediately acknowledged. The thread protecting the resource starts with the behaviour onlyreader(z), defined in Table 5, meaning the system has no pending requests for reading or writing. The behaviour onlyreader(x) encodes the state of the controller when there is no pending write and x readers. In a state with x pending readers, when a startWrite request from the thread id is received, the controller thread switches to the behaviour pwrite(id, x), meaning that the thread id is waiting to write and that we should wait for x endRead requests before acknowledging the request to write. A thread willing to read on the protected resource should repeatedly try to send its request on the register req then poll the register allow, e.g., with the behaviour askRead(id ).read allow with id ⇒ · · · where askRead (id ) is a shorthand for read req with none ⇒ req := startRead(id ). The code for a thread willing to end a read session is similar. It is simple to change our encoding so that multiple requests are stored in a fifo queue instead of a one place buffer. B B.1 Proofs Preservation of Control Points Instances Proposition 31 8 Suppose (B, s, i) → (B ′ , s′ , i′ ) and that for all thread indexes j ∈ Zn , B1 (j) is an instance of a control point. Then for all j ∈ Zn , we have that B1′ (j) is an instance of a control point. 29 Table 5: Code for the Readers-Writers pattern onlyreader (x) = match x with s(x′ ) then read req with endRead ⇒ next.onlyreader (x′ ) | startWrite(y) ⇒ next.pwrite(y, s(x′ )) | startRead(y) ⇒ next.allow := y.onlyreader (s(s(x′ ))) | [ ] ⇒ onlyreader (s(x′ )) else read req with startWrite(y) ⇒ next.allow := y.pwrite(y, z) | startRead(y) ⇒ next.allow := y.onlyreader (s(z)) | [ ] ⇒ onlyreader (z) = match x with s(x′ ) then match x′ with s(x′′ ) then read req with endRead ⇒ next.pwrite(id , s(x′′ )) | [ ] ⇒ pwrite(id , s(s(x′′ ))) else read req with endRead ⇒ next.allow := id .pwrite(id , z) | [ ] ⇒ pwrite(id , s(z)) else read req with endWrite ⇒ next.onlyreader (z) | [ ] ⇒ pwrite(id , z) pwrite(id , x) Proof. Let (f (p), be, i) be a control point of an expression body or of a behaviour. In Table 6, we reformulate the evaluation and the reduction by replacing expression bodies or behaviours by triples (f (p), be, σ) where (f (p), be, i) is a control point and σ is a substitution mapping the variables in p to values. By convention, we take σ(r) = r if r is a register. We claim that the evaluation and reduction in Table 6 are equivalent to those presented in Section 2 in the following sense: 1. (f (p), e0 , σ) ⇓ v iff σe0 ⇓ v. X X 2. (f + (p), b0 , s, σ) → (g + (q), b′0 , s′, σ ′ ) iff σb0 → σ ′ b′0 . In the following proofs we will refer to the rules in Table 6. The revised formulation X makes clear that if b is an instance of a control point and (b, s)→(b′ , s′ ) then b′ is an instance. It remains to check that being an instance is a property preserved at the level of system reduction. We proceed by case analysis on the last reduction rule used in the derivation of (B, s, i) → (B ′ , s′ , i′ ). (s1 ) One of the threads performs one step. The property follows by the analysis on behaviours. (s2 ) One of the threads performs one step. Moreover, the threads in waiting status take the 30 Table 6: Expression body evaluation and behaviour reduction revised (e0 ) (e1 ) (f (p), x, σ) ⇓ σ(x) (f (p), r, σ) ⇓ r (f (p), ei , σ) ⇓ vi i ∈ 1..n, (f (p), ei , σ) ⇓ vi i ∈ 1..n g(x) = eb, (g(x), eb, [v/x]) ⇓ v (e3 ) (e2 ) (f (p), c(e), σ) ⇓ c(v) (f (p), g(e), σ) ⇓ v σ(x) = d(. . .), σ(x) = c(v), (f (p), eb 2 , σ) ⇓ v (f ([c(x)/x]p), eb 1 , [v/x] ◦ σ) ⇓ v     (e5 ) (e4 ) match x match x f (p), f (p), with c(x) with c(x) , σ ⇓ v , σ ⇓ v then eb 1 else eb 2 then eb 1 else eb 2 (b1 ) S (f + (p), stop, σ, s) → (f + (p), stop, σ, s) (b2 ) (b3 ) (b4 ) R (f + (p), yield .b, σ, s) → (f + (p), b, σ, s) N (f + (p), next .g(e), σ, s) → (f + (p), g(e), σ, s) X (f + ([c(x)/x]p), b1 , [v/x]◦ σ, s) → (f1+ (p′ ), b′ , σ ′ , s′ ) match x with c(x) X f + (p), , σ, s → (f1+ (p′ ), b′ , σ ′ , s′ ) then b1 else b2 σ(x) =  c(v), X + ′ + ′ ′ ′ σ(x)  = d(. . .), c 6= d, (f (p), b2 , σ,  s) → (f1 (p ), b , σ , s ) (b5 ) match x with c(x) X f + (p), , s, σ → (f1+ (p′ ), b′ , σ ′ , s′ ) then b1 else b2 no pattern matches s(σ(̺)) (b6 ) W (f + (p), read ̺ with . . . , σ, s) → (f + (p), read ̺ with . . . , σ, s) X (b7 ) σ1 (p) = s(σ(̺)), (f + ([p/y]p), b, σ1 ◦ σ, s) → (f1+ (p′ ), b′ , σ ′ , s′ ) X (f + (p), read hyi ̺ with · · · | p ⇒ b | . . . , σ, s) → (f1+ (p′ ), b′ , σ ′ , s′ ) σe ⇓ v, g(x) = b, (b8 ) X (g+ (x, yg ), b, [v/x], s) → (f1+ (p′ ), b′ , σ ′ , s′ ) X (f + (p), g(e), σ, s) → (f1+ (p′ ), b′ , σ ′ , s′ ) X (b9 ) σe ⇓ v, (f + (p), b, σ, s[v/σ(̺)]) → (f1+ (p′ ), b′ , σ ′ , s′ ) X (f + (p), ̺ := e.b, σ, s) → (f1+ (p′ ), b′ , σ ′ , s′ ) 31 [ ] ⇒ g(e) branch of the read instructions that were blocking. A thread read ̺ . . . | [ ] ⇒ g(e) in waiting status is an instance of a control point (f + (p), read ̺ . . . | [ ] ⇒ g(e0), j). By (C7 ), (f + (p), g(e0 ), 2) is a control point, and g(e) is one of its instances. ✷ B.2 Evaluation of Closed Expressions Proposition 32 12 Let e be a closed expression. Then there is a value v such that e ⇓ v and e ≥ v with respect to the reduction order. As announced, we refer to the rules in Table 6. We recall that the order > or ≥ refers to the reduction order that satisfies the constraints of index 0. We start by proving the following working lemma. Lemma 33 For all well formed triples, (f (p), eb, σ), there is a value v such that (f (p), eb, σ) ⇓ v. Moreover, if eb is an expression then σ(eb) ≥ v else f (σp) ≥ v. Proof. We proceed by induction on the pair (f (σp), eb) ordered lexicographically from left to right. The first argument is ordered according to the reduction order and the second according to the structure of the expression body. eb ≡ x. We apply rule (e0 ) and σ(x) ≥ σ(x). eb ≡ r. We apply rule (e1 ) and σ(r) = r ≥ r. eb ≡ c(e1 , . . . , en ). We apply rule (e2 ). By inductive hypothesis, (f (p), ei , σ) ⇓ vi for i ∈ 1..n and σei ≥ vi . By definition of reduction order, we derive σ(c(e1 , . . . , en )) ≥ c(v1 , . . . , vn ). eb ≡ f (e1 , . . . , en ). We apply rule (e3 ). By inductive hypothesis, (f (p), ei , σ) ⇓ vi for i ∈ 1..n and σei ≥ vi . By the definition of the generated constraints f (p) > g(e), which by definition of reduction order implies that f (σp) > g(σe) ≥ g(v) = g([v/x]x). Thus by inductive hypothesis, g(x, eb, [v/x]) ⇓ v. We conclude by showing by case analysis that g(σe) ≥ v. • eb is an expression. By the constraint we have g(x) > eb, and by inductive hypothesis [v/x]eb ≥ v. So g(σe) ≥ g(v) > [v/x]eb ≥ v. • eb is not an expression. Then by inductive hypothesis, g(v) ≥ v and we know g(σe) ≥ g(v). eb ≡ match x with c(x) . . . . We distinguish two cases. • σ(x) = c(v). Then rule (e4 ) applies. Let σ ′ = [v/x]◦σ. Note that σ ′ ([c(x)/x]p) = σp. By inductive hypothesis, we have that (f ([c(x)/x]p), eb 1 , σ ′ ) ⇓ v. We show by case analysis that f (σp) ≥ v. 32 – eb 1 is an expression. By inductive hypothesis, σ ′ (eb 1 ) ≥ v. By the constraint, f ([c(x)/x]p) > eb 1 . Hence, f (σp) = f (σ ′ [c(x)/x]p) > σ ′ (eb 1 ). – eb 2 is not an expression. By inductive hypothesis, we have that f (σp) equals f (σ ′ [c(x)/x]p) ≥ v. • σ(x) = d(. . .) with c 6= d. Then rule (e5 ) applies and an argument simpler than the one above allows to conclude. ✷ Relying on Lemma 33 we can now prove Proposition 12, that if e is a closed expression and e ⇓ v then e ≥ v in the reduction order. Proof. We proceed by induction on the structure of e. e is value v. Then v ⇓ v and v ≥ v. e ≡ c(e1 , . . . , en ). By inductive hypothesis, ei ⇓ vi and ei ≥ vi for i ∈ 1..n. By definition of reduction order, c(e) ≥ c(v). e ≡ f (e1 , . . . , en ). By inductive hypothesis, ei ⇓ vi and ei ≥ vi for i ∈ 1..n. Suppose f (x) = eb. By Lemma 33, (f (x), eb, [v/x]) ⇓ v and either f (v) ≥ v or f (x) > eb and σ(eb) ≥ v. We conclude by a simple case analysis. ✷ B.3 Progress Proposition 34 13 Let b be an instance of a control point. Then for all stores s, there X exists a store s′ and a status X such that (b, s) → (b′ , s′ ). Proof. We start by defining a suitable well-founded order. If b is a behaviour, then let nr (b) be the maximum number of reads that b may perform in an instant. Moreover, let ln(b) be the length of b inductively defined as follows: ln(stop) = ln(f (e)) = 0 ln(yield .b) = ln(̺ := e.b) = 1 + ln(b) ln(next.f (e)) = 2 ln(match x with c(x) then b1 else b2 ) = 1 + max (ln(b1 ), ln(b2 )) ln(read ̺ with . . . | pi ⇒ bi | . . . | [ ] ⇒ f (e)) = 1 + max (. . . , ln(bi ), . . .) If the behaviour b is an instance of the control point γ ≡ (f + (p), b0 , i) via a substitution σ then we associate with the pair (b, γ) a measure: µ(b, γ) =def (nr(b), f + (σp), ln(b)) . We assume that measures are lexicographically ordered from left to right, where the order on the first and third component is the standard order on natural numbers and the order on the second component is the reduction order considered in study of the termination conditions. This is a well-founded order. Now we show the assertion by induction on µ(b, γ). We proceed by case analysis on the structure of b. b ≡ stop. Rule (b1 ) applies, with X = S, and the measure stays constant. 33 b ≡ yield .b′ . Rule (b2 ) applies, with X = R, and the measure decreases because ln(b) decreases. b ≡ next.b′ . Rule (b3 ) applies, with X = N, and the measure decreases because ln(b) decreases. b ≡ match . . . . creases. Rules (b4 ) or (b5 ) apply and the measure decreases because ln(b) de- b ≡ read . . . . If no pattern matches then rule (b6 ) applies and the measure is left unchanged. If a pattern matches then rule (b7 ) applies and the measure decreases because nr (b) decreases and then the induction hypothesis applies. b ≡ g(e). Rule (b8 ) applies to (f + (p), g(e0 ), σ), assuming e = σe0 . By Proposition 12, we know that e ⇓ v and e ≥ v in the reduction order. Suppose g is associated to the declaration g(x) = b. The constraint associated with the control point requires f + (p) > g + (e0 , yg ). Then using the properties of reduction orders we observe: f + (σp) > g + (σe0 , yg ) = g + (e, yg ) ≥ g + (v, yg ) Thus the measure decreases because f + (σp) > g + (v, yg ), and then the induction hypothesis applies. b ≡ ̺ := e.b′ . By Proposition 12, we have e ⇓ v. Hence rule (b9 ) applies, the measure decreases because ln(b) decreases, and then the induction hypothesis applies. ✷ Remark 35 We point out that in the proof of proposition 13, if X = R then the measure decreases and if X ∈ {N, S, W } then the measure decreases or stays the same. We use this observation in the following proof of Theorem 14. B.4 Termination of the Instant Theorem 36 14 All sequences of system reductions involving only rule (s1 ) are finite. Proof. We order the status of threads as follows: R > N, S, W . With a behaviour B1 (i) coming with a control point γi , we associate the pair µ′ (i) = (µ(B1 (i), γi ), B2 (i)) where µ is the measure defined in the proof of Proposition 13. Thus µ′ (i) can be regarded as a quadruple with a lexicographic order from left to right. With a system B of n threads we associate the measure µB =def (µ′ (0), . . . , µ′(n − 1)) that is a tuple. We compare such tuples using the product order. We prove that every system reduction sequence involving only rule (s1 ) terminates by proving that this measure decreases during reduction. We recall the rule below: X (B1 (i), s) → (b′ , s′ ), B2 (i) = R, B ′ = B[(b′ , X)/i], N (B ′ , s′ , i) = k (B, s, i) → (B ′ [(B1′ (k), R)/k], s′ , k) 34 Let B ′′ = B ′ [(B1′ (k), R)/k]. We proceed by case analysis on X and B2′ (k). If B2′ (k) = R then µ′ (k) is left unchanged. The only other case is B2′ (k) = W . In this case the conditions on the scheduler tell us that i 6= k. Indeed, the thread k must be blocked on a read r instruction and it can only be scheduled if the value stored in r has been modified, which means than some other thread than k must have modified r. For the same reason, some pattern in the read r instruction of B1 (k) matches s′ (r), which means that the number of reads that B1 (k) may perform in the current instant decreases and that µ′ (k) also decreases. X By hypothesis we have (B1 (i), s) → (b′ , s′ ), hence by Remark 35, µ′ (i) decreases or stays the same. By the previous line of reasoning µ′ (k) decreases and the other measures µ′ (j) stay the same. Hence the measure µB decreases, as needed. ✷ B.5 Bounding the Size of Values for Threads Theorem 37 19 Given a system of synchronous threads B, suppose that at the beginning of the instant B1 (i) = f (v) for some thread index i. Then the size of the values computed by the thread i during an instant is bounded by qf + (v,u) where u are the values contained in the registers at the time they are read by the thread (or some constant value, if they are not read at all). In Table 6, we have defined the reduction of behaviours as a big step semantics. In Table 7 we reformulate the operational semantics following a small step approach. First, note that there are no rules corresponding to (b1 ), (b3 ) or (b6 ) since these rules either terminate or suspend the computation of the thread in the instant. Second, the reduction makes abstraction of the memory and the scheduler. Instead, the reduction relation is parameterized on an assignment δ associating values with the labels of the read instructions. The assignment δ is a kind of oracle that provides the thread with the finitely many values (because of the read once condition) it may read within the current instant. The assignment δ provides a safe abstraction of the store s used in the transition rules of Table 6. Note that the resulting system represents more reductions than can actually occur in the original semantics within an instant. Namely, a thread can write a value v in r and then proceed to read from r a value different from v without yielding the control. This kind of reduction is impossible in the original semantics. However, since we do not rely on a precise monitoring of the values written in the store, this loss of precision does not affect our analysis. Next we prove that if (f + (p), b, σ) →δ (g + (q), b′ , σ ′ ) then qf + (σ′′ ◦σ(p)) ≥ qg+ (σ′ (q)) over the non-negative reals, where σ ′′ is either the identity or the restriction of δ to the label of the read instruction in case (b′ 7 ). Proof. By case analysis on the small step rules. Cases (b′ 2 ), (b′ 5 ) and (b′ 9 ) are immediate. (b′ 4 ) The assertion follows by a straightforward computation on substitutions. (b′ 7 ) Then σ ′′ (y) = δ(y) = [σ1 (p)/y] and recalling that patterns are linear, we note that: f + ((σ ′′ ◦ σ)(p)) = f + ((σ1 ◦ σ)[p/y](p)). 35 Table 7: Small step reduction within an instant (b′ 2 ) (f + (p), yield .b, σ) →δ (f + (p), b, σ) match x with c(x)
, σ →δ (f + ([c(x)/x]p), b1 , [v/x] ◦ σ) if (1) (b′ 4 ) f + (p), then b1 else b2 match x with c(x)
(b′ 5 ) f + (p), , σ →δ (f + (p), b2 , σ) if σ(x) = d(. . .), c 6= d then b1 else b2 (b′ 7 ) (f + (p), read hyi ̺ with · · · | p ⇒ b | . . . , σ) →δ (f + ([p/y]p), b, σ1 ◦ σ) if (2) (b′ 8 ) (f + (p), g(e), σ) →δ (g+ (x, yg ), b, [v/x]) if σe ⇓ v and g(x) = b (b′ 9 ) (f + (p), ̺ := e.b, σ) →δ (f + (p), b, σ) if σe ⇓ v where: (1) ≡ σ(x) = c(v) and (2) ≡ σ1 (p) = δ(y). (b′ 8 ) By the properties of quasi-interpretations, we know that qσ(e) ≥ qv . By the constraints generated by the control points, we derive that qf + (p) ≥ qg+ (e,yg ) over the nonnegative reals. By the substitutivity property of quasi-interpretations, this implies that qf + (σ(p)) ≥ qg+ (σ(e,yg )) . Thus we derive, as required: qf + (σ(p)) ≥ qg+ (σ(e,yg )) ≥ qg+ (v,yg ) . ✷ It remains to support our claim that all values computed by the thread i during an instant have a size bounded by qf (v,u) where u are either the values read by the thread or some constant value. Proof. By inspecting the shape of behaviours we see that a thread computes values either when writing into a register or in recursive calls. We consider in turn the two cases. Writing Suppose (f + (p, yf ), b, σ) →∗δ (g + (q), ̺ := e.b′ , σ ′ ) by performing a series of reads recorded by the substitution σ ′′ . Then the invariant we have proved above implies that: qf + ((σ′′ ◦σ)(p,yf )) ≥ qg+ (σ′ q) over the non-negative reals. If some of the variables in yf are not instantiated by the substitution σ ′′ , then we may replace them by some constant. Next, we observe that the constraint of index 1 associated with the control point requires that qg+ (q) ≥ qe and that if σ(e) ⇓ v then this implies qg+ (σ′ (q)) ≥ qσ′ (e) ≥ qv ≥ |v|. Recursive call Suppose (f + (p, yf ), b, σ) →∗δ (g + (q), h(e), σ ′ ) by performing a series of reads recorded by the substitution σ ′′ . Then the invariant we have proved above implies that: qf + ((σ′′ ◦σ)(p,yf )) ≥ qg+ (σ′ (q)) over the non-negative reals. Again, if some of the variables in yf are not instantiated by the substitution σ ′′ , then we may replace them by some constant value. Next we observe that the constraint of index 0 associated with the control point requires that qg+ (q) ≥ qh+ (e,yh ) . Moreover, if σ ′ (e) ⇓ v then qg+ (σ′ (q)) ≥ qh+ (σ′ (e,yh )) ≥ qh+ (v,yh ) ≥ qvi ≥ |vi |, where vi is any of the values in v. The last inequation relies on the monotonicity property of assignments, see property (3) in Definition 16, that is ✷ qh+ (z1 , . . . , zn ) ≥ zj for all j ∈ 1..n. B.6 Bounding the Size of Values for Systems Corollary 38 20 Let B be a system with m distinct read instructions and n threads. 36 Suppose B1 (i) = fi (vi ) for i ∈ Zn . Let c be a bound of the size of the largest parameter of the functions fi and the largest default value of the registers. Suppose h is a function bounding all the quasi-interpretations, that is, for all the functions fi+ we have h(x) ≥ qfi+ (x, . . . , x) over the non-negative reals. Then the size of the values computed by the system B during an instant is bounded by hn·m+1 (c). Proof. Because of the read once condition, during an instant a system can perform a (successful) read at most n · m times. We proceed by induction on the number k of reads the system has performed so far to prove that the size of the values is bounded by hk+1 (c). k = 0 If no read has been performed, then Theorem 19 can be applied to show that all values have size bound by h(c). k > 0 Inductively, the size of the values in the parameters and the registers is bounded by hk (c). Theorem 19 says that all the values that can be computed before performing a new read have a size bound by h(hk (c)) = hk+1 (c). ✷ B.7 Combination of LPO and Polynomial Quasi-interpretations Theorem 39 24 If a system B terminates by LPO and admits a polynomial quasi-interpretation then the computation of the system in an instant runs in space polynomial in the size of the parameters of the threads at the beginning of the instant. Proof. We can always choose a polynomial for the function h in corollary 20. Hence, hnm+1 is also a polynomial. This shows that the size of all the values computed by the system is bounded by a polynomial. The number of values in a frame depends on the number of formal parameters and local variables and it can be statically bound. It remains to bound the number of frames on the stack. Note that behaviours are tail recursive. This means that the stack of each thread contains a frame that never returns a value plus possibly a sequence of frames that relate to the evaluation of expressions. From this point on, one can follow the proof in [10]. The idea is to exploit the characteristics of the LPO order: a nested sequence of recursive calls f1 (v1 ), . . . , fn (vn ) must satisfy f1 (v1 ) > · · · > fn (vn ), where > is the LPO order on terms. Because of the polynomial bound on the size of the values and the characteristics of the LPO on constructors, one can provide a polynomial bound on the length of such strictly decreasing sequences and therefore a polynomial bound on the size of the stack needed to execute the system. ✷ B.8 Compiled Code is Well-shaped Theorem 40 30 The shape analysis succeeds on the compilation of a well-formed program. Let be be either a behaviour or an expression body, η be a sequence of variables, and E be a sequence of expressions. We say that the triple (be, η, E) is compatible if for all variables x free in be, the index i(x, η) is defined and if η[k] = x then E[k] = x. Moreover, 37 we say that the triple is strongly compatible if it is compatible and |η| = |E|. In the following we will neglect typing issues that offer no particular difficulty. First we prove the following lemma. Lemma 41 If (e, η, E) is compatible then the shape analysis of C ′ (e, η) starting from the shape E succeeds and produces a shape E · e. Proof. By induction on the structure of e. e ≡ x Then C ′ (x, η) = load i(x, η). We know that i(x, η) is defined and η[k] = x implies E[k] = x. So the shape analysis succeeds and produces E · x. e ≡ c(e1 , . . . , en ) Then C ′ (c(e1 , . . . , en ), η) = C ′ (e1 , η) · · · C ′ (en , η)(build c n). We note that if e′ is a subexpression of e, e′′ is another expression, and (e, η, E) is compatible then (e′ , η, E · e′′ ) is compatible too. Thus we can apply the inductive hypothesis to e1 , . . . , en and derive that the shape analysis of C ′ (e1 , η) starting from E succeeds and produces E · e1 ,. . . , and the shape analysis of C ′ (en , η) starting from E · e1 · · · en−1 succeeds and produces E · e1 · · · en . Then by the definition of shape analysis of build we can conclude. e ≡ f (e1 , . . . , en ) An argument similar to the one above applies. Next we generalise the lemma to behaviours and expression bodies. ✷ Lemma 42 If (be, η, E) is strongly compatible then the shape analysis of C(be, η) starting from the shape E succeeds. Proof. be ≡ e We have that C(e, η) = C ′ (e, η)·return and the shape analysis on C ′ (e, η) succeeds, producing at least one expression. be ≡ match x with c(y) then eb 1 else eb 2 Following the definition of the compilation function, we distinguish two cases: • η ≡ η ′ · x: Then C(be, η) = (branch c j) · C(eb 1 , η ′ · y) · (j : C(eb 2 , η) ). By the hypothesis of strong compatibility, E ≡ E ′ · x and by definition of shape analysis on branch we get on the then branch a shape [c(y)/x]E ′ · y up to variable renaming. We observe that (eb 1 , η ′ · y, [c(y)/x]E ′ · y) are strongly compatible (note that here we rely on the fact that η ′ and E ′ have the same length). Hence, by inductive hypothesis, the shape analysis on C(eb 1 , η ′ · y) succeeds. As for the else branch, we have a shape E ′ · x and since (eb 2 , η ′ · x, E ′ · x) are strongly compatible we derive by inductive hypothesis that the shape analysis on C(eb 2 , η) succeeds. • η 6≡ η ′ · x: The compiled code starts with (load i(x, η)) which produces a shape E · x. Then the analysis proceeds as in the previous case. be ≡ stop The shape analysis succeeds. be ≡ f (e1 , . . . , en ) By lemma 41, we derive that the shape analysis of C ′ (e1 , η)· . . .·C ′ (en , η) 38 succeeds and produces E · e1 · · · en . We conclude applying the definition of the shape analysis for tcall. be ≡ yield .b The instruction yield does not change the shape and we can apply the inductive hypothesis on b. be ≡ next.g(e) The instruction next does not change the shape and we can apply the inductive hypothesis on g(e). be ≡ ̺ := e.b By lemma 41, we have the shape E · e. By definition of the shape analysis on write, we get back to the shape E and then we apply the inductive hypothesis on b. be ≡ match . . . The same argument as for expression bodies applies. be ≡ read ̺ with c1 (y1 ) ⇒ b1 | . . . | cn (yn ) ⇒ bn | [ ] ⇒ g(e) We recall that the compiled code is: j0 : (read i(̺, η)) · (branch c1 j1 ) · C(b1 , η · y1 ) · · · jn−1 : (branch cn jn ) · C(bn , η · yn ) · jn : (wait j0 ) · C(g(e), η) The read instruction produces a shape E · y. Then if a positive branch is selected, we have a shape E · yk for k ∈ 1..n. We note that the triples (bk , η · yk , E · yk ) are strongly compatible and therefore the inductive hypothesis applies to C(bk , η · yk ) for k ∈ 1..n. On the other hand, if the last default branch [ ] is selected then by definition of the shape analysis on wait we get back to the shape E and again the inductive hypothesis applies to C(g(e), η). The case where a pattern can be a variable is similar. To conclude the proof we notice that for every function definition f (x) = be, taking η = x = E we have that (be, η, E) are strongly compatible and thus by lemma 42 the shape analysis succeeds on C(be, η) starting from E. ✷ 39 

The Price of Selection in Differential Privacy Mitali Bafna* Jonathan Ullman† arXiv:1702.02970v1 [cs.DS] 9 Feb 2017 February 13, 2017 Abstract In the differentially private top-k selection problem, we are given a dataset X ∈ {±1}n×d , in which each row belongs to an individual and each column corresponds to some binary attribute, and our goal is to find a set of k  d columns whose means are approximately as large as possible. Differential privacy requires that our choice of these k columns does not depend too much on any on individual’s dataset. This problem can be solved using the well known exponential mechanism and composition properties of differential privacy. In the high-accuracy regime, where we require the p error of the selection procedure to be to be smaller than the so-called sampling error α ≈ ln(d)/n, this procedure succeeds given a dataset of size n & k ln(d). We prove a matching lower bound, showing that a dataset of size n & k ln(d) is necessary for private top-k selection in this high-accuracy regime. Our lower bound is the first to show that selecting the k largest columns requires more data than simply estimating the value of those k columns, which can be done using a dataset of size just n & k. * IIT Madras, Department of Computer Science and Engineering. This research was performed while the author was a visiting scholar at Northeastern University. mitali.bafna@gmail.com † Northeastern University, College of Computer and Information Science. jullman@ccs.neu.edu 1 Introduction The goal of privacy-preserving data analysis is to enable rich statistical analysis of a sensitive dataset while protecting the privacy of the individuals who make up that dataset. It is especially desirable to ensure differential privacy [DMNS06], which ensures that no individual’s information has a significant influence on the information released about the dataset. The central problem in differential privacy research is to determine precisely what statistics can be computed by differentially private algorithms and how accurately they can be computed. The seminal work of Dinur and Nissim [DN03] established a “price of privacy”: If we release the answer to & n statistics on a dataset of n individuals, and we do so with error that is asymptotically √ smaller than the sampling error of ≈ 1/ n, then an attacker can reconstruct nearly all of the sensitive information in the dataset, violating any reasonable notion of privacy. For example, if we have a dataset X = (x1 , . . . , xn ) ∈ {±1}n×d and we want to privately approximate its marginal √ 1 Pn vector q = n i=1 xi , then it is suffices to introduce error of magnitude Θ( d/n) to each entry of q [DN03, DN04, BDMN05, DMNS06], and this amount of error is also necessary [BUV14, SU15]. Thus, when d  n, the error must be asymptotically larger the sampling error. Top-k Selection. In many settings, we are releasing the marginals of the dataset in order to find a small set of “interesting” marginals, and we don’t need the entire vector. For example, we may be interested in finding only the attributes that are unusually frequent in the dataset. Thus, an appealing approach to overcome the limitations on computing marginals is to find only the top-k (approximately) largest coordinates of the marginal vector q, up to some error α.1 Once we find these √ k coordinates, we can approximate the corresponding marginals with additional error O( k/n). But, how much error must we have in the top-k selection itself? The simplest way to solve this problem is to greedily find k coordinates using the differentially √ private exponential mechanism [MT07]. This approach finds the top-k marginals up to error . k log(d)/n. The sparse vector algorithm [DNPR10, RR10, HR10] would provide similar guarantees. Thus, when k  d, we can find the top-k marginals and approximate their values with much less error than √ approximating the entire vector of marginals. However, the bottleneck in this approach is the k log(d)/n error in the selection procedure, and this log(d) factor is significant in very p high-dimensional datasets. For comparison, the sampling error for top-k selection is ≈ log(d)/n so the error introduced is asymptotically larger than the sampling error when k log(d)  n. However, the best known lower bound for top-k selection follows by scaling down the lower bounds for √ releasing the entire marginal vector, and say that the error must be & k/n. Top-k selection is a special case of fundamental data analysis procedures like variable selection and sparse regression. Moreover, private algorithms for selection problems underlie many powerful results in differential privacy: private control of false discovery rate [DSZ15], algorithms for answering exponentially many queries [RR10, HR10, GRU12, JT12, Ull15], approximation algorithms [GLM+ 10], frequent itsemset mining [BLST10], sparse regression [ST13], and the optimal analysis of the generalization error of differentially private algorithms [BNS+ 16]. Therefore it is important to precisely understand optimal algorithms for differentially private top-k selection. Our main result says that existing differentially private algorithms for top-k selection are essentially optimal in the high-accuracy regime where the error is required to be asymptotically smaller than the sampling error. 1 Here, the algorithm has error α if it returns a set S ⊆ {1, . . . , d} consisting of of k coordinates, and for each coordinate j ∈ S, qj ≥ τ − α, where τ is the k-th largest value among all the coordinates {q1 , . . . , qd }. 1 Theorem 1.1 (Sample Complexity Lower Bound for Approximate Top-k). There exist functions p n = Ω(k log(d)) and α = Ω( log(d)/n) such that for every d and every k = d o(1) , there is no differentially private algorithm M that takes an arbitrary dataset X ∈ {±1}n×d and (with high probability) outputs an α-accurate top-k marginal vector for X. Tracing Attacks. Our lower bounds for differential privacy follow from a tracing attack [HSR+ 08, SOJH09, BUV14, SU15, DSS+ 15, DSSU17]. In a tracing attack, the dataset X consists of data for n individuals drawn iid from some known distribution over {±1}d . The attacker is given data for a target individual y ∈ {±1}d who is either one of the individuals in X (“IN”), or is an independent draw from the same distribution (“OUT”). The attacker is given some statistics about X (e.g. the top-k statistics) and has to determine if the target y is in or out of the dataset. Tracing attacks are a significant privacy violation, as mere presence in the dataset can be sensitive information, for example if the dataset represents the case group in a medical study [HSR+ 08]. Our results give a tracing attack for top-k statistics in the case where the dataset is drawn uniformly at random. For simplicity, we state the properties of our tracing attack for the case of the exact top-k marginals. We refer the reader to Section 4 for a detailed statement in the case of approximate top-k marginals, which is what we use to establish Theorem 1.1. Theorem 1.2 (Tracing Attack for Exact Top-k). For every ρ > 0, every n ∈ N, and every k  d  2n such that k log(d/k) ≥ O(n log(1/ρ)), there exists an attacker A : {−1, 1}d × {0, 1}d → {IN, OUT} such that the following holds: If we choose X = (x1 , . . . , xn ) ∈ {±1}n×d uniformly at random, and t(X) is the exact top-k vector2 of X, then 1. If y ∈ {±1}d is uniformly random and independent of X, then P [A(y, t(X)) = OUT] ≥ 1 − ρ, and 2. for every i ∈ [n], P [A(xi , t(X)) = IN] ≥ 1 − ρ. While the assumption of uniformly random data is restrictive, it is still sufficient to provide a lower bound for differential privacy. Tracing attacks against algorithms that release the entire marginal vector succeed under weaker assumptions—each column can have a different and essentially arbitrary bias as long as columns are independent. However, for top-k statistics, a stronger assumption on the column biases is necessary—if the column biases are such that t(X) contains a specific set of columns with overwhelming probability, then t(X) reveals essentially no information about X, so tracing will fail. Under the weaker assumption that some unknown set of k columns “stand out” by having significantly larger bias than other columns, we can use the propose-test-release framework [DL09] to find the exact top-k vector when n & log(d). An interesting future direction is to characterize which distributional assumptions are sufficient to bypass our lower bound. We remark that, since our attack “traces” all rows of the dataset (i.e. A(xi , t(X)) = IN for every i ∈ [n]), the attack bears some similarities to a reconstruction attack [DN03, DMT07, DY08, KRSU10, KRS13, NTZ13]. However, the focus on high-dimensional data and the style of analysis is much closer to the literature on tracing attacks. 1.1 Proof Overview Our results use a variant of the inner product attack introduced in [DSS+ 15] (and inspired by the work on fingerprinting codes [BS98, Tar08] and their connection to privacy [Ull13, BUV14, SU15]). 2 Due to the presence of ties, there is typically not a unique top-k. For technical reasons, and for simplicity, we let t(X) denote the unique lexicographically first top-k vector and refer to it as “the” top-k vector. 2 Given a target individual y ∈ {±1}d , and a top-k vector t ∈ {±1}d , the attack is    if hy, ti ≥ τ IN A(y, t) =   OUT otherwise √ where τ = Θ( k) is an appropriately chosen threshold. The key to the analysis is to show that, when X = (x1 , . . . , xn ) ∈ {±1}n×d and y ∈ {±1}d are chosen uniformly at random, t(X) is an accurate top-k vector of X, then E [hy, t(X)i] = 0 and ∀i ∈ [n] E [hxi , t(X)i] > 2τ. If we can establish these two facts then Theorem 1.2 will follow from concentration inequalities for the two inner products. Suppose t(X) is the exact top-k vector. Since each coordinate of y is uniform in {±1} and independent of X, we can write i P h i h i P h E [hy, t(X)i] = j E yj · t(X)j = j E yj E t(X)j = 0. Moreover, for every fixed vector t ∈ {±1}d with k non-zero coordinates, hy, ti is a sum of k indepen√ dent, bounded random variables. Therefore, by Hoeffding’s inequality we have that hy, ti = O( k) with high probability. Since y, X are independent, this bound also holds with high probability √ when X is chosen randomly and t(X) is its top-k vector. Thus, for an appropriate τ = Θ( k), A(y, t(X)) = OUT with high probability. Now, consider the case where y = xi is a row of X, and we want to show that E [hxi , t(X)i] is sufficiently large. Since X is chosen uniformly at random. One can show that, when k  d  2n , the p top-k largest marginals of X are all at least γ = Ω( log(d/k)/n). Thus, on average, when t(X)j = 1, we can think of xi,j ∈ {±1} as a random variable with expectation ≥ γ. Therefore, E [hxi , t(X)i] = E hP i  p  x ≥ kγ = Ω k log(d/k)/n j:t(X)j =1 i,j Even though xi and t(X) are not independent, and do not √ entries, we show that p have independent with high probability over the choice of X, hxi , t(X)i ≥ k log(d/k)/n − O( k) with high probability. Thus, if k log(d/k) & n, we have that A(xi , t(X)) = IN with high probability. Extension to Noisy Top-k. The case of α-approximate top-k statistics does not change the analysis of hy, ti in that case that y is independent of x, but does change the analysis of hxi , ti when xi is a row of X. It is not too difficult to show that for a random row xi , E[hxi , t̂i] & k(γ − α), but it is not necessarily true that hxi , ti is large for every row xi . The problem is that for relevant choices of α, a random dataset has many more than k marginals that are within α of being in the top-k, and the algorithm could choose a subset of k of these to prevent a particular row xi from being traced. For example, if there are 3k columns of X that could be chosen in an α-accurate top-k vector, then with high probability, there exists a vector t specifying k of the columns on which xi = 0, which ensures that hxi , ti = 0. We can, however, show that hxi , t̂i > τ for at least (1 − c)n rows of X for an arbitrarily small constant c > 0. This weaker tracing guarantee is still enough to rule out (ε, δ)-differential privacy for any reasonable setting of ε, δ (Lemma 2.5), which gives us Theorem 1.1. The exact statement and parameters are slightly involved, so we refer the reader to Section 4 for a precise statement and analysis of our tracing attack in the case of approximate top-k statistics (Theorem 4.1). 3 2 Preliminaries Definition 2.1 (Differential Privacy). For ε ≥ 0, ρ ∈ [0, 1] we say that a randomized algorithm M : {±1}n×d → R is (ε, δ)-differentially private if for every two datasets X, X 0 ∈ {±1}n×d , such that X, X 0 differ in at most one row, we have that,   P [M(X) ∈ S] ≤ eε · P M(X 0 ) ∈ S + δ. ∀S ⊆ R Definition 2.2 (Marginals). For a dataset X ∈ {±1}n×d , its marginal vector q(X) = (q1 (X), . . . , qd (X)) is P the average of the rows of X. That is, qj (X) = n1 ni=1 Xi,j . We use the notation q(1) (X) ≥ q(2) (X) ≥ . . . ≥ q(d) (X) to refer to the sorted marginals. We will also define π : [d] → [d] to be the lexicographically first permutation that puts the marginals in sorted order. That is, we define π so that qπ(j) = q(j) and if j < j 0 are such that qj = qj 0 , then π(j) < π(j 0 ). Definition 2.3 (Accurate Top-k Vector). Given a dataset X ∈ {±1}n×d and a parameter α ≥ 0, a vector t̂ ∈ {0, 1}d is an α−accurate top-k vector of X if, 1. t̂ has exactly k non-zero coordinates, and 2. (t̂i = 1) ⇒ (qi (X) ≥ q(k) (X) − α). When α = 0, we define the exact top-k vector of X as t(X) ∈ {0, 1}d to be the lexicographically first 0-accurate top-k vector.3 Specifically, we define t(X) so that (t(X)j = 1) ⇔ j ∈ {π(1), . . . , π(k)}. We refer to these set of columns as the top-k columns of X. For comparison with our results, we state a positive result for privately releasing an αapproximate top-k vector, which is an easy consequence of the exponential mechanism and composition theorems for differential privacy Theorem 2.4. For every n, d, k ∈ N, and ε, δ, β ∈ (0, 1), there is an (ε, δ)-differentially private algorithm that takes as input a dataset X ∈ {±1}n×d , and with probability at least 1 − β, outputs an α-accurate top-k vector of X, for p   k · ln(1/δ) · ln(kd/β)   α = O   εn 2.1 Tracing Attacks Intuitively, tracing attacks violate differential privacy because if the target individual y is outside the dataset, then A(y, M(X)) reports OUT with high probability, whereas if y were added to the dataset to obtain X 0 , A(y, M(X 0 )) reports IN with high probability. Therefore M(X), M(X 0 ) must have very different distributions, which implies that M is not differentially private. The next lemma formalizes and quantifies this property. 3 Due to ties, there may not be a unique 0-accurate top-k vector of X. For technical reasons we let t(X) be the unique lexicographically first 0-accurate top-k vector, so we are justified in treating t(X) as a function of X. 4 Lemma 2.5 (Tracing Violates DP). Let M : {±1}n×d → R be a (possibly randomized) algorithm. Suppose there exists an algorithm A : {±1}d × R → {IN, OUT} such that when X ∼ {±1}n×d and y ∼ {±1}d are independent and uniformly random, 1. (Soundness) P [A(y, M(X)) = IN] ≤ ρ 2. (Completeness) P [# {i | A(xi , M(X)) = IN} ≥ n − m] ≥ 1 − ρ. Then M is not (ε, δ)-differentially private for any ε, δ such that eε ρ + δ < 1 − ρ − m n . If ρ < 1/4 and m > 3n/4, then there are absolute constants ε0 , δ0 > 0 such that M is not (ε0 , δ0 )-differentially private. The constants ε0 , δ0 can be made arbitrarily close to 1 by setting ρ and m/n to be appropriately small constants. Typically differentially private algorithms typically satisfy (ε, δ)-differential privacy where ε = o(1), δ = o(1/n), so ruling out differential privacy with constant (ε, δ) is a strong lower bound. 2.2 Probabilistic Inequalities We will make frequent use of the following concentration and anticoncentration results for sums of independent random variables. Lemma 2.6 (Hoeffding Bound). Let Z1 , . . . , Zn be independent random variables supported on {±1}, and P let Z = n1 ni=1 Zi . Then ∀ν > 0 1 2 P [Z − E[Z] ≥ ν] ≤ e− 2 ν n . Hoeffding’s bound on the upper tail also applies to random variables that are negative-dependent, which in this case means that setting any set of the variables B to +1 only makes the variables in [n] \ B more likely to be −1 [PS97]. Similarly, if the random variables are positive-dependent (their negations are negative-dependent), then Hoeffding’s bound applies to the lower tail. Theorem 2.7 (Chernoff Bound). Let Z1 , . . . , Zn be a sequence of independent {0, 1}-valued random P variables, let Z = ni=1 Zi , and let µ = E[Z]. Then ν2 1. (Upper Tail) ∀ν > 0 P [Z ≥ (1 + ν)µ] ≤ e− 2+ν µ , and 2. (Lower Tail) ∀ν ∈ (0, 1) P [Z ≤ (1 − ν)µ] ≤ e− 2 ν µ . 1 2 Theorem 2.8 (Anticoncentration [LT13]). Let Z1 , . . . , Zn be independent and uniform in {±1}, and let P Z = n1 ni=1 Zi . Then for every β > 0, there exists Kβ > 1 such that for every n ∈ N, Kβ 1 ∀v ∈ √ , n Kβ " 3 # P [Z ≥ ν] ≥ e− 1+β 2 2 ν n . Tracing Using the Top-k Vector Given a (possibly approximate) top-k vector t of a dataset X, and a target individual y, we define the following inner product attack. 5 Aρ,d,k (y, t) : p Input: y ∈ {±1}d and t ∈ {0, 1}d . Let τ = 2k ln(1/ρ). If hy, ti > τ, output IN; else output OUT. In this section we will analyze this attack when X ∈ {±1}n×d is a uniformly random matrix, and t = t(X) is the exact top-k vector of X. In this case, we have the following theorem. Theorem 3.1. There is a universal constant C ∈ (0, 1) such that if ρ > 0 is any parameter and n, d, k ∈ N satisfy d ≤ 2Cn , k ≤ Cd and k ln(d/2k) ≥ 8n ln(1/ρ), then Aρ,d,k has the following properties: If X ∼ {±1}n×d , y ∼ {±1}d are independent and uniform, and t(X) is the exact top-k vector of X, then h i 1. (Soundness) P Aρ,d,k (y, t(X)) = IN ≤ ρ, and h i 2. (Completeness) for every i ∈ [n], P Aρ,d,k (xi , t(X)) = OUT < ρ + e−k/4 . We will prove the soundness and completeness properties separately in Lemmas 3.2 and 3.3, respectively. The proof of soundness is straightforward. Lemma 3.2 (Soundness). For every ρ > 0, n ∈ N, and k ≤ d ∈ N, if X ∼ {±1}n×d , y ∼ {±1}d are independent and uniformly random, and t(X) is the exact top-k vector, then q   P hy, t(X)i ≥ 2k ln(1/ρ) ≤ ρ. p Proof. Recall that τ := 2k ln(1/ρ). Since X, y are independent, we have P [hy, t(X)i ≥ τ] h i P hy, t(X)i ≥ τ t(X) = IT · P [t(X) = IT ] X,y = X T ⊆[d]:|T |=k = X T ⊆[d]:|T |=k ≤ max T ⊆[d]:|T |=k X,y X   X    P  yj ≥ τ  · P [t(X) = IT ]  X y  j∈T   X    P  yj ≥ τ   y  (X, y are independent) j∈T For every fixed T , the random variables {yj }j∈T are independent and uniform on {±1}, so by Hoeffding’s inequality,   q X    P  yj ≥ 2k ln(1/ρ) ≤ ρ.   j∈T This completes the proof of the lemma. 6 We now turn to proving the completeness property, which will following immediately from the following lemma. Lemma 3.3 (Completeness). There is a universal constant C ∈ (0, 1) such that for every ρ > 0, n ∈ N, d ≤ 2Cn , and k ≤ Cd, if X ∼ {±1}n×d is chosen uniformly at random, t(X) is the exact top-k vector, and xi is any row of X, r   q   ln(d/2k) P hxi , t(X)i ≤ k − 2k ln(1/ρ) ≤ ρ + e−k/4 . n To see how the completeness p property of p Theorem 3.1 follows from the lemma, observe if k ln(d/2k) ≥ 8n ln(1/ρ), then k ln(d/2k)/n − 2k ln(1/ρ) ≥ τ. Therefore Lemma 3.3 implies that h i −k/4 P [hxi , t(X)i < τ] ≤ ρ + e , so P Aρ,d,k (xi , t(X)) = IN ≥ 1 − ρ − e−k/4 . Before proving the lemma, we will need a few claims about the distribution of hxi , t(X)i. The first claim asserts that, although X ∈ {±1}n×d is uniform, the k columns of X with the largest marginals are significantly biased. Claim 3.4. There is a universal constant C ∈ (0, 1), such that for every n ∈ N, d ≤ 2Cn and k ≤ Cd, if X ∈ {±1}n×d is drawn uniformly at random, then r     ln(d/2k)  ≤ e−k/4 . P q(k) (X) <  n Proof of Claim 3.4. For every j ∈ [d], define Ej to be the event that 1X qj = xij > n r i∈[n] ln(d/2k) . n We would like to apply Theorem 2.8 to the random variable " p K 1 ln (d/2k) /n ∈ √1 , n K1 1P n i xij . To do so, we need # where K1 is the universal constant from that theorem (applied with β = 1). These inequalities will be satisfied as long as d ≤ 2Cn , and k ≤ Cd for a suitable universal constant C ∈ (0, 1). Applying Theorem 2.8 gives r   h i  ln(d/2k)  2k  ∀j ∈ [d] P Ej = P qj > .  ≥ n d P By linearity of expectation, we have that E[ j Ej ] ≥ 2k. Since the columns of X are independent, and the events Ej only depend on a single column of X, the events Ej are also independent. P Therefore, we can apply a Chernoff bound (Theorem 2.7) to j Ej to get   d X    P  Ej < k  ≤ e−k/4 .   j=1 q P If j Ej ≥ k, then there exist k values qj that are larger than value. This completes the proof of the claim. 7 ln(d/2k) , n so q(k) is also at least this The previous claim establishes that if we restrict X to its top-k columns, the resulting matrix Xt ∈ {±1}n×k is a random matrix whose mean entry is significantly larger than 0. This claim is enough to establish that the inner product hxi , t(X)i is large in expectation over X. However, since XT its columns are not necessarily independent, which prevents us from applying concentration to get the high probability statement we need. However, the columns of Xt are independent if we condition on the value and location of the (k + 1)-st marginal. Claim 3.5. Let X ∈ {±1}n×d be a random matrix from a distribution with independent columns, and let
t(X) be its marginals. For every q ∈ [−1, 1], k, j ∈ [d], T ∈ [d] , the conditional distribution k X | (q(k+1) = q) ∧ (π(k + 1) = j) ∧ (t(X) = IT ) also has independent columns. Proof of Claim 3.5. Suppose we condition on the value of the (k + 1)-st marginal, q(k+1) = q, its location, π(k + 1) = j, and the set of top-k marginals t = IT . By definition of the (exact, lexicographically first) top-k vector, we have that if ` < j, then ` ∈ T if and only if q` ≥ q. Similarly, if ` > j, then ` ∈ T if and only if q` > q. Since we have conditioned on a fixed tuple (q, j, T ), the statements q` > q and q` ≥ q now depend only on the `-th column. Thus, since the columns of X are independent, they remain independent even when condition on any tuple (q, j, T ). Specifically, if ` < j and ` ∈ T , P then column ` is drawn independently from the conditional distribution ((u1 , . . . , un ) | n1 i ui ≥ q), where (u1 , . . . , un ) ∈ {±1}n are chosen independently and uniformly at random. Similarly, if ` > j P and ` ∈ T , then column ` is drawn independently from ((u1 , . . . , un ) | n1 i ui > q). Now we are ready to prove Lemma 3.3. q Proof of Lemma 3.3. For convenience, define γ = of X. We can write ln(d/2k) n p and τc = kγ − 2k ln(1/ρ). Fix any row xi P [hxi , t(X)i < τc ] h i h i ≤ P hxi , t(X)i < τc q(k+1) ≥ γ + P q(k+1) < γ h i ≤ P hxi , t(X)i < τc q(k+1) ≥ γ + e−k/4 (Claim 3.4) h i ≤ max P hxi , t(X)i < τc (q(k+1) = q) ∧ (π(k + 1) = j) ∧ (t(X) = IT ) + e−k/4 (1) q≥γ,j∈[d],T ∈([d] k ) Let Gq,j,T be the event (q(k+1) = q) ∧ (π(k + 1) = j) ∧ (t(X) = IT ). By linearity of expectation, we can write h i E hxi , t(X)i < τc Gq,j,T ≥ kq ≥ kγ. P Using Claim 3.5, we have that `:t(X)` =1 xi` conditioned on Gδ,j,T is a sum of independent {±1}valued random variables. Thus, q   h i P hxi , t(X)i < τc | Gq,j,T = P hxi , t(X)i < kγ − 2k ln(1/ρ) | Gq,j,T q   ≤ P hxi , t(X)i < kq − 2k ln(1/ρ) | Gq,j,T (q ≥ γ) (Hoeffding) ≤ρ Combining with (1) completes the proof. 8 4 Tracing Using an Approximate Top-k Vector In this section we analyze the inner product attack when it is given an arbitrary approximate top-k vector. Theorem 4.1. For every ρ > 0, there exist universal constants C, C 0 ∈ (0, 1) (depending only on ρ) such that if d ∈ N is sufficiently large and n, d, k ∈ N and α ∈ (0, 1) satisfy r ln(d/2k) Cn C 0 d ≤ 2 , k ≤ d , n = C k ln(d/2k), and α ≤ C , n and t̂ : {±1}n×d → {0, 1}d is any randomized algorithm such that h i ∀X ∈ {±1}n×d P t̂(X) is an α-approximate top-k vector for X ≥ 1 − ρ, then Aρ,d,k (Section 3) has the following properties: If X ∼ {±1}n×d , y ∈ {±1}d are independent and uniform, then h i 1. (Soundness) P Aρ,d,k (y, t(X)) = IN ≤ ρ, and h n o i 2. (Completeness) P # i ∈ [n] Aρ,d,k (xi , t̂(X)) = IN < (1 − e2 ρ)n < 2ρ + 2e−k/6 . The proof of the soundness property is nearly identical to the case of exact statistics so we will focus only on proving the completeness property. Intuitively, the proof of completeness takes the same general form as it did for the case of exact top-k statistics. First, for some parameter γ > 0, with high probability X has at least k marginals that are at least γ. Therefore, any marginal j contained in any α-accurate top-k vector has value qj ≥ λ := γ − α. From this, we can conclude that the expectation of hxi , t̂(X)i ≥ k(γ − α) where the expectation is taken over the choices of X, t̂(X), and i. However, unlike the case of the exact marginals, the k columns selected by t̂ may be significantly correlated so that for some choices of i, hxi , t̂(X)i is small with high probability. At a high level we solve this problem as follows: first, we restrict to the set of dλ columns j such that qj ≥ γ − α, which remain mutually independent. Then we argue that for every fixed α-accurate top-k vector specifying a subset of k of these columns, with overwhelming probability the inner product is large for most choices of i ∈ [n]. Finally, we
take a union bound over all dkλ possible choices of α-accurate top-k vector. To make the union bound tolerable, we need that with high probability dλ is not too big. Our choice of γ was such that only about k columns are above γ, therefore if we take λ very close to γ, we will also be able to say that dλ is not too much bigger than k. By assuming that the top-k vector is α-accurate for α  γ, we get that λ = γ − α is very close to γ. Before stating the exact parameters and conditions in Lemma 4.4, we will need to state and prove a few claims about random matrices. Claim 4.2. For every β > 0, there is a universal constant C ∈ (0, 1) (depending only on β), such that q for every n ∈ N, d ≤ 2Cn and k ≤ Cd, if X ∈ {±1}n×d is drawn uniformly at random, then for γ := 2 1+β · ln(d/2k) , n we have h i P q(k) (X) < γ ≤ e−k/4 . The above claim is just a slightly more general version of Claim 3.4 (in which we have fixed β = 1), so we omit its proof. 9 Claim 4.3. For every n, d ∈ N and every λ ∈ (0, 1), if X ∈ {±1}n×d is drawn uniformly at random, then for dλ := 2d exp(− 12 λ2 n) h n o i P # j qj > λ > dλ ≤ e−dλ /6 . Proof of Claim 4.3. For every j ∈ [d], define Ej to be the event that qj = P are independent, applying Hoeffding’s bound to i xij gives, ∀j ∈ [d] 1P n i xij > λ. Since the xij ’s h i h i 2 P Ej = P qj > λ ≤ e−λ n/2 . P 2 By linearity of expectation, we have that E[ j Ej ] ≤ de−λ n/2 = 12 dλ . Since the columns of X are P independent, we can apply a Chernoff bound (Theorem 2.7) to j Ej , which gives P hP d i −dλ /6 E > d . λ ≤e j=1 j This completes the proof of the claim. Now we are ready to state our exact claim about the completeness of the attack when given an α-accurate top-k vector. Lemma 4.4 (Completeness). For every ρ > 0, there exist universal constants C2 , C3 , C4 , C5 ∈ (0, 1) (depending only on ρ) such that if n, d, k ∈ N and α ∈ (0, 1) satisfy, r ln(2d) , 4k ≤ min{(2d)C2 , 4C4 d}, 8n ln(1/ρ) = C32 k ln(2d), d ≤ 2C4 n α ≤ C5 n and t̂ is an algorithm that, for every X ∈ {±1}n×d , outputs an α-accurate top-k vector with probability at least 1 − ρ, then for a uniformly random X ∈ {±1}n×d , we have h n o i P # i ∈ [n] | hxi , t̂(X)i ≥ τc < (1 − e2 ρ)n < 2ρ + e−k/4 + e−k/6 , q where τc := C3 k ln(2d) n p − 2k ln(1/ρ). To see how the completeness property of Theorem 4.1 follows from the lemma, observe if 8n ln(1/ρ) = C32 k ln(2d), then r τc = C 3 k ln(2d) − n q 2k ln(1/ρ) = q 2k ln(1/ρ) = τ where τ is the threshold in Aρ,d,k . Therefore Lemma 4.4 implies that h n o i P # i ∈ [n] Aρ,d,k (xi , t̂(X)) = IN < (1 − e2 ρ)n < ρ + e−k/4 + e−k/6 . The universal constants C, C 0 will be C = min{C2 , C4 , C5 } − δ for an arbitrarily small δ > 0, and = C32 . As long as d is sufficiently large the conditions k ≤ d C in Theorem 4.1 will imply the corresponding condition in the above lemma. C0 10 Proof of Lemma 4.4. First, we will condition everything on the event Gα := {t̂ = t̂(X) is an α-accurate top-k vector of X}. By assumption, for every X ∈ {±1}n×d , P [Gα ] ≥ 1 − ρ. For convenience define the constant c := e2 ρ, so that the lemma asserts that, with high probability, A(xi , t̂(X)) = IN for at least (1 − c)n rows xi . As in the proof of completeness for the case of exact top-k, we will first condition on the event that at least k marginals are above the threshold γ. Now, by Claim 4.2, with an appropriate choice of r s ln(d/2k) 2 c γ := , β := · c 16 ln(1/ρ) 1 + 16 ln(1/ρ) n and the assumptions that k ≤ C4 d and d ≤ 2C4 n for some universal constant C4 depending only on β, the event   r     ln(2d)   Gγ :=  q (X) ≥ γ = C , (k) 1    n  will hold with probability 1 − e−k/4 . Here we define the universal constants s 2 c C1 := C2 := . c 1 + 8 ln(1/ρ) 2c + 4 ln(1/ρ) depending only on ρ. These constants were chosen so that provided 4k ≤ (2d)C2 , the inequality in the definition of Gγ will be satisfied. In light of the above analysis, we condition the rest of the analysis on the event Gγ , which is h i satisfies P Gγ ≥ 1 − e−k/4 . If we condition on Gα and Gγ , then for any marginal j chosen by t̂ (i.e. t̂j = 1), then we can say that qj ≥ λ for any λ ≤ γ − α. Now, we define the constants s C3 := 2 c 1 + 4 ln(1/ρ) C5 := C1 − C3 > 0, where one can −C3 > 0 holds for all choices of c. Now by our assumption q verify that the inequality C1q that α < C5 ln(2d) n , we can define λ := C3 ln(2d) n . For any matrix X ∈ {±1}n×d , we can define Sλ = Sλ (X) ⊆ {1, . . . , d} to be the set of columns of X whose marginals are greater than λ. The analysis above says that, conditioned onGγ and Gα , if t̂j = 1, then j ∈ Sλ . Note that, if X is chosen uniformly at random, and we define X≥λ ∈ {±1}n×|Sλ | to be the restriction of X to the columns contained in Sλ , then the columns of X≥λ remain independent. The size of Sλ is a random variable supported on {0, 1, . . . , d}. In our analysis we will need to condition on the even that |Sλ |  d. Using Claim 4.3 we have that if dλ := 2de− λ2 n 2 then the event GS := {|Sλ (X)| ≤ dλ } 11 satisfies P [GS ] ≥ 1 − e−dλ /6 ≥ 1 − e−k/6 where we have used the fact that dλ ≥ k. This fact is not difficult to verify form our choice of parameters. Intuitively, since λ ≤ γ, and there are at least k marginals larger than γ, there must also typically be at least k marginals larger than λ. We condition the remainder of the analysis on the event GS . Later in the proof we require that the size of Sλ is small with high probability. Using Claim 4.3 λ2 n we can say that the size of Sλ is at most dλ = 2de− 2 with probability at least, 1 − e−dλ /6 . When q(k) ≥ γ, the number of marginals greater than λ would be at least k. So dλ > k and the error probability e−dλ /6 is at most e−k/6 . We will henceforth condition on the event that |Sλ (X)| ≤ dλ . We will say that p the attack A fails on t̂ when we fail to trace more than cn rows, i.e. A fails when {i : hxi , t̂i < kλ − 2k ln(1/ρ)} > cn = e2 ρn. Formally we have that, h i h i h i P A fails on t̂ ≤ P A fails on t̂ ∧ Gα ∧ Gγ ∧ GS + P ¬Gα ∨ ¬Gγ ∨ ¬GS h i ≤ P A fails on t̂ ∧ Gα ∧ Gγ ∧ GS + ρ + e−k/4 + e−k/6 (2) Thus, to complete the proof, it suffices to show that h i h i P A fails on t̂ ∧ Gα ∧ Gγ ∧ GS = P (A fails on t̂) ∧ (t̂ is α-accurate) ∧ (q(k) ≥ γ) ∧ (|Sλ | ≤ dλ ) h i ≤ P (A fails on t̂) ∧ (t̂ ⊆ Sλ ) ∧ (|Sλ | ≤ dλ ) " ! ! # Sλ ≤ P ∃v ∈ A fails on v ∧ (|Sλ | ≤ dλ ) k
where we have abused notation and written t̂ ⊆ Sλ to mean that t̂j = 1 =⇒ j ∈ Sλ , and used v ∈ Skλ to mean that v is a subset of Sλ of size k. h exactly i Sλ
We will now upper bound P (∃v ∈ k A fails on v) ∧ (|Sλ | ≤ dλ ) . Observe that, since the columns of X are identically distributed, this probability is independent of the specific choice of Sλ and depends only on |Sλ |. Further, decreasing the size of Sλ only decreases the probability. Thus, we will fix a set S of size exactly assume Sλi = S. Thus, for our canonical choice of set h dλ and S
S = {1, . . . , dλ }, we need to bound P ∃v ∈ k A fails on v . Consider a fixed vector v ⊆ S. That is, a vector v ∈ {0, 1}d such that vj = 1 =⇒ j ∈ S. Define the event Ei,v to be the event that hxi , vi is too small for some specific row i and some specific vector v ⊆ S. That is, q   Ei,v := hxi , vi < τc := kλ − 2k ln(1/ρ) . Since the columns of XS are independent, for a fixed i and v, by Hoeffding’s inequality gives hP i   P Ei,v = P j:vj =1 xi,j < τc ≤ ρ. We have proved that the probability that hxi , vi is small, is small for a given row. We want to bound the probability that hxi , vi is small for an entire set of rows R ⊆ [n]. Unfortunately, since we require that qj ≥ λ for every column j ∈ S, the rows xi are no longer independent. However, the rows satisfy a negative-dependence condition, captured in the following claim. Claim 4.5. For every R ⊆ [n],   ^  P  Ei,v  ≤ ρ|R| . i∈R 12 To maintain the flow of the analysis, we defer the proof of this claim to Section 4.1 By definition, A fails on v only if there exists a set R of exactly cn = e2 ρn rows such that Taking a union bound over all such sets R and all v , we have " ! # ! ! S dλ n P ∃v ∈ A fails on v ≤ · · ρcn k k cn !k   enρ cn edλ · ≤ k cn V i∈R Ei,v . ≤ dλk · e−cn where we have used the identity 2 − λ2 n a
b b ≤ ( ea b ) . We have already set the parameter λ, and set dλ = 2de . Thus, all that remains is to show that for our choice of parameters dλk · e−cn ≤ ρ, which is equivalent to cn ≥ ln(1/ρ) + k ln(dλ ). Substituting our choice of λ gives the condition kλ2 n ≥ ln(1/ρ) + k ln(2d) − cn 2 C32 k ln(2d) , 8 ln(1/ρ) One can check that, for our choice of n = and our choice of λ = C3 been defined above, the preceding equation is satisfied. Thus, we have established that " ! # S P ∃v ∈ A fails on v ≤ dλk · e−cn ≤ ρ. k q ln(2d) n where C3 has As we have argued above, this implies that h i P A fails on t̂ ≤ 2ρ + e−k/4 + e−k/6 This completes the proof of the completeness lemma. 4.1 Proof of Claim 4.5 Recall that, for a given X ∈ {±1}n×d , Ei,v is the event that hxi , vi < τc for a specific row i and a specific vector v ⊆ S, where S = Sλ is the set of columns j of X such that qj ≥ λ. Thus, we can think of XS ∈ {±1}n×|S| as a matrix with |S| independent columns that are uniformly random subject to the constraint that each column’s mean is at least λ. Since, flipping some entries of XS from −1 to +1 eS in which each column’s mean is can only increase hxi , vi, we will in fact use the distribution X exactly λn. Thus, when we refer to the probabilities of events involving random variables xi,j , we will use this distribution on XS as the probability space. Additionally, since v is fixed, and the probability is the same for all v, we will simply write Ei to cut down on notational clutter. For a specific set R ⊆ [n], we need to calculate   ^  P  Ei  . XS  i∈R 13 We can write       ^ X ^     P  Ei  = P   xij < τc     e e X X S i∈R S i∈R j∈v    X    ≤ P  xij < |R|τc  .  eS  X (3) i∈R,j∈v The key property of XS is that its entries xij are positively correlated. That is, for every set I ⊂ [n] × S of variables xij , we have, h i Y h i P ∀(i, j) ∈ I xij = −1 ≤ P xij = −1 . (4) eS X (i,j)∈I eS X eS are independent if we partition the elements of I into sets I1 , . . . , Ik , Since the columns of X where each set Il has pairs of I which come from the column l, then, i h i Y h P ∀(i, j) ∈ I xij = −1 = P ∀(i, j) ∈ Il xij = −1 . eS X l∈[k] eS X n o So it is enough to show that equation 4 holds when I = (i1 , l), . . . , (ip , l) . For simplicity of notation we will refer to these elements of I as {1, . . . , p}. We have that, p Y h i P [∀a ∈ I xa = −1] = P xa = −1 (∀b ∈ {1, . . . , a − 1} , xb = −1) . (5) a=1 We will show that each of the terms in the product is smaller than P [xa = −1]. For a fixed a ∈ I, let B be the set {1, . . . , a − 1} and let E be the event that (∀b ∈ B, xb = −1). Since every column of XS sums to nλ, we have    X    E  xil B = nλ.   i∈[n] On the other hand, since the bits in B are all set to −1 and all the other bits in column l are equal in expectation,    X   h i   E  xil B = −|B| + (n − |B|) · E xa B ,   i∈[n] which means that E[xa B] ≥ λ = E[xa ]. h i Since P [xa = −1] = (1 − E[xa ])/2, we get that P xa = −1 B ≤ P [xa = −1]. Substituting this back into (5), we get that the variables are positively correlated. We have that,    X    E  xij  = |R|kλ,   i∈R,j∈v 14 and since Hoeffding’s inequality applies equally well to positively-correlated random variables [PS97], we also have   p       q  |R| 2k ln(1/ρ) 2    X   X        = ρ|R| . P  xij ≤ |R|τc  ≤ P  xij < |R|kλ − |R| 2k ln(1/ρ) ≤ exp −      2|R|k e e  XS XS i∈R,j∈v i∈R,j∈V Substituting this in equation 3, we get that,   ^   P  Ei  ≤ ρ|R| . e X S i∈R eS , we have Finally, we use the fact that, by our definition of the distributions XS , X     ^  ^  P  Ei  ≤ P  Ei  ≤ ρ|R| . XS  e X i∈R S i∈R This completes the proof. Acknowledgements We are grateful to Adam Smith and Thomas Steinke for many helpful discussions about tracing attacks and private top-k selection. References [BDMN05] Avrim Blum, Cynthia Dwork, Frank McSherry, and Kobbi Nissim. Practical privacy: the SuLQ framework. In PODS, 2005. [BLST10] Raghav Bhaskar, Srivatsan Laxman, Adam Smith, and Abhradeep Thakurta. Discovering frequent patterns in sensitive data. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 503–512. ACM, 2010. [BNS+ 16] Raef Bassily, Kobbi Nissim, Adam Smith, Thomas Steinke, Uri Stemmer, and Jonathan Ullman. Algorithmic stability for adaptive data analysis. In Proceedings of the 48th Annual ACM SIGACT Symposium on Theory of Computing, pages 1046–1059. ACM, 2016. [BS98] Dan Boneh and James Shaw. Collusion-secure fingerprinting for digital data. IEEE Transactions on Information Theory, 44(5):1897–1905, 1998. [BUV14] Mark Bun, Jonathan Ullman, and Salil P. Vadhan. Fingerprinting codes and the price of approximate differential privacy. In STOC, 2014. [DL09] Cynthia Dwork and Jing Lei. Differential privacy and robust statistics. In Proceedings of the forty-first annual ACM symposium on Theory of computing, pages 371–380. ACM, 2009. 15 [DMNS06] Cynthia Dwork, Frank McSherry, Kobbi Nissim, and Adam Smith. Calibrating noise to sensitivity in private data analysis. In TCC, 2006. [DMT07] Cynthia Dwork, Frank McSherry, and Kunal Talwar. The price of privacy and the limits of lp decoding. In Proceedings of the thirty-ninth annual ACM symposium on Theory of computing, pages 85–94. ACM, 2007. [DN03] Irit Dinur and Kobbi Nissim. Revealing information while preserving privacy. In PODS, 2003. [DN04] Cynthia Dwork and Kobbi Nissim. Privacy-preserving datamining on vertically partitioned databases. In CRYPTO, 2004. [DNPR10] Cynthia Dwork, Moni Naor, Toniann Pitassi, and Guy N Rothblum. Differential privacy under continual observation. In Proceedings of the forty-second ACM symposium on Theory of computing, pages 715–724. ACM, 2010. [DSS+ 15] Cynthia Dwork, Adam D. Smith, Thomas Steinke, Jonathan Ullman, and Salil P. Vadhan. Robust traceability from trace amounts. In FOCS, 2015. [DSSU17] Cynthia Dwork, Adam Smith, Thomas Steinke, and Jonathan Ullman. Exposed a survey of attacks on private data. Annual Review of Statistics and Its Application, 2017. [DSZ15] Cynthia Dwork, Weijie Su, and Li Zhang. Private false discovery rate control. arXiv preprint arXiv:1511.03803, 2015. [DY08] Cynthia Dwork and Sergey Yekhanin. New efficient attacks on statistical disclosure control mechanisms. In Advances in Cryptology - CRYPTO 2008, 28th Annual International Cryptology Conference, Santa Barbara, CA, USA, August 17-21, 2008. Proceedings, pages 469–480, 2008. [GLM+ 10] Anupam Gupta, Katrina Ligett, Frank McSherry, Aaron Roth, and Kunal Talwar. Differentially private combinatorial optimization. In Proceedings of the twenty-first annual ACM-SIAM symposium on Discrete Algorithms, pages 1106–1125. Society for Industrial and Applied Mathematics, 2010. [GRU12] Anupam Gupta, Aaron Roth, and Jonathan Ullman. Iterative constructions and private data release. In TCC, 2012. [HR10] Moritz Hardt and Guy N. Rothblum. A multiplicative weights mechanism for privacypreserving data analysis. In FOCS, 2010. [HSR+ 08] Nils Homer, Szabolcs Szelinger, Margot Redman, David Duggan, Waibhav Tembe, Jill Muehling, John V Pearson, Dietrich A Stephan, Stanley F Nelson, and David W Craig. Resolving individuals contributing trace amounts of dna to highly complex mixtures using high-density snp genotyping microarrays. PLoS genetics, 4(8):e1000167, 2008. [JT12] Prateek Jain and Abhradeep Thakurta. Mirror descent based database privacy. In Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, pages 579–590. Springer, 2012. 16 [KRS13] Shiva Prasad Kasiviswanathan, Mark Rudelson, and Adam Smith. The power of linear reconstruction attacks. In Proceedings of the Twenty-Fourth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 1415–1433. Society for Industrial and Applied Mathematics, 2013. [KRSU10] Shiva Prasad Kasiviswanathan, Mark Rudelson, Adam Smith, and Jonathan Ullman. The price of privately releasing contingency tables and the spectra of random matrices with correlated rows. In Proceedings of the 42nd ACM Symposium on Theory of Computing, STOC 2010, Cambridge, Massachusetts, USA, 5-8 June 2010, pages 775–784, 2010. [LT13] Michel Ledoux and Michel Talagrand. Probability in Banach Spaces: isoperimetry and processes. Springer Science & Business Media, 2013. [MT07] Frank McSherry and Kunal Talwar. Mechanism design via differential privacy. In Foundations of Computer Science, 2007. FOCS’07. 48th Annual IEEE Symposium on, pages 94–103. IEEE, 2007. [NTZ13] Aleksandar Nikolov, Kunal Talwar, and Li Zhang. The geometry of differential privacy: the sparse and approximate cases. In STOC, 2013. [PS97] Alessandro Panconesi and Aravind Srinivasan. Randomized distributed edge coloring via an extension of the chernoff–hoeffding bounds. SIAM Journal on Computing, 26(2):350–368, 1997. [RR10] Aaron Roth and Tim Roughgarden. Interactive privacy via the median mechanism. In STOC, pages 765–774. ACM, June 5–8 2010. [SOJH09] Sriram Sankararaman, Guillaume Obozinski, Michael I Jordan, and Eran Halperin. Genomic privacy and limits of individual detection in a pool. Nature genetics, 41(9):965– 967, 2009. [ST13] Adam Smith and Abhradeep Thakurta. Differentially private model selection via stability arguments and the robustness of the lasso. J Mach Learn Res Proc Track, 30:819–850, 2013. [SU15] Thomas Steinke and Jonathan Ullman. Between pure and approximate differential privacy. CoRR, abs/1501.06095, 2015. [Tar08] Gábor Tardos. Optimal probabilistic fingerprint codes. Journal of the ACM (JACM), 55(2):10, 2008. [Ull13] Jonathan Ullman. Answering n2+o(1) counting queries with differential privacy is hard. In STOC, 2013. [Ull15] Jonathan Ullman. Private multiplicative weights beyond linear queries. In PODS, 2015. 17 

Priming Neural Networks arXiv:1711.05918v2 [cs.CV] 17 Nov 2017 Amir Rosenfeld , Mahdi Biparva , and John K.Tsotsos Department of Electrical Engineering and Computer Science York University Toronto, ON, Canada, M3J 1P3 {amir@eecs, mhdbprv@cse,tsotsos@cse}.yorku.ca Abstract Visual priming is known to affect the human visual system to allow detection of scene elements, even those that may have been near unnoticeable before, such as the presence of camouflaged animals. This process has been shown to be an effect of top-down signaling in the visual system triggered by the said cue. In this paper, we propose a mechanism to mimic the process of priming in the context of object detection and segmentation. We view priming as having a modulatory, cue dependent effect on layers of features within a network. Our results show how such a process can be complementary to, and at times more effective than simple post-processing applied to the output of the network, notably so in cases where the object is hard to detect such as in severe noise. Moreover, we find the effects of priming are sometimes stronger when early visual layers are affected. Overall, our experiments confirm that top-down signals can go a long way in improving object detection and segmentation. Figure 1: Visual priming: something is hidden in plain sight in this image. It is unlikely to notice it without a cue on what it is (for an observer that has not seen this image before). Once a cue is given, perception is modified to allow successful detection. See footnote at bottom of this page for the cue, and supplementary material for the full answer. 1. Introduction ing, (2) priming and (3) pruning. Freely viewing the image, the default strategy, likely reveals nothing more than a dry grassy field near a house. Introducing a cue about a target in the image1 results in one of two possibilities. The first, also known as priming, is modification to the computation performed when viewing the scene with the cue in mind. The second, which we call pruning - is a modification to the decision process after all the computation is finished. When the task is to detect objects, this can mean retaining all detections match the cue, even very low confidence ones and discarding all others. While both are viable ways to incorporate the knowledge brought on by the cue, Psychophysical and neurophysiological studies of the human visual system confirm the abundance of top-down effects that occur when an image is observed. Such topdown signals can stem from either internal (endogenous) processes of reasoning and attention or external (exogenous) stimuli- i.e. cues - that affect perception (cf. [35], Chapter 3 for a more detailed breakdown). External stimuli having such effects are said to prime the visual system, and potentially have a profound effect on an observer’s perception. This often results in an “Aha!” moment for the viewer, as he/she suddenly perceives the image differently; Fig. 1 shows an example of such a case. We make here the distinction between 3 detection strategies: (1) free view- 1 Object 1 in image: Ø5O priming often highly increases the chance of detecting the cued object. Viewing the image for an unlimited amount of time and pruning the results is less effective; in some cases, detection is facilitated only by the cue. We claim that priming allows the cue to affect the visual process from early layers, allowing detection where it was previously unlikely to occur in free-viewing conditions. This has also recently gained some neurophysiological evidence [2]. In this paper, we propose a mechanism to mimic the process of visual priming in deep neural networks in the context of object detection and segmentation. The mechanism transforms an external cue about the presence of a certain class in an image (e.g., “person”) to a modulatory signal that affects all layers of the network. This modulatory effect is shown via experimentation to significantly improve object detection performance when the cue is present, more so than a baseline which simply applies post-processing to the network’s result. Furthermore, we show that priming early visual layers has a greater effect that doing so for deeper layers. Moreover, the effects of priming are shown to be much more pronounced in difficult images such as very noisy ones. The remainder of the paper is organized as follows: in Sec. 2 we go over related work from computers vision, psychology and neurophysiology. In Sec. 3 we go over the details of the proposed method. In Sec. 4 we elaborate on various experiments where we evaluate the proposed method in scenarios of object detection and segmentation. We finish with some concluding remarks. 2. Related Work Context has been very broadly studied in cognitive neuroscience [4, 3, 23, 37, 38, 24, 16] and in computer vision [12, 10, 34, 33, 27, 39, 22]. It is widely agreed [30] that context plays crucial role for various visual tasks. Attempts have been made to express a tangible definition for context due to the increased use in the computer vision community [34, 33] . Biederman et al. [4] hypothesizes object-environments dependencies into five categories: probability, interposition, support, familiar size, position. Combinations of some of these categories would form a source of contextual information for tasks such as object detection [33, 30], semantic segmentation [14], and pose estimation [6]. Context consequently is the set of sources that partially or collectively influence the perception of a scene or the objects within [32]. Visual cues originated from contextual sources, depending on the scope they influence, further direct visual tasks at either global or local level [34, 33]. Global context such as scene configuration, imaging conditions, and temporal continuity refers to cues abstracted across the whole scene. On the other hand, local context such as semantic relationships and local-surroundings characterize associations Input Image Feature Extraction Feature Extraction Feature Extraction Task Prediction Feature Extraction Task Prediction (a) Feedforward Input Image Feature Extraction Feature Extraction Cue (b) Pruning Input Image Feature Extraction Feature Extraction Feature Extraction Task Prediction Cue (c) Priming Figure 2: A neural network can be applied to an input in an either unmodified manner (top), pruning the results after running (middle) or priming the network via an external signal (cue) in image to affect all layers of processing (bottom). among various parts of similar scenes. Having delineated various contextual sources, the general process by which the visual hierarchy is modulated prior to a particular task is referred to as visual priming [35, 26]. A cue could be provided either implicitly by a contextual source or explicitly through other modalities such as language. There has been a tremendous amount of work on using some form of top-down feedback to contextually prime the underlying visual representation for various tasks [37, 38, 24, 16]. The objective is to have signals generated from some task such that they could prepare the visual hierarchy oriented for the primary task. [30] proposes contextual priming and feedback for object detection using the Faster R-CNN framework [29]. The intuition is to modify the detection framework to be able to generate semantic segmentation predictions in one stage. In the second stage, the segmentation primes both the object proposal and classification modules. Instead of relying on the same modality for the source of priming, [9, 25] proposes to modulate features of a visual hierarchy using the embeddings of the language model trained on the task of visual question answering [1, 17]. In other words, using feature-wise affine transformations, [25] multiplicatively and additively modulates hidden activities of the visual hierarchy using the top-down priming signals generated from the language model, while [30] append directly the semantic segmentation predictions to the visual hierarchy. Recently, [14] proposes to modulate convolutional weight parameters of a neural networks using segmentation-aware masks. In this regime, the weight parameters of the model are directly approached for the purpose of priming. Although all these methods modulate the visual representation, none has specifically studied the explicit role of category cues to prime the visual hierarchy for object detection and segmentation. In this work, we strive to introduce a consistent parametric mechanism into the neural network framework. The proposed method allows every portion of the visual hierarchy to be primed for tasks such as object detection and semantic segmentation. It should be noted that this use of priming was defined as part of the Selective Tuning (ST) model of visual attention [36]. Other aspects of ST have recently appeared as part of classification and localization networks as well [5, 41], and our work explores yet another dimension of the ST theory. 3. Approach Assume that we have some network N to perform a task such as object detection or segmentation on an image I. In addition, we are given some cue h ∈ Rn about the content of the image. We next describe pruning and priming, how they are applied and how priming is learned. We assume that h is a binary encoding of them presence of some target(s) (e.g, objects) - though this can be generalized to other types of information. For instance, an explicit specification of color, location, orientation, etc, or an encoded features representation as can be produced by a vision or language model. Essentially, one can either ignore this cue, use it to post-process the results, or use it to affect the computation. These three strategies are presented graphically in Fig. 2. Pruning. In pruning, N is fed an image and we use h to post-process the result. In object detection, all bounding boxes output by N whose class is different than indicated by h are discarded. For segmentation, assume N outputs a score map of size C × h × w , where L is the number of classes learned by the network, including a background class. We propose two methods of pruning, with complementary effects. The first type increases recall by ranking the target class higher: for each pixel (x,y), we set the value of all score maps inconsistent with h to be −∞ , except that of the background. This allows whatever detection of the hinted class to be ranked higher than other which previously masked it. The second type simply sets each pixels which was not assigned by the segmentation the target class to the background class. This decreases recall but increases the precision. These types of pruning are demonstrated in Fig. 8 and discussed below. Priming. Our approach is applicable to any network N with a convolutional structure, such as a modern network for object detection, e.g. [20]. To enable priming, we freeze all weights in N and add a parallel branch Np . The role of Np is to transform an external cue h ∈ Rn to modulatory signals which affect all or some of the layers of N . Namely, let Li be some layer of N. Denote the output of Li by xi ∈ Rci ×hi ×wi where ci is the number of feature planes and hi , wi are the height and width of the feature planes. Denote the jth feature plane of xi by xij ∈ Rhi ×wi . Np modulates each feature plane xij by applying to the fij (xij , h) = x̂ij (1) The function fij always operates in a spatially-invariant manner - for each element in a feature plane, the same function is applied. Specifically, we use a simple residual function, that is x̂ij = αij · xij + xij (2) Where the coefficients αi = [αi1 , . . . , αici ]T are determined by a linear transformation of the cue: αi = Wi ∗ h (3) An overall view of the proposed method is presented in Fig. 3. Types of Modulation The modulation in eq. 2 simply adds a calculated value to the feature plane. We have experimented with other types of modulation, namely nonresidual ones (e.g, purely multiplicative), as well as following the modulated features with a non-linearity (ReLU), or adding a bias term in addition to the multiplicative part. The single most important dominant ingredient to reach good performance was the residual formulation - without it, training converged to very poor results. The formulation in eq. 2 performed best without any of the above listed modifications. We note that an additive model, while having converged to better results, is not fully consistent with biologically plausible models ([36]) which involve suppression/selection of visual features, however, it may be considered a first approximation. Types of Cues The simplest form of a cue h is an indicator vector of the object(s) to be detected, i.e, a vector of 20 zeros and 1 in the coordinate corresponding to “horse”, assuming there are 20 possible object classes, such as in Pascal [11]. We call this a categorical cue because it explicitly carries semantic information about the object. This means that when a single class k is indicated, αi becomes the kth column of Wi . o + o W + Task Prediction + o W W Cue Figure 3: Overall view of the proposed method to prime deep neural networks. A cue about some target in the image is given by and external source or some form of feedback. The process of priming involves affecting each layer of computation of the network by modulating representations along the path. 3.1. Training Multiple Cues Per Image. Contemporary object detection and segmentation benchmarks [19, 11] often contain more than one object type per image. In this case, we may set each coordinate in h to 1 iff the corresponding class is present in the image. However, this tends to prevent Np from learning to modulate the representation of N in a way which allows it to suppress irrelevant objects. Instead, if an image contains k distinct object classes, we duplicate the training sample k times and for each duplicate set the ground truth to contain only one of the classes. This comes at the expense of a longer training time, depending on the average number k over the dataset. 4. Experiments We evaluate our method on two tasks: object detection and object class segmentation. In each case, we take a pre-trained deep neural network and explore how it is affected by priming or pruning. Our goal here is not necessarily to improve state-of-the-art results but rather to show how usage of top-down cues can enhance performance. Our setting is therefore different than standard object-detection/segmentation scenarios: we assume that some cue about the objects in the scene is given to the 0.8 0.865 0.6 mAP mAP To learn how to utilize the cue, we freeze the parameters of our original network N and add the network block Np . During training, with each training example (Ii , yi ) fed into N we feed hi into Np , where Ii is an image, yi is the ground-truth set of bounding boxes and hi is the corresponding cue. The output and loss functions of the detection network remain the same, and the error is propagated through the parameters of Np . Fig. 3 illustrates the network. Np is very lightweight with respect to N , as it only contains parametersPto transform from the size of the cue h to at most K = i ki where ki is the number of output feature planes in each layer of the network. H 0.870 0.860 0.855 0.4 0.2 0000 0001 0010 0011 0100 0111 1000 1100 1110 1111 H (a) 0 10 20 30 40 50 60 70 80 90 100 noise 1111 1110 1100 1000 0111 0100 0011 0010 0001 0000 (b) Figure 4: (a) Performance gains by priming different parts of the SSD objects detector. Priming early parts of the network causes the most significant boost in performance. black dashed line shows performance by pruning. (b) Testing variants of priming against increasing image noise. The benefits of priming become more apparent in difficult viewing conditions. The x axis indicates which block of the network was primed (1 for primed, 0 for not primed). network and the goal is to find how it can be utilized optimally. Such information can be either deduced from the scene, such as in contextual priming [30, 18] or given by an external source, or even be inferred from the task, such as in question answering [1, 17]. Our experiments are conducted on the Pascal VOC [11] 2007 and 2012 datasets. For priming object detection networks we use pre-trained models of SSD [20] and yolo-v2 [28] and for segmentation we use the FCN-8 segmentation network of [21] and the DeepLab network of [7]. We use the YellowFin optimizer [40] in all of our experiments, with a learning rate of either 0.1 or 0.01 (depending on the task). 4.1. Object Detection We begin by testing our method on object detection. Using an implementation of SSD [20], we apply a pre-trained detector trained on the trainval sets of Pascal 2012+2007 to the test set of Pascal 2007. We use the SSD-300 variant as described in the paper. In this experiment, we trained and tested on what we cal PAS# : this is a reduced version of Pascal-2007 containing only images with a single object class (but possibly multiple instances). We use this 4.1.1 Deep vs Shallow Priming We proceed to the main result, that is, how priming affects detection. The SSD object detector contains four major components: (1) a pre-trained part made up of some of the layers of vgg-16 [31] (a.k.a the “base network” in the SSD paper), (2) some extra convolutional layers on top of the vgg-part, (3) a localization part and (4) a class confidence part. We name these part vgg, extra, loc and conf respectively. To check where priming has the most significant impact, we select different subsets of these components and denote them by 4-bit binary vectors si ∈ {0, 1}4 , where the bits correspond from left to right to the vgg,extra,localization and confidence parts. For example, s = 1000 means letting Np affect only the earliest (vgg) part of the detector, while all other parts remain unchanged by the priming (except indirectly affecting the deeper parts of the net). We train Np on 10 different configurations: these include priming from the deepest layers to the earliest: 1111, 0111, 0011, 0001 and from the earliest layer to the deepest: 1000, 1100, 1110. We add 0100 and 0010 to check the effect of exclusive control over middle layers and finally 0000 as the default configuration in which Np is degenerate and the result is identical to pruning. Fig 4 (a) shows the effect of priming each of these subsets of layers on PAS# . Priming early layers (those at the bottom of the network) has a much more pronounced effect than priming deep layers. The single largest gain by priming a single component is for the vgg part: 1000 boosts performance from 85% to 87.1%. A smaller gain is attained by the extra component: 86.1% for 0100. The performance peaks at 87.3% for 1110, though this is only marginally higher than attained by 1100 - priming only the first two parts. 4.1.2 Ablation Study Priming the earliest layers (vgg+extra) of the SSD object detector brings the most significant boost in performance. The first component described above contains 15 convolutional layers and the second contains 8 layers, an overall total of 23. To see how much we can gain with priming on the first few layers, we checked the performance mAP 0.87 0.86 08 15_07 15_06 15_05 15_04 15_03 15_02 15_01 15_00 15_00 14_00 13_00 12_00 11_00 10_00 09_00 08_00 07_00 06_00 05_00 04_00 03_00 02_00 01- _ -- reduced dataset to test various aspects of our method, as detailed in the following subsections. Without modification, the detector attains a mAP (mean-average precision) of 81.4% on PAS# (77.4% on the full test set of Pascal 2007). Using simple pruning as described above, this increases to 85.2%. This large boost in performance is perhaps not surprising, since pruning effectively removes all detections of classes that do not appear in the image. The remaining errors are those of false alarms of the “correct” class or misdetections. configuration Figure 5: Effects of early priming: we show how mAP increases when we allow priming to affect each time another layer, from the very bottom of the network. Priming early layers has a more significant effect than doing so for deeper ones. The numbers indicate how many layers were primed from the first,second blocks of the SSD network, respectively. on PAS# when training on the first k layers only, for each k ∈ {1, 2, . . . 23}. Each configuration was trained for 4000 iterations. Fig. 5 shows the performance obtained by each of these configurations, where i j in the x-axis refers to having trained the first i layers and the first j layers of the first and second parts respectively. We see that the very first convolutional layer already boosts performance when primed. The improvement continues steadily as we add more layers and fluctuates around 87% after the 15th layer. The fluctuation is likely due to randomness in the training process. This further shows that priming has strong effects when applied to very early layers of the network. 4.1.3 Detection in Challenging Images As implied by the introduction, perhaps one of the cases where the effect of priming is stronger is when facing a challenging image, such as adverse imaging conditions, low lighting, camouflage, noise. As one way to test this, we compared how priming performs under noise. We took each image in the test set of Pascal 2007 and added random Gaussian noise chosen from a range of standard deviations, from 0 to 100 in increments of 10. The noisy test set of PAS# with variance σ is denoted PAS# N(σ) . For each σ, we measure the mAP score attained by either pruning or priming. Note that none of our experiments involved training with only images - these are only used for testing. We plot the results in Fig. 4 (b). As expected, both methods suffer from decreasing accuracy as the noise increases. However, priming is more robust to increasing levels of noise; the difference between the two methods peaks at a moderate level of noise, that is, σ = 80, with an advantage of 10.7% in mAP: 34.8% compared to 24.1% by pruning. The gap decreases gradually to 6.1% (26.1% vs 20%) for a noise level of σ = 100. We believe that this is due to the early-layer effects of priming on the network, selecting features from the bottom up thresh=0.1 thresh=0.01 thresh=0.5 thresh=0.2 thresh=0.1 thresh=0.01 σ = 25 σ = 50 σ = 100 σ = 100 σ = 50 σ = 25 σ=0 thresh=0.2 σ=0 thresh=0.5 (a) (b) Figure 6: Priming vs. Pruning. Priming a detector allows it to find objects in images with high levels of noise while mostly avoiding false-alarms. Left to right (a,b): decreasing detection thresholds (increasing sensitivity). Top to bottom: increasing levels of noise. Priming (blue dashed boxes) is able to detect the horse (a) across all levels of noise, while pruning (red dashed boxes) does not. For the highest noise level, the original classifier does not detect the horse at all - so pruning is ineffective. (b) Priming enables detection of the train for all but the most severe level of noise. Decreasing the threshold for pruning only produces false alarms. We recommend viewing this figure in color on-line. to match the cue. Fig 6 shows qualitative examples, comparing priming versus pruning: we increase the noise from top to bottom and decrease the threshold (increase the sensitivity) from left to right. We show in each image only the top few detections of each method to avoid clutter. Priming allows the detector to find objects in images with high levels of noise (see lower rows of a,b). In some cases priming proves to be essential for the detection: lowering the un-primed detector’s threshold to a minimal level does not increase the recall of the desired object (a, 4th row); in fact, it only increases the number of false alarms (b, 2nd row, last column). Priming, on the other hand, is often less sensitive to a low threshold and the resulting detection persists along a range thereof. object classes c1 , . . . ck we split the training example for I to k different tuples < Ii , hi , gti >, i ∈ {1 . . . k}, where Ii are all identical to I, hi indicate the presence of class ci and gti is the ground-truth gt reduced to contain only the objects of class ci - meaning the bounding boxes for detection, or the masks for segmentation. This explicitly coerces the priming network Np to learn how to force the output to correspond to the given cue, as the input image remains the same but the cue and desired output change together. We refer to this method multi-cue aware training (CAT for short) , and refer to the unchanged training scheme as regular training. 4.2. Cue Aware Training Here, we test the multi-cue training method on object class segmentation. We begin with the FCN-8 segmentation network of [21]. We train on the training split of SBD (Berkeley Semantic Boundaries Dataset and Benchmark) dataset [13], as is done in [42, 7, 8, 21]. We base our code on an unofficial PyTorch2 implementation3 . Testing is done of In this section, we also test priming on an object detection task as well as segmentation with an added ingredient - multi-cue training and testing. In Sec. 4.1 we limited ourselves to the case where there is only one object class per image. This limitation is often unrealistic. To allow multiple priming cues per image, we modify the training process as follows: for each training sample < I, gt > containing 4.2.1 Multi-Cue Segmentation 2 http://pytorch.org/ 3 https://github.com/wkentaro/pytorch-fcn Figure 7: Effect of priming a segmentation network with different cues. In each row, we see an input image and the output of the network when given different cues. Top row: cues are resp. bottle, diningtable, person. Given a cue (e.g, bottle), the network becomes more sensitive to bottle-like image structures while suppressing others. This happens not by discarding results but rather by affecting computation starting from the early layers. the validation set of Pascal 2011, taking care to avoid overlapping images between the training set defined by [13] 4 , which leaves us with 736 validation images. The baseline results average IOU score of 65.3%. As before, we let the cue be a binary encoding of the classes present in the image. We train and test the network in two different modes: one is by setting for each training sample (and testing) the cue so hi = 1 if the current image contains at least one instance of class i and 0 otherwise. The other is the multi-cue method we describe earlier, i.e , splitting each sample to several cues with corresponding ground-truths so each cue is a one-hot encoding, indicating only a single class. For both training strategies, testing the network with a cue creates a similar improvement in performance, from 65.3% to 69% for regular training and to 69.2% for multi-cue training. The main advantage of the multi-cue training is that it allows the priming network Np to force N to focus on different objects in the image. This is illustrated in Fig. 7. The top row of the figure shows from left to right an input image and the resulting segmentation masks when the network is cued with classes bottle, diningtable and person. The bottom row is cued with bus, car, person. The cueaware training allows the priming network to learn how to suppress signals relating to irrelevant classes while retaining the correct class from the bottom-up. Types of Pruning. As mentioned in Sec. 3, we examine two types of pruning to post-process segmentation results. One type removes image regions which were wrongly labeled as the target class, replacing them with background and the other increases the recall of previously missed seg4 for details, please refer to https://github.com/shelhamer/ fcn.berkeleyvision.org/tree/master/data/pascal (a) input (b) gt (c) regular (d) prune-2 (e) prune-1 (f) priming Figure 8: Comparing different methods of using a cue to improve segmentation: From left to right: input image (with cue overlayed), ground-truth (all classes), unprimed segmentation, pruning type-2, pruning type-1, and priming. In each image, we aid the segmentation network by adding a cue (e.g, “plane”). White regions are marked as “don’t care” in the ground truth. mentation regions by removing all classes except the target class and retaining pixels where the target class scored higher than the background. The first type increases precision but cannot increase recall. The second type increases recall but possibly hinders precision. We found that both types results in a similar overall mean-IOU. Figure 8 shows some examples where both types of pruning result in segmentations inferior to the one resulting by priming: postprocessing can increase recall by lowering precision (first row, column d) or increase precision by avoiding falsedetections (second and fourth row, column e), priming (column f) increases both recall and precision. The second, and fourth rows missing parts of the train/bus are recovered while removing false classes. The third and fifth rows previously undetected small objects are now detected. The person (first row) is segmented more accurately. DeepLab. Next, we use the DeepLab [7] network for semantic-segmentation with ResNet-101 [15] as a base network. We do not employ a CRF as post-processing. The mean-IOU of the baseline is 76.3%. Using Priming, increases this to 77.15%. While in this case priming does not improve as much as in the other cases we tested, we find that it is especially effective at enabling the network to discover small objects which were not previously segmented by the non-primed version: the primed network discovers ing method for 25 epochs. When using only pruning, performance on the test-set improves to 78.2% mAP. When we include priming as well, this goes up to 80.6%, 5. Conclusion We have presented a simple mechanism to prime neural networks, as inspired by psychological top-down effects known to exist in human observers. We have tested the proposed method on two tasks, namely object detection and segmentation, using two methods for each task, and comparing it to simple post-processing of the output. Our experiments confirm that as is observed in humans, effective usage of a top-down signal to modulate computations from early layers not only improves robustness to noise but also facilitates better object detection and segmentation, enabling detection of objects which are missed by the baselines without compromising precision, notably so for small objects and those having an atypical appearance. References Figure 9: Priming a network allows discovery of small objects which are completely missed by the baseline, or ones with uncommon appearance (last row). From left to right: input image, ground-truth, baseline segmentation [7], primed network. 57 objects which were not discovered by the unprimed network, whereas the latter discovers only 3 which were not discovered by the former. Fig. 9 shows some representative examples of where priming was advantageous. Note how the bus, person, (first three rows) are segmented by the primed network (last column). We hypothesize that the priming process helps increase the sensitivity of the network to features relevant to the target object. The last row shows a successful segmentation of potted plants with a rather atypical appearance. 4.2.2 Multi-Cue Object Detection We apply the CAT method to train priming on object detection as well. For this experiment, we use the YOLOv2 method of [28]. The base network we used is a port of the original network, known as YOLOv2 544 × 544. Trained on the union of Pascal 2007 and 2012 datasets, it is reported by the authors to obtain 78.6% mAP on the test set of Pascal 2007. The implementation we use5 reaches a slightly lower 76.8%, with a PyTorch port of the network weights released by the authors. We use all the convolutional layers of DarkNet (the base network of YOLOv2 ) to perform priming. We freeze all network parameters of the original detection network and train a priming network with the multi-cue train5 https://github.com/marvis/pytorch-yolo2 [1] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In Proceedings of the IEEE International Conference on Computer Vision, pages 2425–2433, 2015. 2, 4 [2] Mandy V Bartsch, Kristian Loewe, Christian Merkel, Hans-Jochen Heinze, Mircea A Schoenfeld, John K Tsotsos, and Jens-Max Hopf. Attention to color sharpens neural population tuning via feedback processing in the human visual cortex hierarchy. Journal of Neuroscience, 37(43):10346–10357, 2017. 1 [3] Irving Biederman. On the semantics of a glance at a scene. 1981. 2 [4] Irving Biederman, Robert J Mezzanotte, and Jan C Rabinowitz. Scene perception: Detecting and judging objects undergoing relational violations. Cognitive psychology, 14(2):143–177, 1982. 2 [5] Mahdi Biparva and John Tsotsos. STNet: Selective Tuning of Convolutional Networks for Object Localization. In The IEEE International Conference on Computer Vision (ICCV), Oct 2017. 2 [6] Joao Carreira, Pulkit Agrawal, Katerina Fragkiadaki, and Jitendra Malik. Human pose estimation with iterative error feedback. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 4733–4742, 2016. 2 [7] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. arXiv preprint arXiv:1606.00915, 2016. 4, 4.2.1, 4.2.1, 9 [8] Jifeng Dai, Kaiming He, and Jian Sun. Boxsup: Exploiting bounding boxes to supervise convolutional networks for semantic segmentation. In Proceedings of the IEEE International Conference on Computer Vision, pages 1635–1643, 2015. 4.2.1 [9] Harm de Vries, Florian Strub, Jérémie Mary, Hugo Larochelle, Olivier Pietquin, and Aaron Courville. Modulating early visual processing by language. arXiv preprint arXiv:1707.00683, 2017. 2 [10] Santosh K Divvala, Derek Hoiem, James H Hays, Alexei A Efros, and Martial Hebert. An empirical study of context in object detection. In Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on, pages 1271–1278. IEEE, 2009. 2 [11] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303–338, 2010. 3, 3.1, 4 [12] Carolina Galleguillos and Serge Belongie. Context based object categorization: A critical survey. Computer vision and image understanding, 114(6):712– 722, 2010. 2 [13] Bharath Hariharan, Pablo Arbeláez, Lubomir Bourdev, Subhransu Maji, and Jitendra Malik. Semantic contours from inverse detectors. In Computer Vision (ICCV), 2011 IEEE International Conference on, pages 991–998. IEEE, 2011. 4.2.1 [14] Adam W Harley, Konstantinos G Derpanis, and Iasonas Kokkinos. Segmentation-Aware Convolutional Networks Using Local Attention Masks. arXiv preprint arXiv:1708.04607, 2017. 2 [15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770–778, 2016. 4.2.1 [16] Andrew Hollingworth. Does consistent scene context facilitate object perception? Journal of Experimental Psychology: General, 127(4):398, 1998. 2 [17] Justin Johnson, Bharath Hariharan, Laurens van der Maaten, Li Fei-Fei, C Lawrence Zitnick, and Ross Girshick. CLEVR: A diagnostic dataset for compositional language and elementary visual reasoning. arXiv preprint arXiv:1612.06890, 2016. 2, 4 [18] Harish Katti, Marius V Peelen, and SP Arun. Object detection can be improved using humanderived contextual expectations. arXiv preprint arXiv:1611.07218, 2016. 4 [19] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In European conference on computer vision, pages 740–755. Springer, 2014. 3.1 [20] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In European conference on computer vision, pages 21–37. Springer, 2016. 3, 4, 4.1 [21] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3431– 3440, 2015. 4, 4.2.1 [22] Kevin P Murphy, Antonio Torralba, and William T Freeman. Using the forest to see the trees: A graphical model relating features, objects, and scenes. In Advances in neural information processing systems, pages 1499–1506, 2004. 2 [23] Aude Oliva and Antonio Torralba. The role of context in object recognition. Trends in cognitive sciences, 11(12):520–527, 2007. 2 [24] tephen E Palmer. The effects of contextual scenes on the identification of objects. Memory & Cognition, 3(5):519–526, 1975. 2 [25] Ethan Perez, Florian Strub, Harm de Vries, Vincent Dumoulin, and Aaron Courville. FiLM: Visual Reasoning with a General Conditioning Layer. arXiv preprint arXiv:1709.07871, 2017. 2 [26] Michael I Posner, Mary Jo Nissen, and William C Ogden. Attended and unattended processing modes: The role of set for spatial location. Modes of perceiving and processing information, 137:158, 1978. 2 [27] Andrew Rabinovich, Andrea Vedaldi, Carolina Galleguillos, Eric Wiewiora, and Serge Belongie. Objects in context. In Computer vision, 2007. ICCV 2007. IEEE 11th international conference on, pages 1–8. IEEE, 2007. 2 [28] Joseph Redmon and Ali Farhadi. YOLO9000: Better, Faster, Stronger. arXiv preprint arXiv:1612.08242, 2016. 4, 4.2.2 [40] Jian Zhang, Ioannis Mitliagkas, and Christopher Ré. YellowFin and the Art of Momentum Tuning. arXiv preprint arXiv:1706.03471, 2017. 4 [29] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster R-CNN: Towards real-time object detection with region proposal networks. In Advances in neural information processing systems, pages 91–99, 2015. 2 [41] Jianming Zhang, Zhe Lin, Jonathan Brandt, Xiaohui Shen, and Stan Sclaroff. Top-down Neural Attention by Excitation Backprop. In European Conference on Computer Vision, pages 543–559. Springer, 2016. 2 [30] Abhinav Shrivastava and Abhinav Gupta. Contextual priming and feedback for faster r-cnn. In European Conference on Computer Vision, pages 330–348. Springer, 2016. 2, 4 [31] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 4.1.1 [32] Thomas M Strat. Employing contextual information in computer vision. DARPA93, pages 217–229, 1993. 2 [33] Antonio Torralba. Contextual priming for object detection. International journal of computer vision, 53(2):169–191, 2003. 2 [34] Antonio Torralba, Kevin P Murphy, William T Freeman, and Mark A Rubin. Context-based vision system for place and object recognition. In null, page 273. IEEE, 2003. 2 [35] John K Tsotsos. A computational perspective on visual attention. MIT Press, 2011. 1, 2 [36] John K. Tsotsos, Scan M. Culhane, Winky Yan Kei Wai, Yuzhong Lai, Neal Davis, and Fernando Nuflo. Modeling visual attention via selective tuning. Artificial Intelligence, 78(1–2):507–545, 1995. Special Volume on Computer Vision. 2, 3 [37] Endel Tulving, Daniel L Schacter, et al. Priming and human memory systems. Science, 247(4940):301– 306, 1990. 2 [38] Gagan S Wig, Scott T Grafton, Kathryn E Demos, and William M Kelley. Reductions in neural activity underlie behavioral components of repetition priming. Nature neuroscience, 8(9):1228–1233, 2005. 2 [39] Jian Yao, Sanja Fidler, and Raquel Urtasun. Describing the scene as a whole: Joint object detection, scene classification and semantic segmentation. In Computer Vision and Pattern Recognition (CVPR), 2012 IEEE Conference on, pages 702–709. IEEE, 2012. 2 [42] Hengshuang Zhao, Jianping Shi, Xiaojuan Qi, Xiaogang Wang, and Jiaya Jia. Pyramid scene parsing network. arXiv preprint arXiv:1612.01105, 2016. 4.2.1 

Object category learning and retrieval with weak supervision arXiv:1801.08985v1 [cs.CV] 26 Jan 2018 Steven Hickson, Anelia Angelova, Irfan Essa, Rahul Sukthankar Google Brain / Google Research (shickson, anelia, irfanessa, sukthankar)@google.com Abstract We consider the problem of retrieving objects from image data and learning to classify them into meaningful semantic categories with minimal supervision. To that end, we propose a fully differentiable unsupervised deep clustering approach to learn semantic classes in an end-to-end fashion without individual class labeling using only unlabeled object proposals. The key contributions of our work are 1) a kmeans clustering objective where the clusters are learned as parameters of the network and are represented as memory units, and 2) simultaneously building a feature representation, or embedding, while learning to cluster it. This approach shows promising results on two popular computer vision datasets: on CIFAR10 for clustering objects, and on the more complex and challenging Cityscapes dataset for semantically discovering classes which visually correspond to cars, people, and bicycles. Currently, the only supervision provided is segmentation objectness masks, but this method can be extended to use an unsupervised objectness-based object generation mechanism which will make the approach completely unsupervised. 1 Introduction Unsupervised discovery of common patterns is a long standing task for artificial intelligence as shown in Barlow (1989); Bengio, Courville, and Vincent (2012). Recent deep learning approaches have offered major breakthroughs in classification into multiple categories with millions of labeled examples (e.g. Krizhevsky (2009); Szegedy et al. (2015); He et al. (2016) and many others). These methods rely on a lot of annotated data for training in order to perform well. Unfortunately, labeling is an inefficient and expensive progress, so learning from unlabeled data is desirable for many complex tasks. At the same time, much of human knowledge and learning is obtained by unsupervised observations Grossberg (1994). The goal of this work is to show that semantically meaningful classes can be learned with minimal supervision. Given a set of objectness proposals, we use the activations of foreground objects in order to learn deep features to cluster the available data while simultaneously learning the embedding in an end-to-end manner. More specifically, we propose a differentiable clustering approach that learns better separability of classes and embedding. The main idea is to store the potential cluster means as weights in a neural network at the higher levels of feature representation. This allows them to be learned jointly with the potential feature representation. This differentiable clustering approach is integrated with Deep Neural Networks (e.g. Szegedy et al. (2015)) to learn semantic classes in an end-to-end fashion without manual class labeling. The idea of doing this ‘end-to-end’ is that gradient descent can not only learn good weights for clustering, it can also change the embedding to allow for better clustering without the use of labels. We see that this leads to better feature representation. Our results show also that different object categories emerge and can later be retrieved from test images never before seen by the network, resulting in clusters of meaningful categories, such as cars, persons, bicycles. 31st Conference on Neural Information Processing Systems (NIPS 2017), Long Beach, CA, USA. In this work we use given segmentation objectness masks, which are candidate objects without labels. This can be extended by using an independent objectness-based object generation mechanism Pathak et al. (2017); Faktor and Irani (2014) or by using unsupervised motion segmentation in videos or structure from motion Vijayanarasimhan et al. (2017). 2 Related Work Unsupervised learning (Barlow (1989)) and unsupervised deep learning (Bengio, Courville, and Vincent (2012), Bengio (2012), Bengio and others (2009)) are central topics to Machine Learning. Unsupervised deep learning has been shown to improve results on classification tasks per Erhan et al. (2010), especially given small datasets and complicated high dimensional data such as video. This has been explored by many representations including sequence to sequence learning and textual representations (Radford, Jozefowicz, and Sutskever (2017), Ramachandran, Liu, and Le (2016)). Our work focuses on unsupervised deep learning for discovering visual object categories. This has also been shown to improve results such as in Doersch and Zisserman (2017). Unsupervised discovery of visual objects has been a large topic of interest in computer vision (Sivic et al. (2005); Russell et al. (2006); Singh, Gupta, and Efros (2012); Bach and Jordan (2005); Kwak et al. (2015); Pathak et al. (2017)). Building specialized, deep embeddings to help computer vision tasks is also a popular approach such as in Agrawal, Carreira, and Malik (2015). Transfer learning from supervised tasks has proven to be very successful. Further, Agrawal, Carreira, and Malik (2015) propose learning the lower dimensional embedding through unsupervised learning and show improved performance when transfered to other supervised tasks. Despite the popularity of building different embeddings, there is little work investigating the use of clustering to modify the embedding in an end-to-end deep learning framework. Bottou and Bengio (1995) investigate a differentiable version of the kmeans algorithm and examine its convergence properties. Our work focuses on learnable feature representations (instead of fixed ones as in Bottou and Bengio (1995)) and introduces memory units for the task. 3 Unsupervised deep clustering Our unsupervised deep clustering is inspired by Bottou and Bengio (1995), who consider differentiable clustering algorithms. We differ from this approach because the features we cluster also change with backpropogation. In our work, we add a kmeans-like loss that is integrated end-to-end. Our idea is to store the potential cluster means as weights in the network and thus have them be learned. The proposed clustering is done simultaneously while building an embedding. Given information of a potential object vs background (binary labels), clustering in a differentiable way provides a better embedding for the input data. We show that this method can be used for meaningful semantic retrieval of related objects. 3.1 Embedding with clustering We train a convolutional neural network (CNN) to predict foreground and background using oracle labels of patches of objects and background images. Concurrently, we learn the clustering of objects by imposing constraints that will force the embedding to be partitioned into multiple semantically coherent clusters of objects without explicit labels for different objects. For our experiments, we use random initialization on the fully-connected layers (the last two layers) and we add the differentiable clustering module after the second to last layer. Note that we only cluster the foreground labels as background activations are not of interest for clustering; the classifier can predict foreground vs background with high accuracy (above 90%). The objective function is shown in Equation 1. N 1 X Lk = mink [(xn − wk )2 ] 2N n=1 2 (1) In this equation, N is the number of samples, k is the number of defined clusters, w is the “weight” (theoretically and typically the mean of the cluster) for each k, and x is the activations from the fully connected layer before the classification fully connected layer. This is differentiable and the gradient descent algorithm is shown in Equation 2. δwk = wk0 − wk = N  X lr (xn − wk ) if k = s(xn , w) 0 otherwise n=1 (2) where s(xn , w) = argmin Pk [xn ] and lr is the learning rate. We also add L2 regularization over the weights to the loss L2 = j wj2 . Furthermore, we use a custom clustering regularization loss LC that enforces that the clusters are evenly distributed as defined in Equation 3 and Equation 4. K K 1 XX LC = |countk − countj | NK (3) k=0 j=k countk = N  X 0 n=0 1 if argmink [xn ] = 0 if argmink [xn ] = 1 (4) The final loss to be optimized is shown in (Equation 5) L = Lk + αr L2 + αc LC (5) where αr and αc are hyperparameters which are tuned during the training. For our method, we use αr = 0.25 and αc = 1. We apply this loss to every point that is labeled as potentially an object and ignore the background ones when clustering. This way we learn foreground vs background and then learn clustering of the foreground activations. Optimization was performed with a ‘RMSProp’ optimizer, with a learning rate of 0.045, momentum 0.9, decay factor 0.9, and  of 1.0. 4 Experimental evaluation We experiment with a toy example using CIFAR10 and a more challenging example using Cityscapes. 4.1 CIFAR10 dataset We first test the proposed unsupervised clustering approach on the CIFAR10 Krizhevsky (2009) dataset. The goal of this experiment is to test if clustering can uncover separate categories in a simple toy problem with a two class setting. Clusters Automobile Dog Cluster 0 68.5% 17.9% Cluster 1 31.5% 82.1% Table 1: Unsupervised clustering results on CIFAR10 for discovery of two classes. Per cluster accuracy for each of the two given classes on the test set (class labels are unknown during training). We selected as an example the dog and automobile classes to label as foreground. We then train a network from scratch based on the Network in Network architecture (NiN) of Lin, Chen, and Yan (2013) from scratch for our experiments. All other classes of CIFAR are considered background for this experiment. By attaching our modified clustering objective function to the next to last layer, we attempt to cluster dog and automobile without labels. We can see in our simple experimental results that classes are naturally clustered with the majority of examples correctly assigned. Table 1 shows quantitative results on the test set. As seen 68.5% of the automobile classes and 82.1% of the dog examples are correctly assigned to separate clusters. Note that in these cases, the concepts and classes of dog and automobile are unknown to the training algorithm and we are just looking at them after clustering for evaluation. 3 Classes Cluster 0 Cluster 1 Person 4320 138 Rider 676 138 Car 1491 4399 Truck 60 69 Bus 49 89 Train 17 16 Motorcycle 88 205 Bicycle 795 787 Table 2: Unsupervised clustering of objects from Cityscapes using our method. The table shows number of examples assigned to each learned cluster (for K=2). 4.2 Cityscapes dataset The Cityscapes dataset (Cordts et al. (2016)) is a large-scale dataset that is used for evaluating various classification, detection, and segmentation algorithms related to autonomous driving. It contains 2975 training, 500 validation, and 1525 test images, where the test set is provided for the purposes of the Cityscape competition only. In this work, we used the training set for training and the validation set for testing and visualizing results (as the test set has no annotation results). Annotation is provided for classes which represent moving agents in the scene, such as pedestrian, car, motorcycle, bicycle, bus, truck, rider, train. In this work we only use foreground/background labels and intend to discover semantic groups from among the moving objects. 4.3 Weakly supervised discovery of classes In this experiment we considered the larger, real-life dataset, Cityscapes (Cordts et al. (2016)), described above to see if important class categories, e.g. the moving objects in the scene can be clustered into semantically meaningful classes. We extract the locations and extents of the moving objects and use that as weak supervision. Note the classes are uneven and car and person dominate. We show results clustering 8 categories into 2 and 3 clusters despite the rarity of some of them (such as bicycle). All results below are presented on the validation set. We report the results in terms of the number of object patches extracted from the available test images. For this dataset, the CNN architecture is based on the Inception architecture proposed by Szegedy et al. (2015). Since there are a small number of examples, we pre-train only the convolutional layers of the network. Results on clustering the 8 classes of moving objects into 2 and 3 clusters are presented in Table 2 and Table 3 respectively for the learned embedding by the proposed approach and the baseline embedding. The baseline embedding is calculated by fine-tuning the same architecture in the same manner, but without our loss (Equation 5); it uses the same amount of information as input as our embedding. For this experiment, we apply standard kmeans on both activations after training is completed. We see here that our method provides better clustering for the two dominant classes in the dataset (car and person). On the other hand, the baseline embedding clusters on one class only, similar to the two class case. We have consistently observed this behavior for different runs and hypothesize this is due to the sparse nature of the baseline embedding and it’s activations. Figure 1 visualizes the three retrieved clusters (color-coded) when clustering into 3 clusters with our approach. We can see that people (in blue) and cars (in green) are often correctly retrieved. Bikes are more rare and may be more often mistaken, for example in cases where a portion of the patch contains part of a car, or since the bicycle very often has a person riding it. Still this is exciting result, given that it is learned by not providing a single class label during training. 5 Conclusions We propose a differentiable clustering objective which learns to separate classes during learning and build a better embedding. The key idea is to be able to learn the clusters which are stored as weights, and simultaneously learn the feature representation and the clustering of the data. Our results show that the proposed approach is useful for extracting semantically related objects. 4 Our method Baseline Classes Cluster 0 Cluster 1 Cluster 2 Cluster 0 Cluster 1 Cluster 2 Person 151 4315 17 4482 1 0 Rider 258 551 7 816 0 0 Car 5195 950 180 6312 13 0 Truck 89 39 5 131 2 0 Bus 127 20 5 152 0 0 Train 25 9 1 35 0 0 Motorcycle 127 76 4 207 0 0 Bicycle 1128 541 450 2119 0 0 Table 3: Unsupervised clustering on the Cityscapes dataset with 3 clusters. The table shows the number of examples assigned to each learned cluster. Our method (left) and baseline (right). Our method results in 69.98% accuracy. Figure 1: Visualization of clusters learned by our method (for K=3). From the figure, the green class is responsible for retrieving cars, the blue one persons, and the red one bicycles. We can see that both cars and persons are discovered well but bicycles, a rarer class, can be confused with a person or with a partially visible car in the background. 5 References Agrawal, P.; Carreira, J.; and Malik, J. 2015. Learning to see by moving. CVPR. Bach, F. R., and Jordan, M. I. 2005. Learning spectral clustering. NIPS. Barlow, H. 1989. Unsupervised learning. Neural computation. Bengio, Y., et al. 2009. Learning deep architectures for ai. Foundations and trends R in Machine Learning 2(1):1–127. Bengio, Y.; Courville, A. C.; and Vincent, P. 2012. Unsupervised feature learning and deep learning: A review and new perspectives. CoRR, abs/1206.5538. Bengio, Y. 2012. Deep learning of representations for unsupervised and transfer learning. In Proceedings of ICML Workshop on Unsupervised and Transfer Learning, 17–36. Bottou, L., and Bengio, Y. 1995. Convergence properties of the k-means algorithms. In Advances in neural information processing systems, 585–592. Cordts, M.; Omran, M.; Ramos, S.; Rehfeld, T.; Enzweiler, M.; Benenson, R.; Franke, U.; Roth, S.; and Schiele, B. 2016. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 3213–3223. Doersch, C., and Zisserman, A. arXiv:1708.07860. 2017. Multi-task self-supervised visual learning. arXiv preprint Erhan, D.; Bengio, Y.; Courville, A.; Manzagol, P.-A.; Vincent, P.; and Bengio, S. 2010. Why does unsupervised pre-training help deep learning? Journal of Machine Learning Research 11(Feb):625–660. Faktor, A., and Irani, M. 2014. Video segmentation by non-local consensus voting. BMVC. Grossberg, S. 1994. 3-d vision and figure-ground separation by visual cortex. Perception and Psychophysics. He, K.; Zhang, X.; Ren, S.; and Sun, J. 2016. Deep residual learning for image recognition. CVPR. Krizhevsky, A. 2009. Learning multiple layers of features from tiny images. Kwak, S.; Cho, M.; Laptev, I.; Ponce2, J.; and Schmid, C. 2015. Unsupervised object discovery and tracking in video collections. ICCV. Lin, M.; Chen, Q.; and Yan, S. 2013. Network in network. arXiv preprint arXiv:1312.4400. Pathak, D.; Girshick, R.; Dollar, P.; Darrell, T.; and Hariharan, B. 2017. Learning features by watching objects move. CVPR. Radford, A.; Jozefowicz, R.; and Sutskever, I. 2017. Learning to generate reviews and discovering sentiment. arXiv preprint arXiv:1704.01444. Ramachandran, P.; Liu, P. J.; and Le, Q. V. 2016. Unsupervised pretraining for sequence to sequence learning. arXiv preprint arXiv:1611.02683. Russell, B. C.; Efros, A. A.; Sivic, J.; Freeman, W. T.; and Zisserman, A. 2006. Using multiple segmentations to discover objects and their extent in image collections. CVPR. Singh, S.; Gupta, A.; and Efros, A. A. 2012. Unsupervised discovery of mid-level discriminative patches. ECCV. Sivic, J.; Russell, B. C.; Efros, A. A.; Zisserman, A.; and Freeman, W. T. 2005. Discovering objects and their location in images. ICCV. Szegedy, C.; Liu, W.; Jia, Y.; Sermanet, P.; Reed, S.; Anguelov, D.; Erhan, D.; Vanhoucke, V.; and Rabinovich, A. 2015. Going deeper with convolutions. CVPR. Vijayanarasimhan, S.; Ricco, S.; Schmid, C.; Sukthankar, R.; and Fragkiadaki, K. 2017. Sfm-net: Learning of structure and motion from video. 6 

arXiv:1608.01654v2 [cs.PL] 7 Nov 2016 Hypercollecting Semantics and its Application to Static Analysis of Information Flow Mounir Assaf David A. Naumann Julien Signoles Stevens Institute of Technology, Hoboken, US first.last@stevens.edu Stevens Institute of Technology, Hoboken, US first.last@stevens.edu Software Reliability and Security Lab, CEA LIST, Saclay, FR first.last@cea.fr Éric Totel Frédéric Tronel CIDRE, CentraleSupélec, Rennes, FR first.last@centralesupelec.fr CIDRE, CentraleSupélec, Rennes, FR first.last@centralesupelec.fr Abstract We show how static analysis for secure information flow can be expressed and proved correct entirely within the framework of abstract interpretation. The key idea is to define a Galois connection that directly approximates the hyperproperty of interest. To enable use of such Galois connections, we introduce a fixpoint characterisation of hypercollecting semantics, i.e. a “set of sets” transformer. This makes it possible to systematically derive static analyses for hyperproperties entirely within the calculational framework of abstract interpretation. We evaluate this technique by deriving example static analyses. For qualitative information flow, we derive a dependence analysis similar to the logic of Amtoft and Banerjee (SAS’04) and the type system of Hunt and Sands (POPL’06). For quantitative information flow, we derive a novel cardinality analysis that bounds the leakage conveyed by a program instead of simply deciding whether it exists. This encompasses problems that are hypersafety but not k-safety. We put the framework to use and introduce variations that achieve precision rivalling the most recent and precise static analyses for information flow. Categories and Subject Descriptors D.2.4 [Software Engineering]: Software/Program Verification–Assertion checkers; D.3 [Programming Languages]; F.3.1 [Logics and meanings of programs]: Semantics of Programming Language Keywords static analysis, abstract interpretation, information flow, hyperproperties 1. Introduction Most static analyses tell something about all executions of a program. This is needed, for example, to validate compiler optimizations. Functional correctness is also formulated in terms of a predicate on observable behaviours, i.e. more or less abstract execution traces: A Permission to make digital or hard copies of part or all 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, to republish, to post on servers, or to redistribute to lists, contact the Owner/Author(s). Request permissions from permissions@acm.org or Publications Dept., ACM, Inc., fax +1 (212) 869-0481. POPL ’17, January 18 - 20, 2017, Paris, France Copyright c 2017 held by owner/author(s). Publication rights licensed to ACM. ACM 978-1-4503-4660-3/17/01. . .$15.00 DOI: http://dx.doi.org/10.1145/http://dx.doi.org/10.1145/3009837.3009889 program is correct if all its traces satisfy the predicate. By contrast with such trace properties, extensional definitions of dependences involve more than one trace. To express that the final value of a variable x may depend only on the initial value of a variable y, the requirement—known as noninterference in the security literature (Sabelfeld and Myers 2003)—is that any two traces with the same initial value for y result in the same final value for x. Sophisticated information flow policies allow dependences subject to quantitative bounds—and their formalisations involve more than two traces, sometimes unboundedly many. For secure information flow formulated as decision problems, the theory of hyperproperties classifies the simplest form of noninterference as 2-safety and some quantitative flow properties as hypersafety properties (Clarkson and Schneider 2010). A number of approaches have been explored for analysis of dependences, including type systems, program logics, and dependence graphs. Several works have used abstract interpretation in some way. One approach to 2-safety is by forming a product program that encodes execution pairs (Barthe et al. 2004; Terauchi and Aiken 2005; Darvas et al. 2005), thereby reducing the problem to ordinary safety which can be checked by abstract interpretation (Kovács et al. 2013) or other means. Alternatively, a 2-safety property can be checked by dedicated analyses which may rely in part on ordinary abstract interpretations for trace properties (Amtoft et al. 2006). The theory of abstract interpretation serves to specify and guide the design of static analyses. It is well known that effective application of the theory requires choosing an appropriate notion of observable behaviour for the property of interest (Cousot 2002; Bertrane et al. 2012, 2015). Once a notion of “trace” is chosen, one has a program semantics and “all executions” can be formalized in terms of collecting semantics, which can be used to define a trace property of interest, and thus to specify an abstract interpretation (Cousot and Cousot 1977, 1979; Cousot 1999). The foundation of abstract interpretation is quite general, based on Galois connections between semantic domains on which collecting semantics is defined. Clarkson and Schneider (2010) formalize the notion of hyperproperty in a very general way, as a set of sets of traces. Remarkably, prior works using abstract interpretation for secure information flow do not directly address the set-of-sets dimension and instead involve various ad hoc formulations. This paper presents a new approach of deriving information flow static analyses within the calculational framework of abstract interpretation. First contribution. We lift collecting semantics to sets of trace sets, dubbed hypercollecting semantics, in a fixpoint formulation which is not simply the lifted direct image. This can be composed with Galois connections that specify hyperproperties beyond 2safety, without recourse to ad hoc additional notions. On the basis of this foundational advance, it becomes possible to derive static analyses entirely within the calculational framework of abstract interpretation (Cousot and Cousot 1977, 1979; Cousot 1999). Second contribution. We use hypercollecting semantics to derive an analysis for ordinary dependences. This can be seen as a rational reconstruction of both the type system of Hunt and Sands (2006, 2011) and the logic of Amtoft and Banerjee (2004). They determine, for each variable x, a conservative approximation of the variables y whose initial values influence the final value of x. Third contribution. We derive a novel analysis for quantitative information flow. This shows the benefit of taking hyperproperties seriously by means of abstract interpretation. For noninterference, once the variables y on which x depends have fixed values, there can be only one final value for x. For quantitative information flow, one is interested in measuring the extent to which other variables influence x: for a given range of variation for the “high inputs”, what is the range of variation for the final values of x? We directly address this question as a hyperproperty: given a set of traces that agree only on the low inputs, what is the cardinality of the possible final values for x? Using the hypercollecting semantics, we derive a novel cardinality abstraction. We show how it can be used for analysis of quantitative information problems including a bounding problem which is not k-safety for any k. The calculational approach disentangles key design decisions and it enabled us to identify opportunities for improving precision. We assess the precision of our analyses and provide a formal characterisation of precision for a quantitative information flow analysis vis a vis qualitative. Versions of our analyses rival state of the art analyses for qualitative and quantitative information flow. Our technical development uses the simplest programming language and semantic model in which the ideas can be exposed. One benefit of working entirely within the framework of abstract interpretation is that a wide range of semantics and analyses are already available for rich programming languages. Outline. Following the background (Section 2), we introduce domains and Galois connections for hyperproperties (Section 3) and hypercollecting semantics (Section 4). Hyperproperties for information flow are defined in Section 5. We use the framework to derive the static analyses in Section 6 and Section 7. Section 8 uses examples to evaluate the precision of the analyses, and shows how existing analyses can be leveraged to improve precision. We discuss related work (Section 9) and conclude. Appendices provide detailed proofs for all results, as well as a table of symbols. 2. Background: Collecting Semantics, Galois Connections The formal development uses deterministic imperative programs over integer variables. Let n range over literal integers Z, x over variables, and ⊕ (resp. cmp) over some arithmetic (resp. comparison) operators. c ::= skip | x := e | c1 ; c2 | if b then c1 else c2 | while b do c e ::= n | x | e1 ⊕ e2 | b b ::= e1 cmp e2 Different program analyses may consider different semantic domains as needed to express a given class of program properties. For imperative programs, the usual domains are based on states σ ∈ States that map each variable to a value (Winskel 1993). Some P(States∗ ) P(Trc) P(States) Figure 1. Fragment of the hierarchy of semantic domains abstraction (−−−−−→) program properties require the use of traces that include intermediate states; others can use more abstract domains. For information flow properties involving intermediate outputs, or restricted to explicit data flow (Schoepe et al. 2016), details about intermediate steps are needed. By contrast, bounding the range of variables can be expressed in terms of final states. As another example, consider determining which variables are left unchanged: To express this, we need both initial and final states. In this paper we use the succinct term trace for elements of Trc defined by Trc , States × States, interpreting t ∈ Trc as an initial and final state. In the literature, these are known as relational traces, by contrast with maximal trace semantics using the set States∗ of finite sequences. A uniform framework describes the relationships and correspondences between these and many other semantic domains using Galois connections (Cousot 2002). Three of these domains are depicted in Figure 1. Given partially ordered sets C, A, the monotone functions α ∈ C → A and γ ∈ A → C comprise a Galois connection, γ − − (A, v), provided they satisfy a proposition we write (C, ≤) ← −− α→ α(c) v a iff c ≤ γ(a) for all c ∈ C, a ∈ A. For example, to specify an analysis that determines which variables are never changed, let A be sets of variables. Define α ∈ P(Trc) → P(Vars) by α(T ) = {x | ∀(σ, σ 0 ) ∈ T, σ(x) = σ 0 (x)} and γ(X) = {(σ, σ 0 ) | ∀x ∈ X, σ(x) = σ 0 (x)}. Then γ − − (P(V ar), ⊇). (P(Trc), ⊆) ← −− α→ For the hierarchy of usual domains, depicted in Figure 1, the connections are defined by an “element-wise abstraction”. Define elt ∈ States∗ → Trc by elt(σ0 σ1 . . . σn ) , (σ0 , σn ). This lifts to an abstraction P(States∗ ) → P(Trc). Lemma 1 Element-wise abstraction. Let elt ∈ C → A be a function between sets. Let αelt (C) , {elt(c) | c ∈ C} and γelt −− −− − (P(A), ⊆). γelt (A) , {c | elt(c) ∈ A}. Then (P(C), ⊆) ← −− → α− elt The domain P(States), which suffices to describe the final reachable states of a program, is an abstraction of the relational domain P(Trc), by elt(σ, τ ) , τ . In this paper we focus on the domain Trc because it is the simplest that can express dependences. Program semantics. We define both the denotational semantics JcK ∈ Trc⊥ → Trc⊥ of commands and the denotational semantics JeK ∈ Trc → Val of expressions. Here Val , Z and Trc⊥ adds bottom element ⊥ using the flat ordering. JcK ∈ Trc⊥ → Trc⊥ Standard semantics of commands JcK⊥ , ⊥ Jx := eK(σ, τ ) , (σ, τ [x 7→ JeK(σ, τ )]) Jc1 ; c2 Kt , Jc2 K ◦ Jc1 Kt JskipKt , t ( Jc1 Kt Jif b then c1 else c2 Kt , Jc2 Kt if JbKt = 1 if JbKt = 0 Jwhile b do cKt , (lfp4̇ (λt.⊥) F)(t) ( t where F(w)(t) , w ◦ JcKt if JbKt = 0 otherwise Let t be a trace (σ, τ ). The denotation JeKt evaluates e in the “current state”, τ . (In Sect. 5 we also use JeKpre t which evaluates e in the initial state, σ.) The denotation JcKt is (σ, τ 0 ) where execution of c in τ leads to τ 0 . The denotation is ⊥ in case c diverges from τ . Boolean expressions evaluate to either 0 or 1. We assume programs do not go wrong. We denote by 4̇ the point-wise lifting to Trc⊥ → Trc⊥ of the approximation order 4 on Trc⊥ . The terminating computations of c can be written as its image on the initial traces: {JcKt | t ∈ IniTrc and JcKt 6= ⊥} where IniTrc , {(σ, σ) | σ ∈ States} To specify properties that hold for all executions we use collecting semantics which lifts the denotational semantics to arbitrary sets T ∈ P(Trc) of traces. The idea is that ⦃c⦄T is the direct image of JcK on T . To be precise, in this paper we focus on terminationinsensitive properties, and thus ⦃c⦄T is the set of non-⊥ traces t0 such that JcKt = t0 for some t ∈ T . Later we also use the collecting semantics of expressions: ⦃e⦄T , {JeKt | t ∈ T }. Importantly, the collecting semantics ⦃c⦄ ∈ P(Trc) → P(Trc) can be defined compositionally using fixpoints (Cousot 2002, Sec. 7). For conditional guard b, write ⦃ grdb ⦄ for the filter defined by ⦃ grdb ⦄T , {t ∈ T | JbKt = 1}. ⦃c⦄ ∈ P(Trc) → P(Trc) Collecting semantics ⦃x := e⦄T , {Jx := eKt | t ∈ T } ⦃c1 ; c2 ⦄T , ⦃c2 ⦄ ◦ ⦃c1 ⦄T ⦃skip⦄T , T ⦃if b then c1 else c2 ⦄T , ⦃c1 ⦄ ◦ ⦃ grdb ⦄T ∪ ⦃c2 ⦄ ◦ ⦃ grd¬b ⦄T   ⦃while b do c⦄T , ⦃ grd¬b ⦄ lfp⊆ T ⦃if b then c else skip⦄ The clause for while loops uses the denotation of a constructed conditional command as a definitional shorthand—its denotation is compositional. γ − − (A, v), such as Given a Galois connection (P(Trc), ⊆) ← −− α→ the one for unmodified variables, the desired analysis is specified as α ◦ ⦃c⦄ ◦ γ. Since it is not computable in general, we only require an approximation f ] ∈ A → A that is sound in this sense: α ◦ ⦃c⦄ ◦ γ v̇ f ] (1) where v̇ denotes the point-wise lifting of the partial order v. To explain the significance of this specification, suppose one wishes to prove program c satisfies a trace property T ∈ P(Trc), i.e. to prove that ⦃c⦄(IniTrc) ⊆ T . Given eq. (1) it suffices to find an abstract value a that approximates IniTrc, i.e. IniTrc ⊆ γ(a), and show that γ(f ] (a)) ⊆ T (2) ] ˙ eq. (1) is equivalent to ⦃c⦄ ◦ γ ⊆ γ ◦ f by a property of Galois connections. So eq. (2) implies ⦃c⦄(γ(a)) ⊆ T which (by monotonicity of ⦃c⦄) implies ⦃c⦄(IniTrc) ⊆ ⦃c⦄(γ(a)) ⊆ T . The beauty of specification eq. (1) is that f ] can be obtained as an abstract interpretation ⦃c⦄] , derived systematically for all c by calculating from the left side of eq. (1) as shown by Cousot (1999). 3. Domains and Galois Connections for Hyperproperties To express hyperproperties, we need Galois connections for domains that involve sets of sets of observable behaviours. This section spells out how such powerset domains form a hierarchy as illustrated along the top of Figure 2. We describe how dependences and P(P(States∗ )) P(P(Trc)) P(P(States)) P(States∗ ) P(Trc) P(States) abstraction Figure 2. Extended hierarchy of semantic domains (−−−−−→) cardinalities for quantitative information flow can be formulated as Galois connections. We spell out a methodology whereby the standard notions and techniques of abstract interpretation can be applied to specify and derive—in the same form as Equation (1)— static analyses for hyperproperties. As a first example, consider the condition: the final value of x depends only on the initial value of y. Its expression needs, at least, two traces: If two traces, denoted by (σ, σ 0 ) and (τ, τ 0 ), agree on the initial value of y then they agree on the final value of x. That is, σ(y) = τ (y) implies σ 0 (x) = τ 0 (x). This must hold for any two traces of the program. This is equivalent to the following: For all sets T of traces, if traces in T all agree on the initial value of y then they all agree on the final value of x. Later we extend this example to an analysis that infers which dependences hold. Consider the problem of quantifying information flow with mincapacity (Smith 2009). For a program on two integer variables h, l, the problem is to infer how much information is conveyed via l about h: considering some traces that agree on the initial value of l, how many final values are possible for l. For example, the program l := (h mod 2) + l has two final values for l, for each initial l, though there are many possible initial values for h. This cardinality problem generalizes prior work on quantitative flow analysis, where typically low inputs are not considered. Whereas the simple dependence problem can be formulated in terms of 2 traces, the cardinality problem involves trace sets of unbounded size. In the terminology of hyperproperties, it is not a k-safety hyperproperty for any k (Yasuoka and Terauchi 2011, Sec. 3), although it is hypersafety (Clarkson and Schneider 2010). For a fixed k, the problem “variable l has at most k − 1 final values” is k-safety, which means it can be formulated in terms of sets with at most k traces. It turns out that by using Galois connections on sets of sets, we can develop a general theory that encompasses many hyperproperties and which enables derivation of interesting abstract interpreters. For our applications, we use relational traces as the notion of observable behavior, and thus P(P(Trc)). The approach works as well for other notions, so there is a hierarchy of domains as shown at the top of Figure 2, in parallel with the ordinary hierarchy shown along the bottom. The abstractions of this hierarchy are obtained by lifting each abstraction between two standard collecting semantics (Cousot 2002) to their hypercollecting versions, by element-wise abstraction (Lemma 1). For instance, Lemma 1 justifies the abstraction between P(P(Trc)) and P(P(States)), by lifting the abstraction between P(Trc) and P(States) (Cousot 2002, Sec. 8). Additionally, the diagonal lines in Figure 2 represent abstractions between hypercollecting semantics defined over some form of observations and the corresponding collecting semantics defined over the same observations. Lemma 2 . Let C be a set. Define αhpp (C) , ∪C∈C C and γhpp (C) , P(C). These form a Galois connection: γhpp −− −−− − (P(C), ⊆) (P(P(C)), ⊆) ← −− → α −− hpp It is noted by Clarkson and Schneider (2010) that any trace property can be lifted to a unique hyperproperty; this lifting is exactly the concretisation γhpp of Lemma 2. Although the model of Clarkson and Schneider (2010) is quite general, it does focus on infinite traces. But hyperproperties can be formulated in terms of other notions of observation, as illustrated in Figure 2. Cardinality abstraction. To lay the groundwork for our quantitative information flow analysis, we consider abstracting a set of values by its cardinality. Cardinality is one ingredient in many quantitative information flow analyses estimating the amount of sensitive information a program may leak (Smith 2009; Backes et al. 2009; Braun et al. 2009; Köpf and Rybalchenko 2013; Mardziel et al. 2013; Doychev et al. 2013). The lattice of abstract representations we consider is the set [0, ∞] , N ∪ {∞} where ∞ denotes an infinite cardinal number. We use the natural order ≤, and max as a join. Consider the abstraction operator crdval ∈ P(Val) → [0, ∞] computing cardinality and given by crdval(V ) , |V |. This operator crdval is not additive, i.e. it does not preserve joins; e.g. crdval({1, 2} ∪ {2, 3}) 6= max(crdval({1, 2}), crdval({2, 3})). Thus, there exists no associated concretisation f for which crdval is the lower adjoint in a Galois connection. Yet, we can lift the abstraction operator crdval to a Galois connection over P(P(Val)) through what is called a supremus abstraction (Cousot 2002, p.52). Lemma 3 Supremus abstraction. Let elt ∈ C → A be a function from a set C, with codomain forming a complete lattice (A, v). Let αelt (C) , tc∈C elt(c) and γelt (a) , {c ∈ C | elt(c) v a}. Then γelt −− −− − (A, v) (P(C), ⊆) ← −− → α− elt For example, define αcrdval (V) , maxV ∈V crdval(V ) and γcrdval (n) , {V | crdval(V ) ≤ n}. Thus we obtain a Galois γcrdval ← −− −− −− −− − − ([0, ∞] , ≤). connection (P(P(Val)), ⊆) − − → αcrdval As another example let us consider, in simplified form, an ingredient in dependency or noninterference analysis. For program variable x, agreex ∈ P(States) → {tt, ff} determines whether a set of states contains only states that all agree on x’s value: agreex (Σ) , (∀σ, σ 0 ∈ Σ, JxKσ = JxKσ 0 ) Function agreex is not additive, so it is not part of a Galois connection from P(States) to {tt, ff}. The same problem arises with agreements on multiple variables, and with more concrete domains like the finite maximal trace semantics P(States∗ ). We lift the operator agreex to a Galois connection over P(P(States)). A supremus abstraction yields αagreex (S) , (∀Σ ∈ S, agreex (Σ)) γagreex (bv) , {Σ | agreex (Σ) ⇐= bv} abstract concretisation/denotation of a security requirement yields a set of sets of traces, namely a hyperproperty. Hints of this intuition appear in the literature (McLean 1994; Volpano 1999; Rushby 2001; Zakinthinos and Lerner 1997); e.g. security policies “are predicates on sets of traces (i.e. they are higher order)” (Rushby 2001, p.2). However, only recently has a comprehensive framework proposed a sharp characterisation of security policies as hyperproperties (Clarkson and Schneider 2008, 2010). Abstract interpretation of hyperproperties. The basic methodology for the verification of a hyperproperty HP, may be described as follows: Step 1. Design approximate representations forming a complete lattice A, choose a collecting semantics C among the extended hierarchy (set of sets domains, e.g. P(P(Trc))), and define α, γ γ − − (A, v). for a Galois connection (C, ≤) ← −− α→ Step 2. Compute an approximation a ∈ A of the semantics C ∈ C of the program P of interest. Step 3. Prove that the inferred approximation a implies that P satisfies HP. The concretisation γ(a) is a set of trace sets, of which the program’s trace set is a member—by contrast to approximations of trace properties, which infer a single trace set of which the program trace set is a subset. Then, it suffices to prove γ(a) ⊆ HP. Step 1 is guided by the need to have γ(a) ⊆ HP, i.e. a describes a hyperproperty that implies HP. The calculational design (Cousot 1999) of abstract domains greatly systematises Step 2, by relying on the Galois connection defined in Step 1. Collecting semantics can be adapted to the additional structure of sets, as we show in Section 4. 4. Hypercollecting Semantics In the following, we introduce a hypercollecting semantics defined over sets T ∈ P(P(Trc)) of sets of traces. This is used in subsequent sections to derive static analyses. Here is Step 2 of the methodology, spelled out in detail. Given γ − − (A, v] ) built by the a Galois connection (P(P(Trc)), ⊆) ← −− α→ supremus abstraction, and an approximation a of the initial traces (i.e. IniTrc is in γ(a)), find an approximation a0 ∈ A of the analysed program c, i.e. ⦃c⦄ IniTrc is in γ(a0 ). Then prove that the program satisfies the hyperproperty HP of interest, i.e. γ(a0 ) ⊆ HP. In order to compute a0 , we define a hypercollecting semantics LcM ∈ P(P(Trc)) → P(P(Trc)). That will serve to derive—in the manner of Equation (1)—a static analysis that is correct by construction. Hypercollecting semantics γagreex −− −− −− −− − − ({tt, ff}, ⇐=). so that (P(P(States)), ⊆) ← −− → αagree x These examples are consistent with the many formulations of noninterference (e.g. (Goguen and Meseguer 1982; Volpano and Smith 1997; Giacobazzi and Mastroeni 2004; Amtoft and Banerjee 2004; Hunt and Sands 2006)) that motivated the characterisation of information-flow security requirements as hyperproperties (Clarkson and Schneider 2010). Concretising an abstract value a can be seen as defining the denotation of a type expression (as in, for instance, Benton (2004, Sec. 3.3.1) and Hunt and Sands (1991)), i.e. defining the set of objects that satisfy the description a. Thus, concretising tt, when tt is interpreted as “satisfies a property requirement”, naturally yields a set of traces. Concretising tt, where tt is interpreted as “satisfies a security requirement”, yields a set of sets of traces. Intuitively, the most abstract denotation/concretisation of a property requirement is defined in terms of a set of traces. The most L c M ∈ P(P(Trc)) → P(P(Trc)) Lx := eMT , {⦃x := e⦄T | T ∈ T} Lc1 ; c2 MT , Lc2 M ◦ Lc1 MT LskipMT , T Lif b then c1 else c2 MT , {⦃c1 ⦄ ◦ ⦃ grdb ⦄T ∪ ⦃c2 ⦄ ◦ ⦃ grd¬b ⦄T | T ∈ T}   Lwhile b do c MT , Lgrd¬b M lfp⊆ T Lif b then c else skipM Lgrdb MT , {⦃ grdb ⦄T | T ∈ T} Recall from Section 2 that standard collecting semantics is a fixpoint-based formulation that captures the direct image on sets of the underlying program semantics – this is proved, for example, by Cachera and Pichardie (2010); Assaf and Naumann (2016). The fixpoint formulation at the level of sets-of-sets we use is not simply the direct image of the standard collecting semantics. The direct image of the standard collecting semantics would yield a set of (inner) fixpoints over sets of traces, whereas an outer fixpoint over sets of sets of traces enables straightforward application of the fixpoint transfer theorem. Theorem 1 . For all c and all T ∈ P(Trc), ⦃c⦄T is in L c M{T }. For a singleton {T }, the set LcM{T } ∈ P(P(Trc)) is not necessarily a singleton set containing only the element ⦃c⦄T . If c is a loop, LcM{T } yields a set of sets R of traces, where each set R of traces contains only traces that exit the loop after less than k iterations, for k ∈ N. We prove this theorem as corollary of the following: ∀T ∈ P(P(Trc)), {⦃c⦄T | T ∈ T} ⊆ LcMT This is proved by structural induction on commands. For loops, there is a secondary induction on iterations of the loop body. In summary, suppose one wishes to prove program c satisfies hyperproperty HP ∈ P(P(Trc)), i.e. one wishes to prove that ⦃c⦄(IniTrc) ∈ HP. Suppose we have an approximation f ] of the hypercollecting semantics, similarly to eq. (1), i.e. ] α ◦ LcM ◦ γ v̇ f ] (3) Given eq. (3) it suffices to find an abstract value a that approximates IniTrc, i.e. IniTrc ∈ γ(a), and show that: γ(f ] (a)) ⊆ HP (4) ˙ γ ◦ f ] by a property Why? Equation (3) is equivalent to LcM γ ⊆ of Galois connections. So we have ⦃c⦄(IniTrc) ∈ LcM(γ(a)) ⊆ γ(f ] (a)) ⊆ HP using IniTrc ∈ γ(a), the Theorem, and eq. (4). ◦ 5. Information Flow This section gives a number of technical definitions which build up to the definition of Galois connections with which we specify information flow policies explicitly as hyperproperties. When a fixed main program is considered, we refer to it as P and its variables as VarP . Our analyses are parametrised by the program P to analyse, and an initial typing context Γ ∈ VarP → L mapping each variable to a security level l ∈ L for its initial value. We assume (L, v, t, u) is a finite lattice. In the most concrete case, L may be defined as the universal flow lattice, i.e. the powerset of variables P(VarP ), from which all other information flow types can be inferred through a suitable abstraction (Hunt and Sands 2006, Sec. 6.2); the initial typing context is then defined as λx.{x}. Initial l-equivalence and variety. A key notion in information flow is l-equivalence. Two states are l-equivalent iff they agree on the values of variables having security level at most l. We introduce the same notion over a set of traces, requiring that the initial states are l-equivalent. Let us first denote by JeKpre ∈ Trc → Val the evaluation of expression e in the initial state σ of a trace (σ, τ ) ∈ Trc—unlike JeK ∈ Trc → Val which evaluates expression e in the final state τ . Then, we denote by T |=Γ l the judgement that all traces in a set T ⊆ Trc are initially l-equivalent, i.e. they all initially agree on the value of variables up to a security level l ∈ L. For example, in the case that L is the universal flow lattice, T |=Γ {x, y} means ∀t1 , t2 ∈ T, JxKpre t1 = JxKpre t2 ∧ JyKpre t1 = JyKpre t2 . Initial l-equivalence T |=Γ l iff. ∀t1 , t2 ∈ T, ∀x ∈ VarP , Γ(x) v l =⇒ JxKpre t1 = JxKpre t2 T |=Γ l The notion of variety (Cohen 1977) underlies most definitions of qualitative and quantitative information flow security. Information is transmitted from a to b over execution of program P if by “varying the initial value of a (exploring the variety in a), the resulting value in b after P’s execution will also vary (showing that variety is conveyed to b)” (Cohen 1977). We define the l-variety of expression e, as the set of sets of values e may take, when considering only initially l-equivalent traces. The variety is defined first as a function Ol ⦃e⦄ ∈ P(Trc) → P(P(Val)) on trace sets, from which we obtain a function Ol LeM ∈ P(P(Trc)) → P(P(Val)), on sets of trace sets. Intuitively, l-variety of expression e is the variety that is conveyed to e by varying only the input values of variables having a security level l0 such that ¬(l0 v l). Ol ⦃e⦄ l-variety Ol LeM Ol ⦃e⦄ ∈ P(Trc) → P(P(Val)) Ol ⦃e⦄T , {⦃e⦄R | R ⊆ T and R |=Γ l} Ol LeM ∈ P(P(Trc)) → P(P(Val)) Ol LeMT , ∪T ∈T Ol ⦃e⦄T Each set V ∈ Ol ⦃e⦄T of values results from initially lequivalent traces (R |=Γ l for R ⊆ T ). Thus, expression e does not leak sensitive information to attackers having a security clearance l ∈ L if Ol ⦃e⦄T is a set of singleton sets. Indeed, sensitive data for attackers with security clearance l ∈ L is all data having a security level l0 for which attackers do not have access (i.e. ¬(l0 v l) (Denning and Denning 1977)). Thus, if Ol ⦃e⦄T is a set of singleton sets, this means that no matter how sensitive information varies, this variety is not conveyed to expression e. Besides a pedagogical purpose, we define l-variety Ol ⦃e⦄ (resp. Ol LeM) instead of simply lifting the denotational semantics JeK of expressions to sets of traces (resp. sets of sets of traces) since we want to build modular abstractions of traces by relying on underlying abstractions of values. Thus, l-variety enables us to pass information about initially l-equivalent traces to the underlying domain of values by keeping disjoint values that originate from traces that are not initially l-equivalent. Specifying information flow. We now have the ingredients needed to describe information flow for command c, with respect to typing context Γ ∈ VarP → L. A quantitative security metric, introduced by Smith (2009, 2011), relies on min-entropy and mincapacity (Rényi 1961) in order to estimate the leakage of a program. Let us assume a program P that is characterized by a set TP ∈ P(Trc) of traces, i.e. TP , ⦃ P ⦄ IniTrc. For simplicity, assume attackers only observe the value of a single variable x ∈ VarP . (The generalization to multiple variables is straightforward.) The leakage of P, as measured by min-capacity, to attackers having security clearance l ∈ L is defined by MLl , log2 ◦ αcrdval ◦ Ol ⦃x⦄TP (The definition of αcrdval follows Lemma 3.) For our purposes, it suffices to know that this quantity aims to measure, in bits, the remaining uncertainty about sensitive data for attackers with security clearance l. Refer to the original work (Smith 2009) for more details. Leaving aside the logarithm in the definition of MLl , a quantitative security requirement may enforce a limit on the amount of information leaked to attackers with security clearance l ∈ L, by requiring that the l-cardinality of variable x is less than or equal to some non-negative integer k. We denote by SR(l, k, x) the hyperproperty that characterises this security requirement, i.e. the set of program denotations satisfying it: SR(l, k, x) , {T ∈ P(Trc) | αcrdval ◦ Ol ⦃x⦄T ≤ k} Note that SR implicitly depends on the choice of initial typing Γ, as does Ol ⦃x⦄T . The termination-insensitive noninterference policy “the final value of x depends only on the initial values of variables labelled at most l” corresponds to the hyperproperty SR(l, 1, x). Therefore, the program P satisfies SR(l, 1, x) if αcrdval ◦ Ol ⦃x⦄TP ≤ 1. Let T = LPM{IniTrc}. Since TP is in T (Theorem 1), then P satisfies SR(l, 1, x) if αcrdval ◦ Ol LxMT ≤ 1, by monotony of αcrdval and by Ol ⦃x⦄TP ⊆ Ol LxMT from the definition of Ol L−M. 6. Dependences We rely on abstract interpretation to derive a static analysis similar to existing ones inferring dependences (Amtoft and Banerjee 2004; Hunt and Sands 2006; Amtoft et al. 2006; Hunt and Sands 2011). Recall that our analyses are parametrised on a security lattice L and program P. We denote by l ; x an atomic dependence constraint, with l ∈ L and x ∈ VarP , read as “agreement up to security level l leads to agreement on x”. It is an atomic pre-post contract expressing that the final value of x must only depend on initial values having at most security level l. Said otherwise, l ; x states the noninterference of variable x from data that is sensitive for attackers with security clearance l, i.e. all inputs having security level l0 such that ¬(l0 v l). Dependences are similar to information flow types (Hunt and Sands 2006) and are the dual of independences assertions (Amtoft and Banerjee 2004). Both interpretations are equivalent (Hunt and Sands 2006, Sec. 5). D ∈ Dep Dep Lattice of dependence constraints Given a lattice L and program P, define Dep , P({l ; x | l ∈ L, x ∈ VarP }) D 1 v\ D 2 , D 1 ⊇ D 2 agree agree(V ) ∈ , αagree αagree (V) ∈ , γagree γagree (bv) ∈ , αagree deptr deptr(T ) ∈ , αdeptr αdeptr (T) ∈ , γdeptr γdeptr (D) ∈ , γagree P(Val) → {tt, ff} (∀v1 , v2 ∈ V, v1 = v2 ) P(P(Val)) → {tt, ff} ∧V ∈V agree(V ) {tt, ff} → P(P(Val)) {V ∈ P(Val) | agree(V ) ⇐= bv} αdeptr γdeptr P(Trc) → Dep {l ; x | l ∈ L, x ∈ VarP , αagree (Ol ⦃x⦄T )} P(P(Trc)) → Dep t\ T ∈T deptr(T ) Dep → P(P(Trc)) {T | deptr(T ) v\ D} γdeptr −− −−−− − (Dep, v\ ) (P(P(Trc)), ⊆) ← − − → α −−− deptr Note that deptr(T ) is the set of dependences l ; x for which αagree (Ol ⦃x⦄T ) holds. For instance, the initial typing context Γ ∈ VarP → L determines the initial dependences of a program: αdeptr ({IniTrc}) = {l ; x | l ∈ L, x ∈ VarP and αagree (Ol ⦃x⦄ IniTrc)} = {l ; x | l ∈ L, x ∈ VarP and Γ(x) v l} l We derive an approximation OD LeM\ of l-variety Ol LeM. This l \ approximation OD LeM ∈ Dep → {tt, ff}, called l-agreement of expression e, determines whether a set D of dependence constraints guarantees that no variety is conveyed to expression e when the inputs up to security level l are fixed. Notice that we use symbol \ and subscript D here, for contrast with similar notation using ] and subscript C in later sections. l OD LeM\ ∈ Dep → {tt, ff} l OD LnM\ D , tt In the rest of this section, L and P are fixed, together with a typing context Γ ∈ VarP → L. The semantic characterisation of dependences is tightly linked to variety. An atomic constraint l ; x holds if no variety is conveyed to x when the inputs up to security level l are fixed. We use this intuition to define the Galois connections linking the hypercollecting semantics and the lattice Dep, by instantiating the supremus abstraction in Lemma 3. The agreement abstraction approximates a set V ∈ P(P(Val)) by determining whether it contains variety. agree deptr Dependence abstraction l-agreement of expressions D1 t\ D2 , D1 ∩ D2 Agreements abstraction of level at most l. So αagree (Ol ⦃x⦄T ) holds just if there is at most one final value. l OD LxM\ D , (l ; x ∈ D) l l l OD Le1 ⊕ e2 M\ D , OD Le1 M\ D ∧ OD Le2 M\ D l l l OD Le1 cmp e2 M\ D , OD Le1 M\ D ∧ OD Le2 M\ D l Deriving the clauses defining OD L−M\ amounts to a constructive proof of the following. l Lemma 4 . OD LeM\ is sound: ∀e, ∀l, ∀D, αagree ◦ l Ol LeM ◦ γdeptr (D) ⇐= OD LeM\ D . Dependence abstract semantics. We derive a dependence abstract semantics LcM\ by approximating the hypercollecting semantics LcM. This abstract semantics LcM\ ∈ Dep → Dep over-approximates the dependence constraints that hold after execution of a command c, on inputs satisfying initial dependence constraints. We assume a static analysis approximating the variables that a command modifies. Mod ∈ Com → P(V ar) Modifiable variables −−−−− ({tt, ff}, ⇐=) (P(P(Val)), ⊆) ← −− α−−−→ For all c, x, if there exists t, t ∈ Trc such that JcKt = t0 and JxKpre t0 6= JxKt0 , then x ∈ Mod(c). Note that γagree (tt) is {V ∈ P(Val) | agree(V )} and γagree (ff) is P(Val). Also, agree(V ) iff |V | ≤ 1. The dependence abstraction approximates a set T ∈ P(P(Trc)) by a dependence constraint D ∈ Dep. Recall that Ol ⦃x⦄T is the set of final values for variable x in traces t ∈ T that agree on inputs The abstract semantics of assignments x := e discards all atomic constraints related to variable x in the input set D of constraints, and adds atomic constraints l ; x if D guarantees l-agreement for expression e. For conditionals, for each security level l, if the input set D guarantees l-agreement of the conditional guard, γagree agree 0 the abstract semantics computes the join over the dependences of both conditional branches, after projecting to only those atomic constraints related to l (notation π l (−)). If D does not guarantee l-agreement of the conditional guard, atomic constraints related to both l and variables possibly modified are discarded. Intuitively, if D guarantees l-agreement of the conditional guard, then l-agreement over some variable x in both branches guarantees l-agreement over x after the conditional command. Otherwise, the only l-agreements that are guaranteed after the conditional are those that hold before the conditional for variables that are not modified. LcM\ ∈ Dep → Dep Dependence abstract semantics LskipM\ D , D Comparison with previous analyses. Our dependence analysis is similar to the logic of Amtoft and Banerjee (2004) as well as the flowsensitive type system of Hunt and Sands (2006). The relationship between our sets D ∈ Dep of dependence constraints and the type environments ∆ ∈ VarP → L of Hunt and Sands can be formalised by the abstraction: Lc1 ; c2 M\ D , Lc2 M\ ◦ Lc1 M\ D Lx := eM\ D , l {l ; y ∈ D | y 6= x} ∪ {l ; x | l ∈ L, OD LeM\ D} Lif b then c1 else c2 M\ D , let D1 = Lc1 M\ D in let D2 = Lc2 M\ D in let W ( = Mod(if b then c1 else c2 ) in l S π l (D1 ) t\ π l (D2 ) LbM\ D if OD l / W } otherwise l∈L {l ; x ∈ π (D) | x ∈ αhs αhs (D) ∈ , γhs γhs (∆) ∈ , Dep → VarP → L λx. u {l | l ; x ∈ D} (VarP → L) → Dep {l ; x | x ∈ VarP , l ∈ L, ∆(x) v l} This is in fact an isomorphism because of the way we interpret dependences. Indeed, if l ; x holds, then also l0 ; x for all l0 ∈ L such that l v l0 (cf. Corollary 4 in Appendix G.2). This observation suggests reformulating the sets D ∈ Dep of dependence constraints to contain only elements with minimal level, but we refrain from doing so for simplicity of presentation. Our dependence analysis is at least as precise as the type system of Hunt and Sands. To state this result, we denote by ⊥L the bottom element of the lattice L. We also assume that the modified variables is precise enough to simulate the same effect as the program counter used in the type system: Mod(c) is a subset of the variables that are targets of assignments in c. Theorem 3 . For all c, D0 , D ∈ Dep, ∆0 , ∆ ∈ VarP → L, where ⊥L ` ∆0 {c}∆, and D = LcM\ D0 , it holds that: ] \ Lwhile b do cM\ D , lfpv D Lif b then c1 else c2 M π l (D) , {l ; x ∈ D | x ∈ VarP } αhs (D0 ) v̇ ∆0 =⇒ αhs (D) v̇ ∆ . Theorem 2 . The dependence semantics is sound: αdeptr \ ◦ \ LcM ◦ γdeptr v̇ LcM\ . 7. We denote by v̇ the point-wise lifting of the partial order v\ . We can derive this abstract semantics by directly approximating the relational hypercollecting semantics LcM through the dependence Galois connection (αdeptr , γdeptr ). The derivation is by structural induction on commands. It leverages mathematical properties of Galois connections. We start with the specification of the best abstract transformer αdeptr ◦ LcM ◦ γdeptr ∈ Dep → Dep, and successively approximate it to finally obtain the definition of the dependence abstract semantics for each form of command. The derivation is the proof, and the obtained definition of the abstract semantics is correct by construction. Let us showcase the simplest derivation for a sequence of commands in order to illustrate this process: αdeptr ◦ Lc1 ; c2 M ◦ γdeptr = HBy definition of the hypercollecting semanticsI αdeptr ◦ \ Lc2 M ◦ Lc1 M ◦ γdeptr v̇ HBy γdeptr αdeptr ◦ \ ◦ αdeptr is extensive I Lc2 M ◦ γdeptr ◦ αdeptr ◦ Lc1 M ◦ γdeptr v̇ HBy induction hypothesis αdeptr Lc2 M\ ◦ Lc1 M\ ◦ Lattice of cardinality constraints Card C ∈ Card For a program P and lattice L, we say C is a valid set of constraints iff ∀x ∈ VarP , ∀l ∈ L, ∃!n ∈ [0, ∞] , l ; x#n ∈ C . Let Card be the set of valid sets of constraints. It is a complete lattice: C1 v] C2 iff ∀l ; x#n1 ∈ C1 , ∃n2 , l ; x#n2 ∈ C2 ∧ n1 ≤ n2 C1 t] C2 , {l ; x# max(n1 , n2 ) | l ; x#n1 ∈ C1 , l ; x#n2 ∈ C2 } \ LcM ◦ γdeptr v̇ LcM\ I , HTake this last approximation as the definition.I Lc1 ; c2 M\ Cardinality Abstraction Dependence analysis is only concerned with whether variety is conveyed. We refine this analysis by deriving a cardinality abstraction that enumerates variety. We denote by l ; x#n an atomic cardinality constraint where l ∈ L, x ∈ VarP and n ∈ [0, ∞], read as “agreement up to security level l leads to a variety of at most n values in variable x”. Alternatively, we can leverage Galois connections to give the analysis as an approximation of the cardinality analysis. We work this out by Lemmas 6 and 7, introduced in Section 7. In the rest of this section, L and P are fixed, together with a typing context Γ ∈ VarP → L. A valid constraint set is essentially a function from l and x to n. So v] is essentially a pointwise order on functions, and we ensure that v] is antisymmetric. The cardinality abstraction relies on the abstraction αcrdval , introduced in Section 3, in order to approximate l-variety of a variable into a cardinality n ∈ [0, ∞]. crdtr Cardinality abstraction crdtr crdtr(T ) ∈ , αcrdtr αcrdtr (T) ∈ , γcrdtr γcrdtr (C ) ∈ , αcrdtr γcrdtr P(Trc) → Card {l ; x#n | l ∈ L, x ∈ VarP , n = αcrdval (Ol ⦃x⦄T ) } P(P(Trc)) → Card t] T ∈T crdtr(T ) Card → P(P(Trc)) {T | crdtr(T ) v] C } Theorem 4 . The cardinality abstract semantics is sound: γcrdtr −− −−−− (Card, v ) (P(P(Trc)), ⊆) ← −− α −−→ ] crdtr The cardinality abstraction enables us to derive an approximation l l OC LeM] of l-variety Ol LeM. This approximation OC LeM] ∈ Card → [0, ∞], called l-cardinality of expression e, enumerates the l-variety conveyed to expression e assuming a set C ∈ Card of cardinality constraints holds. Note that the infinite cardinal ∞ is absorbing, i.e. ∀n, ∞ × n , ∞. l OC LeM] ∈ Card → [0, ∞] l-cardinality of expressions l OC LnM] C , 1 l OC LxM] C , n where l ; x#n ∈ C l l l OC Le1 ⊕ e2 M] C , OC Le1 M] C × OC Le2 M] C   l l l OC Le1 cmp e2 M] C , min 2, OC Le1 M] C × OC Le2 M] C l LeM] is sound: Lemma 5 . OC ∀e, ∀l, αcrdval ◦ l ˙ OC Ol LeM ◦ γcrdtr ≤ LeM] . We now derive a cardinality abstract semantics by approximating the relational hypercollecting semantics of Section 4. It uses definitions to follow. Cardinality abstract semantics LskipM] C , C abstract semantics of conditionals is also similar to dependences: if the conditional guard does not convey l-variety, then all initially l-equivalent traces follow the same execution path and the join operator (defined as max over cardinality) over both conditional branches over-approximates the l-cardinality after the conditional. Otherwise, the l-cardinality over both conditional branches have to be summed—for the variables that may be modified in the conditional branches—to soundly approximate the l-cardinality after the conditional. LcM] ∈ Card → Card Lc1 ; c2 M] C , Lc2 M] ◦ Lc1 M] C Lx := eM] C , {l ; y#n ∈ C | y 6= x} l ∪{l ; x#n | l ∈ L, x ∈ VarP , n = OC LeM] C } Lif b then c1 else c2 M] C , let C1 = Lc1 M] C in let C2 = Lc2 M] C in let W ( = Mod(if b then c1 else c2 ) in l S π l (C1 ) t] π l (C2 ) if OC LbM] C = 1 l ] l l∈L π (C1 ) t add(W,π l (C )) π (C2 ) otherwise ] ] Lwhile b do cM] C , lfpv C Lif b then c1 else c2 M π l (C ) , {l ; x#n ∈ CS | x ∈ VarP , n ∈ [0, ∞]} C1 t] add(W,C0 ) C2 , x∈Var S P \W {l ; x#n ∈ C0 } ∪ x∈W {l ; x#(n1 +n2 ) | l ; x#nj ∈ Cj , j = 1, 2} The abstract semantics of assignments x := e is similar in spirit to the one for dependences: discard atomic constraints related to x, and add new ones by computing l-cardinality of expression e. The αcrdtr ◦ ] LcM ◦ γcrdtr v̇ LcM] . The lattice Card is complete, although not finite. We may define a widening operator ∇ ∈ Card × Card → Card to ensure convergence of the analysis (Cousot and Cousot 1992)(Nielson et al. 1999)(Cortesi and Zanioli 2011, Sec. 4). C1 ∇C2 , {l ; x#n | l ; x#n1 ∈ C1 , l ; x#n2 ∈ C2 , n = n1 ∇n2 } n1 ∇n2 , if (n2 ≤ n1 ) then n1 else ∞ The occurrence of widening depends on the iteration strategy employed by the static analyser. Widening accelerates or forces the convergence of fixpoint computations. In the simplest setting, the analyser passes as arguments to the widening operator the old set C1 of cardinality as well as the new set C2 that is computed. For each atomic cardinality constraint, the widening operator then compares the old cardinality n1 to the new cardinality n2 . If the cardinality is still strictly increasing (n2 > n1 ), the widening forces the convergence by setting it to ∞. If the cardinality is decreasing, the widening operator sets it to the maximum cardinality n1 in order to force convergence and ensure the sequence of computed cardinalities is stationary. Min-capacity leakage. So far, we showed how one can derive static analyses of hyperproperties—the abstract representations themselves are interpreted as hyperproperties—by approximating hypercollecting semantics. Let us now recall the security requirement SR(l, k, x) introduced in Section 4 in order to illustrate how these analyses may prove that a program satisfies a hyperproperty, i.e. Step 3 of the methodology in Section 3 (see also Equation (4)). Consider a program P characterised by a set TP ∈ P(Trc) of traces, i.e. TP is ⦃ P ⦄ IniTrc. How do we prove that P satisfies the hyperproperty SR(l, k, x)? We can use the cardinality analysis to prove that variable x has a l-cardinality that is at most k. Indeed, if C approximates TP (i.e. αcrdtr ({TP }) v] C ) then αcrdval ◦ l Ol ⦃x⦄TP ≤ OC LxM] C . Thus, if the inferred l-cardinality of C is at most k then program P is guaranteed to satisfy the hyperproperty SR(l, k, x). We have {TP } ⊆ γcrdtr (C ) since C approximates TP (i.e. αcrdtr ({TP }) v] C ). And we have γcrdtr (C ) ⊆ SR(l, k, x) l by assumption OC LxM] C ≤ k. Hence TP ∈ SR(l, k, x). The hyperproperty SR(l, k, x) is a (k + 1)-safety hyperproperty (Clarkson and Schneider 2010), i.e. it requires exhibiting at most k + 1 traces in order to prove that a program does not satisfy SR(l, k, x). For example, termination-insensitive noninterference for security level l, which corresponds to the hyperproperty SR(l, 1, x), is 2-safety. A k-safety hyperproperty of a program can be reduced to a safety property of a k-fold product program (Barthe et al. 2004; Terauchi and Aiken 2005; Darvas et al. 2005; Clarkson and Schneider 2010). Various quantitative information flow properties are not k-safety. For example, the bounding problem that the cardinality analysis targets, namely min-capacity leakage is not a k-safety hyperproperty for any k (Yasuoka and Terauchi 2011, Sec. 3). Instead, this bounding problem is hypersafety (Clarkson and Schneider 2010). Cardinalities vs. dependences. Just as quantitative security metrics are the natural generalisations of qualitative metrics such as noninterference, the cardinality abstraction is a natural generalisation of dependence analysis. Instead of deciding if variety is conveyed, the cardinality analysis enumerates this variety. In other words, dependences are abstractions of cardinalities. We can factor the Galois connections, e.g. (αagree , γagree ) is (αlqone ◦ αcrdval , γcrdval ◦ γlqone ) for suitable (αlqone , γlqone ). improve precision. For simplicity, we consider a two point lattice {L, H} and an initial typing context where variables yi are the only low variables (Γ(yi ) = L). As is usual, low may flow to high (L v H). Consider the following program. 1 2 3 Lemma 6 . (αagree , γagree ) is the composition of two Galois connections (αcrdval , γcrdval ) and (αlqone , γlqone ) : 4 γlqone γ Listing 1. Leaking 1 bit of secret crdval −− − −−− − ([0, ∞] , ≤) ← −−−−− ({tt, ff}, ⇐=) (P(P(Val)), ⊆) ← −− → −− α −−− α−−−→ crdval lqone The cardinality abstraction determines that x has at most 2 values after the execution of the program in Listing 1, for initially L-equivalent traces. For fixed low inputs, x has one value in the then branch and one value in the else branch, and these cardinalities get summed after the conditional since the conditional guard may evaluate to 2 different values. Thus, the cardinality abstraction proves that this example program satisfies the hyperproperty SR(L, 2, x). with: ( tt ff ( 1 γlqone (bv) , ∞ αlqone (n) , if n ≤ 1 , and otherwise. if bv = tt otherwise. Stronger trace properties. Another way of proving a hyperproperty is by proving a stronger trace property. If a program is proven to satisfy a trace property T ∈ P(Trc), then proving that T is stronger than hyperproperty H ∈ P(P(Trc))—in the sense that γhpp (T ) ⊆ H—guarantees the program satisfies the hyperproperty H. For instance, by proving for some program that an output variable x ranges over an interval of integer values whose size is k, we can prove that program satisfies SR(L, k, x). However, approximating a hyperproperty by a trace property may be too coarse for some programs, as we can illustrate with an interval analysis (Cousot and Cousot 1977) on the example program in Listing 1. Such an interval analysis loses too much precision in the initial state of this program, since it maps all low input variables y1 , y2 and y3 to [−∞, +∞]. After the conditional, it determines that x belongs to the interval [−∞, +∞], which is a coarse overapproximation. Also, a polyhedron (Cousot and Halbwachs 1978) does not capture the disjunction that is needed for this example program (x = y2 or x = y3 ). Both abstract domains and many more existing ones are not suitable for the task of inferring cardinalities or dependences because they are convex. Using them as a basis to extract counting information delivers an over-approximation of the leakage, but a coarse one, especially in the presence of low inputs. A disjunction of two polyhedra —through powerset domains, disjunctive postconditions, or partitioning (Bourdoncle 1992)— is as precise as the cardinality analysis for this example. However, disjunctions are not tractable in general. As soon as one fixes a maximum number of disjunctive elements (as in the quantitative information flow analysis of Mardziel et al. (2011, 2013)) or defines a widening operator to guarantee convergence, one loses the relative precision wrt. classical dependence analyses (Amtoft and Banerjee 2004; Hunt and Sands 2006) that the cardinality analysis guarantees (Cf. Corollary 1). Future work will investigate relying on cardinality analysis as a strategy guiding trace partitioning (Rival and Mauborgne 2007). Combining our analyses with existing domains will also deliver better precision. Consider the following program. Lemma 7 . (αdeptr , γdeptr ) is the composition of two Galois connections (αcrdtr , γcrdtr ) and (αlqonecc , γlqonecc ) : γlqonecc γ crdtr −− −−−− (Card, v] ) ← −− −−−−− − (Dep, v\ ) (P(P(Trc)), ⊆) ← −− −− → α −−→ α −−−− crdtr lqonecc with: αlqonecc (C ) , {l ;Sx | l ; x#n ∈ C and αlqone (n)} γlqonecc (D) , {l ; x#n | n = γlqone (l ; x ∈ D)} l∈L,x∈VarP We use Lemmas 6 and 7 to abstract further the cardinality abstract semantics and derive the correct by construction dependence analysis of Section 6. This derivation, which can be found in Appendix G, proves Lemma 4 and Theorem 2 stated earlier. As a corollary and by Theorem 3, this also proves the precision of the cardinality analysis relative to Amtoft and Banerjee’s logic (Amtoft and Banerjee 2004) as well as Hunt and Sands’ type system (Hunt and Sands 2006, 2011). Corollary 1 No leakage for well-typed programs. For all c, C0 , C ∈ Card, ∆0 , ∆ ∈ VarP → L, where ⊥L ` ∆0 {c}∆, and C = LcM] C0 , it holds that: αhs αlqonecc (C0 ) v̇ ∆0 =⇒   l ∀x ∈ VarP , l ∈ L, ∆(x) v l =⇒ OC LxM] ≤ 1 ◦ The cardinality analysis determines that there is no leakage for programs that are “well-typed” by the flow-sensitive type system of Hunt and Sands. By “well-typed”, we mean that the final typing environment that is computed by the type system allows attackers with security clearance l ∈ L to observe a variable x ∈ VarP . To the best of our knowledge, the cardinality abstraction is the first approximation-based analysis for quantitative information flow that provides a formal precision guarantee wrt. traditional analyses for qualitative information flow. This advantage makes the cardinality analysis appealing even when interested in proving a qualitative security policy such as non-interference, since the cardinality abstraction provides quantitative information that may assist in making better informed decisions if declassification is necessary. Nonetheless, we need further experimentation to compare to other quantitative analyses —see Section 9. 8. Towards More Precision This section introduces examples to evaluate the precision of the analyses, and shows how existing analyses can be leveraged to if (y1 ≥ secret ) then x := y2 else x := y3 1 2 if (y1 ≥ secret ) then x := y2 else x := y3 ; o := x * y4 Listing 2. Leaking x The cardinal abstraction determines that variable o leaks the two possible values of x: for fixed low inputs, x has two possible values whereas y4 has one possible value. Relational abstract domains such as polyhedra (Cousot and Halbwachs 1978) or octogons (Miné 2006a) do not support non-linear expressions, and therefore are unable to compute a precise bound of the leakage for variable o. Consider an analysis with a disjunction {x = y2 ∨ x = y3 } of polyhedra and linearisation over intervals (Miné 2006b). Linearisation of expressions y2 ∗y4 and y3 ∗y4 will compute the following constraints for variable o: {(o = y2 ∗ [−∞, +∞]) ∨ (o = y3 ∗ [−∞, +∞])} if linearisation happens for the right side of expressions, or constraint {(o = [−∞, +∞] ∗ y4 ) ∨ (o = [−∞, +∞] ∗ y4 )} if linearisation happens for the left side expressions. Two more combinations of constraints are possible, but none will deduce that variable o has at most 2 values, because the underlying domain of intervals lacks the required precision. Linearisation over both intervals and cardinalities delivers better precision. We can now also improve the dependence abstraction: αlqonecc αcrdval ◦ l Ol Lx2 M ◦ Lgrdx1 ==x2 M ◦ γcrdtr (C ) ≤ OC Lx1 M] C Therefore, we can deduce that: αcrdtr ◦ Lgrdx1 ==x2 M ◦ γcrdtr (C ) v {l ; x#n ∈ C | x 6= x1 , x 6= x2 } ] ∪ {l ; x1 # min(n1 , n2 ), l ; x2 # min(n1 , n2 ) | l ; x1 #n1 ∈ C , l ; x2 #n2 ∈ C } , Lgrdx1 ==x2 M] C For other comparison operators, we use as before Lgrdb M] C , C . l ; x1 #n1 ∈ γlqonecc (D), l ; x2 #n2 ∈ γlqonecc (D)}) For other comparison operators, we also use Lgrdb M\ D , D. With these new definitions, we can update the abstract semantics of conditionals and loops, for both dependences and cardinalities, to leverage the transfer functions Lgrd− M\ and Lgrd− M] . Improved dependences abstract semantics LcM\ ∈ Dep → Dep Lif b then c1 else c2 M\ D , let D1 = Lgrdb M\ ◦ Lc1 M\ D in let D2 = Lgrd¬b M\ ◦ Lc2 M\ D in let W ( = Mod(if b then c1 else c2 ) in l S π l (D1 ) t\ π l (D2 ) LbM\ D if OD l / W } otherwise l∈L {l ; x ∈ π (D) | x ∈ The cardinality abstraction determines that initially L-equivalent memories lead to a variety of at most 2 in the pointer p after the conditional, whereas both y2 and y3 have a variety of 1. Assuming an aliasing analysis determines that p may point to y2 or y3 , the cardinality analysis determines that variable o has a variety of at most 2, for initially L-equivalent memories. l Ol Lx1 M ◦ Lgrdx1 ==x2 M ◦ γcrdtr (C ) ≤ OC Lx2 M] C γlqonecc (D) , Lgrdx1 ==x2 M\ D Listing 3. Leaking 1 bit of secret ◦ ◦ v {l ; x ∈ D | x 6= x1 , x 6= x2 } ∪ {l ; x1 , l ; x2 | l ; x1 ∈ D or l ; x2 ∈ D} \ if (y1 ≥ secret ) then p := &y2 else p := &y3 o := * p αcrdval Lgrdx1 ==x2 M] v αlqonecc ({l ; x#n ∈ γlqonecc (D) | x 6= x1 , x 6= x2 }) ∪ αlqonecc ({l ; x1 # min(n1 , n2 ), l ; x2 # min(n1 , n2 ) | \ Scaling to richer languages. We can rely on existing abstract domains to support richer language constructs, e.g. pointers and aliasing. Consider the following variation of Listing 1. Improving precision. To improve precision of the cardinality abstraction, we can augment it with existing abstract domains. One shortcoming of the cardinality analysis is the fact that it is not relational. Assuming attackers with security clearance L observe both variables x and o after execution of the program in Listing 2, the cardinality abstraction leads us to compute a leakage of two bits: four different possible values, instead of only 2 possible values for initially L-equivalent memories. Relying on a relational domain with linearisation (Miné 2006b) over cardinalities captures the required constraints {L ; x#2, L ; o#1 ∗ x} to compute a leakage of only one bit; these constraints are to be interpreted as “initially L-equivalent memories result in o being equal to one fixed integer times x, and x having at most 2 values”. We leave these extensions of cardinality analysis —and its abstraction as dependence analysis— for future work. In the following, we focus on one particular improvement to both previous analyses in order to gain more precision. We uncovered this case while deriving the analyses, by relying on the calculational framework of abstract interpretation. Indeed, notice that the following holds: ◦ Lwhile b do cM\ D , Lgrd¬b M\ ] ◦ \ lfpv D Lif b then c1 else c2 M Improved cardinality abs. semantics LcM] ∈ Card → Card Lif b then c1 else c2 M] C , let C1 = Lgrdb M] ◦ Lc1 M] C in let C2 = Lgrd¬b M] ◦ Lc2 M] C in let W ( = Mod(if b then c1 else c2 ) in l S π l (C1 ) t] π l (C2 ) LbM] C = 1 if OC l ] l l∈L π (C1 ) t add(W,π l (C )) π (C2 ) otherwise Lwhile b do cM] C , Lgrd¬b M\ ] ◦ ] lfpv C Lif b then c1 else c2 M To illustrate the benefits of this improvement, consider the following example. 1 2 3 4 5 while ( secret != y3 ) do { x := x +1; secret := secret - 1; } o := secret ; Listing 4. Improved precision The cardinality analysis determines that initially L-equivalent memories result in x having an infinity of values: the L-cardinality of x grows until it is widened to ∞. In contrast, cardinalities also determine that variables o and secret have only 1 value, assuming Lequivalent memories. This is because of the reduction that concerns variable secret after the while loop, specifically Lgrdsecret==y3 M\ . Similarly, the improved dependence analysis also determines that both variables secret and o are low. These are sound precision gains for termination-insensitive noninterference; Askarov et al. (2008) discusses the guarantees provided by this security requirement. Remarkably, this has been overlooked by many previous analyses. In fact, this simple improvement makes our dependence analysis strictly more precise than Amtoft and Banerjee (2004)’s and Hunt and Sands (2006, 2011)’s analyses and incomparable to the more recent dependence analysis of Müller et al. (2015). Combination with intervals. Consider now the following example inspired from Müller et al. (2015). 0 1 2 3 4 1 2 3 4 5 6 7 if ( secret == 0) then { x := 0; y := y + 1; } else { x := 0; } 5 6 7 8 9 10 11 Listing 5. Example program from Müller et al. (2015) The analysis of Müller et al. (2015) determines that x is low, whereas the cardinality abstraction determines that L-equivalent memories result in at most 2 values for variable x, because it does not track the actual values of variables. We can combine cardinality with an interval analysis to be more precise in such cases, through a reduced product (Cousot and Cousot 1979; Granger 1992; Cortesi et al. 2013). Assume a set StInt of interval environments provided with ˙ ],Int . Assume also the usual partial order that we denote by ≤ Int Int a Galois connection (α , γ ) enabling the derivation of an interval analysis as an approximation of a standard collecting semantics defined over P(Trc). We can lift this Galois connection to P(P(Trc)) to obtain a Galois connection by compositing with (αhpp , γhpp ), to obtain (α0 , γ 0 ) , (αInt ◦ αhpp , γ Int ◦ γhpp ) with: γhpp γ Int hpp αInt ˙ ],Int ) −− −−− − (P(Trc), ⊆) ← −− −− − − (StInt, ≤ (P(P(Trc)), ⊆) ← −− → −− → α −− A Granger’s reduced product Granger (1992) for the cardinality abstraction and an interval analysis may be defined as a pair of functions toint ∈ Card × StInt → StInt and tocard ∈ Card × StInt → Card verifying the following conditions: 1. soundness: γ 0 (toint(C , ı)) ∩ γcrdtr (C ) γ 0 (ı) ∩ γcrdtr (tocard(C , ı)) 2. reduction: toint(C , ı) tocard(C , ı) = γ 0 (ı) ∩ γcrdtr (C ) = γ 0 (ı) ∩ γcrdtr (C ) ˙ ],Int ı ≤ v] C Let us denote by size the function that returns the size of an interval. One such Granger’s reduced product can be defined as: tocard tocard(C , ı) ∈ , toint toint(C , ı) ∈ , Card × StInt → Card {l ; x#n0 | l ; x#n ∈ C and n0 = min (n, size ı(x))} Card × StInt → Card ı Once enhanced with this reduced product, the cardinality analysis determines for the program in Listing 5, that L-equivalent memories result in at most one possible value for variable x. The dependence analysis can be improved similarly, with a reduction function defined as follows: todep ∈ Dep × StInt → Dep todep(D, ı) , D ∪ {l ; x | l ∈ L and size ı(x) = 1} Once extended with a reduced product with intervals, the dependence analysis is also able to determine that variable x is low for the program in Listing 5. 12 13 //L ; h#∞, L ; y1 #1, L ; y2 #1, L ; y3 #1 y1 := 1; //L ; y1 #1 if ( h == y1 ) then { skip ; //L ; h#1, L ; y1 #1, L ; y2 #1 } else { y2 := 5; //L ; y1 #1, L ; y2 #1 while (y2 != 1) do { y2 := y2 -1; //L ; y2 #1 y1 := y2 ; //L ; y1 #1 } //L ; y1 #1, L ; y2 #1 } //L ; h#∞, L ; y1 #2, L ; y2 #2, L ; y3 #1 o := y1 * y3 ; //L ; o#2 Listing 6. No leakage for variable o More reduced products. As a final example, let us consider Listing 6, inspired by Besson et al. (2016, program 7), that we annotate with the result of the improved cardinality abstraction. To the best of our knowledge, no existing automated static analysis determines that variable o is low at the end of this program. Also, no prior monitor but the one recently presented by Besson et al. (2016) accepts all executions of this program, assuming attackers with clearance L can observe variable o. For initially L-equivalent memories, the cardinality abstraction determines that variables y1 , y2 and o have at most two values. This result is precise for y2 , but not precise for y1 and o. As a challenge, let us see what is required to gain more precision to determine that both variables y1 and o have at most 1 possible value – they are low. To tackle this challenge, we need to consider cardinality combined with an interval analysis and a simple relational domain tracking equalities. With the equality y1 = y2 at the exit of the loop, both y1 and y2 will be reduced to the singleton interval [1, 1]. After the conditional, we still deduce that y2 has at most 2 different values thanks to the cardinality abstraction. Using intervals, we deduce that variable y1 has only one value (singleton interval [1, 1]). And finally, at the last assignment the cardinalities abstraction determines that variable o has only one possible value. Similarly, this same combination of analyses can be put to use to let the dependence analysis reach the desired precision. 9. Related Work Although noninterference has important applications, for many security requirements it is too strong. That is one motivation for research in quantitative information flow analysis. In addition, a number of works investigate weakenings of noninterference and downgrading policies that are conditioned on events or data values (Askarov and Sabelfeld 2007; Banerjee et al. 2008; Sabelfeld and Sands 2009; Mastroeni and Banerjee 2011). Assaf (2015, Chapter 4) proposes to take the guarantees provided by termination-insensitive noninterference (Askarov et al. 2008) as an explicit definition for security; this Relative Secrecy requirement is inspired by Volpano and Smith (2000) who propose a type-system preventing batch-job programs from leaking secrets in polynomial time. Giacobazzi and Mastroeni (2004) introduce abstract noninterference, which generalizes noninterference by means of abstract interpretations that specify, for example, limits on the attacker’s power and the extent of partial releases (declassification). The survey by Mastroeni (2013) further generalizes the notion and highlights, among other things, its applicability to a range of underlying semantics. The Galois connections in this work are at the level of trace sets, not sets of sets. Abstract noninterference retains the explicit 2-run formulation (Volpano et al. 1996; Sabelfeld and Myers 2003): from two related initial states, two executions lead to related final states. The relations are defined in terms of abstract interpretations of the individual states/executions. Mastroeni and Banerjee (2011) show how to infer indistinguishability relations—modelling attackers’ observations—to find the best abstract noninterference policy that holds. The inference algorithm iteratively refines the relation by using counter-examples and abstract domain completion (Cousot and Cousot 1979). Set-of-sets structures occur in work on abstraction for nondeterministic programs, but in those works one level of sets are powerdomains for nondeterminacy; the properties considered are trace properties (Schmidt 2009, 2012). Hunt and Sands (1991) develop a binding time analysis and a strictness analysis (Hunt 1990) based on partial equivalence relations: Their concretisations are sets of equivalence classes. Cousot and Cousot (1994) point out that this analysis could be achieved by a collecting semantics over sets-of-sets, defined simply as a direct image. To the best of our knowledge this has not been explored further in the literature, except in unpublished work on which this paper builds (Assaf 2015; Assaf et al. 2016b). Clarkson et al. (2014); Finkbeiner et al. (2015) extend temporal logic with means to quantify over multiple traces in order to express hyperproperties, and provide model checking algorithms for finite space systems. Agrawal and Bonakdarpour (2016) introduce a technique for runtime verification of k-safety properties. The dependences analysis we derive is similar to the information flow logic of Amtoft and Banerjee (2004) and the equivalent flow-sensitive type system of Hunt and Sands (2006). Amtoft and Banerjee use the domain P(Trc) and on the basis of a relational logic they validate a forward analysis. In effect their interpretation of “independences” is a Galois connection with sets of sets, but the analysis is not formulated or proved correct as an abstract interpretation. To deal with dynamically allocated state, Amtoft et al. (2006) augment the relational assertions of information flow logic with region assertions, which can be computed by abstract interpretation. This is used both to express agreement relations between the two executions and to approximate modifiable locations. This approach is generalized in Banerjee et al. (2016) to a relational Hoare logic for object-based programs that encompasses information flow properties with conditional downgrading (Banerjee et al. 2008). Müller et al. (2015) give a backwards analysis that infers dependencies and is proved strictly more precise than (Hunt and Sands 2006; Amtoft and Banerjee 2004). This is achieved by product construction that facilitates inferring relations between variables in executions that follow different control paths. Correctness of the analysis is proved by way of a relational Hoare logic. The variations of our proposed analyses, in Section 8, rivals theirs in terms of precision—they are incomparable. Our dependence analysis relies on an approximation of the modifiable variables, to soundly track implicit flows due to control flow, instead of labelling a program counter variable pc to account for implicit flows (Sabelfeld and Myers 2003). Zanioli and Cortesi (2011) also derive a similar analysis through a syntactic Galois connection— a syntactic assignment z := x ∗ y is abstracted into a propositional formula x → z∧y → z denoting an information flow from variables x and y to variable z. The soundness of this analysis wrt. a semantic property such as noninterference requires more justification, though it is remarkable that the concretisation of propositional formula yields, roughly speaking, a set of program texts. Zanotti (2002) also provides an abstract interpretation account of a flow-insensitive type system (Volpano et al. 1996) enforcing noninterference by guaranteeing a stronger safety property, namely that sensitive locations should not influence public locations (Boudol 2008). Kovács et al. (2013) explicitly formulate termination-insensitive noninterference as an abstract interpretation, namely the “merge over all twin computations” that makes explicit both the 2-safety aspect and the need for an analysis to relate some aligned intermediate states. Their analysis, like many others, is based on reducing the problem to a safety property of product programs. Sousa and Dillig (2016) implement an algorithm that automates reasoning in a Hoare logic for k-safety, implicitly constructing product programs; the performance compares favorably with explicit construction of product programs. Program dependency graphs are another approach to dependency, shown to be correct for noninterference by Wasserrab et al. (2009) using slicing and a simulation argument. Denning (1982, Chap. 5) proposes the first quantitative measure of a program’s leakage in terms of Shannon entropy (Shannon 1948). Other quantitative metrics emerge in the literature (Braun et al. 2009; Clarkson et al. 2009; Smith 2009; Dwork 2011; Smith 2011; Alvim et al. 2012). These quantitative security metrics model different scenarios suitable for different policies. Most existing static analyses for quantitative information flow leverage existing model checking tools and abstract domains for safety; they prove that a program satisfies a quantitative security requirement by proving a stronger safety property. In contrast, the cardinal abstraction proves a hyperproperty by inferring a stronger hyperproperty satisfied by the analysed program. This is key to target quantitative information flow in mutlilevel security lattices, beyond the 2-point lattice {L, H}. Backes et al. (2009) synthesize equivalence classes induced by outputs over low equivalent memories by relying on software model checkers, in order to bound various quantitative metrics. Heusser and Malacaria (2009) also rely on a similar technique to quantify information flow for database queries. Köpf and Rybalchenko (2010) note that the exact computation of information-theoretic characteristics is prohibitively hard, and propose to rely on approximation-based analyses, among which are randomisation techniques and abstract interpretation ones. They also propose to rely on a self-composed product program to model a scenario where attackers may refine their knowledge by influencing the low inputs. Klebanov (2014) relies on similar techniques to handle programs with low inputs, and uses polyhedra to synthesize linear constraints (Cousot and Halbwachs 1978) over variables. Mardziel et al. (2013) decide whether answering a query on sensitive data augments attackers’ knowledge beyond a certain threshold, by using probabilistic polyhedra. 10. Conclusion Galois connection-based semantic characterisations of program analyses provide new perspectives and insights that lead to improved techniques. We have extended the framework to fully encompass hyperproperties, through a remarkable form of hypercollecting semantics that enables calculational derivation of analyses. This new foundation raises questions too numerous to list here. One promising direction is to combine dependence and cardinality analysis with existing abstract domains, e.g. through advanced symbolic methods (Miné 2006b), and partitioning (Handjieva and Tzolovski 1998; Rival and Mauborgne 2007). Static analysis of secure information flow has yet to catch up with recent advances in dynamic information flow monitoring (Besson et al. 2013; Bello et al. 2015; Hedin et al. 2015; Assaf and Naumann 2016; Besson et al. 2016). We discussed, in Section 8, how existing static analyses may be of use to statically secure information flow. It seems likely that hypercollecting semantics will also be of use for dynamic analyses. Acknowledgments Thanks to Anindya Banerjee and the anonymous reviewers for thoughtful comments and helpful feedback. This work was partially supported by NSF awards CNS-1228930 and CCF-1649884, ANR project AnaStaSec ANR-14-CE28-0014 and a CFR CEA Phd Fellowship. References S. Agrawal and B. Bonakdarpour. Runtime verification of k-safety hyperproperties in HyperLTL. In IEEE Computer Security Foundations Symposium, pages 239–252, 2016. M. S. Alvim, K. Chatzikokolakis, C. Palamidessi, and G. Smith. Measuring information leakage using generalized gain functions. In IEEE Computer Security Foundations Symposium, pages 265–279, 2012. T. Amtoft and A. Banerjee. Information flow analysis in logical form. In Static Analysis Symposium, pages 100–115, 2004. T. Amtoft, S. Bandhakavi, and A. Banerjee. A logic for information flow in object-oriented programs. In ACM Symposium on Principles of Programming Languages, pages 91–102, 2006. A. Askarov and A. Sabelfeld. Gradual release: Unifying declassification, encryption and key release policies. In IEEE Symposium on Security and Privacy, 2007. A. Askarov, S. Hunt, A. Sabelfeld, and D. Sands. Termination-insensitive noninterference leaks more than just a bit. In European Symposium on Research in Computer Security, volume 5283 of LNCS, 2008. M. Assaf. From Qualitative to Quantitative Program Analysis : Permissive Enforcement of Secure Information Flow. PhD thesis, Université de Rennes 1, May 2015. https://hal.inria.fr/tel-01184857. M. Assaf and D. Naumann. Calculational design of information flow monitors. In IEEE Computer Security Foundations Symposium, pages 210–224, 2016. M. Assaf, D. Naumann, J. Signoles, É. Totel, and F. Tronel. Hypercollecting semantics and its application to static analysis of information flow. Technical report, Apr. 2016a. URL https://arxiv.org/abs/1608. 01654. M. Assaf, J. Signoles, É. Totel, and F. Tronel. The cardinal abstraction for quantitative information flow. In Workshop on Foundations of Computer Security (FCS), June 2016b. https://hal.inria.fr/hal01334604. M. Backes, B. Köpf, and A. Rybalchenko. Automatic discovery and quantification of information leaks. In IEEE Symposium on Security and Privacy, pages 141–153. IEEE, 2009. A. Banerjee, D. A. Naumann, and S. Rosenberg. Expressive declassification policies and modular static enforcement. In IEEE Symposium on Security and Privacy, pages 339–353, 2008. A. Banerjee, D. A. Naumann, and M. Nikouei. Relational logic with framing and hypotheses. In 36th IARCS Annual Conference on Foundations of Software Technology and Theoretical Computer Science, 2016. To appear. G. Barthe, P. R. D’Argenio, and T. Rezk. Secure information flow by selfcomposition. In IEEE Computer Security Foundations Workshop, pages 100–114, 2004. L. Bello, D. Hedin, and A. Sabelfeld. Value sensitivity and observable abstract values for information flow control. In Logic for Programming, Artificial Intelligence, and Reasoning (LPAR), pages 63–78, 2015. N. Benton. Simple relational correctness proofs for static analyses and program transformations. In ACM Symposium on Principles of Programming Languages, pages 14–25, 2004. J. Bertrane, P. Cousot, R. Cousot, J. Feret, L. Mauborgne, A. Miné, and X. Rival. Static analysis and verification of aerospace software by abstract interpretation. In AIAA Infotech@Aerospace 2010, 2012. J. Bertrane, P. Cousot, R. Cousot, J. Feret, L. Mauborgne, A. Miné, and X. Rival. Static analysis and verification of aerospace software by abstract interpretation. Foundations and Trends in Programming Languages, 2 (2-3):71–190, 2015. F. Besson, N. Bielova, and T. Jensen. Hybrid information flow monitoring against web tracking. In IEEE Computer Security Foundations Symposium, pages 240–254. IEEE, 2013. F. Besson, N. Bielova, and T. Jensen. Hybrid monitoring of attacker knowledge. In IEEE Computer Security Foundations Symposium, pages 225–238, 2016. G. Boudol. Secure information flow as a safety property. In Formal Aspects in Security and Trust, pages 20–34, 2008. F. Bourdoncle. Abstract interpretation by dynamic partitioning. Journal of Functional Programming, 2(04):407–435, 1992. C. Braun, K. Chatzikokolakis, and C. Palamidessi. Quantitative notions of leakage for one-try attacks. In Mathematical Foundations of Programming Semantics (MFPS), volume 249, pages 75–91, 2009. D. Cachera and D. Pichardie. A certified denotational abstract interpreter. In Interactive Theorem Proving (ITP), pages 9–24, 2010. M. R. Clarkson and F. B. Schneider. Hyperproperties. In IEEE Computer Security Foundations Symposium, pages 51–65, 2008. M. R. Clarkson and F. B. Schneider. Hyperproperties. Journal of Computer Security, 18(6):1157–1210, 2010. M. R. Clarkson, A. C. Myers, and F. B. Schneider. Quantifying information flow with beliefs. Journal of Computer Security, 17:655–701, 2009. M. R. Clarkson, B. Finkbeiner, M. Koleini, K. K. Micinski, M. N. Rabe, and C. Sánchez. Temporal logics for hyperproperties. In Principles of Security and Trust, volume 8414 of LNCS, pages 265–284, 2014. E. Cohen. Information transmission in computational systems. In Proceedings of the sixth ACM Symposium on Operating Systems Principles, pages 133–139, 1977. A. Cortesi and M. Zanioli. Widening and narrowing operators for abstract interpretation. Computer Languages, Systems & Structures, pages 24–42, 2011. A. Cortesi, G. Costantini, and P. Ferrara. A survey on product operators in abstract interpretation. In Semantics, Abstract Interpretation, and Reasoning about Programs: Essays Dedicated to David A. Schmidt on the Occasion of his Sixtieth Birthday, volume 129 of EPTCS, pages 325–336, 2013. P. Cousot. The calculational design of a generic abstract interpreter. In M. Broy and R. Steinbrüggen, editors, Calculational System Design, volume 173, pages 421–506. NATO ASI Series F. IOS Press, Amsterdam, 1999. P. Cousot. Constructive design of a hierarchy of semantics of a transition system by abstract interpretation. Theoretical Computer Science, 277 (1-2):47–103, 2002. P. Cousot and R. Cousot. Abstract interpretation: a unified lattice model for static analysis of programs by construction or approximation of fixpoints. In ACM Symposium on Principles of Programming Languages, pages 238–252, 1977. P. Cousot and R. Cousot. Systematic design of program analysis frameworks. In ACM Symposium on Principles of Programming Languages, pages 269–282, 1979. P. Cousot and R. Cousot. Comparing the galois connection and widening/narrowing approaches to abstract interpretation. In Programming Language Implementation and Logic Programming (PLILP), pages 269–295, 1992. P. Cousot and R. Cousot. Higher-order abstract interpretation (and application to comportment analysis generalizing strictness, termination, projection and per analysis of functional languages). In International Conference on Computer Languages (ICCL), pages 95–112, 1994. P. Cousot and N. Halbwachs. Automatic discovery of linear restraints among variables of a program. In ACM Symposium on Principles of Programming Languages, pages 84–96, 1978. Á. Darvas, R. Hähnle, and D. Sands. A theorem proving approach to analysis of secure information flow. In Security in Pervasive Computing, pages 193–209, 2005. D. E. R. Denning. Cryptography and Data Security. Addison-Wesley Longman Publishing Co., Inc., 1982. D. E. R. Denning and P. J. Denning. Certification of programs for secure information flow. Communications of ACM, 20(7):504–513, 1977. G. Doychev, D. Feld, B. Köpf, L. Mauborgne, and J. Reineke. Cacheaudit: A tool for the static analysis of cache side channels. In USENIX Security Symposium, pages 431–446, 2013. C. Dwork. A firm foundation for private data analysis. Communications of ACM, pages 86–95, 2011. B. Finkbeiner, M. N. Rabe, and C. Sánchez. Algorithms for model checking HyperLTL and HyperCTL ˆ*. In Computer Aided Verification, volume 9206 of LNCS, pages 30–48, 2015. R. Giacobazzi and I. Mastroeni. Abstract non-interference: parameterizing non-interference by abstract interpretation. In ACM Symposium on Principles of Programming Languages, pages 186–197, 2004. J. A. Goguen and J. Meseguer. Security policies and security models. In IEEE Symposium on Security and Privacy, pages 11–20, 1982. P. Granger. Improving the results of static analyses programs by local decreasing iteration. In Foundations of Software Technology and Theoretical Computer Science, volume 652, pages 68–79, 1992. M. Handjieva and S. Tzolovski. Refining dtatic analyses by trace-based partitioning using control flow. In International Static Analysis Symposium, 1998. D. Hedin, L. Bello, and A. Sabelfeld. Value-sensitive hybrid information flow control for a JavaScript-Like language. In IEEE Computer Security Foundations Symposium, pages 351–365, 2015. J. Heusser and P. Malacaria. Applied quantitative information flow and statistical databases. In Formal Aspects in Security and Trust, pages 96–110, 2009. S. Hunt. PERs generalize projections for strictness analysis (extended abstract). In Proceedings of the Third Annual Glasgow Workshop on Functional Programming, 1990. S. Hunt and D. Sands. Binding time analysis: A new PERspective. In Proceedings of the Symposium on Partial Evaluation and SemanticsBased Program Manipulation, PEPM’91, Yale University, New Haven, Connecticut, USA, June 17-19, 1991, pages 154–165, 1991. S. Hunt and D. Sands. On flow-sensitive security types. In ACM Symposium on Principles of Programming Languages, pages 79–90, 2006. S. Hunt and D. Sands. From exponential to polynomial-time security typing via principal types. In ACM Workshop on Programming Languages and Analysis for Security, pages 297–316, 2011. V. Klebanov. Precise quantitative information flow analysis - a symbolic approach. Theoretical Computer Science, 538:124–139, 2014. B. Köpf and A. Rybalchenko. Approximation and randomization for quantitative information-flow analysis. In IEEE Computer Security Foundations Symposium, pages 3–14, 2010. B. Köpf and A. Rybalchenko. Automation of quantitative information-flow analysis. In Formal Methods for Dynamical Systems - 13th International School on Formal Methods for the Design of Computer, Communication, and Software Systems, volume 7938 of LNCS, pages 1–28, 2013. M. Kovács, H. Seidl, and B. Finkbeiner. Relational abstract interpretation for the verification of 2-hypersafety properties. In ACM SIGSAC conference on Computer and Communications Security, pages 211–222, 2013. P. Mardziel, S. Magill, M. Hicks, and M. Srivatsa. Dynamic enforcement of knowledge-based security policies. In IEEE Computer Security Foundations Symposium, pages 114–128. IEEE, 2011. P. Mardziel, S. Magill, M. Hicks, and M. Srivatsa. Dynamic enforcement of knowledge-based security policies using probabilistic abstract interpretation. Journal of Computer Security, 21(4):463–532, 2013. I. Mastroeni. Abstract interpretation-based approaches to security - A survey on abstract non-interference and its challenging applications. In Semantics, Abstract Interpretation, and Reasoning about Programs: Essays Dedicated to David A. Schmidt on the Occasion of his Sixtieth Birthday, volume 129 of EPTCS, pages 41–65, 2013. I. Mastroeni and A. Banerjee. Modelling declassification policies using abstract domain completeness. Mathematical Structures in Computer Science, 21(06):1253–1299, 2011. J. McLean. A general theory of composition for trace sets closed under selective interleaving functions. In IEEE Symposium on Security and Privacy, pages 79–93, 1994. A. Miné. The octagon abstract domain. Higher-order and symbolic computation, 19(1):31–100, 2006a. A. Miné. Symbolic methods to enhance the precision of numerical abstract domains. In Verification, Model Checking, and Abstract Interpretation, pages 348–363, 2006b. C. Müller, M. Kovács, and H. Seidl. An analysis of universal information flow based on self-composition. In IEEE Computer Security Foundations Symposium, pages 380–393, 2015. F. Nielson, H. R. Nielson, and C. Hankin. Principles of Program Analysis. Springer, 1999. A. Rényi. On measures of entropy and information. In the Fourth Berkeley Symposium on Mathematical Statistics and Probability, 1961. X. Rival and L. Mauborgne. The trace partitioning abstract domain. ACM Transactions on Programming Languages and Systems, 29(5):26, 2007. J. Rushby. Security requirements specifications: How and what. In Symposium on Requirements Engineering for Information Security (SREIS), 2001. A. Sabelfeld and A. C. Myers. Language-based information-flow security. IEEE Journal on Selected Areas in Communications, 21(1):5–19, 2003. A. Sabelfeld and D. Sands. Declassification: Dimensions and principles. Journal of Computer Security, 17(5), 2009. D. A. Schmidt. Abstract interpretation from a topological perspective. In Static Analysis, 16th International Symposium, volume 5673 of LNCS, pages 293–308, 2009. D. A. Schmidt. Inverse-limit and topological aspects of abstract interpretation. Theoretical Computer Science, 430:23–42, 2012. D. Schoepe, M. Balliu, B. C. Pierce, and A. Sabelfeld. Explicit secrecy: A policy for taint tracking. In IEEE European Symposium on Security and Privacy, pages 15–30, 2016. C. E. Shannon. A mathematical theory of communication. The Bell System Technical Journal, 27:379–423, 1948. G. Smith. On the foundations of quantitative information flow. In International Conference on Foundations of Software Science and Computational Structures, pages 288–302, 2009. G. Smith. Quantifying information flow using min-entropy. In Quantitative Evaluation of Systems (QEST), 2011 Eighth International Conference on, pages 159–167. IEEE, 2011. M. Sousa and I. Dillig. Cartesian Hoare logic for verifying k-safety properties. In ACM Conference on Programming Language Design and Implementation, pages 57–69, 2016. T. Terauchi and A. Aiken. Secure information flow as a safety problem. In Static Analysis Symposium, pages 352–367, 2005. D. Volpano and G. Smith. Eliminating covert flows with minimum typings. In IEEE Computer Security Foundations Workshop, pages 156–168, 1997. D. Volpano and G. Smith. Verifying secrets and relative secrecy. In ACM Symposium on Principles of Programming Languages, pages 268–276, 2000. D. Volpano, C. Irvine, and G. Smith. A sound type system for secure flow analysis. Journal of Computer Security, 4(2-3):167–187, 1996. D. M. Volpano. Safety versus secrecy. In Static Analysis Symposium, pages 303–311, 1999. D. Wasserrab, D. Lohner, and G. Snelting. On PDG-based noninterference and its modular proof. In ACM Workshop on Programming Languages and Analysis for Security, pages 31–44, 2009. G. Winskel. The Formal Semantics of Programming Languages: an Introduction. Cambridge, 1993. H. Yasuoka and T. Terauchi. On bounding problems of quantitative information flow. Journal of Computer Security, 19(6):1029–1082, 2011. A. Zakinthinos and S. Lerner. A general theory of security properties. In IEEE Symposium on Security and Privacy, pages 94–102, 1997. M. Zanioli and A. Cortesi. Information leakage analysis by abstract interpretation. In SOFSEM 2011: Theory and Practice of Computer Science, pages 545–557, 2011. M. Zanotti. Security typings by abstract interpretation. In Static Analysis Symposium, volume 2477, pages 360–375, 2002. Appendix A. Symbols Val ∞ v ∈ Val V ∈ P(Val) V ∈ P(P(Val)) Trc t ∈ Trc T ∈ P(Trc) T ∈ P(P(Trc)) States σ ∈ States Σ ∈ P(States) S ∈ P(P(States)) States∗ , ∪n∈N Statesn a set of integers an infinite cardinal number an integer a set of values a set of sets of values the set of (relational) traces a trace a set of traces a set of sets of traces the set of states a state a set of states a set of sets of states the set of finite sequence of states VarP L l∈L Γ ∈ VarP → L the set of variables of a program P a multilevel security lattice a security level an initial typing context γ − −A C← −− α→ a Galois connection JcK ∈ Trc → Trc JeK ∈ Trc → Val JeKpre ∈ Trc → Val ⦃c⦄ ∈ P(Trc) → P(Trc) LcM ∈ P(P(Trc)) → P(P(Trc)) denotational semantics of commands value of e in final state value of e in initial state collecting semantics hypercollecting semantics l;x atomic dependence: “agreement up to security level l leads to agreement on x” A set of atomic dependency constraints atomic cardinality: “agreement up to security level l leads to an l-cardinality of n values for x” A valid set of atomic cardinality constraints D ∈ Dep l ; x#n C ∈ Card Appendix B. Background: Collecting Semantics, Galois Connections Lemma 1 Element-wise abstraction. Let elt ∈ C → A be a function between sets. Let αelt (C) , {elt(c) | c ∈ C} and γelt −− −− − (P(A), ⊆). γelt (A) , {c | elt(c) ∈ A}. Then (P(C), ⊆) ← −− → α− elt Proof. Let C ∈ P(C) and A ∈ P(A). αelt (C) ⊆ A ⇐⇒ {elt(c) | c ∈ C} ⊆ A ⇐⇒ ∀c ∈ C, elt(c) ∈ A ⇐⇒ C ⊆ {c | elt(c) ∈ A} ⇐⇒ C ⊆ γelt (A) Appendix C. Domains and Galois Connections for Hyperproperties Lemma 2 . Let C be a set. Define αhpp (C) , ∪C∈C C and γhpp (C) , P(C). These form a Galois connection: γhpp −− −−− − (P(C), ⊆) (P(P(C)), ⊆) ← −− → α −− hpp Proof. This is a special case of the supremus abstraction (Cousot 2002, p.52), that is defined in Lemma 3. Indeed, we can instantiate a supremus γhpp −− −−− − P(C), with αhpp (C) = ∪C∈C C abstraction by taking hpp , id (∈ P(C) → P(C)). We thus obtain a Galois connection P(P(C)) ← −− → α −− hpp and γhpp (C) = {C 0 ∈ P(C) | C 0 ⊆ C} (= P(C)). Notice here that the powerset of a set C, provided with set inclusion as a partial order, is a complete lattice as required by the supremus abstraction. Lemma 3 Supremus abstraction. Let elt ∈ C → A be a function from a set C, with codomain forming a complete lattice (A, v). Let αelt (C) , tc∈C elt(c) and γelt (a) , {c ∈ C | elt(c) v a}. Then γ elt −− − − − (A, v) (P(C), ⊆) ← −− → α− elt Proof. Notice that the assumption that the lattice (A, v, t) is complete guarantees that αelt (C) is well-defined: the set {elt(c) | c ∈ C} does have a supremum. Let C ∈ P(C) and a ∈ A. The proof goes by definitions. αelt (C) v a ⇐⇒ tc∈C elt(c) v a ⇐⇒ ∀c ∈ C, elt(c) v a ⇐⇒ C ⊆ {c ∈ C | elt(c) v a} ⇐⇒ C ⊆ γelt (a) Appendix D. Hypercollecting Semantics Before proving the main result of this section in Theorem 1, we will first prove Lemma 8. Both proofs of Lemma 8 and Theorem 1 are by structural induction. Most cases follow from definitions. The important cases are for while loops and the proof technique is a classical one when using a denotational semantics. E.g., in order to prove equality of two denotations characterised as a fixpoint, it suffices to introduce two sequences that converge towards the fixpoint characterisations and prove equality of these sequences. This ensures that their limits – the denotations characterised as a fixpoint – are equal. Let us now prove Lemma 8 – this lemma is used later in the proof case of while loops for Theorem 1. Lemma 8 . For all commands c, for all sets of traces T ∈ P(Trc), the standard collecting semantics (Section 2) can be expressed as the direct image of the denotational semantics : ⦃c⦄T = {JcKt ∈ Trc | t ∈ T } Proof. The proof proceeds by structural induction on commands. The most important case is the case of while loops. 1 – Case skip: ⦃skip⦄T = T = {JskipKt | t ∈ T } 2 – Case x := e: ⦃x := e⦄T = {Jx := eKt | t ∈ T } 3 – Case c1 ; c2 : ⦃c1 ; c2 ⦄T = ⦃c2 ⦄ ◦ ⦃c1 ⦄T = HBy induction on c1 I ⦃c2 ⦄({Jc1 Kt ∈ Trc | t ∈ T }) = HBy induction on c2 I {Jc2 K ◦ Jc1 Kt ∈ Trc | t ∈ T } = {Jc1 ; c2 Kt ∈ Trc | t ∈ T } 4 – Case if (b) then c1 else c2 : ⦃if (b) then c1 else c2 ⦄T = ⦃c1 ⦄ ◦ ⦃ grdb ⦄T ∪ ⦃c2 ⦄ ◦ ⦃ grd¬b ⦄T = HBy induction hypothesis on both c1 and c2 I {Jc1 Kt ∈ Trc | t ∈ ⦃ grdb ⦄T } ∪ {Jc2 Kt ∈ Trc | t ∈ ⦃ grd¬b ⦄T } = {Jif (b) then c1 else c2 Kt ∈ Trc | t ∈ T } 5 – Case while (b) do c: 5.1 – Let us first prove the following intermediate result: ∀T ∈ P(Trc),   b ◦ ⦃ grd ⦄X . λX. T ∪ ⦃c⦄ {Jwhile (b) do cKt ∈ Trc | t ∈ T } = ⦃ grd¬b ⦄ lfp⊆ ∅ Indeed, let the sequence (xTn )n≥0 be defined as: with F defined as xTn , {F (n) (⊥)(t) ∈ Trc | t ∈ T } ( t F(w)(t) , w ◦ JcKt if JbKt = 0 otherwise Notice that for all t ∈ T , the sequence (F (n) (⊥)(t))n≥0 converges and is equal to the evaluation of the while loop in the state t (i.e. Jwhile b do cKt = F (∞) (⊥)(t)), by definition of the denotational semantics of loops; thus, the sequence xTn converges to {Jwhile b do cKt ∈ Trc | t ∈ T }. Let also the sequences (ynT )n≥0 and (gnT )n≥0 be defined as: ynT , ⦃ grd¬b ⦄gnT T gn+1 , T ∪ ⦃c⦄ ◦ ⦃ grdb ⦄gnT g0T , ∅ ⊆ b b ◦ ◦ Notice that for all T ∈ P(Trc), the sequence gnT converges to lfp⊆ ∅ λX. T ∪ ⦃c⦄ ⦃ grd ⦄X (or written otherwise: lfpT λX. ⦃c⦄ ⦃ grd ⦄X). ⊆ T ¬b b ◦ This also means that the sequence yn converges to ⦃ grd ⦄( lfp∅ λX. T ∪ ⦃c⦄ ⦃ grd ⦄X ). Thus, it suffices to prove that: ∀T ∈ P(Trc), ∀n ∈ N, xTn = ynT . The proof proceeds by induction on n. - xT0 = ∅ = y0T - Let n ∈ N such that: ∀T ∈ P(Trc), xTn = ynT . Then: xTn+1 = {F (n+1) (⊥)(t) ∈ Trc | t ∈ T } = ⦃ grd¬b ⦄T ∪ {F (n) (⊥)(JcKt) ∈ Trc | σ ∈ ⦃ grdb ⦄T } = ⦃ grd¬b ⦄T ∪ {F (n) (⊥)(t) ∈ Trc | t ∈ ⦃c⦄ ◦ ⦃ grdb ⦄T } ◦⦃ grdb = ⦃ grd¬b ⦄T ∪ x⦃c⦄ n ⦄T = HBy induction hypothesisI ◦⦃ grdb ⦄T b ⦄T ⦃ grd¬b ⦄T ∪ yn⦃c⦄ = HBy definition of yn⦃c⦄ ¬b ◦⦃ grd ¬b ⦄T ∪ ⦃ grd   b ◦ = ⦃ grd ⦄ T ∪ gn⦃c⦄ ⦃ grd ⦄T [ = HBecause for all T , gnT = ⦃ grd I b ◦ ⦄gn⦃c⦄ ⦃ grd ⦄T ¬b 0≤k≤n−1 (⦃c⦄ ◦ ⦃ grdb ⦄)(k) (T ) I T ⦃ grd¬b ⦄gn+1 T = yn+1 5.2 – Let us now prove that : b ⊆ ◦ lfp⊆ ∅ λX. T ∪ ⦃c⦄ ⦃ grd ⦄X = lfp∅ λX.T ∪ ⦃if b then c else skip⦄X Indeed, let the sequence (fnT )n≥0 be defined as: f0T , ∅ T fn+1 , T ∪ ⦃if b then c else skip⦄fnT Therefore, by induction on n ∈ N, it holds that fn = gn : - f0T = g0T = ∅. - let n ∈ N, such that fnT = gnT . Then: T gn+1 = Σ ∪ ⦃c⦄ ◦ ⦃ grdb ⦄gnT T = HSince ⦃ grd¬b ⦄gnT ⊆ gnT ⊆ gn+1 I T ∪ ⦃c⦄ ◦ ⦃ grdb ⦄gnT ∪ ⦃ grd¬b ⦄gnT = T ∪ ⦃if b then c else skip⦄gnT = HBy induction hypothesisI T ∪ ⦃if b then c else skip⦄fnT T = fn+1 This concludes our induction on n. Thus, by passing to the limit of both sequences, we obtain the desired result. 5.3 – Finally, we can conclude:   ⦃while b do c⦄T =⦃ grd¬b ⦄ lfp⊆ T ⦃if b then c else skip⦄ = H by intermediate result 5.2 I   b ⦃ grd¬b ⦄ lfp⊆ T ⦃c⦄⦃ grd ⦄ = H by intermediate result 5.1I {⦃while (b) do c⦄t ∈ Trc | t ∈ T } We conclude this proof by structural induction, and Cases 1 to 5. Theorem 1 . For all c and all T ∈ P(Trc), ⦃c⦄T is in L c M{T }. Proof. We prove the theorem as a corollary of this more general result: ∀T ∈ P(P(Trc)), {⦃c⦄T | T ∈ T} ⊆ LcMT This proof proceeds by structural induction on commands. The most important case is the one for while loops; the other ones follow from definition. 1 – Case skip : 2 – Case x := e: LskipMT = {⦃skip⦄T | T ∈ T} ⊇ {⦃skip⦄T | T ∈ T} 3 – Case c1 ; c2 : Lx := eMT = {⦃x := e⦄T | T ∈ T} ⊇ {⦃x := e⦄T | T ∈ T} Lc1 ; c2 MT = Lc2 M ◦ Lc1 MT ⊇ Hby structural induction on c1 , and monotonicity of the hypercollecting semantics LcMI Lc2 M({⦃c1 ⦄T | T ∈ T}) ⊇ Hby structural induction on c2 I {⦃c2 ⦄T 0 | T 0 ∈ {⦃c1 ⦄T | T ∈ T}} = {⦃c2 ⦄ ◦ ⦃c1 ⦄T | T ∈ T} = {⦃c1 ; c2 ⦄T | T ∈ T} 4 – Case if (b) then c1 else c2 : Lif (b) then c1 else c2 MT = {⦃if (b) then c1 else c2 ⦄T | T ∈ T} ⊇ {⦃if (b) then c1 else c2 ⦄T | T ∈ T} 5 – Case while (b) do c: Let (XTn )n∈N be the sequence defined as n o XTn , {F (n) (⊥)(t) ∈ Trc | t ∈ T } | T ∈ T for n ≥ 1, XT0 = ∅ where: F(w)(t) , ( t w ◦ JcKt if JbKt = 0 otherwise Notice that the limit of the sequence xTn , {F (n) (⊥)(t) ∈ Trc | t ∈ T } is the ordinary collecting semantics ⦃while (b) do c⦄T of the while loop, as proved in Lemma 8. Thus, the sequence XTn converges to {⦃while (b) do c⦄T | T ∈ T}. Let also (YTn )n∈N and (GTn )n∈N be the sequences defined as GTn+1 , T ∪ Lif (b) then c else skipMGTn for n ≥ 0, GT0 , ∅ YTn , Lgrd¬b MGTn is the hypercollecting semantics of the while loop (Lwhile (b) do cMT). Notice that the limit of Thus, it suffices to prove that the sequences XTn and YTn verify the following result ∀T ∈ P(P(Trc)), ∀n ∈ N, XTn+1 ⊆ YTn+1 ; passing to the limit in this inequality leads to the required result ∀T ∈ P(P(Trc)), {⦃while (b) do c⦄T | T ∈ T} ⊆ Lwhile (b) do cMT. We prove the following more precise characterisation of the sequences XTn and YTn (this implies XTn+1 ⊆ YTn+1 ): YTn ∀n ∈ N, ∀T ∈ P(P(Trc)), YTn+1 = YTn ∪ XTn+1 The remaining of this proof proceeds by induction on n ∈ N. - case n = 0: YT1 = Lgrd¬b MGT1 = Lgrd¬b MT = {⦃ grd¬b ⦄T | T ∈ T} = {{F (1) (⊥)(t) ∈ Trc | t ∈ T } | T ∈ T} = Hsince YT0 = ∅ and by definition of XT1 I YT0 ∪ XT1 - Let n ∈ N such that YTn+1 = YTn ∪ XTn+1 . Then: YTn+2 = Lgrd¬b MGTn+2   = Lgrd¬b M T ∪ Lif (b) then c1 else skipMGTn+1 = Lgrd¬b MT ∪ Lgrd¬b M ◦ Lif (b) then c1 else skipMGTn+1 = H ∀n ∈ N, GTn+1 = ∪0≤k≤n Lif (b) then c1 else c2 M(k) TI   Lgrd¬b MT ∪ Lgrd¬b M ∪1≤k≤n+1 Lif (b) then c else skipM(k) T = ∪0≤k≤n+1 Lgrd¬b M ◦ Lif (b) then c else skipM(k) T   = ∪0≤k≤n Lgrd¬b M ◦ Lif (b) then c else skipM(k) T ∪ Lgrd¬b M ◦ Lif (b) then c else skipM(n+1) T   = Lgrd¬b M ◦ ∪0≤k≤n Lif (b) then c else skipM(k) T ∪ Lgrd¬b M ◦ Lif (b) then c else skipM(n+1) T = Lgrd¬b MGTn+1 ∪ Lgrd¬b M ◦ Lif (b) then c else skipM(n+1) T = YTn+1 ∪ Lgrd¬b M ◦ Lif (b) then c else skipM(n+1) T = YTn+1 ∪ {⦃ grd¬b ⦄ ◦ ⦃if (b) then c1 else skip⦄(n+1) T | T ∈ T} = Hthe set ⦃ grd¬b ⦄ ◦ ⦃if (b) then c1 else c2 ⦄(n+1) T is the set of traces exiting the loop body after n+1 or less iterations:I H it is equal to {F (n+2) (⊥)(t) | t ∈ T } by definition of FI n o YTn+1 ∪ {F (n+2) (⊥)(t) | t ∈ T } | T ∈ T = YTn+1 ∪ XTn+2 This concludes our induction on n. We conclude this proof by structural induction, and Cases 1 to 5. Appendix E. Dependences γdeptr −− −−−− − Dep. Lemma 9 . (αdeptr , γdeptr ) yields a Galois connection: P(P(Trc)) ← − − → α −−− deptr Proof. The lattice Dep is finite, therefore complete. Thus, this is a Galois connection since it is an instance of the supremus abstraction presented in Lemma 3. γagree −−−−− {tt, ff}. The same reasoning applies for P(P(Val)) ← −− α−−−→ agree The proofs of of both Lemma 4 and Theorem 2 are deferred to Appendix G: as we explain after Lemmas 6 and 7, we derive the dependence abstract semantics as an approximation of the cardinality semantics. Appendix F. Cardinality Abstraction γ crdtr −− −−−− Card. Lemma 10 . (αcrdtr , γcrdtr ) yields a Galois connection: P(P(Trc)) ← −− α −−→ crdtr Proof. The lattice Card is complete, since all subsets of Card have an infimum and a supremum wrt. partial order v] , notably because the closed interval [0, ∞] is complete wrt. partial order ≤. Thus, this is an instance of the supremus abstraction Lemma 3. l Lemma 5 . OC LeM] is sound: ∀e, ∀l, αcrdval ◦ l ˙ OC Ol LeM ◦ γcrdtr ≤ LeM] . Proof. The derivation proof is by structural induction on expressions. In each case we start from the left side and derive the definition on the right side. The interesting case is for binary arithmetic operations. 1 – Case : integer literal n Let l ∈ L, and C ∈ Card. Ol LnM ◦ γcrdtr (C )   = αcrdval ◦ Ol LnM {T | crdtr(T ) v] C } αcrdval ◦ = αcrdval ∪T ∈γcrdtr (C ) {⦃n⦄R | R ⊆ T and R |=Γ l}
= αcrdval ({⦃n⦄R | R ⊆ T, R |=Γ l and T ∈ γcrdtr (C )}) ≤ HNB: precision loss for simplicity of presentation, when C is bottomI αcrdval ({{n}}) = max crdval(V ) V ∈{{n}} =1 l , OC LnM] C l LnM] C is being defined. Here we use , to indicate that OC 2 – Case : variable id Let l ∈ L, and C ∈ Card. Ol LidM ◦ γcrdtr (C )   = αcrdval ∪T ∈γcrdtr (C ) Ol ⦃id⦄T αcrdval ◦ = Hαcrdval preserves joinsI   max αcrdval Ol ⦃id⦄T T ∈γcrdtr (C ) = n H where id ; l#n ∈ αcrdtr ≤ Hαcrdtr ◦ γcrdtr (C )I γcrdtr is reductive : αcrdtr n H where id ; l#n ∈ C I 3 – Case : e1 ⊕ e2 Let l ∈ L, and C ∈ Card. ◦ ◦ γcrdtr (C ) v] C I l , OC LidM] C αcrdval ◦ Ol Le1 ⊕ e2 M ◦ γcrdtr (C ) = αcrdval ({⦃e1 ⊕ e2 ⦄R | R ⊆ T, R |=Γ l and T ∈ γcrdtr (C )}) ≤ αcrdval ({⦃e1 ⦄R | R ⊆ T, R |=Γ l and T ∈ γcrdtr (C )}) × αcrdval ({⦃e2 ⦄R | R ⊆ T, R |=Γ l and T ∈ γcrdtr (C )}) = αcrdval ◦ Ol Le1 M ◦ γcrdtr (C ) × αcrdval ◦ ≤ HBy induction hypothesisI Ol Le2 M ◦ γcrdtr (C ) l l OC Le1 M] C × OC Le2 M] C l , OC Le1 ⊕ e2 M] C 4 – Case : e1 cmp e2 This derivation is similar to case e1 ⊕ e2 , with the difference that booleans evaluate to at most 2 different values, 1 or 0. Ol Le1 cmp e2 M ◦ γcrdtr (C )   l l ≤ min 2, OC Le1 M] C × OC Le2 M] C αcrdval ◦ l , OC Le1 cmp e2 M] C 5 – Case : conclusion We conclude by structural induction on expressions, and cases 1 to 4. Theorem 4 . The cardinality abstract semantics is sound: αcrdtr ◦ ] LcM ◦ γcrdtr v̇ LcM] . Proof. The derivation proof is by structural induction on commands. The interesting case is for conditionals. 1 – Case : skip Let C ∈ Card. αcrdtr LskipM ◦ γcrdtr (C ) ◦ = αcrdtr ◦ ] v Hαcrdtr C γcrdtr (C ) ◦ γcrdtr is reductive: αcrdtr ◦ γcrdtr (C ) v] C I , LskipM] C 2 – Case : c1 ; c2 αcrdtr ◦ Lc1 ; c2 M ◦ γcrdtr (C ) = αcrdtr ] v αcrdtr ◦ Lc2 M ◦ Lc1 M ◦ γcrdtr (C ) αcrdtr is extensive, Lc2 M and αcrdtr are monotoneI Lc2 M ◦ γcrdtr ◦ αcrdtr ◦ Lc1 M ◦ γcrdtr (C ) Hγcrdtr ◦ ◦ v] HBy induction hypothesisI Lc2 M] ◦ Lc1 M] C , Lc1 ; c2 M] C 3 – Case : id := e 3.1 – We first proceed towards an intermediate derivation: αcrdtr = ◦ Lid := eM ◦ γcrdtr (C ) G] crdtr(T ) T ∈Lid:=eM◦γcrdtr (C )   G] = [  T ∈Lid:=eM◦γcrdtr (C ) {l ; x#n | n = αcrdval (O ⦃x⦄T ) } l l∈L,x∈VarP  = [ l∈L,x∈VarP T ∈Lid:=eM◦γcrdtr (C )  l ; x#n | n = {l ; x#n | n = αcrdval (O ⦃x⦄T ) } l  l∈L,x∈VarP =  G] [ max T ∈Lid:=eM◦γcrdtr (C ) αcrdval (Ol ⦃x⦄T ) We now consider two cases: variables that are not modified by the assignment, and variable that are. 3.2 – Case x 6= id:  Notice that ∀l ∈ L, ∀x ∈ VarP , such that x 6= id, ∀T ∈ γcrdtr (C ): Ol ⦃x⦄T = Ol ⦃x⦄ ⦃id := e⦄T
Thus: max αcrdval (Ol ⦃x⦄T ) T ∈Lid:=eM◦γcrdtr (C ) = max T ∈γcrdtr (C ) αcrdval (Ol ⦃x⦄T ) = Hαcrdval preserves joinsI   [ l αcrdval  O ⦃x⦄T  T ∈γcrdtr (C ) = HBy definition of Ol LxMI   αcrdval Ol LxM (γcrdtr (C )) = αcrdval ◦ Ol LxM ◦ γcrdtr (C ) l ≤ HBy soundness of OC LxM] C , Lemma 5I l OC LxM] C = n where l ; x#n ∈ C 3.3 – Case x is id : ∀l ∈ L, we have : max αcrdval (Ol ⦃id⦄T ) T ∈Lid:=eM◦γcrdtr (C ) = max T ∈γcrdtr (C ) = αcrdval ◦ αcrdval (Ol ⦃e⦄T ) Ol LeM ◦ γcrdtr (C ) l ≤ HBy soundness of OC LxM] C , Lemma 5I l OC LeM] C 3.4 – Final derivation: αcrdtr ◦ Lid := eM ◦ γcrdtr (C ) = HRecall the intermediate derivation in Case 3.1I  [ l ; x#n | n = max T ∈Lid:=eM◦γcrdtr (C ) l∈L,x∈VarP αcrdval (Ol ⦃x⦄T )  v] HBy Cases 3.2 and 3.3I   n o [ [
l ]  l ; x#n ∈ C } ∪ {l ; id#OC LeM C  x∈VarP \{id} l∈L HNB: this set of constraints remains valid, owing to exclusion of id on the leftI n o l = l ; x#n ∈ C | x 6= id} ∪ {l ; id#OC LeM] C | l ∈ L , Lid := eM] C 4 – Case if b then c1 else c0 : 4.1 – Intermediate derivation: αcrdtr = ◦ Lif b then c1 else c0 M ◦ γcrdtr (C ) G] crdtr(T ) T ∈Lif b c1 else c0 M◦γcrdtr (C ) =  G] [  T ∈Lif b c1 else c0 M◦γcrdtr (C ) l∈L,x∈VarP n  o l ; x#αcrdval (O ⦃x⦄T )  l  = G] [  l∈L,x∈VarP = n T ∈Lif b c1 else c0 M◦γcrdtr (C )  l ; x# [  o l ; x#αcrdval (O ⦃x⦄T )  l l∈L,x∈VarP max T ∈Lif b c1 else c0 M◦γcrdtr (C ) αcrdval (Ol ⦃x⦄T )  l 4.2 – Case OC LbM] C = 1 : l Let l ∈ L, and assume OC LbM] C = 1. Let x ∈ VarP . 0 ◦ ∀T ∈ Lif b c1 else c0 M γcrdtr (C ), exists T ∈ γcrdtr (C ) such that T 0 = ⦃if b c1 else c0 ⦄T . (Since LcM is not just the lifting of ⦃c⦄ to a set of sets (semantics of loops is not), in general if T 0 ∈ LcMT, we only have the existence of T ∈ T such that T 0 ⊆ ⦃c⦄T . Here, we also rely on the fact that γcrdtr (C ) is a subset-closed. This is merely a convenient shortcut to avoid lengthy details; it should be possible to use only the fact that T 0 ⊆ ⦃c⦄T to perform the same derivation. ) Let T 0 ∈ Lif b c1 else c0 M ◦ γcrdtr (C ), and T ∈ γcrdtr (C ) such that T 0 = ⦃if b c1 else c0 ⦄T . l ˙ OC Since αcrdval ◦ Ol LbM ◦ γcrdtr (C ) ≤ LbM] C (= 1), ∀R ⊆ T such that R |=Γ l, the traces r ∈ R all evaluate b to 1 or (exclusively) to 0; i.e. the sets R ⊆ T such that R |=Γ l are partitioned into the sets evaluating b to 1, and those evaluating b to 0. Therefore, ∀R0 ⊆ T 0 such that R0 |=Γ l, exists R ∈ T and j ∈ {0, 1} such that ⦃b⦄R = {j} and ⦃cj ⦄R = R0 . Thus, αcrdval (Ol ⦃x⦄T 0 ) = αcrdval ({⦃x⦄R0 | R0 ⊆ T 0 and R0 |=Γ l})   [ = αcrdval  {⦃x⦄(⦃if b c1 else c0 ⦄R)} R⊆T and R|=Γ l   = αcrdval  [ [ j∈{0,1} R⊆T and R|=Γ l and ⦃b⦄R={j}   = max αcrdval  j∈{0,1} {⦃x⦄(⦃cj ⦄R)}  [ R⊆T and R|=Γ l and ⦃b⦄R={j} {⦃x⦄(⦃cj ⦄R)} ≤ Hαcrdval is monotoneI   max αcrdval ◦ Ol LxM ◦ Lcj M ◦ γcrdtr (C ) j∈{0,1} ≤ Hαcrdval ◦ Ol LxM is monotone, γcrdtr ◦ αcrdtr extensive I   max αcrdval ◦ Ol LxM ◦ γcrdtr ◦ αcrdtr Lcj M ◦ γcrdtr (C ) j∈{0,1} ≤ HBy induction hypothesisI  max αcrdval ◦ Ol LxM ◦ γcrdtr j∈{0,1} ◦ Lcj M] C  ≤ HBy soundness of abstract variety, Lemma 5I   l max OC LxM] ◦ Lcj M] C j∈{0,1}   = max nj where l ; x#nj ∈ Lcj M] C j∈{0,1} l 4.3 – Case OC LbM] C > 1, x ∈ / Mod(if b c1 else c0 ) : l Let l ∈ L, and assume OC LbM] C > 1. Let x ∈ VarP . Let T 0 ∈ Lif b c1 else c0 M ◦ γcrdtr (C ), and T ∈ γcrdtr (C ) such that T 0 = ⦃if b c1 else c0 ⦄T . Notice first that if x ∈ / Mod(if b c1 else c0 ), then: αcrdval (Ol ⦃x⦄T 0 ) = αcrdval (Ol ⦃x⦄T ) ≤ Hαcrdval is monotoneI αcrdval l 4.4 – Case OC LbM] C > 1, x ∈ Mod(if b c1 else c0 ) : l Let l ∈ L, and assume OC LbM] C > 1. Let x ∈ VarP . ◦ Ol LxM ◦ γcrdtr (C ) l ≤ OC LxM] C = n s.t l ; x#n ∈ C Let T 0 ∈ Lif b c1 else c0 M ◦ γcrdtr (C ), and T ∈ γcrdtr (C ) such that T 0 = ⦃if b c1 else c0 ⦄T . αcrdval (Ol ⦃x⦄T 0 ) = αcrdval (Ol ⦃x⦄ ◦ ⦃if b c1 else c0 ⦄T )   ≤ αcrdval (Ol ⦃x⦄ ⦃c1 ⦄ ◦ ⦃ grdb ⦄T ∪ ⦃c0 ⦄ ◦ ⦃ grd¬b ⦄T     ≤ αcrdval Ol ⦃x⦄ ◦ ⦃c1 ⦄ ◦ ⦃ grdb ⦄T + αcrdval Ol ⦃x⦄ ◦ ⦃c0 ⦄ ◦ ⦃ grd¬b ⦄T ≤ HBy monotonicity, T ∈ γcrdtr (C ) and Theorem 1I αcrdval ◦ l ≤ OC LxM] Ol LxM ◦ Lc1 M ◦ Lgrd¬b M ◦ γcrdtr (C ) + αcrdval Lc1 M] ◦ ◦ l Lgrd¬b M] C + OC LxM] ◦ Lc0 M] ◦ ◦ Ol LxM ◦ Lc0 M ◦ Lgrd¬b M ◦ γcrdtr (C ) Lgrd¬b M] C ≤ HAs a first approximation, we simply use Lgrdb M] C v] C . We refine this in Section 8I l OC LxM] ◦ l Lc1 M] C + OC LxM] ◦ Lc2 M] C = n1 + n2 s.t l ; x#n1 ∈ Lc1 M] C and l ; x#n2 ∈ Lc2 M] C 4.4 – Final derivation: αcrdtr ◦ Lif b then c1 else c0 M ◦ γcrdtr (C ) = HBy the intermediate derivation in case 4.1I   [ l ; x# max αcrdval (Ol ⦃x⦄T ) T ∈Lif b c1 else c0 M◦γcrdtr (C ) l∈L,x∈VarP ( ≤ π l (Lc1 M] C ) t] π l (Lc2 M] C ) π l (Lc1 M] C ) t] add(if b c1 else c0 ,πl (C )) π l (Lc1 M] C ) [ l∈L l LbM] = 1 if OC otherwise with π l (C ) , {l ; x#n ∈ C | x ∈ VarP , n ∈ [0, ∞]} and C1 t] add(com,C0 ) C2 , [ {l ; x#n | n , n1 + n2 s.t l ; x#nj ∈ Cj , j = 1, 2} x∈Mod(com) [ {l ; x#n ∈ C0 } x∈VarP \ Mod(com) 5 – Case while b do c: αcrdtr ◦ Lwhile b do cM] = αcrdtr ◦ γcrdtr (C )   ¬b ⊆ ◦ Lgrd M lfp γcrdtr (C ) Lif b then c else skipM v] Hαcrdtr ,Lgrd¬b M are monotone, γcrdtr ◦ αcrdtr is extensiveI   αcrdtr ◦ Lgrd¬b M ◦ γcrdtr ◦ αcrdtr lfp⊆ γcrdtr (C ) Lif b then c else skipM v] HBy assuming Lgrdb M] is soundI   Lgrd¬b M] ◦ αcrdtr lfp⊆ γcrdtr (C ) Lif b then c else skipM v] HBy the fixpoint transfer theoremI Lgrd¬b M] ] ◦ ] lfpv C Lif b then c1 else c2 M ] v] Hprecision loss for simplicity as a first approximation, Lgrd¬b M] v̇ idI ] ] lfpv C Lif b then c1 else c2 M , Lwhile b do cM] C 6 – Case : conclusion We conclude by structural induction on commands and cases 1 to 5. Appendix G. Appendix G.1 Dependencies reloaded Soundness proof for dependences semantics As noted in the text, we can derive the dependency analysis by calculation from its specification. The derivation looks similar to the one in Appendix F for the cardinality abstraction. So here we choose a different way of proving soundness for dependency analysis. We formulate it as an abstraction of the cardinality abstraction. This is another illustration of the benefit gained from working with hyperproperties entirely within the framework of abstract interpretation. This proof of soundness also implies that the cardinality abstraction is at least as precise as the type system of Hunt and Sands (Hunt and Sands 2006) and the logic of Amtoft and Banerjee (Amtoft and Banerjee 2004), as a corollary of Theorem 3. Lemma 6 . (αagree , γagree ) is the composition of two Galois connections (αcrdval , γcrdval ) and (αlqone , γlqone ) : γcrdval γlqone crdval lqone −− −−−− − ([0, ∞] , ≤) ← −−−−− ({tt, ff}, ⇐=) (P(P(Val)), ⊆) ← −− → −− α −−− α−−−→ with: ( if n ≤ 1 , and otherwise. tt ff ( 1 γlqone (bv) , ∞ αlqone (n) , if bv = tt otherwise. Proof. Notice that: agree(V ) , (∀v1 , v2 ∈ V, v1 = v2 ) = (crdval(V ) ≤ 1) Also, αagree (V) , ∧V ∈V agree(V ) = ∧V ∈V (crdval(V ) ≤ 1)   = max crdval(V ) ≤ 1 V ∈V = αcrdval (V) ≤ 1 = αlqone ◦ αcrdval (V) ( where αlqone (n) , tt if n ≤ 1 ff otherwise And, γagree (bv) , {V ∈ P(Val) | agree(V ) ⇐= bv} = {V ∈ P(Val) | crdval(V ) ≤ γlqone (bv)} where ( 1 if bv = tt γlqone (bv) , ∞ otherwise = γcrdval ◦ γlqone (bv) γlqone −−−−− {tt, ff}: Notice that [0, ∞] ← −− α−−−→ lqone ∀n ∈ [0, ∞] , ∀ bv ∈ {tt, ff}, Thus, we obtain αagree = αlqone ◦ αlqone (n) ⇐= bv iff. n ≤ γlqone (bv) αcrdval , as well as γagree = γcrdval ◦ γlqone : γcrdval γlqone crdval lqone −−−−− ({tt, ff}, ⇐=) −− −−−− − ([0, ∞] , ≤) ← (P(P(Val)), ⊆) ← −− → −− α −−− α−−−→ Lemma 7 . (αdeptr , γdeptr ) is the composition of two Galois connections (αcrdtr , γcrdtr ) and (αlqonecc , γlqonecc ) : γcrdtr γlqonecc crdtr lqonecc −− −−−− (Card, v] ) ← −− −−−−− − (Dep, v\ ) (P(P(Trc)), ⊆) ← −− −− → α −−→ α −−−− with: αlqonecc (C ) , {l ;Sx | l ; x#n ∈ C and αlqone (n)} γlqonecc (D) , {l ; x#n | n = γlqone (l ; x ∈ D)} l∈L,x∈VarP Proof. First, αdeptr (T) = G\ deptr(T ) T ∈T = G\ [ T ∈T l∈L,x∈VarP {l ; x | αagree (Ol ⦃x⦄T )} = HBy the decomposition in Case 1I G\ [ {l ; x | αlqone ◦ αcrdval (Ol ⦃x⦄T )} T ∈T l∈L,x∈VarP   = G\ αlqonecc  T ∈T = G\ [ {l ; x#n | n = αcrdval (O ⦃x⦄T )} l l∈L,x∈VarP Hwith αlqonecc (C ) , {l ; x | l ; x#n ∈ C and αlqone (n)}I αlqonecc ◦ crdtr(T ) T ∈T = Hαlqonecc preserves unionsI ! G] αlqonecc crdtr(T ) T ∈T = αlqonecc ◦ αcrdtr (T) Also, γdeptr (D) = {T | deptr(T ) v\ D} = {T | αlqonecc ◦ = {T | αlqonecc ◦ crdtr(T ) v\ D} crdtr(T ) ⊇ D} = {T | ∀l ; x ∈ D, l ; x ∈ αlqonecc  = T | ∀l ; x ∈ D, ◦ crdtr(T )}   l ; x ∈ αlqonecc ∪l∈L,x∈VarP {l ; x#n | n = αcrdval (Ol ⦃x⦄T ) } = {T | ∀l ; x ∈ D, αlqone (αcrdval (Ol ⦃x⦄T ))} = {T | ∀l ; x ∈ D, αcrdval (Ol ⦃x⦄T ) ≤ 1} = γcrdtr ◦ γlqonecc (D) Hγlqonecc (D) , Therefore, we have αdeptr = αlqonecc ◦ [ l∈L,x∈VarP {l ; x#n | n = γlqone (l ; x ∈ D)}I αcrdtr , and γdeptr = γcrdtr ◦ γlqonecc , with: γcrdtr γlqonecc crdtr lqonecc −− −−−− (Card, v] ) ← −− −−−−− − (Dep, v\ ) (P(P(Trc)), ⊆) ← −− −− → α −−→ α −−−− l Lemma 4 . OD LeM\ is sound: ∀e, ∀l, ∀D, αagree ◦ l Ol LeM ◦ γdeptr (D) ⇐= OD LeM\ D . Proof. l 1 – Derivation of Agreements OD LeM\ up to l as an abstraction of cardinalities up to security level l. αagree ◦ Ol LnM ◦ γdeptr (D) = αlqone Henceforth, we will derive 1.1 – Case : n Let l ∈ L, D ∈ Dep. l OD LeM\ ⇐= ◦ αcrdval ◦ Ol LnM ◦ γdeptr l αlqone ◦ OC LeM] ◦ ◦ γlqonecc (D) γlqonecc (D) l as an abstraction of cardinalities OC LeM] . This derivation goes by structural induction on expressions. αlqone ◦ l OC LnM] ◦ γlqonecc (D) = αlqone (1) = tt l H, OD LnM\ I 1.2 – Case : id Let l ∈ L, D ∈ Dep. αlqone ◦ l OC LidM] ◦ γlqonecc (D) = αlqone (n) where l ; id#n ∈ γlqonecc (D) = (l ; id ∈ D) 1.3 – Case : e1 ⊕ e2 Let l ∈ L, D ∈ Dep. l H, OD LidM\ I l OC Le1 ⊕ e2 M] ◦ γlqonecc (D)  l l = αlqone (OC Le1 M] ◦ γlqonecc (D)) × (OC Le2 M] αlqone ◦ ⇐= αlqone = ◦ l OC Le1 M] l OD Le1 M\ D ∧ γlqonecc (D) ∧ αlqone ◦ l OD Le2 M\ D H, 1.4 – Case : e1 cmp e2 This case is similar to case 1.3. 1.5 – Case: conclusion We conclude by structural induction on expressions. l OD Le1 ◦ ◦ γlqonecc (D)) l OC Le2 M] ◦  γlqonecc (D) ⊕ e2 M DI \ Theorem 2 . The dependence semantics is sound: αdeptr Proof. Recall that we have αdeptr = αlqonecc ◦ ◦ \ LcM ◦ γdeptr v̇ LcM\ . αcrdtr , and γdeptr = γcrdtr ◦ γlqonecc , with: γcrdtr γlqonecc crdtr lqonecc −− −−−−− − Dep −− −−−− Card ← P(P(Trc)) ← −− → −− α −−−− α −−→ Since, αdeptr ◦ LcM ◦ γdeptr (D) = αlqonecc v\ αlqonecc ◦ αcrdtr ◦ LcM] ◦ ◦ LcM ◦ γcrdtr ◦ γlqonecc (D) γlqonecc (D) We will continue the derivation of dependences abstract semantics LcM\ as an abstraction of LcM] . We make explicit 2 derivations, for assignments and conditionals. The other cases are similar to the derivation of the cardinalities abstract semantics. 1 – Case : id := e Lid := eM] ◦ γlqonecc (D)  l = αlqonecc {l ; x#n ∈ γlqonecc (D) | x 6= id} ∪ {l ; id#n | n , OC LeM] αlqonecc ◦ ◦  γlqonecc (D), l ∈ L} l = {l ; x ∈ D | x 6= id} ∪ {l ; id | OD LeM\ } 2 – Case if b then c1 else c2 : αlqonecc = ⇐= ◦ Lif b then c1 else c2 M] γlqonecc (D) ◦ l if OC LbM] ◦ γlqonecc (D) = 1      otherwise let D1 = Lc1 M D in let D2 = Lc2 M\ D in let W ( = Mod(if b c1 else c0 ) in l S π l (D1 ) t\ π l (D2 ) if OD LbM\ D l l∈L π (D) \ {l ; x | x ∈ W } otherwise We conclude by structural induction on commands Appendix G.2 ◦ let C1 = Lc1 M γlqonecc (D) in let C2 = Lc2 M] ◦ γlqonecc (D) in let W =  Mod(if  b c1 else c0 ) in l ] l  π (C1 ) t π (C2 ) π l (C ) S   1 αlqonecc  t] add(W,πl (C )) l∈L     π l (C2 ) ] Precision proof Lemma 11 . For all l, l0 ∈ L, for all T ∈ P(Trc): \ . 0 l v l0 =⇒ Ol ⦃e⦄T ⊆ Ol ⦃e⦄T Proof. Assume l v l0 . Then, for all R ⊆ T , R |=Γ l0 =⇒ R |=Γ l Thus, Therefore, it holds that: {R | R ⊆ T and R |=Γ l0 } ⊆ {R | R ⊆ T and R |=Γ l} 0 Ol ⦃e⦄T ⊆ Ol ⦃e⦄T Corollary 2 . For all l, l0 ∈ L, for all T ∈ P(P(Trc)): 0 l v l0 =⇒ Ol LeMT ⊆ Ol LeMT Proof. This is a direct result from Lemma 11 and definition of Ol LeM. Corollary 3 . For all l, l0 ∈ L, for all id, for all T ∈ P(P(Trc)): l v l0 =⇒ l ; id ∈ αdeptr (T) =⇒ l0 ; id ∈ αdeptr (T)
Proof. 0 Let us assume l v l0 . By Corollary 2, we have Ol LidMT ⊆ Ol LidMT. 0 And by monotonicity of αagree , we have αagree (Ol LidMT) ⇐= αagree (Ol LidMT). l Thus, if l ; id ∈ αdeptr (T), then αagree (O LidMT) = tt and also 0 αagree (Ol LidMT) = tt, thus l0 ; D ∈ αdeptr (T). Corollary 4 . For all l, l0 ∈ L, for all id, for all D ∈ Dep: l v l0 =⇒ γdeptr (D ∪ {l ; id}) = γdeptr (D ∪ {l ; id, l0 ; id}) Proof. 1 – Note that D ∪ {l ; id, l0 ; id} ⊇ D ∪ {l ; id}, thus D ∪ {l ; id, l0 ; id} v\ D ∪ {l ; id} Therefore, by monotony of γdeptr : γdeptr (D ∪ {l ; id, l0 ; id}) ⊆ γdeptr (D ∪ {l ; id}) 2 – Also, let T ∈ γdeptr (D ∪ {l ; id}). We have deptr(T ) v\ D ∪ {l ; id} by definition of γdeptr . Thus, l ; id ∈ αdeptr (T ) and also l0 ; id ∈ αdeptr (T ) by Corollary 3. This also means that deptr(T ) v\ D ∪ {l ; id, l0 ; id}. Finally, T ∈ γdeptr (D ∪ {l ; id, l0 ; id}) and γdeptr (D ∪ {l ; id, l0 ; id}) ⊇ γdeptr (D ∪ {l ; id}) This concludes our proof by Case 1 and 2. Henceforth, we will assume that all D ∈ Dep are well formed, meaning that ∀l, l0 ∈ Dep, l ; id ∈ D =⇒ l0 ; id ∈ D. We conjecture that this can be proven for the dependence analysis we have derived: given a well formed initial set of dependence constraints, the analysis always yields a well formed set of dependence constraints. For simplicity, we will use Corollary 4 to argue that we can still augment any set of dependence constraints to ensure it is well formed by adding the appropriate atomic constraints. An alternative approach would reduce the set of dependence constraints, and change slightly the abstract semantics in order to leverage Corollary 4 and guarantee the same precision, but we refrain from doing so for simplicity. We consider the constructive version of Hunt and Sands’ flow sensitive type system, proposed in (Hunt and Sands 2011). Lemma 12 . For all e, D ∈ Dep, ∆ ∈ VarP → L, l ∈ L such that ∆ ` e : l, it holds that: l αhs (D) v̇ ∆ =⇒ OD LeM\ D = tt Proof. The proof proceeds by structural induction on expressions. 1 – Case n: l By definition of OD LnM\ , we have: l ∀D, ∀l ∈ L, OD LnM\ D = tt 2 – Case id: By definition of the type system, we have ∆(id) = l. Thus: αhs (D) v̇ ∆ =⇒ u{l0 | l0 ; id ∈ D} v l =⇒ Hsince D is assumed well-formedI l ; id ∈ D l =⇒ OD LidM\ D = tt 3 – Case e1 ⊕ e2 : By definition of the type system, there is l1 , l2 such that ∆ ` e1 : l1 and ∆ ` e2 : l2 , with l1 t l2 = l. Thus, by induction on e1 and e2 , and assuming αhs (D) v̇ ∆, we have: l1 l2 OD Le1 M\ D = tt ∧ OD Le2 M\ D = tt Thus, since D is well formed and l1 v l and l2 v l, it holds that: Therefore, also: l l OD Le1 M\ D = tt ∧ OD Le2 M\ D = tt l OD Le1 ⊕ e2 M\ D = tt 4 – Case e1 cmp e2 : This case is similar to Case 3. 5 – We conclude by structural induction and Cases 1 to 4. Let us denote by ⊥L ∈ L the bottom element of the lattice L. Theorem 3 . For all c, D0 , D ∈ Dep, ∆0 , ∆ ∈ VarP → L, where ⊥L ` ∆0 {c}∆, and D = LcM\ D0 , it holds that: αhs (D0 ) v̇ ∆0 =⇒ αhs (D) v̇ ∆ . Proof. The proof goes by structural induction on commands. The conditional case explicitly assumes that the modified variables analysis is precise enough, to enable the simulation of the program counter. This can be achieved by collecting variable names in a while language. 1 – Case skip : this case stems from the premice. 2 – Case id := e : Assume αhs (D0 ) v̇ ∆0 . 2.1 – Case : x 6= id Then, for all x 6= id, ∆(x) = ∆0 (x). Also, αhs (D)(x) = αhs (D0 )(x) v ∆0 (x) = ∆(x). 2.2 – Case : x = id Otherwise, ∆ = ∆0 [id 7→ l], where ∆0 ` e : l. l By Lemma 12, since αhs (D0 ) v̇ ∆0 , we have OD LeM\ D0 = tt. Thus, l ; id ∈ D and : αhs (D)(id) v ∆(id) 2.3 – Finally, by Cases 2.2 and 2.3, we have : αhs (D) v̇ ∆ 3 – Case c1 ; c2 : This case proceeds by induction on both c1 and c2 by remarking that the type system types both command c1 and c2 in ⊥L . 4 – Case if (b) then c1 else c2 : Assume αhs (D0 ) v̇ ∆0 . Let lb , ∆1 , ∆2 such that: lb ` ∆0 {c1 }∆1 and lb ` ∆0 {c2 }∆2 , with ∆ = ∆1 ṫ ∆2 . Also, let ∆01 and ∆02 such that: ⊥L ` ∆0 {c1 }∆01 and ⊥L ` ∆0 {c2 }∆02 , with: ∆1 = ∆01 [id 7→ ∆01 (id) t lb , ∀id ∈ Mod(c1 )], ∆2 = ∆02 [id 7→ ∆02 (id) t lb , ∀id ∈ Mod(c2 )]. Intuitively, the program counter pc can be simulated by a modified variables analysis that is precise enough. For a while language, this can be achieved simply by collecting variable names. Let D1 , D2 ∈ Dep such that: Lc1 M\ D0 = D1 and Lc2 M\ D0 = D2 . Then, assuming W = M od(if (b) then c1 else c2 ), we have:  l \ l l \  π (D 1 ) t π (D2 ) if OD LbM D S  l D= {l ; x ∈ π (D) | l∈L   x∈ / W} otherwise 4.1 – By induction on c1 , we have αhs (D1 ) v̇ ∆01 . 4.2 – By induction on c2 , we have αhs (D2 ) v̇ ∆02 . 4.3 – Assume x 6∈ W , and prove αhs (D)(x) v ∆(x). Since x 6∈ W , we have ∆(x) = ∆0 (x). Therefore, αhs (D0 ) v̇ ∆0 implies ∆(x) ; x ∈ D0 . Thus, since x 6∈ W , we have ∆(x) ; x ∈ D1 , and ∆(x) ; x ∈ D2 (atomic constraints related to variables not explicitly written in c1 are not discarded from D0 , and likewise for those that are not explicitly written in c2 ). Thus, ∆(x) ; x ∈ D, meaning that αhs (D)(x) v ∆(x). 4.4 – Assume x ∈ W amd prove αhs (D)(x) v ∆(x): We have lb v ∆(x) since x is explicitly written in one of the branches at least. Also, by 4.1 and 4.2, we have αhs (D1 )(x) v ∆01 (x), and αhs (D2 )(x) v ∆02 (x). Meaning that αhs (D1 t\ D2 )(x) = αhs (D1 )(x) t αhs (D2 )(x) v ∆01 (x) t ∆02 (x) v ∆(x) Notice that ∆01 and ∆02 are well formed. Thus, exists lx such that lb v lx , such that lx ; x ∈ D1 t\ D2 , and lx v ∆(x). l And since ∀l ∈ L, such that lb v l, we also have OD LbM\ D0 = tt by using Lemma 12 and D0 is well formed. l l \ l Thus, ∀l ∈ L, such that lb v l, π (D) = π (D1 ) t π (D2 ) = π l (D1 t\ D2 ), i.e. lx ; x ∈ D Thus, αhs (D)(x) v ∆(x). 5 – Case : while (b) then c Assume αhs (D0 ) v̇ ∆0 . The output type environment ∆ is defined by: ∆ = lfp λ∆v .let ∆0 s.t. ⊥L t ∆v (a) ` ∆v {c}∆0 in ∆0 ṫ ∆0 Or written differently, ∆ is given by: ∆ = lfp λ∆v .let ∆0 s.t. ⊥L ` ∆v {if (b) then c else skip}∆0 in ∆0 ṫ ∆0 Let (∆n ) be the sequence defined as ∆n+1 = let ∆0 s.t. ⊥L ` ∆n {if (b) then c else skip}∆0 in ∆0 ṫ ∆0 Also, let (Dn ) be the sequence defined as Dn+1 = D0 t\ Lif (b) then c else skipM\ Dn Then, we prove by induction on n that αhs (Dn ) v ∆n . 5.1 – Case n = 0. This case holds by assumption αhs (D0 ) v̇ ∆0 . 5.2 – Case: Assume αhs (Dn ) v ∆n , and prove αhs (Dn+1 ) v ∆n+1 . Let ∆0 such that ⊥L ` ∆n {if (b) then c else skip}∆0 By assumption, we have αhs (Dn ) v ∆n . Thus, by using the same proof in Case 4, we have αhs (Lif (b) then c else skipM\ Dn ) v̇ ∆0 Therefore,   αhs (Lif (b) then c else skipM\ Dn ) ṫ αhs (D0 ) v̇ (∆0 ṫ ∆0 ) Therefore, αhs (Dn+1 ) v ∆n+1 . which proves that both least fixpoints are equal. 6 – Finally, we conclude by Cases 1–5, and structural induction on commands. 

Bounded Model Checking for Probabilistic Programs? Nils Jansen2 , Christian Dehnert1 , Benjamin Lucien Kaminski1 , Joost-Pieter Katoen1 , and Lukas Westhofen1 1 arXiv:1605.04477v2 [cs.PL] 26 Jul 2016 2 RWTH Aachen University, Germany University of Texas at Austin, USA Abstract. In this paper we investigate the applicability of standard model checking approaches to verifying properties in probabilistic programming. As the operational model for a standard probabilistic program is a potentially infinite parametric Markov decision process, no direct adaption of existing techniques is possible. Therefore, we propose an on– the–fly approach where the operational model is successively created and verified via a step–wise execution of the program. This approach enables to take key features of many probabilistic programs into account: nondeterminism and conditioning. We discuss the restrictions and demonstrate the scalability on several benchmarks. 1 Introduction Probabilistic programs are imperative programs, written in languages like C, Scala, Prolog, or ML, with two added constructs: (1) the ability to draw values at random from probability distributions, and (2) the ability to condition values of variables in a program through observations. In the past years, such programming languages became very popular due to their wide applicability for several different research areas [1]: Probabilistic programming is at the heart of machine learning for describing distribution functions; Bayesian inference is pivotal in their analysis. They are central in security for describing cryptographic constructions (such as randomized encryption) and security experiments. In addition, probabilistic programs are an active research topic in quantitative information flow. Moreover, quantum programs are inherently probabilistic due to the random outcomes of quantum measurements. All in all, the simple and intuitive syntax of probabilistic programs makes these different research areas accessible to a broad audience. However, although these programs typically consist of a few lines of code, they are often hard to understand and analyze; bugs, for instance non–termination of a program, can easily occur. It seems of utmost importance to be able to automatically prove properties like “Is the probability for termination of the program ? This work has been partly funded by the awards AFRL # FA9453-15-1-0317, ARO # W911NF-15-1-0592 and ONR # N00014-15-IP-00052 and is supported by the Excellence Initiative of the German federal and state government. at least 90%” or “Is the expected value of a certain program variable at least 5 after successful termination?”. Approaches based on the simulation of a program to show properties or infer probabilities have been made in the past [2,3]. However, to the best of our knowledge there is no work which exploits wellestablished model checking algorithms for probabilistic systems such as Markov decision processes (MDP) or Markov chains (MCs), as already argued to be an interesting avenue for the future in [1]. As the operational semantics for a probabilistic program can be expressed as a (possible infinite) MDP [4], it seems worthwhile to investigate the opportunities there. However, probabilistic model checkers like PRISM [5], iscasMc [6], or MRMC [7] offer efficient methods only for finite models. We make use of the simple fact that for a finite unrolling of a program the corresponding operational MDP is also finite. Starting from a profound understanding of the (intricate) probabilistic program semantics—including features such as observations, unbounded (and hence possibly diverging) loops, and nondeterminism—we show that with each unrolling of the program both conditional reachability probabilities and conditional expected values of program variables increase monotonically. This gives rise to a bounded model-checking approach for verifying probabilistic programs. This enables for a user to write a program and automatically verify it against a desired property without further knowledge of the programs semantics. We extend this methodology to the even more complicated case of parametric probabilistic programs, where probabilities are given by functions over parameters. At each iteration of the bounded model checking procedure, parameter valuations violating certain properties are guaranteed to induce violation at each further iteration. We demonstrate the applicability of our approach using five well-known benchmarks from the literature. Using efficient model building and verification methods, our prototype is able to prove properties where either the state space of the operational model is infinite or consists of millions of states. Related Work. Besides the tools employing probabilistic model checking as listed above, one should mention the approach in [8], where finite abstractions of the operational semantics of a program were verified. However, this was defined for programs without parametric probabilities or observe statements. In [9], verification on partial operational semantics is theoretically discussed for termination probabilities. The paper is organized as follows: In Section 2, we introduce the probabilistic models we use, the probabilistic programming language, and the structured operational semantics (SOS) rules to construct an operational (parametric) MDP. Section 3 first introduces formal concepts needed for the finite unrollings of the program, then shows how expectations and probabilities grow monotonically, and finally explains how this is utilized for bounded model checking. In Section 4, an extensive description of used benchmarks, properties and experiments is given before the paper concludes with Section 5. 2 Preliminaries 2.1 Distributions and Polynomials A probability distribution P over a finite or countably infinite set X is a function µ : X → [0, 1] ⊆ R with x∈X µ(x) = 1. The set of all distributions on X is denoted by Distr (X). Let V be a finite set of parameters over R. A valuation for V is a function u : V → R. Let Q[V ] denote the set of multivariate polynomials with rational coefficients and QV the set of rational functions (fractions of polynomials) over V . For g ∈ Q[V ] or g ∈ QV , let g[u] denote the evaluation of g at u. We write g = 0 if g can be reduced to 0, and g 6= 0 otherwise. 2.2 Probabilistic Models First, we introduce parametric probabilistic models which can be seen as transition systems where the transitions are labelled with polynomials in Q[V ]. Definition 1 (pMDP and pMC). A parametric Markov decision process (pMDP) is a tuple M = (S, sI , Act, P) with a countable set S of states, an initial state sI ∈ S, a finite set Act of actions, and a transition function P : S × Act × S → Q[V ] satisfying for all s ∈ S : Act(s) 6= ∅, where V is a finite set of parameters over R and Act(s) = {α ∈ Act | ∃s0 ∈ S. P(s, α, s0 ) 6= 0}. If for all s ∈ S it holds that |Act(s)| = 1, M is called a parametric discrete-time Markov chain (pMC), denoted by D. At each state, an action is chosen nondeterministically, then the successor states are determined probabilistically as defined by the transition function. Act(s) is the set of enabled actions at state s. As Act(s) is non-empty for all s ∈ S, there are no deadlock states. For pMCs there is only one single action per state and we write the transition probability function as P : S × S → Q[V ], omitting that action. Rewards are defined using a reward function rew : S → R which assigns rewards to states of the model. Intuitively, the reward rew(s) is earned upon leaving the state s. Schedulers. The nondeterministic choices of actions in pMDPs can be resolved using schedulers 3 . In our setting it suffices to consider memoryless deterministic schedulers [10]. For more general definitions we refer to [11]. Definition 2. (Scheduler) A scheduler for pMDP M = (S, sI , Act, P) is a function S : S → Act with S(s) ∈ Act(s) for all s ∈ S. Let Sched M denote the set of all schedulers for M. Applying a scheduler to a pMDP yields an induced parametric Markov chain, as all nondeterminism is resolved, i.e., the transition probabilities are obtained w.r.t. the choice of actions. Definition 3. (Induced pMC) Given a pMDP M = (S, sI , Act, P), the pMC induced by S ∈ Sched M is given by MS = (S, sI , Act, P S ), where P S (s, s0 ) = P(s, S(s), s0 ), 3 for all s, s0 ∈ S . Also referred to as adversaries, strategies, or policies. Valuations. Applying a valuation u to a pMDP M, denoted M[u], replaces each polynomial g in M by g[u]. We call M[u] the instantiation of M at u. A valuation u is well-defined for M if the replacement yields probability distributions at all states; the resulting model M[u] is a Markov decision process (MDP) or, in absence of nondeterminism, a Markov chain (MC). Properties. For our purpose we consider conditional reachability properties and conditional expected reward properties in MCs. For more detailed definitions we refer to [11, Ch. 10]. Given an MC D with state space S and initial state sI , let PrD (¬♦U ) denote the probability not to reach a set of undesired states U from the initial state sI within D. Furthermore, let PrD (♦T | ¬♦U ) denote the conditional probability to reach a set of target states T ⊆ S from the initial state sI within D, given that no state in the set U is reached. We use the standard probability measure on infinite paths through an MC. For threshold λ ∈ [0, 1] ⊆ R, the reachability property, asserting that a target state is to be reached with conditional probability at most λ, is denoted ϕ = P≤λ (♦T | ¬♦U ). The property is satisfied by D, written D |= ϕ, iff PrD (♦T | ¬♦U ) ≤ λ. This is analogous for comparisons like <, >, and ≥. The reward of a path through an MC D until T is the sum of the rewards of the states visited along on the path before reaching T . The expected reward of a finite path is given by its probability times its reward. Given PrD (♦T ) = 1, the conditional expected reward of reaching T ⊆ S, given that no state in set U ⊆ S is reached, denoted ERD (♦T | ¬♦U ), is the expected reward of all paths accumulated until hitting T while not visiting a state in U in between divided by the probability of not reaching a state in U (i.e., divided by PrD (¬♦U )). An expected reward property is given by ψ = E≤κ (♦T | ¬♦U ) with threshold κ ∈ R≥0 . The property is satisfied by D, written D |= ψ, iff ERD (♦T | ¬♦U ) ≤ κ. Again, this is analogous for comparisons like <, >, and ≥. For details about conditional probabilities and expected rewards see [12]. Reachability probabilities and expected rewards for MDPs are defined on induced MCs for specific schedulers. We take here the conservative view that a property for an MDP has to hold for all possible schedulers. Parameter Synthesis. For pMCs, one is interested in synthesizing well-defined valuations that induce satisfaction or violation of the given specifications [13]. In detail, for a pMC D, a rational function g ∈ QV is computed which— when instantiated by a well-defined valuation u for D—evaluates to the actual reachability probability or expected reward for D, i.e., g[u] = PrD[u] (♦T ) or g[u] = ERD[u] (♦T ). For pMDPs, schedulers inducing maximal or minimal probability or expected reward have to be considered [14]. 2.3 Conditional Probabilistic Guarded Command Language We first present a programming language which is an extension of Dijkstra’s guarded command language [15] with a binary probabilistic choice operator, yielding the probabilistic guarded command language (pGCL) [16]. In [17], pGCL was endowed with observe statements, giving rise to conditioning. The syntax of this conditional probabilistic guarded command language (cpGCL) is given by P ::= skip | abort | x := E | P; P | if G then P else P | {P} [g] {P} | {P}  {P} | while (G) {P} | observe (G) Here, x belongs to the set of program variables V; E is an arithmetical expression over V; G is a Boolean expression over arithmetical expressions over V. The probability is given by a polynomial g ∈ Q[V ]. Most of the cpGCL instructions are self–explanatory; we elaborate only on the following: For cpGCL-programs P and Q, {P } [g] {Q} is a probabilistic choice where P is executed with probability g and Q with probability 1−g; analogously, {P }  {Q} is a nondeterministic choice between P and Q; abort is syntactic sugar for the diverging program while (true) {skip}. The statement observe (G) for the Boolean expression G blocks all program executions violating G and induces a rescaling of probability of the remaining execution traces so that they sum up to one. For a cpGCLprogram P , the set of program states is given by S = {σ | σ : V → Q}, i.e., the set of all variable valuations. We assume all variables to be assigned zero prior to execution or at the start of the program. This initial variable valuation σI ∈ S with ∀x ∈ V. σI (x) = 0 is called the initial state of the program. Example 1. Consider the following cpGCL-program with variables x and c: 1 2 3 4 while ( c = 0) { { x := x + 1 } [0.5] { c := 1 } }; observe " x is odd " While c is 0, the loop body is iterated: With probability 1/2 either x is incremented by one or c is set to one. After leaving the loop, the event that the valuation of x is odd is observed, which means that all program executions where x is even are blocked. Properties of interest for this program would, e.g., concern the termination probability, or the expected value of x after termination. 4 2.4 Operational Semantics for Probabilistic Programs We now introduce an operational semantics for cpGCL-programs which is given by an MDP as in Definition 1. The structure of such an operational MDP is schematically depicted below. h i hP, σI i ↓ ↓ ↓ ↓↓ ↓ diverge hsink i Squiggly arrows indicate reaching certain states via possibly multiple paths and states; the clouds indicate that there might be several states of the particular kind. hP, σI i marks the initial state of the program P . In general the states of the operational MDP are of the form hP 0 , σ 0 i where P 0 is the program that is left to be executed and σ 0 is the current variable valuation. All runs of the program (paths through the MDP) are either terminating and eventually end up in the hsink i state, or are diverging (thus they never reach hsink i). Diverging runs occur due to non–terminating computations. A terminating run has either terminated successfully, i.e., it passes a ↓–state, or it has terminated due to a violation of an observation, i.e., it passes the h i–state. Sets of runs that eventually reach h i, or hsink i, or diverge are pairwise disjoint. The ↓–labelled states are the only ones with positive reward, which is due to the fact that we want to capture probabilities of events (respectively expected values of random variables) occurring at successful termination of the program. The random variables of interest are E = {f | f : S → R≥0 }. Such random variables are referred to as post–expectations [16]. Formally, we have: Definition 4 (Operational Semantics of Programs). The operational semantics of a cpGCL program P with respect to a post–expectation f ∈ E is the MDP Mf JP K = (S, hP, σI i, Act, P) together with a reward function rew, where  – S = hQ, σi, h↓, σi Q is a cpGCL program, σ ∈ S ∪ {h i, hsink i} is the countable set of states, – hP, σI i ∈ S is the initial state, – Act = {left, right, none} is the set of actions, and – P is the smallest relation defined by the SOS rules given in Figure 1. The reward function is rew(s) = f (σ) if s = h↓, σi, and rew(s) = 0, otherwise. A state of the form h↓, σi indicates successful termination, i.e., no commands are left to be executed. These terminal states and the h i–state go to the hsink i state. skip without context terminates successfully. abort self–loops, i.e., diverges. x := E alters the variable valuation according to the assignment then terminates successfully. For the concatenation, h↓; Q, σi indicates successful termination of the first program, so the execution continues with hQ, σi. If for P ; Q the execution of P leads to h i, P ; Q does so, too. Otherwise, for hP, σi−→µ, µ is lifted such that Q is concatenated to the support of µ. For more details on the operational semantics we refer to [4]. If for the conditional choice σ |= G holds, P is executed, otherwise Q. The case for while is similar. For the probabilistic choice, a distribution ν is created according to probability p. For {P }  {Q}, we call P the left choice and Q the right choice for actions left, right ∈ Act. For the observe statement, if σ |= G then observe acts like skip. Otherwise, the execution leads directly to h i indicating a violation of the observe statement. Example 2. Reconsider Example 1, where we set for readability P1 = {x := x + 1} [0.5] {c := 1}, P2 = observe(“x is odd”), P3 = {x := x + 1}, and (terminal) h↓, σi −→ hsink i (undesired) (skip) (abort) hskip, σi −→ h↓, σi (assign) h i −→ hsink i habort, σi −→ habort, σi hx := E, σi −→ h↓, σ[x ← JEKσ ]i σ 6|= G hobserve G, σi −→ h i σ |= G (observe1) hobserve G, σi −→ h↓, σi (observe2) (concatenate1) (concatenate2) (concatenate3) (if1) h↓; Q, σi −→ hQ, σi hP, σi −→ µ 0 0 0 0 0 , where ∀P . ν(hP ; Q, σ i) := µ(hP , σ i) hP ; Q, σi −→ ν σ |= G hite (G) {P } {Q}, σi −→ hP, σi (while1) (prob) (if2) σ |= G hwhile (G) {P }, σi −→ hP ; while (G) {P }, σi h{P } [p] {Q}, σi −→ ν (nondet1) hP, σi −→ h i hP ; Q, σi −→ h i σ 6|= G hite (G) {P } {Q}, σi −→ hQ, σi (while2) σ 6|= G hwhile (G) {P }, σi −→ h↓, σi , where ν(hP, σi) := p, ν(hQ, σi) := 1 − p (nondet2) left h{P }  {Q}, σi −−−→ hP, σi right h{P }  {Q}, σi −−−−→ hQ, σi Fig. 1. SOS rules for constructing the operational MDP of a cpGCL program. We use s −→ t to indicate P(s, none, t) = 1, s −→ µ for µ ∈ Distr (S) to indicate left right ∀t ∈ S : P(s, none, t) = µ(t), s −−−→ t to indicate P(s, left, t) = 1, and s −−−−→ t to indicate P(s, right, t) = 1. P4 = {c := 1}. A part of the operational MDP Mf JP K for an arbitrary initial variable valuation σI and post–expectation x is depicted in Figure 2.4 Note that this MDP is an MC, as P contains no nondeterministic choices. The MDP has been unrolled until the second loop iteration, i.e., at state hP, σI [x/2]i, the unrolling could be continued. The only terminating state is h↓, σI [x/1, c/1]i. As our post-expectation is the value of variable x, we assign this value to terminating states, i.e., reward 1 at state h↓, σI [x/1, c/1]i, where x has been assigned 1. At state hP, σI [c/1]i, the loop condition is violated as is the subsequent observation because of x being assigned an even number. 4 3 Bounded Model Checking for Probabilistic Programs In this section we describe our approach to model checking probabilistic programs. The key idea is that satisfaction or violation of certain properties for a program can be shown by means of a finite unrolling of the program. Therefore, we introduce the notion of a partial operational semantics of a program, which we exploit to apply standard model checking to prove or disprove properties. First, we state the correspondence between the satisfaction of a property for a cpGCL-program P and for its operational semantics, the MDP Mf JP K. Intu4 We have tacitly overloaded the variable name x to an expectation here for readability. More formally, by the “expectation x” we actually mean the expectation λσ. σ(x). hP, σI i 1 2 hP3 ; P, σI i hP1 ; P, σI i 1 2 h↓; P, σI [x/1]i hP3 ; P, σI [x/1]i 1 2 hP, σI [x/1]i hP4 ; P, σI i h↓; P, σI [c/1]i 1 2 hP4 ; P, σI [x/1]i hP, σI [c/1]i h↓; P, σI [x/2]i h↓; P, σI [x/1, c/1]i h↓; P2 , σI [c/1]i hP, σI [x/2]i .. . hP, σI [x/1, c/1]i hP2 , σI [c/1]i h↓; P2 , σI [x/1, c/1]i h i hP2 , σI [x/1, c/1]i 1 h↓, σI [x/1, c/1]i hsink i Fig. 2. Partially unrolled operational semantics for program P itively, a program satisfies a property if and only if the property is satisfied on the operational semantics of the program. Definition 5 (Satisfaction of Properties). Given a cpGCL program P and a (conditional) reachability or expected reward property ϕ. We define P |= ϕ iff Mf JP K |= ϕ . This correspondence on the level of a denotational semantics for cpGCL has been discussed extensively in [17]. Note that there only schedulers which minimize expected rewards were considered. Here, we also need maximal schedulers as we are considering both upper and lower bounds on expected rewards and probabilities. Note that satisfaction of properties is solely based on the operational semantics and induced maximal or minimal probabilities or expected rewards. We now introduce the notion of a partial operational MDP for a cpGCL– program P , which is a finite approximation of the full operational MDP of P . Intuitively, this amounts to the successive application of SOS rules given in Figure 1, while not all possible rules have been applied yet. Definition 6 (Partial Operational Semantics). A partial operational semantics for a cpGCL–program P is a sub-MDP Mf JP K0 = (S 0 , hP, σI i, Act, P 0 ) of the operational semantics for P (denoted Mf JP K0 ⊆ Mf JP K) with S 0 ⊆
S. Let  0 Sexp = S \ hQ, σi ∈ S 0 Q 6= ↓, ∃ s ∈ S \ S 0 ∃ α ∈ Act : P hQ, σi, α, s > 0 be the set of expandable states. Then the transition probability function P 0 is for s, s0 ∈ S 0 and α ∈ Act given by ( 0 0 P (s, α, s ) = if s = s0 for s, s0 ∈ Sexp , otherwise . 1, P(s, α, s0 ), Intuitively, the set of non–terminating expandable states describes the states where there are still SOS rules applicable. Using this definition, the only transitions leaving expandable states are self-loops, enabling to have a well-defined probability measure on partial operational semantics. We will use this for our method, which is based on the fact that both (conditional) reachability probabilities and expected rewards for certain properties will always monotonically increase for further unrollings of a program and the respective partial operational semantics. This is discussed in what follows. 3.1 Growing Expectations As mentioned before, we are interested in the probability of termination or the expected values of expectations (i.e. random variables ranging over program states) after successful termination of the program. This is measured on the operational MDP by the set of paths reaching hsink i from the initial state conditioned on not reaching h i [17]. In detail, we have to compute the conditional expected value of post–expectation f after successful termination of program P , given that no observation was violated along the computation. For nondeterministic programs, we have to compute this value either under a minimizing or maximizing scheduler (depending on the given property). We focus our presentation on expected rewards and minimizing schedulers, but all concepts are analogous for the other cases. For Mf JP K we have inf S∈Sched Mf JP K ERM f JP KS (♦hsink i | ¬♦h i) . f Recall that Mf JP KS is the induced MC under scheduler S ∈ Sched M JP K as in Definition 3. Recall also that for ¬♦h i all paths not eventually reaching h i either diverge (collecting reward 0) or pass by a ↓–state and reach hsink i. More importantly, all paths that do eventually reach h i also collect reward 0. Thus: inf S∈Sched Mf JP K ERM f JP KS f = inf S∈Sched = Mf JP K JP KS (♦hsink i ∩ ¬♦h i) Mf JP KS Mf JP K inf S∈Sched ERM (♦hsink i | ¬♦h i) Pr Mf JP KS ER (♦hsink i) Mf JP KS Pr (¬♦ ) (¬♦ ) . Finally, observe that the probability of not reaching h i is one minus the probability of reaching h i, which gives us: f = ERM inf S∈Sched Mf JP K JP KS (♦hsink i) Mf JP KS 1 − Pr . (†) (♦ ) Regarding the quotient minimization we assume “ 00 < 0” as we see 00 —being undefined—to be less favorable than 0. For programs without nondeterminism this view agrees with a weakest–precondition–style semantics for probabilistic programs with conditioning [17]. f S It was shown in [18] that all strict lower bounds for ERM JP K (♦hsink i) are in principle computably enumerable in a monotonically non–decreasing fashion. One way to do so, is to allow for the program to be executed for an increasing number of k steps, and collect the expected rewards of all execution traces that have lead to termination within k computation steps. This corresponds naturally to constructing a partial operational semantics Mf JP K0 ⊆ Mf JP K as in Definition 6 and computing minimal expected rewards on Mf JP K0 . Analogously, it is of course also possible to monotonically enumerate all S f strict lower bounds of PrM JP K (♦ ), since—again—we need to just collect the probability mass of all traces that have led to h i within k computation steps. Since probabilities are quantities bounded between 0 and 1, a lower bound for S S f f PrM JP K (♦ ) is an upper bound for 1 − PrM JP K (♦ ). f S Put together, a lower bound for ERM JP K (♦hsink i) and a lower bound for f S PrM JP K (♦ ) yields a lower bound for (†). We are thus able to enumerate all f S lower bounds of ERM JP K (♦hsink i | ¬♦h i) by inspection of a finite sub–MDP of Mf JP K. Formally, we have: Theorem 1. For a cpGCL program P , post–expectation f , and a partial operational MDP Mf JP K0 ⊆ Mf JP K it holds that inf S∈Sched Mf JP K0 ≤ ERM f JP K0S inf S∈Sched 3.2 (♦hsink i | ¬♦h i) Mf JP K ERM f JP KS (♦hsink i | ¬♦h i) . Model Checking Using Theorem 1, we transfer satisfaction or violation of certain properties from a partial operational semantics Mf JP K0 ⊆ Mf JP K to the full semantics of the program. For an upper bounded conditional expected reward property ϕ = E≤κ (♦T | ¬♦U ) where T, U ∈ S we exploit that Mf JP K0 6|= ϕ =⇒ P 6|= ϕ . (1) That means, if we can prove the violation of ϕ on the MDP induced by a finite unrolling of the program, it will hold for all further unrollings, too. This is because all rewards and probabilities are positive and thus further unrolling can only increase the accumulated reward and/or probability mass. Dually, for a lower bounded conditional expected reward property ψ = E≥λ (♦T | ♦U ) we use the following property: Mf JP K0 |= ψ =⇒ P |= ϕ . (2) The preconditions of Implication (1) and Implication (2) can be checked by probabilistic model checkers like PRISM [5]; this is analogous for conditional reachability properties. Let us illustrate this by means of an example. Example 3. As mentioned in Example 1, we are interested in the probability of termination. As outlined in Section 2.4, this probability can be measured by Pr(♦hsink i | ¬♦h i) = Pr(♦hsink i ∧ ¬♦h i) . Pr(♦h i) We want this probability to be at least 1/2, i.e., ϕ = P≥0.5 (♦hsink i | ¬♦h i). Since for further unrollings of our partially unrolled MDP this probability never decreases, the property can already be verified on the partial MDP Mf JP K0 by f PrM JP K0 (♦hsink i | ¬♦h i) = 1/4 1/2 = 1 , 2 where Mf JP K0 is the sub-MDP from Figure 2. This finite sub-MDP Mf JP K0 is therefore a witness of Mf JP K |= ϕ. 4 Algorithmically, this technique relies on suitable heuristics regarding the size of the considered partial MDPs. Basically, in each step k states are expanded and the corresponding MDP is model checked, until either the property can be shown to be satisfied or violated, or no more states are expandable. In addition, heuristics based on shortest path searching algorithms can be employed to favor expandable states that so far induce high probabilities. Note that this method is a semi-algorithm when the model checking problems stated in Implications (1) and (2) are considering strict bounds, i.e. < κ and > κ. It is then guaranteed that the given bounds are finally exceed. Consider now the case where we want to show satisfaction of ϕ = E≤κ (♦T | ¬♦U ), i.e., Mf JP K0 |= ϕ ⇒ P |= ϕ. As the conditional expected reward will monotonically increase as long as the partial MDP is expandable, the implication is only true if there are no more expandable states, i.e., the model is fully expanded. This is analogous for the violation of upper bounded properties. Note that many practical examples actually induce finite operational MDPs which enables to build the full model and perform model checking. It remains to discuss how this approach can be utilized for parameter synthesis as explained in Section 2.2. For a partial operational pMDP Mf JP K0 and a property ϕ = E≤κ (♦T | ¬♦U ) we use tools like PROPhESY [13] to determine for which parameter valuations ϕ is violated. For each valuation u with Mf JP K0 [u] 6|= ϕ it holds that Mf JP K[u] 6|= ϕ; each parameter valuation violating a property on a partial pMDP also violates it on the fully expanded MDP. 4 Evaluation Experimental Setup. We implemented and evaluated the bounded model checking method in C++. For the model checking functionality, we use the stochastic model checker Storm, developed at RWTH Aachen University, and PROPhESY [19] for parameter synthesis. We consider five different, well-known benchmark programs, three of which are based on models from the PRISM benchmark suite [5] and others taken from other literature (see Appendix A for some examples). We give the running times of our prototype on several instances of these models. Since there is — to the best of our knowledge — no other tool that can analyze cpGCL programs in a purely automated fashion, we cannot meaningfully compare these figures to other tools. As our technique is restricted to establishing that lower bounds on reachability probabilities and the expectations of program variables, respectively, exceed a threshold λ, we need to fix λ for each experiment. For all our experiments, we chose λ to be 90% of the actual value for the corresponding query and choose to expand 106 states of the partial operational semantics of a program between each model checking run. We ran the experiments on an HP BL685C G7 machine with 48 cores clocked with 2.0GHz each and 192GB of RAM while each experiment only runs in a single thread with a time–out of one hour. We ran the following benchmarks5 : Crowds Protocol [21]. This protocol aims at anonymizing the sender of R messages by routing them probabilistically through a crowd of N hosts. Some of these hosts, however, are corrupt and try to determine the real sender by observing the host that most recently forwarded a message. For this model, we are interested in (a) the probability that the real sender is observed more than R/10 times, and (b) the expected number of times that the real sender is observed. We also consider a variant (crowds-obs) of the model in which an observe statement ensures that after all messages have been delivered, hosts different from the real sender have been observed at least R/4 times. Unlike the model from the PRISM website, our model abstracts from the concrete identity of hosts different from the sender, since they are irrelevant for properties of interest. Herman Protocol. In this protocol [22], N hosts form a token-passing ring and try to steer the system into a stable state. We consider the probability that the system eventually reaches such a state in two variants of this model where the initial state is either chosen probabilistically or nondeterministically. Robot. The robot case-study is loosely based on a similar model from the PRISM benchmark suite. It models a robot that navigates through a bounded area of an unbounded grid. Doing so, the robot can be blocked by a janitor that is moving probabilistically across the whole grid. The property of interest is the probability that the robot will eventually reach its final destination. 5 All input programs and log files of the experiments can be downloaded at moves.rwth-aachen.de/wp-content/uploads/conference material/pgcl atva16.tar.gz Table 1. Benchmark results for probability queries. program instance #states #trans. full? λ result actual time (100,60) 877370 1104290 yes 0.29 0.33 0.33 109 crowds (100,80) 106 1258755 no 0.30 0.33 0.33 131 (100,100) 2 · 106 2518395 no 0.30 0.33 0.33 354 (100,60) 878405 1105325 yes 0.23 0.26 0.26 126 crowds-obs (100,80) 106 1258718 no 0.23 0.25 0.24 170 (100,100) 3 · 106 3778192 no 0.23 0.26 0.26 890 (17) 106 1136612 no 0.9 0.99 1 91 herman (21) 106 1222530 no 0.9 0.99 1 142 (13) 1005945 1112188 yes 0.9 1 1 551 herman-nd (17) − − no 0.9 0 1 TO robot 181595 234320 yes 0.9 1 1 24 predator 106 1234854 no 0.9 0.98 1 116 (5) 106 1589528 no 0.75 0.83 0.83 11 coupon (7) 2 · 106 3635966 no 0.67 0.72 0.74 440 (10) − − no 0.57 0 0.63 TO (5) 106 1750932 no 0.85 0.99 0.99 11 coupon-obs (7) 106 1901206 no 0.88 0.91 0.98 15 (10) − − no 0.85 0 0.95 TO (5) 106 1356463 no 3.4e-3 3.8e-3 3.8e-3 9 coupon-classic (7) 106 1428286 no 5.5e-4 6.1e-4 6.1e-4 9 (10) − − no 3.3e-5 0 3.6e-5 TO Predator. This model is due to Lotka and Volterra [23, p. 127]. A predator and a prey population evolve with mutual dependency on each other’s numbers. Following some basic biology principles, both populations undergo periodic fluctuations. We are interested in (a) the probability of one of the species going extinct, and (b) the expected size of the prey population after one species has gone extinct. Coupon Collector. This is a famous example6 from textbooks on randomized algorithms [24]. A collector’s goal is to collect all of N distinct coupons. In every round, the collector draws three new coupons chosen uniformly at random out of the N coupons. We consider (a) the probability that the collector possesses all coupons after N rounds, and (b) the expected number of rounds the collector needs until he has all the coupons as properties of interest. Furthermore, we consider two slight variants: in the first one (coupon-obs), an observe statement ensures that the three drawn coupons are all different and in the second one (coupon-classic), the collector may only draw one coupon in each round. Table 1 shows the results for the probability queries. For each model instance, we give the number of explored states and transitions and whether or not the model was fully expanded. Note that the state number is a multiple of 106 in 6 https://en.wikipedia.org/wiki/Coupon collector%27s problem Table 2. Benchmark results for expectation queries. program instance #states #trans. full? result actual time (100,60) 877370 1104290 yes 5.61 5.61 125 crowds (100,80) 106 1258605 no 7.27 7.47 176 (100,100) 2 · 106 2518270 no 9.22 9.34 383 (100,60) 878405 1105325 yes 5.18 5.18 134 crowds-obs (100,80) 106 1258569 no 6.42 6.98 206 (100,100) 2 · 106 2518220 no 8.39 8.79 462 predator − 3 · 106 3716578 no 99.14 ? 369 (5) 106 1589528 no 4.13 4.13 15 coupon (7) 3 · 106 5379492 no 5.86 6.38 46 (10) − − no 0 10.1 TO (5) 106 1750932 no 2.57 2.57 13 coupon-obs (7) 2 · 106 3752912 no 4.22 4.23 30 (10) − − no 0 6.96 TO (5) 106 1356463 no 11.41 11.42 15 coupon-classic (7) 106 1393360 no 18.15 18.15 21 (10) − − no 0 29.29 TO case the model was not fully explored, because our prototype always expands 106 states before it does the next model checking call. The next three columns show the probability bound (λ), the result that the tool could achieve as well as the actual answer to the query on the full (potentially infinite) model. Due to space constraints, we rounded these figures to two significant digits. We report on the time in seconds that the prototype took to establish the result (TO = 3600 sec.). We observe that for most examples it suffices to perform few unfolding steps to achieve more than 90% of the actual probability. For example, for the largest crowds-obs program, 3 · 106 states are expanded, meaning that three unfolding steps were performed. Answering queries on programs including an observe statement can be costlier (crowds vs. crowds-obs), but does not need to be (coupon vs. coupon-obs). In the latter case, the observe statement prunes some paths early that were not promising to begin with, whereas in the former case, the observe statement only happens at the very end, which intuitively makes it harder for the search to find target states. We are able to obtain non-trivial lower bounds for all but two case studies. For herman-nd, not all of the (nondeterministically chosen) initial states were explored, because our exploration order currently does not favour states that influence the obtained result the most. Similarly, for the largest coupon collector examples, the time limit did not allow for finding one target state. Again, an exploration heuristic that is more directed towards these could potentially improve performance drastically. Table 2 shows the results for computing the expected value of program variables at terminating states. For technical reasons, our prototype currently cannot perform more than one unfolding step for this type of query. To achieve mean- expected value 2 1.5 1 0.5 lower bound on exp. number of draws actual value 0 10000 20000 30000 40000 50000 60000 number of explored states (a) coupon-obs (5) 100 90 80 70 60 50 40 30 20 lower bound on exp. number of goats 10 500000 1x106 1.5x106 2x106 2.5x106 number of explored states expected value 3 2.5 (b) predator Fig. 3. The obtained values approach the actual value from below. (a) after 9 iterations (b) after 13 iterations Fig. 4. Analyzing parametric models yields violating parameter instances. ingful results, we therefore vary the number of explored states until 90% of the actual result is achieved. Note that for the predator program, the actual value for the query is not known to us, so we report on the value at which the result only grows very slowly. The results are similar to the probability case in that most often a low number of states suffices to show meaningful lower bounds. Unfortunately — as before — we can only prove a trivial lower bound for the largest coupon collector examples. Figure 3 illustrates how the obtained lower bounds approach the actual expected value with increasing number of explored states for two case studies. For example, in the left picture one can observe that exploring 60000 states is enough to obtain a very precise lower bound on the expected number of rounds the collector needs to gather all five coupons, as indicated by the dashed line. Finally, we analyze a parametric version of the crowds model that uses the parameters f and b to leave the probabilities (i) for a crowd member to be corrupt (b) and (ii) of forwarding (instead of delivering) a message (f ) unspecified. In each iteration of our algorithm, we obtain a rational function describing a lower bound on the actual probability of observing the real sender of the message more than once for each parameter valuation. Figure 4 shows the regions of the parameter space in which the protocol was determined to be unsafe (after iterations 9 and 13, respectively) in the sense that the probability to identify the real sender exceeds 12 . Since the results obtained over different iterations are monotonically increasing, we can conclude that all parameter valuations that were proved to be unsafe in some iteration are in fact unsafe in the full model. This in turn means that the blue area in Figure 4 grows in each iteration. 5 Conclusion and Future Work We presented a direct verification method for probabilistic programs employing probabilistic model checking. We conjecture that the basic idea would smoothly translate to reasoning about recursive probabilistic programs [25]. In the future we are interested in how loop invariants [26] can be utilized to devise complete model checking procedures preventing possibly infinite loop unrollings. This is especially interesting for reasoning about covariances [27], where a mixture of invariant–reasoning and successively constructing the operational MC would yield sound over- and underapproximations of covariances. To extend the gain for the user, we will combine this approach with methods for counterexamples [28], which can be given in terms of the programming language [29,19]. Moreover, it seem promising to investigate how approaches to automatically repair a probabilistic model towards satisfaction of properties [30,31] can be transferred to programs. References 1. Gordon, A.D., Henzinger, T.A., Nori, A.V., Rajamani, S.K.: Probabilistic programming. In: FOSE, ACM Press (2014) 167–181 2. Sankaranarayanan, S., Chakarov, A., Gulwani, S.: Static analysis for probabilistic programs: inferring whole program properties from finitely many paths. In: PLDI, ACM (2013) 447–458 3. Claret, G., Rajamani, S.K., Nori, A.V., Gordon, A.D., Borgström, J.: Bayesian inference using data flow analysis. In: ESEC/SIGSOFT FSE, ACM Press (2013) 92–102 4. Gretz, F., Katoen, J.P., McIver, A.: Operational versus weakest pre-expectation semantics for the probabilistic guarded command language. Perform. Eval. 73 (2014) 110–132 5. Kwiatkowska, M., Norman, G., Parker, D.: Prism 4.0: Verification of probabilistic real-time systems. In: CAV. Volume 6806 of LNCS, Springer (2011) 585–591 6. Hahn, E.M., Li, Y., Schewe, S., Turrini, A., Zhang, L.: IscasMC: A web-based probabilistic model checker. In: FM. Volume 8442 of LNCS, Springer (2014) 312– 317 7. Katoen, J.P., Zapreev, I.S., Hahn, E.M., Hermanns, H., Jansen, D.N.: The ins and outs of the probabilistic model checker MRMC. Performance Evaluation 68(2) (2011) 90–104 8. Kattenbelt, M.: Automated Quantitative Software Verification. PhD thesis, Oxford University (2011) 9. Sharir, M., Pnueli, A., Hart, S.: Verification of probabilistic programs. SIAM Journal on Computing 13(2) (1984) 292–314 10. Vardi, M.Y.: Automatic verification of probabilistic concurrent finite-state programs. In: FOCS, IEEE Computer Society (1985) 327–338 11. Baier, C., Katoen, J.P.: Principles of Model Checking. The MIT Press (2008) 12. Baier, C., Klein, J., Klüppelholz, S., Märcker, S.: Computing conditional probabilities in Markovian models efficiently. In: TACAS. Volume 8413 of LNCS, Springer (2014) 515–530 13. Dehnert, C., Junges, S., Jansen, N., Corzilius, F., Volk, M., Bruintjes, H., Katoen, J., Ábrahám, E.: Prophesy: A probabilistic parameter synthesis tool. In: CAV. Volume 9206 of LNCS, Springer (2015) 214–231 14. Quatmann, T., Dehnert, C., Jansen, N., Junges, S., Katoen, J.: Parameter synthesis for Markov models: Faster than ever. CoRR abs/1602.05113 (2016) 15. Dijkstra, E.W.: A Discipline of Programming. Prentice Hall (1976) 16. McIver, A., Morgan, C.: Abstraction, Refinement And Proof For Probabilistic Systems. Springer (2004) 17. Jansen, N., Kaminski, B.L., Katoen, J., Olmedo, F., Gretz, F., McIver, A.: Conditioning in probabilistic programming. Electr. Notes Theor. Comput. Sci. 319 (2015) 199–216 18. Kaminski, B.L., Katoen, J.P.: On the hardness of almost-sure termination. In: MFCS. Volume 9234 of LNCS, Springer (2015) 19. Dehnert, C., Jansen, N., Wimmer, R., Ábrahám, E., Katoen, J.: Fast debugging of PRISM models. In: ATVA. Volume 8837 of LNCS, Springer (2014) 146–162 20. Jansen, N., Dehnert, C., Kaminski, B.L., Katoen, J., Westhofen, L.: Bounded model checking for probabilistic programs. CoRR abs/1605.04477 (2016) 21. Reiter, M.K., Rubin, A.D.: Crowds: Anonymity for web transactions. ACM Trans. on Information and System Security 1(1) (1998) 66–92 22. Herman, T.: Probabilistic self-stabilization. Inf. Process. Lett. 35(2) (1990) 63–67 23. Brauer, F., Castillo-Chavez, C.: Mathematical Models in Population Biology and Epidemiology. Texts in Applied Mathematics. Springer New York (2001) 24. Erds, P., Rnyi, A.: On a classical problem of probability theory. Publ. Math. Inst. Hung. Acad. Sci., Ser. A 6 (1961) 215 – 220 25. Olmedo, F., Kaminski, B., Katoen, J.P., Matheja, C.: Reasoning about recursive probabilistic programs. In: LICS. (2016) [to appear]. 26. Gretz, F., Katoen, J.P., McIver, A.: PRINSYS - on a quest for probabilistic loop invariants. In: QEST. Volume 8054 of LNCS, Springer (2013) 193–208 27. Kaminski, B., Katoen, J.P., Matheja, C.: Inferring covariances for probabilistic programs. In: QEST. Volume 9826 of LNCS, Springer (2016) [to appear]. 28. Ábrahám, E., Becker, B., Dehnert, C., Jansen, N., Katoen, J., Wimmer, R.: Counterexample generation for discrete-time Markov models: An introductory survey. In: SFM. Volume 8483 of Lecture Notes in Computer Science, Springer (2014) 65–121 29. Wimmer, R., Jansen, N., Abraham, E., Katoen, J.P.: High-level counterexamples for probabilistic automata. Logical Methods in Computer Science 11(1:15) (2015) 30. Bartocci, E., Grosu, R., Katsaros, P., Ramakrishnan, C.R., Smolka, S.A.: Model repair for probabilistic systems. In: TACAS. Volume 6605 of Lecture Notes in Computer Science, Springer (2011) 326–340 31. Pathak, S., Ábrahám, E., Jansen, N., Tacchella, A., Katoen, J.P.: A greedy approach for the efficient repair of stochastic models. In: NFM. Volume 9058 of LNCS, Springer (2015) 295–309 A Models A.1 1 2 3 4 5 int int int int int coupon-obs (5) coup0 coup1 coup2 coup3 coup4 := := := := := 0; 0; 0; 0; 0; 6 7 8 9 int draw1 := 0; int draw2 := 0; int draw3 := 0; 10 11 int numberDraws := 0; 12 13 14 15 16 17 while (!( coup0 = 1) | !( coup1 = 1) | !( coup2 = 1) | !( coup3 = 1) | !( coup4 = 1) ) { draw1 := unif (0 ,4) ; draw2 := unif (0 ,4) ; draw3 := unif (0 ,4) ; numberDraws := numberDraws + 1; 18 observe ( draw1 != draw2 & draw1 != draw3 & draw2 != draw3 ) ; 19 20 if ( draw1 = 0 | draw2 = 0 | draw3 = 0) { coup0 := 1; } if ( draw1 = 1 | draw2 = 1 | draw3 = 1) { coup1 := 1; } if ( draw1 = 2 | draw2 = 2 | draw3 = 2) { coup2 := 1; } if ( draw1 = 3 | draw2 = 3 | draw3 = 3) { coup3 := 1; } if ( draw1 = 4 | draw2 = 4 | draw3 = 4) { coup4 := 1; } 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 } A.2 1 2 3 4 5 int int int int int coupon (5) coup0 coup1 coup2 coup3 coup4 := := := := := 0; 0; 0; 0; 0; 6 7 8 9 int draw1 := 0; int draw2 := 0; int draw3 := 0; 10 11 int numberDraws := 0; 12 13 14 15 16 17 while (!( coup0 = 1) | !( coup1 = 1) | !( coup2 = 1) | !( coup3 = 1) | !( coup4 = 1) ) { draw1 := unif (0 ,4) ; draw2 := unif (0 ,4) ; draw3 := unif (0 ,4) ; numberDraws := numberDraws + 1; 18 if ( draw1 = 0 | draw2 = 0 | draw3 = 0) { coup0 := 1; } if ( draw1 = 1 | draw2 = 1 | draw3 = 1) { coup1 := 1; } if ( draw1 = 2 | draw2 = 2 | draw3 = 2) { coup2 := 1; } if ( draw1 = 3 | draw2 = 3 | draw3 = 3) { coup3 := 1; } if ( draw1 = 4 | draw2 = 4 | draw3 = 4) { coup4 := 1; } 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 } 1 2 3 4 5 A.3 crowds-obs (100, 60) int int int int int delivered := 0; lastSender := 0; remainingRuns := 60; observeSender := 0; observeOther := 0; 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 while ( remainingRuns > 0) { while ( delivered = 0) { { if ( lastSender = 0) { observeSender := observeSender + 1; } else { observeOther := observeOther + 1; } lastSender := 0; delivered := 1; } [0.091] { { { lastSender :=0; } [1/100] { lastSender := 1; } } [0.8] { lastSender := 0; // When not forwarding , the message is delivered here delivered := 1; } } } // Set up new run . delivered := 0; remainingRuns := remainingRuns - 1; } observe ( observeOther > 15) ; A.4 1 2 3 4 5 int int int int int crowds (100, 60) delivered := 0; lastSender := 0; remainingRuns := 60; observeSender := 0; observeOther := 0; 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 while ( remainingRuns > 0) { while ( delivered = 0) { { if ( lastSender = 0) { observeSender := observeSender + 1; } else { observeOther := observeOther + 1; } lastSender := 0; delivered := 1; } [0.091] { { { lastSender :=0; } [1/100] { lastSender := 1; } } [0.8] { lastSender := 0; // When not forwarding , the message is delivered here delivered := 1; } } } // Set up new run . delivered := 0; remainingRuns := remainingRuns - 1; } A.5 crowds (100, 60) parametric This program is parametric with the parameters f (probability of forwarding the message) and b (probability that a crowd member is bad). 1 2 3 4 5 int int int int int delivered := 0; lastSender := 0; remainingRuns := 60; observeSender := 0; observeOther := 0; 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 while ( remainingRuns > 0) { while ( delivered = 0) { { if ( lastSender = 0) { observeSender := observeSender + 1; } else { observeOther := observeOther + 1; } lastSender := 0; delivered := 1; } [b] { { { lastSender :=0; } [1/100] { lastSender := 1; } } [f] { lastSender := 0; // When not forwarding , the message is delivered here delivered := 1; } } } // Set up new run . delivered := 0; remainingRuns := remainingRuns - 1; } 

arXiv:1704.00997v1 [math.AC] 4 Apr 2017 SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA Abstract. The notion of 2-AGL ring in dimension one which is a natural generalization of almost Gorenstein local ring is posed in terms of the rank of Sally modules of canonical ideals. The basic theory is developed, investigating also the case where the rings considered are numerical semigroup rings over fields. Examples are explored. Contents 1. Introduction 2. Preliminaries 3. 2-AGL rings and Proof of Theorem 1.4 4. 2-AGL rings obtained by idealization 5. The algebra m : m 6. Numerical semigroup rings References 1 5 8 13 14 20 28 1. Introduction The destination of this research is to find a good notion of Cohen-Macaulay local rings of positive dimension which naturally generalizes Gorenstein local rings. In dimension one, the research has started from the works of V. Barucci and R. Fröberg [BF] and the second, the fourth authors, and T. T. Phuong [GMP]. In [BF] Barucci and Fröberg introduced the notion of almost Gorenstein ring in the case where the local rings are one-dimensional and analytically unramified. They explored also numerical semigroup rings over fields and developed a beautiful theory. In [GMP] the authors extended the notion given by [BF] to arbitrary one-dimensional Cohen-Macaulay local rings and showed that their new definition works well to analyze almost Gorenstein rings which are analytically ramified. In [GTT] the second author, R. Takahashi, and N. Taniguchi gave the notion of almost Gorenstein local/graded rings of higher dimension. Their research is still in progress, exploring, for example, the problem of when the Rees algebras of ideals/modules are 2010 Mathematics Subject Classification. 13H10, 13H15, 13A30. Key words and phrases. Cohen-Macaulay ring, Gorenstein ring, almost Gorenstein ring, canonical ideal, parameter ideal, Rees algebra, Sally module. The first author was partially supported by the International Research Supporting Program of Meiji University. The second author was partially supported by JSPS Grant-in-Aid for Scientific Research (C) 25400051. The fourth author was partially supported by JSPS Grant-in-Aid for Scientific Research 26400054. 1 2 TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA almost Gorenstein rings; see [GMTY1, GMTY2, GMTY3, GMTY4, GRTT, T]. One can consult [El] for a deep investigation of canonical ideals in dimension one. The interests of the present research are a little different from theirs and has been strongly inspired by [GGHV, Section 4] and [V2]. Our aim is to discover a good candidate for natural generalization of almost Gorenstein rings. Even though our results are at this moment restricted within the case of dimension one, we expect that a higher dimensional notion might be possible after suitable modifications. However, before entering more precise discussions, let us fix our terminology. Throughout this paper let (R, m) be a Cohen-Macaulay local ring of dimension one and let I be a canonical ideal of R. Assume that I contains a parameter ideal Q = (a) of R as a reduction. We set K = aI = { xa | x ∈ I} in the total ring Q(R) of fractions of R and let S = R[K]. Therefore, K is a fractional ideal of R such that R ⊆ K ⊆ R and S is a module-finite extension of R, where R denotes the integral closure of R in Q(R). We denote by c = R : S the conductor. With this notation the second and the fourth authors and T. T. Phuong [GMP] closely studied the almost Gorenstein property of R. Here let us recall the definition of almost Gorenstein local rings given by [GTT], which works with a suitable modification in higher dimensional cases also. Notice that in our setting, the condition in Definition 1.1 below is equivalent to saying that mK ⊆ R ([GTT, Proposition 3.4]). Definition 1.1 ([GTT, Definition 1.1]). Suppose that R possesses the canonical module KR . Then we say that R is an almost Gorenstein local (AGL for short) ring, if there is an exact sequence 0 → R → KR → C → 0 of R-modules such that mC = (0). Consequently, R is an AGL ring if R is a Gorenstein ring (take C = (0)) and Definition 1.1 certifies that once R is an AGL ring, although it is not a Gorenstein ring, R can be embedded into its canonical module KR and the difference KR /R is little. Let ei (I) (i = 0, 1) denote the Hilbert coefficients of R with respect to I (notice that our canonical ideal I is an m-primary ideal of R) and let r(R) = ℓR (Ext1R (R/m, R)) denote the Cohen-Macaulay type of R. With this notation the following characterization of AGL rings is given by [GMP], which was a starting point of the present research. Theorem 1.2 ([GMP, Theorem 3.16]). The following conditions are equivalent. (1) (2) (3) (4) (5) R is an AGL ring but not a Gorenstein ring. e1 (I) = r(R). e1 (I) = e0 (I) − ℓR (R/I) + 1. ℓR (S/K) = 1, that is S = K : m. ℓR (I 2 /QI) = 1. SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS 3 (6) S = m : m but R is not a DVR. When this is the case, I 3 = QI 2 and ℓR (R/I n+1   n+1 − r(R) ) = (r(R) + ℓR (R/I) − 1) 1 for all n ≥ 1, where ℓR (∗) denotes the length. The aim of the present research is to ask for a generalization of AGL rings in dimension one. For the purpose we notice that Condition (3) in Theorem 1.2 is equivalent to saying that the Sally module of I with respect to Q has rank one. In order to discuss more explicitly, here let us explain the notion of Sally module ([V1]). The results we recall below hold true in Cohen-Macaulay local rings (R, m) of arbitrary positive dimension for all m-primary ideals I and reductions Q of I which are parameter ideals of R ([GNO]). Let us, however, restrict our attention to the case where dim R = 1 and I is a canonical ideal of R. Let T = R(Q) = R[Qt] and R = R(I) = R[It] respectively denote the Rees algebras of Q and I, where t is an indeterminate over R. We set SQ (I) = IR/IT and call it the Sally module of I with respect to Q ([V1]). Then SQ (I) is a finitely generated graded T -module with dimT SQ (I) ≤ 1 whose grading is given by ( (0) if n ≤ 0, [SQ (I)]n = n+1 n I /Q I if n ≥ 1 for each n ∈ Z ([GNO, Lemma 2.1]). Let p = mT and B = T /p (= (R/m)[T ] the polynomial ring). We set rank SQ (I) = ℓTp ([SQ (I)]p ) and call it the rank of SQ (I). Then AssT SQ (I) ⊆ {p} and rank SQ (I) = e1 (I) − [e0 (I) − ℓR (R/I)] ([GNO, Proposition 2.2]). As we later confirm in Section 2 (Theorem 2.5), the invariant rank SQ (I) is equal to ℓR (S/K) and is independent of the choice of canonical ideals I and their reductions Q. By [S3, V1] it is known that Condition (3) in Theorem 1.2 is equivalent to saying that rank SQ (I) = 1, which is also equivalent to saying that SQ (I) ∼ = B(−1) as a graded T -module. According to these stimulating facts, as is suggested by [GGHV, Section 4] it seems reasonable to expect that one-dimensional Cohen-Macaulay local rings R which satisfy the condition rank SQ (I) = 2, that is e1 (I) = e0 (I) − ℓR (R/I) + 2 for canonical ideals I could be a good candidate for generalization of AGL rings. 4 TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA Chasing the expectation, we now give the following. Definition 1.3. We say that R is a 2-almost Gorenstein local (2-AGL for short) ring, if rank SQ (I) = 2. In this paper we shall closely explore the structure of 2-AGL rings to show the above expectation comes true. Let us note here the basic characterization of 2-AGL rings, which starts the present paper. Theorem 1.4. The following conditions are equivalent. (1) (2) (3) (4) (5) (6) (7) R is a 2-AGL ring. There is an exact sequence 0 → B(−1) → SQ (I) → B(−1) → 0 of graded T -modules. K 2 = K 3 and ℓR (K 2 /K) = 2. I 3 = QI 2 and ℓR (I 2 /QI) = 2. R is not a Gorenstein ring but ℓR (S/[K : m]) = 1. ℓR (S/K) = 2. ℓR (R/c) = 2. When this is the case, m·SQ (I) 6= (0), whence the exact sequence given by Condition (2) is not split, and we have   n+1 n+1 − (e0 (I) − ℓR (R/I) + 2) ℓR (R/I ) = e0 (I) 1 for all n ≥ 1. See [HHS] for another direction of generalization of Gorenstein rings. In [HHS] the authors posed the notion of nearly Gorenstein ring and developed the theory. Here let us note that in dimension one, 2-AGL rings are not nearly Gorenstein and nearly Gorenstein rings are not 2-AGL rings (see [HHS, Remark 6.2, Theorem 7.4], Theorems 1.4, 3.6). Here let us explain how this paper is organized. In Section 2 we will summarize some preliminaries, which we need throughout this paper. The proof of Theorem 1.4 shall be given in Section 3. In Section 3 we study also the question how the 2-AGL property of rings is preserved under flat base changes. Condition (7) in Theorem 1.4 is really practical, which we shall show in Sections 5 and 6. In Section 4 we study 2-AGL rings obtained by idealization. We will show that A = R ⋉ c is a 2-AGL ring if and only if so is R, which enables us, starting from a single 2-AGL ring, to produce an infinite family {An }n≥0 of 2-AGL rings which are analytically ramified (Example 4.3). Let v(R) (resp. e(R)) denote the embedding dimension of R (resp. the multiplicity e0m (R) of R with respect to m). We set B = m : m. Then it is known by [GMP, Theorem 5.1] that R is an AGL ring with v(R) = e(R) if and only if B is a Gorenstein ring. In Section 5 we shall closely study the corresponding phenomenon of the 2-AGL property. We will show that if R is a 2-AGL ring with v(R) = e(R), then B contains a unique maximal ideal M such that SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS 5 BN is a Gorenstein ring for all N ∈ Max B \ {M} and BM is an AGL ring which is not a Gorenstein ring. The converse is also true under suitable conditions, including the specific one that K/R is a free R/c-module. Section 6 is devoted to the analysis of the case where R = k[[H]] (k a field) are the semigroup rings of numerical semigroups H. We will give in several cases a characterization for R = k[[H]] to be 2-AGL rings in terms of numerical semigroups H. 2. Preliminaries The purpose of this section is to summarize some auxiliary results, which we later need throughout this paper. First of all, let us make sure of our setting. Setting 2.1. Let (R, m) be a Cohen-Macaulay local ring with dim R = 1, possessing the canonical module KR . Let I be a canonical ideal of R. Hence I is an ideal of R such that I 6= R and I ∼ = KR as an R-module. We assume that I contains a parameter ideal Q = (a) of R as a reduction. Let o I nx K= = |x∈I a a in the total ring Q(R) of fractions of R. Hence K is a fractional ideal of R such that R ⊆ K ⊆ R, where R denotes the integral closure of R in Q(R). Let S = R[K] and c = R : S. We denote by SQ (I) = IR/IT the Sally module of I with respect to Q, where T = R[Qt], R = R[It], and t is an indeterminate over R. Let B = T /mT and ei (I) (i = 0, 1) the Hilbert coefficients of I. We notice that a one-dimensional Cohen-Macaulay local ring (R, m) contains a canonib is a Gorenstein ring, where R b denotes the m-adic completion cal ideal if and only if Q(R) of R ([HK, Satz 6.21]). Also, every m-primary ideal of R contains a parameter ideal as a reduction, once the residue class field R/m of R is infinite. If K is a given fractional ideal of R such that R ⊆ K ⊆ R and K ∼ = KR as an R-module, then taking a non-zerodivisor a ∈ m so that aK ( R, I = aK is a canonical ideal of R such that Q = (a) as a reduction and K = aI . Therefore, the existence of canonical ideals I of R containing parameter ideals as reductions is equivalent to saying that there are fractional ideals K of R such that R ⊆ K ⊆ R and K ∼ = KR as an R-module (cf. [GMP, Remark 2.10]). We have for all n ≥ 0 K n+1 /K n ∼ = I n+1 /QI n as an R-module, whence K/R ∼ = I/Q. Let rQ (I) = min{n ≥ 0 | I n+1 = QI n } be the reduction number of I with respect to Q. Let us begin with the following. Lemma 2.2. The following assertions hold true. (1) rQ (I) = min{n ≥ 0 | K n = K n+1 }. Hence S = K n for all n ≥ rQ (I). 6 TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA (2) Let b ∈ I. Then (b) is a reduction of I if and only if When this is the case, S = R[ Ib ] and rQ (I) = r(b) (I). b a is an invertible element of S. Proof. (1) The first equality is clear, since I = aK. The second one follows from the fact S that S = n≥0 K n . (2) Suppose that (b) is a reduction of I and choose an integer n ≫ 0 so that S = K n n+1 n and I n+1 = bI n . Then since aI n+1 = ab · aI n , we get S = ab S, whence ab is an invertible element of S. The reverse implication is now clear. To see S = R[ Ib ], notice that S ⊇ ab · aI = Ib , because ab ∈ S. Hence S ⊇ R[ Ib ] and by symmetry we get S = R[ Ib ]. To see rQ (I) = r(b) (I), let n = rQ (I). Then K n+1 = S = ab S = ab K n by Assertion (1) , so that I n+1 = bI n . Therefore, rQ (I) ≥ r(b) (I), whence rQ (I) = r(b) (I) by symmetry.  Proposition 2.3 ([GMP, GTT]). The following assertions hold true. (1) c = K : S and ℓR (R/c) = ℓR (S/K). (2) c = R : K if and only if S = K 2 . (3) R is a Gorenstein ring if and only if rQ (I) ≤ 1. When this is the case, I = Q, that is K = R. (4) R is an AGL ring if and only if mK 2 ⊆ K. (5) Suppose that R is an AGL ring but not a Gorenstein ring. Then rQ (I) = 2 and ℓR (K 2 /K) = 1. Proof. (1) See [GMP, Lemma 3.5 (2)]. (2) Since K : K = R ([HK, Bemerkung 2.5 a)]), we have R : K = (K : K) : K = K : K 2 . Because c = K : S by Assertion (1), c = R : K if and only if K : S = K : K 2 . The latter condition is equivalent to saying that S = K 2 ([HK, Definition 2.4]). (3), (5) See [GMP, Theorems 3.7, 3.16]. (4) As K : K = R, mK 2 ⊆ K if and only if mK ⊆ R. By [GTT, Proposition 3.4] the latter condition is equivalent to saying that R is an AGL ring.  Let µR (M) denote, for each finitely generated R-module M, the number of elements in a minimal system of generators of M. Corollary 2.4. The following assertions hold true. (1) K : m ⊆ K 2 if R is not a Gorenstein ring. If R is not an AGL ring, then K : m ( K 2 . (2) mK 2 + K = K : m, if ℓR (K 2 /K) = 2. (3) Suppose that R is not a Gorenstein ring. Then µR (S/K) = r(R/c). Therefore, R/c is a Gorenstein ring if and only if µR (S/K) = 1. Proof. (1) As R is not a Gorenstein ring, K 6= K 2 by Lemma 2.2 (1) and Proposition 2.3 (3). Therefore, K : m ⊆ K 2 , since ℓR ([K : m]/K) = 1 ([HK, Satz 3.3 c)]) and ℓR (K 2 /K) < ∞. Proposition 2.3 (4) implies that K : m 6= K 2 if R is not an AGL ring. SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS 7 (2) Suppose that ℓR (K 2 /K) = 2. Then R is not an AGL ring. Hence mK 2 6⊆ K, while by Assertion (1) K : m ( K 2 . Therefore, since m2 K 2 ⊆ K, we get K ( mK 2 + K ⊆ K : m ( K 2 , whence mK 2 + K = K : m, because ℓR (K 2 /K) = 2. (3) We get µR (S/K) = ℓR (S/(mS + K)) = ℓR ([K : (mS + K)]/(K : S)), where the second equality follows by duality ([HK, Bemerkung 2.5 c)]). Since K : K = R and c = K : S by Proposition 2.3 (1), µR (S/K) = ℓR ([K : (mS + K)]/(K : S)) = ℓR ([(K : mS) ∩ (K : K)] /c) = ℓR ([(K : S) : m] ∩ R] /c) = r(R/c), where the last equality follows from the fact [(K : S) : m] ∩ R = c :R m = (0) :R m/c.  We close this section with the following, which guarantees that rank SQ (I) and S = R[K] are independent of the choice of canonical ideals I of R and reductions Q of I. Assertions (1) and (2) of Theorem 2.5 are more or less known (see, e.g., [GMP, GGHV]). In particular, in [GGHV] the invariant ℓR (I/Q) is called the canonical degree of R and intensively investigated. Let us include here a brief proof in our context for the sake of completeness. Theorem 2.5. The following assertions hold true. (1) ℓR (S/K) = e1 (I) − [e0 (I) − ℓR (R/I)]. Hence rank SQ (I) = ℓR (S/K) = ℓR (R/c). (2) The invariants rQ (I), ℓR (S/K), and ℓR (K/R) are independent of the choice of I and Q. (3) The ring S = R[K] is independent of the choice of I and Q. Proof. (1) We have K/R ∼ = I/Q as an R-module, whence ℓR (K/R) = ℓR (I/Q) = ℓR (R/Q) − ℓR (R/I) = e0 (I) − ℓR (R/I). So, the first equality is clear, because ℓR (S/R) = e1 (I) by [GMP, Lemma 2.1] and ℓR (S/K) = ℓR (S/R) − ℓR (K/R). See [GNO, Proposition 2.2] and Proposition 2.3 (1) for the second and the third equalities. (2), (3) The invariant ℓR (S/R) = e1 (I) is independent of the choice of I, since the first Hilbert coefficient e1 (I) of canonical ideals I is independent of the choice of I ([GMP, Corollary 2.8]). Therefore, because ℓR (I/Q) = e0 (I) − ℓR (R/I) depends only on I, to see that ℓR (S/K) = ℓR (S/R) − ℓR (I/Q) is independent of the choice of I and Q, it is enough to show that ℓR (K/R) = ℓR (I/Q) is independent of the choice of I. Let J be another canonical ideal of R and assume that (b) is a reduction of J. Then, since I ∼ = J as an R-module, J = εI for some invertible element ε of Q(R) ([HK, Satz 2.8]). Let b′ = εa. 8 TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA Then (b′ ) is a reduction of J, r(b′ ) (J) = rQ (I), and ℓR (J/(b′ )) = ℓR (I/Q), clearly. Hence ℓR (J/(b)) = ℓR (J/(b′ )) = ℓR (I/Q), which is independent of I. Because r(b) (J) = r(b′ ) (J) by Lemma 2.2 (2), the reduction number rQ (I) is independent of the choice of canonical ideals I and reductions Q of I. Because R[ aI ] = R[ bJ′ ] = R[ Jb ] where the second equality follows from Lemma 2.2 (2), the ring S = R[K] is independent of the choice of I and Q as well.  3. 2-AGL rings and Proof of Theorem 1.4 The main purpose of this section is to prove Theorem 1.4. Let us maintain Setting 2.1. We begin with the following. Lemma 3.1. The ring R is 2-AGL if and only if K 2 = K 3 and ℓR (K 2 /K) = 2. Proof. If R is a 2-AGL ring, then ℓR (S/K) = 2 by Theorem 2.5 (1), while by Proposition 2.3 (5) ℓR (K 2 /K) ≥ 2 since R is not an AGL ring; therefore S = K 2 . Conversely, if K 2 = K 3 , then K 2 = K n for all n ≥ 2, so that S = K 2 . Hence the equivalence follows.  Before going ahead, let us note basic examples of 2-AGL rings. Later we will give more examples. Let r(R) = ℓR (Ext1R (R/m, R)) denote the Cohen-Macaulay type of R. Example 3.2. Let k[[t]] and k[[X, Y, Z, W ]] denote the formal power series rings over a field k. (1) Consider the rings R1 = k[[t3 , t7 , t8 ]], R2 = k[[X, Y, Z, W ]]/(X 3 − Y Z, Y 2 − XZ, Z 2 − X 2 Y, W 2 − XW ), and R3 = k[[t3 , t7 , t8 ]] ⋉ k[[t]] (the idealization of k[[t]] over k[[t3 , t7 , t8 ]]). Then these rings R1 , R2 , and R3 are 2-AGL rings. The ring R1 is an integral domain, R2 is a reduced ring but not an integral domain, and R3 is not a reduced ring. (2) Let c ≥ 4 be an integer such that c 6≡ 0 mod 3 and set R = k[[t3 , tc+3 , t2c ]]. Then R is a 2-AGL ring such that v(R) = e(R) = 3 and r(R) = 2. We note basic properties of 2-AGL rings. Proposition 3.3. Suppose that R is a 2-AGL ring and set r = r(R). Then we have the following. (1) c = K : S = R : K. (2) ℓR (R/c) = 2. Hence there is a minimal system x1 , x2 , . . . , xn of generators of m such that c = (x21 ) + (x2 , x3 , . . . , xn ). (3) S/K ∼ = R/c and S/R ∼ = K/R ⊕ R/c as R/c-modules. ⊕ℓ (4) K/R ∼ = (R/c) ⊕ (R/m)⊕m as an R/c-module for some ℓ > 0 and m ≥ 0 such that ℓ + m = r − 1. Hence ℓR (K/R) = 2ℓ + m. In particular, K/R is a free R/c-module if and only if ℓR (K/R) = 2(r − 1). SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS 9 (5) µR (S) = r. Proof. (1), (2) We have c = K : S and ℓR (R/c) = 2 by Proposition 2.3 (1). Hence R : K = c, because R : K = (K : K) : K = K : K 2 . The second assertion in Assertion (2) is clear, because m2 ⊆ c and ℓR (m/c) = 1. (3), (4) Because R/c is an Artinian Gorenstein local ring, any finitely generated R/cmodule M contains R/c as a direct summand, once M is faithful. If M is not faithful, then (0) :R/c M ⊇ m/c as ℓR (R/c) = 2, so that M is a vector space over R/m. Therefore, every finitely generated R/c-module M has a unique direct sum decomposition M∼ = (R/c)⊕ℓ ⊕ (R/m)⊕m with ℓ, m ≥ 0 such that µR/c (M) = ℓ + m. Because by Assertion (1) the modules S/R, K/R, and S/K are faithful over R/c, they contain R/c as a direct summand; hence S/K ∼ = R/c, because ℓR (S/K) = ℓR (R/c) by Proposition 2.3 (1). Consequently, the canonical exact sequence 0 → K/R → S/R → S/K → 0 of R/c-modules is split, so that S/R ∼ = K/R ⊕ R/c. Since µR (K/R) = r − 1 > 0, K/R contains R/c as a direct summand, whence K/R ∼ = (R/c)⊕ℓ ⊕ (R/m)⊕m with ℓ > 0 and m ≥ 0, where ℓ + m = r − 1 and ℓR (K/R) = 2ℓ + m. (5) This is now clear, since S/R ∼ = K/R ⊕ R/c.  The 2-AGL rings R such that K/R are R/c-free enjoy a certain specific property, which we will show in Section 5. Here let us note Example 3.4 (resp. Example 3.5) of 2-AGL rings, for which K/R is a free R/c-module (resp. not a free R/c-module). Let V = k[[t]] denote the formal power series ring over a field k. Example 3.4. Let e ≥ 3 and n ≥ 2 be integers. Let R = k[[te , {ten+i }1≤i≤e−2 , t2en−(e+1) ]] P and m the maximal ideal of R. Let K = R + 1≤i≤e−2 Rt(n−2)e+i . Then we have the following. (1) I = t2(n−1)e K is a canonical ideal of R containing (t2(n−1)e ) as a reduction. (2) R is a 2-AGL ring such that m2 = te m and r(R) = e − 1. (3) K/R ∼ = (R/c)⊕2(e−2) as an R/c-module. Example 3.5. Let e ≥ 4 be an integer. Let R = k[[te , {te+i }3≤i≤e−1 , t2e+1 , t2e+2 ]] and m P the maximal ideal of R. Let K = R + Rt + 3≤i≤e−1 Rti . Then we have the following. (1) (2) (3) (4) I = te+3 K is a canonical ideal of R containing (te+3 ) as a reduction. R is a 2-AGL ring such that m2 = te m and r(R) = e − 1. K/R ∼ = (R/c) ⊕ (R/m)⊕(e−3) as an R/c-module. m : m = k[[t3 , t4 , t5 ]]. 10 TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA We note the following. Theorem 3.6. Suppose that R is a 2-AGL ring. Then the following assertions hold true. (1) R is not an AGL ring. (2) There is an exact sequence 0 → B(−1) → SQ (I) → B(−1) → 0 of graded T -modules. (3) m·SQ (I) 6= (0). Therefore, the above exact sequence is not split. Proof. (1) Since ℓR (K 2 /K) = 2 by Lemma 3.1, by Proposition 2.3 (5) R is not an AGL ring. (2) We have mI 2 6⊆ QI by Proposition 2.3 (4), since I 2 /QI ∼ = K 2 /K. Therefore, as ℓR (I 2 /QI) = 2, we get ℓR (I 2 /[mI 2 + QI]) = ℓR ([mI 2 + QI]/QI) = 1. Let us write I 2 = QI + (f ) for some f ∈ I 2 . Then since ℓR ([mI 2 + QI]/QI) = 1 and mI 2 + QI = QI + mf , mI 2 + QI = QI + (αf ) for some α ∈ m. We set g = αf . Now remember that SQ (I) = T ·[SQ (I)]1 = T ·f t, because I 3 = QI 2 ([GNO, Lemma 2.1 (5)]), where ∗ denotes the image in SQ (I). Therefore, because (0) :T gt = mT , we get an exact sequence ϕ 0 → B(−1) − → SQ (I) → C → 0 of graded T -modules, where ϕ(1) = gt. Let ξ denote the image of f t in C = SQ (I)/T ·gt. Then C = T ξ and (0) :T C = mT , whence C ∼ = B(−1) as a graded T -module, and the result follows.  (3) Since [SQ (I)]1 ∼ = I 2 /QI as an R-module, we get m·SQ (I) 6= (0). We are in a position to prove Theorem 1.4. Proof of Theorem 1.4. (1) ⇒ (2) See Theorem 3.6. (1) ⇔ (3) See Lemma 3.1. (3) ⇔ (4) Remember that K n+1 /K n ∼ = I n+1 /QI n for all n ≥ 0. (2) ⇒ (1) We have rank SQ (I) = ℓTp ([SQ (I)]p ) = 2·ℓTp (Bp ) = 2, where p = mT . (1) ⇔ (6) ⇔ (7) See Theorem 2.5 (1). (1) ⇒ (5) By Theorem 3.6 (1) R is not an AGL ring. Hence K : m ( K 2 = S by Corollary 2.4 (1). Because ℓR ((K : m)/K) = 1 and ℓR (S/(K : m)) + ℓR ((K : m)/K) = ℓR (S/K) = 2, we get ℓR (S/(K : m)) = 1. SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS 11 See Theorem 3.6 (3) for the former part of the last assertion. To see the latter part, notice that   ℓR (R/I n+1) = ℓR (R/Qn+1 ) − ℓR (I n+1 /Qn I) + ℓR (Qn I/Qn+1 )   n+1 − [ℓR ([SQ (I)]n ) + ℓR (I/Q)] = e0 (I) 1   n+1 − [2 + ℓR (I/Q)] = e0 (I) 1 for all n ≥ 1, where the last equality follows from the exact sequence given by Condition (2). Thus   n+1 n+1 − [e0 (I) − ℓR (R/I) + 2] ℓR (R/I ) = e0 (I) 1 for all n ≥ 1.  Let us note a consequence of Theorem 1.4. Proposition 3.7. Suppose that r(R) = 2. Then the following conditions are equivalent. (1) R is a 2-AGL ring. (2) c = R : K and ℓR (K/R) = 2. (3) S = K 2 and ℓR (K/R) = 2. When this is the case, K/R ∼ = R/c as an R-module. Proof. (1) ⇒ (2) By Proposition 3.3 (4) K/R ∼ = R/c, since µR (K/R) = r(R) − 1 = 1. Hence ℓR (K/R) = ℓR (R/c) = 2. (2) ⇔ (3) Remember that S = K 2 if and only if c = R : K; see Proposition 2.3 (2). (2) ⇒ (1) Since µR (K/R) = 1, K/R ∼ = R/c, so that ℓR (R/c) = ℓR (K/R) = 2. Hence R is a 2-AGL ring by Theorem 1.4.  Let us explore the question of how the 2-AGL property is preserved under flat base changes. Let (R1 , m1 ) be a Cohen-Macaulay local ring of dimension one and let ϕ : R → R1 be a flat local homomorphism of local rings such that R1 /mR1 is a Gorenstein ring. Hence dim R1 /mR1 = 0 and KR1 ∼ = R1 ⊗R K as an R1 -module ([HK, Satz 6.14]). Notice that R1 ⊆ R1 ⊗R K ⊆ R1 ⊗R R ⊆ R1 in Q(R1 ). We set K1 = R1 ⊗R K. Then R1 also satisfies the conditions stated in Setting 2.1 and we have following. Proposition 3.8. For each n ≥ 0 the following assertions hold true. (1) K1n = K1n+1 if and only if K n = K n+1 . (2) ℓR1 (K1n+1 /K1n ) = ℓR1 (R1 /mR1 )·ℓR (K n+1 /K n ). 12 TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA Proof. The equalities follow from the isomorphisms K1n ∼ = R1 ⊗R K n , K1n+1 /K1n ∼ = R1 ⊗R (K n+1 /K n ) of R1 -modules.  We furthermore have the following. Theorem 3.9. The following conditions are equivalent. (1) R1 is a 2-AGL ring. (2) Either (i) R is an AGL ring and ℓR1 (R1 /mR1 ) = 2 or (ii) R is a 2-AGL ring and mR1 = m1 . Proof. Suppose that R1 is a 2-AGL ring. Then K12 = K13 and ℓR1 (K12 /K1 ) = 2. Therefore, K 2 = K 3 and ℓR1 (K12 /K1 ) = ℓR1 (R1 /mR1 )·ℓR (K 2 /K) = 2 by Proposition 3.8. We have ℓR (K 2 /K) = 1 (resp. ℓR (K 2 /K) = 2) if ℓR1 (R1 /mR1 ) = 2 (resp. ℓR1 (R1 /mR1 ) = 1), whence the implication (1) ⇒ (2) follows. The reverse implication is now clear.  Example 3.10. Let n ≥ 1 be an integer and let R1 = R[X]/(X n + α1 X n−1 + · · · + αn ), where R[X] denotes the polynomial ring and αi ∈ m for all 1 ≤ i ≤ n. Then R1 is a flat local R-algebra with m1 = mR1 + (x) (here x denotes the image of X in R1 ) and R1 /mR1 = (R/m)[X]/(X n ) is a Gorenstein ring. Since ℓR1 (R1 /mR1 ) = n, taking n = 1 (resp. n = 2), we get R1 is an AGL ring (resp. R1 is a 2-AGL ring). Notice that if R is an integral domain and 0 6= α ∈ m, then R1 = R[X]/(X 2 − αX) is a reduced ring but not an integral domain. The ring R2 of Example 3.2 (1) is obtained in this manner, taking n = 2 and α = t3 , from the AGL ring R = k[[t3 , t4 , t5 ]]. We say that R has minimal multiplicity, if v(R) = e(R). When m contains a reduction (α), this condition is equivalent to saying that m2 = αm ([L, S2]). Proposition 3.11. Suppose that e(R) = 3 and R has minimal multiplicity. Then R is a 2-AGL ring if and only if ℓR (K/R) = 2. Proof. Thanks to Theorem 3.9, passing to R1 = R[X]mR[X] if necessary, we can assume that the residue class field R/m of R is infinite. Since v(R) = e(R) = 3, r(R) = 2. Therefore, by Corollary 3.7 we have only to show that S = K 2 , once ℓR (K/R) = 2. Choose a non-zerodivisor b of R so that J = bK ( R. Then, since R/m is infinite, J contains an element c such that J 3 = cJ 2 (see [S1, ES]; remember that µR (J 3 ) ≤ e(R) = 3). Hence K 2 = K 3 by Lemma 2.2 (1) and Theorem 2.5 (2).  We close this section with the following. Remark 3.12. Let r(R) = 2 and assume that R is a homomorphic image of a regular local ring T of dimension 3. If R is a 2-AGL ring, then R has a minimal T -free resolution SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS of the form " X 2 g1 Y g2 Z g3 13 # 0 → T ⊕2 −−−−−→ T ⊕3 → T → R → 0, where X, Y, Z is a regular system of parameters of T . In fact, let M 0 → T ⊕2 − → T ⊕3 → T → R → 0 be a minimal T -free resolution of R. Then, since K ∼ = Ext2T (R, T ), taking the T -dual of the resolution, we get a minimal T -free resolution tM τ →K→0 0 → T → T ⊕3 −→ T ⊕2 − of K. Because µR (K/R) = 1, without loss of generality, we may assume that τ (e2 ) = 1,   where e2 = ( 01 ). Therefore, writing t M = fg11 fg22 fg33 , we get K/R ∼ = T /(f1 , f2 , f3 ) as a T -module. Let C = K/R and q = (f1 , f2 , f3 ). Then since ℓT (T /q) = ℓR (C) = 2, after suitable elementary column-transformations in the matrix t M we get that f1 = X 2 , f2 = Y, f3 = Z for some regular system X, Y, Z of parameters of T . The converse of the assertion in Remark 3.12 is not true in general. In the case where R is the semigroup ring of a numerical semigroup, we shall give in Section 6 a complete   description of the assertion in terms of the matrix t M = fg11 fg22 fg33 (see Theorem 6.4 and its consequences). 4. 2-AGL rings obtained by idealization In this section we study the problem of when the idealization A = R ⋉ c of c = R : S over R is a 2-AGL ring. To do this we need some preliminaries. For a moment let R be an arbitrary commutative ring and M an R-module. Let A = R ⋉ M be the idealization of M over R. Hence A = R ⊕ M as an R-module and the multiplication in A is given by (a, x)(b, y) = (ab, bx + ay) where a, b ∈ R and x, y ∈ M. Let K be an R-module and set L = HomR (M, K) ⊕ K. We consider L to be an A-module under the following action of A (a, x) ◦ (f, y) = (af, f (x) + ay), where (a, x) ∈ A and (f, y) ∈ L. Then it is standard to check that the map HomR (A, K) → L, α 7→ (α ◦ j, α(1)) is an isomorphism of A-modules, where j : M → A, x 7→ (0, x) and 1 = (1, 0) denotes the identity of the ring A. We are now back to our Setting 2.1. Let A = R ⋉ c and set L = S × K. Then A is a one-dimensional Cohen-Macaulay local ring and KA = HomR (A, K) ∼ = L = S × K, = HomR (c, K) × K ∼ 14 TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA because c = K : S by Proposition 2.3 (1). Therefore A = R ⋉ c ⊆ L = S × K ⊆ R ⋉ Q(R). Because Q(A) = Q(R) ⋉ Q(R) and A = R ⋉ Q(R), our idealization A = R ⋉ c satisfies the same assumption as in Setting 2.1 and we have the following. Proposition 4.1. The following assertions hold true. (1) (2) (3) (4) Ln = S ⋉ S for all n ≥ 2, whence A[L] = S ⋉ S. ℓA (A[L]/L) = ℓR (S/K). A : A[L] = c × c. v(A) = v(R) + µR (c) and e(A) = 2·e(R). Proof. (1) Since Ln = (S × K)n = S n × S n−1 K, we have Ln = S × S for n ≥ 2. (2) We get ℓA (A[L]/L) = ℓR ((S ⊕ S)/(S ⊕ K)) = ℓR (S/K). (3) This is straightforward, since A[L] = S ⋉ S. (4) To see the first assertion, remember that m × c is the maximal ideal of A and that (m × c)2 = m2 × mc. For the second one, notice that mA is a reduction of m × c and that A = R ⊕ c as an R-module. We then have e(A) = e0m (A) = 2·e0m (R).  By Proposition 4.1 (2) we readily get the following. Theorem 4.2. A = R ⋉ c is a 2-AGL ring if and only if so is R. Example 4.3. Suppose that R is a 2-AGL ring, for instance, take R = k[[t3 , t7 , t8 ]] (see Example 3.2 (1)). We set ( R if n = 0, An = An−1 ⋉ cn−1 if n ≥ 1, that is A0 = R and for n ≥ 1 let An be the idealization of cn−1 over An−1 , where cn−1 = An−1 : An−1 [Kn−1 ] and Kn−1 denotes a fractional ideal of An−1 such that An−1 ⊆ Kn−1 ⊆ An−1 and Kn−1 ∼ = KAn−1 as an An−1 -module. We then have an infinite family {An }n≥0 of analytically ramified 2-AGL rings such that e(An ) = 2n ·e(R) for each n ≥ 0. Since c = t6 k[[t]] ∼ = k[[t]] for R = k[[t3 , t7 , t8 ]], the ring R3 = k[[t3 , t7 , t8 ]]⋉k[[t]] of Example 3.2 (1) is obtained in this manner. 5. The algebra m : m We maintain Setting 2.1 and set B = m : m. By [GMP, Theorem 5.1] B is a Gorenstein ring if and only if R is an AGL ring of minimal multiplicity. Our purpose of this section is to explore the structure of the algebra B = m : m in connection with the 2-AGL property of R. Let us begin with the following. SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS 15 Proposition 5.1. Suppose that there is an element α ∈ m such that m2 = αm and that R is not a Gorenstein ring. Set L = BK. Then the following assertions hold true. (1) B = R : m = mα and K : B = mK. (2) L = K : m, L ∼ = mK as a B-module, and B ⊆ L ⊆ B. (3) S = B[L] = B[K]. Proof. Since R is not a DVR (resp. m2 = αm), we have B = R : m (resp. B = mα ). Because K : mK = R : m = B, we get mK = K : B. We have K : L = R : B = m, ∼ mK as a B-module. We have since R ( B. Therefore, L = K : m. Clearly, L = mK α = 2 S ⊆ B[K] ⊆ B[L]. Because B ⊆ L = K : m ⊆ K by Corollary 2.4 (1), B[L] ⊆ S, whence S = B[L] = B[K] as claimed.  We have the following. Theorem 5.2. Suppose that R is a 2-AGL ring. Assume that there is an element α ∈ m such that m2 = αm. Set L = BK. Then the following assertions hold true. (1) ℓR (L2 /L) = 1. (2) Let M = (0) :B (L2 /L). Then M ∈ Max B, R/m ∼ = B/M, and BM is an AGL ring which is not a Gorenstein ring. (3) If N ∈ Max B \ {M}, then BN is a Gorenstein ring. Therefore, BN is an AGL ring for every N ∈ Max B. Proof. Because S = K 2 and S ⊇ L ⊇ K by Proposition 5.1, we have S = L2 , while ℓR (L2 /L) = 1 as L = K : m. Hence 0 < ℓB/M (L2 /L) = ℓB (L2 /L) ≤ ℓR (L2 /L) = 1, so that M ∈ Max B, R/m ∼ = B/M, and L2 /L ∼ = B/M. Because L ∼ = K : B as a B-module by Proposition 5.1, we get LM ∼ = KBM as a BM -module ([HK, Satz 5.12]). Therefore, since 2 ℓBM (LM /LM ) = 1, by [GMP, Theorem 3.16] BM is an AGL ring which is not a Gorenstein ring. If N ∈ Max B and if N 6= M, then (L2 /L)N = (0), so that BN is a Gorenstein ring by [GMP, Theorem 3.7].  Let us note a few consequences of Theorem 5.2. Corollary 5.3. Assume that m2 = αm for some α ∈ m and that B is a local ring with maximal ideal n. Then the following conditions are equivalent. (1) R is a 2-AGL ring. (2) B is a non-Gorenstein AGL ring and R/m ∼ = B/n. When this is the case, S is a Gorenstein ring, provided v(B) = e(B). 16 TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA Proof. By Theorem 5.2 we have only to show the implication (2) ⇒ (1). Let L = BK. Then L = K : m, L ∼ = KB , and S = B[L] by Proposition 5.1. Because B is a nonGorenstein AGL ring, ℓB (B[L]/L) = 1 by [GMP, Theorem 3.16], so that ℓR (S/K) = ℓR (S/L) + ℓR (L/K) = ℓB (B[L]/L) + ℓR ((K : m)/K) = 2, where the second equality follows from the fact that R/m ∼ = B/n. Hence R is a 2-AGL ring. The last assertion is a direct consequence of [GMP, Theorem 5.1].  If R is the semigroup ring of a numerical semigroup, the algebra B = m : m is also the semigroup ring of a numerical semigroup, so that B is always a local ring with R/m ∼ = B/n, where n denotes the maximal ideal of B. Hence by Corollary 5.3 we readily get the following. Let k[[t]] be the formal power series ring over a field k. Corollary 5.4. Let H = ha1 , a2 , . . . , aℓ i be a numerical semigroup and R = k[[ta1 , ta2 , . . . , taℓ ]] the semigroup ring of H. Assume that R has minimal multiplicity. Then R is a 2-AGL ring if and only if B = m : m is an AGL ring which is not a Gorenstein ring. When this is the case, S is a Gorenstein ring, provided v(B) = e(B). Proof. Remember that m2 = te m, where e = min{ai | 1 ≤ i ≤ ℓ}.  If v(R) < e(R), the ring S is not necessarily a Gorenstein ring, even though R is a 2-AGL ring and B is an AGL ring with v(B) = e(B) ≥ 3. Let us note one example. Example 5.5. Let R = k[[t5 , t7 , t9 , t13 ]] and set K = R+Rt3 . Then we have the following. (1) K ∼ = KR as an R-module and I = t12 K is a canonical ideal of R with (t12 ) a reduction. (2) (3) (3) (4) Hence r(R) = 2. S = k[[t3 , t5 , t7 ]] and c = (t10 ) + (t7 , t9 , t13 ). R is a 2-AGL ring with v(R) = 4 and e(R) = 5. K/R ∼ = R/c as an R-module. B = k[[t5 , t7 , t8 , t9 , t11 ]] and B is an AGL ring, possessing minimal multiplicity 5. The ring B does not necessarily have minimal multiplicity, even though B is a local ring and R is a 2-AGL ring of minimal multiplicity. Let us note one example. Example 5.6. Let R = k[[t4 , t9 , t11 , t14 ]] and set K = R + Rt3 + Rt5 . Then we have the following. (1) K ∼ = KR as an R-module and I = t11 K is a canonical ideal of R with (t11 ) a reduction. Hence r(R) = 3. (2) R is a 2-AGL ring with m2 = t4 m. (3) ℓR (K/R) = 3 and K/R ∼ = R/c ⊕ R/m as an R-module. 4 5 7 (4) B = k[[t , t , t ]]. SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS 17 In Theorem 5.2, if K/R is a free R/c-module, then B = m : m is necessarily a local ring. To state the result, we need further notation. Suppose that R is a 2-AGL ring and set r = r(R). Then since by Proposition 3.3 (4) K/R ∼ = (R/c)⊕ℓ ⊕ (R/m)⊕m with integers ℓ > 0 and m ≥ 0 such that ℓ + m = r − 1, there are elements f1 , f2 , . . . , fℓ and g1 , g2 , . . . , gm of K such that K/R = ℓ X i=1 R·fi m MX j=1 R·gj , ℓ X R·fi ∼ = (R/c)⊕ℓ , and i=1 m X R·gj ∼ = (R/m)⊕m j=1 P P where ∗ denotes the image in K/R. We set F = ℓi=1 Rfi and U = m j=1 Rgj . Let us 2 write c = (x1 ) + (x2 , x3 , . . . , xn ) for some minimal system {xi }1≤i≤n of generators of m (see Proposition 3.3 (2)). With this notation we have the following. Proposition 5.7. The following assertions hold true. (1) Suppose m = 0, that is K/R is a free R/c-module. Then B is a local ring with maximal ideal mS and R/m ∼ = B/mS. q (2) Suppose that U ⊆ mS for some q > 0. Then B is a local ring. Proof. We divide the proof of Proposition 5.7 into a few steps. Notice that B = R : m, since R is not a DVR. Claim 1. The following assertions hold true. (1) m2 S ⊆ R. (2) mS ⊆ J(B), where J(B) denotes the Jacobson radical of B. Proof. Because m2 K 3 = m2 K 2 ⊆ K, we have m2 K 2 ⊆ K : K = R. Hence m2 S ⊆ R, so that mS ⊆ R : m = B. Let M ∈ Max B and choose N ∈ Max S such that M = N ∩ B. Then because mS ⊆ N, mS ⊆ N ∩ B = M, whence mS ⊆ J(B).  We consider Assertion (2) of Proposition 5.7. Since gjq ∈ mS for all 1 ≤ j ≤ m, the ring B/mS = (R/m)[g1 , g2 , . . . , gm ] is a local ring, where gj denotes the image of gj in B/mS. Therefore B is a local ring, since mS ⊆ J(B) by Claim 1 (2). To prove Assertion (1) of Proposition 5.7, we need more results. Suppose that m = 0; hence U = (0) and ℓ = r − 1. Claim 2. The following assertions hold true. (1) x1 fi 6∈ R for all 1 ≤ i ≤ ℓ. (2) (R : m) ∩ K = R + x1 F and ℓR ([(R : m) ∩ K]/R) = ℓ. (3) B = R + x1 F + Rh for some h ∈ mK 2 . 18 TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA P Proof. (1) Remember that K/R = ℓi=1 R·fi ∼ = (R/c)⊕ℓ . We then have mfi 6= (0) but cfi = (0) for each 1 ≤ i ≤ ℓ. Hence x1 fi 6∈ R, because c = (x21 ) + (xj | 2 ≤ j ≤ n) and m = (x1 ) + c. (2) Because (0) :R/c m is generated by the image of x1 , we have (0) :K/R m = ℓ X R·x1 fi , i=1 whence (R : m) ∩ K = R + x1 F and ℓR ([(R : m) ∩ K]/R) = ℓ. (3) Notice that by Claim 1 (1) R ⊆ mK + R ⊆ (R : m) ∩ K ⊆ R : m and that ℓR ([(R : m) ∩ K]/R) = ℓ = r − 1 by Assertion (2). We then have ℓR ((R : m)/[(R : m) ∩ K]) = 1, because ℓR ((R : m)/R) = r. Hence R : m = [(R : m) ∩ K] + Rg for some g ∈ R : m. On the other hand, because K ( mK 2 + K ( K 2 by Corollary 2.4 (1) and ℓR (K 2 /K) = 2, we get ℓR (K 2 /(mK 2 + K)) = ℓR ((mK 2 + K)/K) = 1. Consequently, K 2 = K + Rξ for some ξ ∈ K 2 . Let us write g = ρ + βξ with ρ ∈ K and β ∈ m. Then since βξ ∈ mK 2 ⊆ R : m by Claim 1 (1) and g ∈ R : m, we get ρ ∈ (R : m) ∩ K, whence setting h = βξ, we have B = R : m = [(R : m) ∩ K] + Rg = [(R : m) ∩ K] + Rh as claimed.  Let us finish the proof of Assertion (1) of Proposition 5.7. In fact, by Claim 2 (3) we have B ⊆ R + mS, which implies R/m ∼ = B/mS, whence mS ∈ Max B. Therefore, B is a local ring with unique maximal ideal mS, since mS ⊆ J(B) by Claim 1 (2). This completes the proof of Proposition 5.7.  Under some additional conditions, the converse of Theorem 5.2 is also true. Theorem 5.8. Suppose that the following conditions are satisfied. (1) e(R) = e ≥ 3 and R is not an AGL ring, (2) B is an AGL ring with e(B) = e, and (3) there is an element α ∈ m such that m2 = αm and n2 = αn, where n denotes the maximal ideal of B. Then R is a 2-AGL ring and K/R is a free R/c-module. Proof. We have B = R : m = ma . Let L = BK. Then by Proposition 5.1 (3) S = B[L] = B[K]. As B is an AGL ring, we have S = n : n by [GMP, Theorem 3.16], whence S = αn . Consequently, R ⊆ B = mα ⊆ S = an . Let us write m = (α, x2 , x3 , . . . , xe ). We set yi = xαi for each 2 ≤ i ≤ e. Claim 3. We can choose the elements {xi }2≤i≤e of m so that yi ∈ n for all 2 ≤ i ≤ e. SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS 19 Proof. Since by Conditions (1) and (2) ℓB (B/αB) = e(B) = e = e(R) = ℓR (R/αR) = ℓR (B/αB), we get the isomorphism R/m ∼ = B/n of fields. Let 2 ≤ i ≤ e be an integer and choose i ∈ n, replacing xi with xi − αci , ci ∈ R so that yi ≡ ci mod n. Then since yi − ci = xi −αc α we have yi ∈ n for all 2 ≤ i ≤ e.  We now notice that B = have m α = R+ e X Pe i=2 R· xαi and xi α ∈ n for each 2 ≤ i ≤ e. We then e e X xi X xi xi R· = Rα + R· , = m+ α α α i=2 i=2 i=2 P where the last equality follows from the fact that m = Rα + ei=2 Rxi . Thus n = (n ∩ R) + R· e X xi n R· 2 , S = =R+ α α i=2 x x whence α2 ∈ c. Let 2 ≤ i, j ≤ e be integers. Then xi · αj2 = xαi · αj ∈ n2 = αn and P P x n = Rα+ ei=2 R· xαi , so that xi · αj2 ∈ Rα2 + ei=2 Rxi , which shows (α2 , x2 , x3 , . . . , xe ) ⊆ c. Therefore, c = (α2 , x2 , x3 , . . . , xe ) because c ( m (remember that R is not an AGL ring), whence ℓR (R/c) = 2, so that R is a 2-AGL ring. Because S = αn and B = mα and because R/m ∼ = B/n, we get ℓR (S/B) = ℓR (S/n) − ℓR (B/n) = ℓR (S/αS) − 1 = e − 1 and ℓR (B/R) = ℓR (B/m) − ℓR (R/m) = ℓR (B/αB) − 1 = e − 1. Therefore, ℓR (S/R) = 2e − 2, whence ℓR (K/R) = 2e − 4 = 2(e − 2) because ℓR (S/K) = 2. Consequently, by Proposition 3.3 (4) K/R is a free R/c-module, since µR (K/R) = e − 2 (notice that r(R) = e − 1), which completes the proof of Theorem 5.8.  However, the ring B is not necessarily a local ring in general, although R is a 2-AGL ring with v(R) = e(R). Let us note one example. Example 5.9. Let V = k[[t]] be the formal power series ring over an infinite field k. We consider the direct product A = k[[t3 , t7 , t8 ]] × k[[t3 , t4 , t5 ]] of rings and set R = k·(1, 1) + J(A) where J(A) denotes the Jacobson radical of A. Then R is a subring of A and a one-dimensional Cohen-Macaulay complete local ring with J(A) the maximal ideal. We have the ring R is a 2-AGL ring and v(R) = e(R) = 6. However m : m = k[[t3 , t4 , t5 ]] × V which is not a local ring, so that K/R is not a free R/c-module. 20 TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA 6. Numerical semigroup rings Let k be a field. In this section we study the case where R = k[[H]] is the semigroup ring of a numerical semigroup H. First of all, let us fix the notation, according to the terminology of numerical semigroups. Setting 6.1. Let 0 < a1 , a2 , . . . , aℓ ∈ Z (ℓ > 0) be positive integers such that GCD (a1 , a2 , . . . , aℓ ) = 1. We set ( ℓ ) X H = ha1 , a2 , . . . , aℓ i = ci ai | 0 ≤ ci ∈ Z for all 1 ≤ i ≤ ℓ i=1 and call it the numerical semigroup generated by the numbers {ai }1≤i≤ℓ . Let V = k[[t]] be the formal power series ring over a field k. We set R = k[[H]] = k[[ta1 , ta2 , . . . , taℓ ]] in V and call it the semigroup ring of H over k. The ring R is a one-dimensional CohenMacaulay local domain with R = V and m = (ta1 , ta2 , . . . , taℓ ). We set T = k[ta1 , ta2 , . . . , taℓ ] in k[t]. Let P = k[X1 , X2 , . . . , Xℓ ] be the polynomial ring over k. We consider P to be a Z-graded ring such that P0 = k and deg Xi = ai for 1 ≤ i ≤ ℓ. Let ϕ : P → T denote the homomorphism of graded k-algebras defined by ϕ(Xi ) = tai for each 1 ≤ i ≤ ℓ. In this section we are interested in the question of when R = k[[H]] is a 2-AGL ring. To study the question, we recall some basic notion on numerical semigroups. Let c(H) = min{n ∈ Z | m ∈ H for all m ∈ Z such that m ≥ n} be the conductor of H and set f(H) = c(H) − 1. Hence f(H) = max (Z \ H), which is called the Frobenius number of H. Let PF(H) = {n ∈ Z \ H | n + ai ∈ H for all 1 ≤ i ≤ ℓ} denote the set of pseudo-Frobenius numbers of H. Therefore, f(H) equals the a-invariant of the graded k-algebra k[ta1 , ta2 , . . . , taℓ ] and ♯PF(H) = r(R) ([GW, Example (2.1.9), Definition (3.1.4)]). We set f = f(H) and X K= Rtf −c c∈PF(H) in V . Then K is a fractional ideal of R such that R ⊆ K ⊆ R and X Rt−c K∼ = KR = c∈PF(H) as an R-module ([GW, Example (2.1.9)]). Let us refer to K as the fractional canonical ideal of R. Notice that tf 6∈ K but mtf ⊆ R, whence K : m = K + Rtf . Let us begin with the following. SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS 21 D E ∨ Theorem 6.2. Suppose that ℓ ≥ 3 and aj 6∈ a1 , . . . , aj , . . . , aℓ for all 1 ≤ j ≤ ℓ. Assume that r(R) = 2 and let K = R + Rta for some 0 < a ∈ Z. Then the following conditions are equivalent. (1) R is a 2-AGL ring. (2) 3a ∈ H and f = 2a + ai for some 1 ≤ i ≤ ℓ. Proof. (1) ⇒ (2) We have t3a ∈ K 2 as K 2 = K 3 . If 2a ∈ H, then K 2 = K and R is a Gorenstein ring (Proposition 2.3 (3)). Hence 2a 6∈ H, so that 3a ∈ H. Because P K : m = mK 2 + K by Corollary 2.4 (2), we get K + Rtf = K : m = K + ℓj=1 Rt2a+aj . Therefore, because ℓR ((K : m)/K) = 1, K + Rtf = K + Rt2a+ai for some 1 ≤ i ≤ ℓ, whence f = 2a + ai . (2) ⇒ (1) We get K 3 = K 2 = K + Rt2a , since 3a ∈ H. Let L = mK 2 + K. Hence P L = K + ℓj=1 Rt2a+aj and L ( K 2 because R is not a Gorenstein ring. Notice that ℓR (K 2 /L) = 1, since µR (K 2 /K) = 1. We furthermore have the following. Claim 4. L = K : m. Proof of Claim 4. We have only to show L ⊆ K : m = K + Rt2a+ai . Let 1 ≤ j ≤ ℓ and assume that t2a+aj 6∈ K + Rt2a+ai . Then aj < ai since f = 2a + ai = c(H) − 1. Because 2a + aj 6∈ H, tf −(2a+aj ) = tai −aj ∈ K = R + Rta . Hence ai − aj ∈ H or ai − aj − a ∈ H. Suppose that ai − aj − a ∈ H. Then, setting h = ai − aj − a ∈ H, we get ai = aj + a + h whence f = 2a + ai = 3a + aj + h ∈ H, which is impossible. Hence ai − aj ∈ H. Let us write ℓ X ai − aj = mk ak k=1 with 0 ≤ mk ∈ Z. Then mk = 0 if ak ≥ ai , since ai − aj < ai . Therefore D E X ∨ ai = aj + mk ak ∈ a1 , . . . , ai , . . . , aℓ ak <ai which contradicts the assumption that aj 6∈ 2a+ai L = K + Rt . D ∨ a1 , . . . , aj , . . . , aℓ E for all 1 ≤ j ≤ ℓ. Thus  We have ℓR (K 2 /K) = ℓR (K 2 /L) + ℓR (L/K) = 2 because ℓR (L/K) = 1 by Claim 4, so that R is a 2-AGL ring.  Let us recover Example 3.2 (2) in the present context. Corollary 6.3. Let c ≥ 4 such that c 6≡ 0 mod 3 and set H = h3, c + 3, 2ci. Then R is a 2-AGL ring such that r(R) = 2 and K/R ∼ = R/c as an R-module. Proof. We set a = c − 3. Then f = 2a + 3 and K = R + Rta .  22 TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA Suppose that ℓ = 3. We set a = Ker ϕ, where ϕ : P = k[X1 , X2 , X3 ] → T = k[ta1 , ta2 , ta3 ] is the homomorphism of k-algebras defined by ϕ(Xi ) = tai for i = 1, 2, 3. Let us write X = X1 , Y = X2 , and Z = X3 for short. If T is not a Gorenstein ring, then by  Xα Y β Zγ [H] it is known that a = I2 Y β ′ Z γ′ X α′ for some integers α, β, γ, α′, β ′, γ ′ > 0, where  α β γ  I2 Xβ ′ Y γ′ Zα′ denotes the ideal of P generated by the 2 × 2 minors of the matrix  αY βZ γX X Y Z . With this notation we have the following. ′ ′ ′ Y β Zγ Xα Theorem 6.4. Suppose that H is 3-generated, that is ℓ = 3. Then the following conditions are equivalent. (1) R is a 2-AGL ring.  2 (2) After a suitable permutation of a1 , a2 , a3 , a = I2 YXβ ′ α′ , β ′ , γ ′ such that α′ ≥ 2 and β ′ , γ ′ > 0. Y Z ′ ′ Zγ Xα  for some integers To prove Theorem 6.4, we need a result of [GMP, Section 4]. Throughout, let H = ha1 , a2 , a3 i and assume that T is not a Gorenstein ring. Hence the ideal a is generated by the 2 × 2 minors of the matrix  α  X Y β Zγ ′ ′ ′ Y β Zγ Xα ′ ′ ′ ′ where 0 < α, β, γ, α′, β ′, γ ′ ∈ Z. Let ∆1 = Z γ+γ − X α Y β , ∆2 = X α+α − Y β Z γ , and ′ ′ ∆3 = Y β+β − X α Z γ . Then a = (∆1 , ∆2 , ∆3 ) and thanks to the theorem of Hilbert– Burch ([E, Theorem 20.15]), the graded ring T = P/a possesses a graded minimal P -free resolution of the form 0 −→ P (−m) ⊕ P (−n) ′  Xα Y β  Y β Z γ′  ′ Zγ Xα  −→ P (−d1 ) ⊕ [∆1 ∆2 ∆3 ] ε P (−d2 ) −→ P −→ T −→ 0, ⊕ P (−d3 ) where d1 = deg ∆1 = a3 (γ + γ ′ ), d2 = deg ∆2 = a1 (α + α′ ), d3 = deg ∆3 = a2 (β + β ′), m = a1 α + d1 = a2 β + d2 = a3 γ + d3 , and n = a1 α′ + d3 = a2 β ′ + d1 = a3 γ ′ + d2 . Therefore (E) n − m = a2 β ′ − a1 α = a3 γ ′ − a2 β = a1 α′ − a3 γ. Let KP = P (−d) denote the graded canonical module of P where d = a1 + a2 + a3 . Then, taking the KP –dual of the above resolution, we get the minimal presentation (♯) P (d1 − d)   Xα Y β Zγ P (m − d) ⊕ ′ ′ ′ ε Y β Zγ Xα ⊕ P (d2 − d) −→ KT −→ 0 −→ P (n − d) ⊕ P (d3 − d) SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS 23 of the graded canonical module KT = Ext2P (T, KP ) of T . Therefore, because KT = P −c ([GW, Example (2.1.9)]), we have ℓk ([KT ]i ) ≤ 1 for all i ∈ Z, whence c∈PF(H) T t m 6= n. After the permutation of a2 and a3 if necessary, we may assume without loss of generality that m < n. Then the presentation (♯) shows that PF(H) = {m − d, n − d} and f = n − d. We set a = n − m. Hence a > 0, f = a + (m − d), and K = R + Rta . With this notation we have the following. Remember that R is the MTM -adic completion of the local ring TM , where M = (tai | i = 1, 2, 3) denotes the graded maximal ideal of T . Proposition 6.5 ([GMP, Theorem 4.1]). ℓR (K/R) = αβγ. Therefore, if R is a 2-AGL ring, then ℓR (K/R) = 2 by Proposition 3.7, so that α = 2 and β = γ = 1 by Proposition 6.5 after a suitable permutation of a1 , a2 , a3 if necessary. Consequently, Theorem 6.4 is reduced to the following.  2  X Y Z Theorem 6.6. Let m < n and assume that a = I2 Y β ′ Z γ′ X α′ with α′ , β ′, γ ′ > 0. Then R is a 2-AGL ring if and only if α′ ≥ 2. When this is the case, f = 2a + a1 , where a=n−m Proof of Theorem 6.6. Notice that R is not an AGL ring, since ℓR (K/R) = 2 by Proposition 6.5. We get by equations (E) above (i) a2 β ′ = 2a1 + a, (ii) a3 γ ′ = a2 + a, and (iii) a1 α′ = a3 + a. Suppose that α′ ≥ 2. Then 3a = a2 (β ′ − 1) + a1 (α′ − 2) + a3 (γ ′ − 1) ∈ H. Hence S = K 2 = R + Rta + Rt2a . Therefore, because (iv) 2a1 + 2a = (2a1 + a) + a = a2 β ′ + (a3 γ ′ − a2 ) = a2 (γ ′ − 1) + a3 γ ′ ∈ H, (v) a2 + 2a = (a3 γ ′ − a2 ) + (a1 α′ − a3 ) + a2 = a3 (γ ′ − 1) + a1 α′ ∈ H, and (vi) a3 + 2a = (a1 α′ − a3 ) + (a2 β ′ − 2a1 ) + a3 = a1 (α′ − 2) + a2 β ′ ∈ H, we get that (t2a1 )+(ta2 , ta3 ) ⊆ K : S = c by Proposition 2.3 (1) and that (2a+a1 )+ai ∈ H for i = 1, 2, 3. Hence 2a + a1 ∈ PF(H) if 2a + a1 6∈ H. Now notice that mK 2 + K = K + Rt2a+a1 because 2a + ai ∈ H for i = 2, 3 by equations (v) and (vi), whence t2a+a1 6∈ K because mK 2 6⊆ K by Proposition 2.3 (4). In particular, 2a + a1 6∈ H. Therefore, ta1 6∈ c, so that c 6= m and c = (t2a1 ) + (ta2 , ta3 ). Thus R is a 2-AGL ring, because ℓR (R/c) = 2. Notice that 2a + a1 ∈ PF(H) = {f − a, f } and we get f = 3a + a1 ∈ H if f 6= 2a + a1 , which is impossible as 3a ∈ H. Hence f = 2a + a1 . Conversely, assume that R is a 2-AGL ring. Then 2a 6∈ H, since K 6= K 2 . Therefore P 3a ∈ H, since t3a ∈ K 2 . Because mK 2 + K = K + 3j=1 Rt2a+aj and a2 + 2a ∈ H by equation (v), we get
K + Rtf = K : m = mK 2 + K = K + Rt2a+a1 + Rt2a+a3 , 24 TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA where the second equality follows from Corollary 2.4 (2). Therefore, if t2a+a3 6∈ K, then f = 2a + a3 , so that PF(H) = {a + a3 , 2a + a3 }, which is absurd because a + a3 ∈ H by equation (iii). Thus t2a+a3 ∈ K, so that mK 2 +K = K +Rt2a+a1 and f = 2a+a1 . Suppose now that α′ = 1. Then a1 = a + a3 by equation (iii), whence f = 2a + a1 = 3a + a3 ∈ H because 3a ∈ H. This is a required contradiction, which completes the proof of Theorem 6.4 as well as that of Theorem 6.6.  When H is 3-generated and e(R) = min{a1 , a2 , a3 } is small, we have the following structure theorem of H for R to be a 2-AGL ring. Corollary 6.7. Let ℓ = 3. (1) Suppose that min{a1 , a2 , a3 } = 3. Then the following conditions are equivalent. (a) R is a 2-AGL ring. (b) H = h3, c + 3, 2ci for some c ≥ 4 such that c 6≡ 0 mod 3. (2) If min{a1 , a2 , a3 } = 4, then R is not a 2-AGL ring. (3) Suppose that min{a1 , a2 , a3 } = 5. Then the following conditions are equivalent. (a) R is a 2-AGL ring. (b) (i) H = h5, 3c + 8, 2c + 2i for some c ≥ 2 such that c 6≡ 4 mod 5 or (ii) H = h5, c + 4, 3c + 2i for some c ≥ 2 such that c 6≡ 1 mod 5. Proof. Let e = min{a1 , a2 , a3 }. Suppose that R is a 2-AGL ring. Then by Theorem 6.4, after a suitable permutation of a1 , a2 , a3 we get  2  a = I2 YXβ ′ ZYγ′ XZα′ for some integers α′ , β ′ , γ ′ such that α′ ≥ 2 and β ′ , γ ′ > 0. Remember that a1 = β ′ γ ′ + β ′ + 1, ′ ′ ′ because a1 = ℓR (R/ta1 R) = ℓk (k[Y, Z]/(Y β +1 , Y β Z, Z γ +1 ). We similarly have that a2 = 2γ ′ + α′ γ ′ + 2 ≥ 6, a3 = α′ β ′ + α′ + 2 ≥ 6 since α′ ≥ 2. Therefore, e = a1 = β ′ γ ′ + β ′ + 1, if e ≤ 5. (1) (a) ⇒ (b) We have β ′ = γ ′ = 1. Hence a2 = α′ + 4 and a3 = 2α′ + 2, that is H = h3, c + 3, 2ci, where c = α′ +1. We have c 6≡ 0 mod 3 because GCD (3, c+3, 2c) = 1, whence c ≥ 4.
2 (b) ⇒ (a) See Corollary 6.3 or notice that a = I2 XY YZ XZc−1 and apply Theorem 6.4 (2) We have a1 = β ′ γ ′ + β ′ + 1 = 4, so that β ′ = 1 and γ ′ = 2. Hence a2 = 2α′ + 6 and a3 = 2α′ + 2, which is impossible because GCD (a1 , a2 , a3 ) = 1. (3) (a) ⇒ (b) We set c = α′ . Then either β ′ = 1 and γ ′ = 3 or β ′ = 2 and γ ′ = 1. For the former case we get (i) H = h5, 3c + 8, 2c + 2i, where c 6≡ 4 mod 5. For the latter case we get (ii) H = h5, c + 4, 3c + 2i, where c 6≡ 1 mod 5.

2 2 Y Z (b) ⇒ (a) For Case (i) we have a = I2 XY ZY3 XZc and for Case (ii) a = I2 X , 2 c Y Z X whence R is a 2-AGL ring by Theorem 6.4.  SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS 25 Even though R is a 2-AGL ring, K/R is not necessarily a free R/c-module (Example 3.5). Here let us note a criterion for the freeness of K/R. Proposition 6.8. Let r = r(R) ≥ 2 and write PF(H) = {c1 < c2 < · · · < cr = f }. Assume that R is a 2-AGL ring. Then the following conditions are equivalent. (1) K/R ∼ = (R/c)⊕(r−1) as an R-module. (2) There is an integer 1 ≤ j ≤ ℓ such that f + aj = ci + cr−i for all 1 ≤ i ≤ r − 1. P Proof. (2) ⇒ (1) We have K = ri=1 Rtf −ci and ℓR (K/(mK + R)) = µR (K/R) = r − 1. Pr−1 ci Rt ⊆ mK + R, whence Because tcr−i = tf −ci +aj ∈ mK + R for all 1 ≤ i < r, R + i=1 ℓR (K/R) ≥ ℓR (K/(mK + R)) + ℓR ((R + r−1 X Rtci )/R) = 2(r − 1). i=1 Thus K/R is a free R/c-module by Proposition 3.3 (4). D E ∨ (1) ⇒ (2) We may assume that aj 6∈ a1 , . . . , aj , . . . , aℓ for all 1 ≤ j ≤ ℓ. Hence m is minimally generated by the elements {tai }1≤i≤ℓ . Therefore, since ℓR (R/c) = 2, by Proposition 3.3 (2) c = (t2aj ) + (tai | 1 ≤ i ≤ ℓ, i 6= j) for some 1 ≤ j ≤ ℓ. Because K/R is minimally generated by {tf −ci }1≤i≤r−1 where tf −ci denotes the image of tf −ci in K/R and because K/R is a free R/c-module, the homomorphism (♯) ϕ : (R/c)⊕(r−1) → K/R, ei 7→ tf −ci for each 1 ≤ i ≤ r − 1 of R/c-modules has to be an isomorphism, where {ei }1≤i≤r−1 denotes the standard basis of (R/c)⊕(r−1) . Now remember that taj 6∈ c, which shows via the isomorphism (♯) above that taj ·tf −ci 6∈ R for all 1 ≤ i ≤ r − 1, while we have t2aj ·tf −ci ∈ R and tak ·tf −ci ∈ R for all k 6= j. Therefore, f − ci + aj 6∈ H but (f − ci + aj ) + am ∈ H for all 1 ≤ m ≤ ℓ, so that f − ci + aj ∈ PF(H) for all 1 ≤ i ≤ r − 1. Notice that f − c1 + aj ≤ f because f − c1 + aj 6∈ H and that f − c1 + aj < f because c1 6= aj . Therefore, because {f − cr−1 + aj < · · · < f − c2 + aj < f − c1 + aj } ⊆ PF(H) = {c1 < · · · < cr−1 < cr = f } and f − c1 + aj < f , we readily get that f + aj = ci + cr−i for all 1 ≤ i ≤ r − 1.  We close this paper with a broad method of constructing 2-AGL rings from a given symmetric numerical semigroup. Let H be a numerical semigroup and assume that H is symmetric, that is R = k[[H]] is a Gorenstein ring. For the rest of this paper we fix an arbitrary integer 0 < e ∈ H. Let αi = min{h ∈ H | h ≡ i mod e} for each 0 ≤ i ≤ e − 1. We set Ape (H) = {αi | 0 ≤ i ≤ e − 1}. We then have Ape (H) = {h ∈ H | h − e 6∈ H}, which is called the Apery set of H mod e. Let us write Ape (H) = {h0 = 0 < h1 < · · · < he−1 }. We then have he−1 = hi + he−1−i for all 0 ≤ i ≤ e − 1, because H is symmetric. Notice that H = he, h1 , . . . , he−2 i and he−1 = h1 + he−2 . Let n ≥ 0 be an integer and set Hn = he, h1 + ne, h2 + ne, . . . , he−1 + nei . 26 TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA Notice that H0 = H and for each n > 0, e < h1 + ne < · · · < he−2 + ne < he−1 + ne and GCD (e, h1 + ne, . . . , he−2 + ne, he−1 + ne) = 1. We set Rn = k[[Hn ]], Sn = Rn [Kn ], and cn = Rn : Sn , where Kn denotes the fractional canonical ideal of Rn . Let mn = (te ) + (thi +ne | 1 ≤ i ≤ e − 1) be the maximal ideal of Rn . With this notation we have the following. Theorem 6.9. For all n ≥ 0 the following assertions hold true. (1) Kn2 = Kn3 . (2) ℓRn (Kn2 /Kn ) = n. (3) Kn /Rn ∼ = (Rn /cn )⊕(e−2) as an Rn -module. Hence R2 is a 2-AGL ring for any choice of the integer 0 < e ∈ H. Proof. We may assume n > 0. We begin with the following. Claim 5. The following assertions hold true. (1) h + ne ∈ Hn for all h ∈ H. (2) v(Rn ) = e(Rn ) = e. Proof of Claim 6.9. (1) Let h = hi + qe with 0 ≤ i ≤ e − 1 and q ≥ 0. Then h + ne = (hi + ne) + qe ∈ Hn . (2) Let 1 ≤ i, j ≤ e−1. Then (hi + ne) + (hj + ne) −e = [(hi + hj ) + ne] + (n−1)e ∈ Hn  by Assertion (1). Therefore, m2n = te mn . Consequently, by Claim 5 (2) we get that {e} ∪ {hi + ne}1≤i≤e−1 is a minimal system of generators of Hn , whence PF(Hn ) = {h1 + (n − 1)e, h2 + (n − 1)e, . . . , he−1 + (n − 1)e}. Pe−2 Rn thj , so that Sn = Rn [Kn ] = R. Therefore, Kn = j=0 Let 0 ≤ i, j ≤ e − 1 and write hi + hj = hk + qe with 0 ≤ k ≤ e − 1 and q ≥ 0. If k ≤ e − 2, then thi thj = (te )q thk ∈ Kn , which shows Kn2 = Kn + Rn the−1 (remember that Pe−2 Rn ·the−1 +hi . If 1 ≤ i ≤ e−2 and the−1 +hi 6∈ Kn , he−1 = h1 + he−2 ). Hence Kn3 = Kn2 + i=1 then the−1 +hi ∈ Rn the−1 ⊆ Kn2 as we have shown above. Hence Kn2 = Kn3 , which proves Assertion (1) of Theorem 6.9. Pe−1 Rn thj . Now notice that by Because Sn = R, we have mn R = te R, so that R = j=0 Claim 5 (1) (hi + ne) + hj = (hi + hj ) + ne ∈ Hn SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS 27 for all 0 ≤ i, j ≤ e−1 and we get thi +ne ∈ cn , whence tne Rn +(thi +ne | 1 ≤ i ≤ e−1)Rn ⊆ cn , while (n − 1)e + hj 6∈ Hn for all 1 ≤ j ≤ e − 1, so that t(n−1)e 6∈ cn . Thus cn = tne Rn + (thi +ne | 1 ≤ i ≤ e − 1)Rn = (thi +ne | 0 ≤ i ≤ e − 1)Rn and hence ℓRn (Rn /cn ) = n. Therefore, ℓRn (Kn /Rn ) = n by Proposition 2.3 (1), which proves Assertion (2) of Theorem 6.9. To prove Assertion (3) of Theorem 6.9, it suffices by Assertion (2) that ℓRn (Kn /Rn ) = n(e − 2), because cn = Rn : Kn by Proposition 2.3 (2) and µRn (Kn /Rn ) = e − 2 (notice that r(Rn ) = e − 1 by Claim 5 (2)). We set Lq = mqn Kn + Rn . We then have by Pe−2 induction on q that Lq = Rn + j=1 Rn thj +qe for all 0 ≤ q ≤ n. In fact, let 0 ≤ q < n and assume thath our assertion holds true for q. Then since Lq+1 = mn Lq + Rn , we get i Pe−2 Pe−2 hj +qe Rn thj +(q+1)e , because for . Therefore, Lq+1 = Rn + j=1 Lq+1 = Rn + mn j=1 Rn t all 1 ≤ i ≤ e − 1 and 1 ≤ j ≤ e − 2 (hi + ne) + (hj + qe) = [(hi + hj ) + ne] + qe ∈ Hn by Claim 5 (1). Hence we obtain a filtration Kn = L0 ⊇ L1 ⊇ · · · ⊇ Ln = Rn , Pe−2 where Lq = Lq+1 + j=1 Rn thj +qe and mn ·(Lq /Lq+1 ) = (0) for 0 ≤ q < n. Consequently, to see that ℓRn (Kn /Rn ) = n(e − 2), it is enough to show the following. Claim 6. ℓk (Lq /Lq+1 ) = e − 2 for all 0 ≤ q < n. Proof of Claim 6. Let 0 ≤ q < n and let {cj }1≤j≤e−2 be elements of the field k such that Pe−2 hj +qe ∈ Lq+1 . Suppose cj 6= 0 for some 1 ≤ j ≤ e − 2. Then thj +qe ∈ Lq+1 . Hence j=1 cj t hj + qe ∈ Hn or (hj + qe) − (hm + (q + 1)e) ∈ Hn for some 1 ≤ m ≤ e − 2. We get hj + qe 6∈ Hn , since hj + (n − 1)e 6∈ Hn . On the other hand, if (hj + ne) − (hm + ne + e) = (hj + qe) − (hm + (q + 1)e) ∈ Hn , then 1 ≤ m < j ≤ e − 2. Let us write (hj + ne) − (hm + ne + e) = α0 e + α1 (h1 + ne) + · · · + αe−1 (he−1 + ne) with integers 0 ≤ αp ∈ Z. Then αj = 0 since (hj + ne) − (hm + ne + e) < hj + ne, so that ∨ hj + ne = (α0 + 1)e + α1 (h1 + ne) + · · · + αj (hj + ne) + · · · + · · · + αe−1 (he−1 + ne), which violates the fact that {e} ∪ {hi + ne}1≤i≤e−1 is a minimal system of generators of Hn . Thus cj = 0 for all 1 ≤ j ≤ e − 2, whence ℓk (Lq /Lq+1 ) = e − 2 as claimed.  Therefore, ℓRn (Kn /Rn ) = n(e − 2), so that Kn /Rn ∼ = (Rn /cn )⊕(e−2) as an Rn -module, which completes the proof of Theorem 6.9.  28 TRAN DO MINH CHAU, SHIRO GOTO, SHINYA KUMASHIRO, AND NAOYUKI MATSUOKA References [BF] [E] [El] [ES] [GGHV] [GMP] [GNO] [GMTY1] [GMTY2] [GMTY3] [GMTY4] [GRTT] [GTT] [GW] [H] [HHS] [HK] [L] [S1] [S2] [S3] [T] [V1] [V2] V. Barucci and R. Fröberg, One-dimensional almost Gorenstein rings, J. Algebra, 188 (1997), 418–442. D. Eisenbud, Commutative Algebra with a View Toward Algebraic Geometry, GTM, 150, Springer-Verlag, 1994. J. Elias, On the canonical ideals of one-dimensional Cohen-Macaulay local rings, Proc. Edinburgh Math. Soc., 59 (2016), 77-90. P. Eakin and A. Sathaye, Prestable ideals, J. Algebra, 41 (1976), 439–454. L. Ghezzi, S. Goto, J. Hong, and W. V. Vasconcelos, Invariants of Cohen-Macaulay rings associated to their canonical ideals, arXiv:1701.05592v1. S. Goto, N. Matsuoka, and T. T. Phuong, Almost Gorenstein rings, J. Algebra, 379 (2013), 355-381. S. Goto, K. Nishida, and K. Ozeki, Sally modules of rank one, Michigan Math. J., 57 (2008), 359–381. S. Goto, N. Matsuoka, N. Taniguchi, and K.-i. Yoshida, The almost Gorenstein Rees algebras of parameters, J. Algebra, 452 (2016), 263–278. S. Goto, N. Matsuoka, N. Taniguchi, and K.-i. Yoshida, The almost Gorenstein Rees algebras over two-dimensional regular local rings, J. Pure Appl. Algebra, 220 (2016), 3425– 3436. S. Goto, N. Matsuoka, N. Taniguchi, and K.-i. Yoshida, On the almost Gorenstein property in Rees algebras of contracted ideals, arXiv:1604.04747v2. S. Goto, N. Matsuoka, N. Taniguchi, and K.-i. Yoshida, The almost Gorenstein Rees algebras of pg -ideals, good ideals, and powers of the maximal ideals, arXiv:1607.05894v2. S. Goto, M. Rahimi, N. Taniguchi, and H. L. Truong, When are the Rees algebras of parameter ideals almost Gorenstein graded rings?, Kyoto J. Math. (to appear). S. Goto, R. Takahashi and N. Taniguchi, Almost Gorenstein rings -towards a theory of higher dimension, J. Pure Appl. Algebra, 219 (2015), 2666–2712. S. Goto and K. Watanabe, On graded rings I, J. Math. Soc. Japan, 30 (1978), no. 2, 179–213. J. Herzog, Generators and relations of Abelian semigroups and semigroup rings, Manuscripta Math., 3 (1970), 175–193. J. Herzog, T. Hibi, and D. I. Stamate, The trace ideal of the canonical module, Preprint 2017, arXiv:1612.02723v1. J. Herzog and E. Kunz, Dear kanonische Modul eines-Cohen-Macaulay-Rings, Lecture Notes in Mathematics, 238, Springer-Verlag, 1971. J. Lipman, Stable ideals and Arf rings, Amer. J. Math., 93 (1971), 649–685. J. Sally, Numbers of generators of ideals in local rings, Lecture Notes in Pure and Applied Mathematics, 35 , Marcel Dekker, INC., 1978. J. Sally, Cohen-Macaulay local rings of maximal embedding dimension, J. Algebra, 56 (1979), 168–183. J. Sally, Hilbert coefficients and reduction number 2, J. Algebraic Geom., 1 (1992), 325–333. N. Taniguchi, On the almost Gorenstein property of determinantal rings, arXiv:1701.06690v1. W. V. Vasconcelos, Hilbert functions, analytic spread, and Koszul homology, Commutative algebra: Syzygies, multiplicities, and birational algebra (South Hadley, 1992), Contemp. Math., 159, 410–422, Amer. Math. Soc., Providence, RI, 1994. V. W. Vasconcelos, The Sally modules of ideals: a survey, Preprint 2017, arXiv:1612.06261v1. SALLY MODULES OF CANONICAL IDEALS IN DIMENSION ONE AND 2-AGL RINGS 29 Thai Nguyen University of Education, Khoa Toan, truong DHSP Thai Nguyen E-mail address: trandominhchau@gmail.com Department of Mathematics, School of Science and Technology, Meiji University, 1-1-1 Higashi-mita, Tama-ku, Kawasaki 214-8571, Japan E-mail address: shirogoto@gmail.com Department of Mathematics and Informatics, Graduate School of Science and Technology, Chiba University, Chiba-shi 263, Japan E-mail address: polar1412@gmail.com Department of Mathematics, School of Science and Technology, Meiji University, 1-1-1 Higashi-mita, Tama-ku, Kawasaki 214-8571, Japan E-mail address: naomatsu@meiji.ac.jp 

arXiv:1705.03980v2 [math.AC] 25 Jan 2018 AUSLANDER MODULES PEYMAN NASEHPOUR Dedicated to my father, Maestro Nasrollah Nasehpour Abstract. In this paper, we introduce the notion of Auslander modules, inspired from Auslander’s zero-divisor conjecture (theorem) and give some interesting results for these modules. We also investigate torsion-free modules. 0. Introduction Auslander’s zero-divisor conjecture in commutative algebra states that if R is a Noetherian local ring, M is a nonzero R-module of finite type, and finite projective dimension, and r ∈ R is not a zero-divisor on M , then r is not a zero-divisor on R [6, p. 8] and [5, p. 496]. This “conjecture” is in fact a theorem, after Peskine and Szpiro in [15] showed that Auslander’s zero-divisor theorem is a corollary of their new intersection theorem and thereby proved it for a large class of local rings. Also see [16, p. 417]. Note that its validity without any restrictions followed when Roberts [17] proved the new intersection theorem in full generality. Also see Remark 9.4.8 in [1]. Let M be an arbitrary unital nonzero module over a commutative ring R with a nonzero identity. Inspired from Auslander’s zero-divisor theorem, one may ask when the inclusion ZR (R) ⊆ ZR (M ) holds, where by ZR (M ), we mean the set of all zero-divisors of the R-module M . In Definition 1.1, we define an R-module M to be Auslander if ZR (R) ⊆ ZR (M ) and in Proposition 1.2, we give a couple of examples for the families of Auslander modules. The main theme of §1 is to see under what conditions if M is an Auslander R-module, then the S-module M ⊗R S is Auslander, where S is an R-algebra (see Theorem 1.4, Theorem 1.6, and Theorem 1.10). For example, in Corollary 1.11, we show that if M is an Auslander R-module, B a content R-algebra, and M has property (A), then M ⊗R B is an Auslander B-module. For the definition of content algebras refer to [14, Section 6]. On the other hand, let us recall that an R-module M is torsion-free if the natural map M → M ⊗ Q is injective, where Q is the total quotient ring of the ring R [1, p. 19]. It is easy to see that M is a torsion-free R-module if and only if ZR (M ) ⊆ ZR (R). In §2, we investigate torsion-free property under polynomial and power series extensions (see Theorem 2.1 and Theorem 2.2). We also investigate torsion-free Auslander modules (check Proposition 2.5, Theorem 2.7, and Theorem 2.9). In this paper, all rings are commutative with non-zero identities and all modules are unital. 2010 Mathematics Subject Classification. 13A15, 13B25, 13F25. Key words and phrases. Auslander modules, Auslander’s zero-divisor conjecture, content algebras, torsion-free modules. 1 2 PEYMAN NASEHPOUR 1. Auslander Modules We start the first section by defining Auslander modules: Definition 1.1. We define an R-module M to be an Auslander module, if r ∈ R is not a zero-divisor on M , then r is not a zero-divisor on R, or equivalently, if the following property holds: ZR (R) ⊆ ZR (M ). Let us recall that if M is an R-module, the content of m ∈ M , denoted by c(m), is defined to be the following ideal: \ c(m) = {I ∈ Id(R) : m ∈ IM }, where by Id(R), we mean the set of all ideals of R. The R-module M is said to be a content R-module, if m ∈ c(m)M , for all m ∈ M [14]. In the following, we give some families of Auslander modules: Proposition 1.2 (Some Families of Auslander Modules). Let M be an R-module. Then the following statements hold: (1) If R is a domain, then M is an Auslander R-module. (2) If M is a flat and content R-module such that for any s ∈ R, there is an x ∈ M such that c(x) = (s). Then M is an Auslander R-module. (3) If M is an R-module such that Ann(M ) = (0), then M is an Auslander R-module. (4) If for any nonzero s ∈ R, there is an x ∈ M such that s · x 6= 0, i.e. Ann(M ) = (0), then HomR (M, M ) is an Auslander R-module. (5) If N is an R-submodule of an R-module M and N is Auslander, then M is also Auslander. (6) If M is an Auslander R-module, then M ⊕ M ′ is an Auslander R-module for any R-module M ′ . In particular, if {Mi }i∈Λ is a family of R-modules and there L is an i ∈QΛ, say i0 , such that Mi0 is an Auslander R-module, then i∈Λ Mi and i∈Λ Mi are Auslander R-modules. Proof. The statement (1) is obvious. We prove the other statements: (2): Let r ∈ ZR (R). By definition, there is a nonzero s ∈ R such that r · s = 0. Since in content modules c(x) = (0) if and only if x = 0 [14, Statement 1.2] and by assumption, there is a nonzero x ∈ M such that c(x) = (s), we obtain that r · c(x) = (0). Also, since M is a flat and content R-module, by [14, Theorem 1.5], r · c(x) = c(r · x). This implies that r ∈ ZR (M ). (3): Suppose that r · s = 0 for some nonzero s in R. By assumpstion, there exists an x in M such that s · x 6= 0, but r · (s · x) = 0, and so r is a zero-divisor on M . (4): Let r ∈ ZR (R). So, there is a nonzero s ∈ R such that r · s = 0. Define fs : M −→ M by fs (x) = s · x. By assumption, fs is a nonzero element of HomR (M, M ). But rfs = 0. This means that r ∈ ZR (Hom(M, M )). (5): The proof is straightforward, if we consider that Z(N ) ⊆ Z(M ). The statement (6) is just a corollary of the statement (5).  Proposition 1.3. Let M be an Auslander R-module and S a multiplicatively closed subset of R contained in R − ZR (M ). Then, MS is an Auslander RS -module. Proof. Let ZR (R) ⊆ ZR (M ) and S be a multiplicatively closed subset of R such that S ⊆ R − ZR (M ). Take r1 /s1 ∈ ZRS (RS ). So there exists an r2 /s2 6= 0/1 such AUSLANDER MODULES 3 that (r1 · r2 )/(s1 · s2 ) = 0/1. Since S ⊆ R − ZR (R), we have r1 · r2 = 0, where r2 6= 0. But ZR (R) ⊆ ZR (M ), so r1 ∈ ZR (M ). Consequently, there is a nonzero m ∈ M such that r1 · m = 0. Since S ⊆ R − ZR (M ), m/1 is a nonzero element of MS . This point that r1 /s1 · m/1 = 0/1, causes r1 /s1 to be an element of ZRS (MS ) and the proof is complete.  Let us recall that an R-module M has property (A), if each finitely generated ideal I ⊆ ZR (M ) has a nonzero annihilator in M [11, Definition 10]. Examples of modules having property (A) include modules having very few zero-divisors [11, Definition 6]. Especially, finitely generated modules over Noetherian rings have property (A) [7, p. 55]. Homological aspects of modules having very few zerodivisors have been investigated in [12]. Finally, we recall that if R is a ring, G a monoid, and f = r1 X g1 + · · · + rn X gn is an element of the monoid ring R[G], then the content of f , denoted by c(f ), is the finitely generated ideal (r1 , . . . , rn ) of R. Theorem 1.4. Let the R-module M have property (A) and G be a commutative, cancellative, and torsion-free monoid. Then, M [G] is an Auslander R[G]-module if and only if M is an Auslander R-module. Proof. (⇒): Let r ∈ ZR (R). So, r ∈ ZR[G] (R[G]) and by assumption, r ∈ ZR[G] (M [G]). Clearly, this means that there is a nonzero g in ZR[G] (M [G]) such that rg = 0. Therefore, there is a nonzero m in M such that rm = 0. (⇐): Let f ∈ ZR[G] (R[G]). By [11, Theorem 2], there is a nonzero element r ∈ R such that f · r = 0. This implies that c(f ) ⊆ ZR (R). But M is an Auslander R-module, so ZR (R) ⊆ ZR (M ), which implies that c(f ) ⊆ ZR (M ). On the other hand, M has property (A). Therefore, c(f ) has a nonzero annihilator in M . Hence, f ∈ ZR[G] (M [G]) and the proof is complete.  Note that a semimodule version of Theorem 1.4 has been given in [10]. It is good to mention that if R is a ring and f = a0 + a1 X + · · · + an X n + · · · is an element of R[[X]], then Af is defined to be the ideal of R generated by the coefficients of f , i.e. Af = (a0 , a1 , . . . , an , . . .). One can easily check that if R is Noetherian, then Af = c(f ). The following lemma is a generalization of Theorem 5 in [4]: Lemma 1.5. Let R be a Noetherian ring, M a finitely generated R-module, f ∈ R[[X]], g ∈ M [[X]] − {0}, and f g = 0. Then, there is a nonzero constant m ∈ M such that f · m = 0. Proof. Define c(g), the content of g, to be the R-submodule of M generated by its coefficients. If c(f )c(g) = (0), then choose a nonzero m ∈ c(g). Clearly, f · m = 0. Otherwise, by Theorem 3.1 in [2], one can choose a positive integer k, such that c(f )c(f )k−1 c(g) = 0, while c(f )k−1 c(g) 6= 0. Now for each nonzero element m in c(f )k−1 c(g), we have f · m = 0 and the proof is complete.  Theorem 1.6. Let R be a Noetherian ring and the R-module M have property (A). Then, M [[X]] is an Auslander R[[X]]-module if and only if M is an Auslander Rmodule. Proof. By Lemma 1.5, the proof is just a mimicking of the proof of Theorem 1.4.  4 PEYMAN NASEHPOUR Since finitely generated modules over Noetherian rings have property (A) [7, p. 55], we have the following corollary: Corollary 1.7. Let R be a Noetherian ring and M be a finitely generated R-module. Then, M [[X]] is an Auslander R[[X]]-module if and only if M is an Auslander Rmodule. Remark 1.8 (Ohm-Rush Algebras). Let us recall that if B is an R-algebra, then B is said to be an Ohm-Rush R-algebra, if f ∈ c(f )B, for all f ∈ B [3, Definition 2.1]. It is easy to see that if P is a projective R-algebra, then P is an OhmRush R-algebra [14, Corollary 1.4]. Note that if R is a Noetherian ring and f = a0 + a1 X + · · · + an X n + · · · is an element of R[[X]], then Af = c(f ), where Af is the ideal of R generated by the coefficients of f . This simply implies that R[[X]] is an Ohm-Rush R-algebra. Now we go further to define McCoy algebras, though we don’t go through them deeply in this paper. McCoy semialgebras (and algebras) and their properties have been discussed in more details in author’s recent paper on zero-divisors of semimodules and semialgebras [10]. Definition 1.9. We say that B is a McCoy R-algebra, if B is an Ohm-Rush Ralgebra and f · g = 0 with g 6= 0 implies that there is a nonzero r ∈ R such that c(f ) · r = (0), for all f, g ∈ B. Since any content algebra is a McCoy algebra [14, Statement 6.1], we have plenty of examples for McCoy algebras. For instance, if G is a torsion-free abelian group and R is a ring, then R[G] is a content - and therefore, a McCoy - R-algebra [13]. For other examples of McCoy algebras, one can refer to content algebras given in Examples 6.3 in [14]. Now we proceed to give the following general theorem on Auslander modules: Theorem 1.10. Let M be an Auslander R-module and B a faithfully flat McCoy R-algebra. If M has property (A), then M ⊗R B is an Auslander B-module. Proof. Let f ∈ ZB (B). So by definition, there is a nonzero r ∈ R such that c(f ) · r = (0). This implies that c(f ) ⊆ ZR (R). But M is an Auslander R-module. Therefore, c(f ) ⊆ ZR (M ). Since c(f ) is a finitely generated ideal of R [14, p. 3] and M has property (A), there is a nonzero m ∈ M such that c(f ) · m = (0). This means that c(f ) ⊆ AnnR (m). Therefore, c(f )B ⊆ AnnR (m)B. Since any McCoy R-algebra is by definition an Ohm-Rush R-algebra, we have that f ∈ c(f )B. Our claim is that AnnR (m)B = AnnB (1 ⊗ m) and here is the proof: Since 0 −→ R/ AnnR (m) −→ M is an R-exact sequence and B is a faithfully flat R-module, we have the following B-exact sequence: 0 −→ B/ AnnR (m)B −→ M ⊗R B, with AnnR (m)B = Ann(m ⊗R 1B ). This means that f ∈ ZB (M ⊗R B) and the proof is complete.  Corollary 1.11. Let M be an Auslander R-module and B a content R-algebra. If M has property (A), then M ⊗R B is an Auslander B-module. AUSLANDER MODULES 5 Proof. By definition of content algebras [14, Section 6], any content R-algebra is faithfully flat. Also, by [14, Statement 6.1], any content R-algebra is a McCoy R-algebra.  Question 1.12. Is there any faithfully flat McCoy algebra that is not a content algebra? 2. Torsion-Free Modules Let us recall that if R is a ring, M an R-module, and Q the total ring of fractions of R, then M is torsion-free if the natural map M → M ⊗ Q is injective [1, p. 19]. It is starightforward to see that M is a torsion-free R-module if and only if ZR (M ) ⊆ ZR (R). Therefore, the notion of Auslander modules defined in Definition 1.1 is a kind of dual to the notion of torsion-free modules. The proof of the following theorem is quite similar to the proof of Proposition 1.4. Therefore, we just mention the proof briefly. Theorem 2.1. Let the ring R have property (A) and G be a commutative, cancellative, and torsion-free monoid. Then, the R[G]-module M [G] is torsion-free if and only if the R-module M is torsion-free. Proof. (⇒): Let r ∈ ZR (M ). Clearly, this implies that r ∈ ZR[G] (M [G]). But the R[G]-module M [G] is torsion-free. Therefore, ZR[G] (M [G]) ⊆ ZR[G] (R[G]). So, r ∈ ZR (R). (⇐): Let f ∈ ZR[G] (M [G]). By [11, Theorem 2], there is a nonzero m ∈ M such that c(f ) · m = 0, which means that c(f ) ⊆ ZR (M ). Since M is torsion-free, c(f ) ⊆ ZR (R), and since R has property (A), f ∈ ZR[G] (R[G]) and the proof is complete.  Theorem 2.2. Let R be a Noetherian ring and M be a finitely generated R-module. Then, the R[[X]]-module M [[X]] is torsion-free if and only if the R-module M is torsion-free. Proof. (⇒): Its proof is similar to the proof of Theorem 2.1 and therefore, we don’t bring it here. (⇐): Let f ∈ ZR[[X]] (M [[X]]). By Lemma 1.5, there is a nonzero element m ∈ M such that f · m = 0. By Remark 1.8, this implies that c(f ) ⊆ ZR (M ). But M is torsion-free, so ZR (M ) ⊆ ZR (R), which implies that c(f ) ⊆ ZR (R). On the other hand, since every Noetherian ring has property (A) (check [7, Theorem 82, p. 56]), c(f ) has a nonzero annihilator in R. This means that f ∈ ZR[[X]] (R[[X]]), Q.E.D.  We continue this section by investigating torsion-free Auslander modules. Remark 2.3. In the following, we show that there are examples of modules that are Auslander but not torsion-free and also there are some modules that are torsion-free but not Auslander. (1) Let R be a ring and S ⊆ R − ZR (R) a multiplicatively closed subset of R. Then, it is easy to see that ZR (R) = ZR (RS ), i.e. RS is a torsion-free Auslander R-module. 6 PEYMAN NASEHPOUR (2) Let D be a domain and M a D-module such that ZD (M ) 6= {0}. Clearly, ZD (D) = {0} and therefore, M is Auslander, while M is not torsion-free. For example, if D is a domain that is not a field, then D has an ideal I such that I 6= (0) and I 6= D. It is clear that ZD (D/I) ⊇ I. (3) Let k be a field and consider the ideal I = (0)⊕k of the ring R = k ⊕k. It is easy to see that ZR (R) = ((0)⊕k)∪(k⊕(0)), while ZR (R/I) = (0)⊕k. This means that the R-module R/I is torsion-free, while it is not Auslander. Proposition 2.4 (Some Families of Torsion-free Auslander Modules). Let M be an R-module. Then, the following statements hold: (1) If R is a domain and M is a flat R-module, then M is torsion-free Auslander R-module. (2) If M is a flat and content R-module such that for any s ∈ R, there is an x ∈ M such that c(x) = (s). Then M is a torsion-free Auslander R-module. (3) If R is a Noetherian ring and M is a finitely generated flat R-module and for any nonzero s ∈ R, there is an x ∈ M such that s · x 6= 0. Then HomR (M, M ) is a torsion-free Auslander R-module. (4) If M is an Auslander R-module, and M and M ′ are both flat modules, then M ⊕ M ′ is a torsion-free Auslander R-module. In particular, if {Mi }i∈Λ is a family of flat R-modules and L there is an i ∈ Λ, say i0 , such that Mi0 is an Auslander R-module, then i∈Λ Mi is a torsion-free Auslander R-module. (5) If R is a coherent ring and {Mi }i∈Λ is a family of flat R-modules and Q there is an i ∈ Λ, say i0 , such that Mi0 is an Auslander R-module, then i∈Λ Mi is a torsion-free Auslander R-module. Proof. It is trivial that every flat module is torsion-free. By considering Proposition 1.2, the proof of statements (1) and (2) is straightforward. The proof of statement (3) is based on Theorem 7.10 in [9] that says that each finitely generated flat module over a local ring is free. Now if R is a Noetherian ring and M is a flat and finitely generated R-module, then M is a locally free Rmodule. This causes HomR (M, M ) to be also a locally free R-module and therefore, HomR (M, M ) is R-flat and by Proposition 1.2, a torsion-free Auslander R-module. The proof of the statements (4) and (5) is also easy, if we note that the direct sum of flat modules is flat [8, Proposition 4.2] and, if R is a coherent ring, then the direct product of flat modules is flat [8, Theorem 4.47].  Proposition 2.5. Let both the ring R and the R-module M have property (A) and G be a commutative, cancellative, and torsion-free monoid. Then, M [G] is a torsion-free Auslander R[G]-module if and only if M is a torsion-free Auslander R-module. Proof. By Theorem 1.4 and Theorem 2.1, the statement holds.  Corollary 2.6. Let R be a Noetherian ring and M a finitely generated R-module, and G a commutative, cancellative, and torsion-free monoid. Then, M [G] is a torsion-free Auslander R[G]-module if and only if M is a torsion-free Auslander R-module. Theorem 2.7. Let M be a flat Auslander R-module and B a faithfully flat McCoy R-algebra. If M has property (A), then M ⊗R B is a torsion-free Auslander Bmodule. AUSLANDER MODULES 7 Proof. By Theorem 1.10, ZB (B) ⊆ ZB (M ⊗R B). On the other hand, since M is a flat R-module, by [8, Proposition 4.1], M ⊗R B is a flat B-module. This implies that ZB (M ⊗R B) ⊆ ZB (B) and the proof is complete.  Corollary 2.8. Let M be a flat Auslander R-module and B a content R-algebra. If M has property (A), then M ⊗R B is a torsion-free Auslander B-module. Theorem 2.9. Let R be a Noetherian ring and M a finitely generated R-module. Then, M [[X]] is a torsion-free Auslander R[[X]]-module if and only if M is a torsion-free Auslander R-module. Proof. Since M is finite and R is Noetherian, M is also a Noetherian R-module. This means that both the ring R and the module M have property (A). Now by Theorem 1.6 and Theorem 2.2, the proof is complete.  Acknowledgements The author was partly supported by the Department of Engineering Science at Golpayegan University of Technology and wishes to thank Professor Winfried Bruns for his invaluable advice. The author is also grateful for the useful comments by the anonymous referee. References [1] W. Bruns and J. Herzog, Cohen-Macaulay Rings, revised edn., Cambridge University Press, Cambridge, 1998. [2] N. Epstein and J. Shapiro, A Dedekind-Mertens theorem for power series rings, Proc. Amer. Math. Soc. 144 (2016), 917–924. [3] N. Epstein and J. Shapiro, The Ohm-Rush content function, J. Algebra Appl., 15, No. 1 (2016), 1650009 (14 pages). [4] D. E. Fields, Zero divisors and nilpotent elements in power series rings, Proc. Amer. Math. Soc. 27 (1971), no. 3, 427–433. [5] M. Hochster, Intersection Problems and Cohen-Macaulay Modules, in Algebraic Geometry: Bowdoin 1985 (part 2) (1987), 491–501. [6] M. Hochster, Topics in the homological theory of modules over commutative rings, CBMS Regional Conf. Ser. in Math., 24, Amer. Math. Soc, Providence, RI, 1975. [7] I. Kaplansky, Commutative Rings, Allyn and Bacon, Boston, 1970. [8] T. Y. Lam, Lectures on Modules and Rings, Springer-Verlag, Berlin, 1999. [9] H. Matsumura, Commutative Ring Theory, Vol. 8., Cambridge University Press, Cambridge, 1989. [10] P. Nasehpour, On zero-divisors of semimodules and semialgebras, arXiv preprint (2017), arXiv:1702.00810. [11] P. Nasehpour, Zero-divisors of semigroup modules, Kyungpook Math. J., 51 (1) (2011), 37–42. [12] P. Nasehpour and Sh. Payrovi, Modules having few zero-divisors, Comm. Algebra 38 (2010), 3154–3162. [13] D. G. Northcott, A generalization of a theorem on the content of polynomials, Proc. Cambridge Phil. Soc. 55 (1959), 282–288. [14] J. Ohm and D. E. Rush, Content modules and algebras, Math. Scand. 31 (1972), 49–68. [15] C. Peskine and L. Szpiro, Dimension projective finie et cohomologie locale, Publ. Math. IHES 42, (1973), 47-119. [16] P. Roberts, Intersection theorems, in: Commutative Algebra, in: Math. Sci. Res. Inst. Publ., Vol. 15, Springer-Verlag, Berlin, 1989, 417–436. [17] P. Roberts, Le théoreme d’intersection, CR Acad. Sci. Paris Ser. I Math 304 (1987), 177–180. 8 PEYMAN NASEHPOUR Department of Engineering Science, Golpayegan University of Technology, Golpayegan, Iran E-mail address: nasehpour@gut.ac.ir, nasehpour@gmail.com 

arXiv:1406.0915v2 [math.AT] 11 May 2016 A VANISHING THEOREM FOR THE p-LOCAL HOMOLOGY OF COXETER GROUPS TOSHIYUKI AKITA A BSTRACT. Given an odd prime number p and a Coxeter group W such that the order of the product st is prime to p for every Coxeter generators s,t of W , we prove that the p-local homology groups Hk (W, Z(p) ) vanish for 1 ≤ k ≤ 2(p − 2). This generalize a known vanishing result for symmetric groups due to Minoru Nakaoka. 1. I NTRODUCTION Coxeter groups are important objects in many branches of mathematics, such as Lie theory and representation theory, combinatorial and geometric group theory, topology and geometry. Since the pioneering work of Serre [21], Coxeter groups have been studied in group cohomology as well. See the book by Davis [9] and §2.2 of this paper for brief outlooks. In this paper, we will study the p-local homology of Coxeter groups for odd prime numbers p. For an arbitrary Coxeter group W , its integral homology group Hk (W, Z) is known to be a finite abelian group for all k > 0, and hence it decomposes into a finite direct sum of p-local homology groups each of which is a finite abelian p-group: Hk (W, Z)  M Hk (W, Z(p) ). p According to a result of Howlett [13], the first and second p-local homology groups, H1 (W, Z(p) ) and H2 (W, Z(p) ), are trivial for every odd prime number p. On the other hand, the symmetric group of n letters Sn (n ≥ 2) is the Coxeter group of type An−1 . Much is known about the (co)homology of symmetric groups. Most notably, in his famous two papers, Nakaoka proved the homology stability for symmetric groups [16] and computed the stable mod p homology [17]. As a consequence of his results, Hk (Sn , Z(p) ) vanishes for 1 ≤ k ≤ 2(p − 2) (see Theorem 2.4 below). The purpose of this paper is to generalize vanishing of Hk (Sn , Z(p) ) to all Coxeter groups: Theorem 1.1. Let p be an odd prime number and W a p-free Coxeter group. Then Hk (W, Z(p) ) = 0 holds for 1 ≤ k ≤ 2(p − 2). Here a Coxeter group W is said to be p-free if the order of the product st is prime to p for every distinct Coxeter generators s,t of W . We should remark that, for p ≥ 2010 Mathematics Subject Classification. Primary 20F55, 20J06; Secondary 55N91. Key words and phrases. Coxeter groups, Group homology. 1 2 T. AKITA 5, the p-freeness assumption is necessary and the vanishing range 1 ≤ k ≤ 2(p − 2) is best possible. The situation is somewhat different for p = 3. See §5.4. The proof of Theorem 1.1 consists of two steps, a case by case argument for finite p-free Coxeter groups with relatively small rank, and the induction on the number of generators. The induction is made possible by means of the equivariant homology of Coxeter complexes and the Leray spectral sequence converging to the equivariant homology. Now we will introduce the content of this paper very briefly. In §2.1, we will recall definitions and relevant facts concerning of Coxeter groups. Known results about homology of Coxeter groups and their consequences will be reviewed more precisely in §2.2. After the consideration of the equivariant homology of Coxeter complexes in §3. the proof of Theorem 1.1 will be given in §4. The final section §5 consists of miscellaneous results. There we will give some classes of Coxeter groups such that all the p-local homology groups vanish. Notation. Throughout this paper, p is an odd prime number unless otherwise stated. Z(p) is the localization of Z at the prime p (the ring of p-local integers). For a finite abelian group A, its p-primary component is denoted by A(p) . Note that A(p)  A ⊗Z Z(p) . For a group G and a (left) G-module M, the co-invariant of M is denoted by MG (see [8, II.2] for the definition). For a prime number p ≥ 2, we denote the cyclic group of order p and the field with p elements by the same symbol Z/p. 2. P RELIMINARIES 2.1. Coxeter groups. We recall definitions and relevant facts concerning of Coxeter groups. Basic references are [1, 7, 9, 14]. See also [12] for finite Coxeter groups. Let S be a finite set and m : S × S → N ∪ {∞} a map satisfying the following conditions: (1) m(s, s) = 1 for all s ∈ S (2) 2 ≤ m(s,t) = m(t, s) ≤ ∞ for all distinct s,t ∈ S. The map m is represented by the Coxeter graph Γ whose vertex set is S and whose edges are the unordered pairs {s,t} ⊂ S such that m(s,t) ≥ 3. The edges with m(s,t) ≥ 4 are labeled by those numbers. The Coxeter system associated to Γ is the pair (W, S) where W = W (Γ) is the group generated by s ∈ S and the fundamental relations (st)m(s,t) = 1 (m(s,t) < ∞): W := hs ∈ S | (st)m(s,t) = 1(m(s,t) < ∞)i. The group W is called the Coxeter group of type Γ, and elements of S are called Coxeter generators of W . The cardinality of S is called the rank of W and is denoted by |S| or rankW . Note that the order of the product st is precisely m(s,t). For a subset T ⊆ S (possibly T = ∅), the subgroup WT := hT i of W generated by elements t ∈ T is called a parabolic subgroup. In particular, WS = W and W∅ = {1}. It is known that (WT , T ) is a Coxeter system. A Coxeter group W is called irreducible if its defining graph Γ is connected, otherwise called reducible. For a reducible Coxeter group W (Γ) of type Γ, if Γ p-LOCAL HOMOLOGY OF COXETER GROUPS 3 consists of the connected components Γ1 , Γ2 , . . . , Γr , then W (Γ) is the direct product of parabolic subgroups W (Γi )’s (1 ≤ i ≤ r), each of which is irreducible: W (Γ) = W (Γ1 ) ×W (Γ2 ) × · · · ×W (Γr ). Coxeter graphs for finite irreducible Coxeter groups are classified. There are four infinite families An (n ≥ 1), Bn (n ≥ 2), Dn (n ≥ 4), I2 (q) (q ≥ 3), and six exceptional graphs E6 , E7 , E8 , F4 , H3 and H4 . The subscript indicates the rank of the resulting Coxeter group. See Appendix for the orders of finite irreducible Coxeter groups. Here we follow the classification given in the book by Humphreys [14] and there are overlaps A2 = I2 (3), B2 = I2 (4). Note that W (An ) is isomorphic to the symmetric group of n + 1 letters, while W (I2 (q)) is isomorphic to the dihedral group of order 2q. Finally, given an odd prime number p, we define a Coxeter group W to be p-free if m(s,t) is prime to p for all s,t ∈ S. Here ∞ is prime to all prime numbers by the convention. For example, the Coxeter group W (I2 (q)) is p-free if and only if q is prime to p, while the Coxeter group W (An ) (n ≥ 2) is p-free for p ≥ 5. For every finite irreducible Coxeter group W , the range of odd prime numbers p such that W is p-free can be found in Appendix. Note that parabolic subgroups of p-free Coxeter groups are also p-free. Henceforth, we omit the reference to the Coxeter graph Γ and the set of Coxeter generators S if there is no ambiguity. 2.2. Known results for homology of Coxeter groups. In this subsection, we will review some of known results concerning the homology of Coxeter groups which are related to our paper. A basic reference for (co)homology of groups is [8]. In the beginning, Serre [21] proved that every Coxeter group W has finite virtual cohomological dimension and is a group of type WFL (see [8, Chapter VIII] for definitions). This implies, in particular, that Hk (W, Z) is a finitely generated abelian group for all k. On the other hand, the rational homology of any Coxeter groups are known to be trivial (see [5, Proposition 5.2] or [9, Theorem 15.1.1]). Combining these results, we obtain the following result: Proposition 2.1. For any Coxeter groups W , the integral homology group Hk (W, Z) is a finite abelian group for all k > 0. Consequently, Hk (W, Z) (k > 0) decomposes into a finite direct sum (2.1) Hk (W, Z)  M Hk (W, Z)(p) p where p runs the finite set of prime numbers dividing the order of Hk (W, Z). The universal coefficient theorem implies (2.2) Hk (W, Z)(p)  Hk (W, Z) ⊗Z Z(p)  Hk (W, Z(p) ) (see [6, Corollary 2.3.3]). It turns out that the study of the integral homology groups of Coxeter groups reduces to the study of the p-local homology groups. Later, we will prove that Hk (W, Z(p) ) = 0 (k > 0) if W has no p-torsion (Proposition 5.3). The first and second integral homology of Coxeter groups are known. 4 T. AKITA Proposition 2.2. For any Coxeter groups W , we have H1 (W, Z)  (Z/2)n1 (W ) and H2 (W, Z)  (Z/2)n2 (W ) for some non-negative integers n1 (W ), n2 (W ). The claim for H1 (W, Z) is obvious because H1 (W, Z) = W /[W,W ] and W is generated by elements of order 2. The statement for H2 (W, Z) was proved by Howlett [13] (following earlier works by Ihara and Yokonuma [15] and Yokonuma [25]). The nonnegative integers n1 (W ), n2 (W ) can be computed from the Coxeter graph for W . As for n1 (W ), let GW be the graph whose vertices set is S and whose edges are unordered pair {s,t} ⊂ S such that m(s,t) is a finite odd integer. Then it is easy to see that n1 (W ) agrees with the number of connected components of GW . In particular, n1 (W ) ≥ 1 and hence H1 (W, Z) = H1 (W, Z(2) ) , 0. For the presentation of n2 (W ), see [13, Theorem A] or [14, §8.11]. As a consequence of Proposition 2.2, we obtain the following result: Corollary 2.3. Let p be an odd prime number. For any Coxeter groups W , we have H1 (W, Z(p) ) = H2 (W, Z(p) ) = 0. The corollary does not hold for the third homology or higher. Indeed, for the Coxeter group W (I2 (q)) of type I2 (q), which is isomorphic to the dihedral group of order 2q as mentioned before, it can be proved that, if p divides q, then Hk (W (I2 (q)), Z(p) ) , 0 whenever k ≡ 3 (mod 4) (see [20, Theorem 2.1] and §5.1 below). This observation also shows the necessity of the p-freeness assumption in our results for p ≥ 5. Finally, we will recall a consequence of results of Nakaoka [16, 17] which was mentioned in the introduction. Theorem 2.4 (Nakaoka [16, 17]). Let Sn be the symmetric group of n letters. Then Hk (Sn , Z(p) ) = 0 (1 ≤ k ≤ 2(p − 2)) for all odd prime numbers p. Proof. In his paper [16], Nakaoka proved the homology stability for symmetric groups. Namely, for 2 ≤ m ≤ n ≤ ∞, the homomorphism Hk (Sm , A) → Hk (Sn , A) induced by the natural inclusion Sm ֒→ Sn is injective for all k, and is an isomorphism if k < (m + 1)/2, where A is an abelian group with the trivial Sn -action, and S∞ is the infinite symmetric group [16, Theorem 5.8 and Corollary 6.7]. He also computed the mod p homology of S∞ in [17, Theorem 7.1], from which we deduce that Hk (S∞ , Z/p) = 0 for 1 ≤ k ≤ 2(p − 2) and that H2p−3 (S∞ , Z/p) , 0. Combining these results, we see that Hk (Sn , Z/p) = 0 (1 ≤ k ≤ 2(p − 2)) for all n. Applying the universal coefficient theorem, the theorem follows.  Theorem 1.1, together with Corollary 2.3 for p = 3, generalize Theorem 2.4 to all Coxeter groups. For further results concerning of (co)homology of Coxeter groups, we refer the book by Davis [9] and papers [2–4, 10, 11, 18, 20, 23] as well as references therein. 3. C OXETER COMPLEXES AND THEIR EQUIVARIANT HOMOLOGY 3.1. Coxeter complexes. We recall the definition and properties of Coxeter complexes which are relevant to prove Theorem 1.1. A basic reference for Coxeter complexes is [1, Chapter 3]. Given a Coxeter group W , the Coxeter complex XW of W is the poset of cosets wWT (w ∈ W, T ( S), ordered by reverse inclusion. It is p-LOCAL HOMOLOGY OF COXETER GROUPS 5 known that XW is actually an (|S|−1)-dimensional simplicial complex (see [1, Theorem 3.5]). The k-simplices of XW are the cosets wWT with k = |S| − |T | − 1. A coset wWT is a face of w′WT ′ if and only if wWT ⊇ w′WT ′ . In particular, the vertices are cosets of the form wWS\{s} (s ∈ S, w ∈ W ), the maximal simplices are the singletons wW∅ = {w} (w ∈ W ), and the codimension one simplices are cosets of the form wW{s} = {w, ws} (s ∈ S, w ∈ W ). In what follows, we will not distinguish between XW and its geometric realization. There is a simplicial action of W on XW by left translation w′ · wWT := w′ wWT . The isotropy subgroup of a simplex wWT is precisely wWT w−1 , which fixes wWT pointwise. Next, consider the subcomplex ∆W = {WT | T ( S} of XW , which consists of a single (|S| − 1)-simplex W∅ and its faces. Since the type function XW → S, wWT 7→ S \ T is well-defined (see [1, Definition 3.6]), ∆W forms the set of representatives of W -orbits of simplices of XW . The following fact is well-known. Proposition 3.1. If W is a finite Coxeter group, then XW is a triangulation of the (|S| − 1)-dimensional sphere S|S|−1 . See [1, Proposition 1.108] for the proof. Alternatively, W can be realized as an orthogonal reflection group on the |S|-dimensional Euclidean space R|S| and hence it acts on the unit sphere S|S|−1 . Each s ∈ S acts on S|S|−1 as an orthogonal reflection. The Coxeter complex XW coincides with the equivariant triangulation of S|S|−1 cut out by the reflection hyperplanes for W . In case W is infinite, Serre proved the following result: Proposition 3.2 ([21, Lemma 4]). If W is an infinite Coxeter group, then XW is contractible. 3.2. Equivariant homology of Coxeter complexes. Given a Coxeter group W , let HkW (XW , Z(p) ) be the k-th equivariant homology group of XW (see [8, Chapter VII] for the definition). If XW is infinite, then XW is contractible so that the equivariant homology is isomorphic to the homology of W : Proposition 3.3. If W is an infinite Coxeter group, then HkW (XW , Z(p) )  Hk (W, Z(p) ) for all k. If W is finite, then HkW (XW , Z(p) ) may not be isomorphic to Hk (W, Z(p) ), however, they are isomorphic if k is relatively small: Proposition 3.4. If W is a finite Coxeter group, then HkW (XW , Z(p) )  Hk (W, Z(p) ) for k ≤ rankW − 1. Proof. Consider the spectral sequence W Ei,2 j = Hi (W, H j (XW , Z(p) )) ⇒ Hi+ j (XW , Z(p) ) 6 T. AKITA 2  H (W, Z ) for all i. Since X is homeomorphic (see [8, VII.7]) and note that Ei,0 i W (p) W |S|−1 2 to S , we have Ei, j = 0 for j , 0, |S| − 1. Hence Hk (XW , Z(p) )  Hk (W, Z(p) ) for k ≤ |S| − 2. Now 2 E0,|S|−1 = H0 (W, H|S|−1 (XW , Z(p) )) = H|S|−1 (XW , Z(p) )W where the RHS is the co-invariant of H|S|−1 (XW , Z(p) ) as a W -module (see [8, III.1]). Since each s ∈ S acts on XW ≈ S|S|−1 as an orthogonal reflection as mentioned in §3.1, it acts on H|S|−1 (XW , Z(p) )  Z(p) as the multiplication by −1. It follows that the co-invariant H|S|−1 (XW , Z(p) )W is isomorphic to the quotient group of Z(p) by the subgroup generated by r − (−1)r = 2r (r ∈ Z(p) ). But this subgroup is nothing but the whole group Z(p) because 2 is invertible in Z(p) . This proves 2 E0,|S|−1 = 0 and hence W H|S|−1 (XW , Z(p) )  H|S|−1 (W, Z(p) ) as desired.  4. P ROOF OF T HEOREM 1.1 We will prove Theorem 1.1 by showing the following two claims: Claim 1. If W is a finite p-free Coxeter group with rankW ≤ 2(p − 2), then Hk (W, Z(p) ) = 0 for 1 ≤ k ≤ 2(p − 2). Claim 2. Claim 1 implies Theorem 1.1. The first claim is equivalent to Theorem 1.1 for finite p-free Coxeter groups with rankW ≤ 2(p − 2), and will be proved by a case by case argument. The second claim will be proved by the induction on rankW by using the equivariant homology of Coxeter complexes. Let us prove Claim 2 first. 4.1. Proof of Claim 2. For every Coxeter group W , there is a spectral sequence (4.1) Ei,1 j = M σ ∈Si W H j (Wσ , Z(p) ) ⇒ Hi+ j (XW , Z(p) ), where Si is the set of representatives of W -orbits of i-simplices of XW , and Wσ is the isotropy subgroup of an i-simplex σ (see [8, VII.7]). It is the Leray spectral sequence for the natural projection EW ×W XW → XW /W . Note that Z(p) in H j (Wσ , Z(p) ) is the trivial Wσ -module because Wσ fixes σ pointwise. We may choose the subset {WT | T ( S, |T | = |S| − i − 1} (the set of i-simplices of ∆W ) as Si , and the spectral sequence can be rewritten as (4.2) Ei,1 j = M W H j (WT , Z(p) ) ⇒ Hi+ j (XW , Z(p) ). T (S |T |=|S|−i−1 2 = 0 for i , 0 and E 2  Z . Lemma 4.1. In the spectral sequence (4.2), Ei,0 (p) 0,0 p-LOCAL HOMOLOGY OF COXETER GROUPS 7 2  H (∆ , Z ) for all i, which implies the lemma because Proof. We claim Ei,0 i W (p) ∆W is an (|S| − 1)-simplex and hence contractible. Although such a claim may be familiar to experts, we write down the proof for completeness. To show the claim, we recall the construction of the spectral sequence (4.1) given in [8, VII.7]. At the 1 -term of (4.1) is given by first stage, the Ei,0 1 Ei,0 = H0 (W,Ci (XW , Z(p) )) = Ci (XW , Z(p) )W , which is isomorphic to the one in (4.1) due to Eckmann-Shapiro lemma. The differ1 → E1 ential d 1 : Ei,0 i−1,0 is the map induced by the boundary operator Ci (XW , Z(p) ) → Ci−1 (XW , Z(p) ). On the other hand, the composition (4.3) Ci (∆W , Z(p) ) ֒→ Ci (XW , Z(p) ) ։ Ci (XW , Z(p) )W is an isomorphism, where the first map is induced by the inclusion ∆W ֒→ XW and the second map is the natural projection, because the subcomplex ∆W forms the set of representatives of W -orbits of simplices of XW . Moreover, the isomorphism (4.3) is compatible with the boundary operator of Ci (∆W , Z(p) ) and the differential on Ci (XW , Z(p) )W . In other words, (4.3) yields a chain isomorphism of chain complexes (Ci (∆W , Z(p) ), ∂ ) → (Ci (XW , Z(p) )W , d 1 ). The claim follows immediately.  Proof of Claim 2. We argue by the induction on |S|. When W is finite, we may assume |S| > 2(p − 2), for we suppose that Claim 1 holds. Consider the spectral sequence (4.2). Observe first that all WT ’s appearing in (4.2) are p-free and satisfy |T | < |S|. By the induction assumption, we have H j (WT , Z(p) ) = 0 (1 ≤ j ≤ 2(p − 2)) for all T ( S, which implies Ei,1 j = Ei,2 j = 0 for 1 ≤ j ≤ 2(p − 2). Moreover, 2 = 0 for i > 0 by Lemma 4.1. This proves H W (X , Z ) = 0 for 1 ≤ k ≤ Ei,0 W (p) k 2(p − 2). Now Claim 2 follows from Proposition 3.3 and 3.4.  4.2. Proof of Claim 1. Given an odd prime p, if W is a finite p-free Coxeter group with rankW ≤ 2(p − 2), then W decomposes into the direct product of finite irreducible p-free Coxeter groups W  W1 × · · · × Wr with Σri=1 rankWi = rankW . Since Z(p) is PID, we may apply the Künneth theorem to conclude that Claim 1 is equivalent to the following claim: Claim 3. If W is a finite irreducible p-free Coxeter group with rankW ≤ 2(p − 2), then Hk (W, Z(p) ) vanishes for 1 ≤ k ≤ 2(p − 2). We prove Claim 3 for each finite irreducible Coxeter group. Firstly, the Coxeter group W (I2 (q)) of type I2 (q) is p-free if and only if q is prime to p. If so, H∗ (W (I2 (q)), Z(p) ) = 0 for ∗ > 0 because the order of W (I2 (q)) is 2q and hence having no p-torsion. Next, we prove the claim for the Coxeter group of type An . To do so, we deal with cohomology instead of homology. We invoke the following elementary lemma: Lemma 4.2. Let G be a finite group and p ≥ 2 a prime. Then Hk (G, Z(p) )  H k+1 (G, Z)(p) for all k ≥ 1. 8 T. AKITA Now Claim 3 for W (An ) can be proved by applying standard arguments in cohomology of finite groups: Lemma 4.3. Hk (W (An ), Z(p) ) = 0 (1 ≤ k ≤ 2(p − 2)) holds for all n with 1 ≤ n ≤ 2(p − 1). Proof. Recall that W (An ) is isomorphic to the symmetric group of n + 1 letters Sn+1 . If n < p then Sn has no p-torsion and hence Hk (Sn , Z(p) ) = 0 for all k > 0. Now suppose p ≤ n ≤ 2p − 1, and let C p be a Sylow p-subgroup of Sn , which is a cyclic group of order p. Then H ∗ (C p , Z)  Z[u]/(pu) where deg u = 2. Let N p be the normalizer of C p in Sn . It acts on C p by conjugation, and the induced map N p → Aut(C p )  (Z/p)× is known to be surjective. Consequently, the invariant H ∗ (C p , Z)Np is the subring generated by u p−1 . Since C p is abelian, the restriction H ∗ (Sn , Z) → H ∗ (C p , Z) induces the isomorphism H k (Sn , Z)(p)  H k (C p , Z)Np for k > 0 by a result of Swan [22, Lemma 1 and Appendix] (see also [24, Lemma 3.4]). This proves, for p ≤ n ≤ 2p − 1, that H k (Sn , Z)(p) = 0 (0 < k < 2p − 2) and H 2p−2 (Sn , Z)(p) , 0. In view of Lemma 4.2, the proposition follows.  Remark 4.4. Since H2p−3 (W (A p−1 ), Z(p) )  H 2p−2 (S p , Z)(p) , 0 for all prime numbers p as was observed in the proof of Lemma 4.3, the vanishing range 1 ≤ k ≤ 2(p − 2) in our theorem is best possible for p ≥ 5. Remark 4.5. Of course, Lemma 4.3 is a direct consequence of Theorem 2.4, however, we avoid the use of Theorem 2.4 for two reasons: Firstly, by doing so, we provide an alternative proof of Theorem 2.4. Secondly, the proof of Lemma 4.3 is much simpler than that of Theorem 2.4, for the latter relies on the homology stability for symmetric groups and the computation of H∗ (S∞ , Z/p). Claim 3 for the Coxeter groups of type Bn and Dn follows from Lemma 4.3 and the following proposition: Proposition 4.6. For any odd prime number p, H∗ (W (Bn ), Z(p) )  H∗ (W (An−1 ), Z(p) ) holds for all n ≥ 2, and H∗ (W (Dn ), Z(p) )  H∗ (W (An−1 ), Z(p) ) holds for all n ≥ 4. Proof. Recall that the Coxeter group W (Bn ) is isomorphic to the semi-direct product (Z/2)n ⋊W (An−1 ) (see [9, §6.7] or [14, §1.1]). In the Lyndon-Hochschild-Serre spectral sequence Ei,2 j = Hi (W (An−1 ), H j ((Z/2)n , Z(p) )) ⇒ Hi+ j (W (Bn ), Z(p) ), one has Ei,2 j = 0 for j , 0 since H j ((Z/2)n , Z(p) ) = 0 for j , 0. This proves H∗ (W (Bn ), Z(p) )  H∗ (W (An−1 ), Z(p) ). On the other hand, W (Dn ) is known to be isomorphic to the semi-direct product (Z/2)n−1 ⋊ W (An−1 ) (see loc. cit.), and the proof for H∗ (W (Dn ), Z(p) )  H∗ (W (An−1 ), Z(p) ) is similar.  p-LOCAL HOMOLOGY OF COXETER GROUPS 9 These observations prove Claim 3 for p ≥ 11, for all finite irreducible Coxeter groups of type other than An , Bn , Dn and I2 (q) have no p-torsion for p ≥ 11. The case p = 3 follows from Corollary 2.3. Now we will prove the cases p = 5 and p = 7. Observe that, apart from Coxeter groups of type An , Bn , Dn and I2 (q), finite irreducible p-free Coxeter groups, with rank at most 2(p − 2) and having ptorsion, are W (E6 ) for p = 5, W (E7 ) and W (E8 ) for p = 7. So the proof of Claim 3 is completed by showing the following lemma: Lemma 4.7. Hk (W (E6 ), Z(5) ) vanishes for 1 ≤ k ≤ 6, while Hk (W (E7 ), Z(7) ) and Hk (W (E8 ), Z(7) ) vanish for 1 ≤ k ≤ 10. Proof. The Coxeter group W (A4 ) is a parabolic subgroup of W (E6 ), and they have a common Sylow 5-subgroup C5 , which is a cyclic group of order 5. The transfer homomorphism to the Sylow 5-subgroup Hk (W (E6 ), Z(5) ) → Hk (C5 , Z(5) ) is injective and factors into a composition of transfer homomorphisms Hk (W (E6 ), Z(5) ) → Hk (W (A4 ), Z(5) ) → Hk (C5 , Z(5) ). In view of Lemma 4.3, we conclude that Hk (W (E6 ), Z(5) ) = 0 for 1 ≤ k ≤ 6, which proves the lemma for W (E6 ). On the other hand, there is a sequence of parabolic subgroups W (A6 ) < W (E7 ) < W (E8 ), and they have a common Sylow 7-subgroup C7 , which is a cyclic group of order 7. The proof of the lemma for W (E7 ) and W (E8 ) is similar.  5. C OXETER GROUPS WITH VANISHING p- LOCAL HOMOLOGY In this final section, we introduce some families of Coxeter groups such that Hk (W, Z(p) ) vanishes for all k > 0. 5.1. Aspherical Coxeter groups. A Coxeter group W is called aspherical in [18] if, for all distinct Coxeter generators s,t, u ∈ S, the inequality 1 1 1 + + ≤1 m(s,t) m(t, u) m(u, s) holds, where 1/∞ = 0 by the convention. The inequality is equivalent to the condition that the parabolic subgroup W{s,t,u} is of infinite order. The (co)homology groups of aspherical Coxeter groups were studied by Pride and Stöhr [18], and the mod 2 cohomology rings of aspherical Coxeter groups were studied by the author [4]. Among other things, Pride and Stöhr obtained the following exact sequence · · · → Hk+1 (W, A) → M s∈S Hk (W{s} , A)⊕n(s) → M Hk (W{s,t} , A) → Hk (W, A) → · · · {s,t}⊂S m(s,t)<∞ terminating at H2 (W, A), where A is a W -module and n(s) is a certain nonnegative integer defined for each s ∈ S [18, Theorem 5]. Since W{s}  Z/2, Hk (W{s} , Z(p) ) = 0 for k > 0. Moreover, if p does not divide m(s,t), then Hk (W{s,t} , Z(p) ) = 0 for k > 0 either. Here no prime numbers p divide ∞ by the convention. Hence we obtain the following result (the statement for k = 1, 2 follows from Corollary 2.3): 10 T. AKITA Proposition 5.1. For any aspherical Coxeter groups W , we have Hk (W, Z(p) )  M Hk (W{s,t} , Z(p) ) {s,t}⊂S p |m(s,t) for all k > 0. Furthermore, Hk (W, Z(p) ) vanishes for all k > 0 if and only if W is p-free. Note that if p divides m(s,t), then ( (Z/m(s,t))(p) Hk (W{s,t} , Z(p) )  0 k ≡ 3 (mod 4) k . 3 (mod 4) for k > 0, where Z/m(s,t) is the cyclic group of order m(s,t) (see [20, Theorem 2.1]). 5.2. Coxeter groups without p-torsion. Next we prove vanishing of the p-local homology of Coxeter groups without p-torsion. Before doing so, we characterize such Coxeter groups in terms of their finite parabolic subgroups. Proposition 5.2. Let p be a prime number. A Coxeter group W has no p-torsion if and only if every finite parabolic subgroup has no p-torsion. Proof. According to a result of Tits, every finite subgroup of W is contained in conjugate of some parabolic subgroup of finite order (see [9, Corollary D.2.9]). The proposition follows at once.  Proposition 5.3. If W is a Coxeter group without p-torsion, then Hk (W, Z(p) ) = 0 for all k > 0. Proof. The claim is obvious for finite Coxeter groups. We prove the proposition for infinite Coxeter groups by the induction on |S|. Let W be an infinite Coxeter group without p-torsion and consider the spectral sequence (4.2). Every proper parabolic subgroup WT of W has no p-torsion, and hence Hk (WT , Z(p) ) = 0 (k > 0) by the induction assumption. This implies Ei,1 j = 0 for j , 0. Moreover Ei,2 j = 0 for (i, j) , (0, 0) by Lemma 4.1, which proves the proposition.  In view of the last proposition, the direct sum decomposition (2.1) can be replaced to the following: Corollary 5.4. For any Coxeter groups W and k > 0, we have Hk (W, Z)  M Hk (W, Z(p) ), p where p runs prime numbers such that W has p-torsion. Remark 5.5. Proposition 5.3 and Corollary 5.4 should be compared with the following general results. Namely, suppose that Γ is a group having finite virtual cohomological dimension vcd Γ. If Γ does not have p-torsion, then H k (Γ, Z)(p) = 0 for k > vcd Γ. Consequently, we have the finite direct product decomposition H k (Γ, Z)  ∏ H k (Γ, Z)(p) p p-LOCAL HOMOLOGY OF COXETER GROUPS 11 which holds for k > vcd Γ, where p ranges over the prime numbers such that Γ has p-torsion. See [8, Chapter X]. 5.3. Right-angled Coxeter groups. A Coxeter group is called right-angled if m(s,t) = 2 or ∞ for all distinct s,t ∈ S. The mod 2 cohomology rings of right-angled Coxeter groups were determined by Rusin [19] (see also [9, Theorem 15.1.4]). In this section, we prove vanishing of p-local homology for a class of Coxeter groups which includes right-angled Coxeter groups. Proposition 5.6. If W is a Coxeter group such that m(s,t) equals to the power of 2 or ∞ for all distinct s,t ∈ S, then Hk (W, Z(p) ) = 0 (k > 0) for all odd prime numbers p ≥ 3. Proof. The finite irreducible Coxeter groups satisfying the assumption are W (A1 ) (of order 2), W (B2 ) (of order 8), and W (I2 (2m ))’s (of order 2m+1 ). Every finite parabolic subgroup of W is isomorphic to a direct product of copies of those groups and hence has the order the power of 2. Consequently, W has no p-torsion by Proposition 5.2. Now the proposition follows from Proposition 5.3.  5.4. 3-free Coxeter groups. In this final section, we look into situations for p = 3 more closely. Firstly, according to Corollary 2.3, H1 (W, Z(3) ) = H2 (W, Z(3) ) = 0 for any Coxeter groups W . This means that Theorem 1.1 remains true for p = 3 without 3-freeness assumption. On the other hand, the finite irreducible 3-free Coxeter groups are W (A1 ),W (B2 ) and W (I2 (q)) such that q is prime to 3, all of which have no 3-torsion. Consequently, every 3-free Coxeter group has no 3-torsion by Proposition 5.2. Applying Proposition 5.3 we obtain the following result: Proposition 5.7. For every 3-free Coxeter group, Hk (W, Z(3) ) = 0 holds for all k > 0. A PPENDIX The following is the table for the Coxeter graph Γ, the order |W (Γ)| of the corresponding Coxeter group W (Γ), the order |W (Γ)| factored into primes, and the range of odd prime numbers p such that W (Γ) is p-free. Γ |W (Γ)| p-freeness A1 2 p≥3 An (n ≥ 2) (n + 1)! p≥5 B2 8 p≥3 n Bn (n ≥ 3) 2 n! p≥5 Dn (n ≥ 4) 2n−1 n! p≥5 E6 72 · 6! 27 · 34 · 5 p≥5 10 4 E7 72 · 8! 2 ·3 ·5·7 p≥5 E8 192 · 10! 214 · 35 · 52 · 7 p≥5 F4 1152 27 · 32 p≥5 H3 120 23 · 3 · 5 p≥7 6 2 2 H4 14400 2 ·3 ·5 p≥7 I2 (q) (q ≥ 3) 2q p6|q 12 T. AKITA Acknowledgement. This study started from questions posed by Takefumi Nosaka concerning of the third p-local homology of Coxeter groups for odd prime numbers p. The author thanks to him for drawing our attention. This study was partially supported by JSPS KAKENHI Grant Numbers 23654018 and 26400077. R EFERENCES [1] Peter Abramenko and Kenneth S. Brown, Buildings, Graduate Texts in Mathematics, vol. 248, Springer, New York, 2008. Theory and applications. MR2439729 (2009g:20055) [2] Toshiyuki Akita, On the cohomology of Coxeter groups and their finite parabolic subgroups, Tokyo J. Math. 18 (1995), no. 1, 151–158, DOI 10.3836/tjm/1270043616. MR1334713 (96f:20078) [3] , On the cohomology of Coxeter groups and their finite parabolic subgroups. II, Group representations: cohomology, group actions and topology (Seattle, WA, 1996), Proc. Sympos. Pure Math., vol. 63, Amer. Math. Soc., Providence, RI, 1998, pp. 1–5, DOI 10.1090/pspum/063/1603123, (to appear in print). MR1603123 (98m:20066) [4] , Aspherical Coxeter groups that are Quillen groups, Bull. London Math. Soc. 32 (2000), no. 1, 85–90, DOI 10.1112/S0024609399006414. MR1718721 (2000j:20068) , Euler characteristics of Coxeter groups, PL-triangulations of closed manifolds, and [5] cohomology of subgroups of Artin groups, J. London Math. Soc. (2) 61 (2000), no. 3, 721–736, DOI 10.1112/S0024610700008693. MR1766100 (2001f:20080) [6] David J. Benson and Stephen D. Smith, Classifying spaces of sporadic groups, Mathematical Surveys and Monographs, vol. 147, American Mathematical Society, Providence, RI, 2008. MR2378355 (2009f:55017) [7] Nicolas Bourbaki, Éléments de mathématique, Masson, Paris, 1981 (French). Groupes et algèbres de Lie. Chapitres 4, 5 et 6. [Lie groups and Lie algebras. Chapters 4, 5 and 6]. MR647314 (83g:17001) [8] Kenneth S. Brown, Cohomology of groups, Graduate Texts in Mathematics, vol. 87, SpringerVerlag, New York, 1982. MR672956 (83k:20002) [9] Michael W. Davis, The geometry and topology of Coxeter groups, London Mathematical Society Monographs Series, vol. 32, Princeton University Press, Princeton, NJ, 2008. MR2360474 (2008k:20091) [10] C. De Concini and M. Salvetti, Cohomology of Coxeter groups and Artin groups, Math. Res. Lett. 7 (2000), no. 2-3, 213–232, DOI 10.4310/MRL.2000.v7.n2.a7. MR1764318 (2001f:20118) [11] Mohamed Errokh and Fulvio Grazzini, Sur la cohomologie modulo 2 des groupes de Coxeter à trois générateurs, C. R. Acad. Sci. Paris Sér. I Math. 324 (1997), no. 7, 741–745, DOI 10.1016/S0764-4442(97)86937-6 (French, with English and French summaries). MR1446573 (98h:20094) [12] L. C. Grove and C. T. Benson, Finite reflection groups, 2nd ed., Graduate Texts in Mathematics, vol. 99, Springer-Verlag, New York, 1985. MR777684 (85m:20001) [13] Robert B. Howlett, On the Schur multipliers of Coxeter groups, J. London Math. Soc. (2) 38 (1988), no. 2, 263–276, DOI 10.1112/jlms/s2-38.2.263. MR966298 (90e:20010) [14] James E. Humphreys, Reflection groups and Coxeter groups, Cambridge Studies in Advanced Mathematics, vol. 29, Cambridge University Press, Cambridge, 1990. MR1066460 (92h:20002) [15] Shin-ichiro Ihara and Takeo Yokonuma, On the second cohomology groups (Schur-multipliers) of finite reflection groups, J. Fac. Sci. Univ. Tokyo Sect. I 11 (1965), 155–171 (1965). MR0190232 (32 #7646a) [16] Minoru Nakaoka, Decomposition theorem for homology groups of symmetric groups, Ann. of Math. (2) 71 (1960), 16–42. MR0112134 (22 #2989) [17] , Homology of the infinite symmetric group, Ann. of Math. (2) 73 (1961), 229–257. MR0131874 (24 #A1721) p-LOCAL HOMOLOGY OF COXETER GROUPS 13 [18] Stephen J. Pride and Ralph Stöhr, The (co)homology of aspherical Coxeter groups, J. London Math. Soc. (2) 42 (1990), no. 1, 49–63, DOI 10.1112/jlms/s2-42.1.49. MR1078174 (91k:20058) [19] David John Rusin, The cohomology of groups generated by involutions, ProQuest LLC, Ann Arbor, MI, 1984. Thesis (Ph.D.)–The University of Chicago. MR2611843 [20] Mario Salvetti, Cohomology of Coxeter groups, Topology Appl. 118 (2002), no. 1-2, 199–208, DOI 10.1016/S0166-8641(01)00051-7. Arrangements in Boston: a Conference on Hyperplane Arrangements (1999). MR1877725 (2003d:20073) [21] Jean-Pierre Serre, Cohomologie des groupes discrets, Prospects in mathematics (Proc. Sympos., Princeton Univ., Princeton, N.J., 1970), Princeton Univ. Press, Princeton, N.J., 1971, pp. 77– 169. Ann. of Math. Studies, No. 70 (French). MR0385006 (52 #5876) [22] Richard G. Swan, The p-period of a finite group, Illinois J. Math. 4 (1960), 341–346. MR0122856 (23 #A188) [23] James Andrew Swenson, The mod-2 cohomology of finite Coxeter groups, ProQuest LLC, Ann Arbor, MI, 2006. Thesis (Ph.D.)–University of Minnesota. MR2709085 [24] C. B. Thomas, Characteristic classes and the cohomology of finite groups, Cambridge Studies in Advanced Mathematics, vol. 9, Cambridge University Press, Cambridge, 1986. MR878978 (88f:20005) [25] Takeo Yokonuma, On the second cohomology groups (Schur-multipliers) of infinite discrete reflection groups, J. Fac. Sci. Univ. Tokyo Sect. I 11 (1965), 173–186 (1965). MR0190233 (32 #7646b) D EPARTMENT OF M ATHEMATICS , H OKKAIDO U NIVERSITY, S APPORO , 060-0810 JAPAN E-mail address: akita@math.sci.hokudai.ac.jp 

DeepLesion: Automated Deep Mining, Categorization and Detection of Significant Radiology Image Findings using Large-Scale Clinical Lesion Annotations arXiv:1710.01766v2 [cs.CV] 10 Oct 2017 Ke Yan? , Xiaosong Wang? , Le Lu, and Ronald M. Summers Department of Radiology and Imaging Sciences, National Institutes of Health Clinical Center, Bethesda, MD 20892 {ke.yan, xiaosong.wang, le.lu, rms}@nih.gov Abstract. Extracting, harvesting and building large-scale annotated radiological image datasets is a greatly important yet challenging problem. It is also the bottleneck to designing more effective data-hungry computing paradigms (e.g., deep learning) for medical image analysis. Yet, vast amounts of clinical annotations (usually associated with disease image findings and marked using arrows, lines, lesion diameters, segmentation, etc.) have been collected over several decades and stored in hospitals’ Picture Archiving and Communication Systems. In this paper, we mine and harvest one major type of clinical annotation data – lesion diameters annotated on bookmarked images – to learn an effective multi-class lesion detector via unsupervised and supervised deep Convolutional Neural Networks (CNN). Our dataset is composed of 33,688 bookmarked radiology images from 10,825 studies of 4,477 unique patients. For every bookmarked image, a bounding box is created to cover the target lesion based on its measured diameters. We categorize the collection of lesions using an unsupervised deep mining scheme to generate clustered pseudo lesion labels. Next, we adopt a regional-CNN method to detect lesions of multiple categories, regardless of missing annotations (normally only one lesion is annotated, despite the presence of multiple co-existing findings). Our integrated mining, categorization and detection framework is validated with promising empirical results, as a scalable, universal or multi-purpose CAD paradigm built upon abundant retrospective medical data. Furthermore, we demonstrate that detection accuracy can be significantly improved by incorporating pseudo lesion labels (e.g., Liver lesion/tumor, Lung nodule/tumor, Abdomen lesions, Chest lymph node and others). This dataset will be made publicly available (under the open science initiative). 1 Introduction Computer-aided detection/diagnosis (CADe/CADx) has been a highly prosperous and successful research field in medical image processing. Many commercial software packages have been developed for clinical usage and screening. Recent advances (e.g., automated classification of skin lesions [3], detection of liver lesion [1], pulmonary embolism [10]) have attracted even more attention to the application of deep learning paradigms to CADe/CADx. Deep learning, namely Convolutional Neural Network ? These two authors contributed equally. 2 (CNN) based algorithms, perform significantly better than conventional statistical learning approaches combined with hand-crafted image features. However, these performance gains are often achieved at the cost of requiring tremendous amounts of training data accompanied with high quality labels. Unlike general computer vision tasks, medical image analysis currently lacks a substantial, large-scale annotated image dataset (comparable to ImageNet [2] and MS COCO [6]),for two main reasons: 1) The conventional methods for collecting image labels via Google search + crowd-sourcing from average users cannot be applied in the medical image domain, as medical image annotation reuqires extensive clinical expertise; 2) Significant inter and intra-observer variability (among even well-trained, experienced radiologists) frequently occurs, and thus may compromise reliable annotation of a large amount of medical images, especially considering the great diversity of radiology diagnosis tasks. Current CADe/CADx methods generally target one particular type of diseases or lesions, such as lung nodules, colon polyps or lymph nodes [7]. Yet, this approach differs from the methods radiologists routinely apply to read medical image studies and compile radiological reports. Multiple findings can be observed and are often correlated. For instance, liver metastases can spread to regional lymph nodes or other body parts. By obtaining and maintaining a holistic picture of relevant clinical findings, a radiologist will be able to make a more accurate diagnosis. However, it remains greatly challenging to develop a universal or multi-purpose CAD framework, capable of detecting multiple disease types in a seamless fashion. Such a framework is crucial to building an automatic radiological diagnosis and reasoning system. In this paper, we attempt to address these challenges by first introducing a new large-scale dataset of bookmarked radiology images, which accommodate lesions from multiple categories. Our dataset, named DeepLesion, is composed of 33,688 bookmarked images from 10,825 studies of 4,477 patients (see samples in Fig. 1). For each bookmarked image, a bounding box is generated to indicate the location of the lesions. Furthermore, we integrate an unsupervised deep mining method to compute pseudo image labels for database self-annotating. Categories of Liver lesion/tumor, Lung nodule/tumor, Abdomen lesions, Chest lymph node and others are identified by our computerized algorithm instead of radiologists’ annotation, which may be infeasible. After obtaining the dataset, we develop an automatic lesion detection approach to jointly localize and classify lesion candidates using discovered multiple categories. Last, how the unsupervisedly-learned pseudo lesion labels affect the deep CNN training strategies and the quantitative performance of our proposed multi-class lesion detector is investigated. 2 Methods In this section, we first describe how our DeepLesion dataset is constructed. Next, we propose an unsupervised deep learning method to mine the latent lesion categories in each image. This method involves an iterative process of deep image feature extraction, image clustering and CNN model retraining. Finally, we present a multi-class object detection approach to detect lesions of multiple categories. 3 (a) (b) (c) Fig. 1. Three sample bookmarked images illustrated with annotation lesion patches (i.e., yellow dashed boxes). The outputs from our proposed multi-category lesion detection framework are shown in colored boxes with LiVer lesion (LV) in Red, Lung Nodule (LN) in Orange, ABdomen lesion (AB) in Green, Chest Lymph node (CL) in magenta and other MiXed lesions (MX) in blue. (a) A spleen metastasis is correctly detected along with several liver and abdomen metastases; (b) Two large lymph nodes in mediastinum are all correctly detected; (c) All three lung nodules are detected despite two small ones not being annotated in this bookmarked image. 2.1 DeepLesion Dataset Radiologists routinely annotate hundreds of clinically meaningful findings in medical images, using arrows, lines, diameters or segmentations to highlight and measure different disease patterns to be reported. These images, called “bookmarked images”, have been collected over close to two decades in our institute’s Picture Archiving and Communication Systems (PACS). Without loss of generality, in this work, we study one type of bookmark in CT images: lesion diameters. Each pair of lesion diameters consists of two lines, one measuring the longest diameter and the second measuring its longest perpendicular diameter in the plane of measurement. We extract the lesion diameter coordinates from the PACS server and convert into corresponding positions in the image plane coordinates, noted as {(x11 , y11 ), (x12 , y12 )}; {(x21 , y21 ), (x22 , y22 )}. A bounding box (lef tx , topy , width, height) is computed to cover a rectangular area enclosing the lesion measurement with 20 pixel padding in each direction, i.e., (xmin −20, ymin − 20, xmax − xmin + 40, ymax − ymin + 40) where xmin = M in(x11 , x12 , x21 , x22 ) and xmax = M ax(x11 , x12 , x21 , x22 ), and similarly for ymin and ymax . The padding range can capture the lesion’s full spatial extent with sufficient image context. We thus generate 33,688 bookmarked radiology images from 10,825 studies of 4,477 unique patients, and each bookmarked image is associated with a bounding box annotation of the enclosed lesion. Sample bookmarked images and bounding boxes are shown in Fig. 1. 2.2 Unsupervised Lesion Categorization The images in our constructed lesion dataset contain several types of lesions commonly observed by radiologists, such as lung nodule/lesion, lymph node, and liver/kidney lesion. However, no detailed precise category labels for each measured lesion have been provided. Obtaining such from radiologists would be highly tedious and timeconsuming, due to the vast size and comprehensiveness of DeepLesion. To address this 4 Fig. 2. Lesion categorization framework via unsupervised and iteratively-optimized deep CNNs. problem, we propose a looped deep optimization procedure for automated category discovery, which generates visually coherent and clinically-semantic image clusters. Our algorithm is conceptually simple: it is based on the hypothesis that the optimization procedure will “converge” to more accurate labels, which will lead to better trained CNN models. Such models, in turn, will generate more representative deep image features, which will allow for creating more meaningful lesion labels via clustering. As a pre-processing step, we crop the lesion patches from the original DICOM slides using the dilated bounding boxes (described in Sec. 2.1) and resize them, prior to feeding them into the CNN model. As shown in Fig. 2, our iterative deep learning process begins by extracting deep CNN features for each lesion patch using the ImageNet [2]) pre-trained VGG-16 [9] network. Next, it applies k-means clustering to the deep feature encoded lesion patches after k is determined via model selection [5]. Next, it fine-tunes the current VGG-16 using the new image labels obtained from k-means. This yields an updated CNN model for the next iteration. The optimization cycle terminates once the convergence criteria have been satisfied. Encoding Lesion Patches using Deep CNN Features: The VGG-16 [9] CNN architecture is adopted for patch encoding and CNN model fine-tuning to facilitate the iterative procedure. The image features extracted from the last fully-connected layer (e.g., FC6/FC7 of VGG-16) are used, as they are able to capture both the visual appearance and the spatial layout of any lesion, with its surrounding context. Convergence in Patch Clustering and Categorization: We hypothesize that the newly generated clusters will converge to the “oracle” label clusters, after undergoing several staged of cluster optimization. Two convergence measurements are employed: Purity [5] and Normalized Mutual Information (NMI). We assess both criteria by computing empirical similarity scores between clustering results from two adjacent iterations. If the similarity score exceeds a pre-defined threshold, the optimal clustering driven categorization of lesion patches has been attained. For each iteration, we randomly shuffle the lesion patches and divide the data into three subsets: training (75%), validation (10%) and testing (15%). Therefore the “improving-then-saturating” trajectory of the CNN classification accuracy on the testing set can also indicate the convergence of the clustering labels (i.e., optimal image labels have been obtained). 2.3 Multi-category Lesion Detection Using the bounding boxes (Sec. 2.1) and their corresponding newly generated pseudocategory labels (Sec. 2.2), we develop a multi-class lesion detector adapted from the 5 Fig. 3. Flow chart of the lesion detection algorithm. Bookmarked clinical annotations provide the ground-truth bounding boxes of lesions for detector training. In detection, the dashed and solid boxes indicate the ground-truth annotation and its predicted lesion detection, respectively. Faster RCNN method [8]. An input image is first processed by several convolutional and max pooling layers to produce feature maps, as shown in Fig. 3. Next, a region proposal network (RPN) parses the feature maps and proposes candidate lesion regions. It estimates the probability of “target/non-target” on a fixed set of anchors (candidate regions) on each position of the feature maps. Furthermore, the location and size of each anchor are fine-tuned via bounding box regression. Afterwards, the region proposals and the feature maps are sent to a Region of Interest (RoI) pooling layer, which resamples the feature maps inside each proposal to a fixed size (we use 7×7 here). These feature maps are then fed into several fully-connected layers that predict the confidence scores for each lesion class and run another bounding box regression for further finetuning. Non-maximum suppression (NMS) is then applied to the feature maps. Finally, the system returns up to five detection proposals with the highest confidence scores (> 0.5), as each image only has one bookmarked clinical annotation. The ImageNet pretrained VGG-16 [9] model is adopted as the backbone of Faster RCNN [8]. It is useful to remove the last pooling layer (pool4) in VGG-16 to enhance the resolution of the feature map and to increase the sampling ratio of positive samples (candidate regions that contain lesions). In our experiments, removing pool4 improves the accuracy by ∼ 15%. It is critical to set the anchor sizes in RPN to fit the size of ground-truth bounding boxes in DeepLesion dataset. Hence, we use anchors of three scales (48, 72, 96) and aspect ratios (1:1, 1:2, 2:1) to cover most of the boxes. For image preparation, we use the ground-truth lesion bounding boxes derived in Sec. 2.1 incorporating enlarged spatial contexts. Each full-slice image in the detection phase is resized, so that the longest dimension is of 512 pixels. We then train the network as demonstrated in Fig. 3 in a multi-task fashion: two classification and two regression losses are jointly optimized. This end-to-end training strategy is more efficient than the four-step method in the original Faster RCNN implementation [8]. During training, each mini-batch has 4 images, and the number of region proposals per image is 32. We use the Caffe toolbox and the Stochastic Gradient Descent (SGD) optimizer. The base 6 (a) (b) (c) (d) (e) (f) Fig. 4. Six sample detection results are illustrated with the annotation lesion patches as yellow dashed boxes. The outputs from our proposed detection framework are shown in colored boxes with LiVer lesion (LV) in Red, Lung Nodule (LN) in Orange, ABdomen lesion (AB) in Green, Chest Lymph node (CL) in magenta and other MiXed lesions (MX) in blue. (a) Four lung lesions are all correctly detected; (b) Two lymph nodes in mediastinum is presented; (c) A Ground Glass Opacity (GGO) and a mass are detected in the lung; (d) An adrenal nodule; (e) Correct detections on both the small abdomen lymph node nearly aorta but also other metastases in liver and spleen; (f) Two liver metastasis are correctly detected. Three lung metastases are detected but erroneously classified as liver lesions . learning rate is set to 0.001, and is reduced by a factor of 10 every 20K iterations. The network generally converges within 60K iterations. 3 Results and Discussion Our lesion categorization method in Sec. 2.2 partitions all lesion patches into five classes k = 5. After visual inspection supervised by a board-certificated radiologist, four common lesion categories are found, namely lung nodule/lesion, liver lesion, chest lymph nodes and abdominal lesions (mainly kidney lesions and lymph nodes), with high purity scores in each category (0.980 for Lung Nodule, 0.955 for Chest Lymph Node, 0.805 for Liver Lesion and 0.995 for Abdomen Lesion). The per-cluster purity scores are estimated through a visual assessment by an experienced radiologist using a set of 800 randomly selected lesion patches (200 images per-category). The remaining bookmarked annotations are treated as a “noisy” mixed lesion class. Our optimization framework converges after six iterations, with a high purity score of 92.3% returned when assessing the statistical similarity or stability of two last iterations. Meanwhile, the top-1 classification accuracy reaches 91.6%, and later fluctuates by ±2% . 7 1 0.8 0.7 0.5 Precision Detection accuracy 0.8 0.6 0.4 0.3 Single-class lesion detection Multi-class lesion detection 0.2 0.6 Single-class lesion detection Liver lesion detection Lung nodule detection Abdomen lesions detection Chest lymph node detection Mixed lesions detection 0.4 0.2 0.1 0 0.1 0 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.9 0.2 0.4 0.6 0.8 1 Recall IoU (a) (b) Fig. 5. (a): Detection accuracy curves with different intersection-over-union (IoU) thresholds. (b): Precision-Recall curves of single-class and five category lesion detection when IoU=0.5. Cluster 1 2 3 4 5 Cluster category Liver lesion Lung nodule abdomen lesions Chest lymph node Mixed lesions Overall Cluster size 774 837 1270 860 1292 5033 Averaged Accuracy: w/o categorization labels (%) 56.04 73.30 48.79 70.77 54.85 59.45 Averaged accuracy: with categorization labels (%) 60.59 76.34 56.22 76.28 58.67 64.30 Table 1. Test detection accuracies of five deep-learned pseudo lesion categories. Note that in DeepLesion dataset, available clinical annotations are only partial (i.e., missing lesion labels). Hence the actual detection accuracies in both configurations are significantly higher than the above reported values since many “false positive” detections are later verified to be true lesions. For lesion detection, all bookmarked images are divided into training (70%), validation (15%), and testing (15%) sets, by random splitting the dataset at the patient level. Although different lesion types may be present in an image, only one clinical annotation per image is available. We adopt a straightforward evaluation criterion: 1) we take the top one detected lesion candidate box (with the highest detection confidence score) per testing image as the detection result; 2) if the intersection-over-union (IoU) between this predicted box and the ground-truth clinical annotation box is larger than 0.5 (as suggested by the PASCAL criterion [4]), the detection is regarded as correct, and vice versa. The lesion category is not considered in this criterion. We denote this evaluation metric as detection accuracy. The proposed multi-class lesion detector merely requires 88 ms to process a 512×512 test bookmarked image on a Titan X Maxwell GPU. Two lesion detection setups or configurations are examined: single-class (all annotation bounding boxes are considered as one abnormality class), and multi-class detection (with pseudo-category labels). Some illustrative results from the multi-category lesion detector on the testing set are shown in Fig. 4. It can be found that our developed detector is able to detect all five types of lesions and simultaneously provide the corresponding lesion category labels. Furthermore, some detection boxes currently considered as 8 false alarms actually belong to true lesions because the lesions bounding boxes are only partially labeled by clinical bookmark annotations. Detailed statistics of the five deeply discovered lesion clusters in the test set are provided in Table 1. This outlines the types of lesions in the clusters that have been verified by radiologists. The averaged accuracy of the single-class detection is 59.45% (testing) and this score becomes 64.3% for multiclass detection (testing). From Table 1, the multi-category detector also demonstrates accuracy improvements of 3∼8% per lesion cluster or category compared against the one-class abnormality detector. Single-class abnormality detection appears to be a more challenging task since it tackles detecting various types of lesions at once. This validates that better lesion detection models can be trained if we can perform unsupervised lesion categorization from a large collection of retrospective clinical data. The default IoU threshold is set as 0.5. Fig. 5 (a) illustrates the detection accuracy curves of both detection models under different IoU thresholds. The multi-category lesion detection achieves the better overall accuracy while being able to assign the lesion labels at the same time. Fig. 5 (b) shows the corresponding detection precision-recall curves. The performances of lung lesion detection and chest lymph node detection significantly outperform the one-class abnormality detection. 4 Conclusion In this paper, we mine, categorize and detect one type of clinical annotations stored in the hospital PACS system as a rich retrospective data source, to build a large-scale Radiology lesion image database. We demonstrate the strong feasibility of employing a new multi-category lesion detection paradigm via unified deep categorization and detection. Highly promising lesion categorization and detection performances, based on the proposed dataset, are reported. To the best of our knowledge, this work is the first attempt of building a scalable, multi-purpose CAD system using abundant retrospective medical data. This is done almost effortlessly since no new arduous image annotation workload is necessary. Our future work include extending bookmarked images to incorporate their successive slices for the scalable and precise lesion volume measurement; extracting and integrating the lesion diagnosis prior from radiology text reports, and improved multi-category detection methods. References 1. A. Ben-Cohen, I. Diamant, E. Klang, M. Amitai, and H. Greenspan. Fully convolutional network for liver segmentation and lesions detection. In MICCAI LABELS-DLMIA, 2016. 2. J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In IEEE CVPR, pages 248–255, 2009. 3. A. Esteva, B. Kuprel, R. A. Novoa, J. Ko, S. M. Swetter, H. M. Blau, and S. Thrun. Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542(7639):115–118, 2017. 4. M. Everingham, A. Eslami, L. Van Gool, C. Williams, J. Winn, and A. Zisserman. The pascal visual object classes challenge: A retrospective. Int. J. Comp. Vis., 111(1):98–136, 2015. 5. R. Gomes, A. Krause, and P. Perona. Discriminative clustering by regularized information maximization. In NIPS, pages 775–783, 2010. 9 6. T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft coco: Common objects in context. In ECCV, pages 740–755, 2014. 7. J. Liu, D. Wang, Z. Wei, L. Lu, L. Kim, E. Turkbey, and R. M. Summers. Colitis detection on computed tomography using regional convolutional neural networks. In IEEE ISBI, 2016. 8. S. Ren, K. He, R. Girshick, and J. Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NIPS, pages 91–99, 2015. 9. K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR. arxiv.org/abs/1409.1556, 2015. 10. N. Tajbakhsh, M. B. Gotway, and J. Liang. Computer-aided pulmonary embolism detection using a novel vessel-aligned multi-planar image representation and convolutional neural networks. In MICCAI, pages 62–69. Springer, 2015. 

SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES OF BIDEGREE (2, 1) arXiv:1211.1648v1 [math.AC] 7 Nov 2012 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI Abstract. Let U ⊆ H 0 (OP1 ×P1 (2, 1)) be a basepoint free four-dimensional vector space. The sections corresponding to U determine a regular map φU : P1 × P1 −→ P3 . We study the associated bigraded ideal IU ⊆ k[s, t; u, v] from the standpoint of commutative algebra, proving that there are exactly six numerical types of possible bigraded minimal free resolution. These resolutions play a key role in determining the implicit equation for φU (P1 × P1 ), via work of Busé-Jouanolou [5], Busé-Chardin [6], Botbol [2] and Botbol-DickensteinDohm [3] on the approximation complex Z. In four of the six cases IU has a linear first syzygy; remarkably from this we obtain all differentials in the minimal free resolution. In particular this allows us to explicitly describe the implicit equation and singular locus of the image. 1. Introduction A central problem in geometric modeling is to find simple (determinantal or close to it) equations for the image of a curve or surface defined by a regular or rational map. For surfaces the two most common situations are when P1 × P1 −→ P3 or P2 −→ P3 . Surfaces of the first type are called tensor product surfaces and surfaces of the latter type are called triangular surfaces. In this paper we study tensor product surfaces of bidegree (2, 1) in P3 . The study of such surfaces goes back to the last century–see, for example, works of Edge [17] and Salmon [26]. Let R = k[s, t, u, v] be a bigraded ring over an algebraically closed field k, with s, t of degree (1, 0) and u, v of degree (0, 1). Let Rm,n denote the graded piece in bidegree (m, n). A regular map P1 × P1 −→ P3 is defined by four polynomials U = Span{p0 , p1 , p2 , p3 } ⊆ Rm,n with no common zeros on P1 × P1 . We will study the case (m, n) = (2, 1), so U ⊆ H 0 (OP1 ×P1 (2, 1)) = V = Span{s2 u, stu, t2 u, s2 v, stv, t2 v}. Let IU = hp0 , p1 , p2 , p3 i ⊂ R, φU be the associated map P1 × P1 −→ P3 and XU = φU (P1 × P1 ) ⊆ P3 . We assume that U is basepoint free, which means that p IU = hs, ti ∩ hu, vi. We determine all possible numerical types of bigraded minimal free resolution for IU , as well as the embedded associated primes of IU . Using approximation complexes, we relate the algebraic properties of IU to the geometry of XU . The next example illustrates our results. Key words and phrases. Tensor product surface, bihomogeneous ideal, Segre-Veronese map. Schenck supported by NSF 1068754, NSA H98230-11-1-0170. 1 2 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI Example 1.1. Suppose U is basepoint free and IU has a unique first syzygy of bidegree (0, 1). Then the primary decomposition of IU is given by Corollary 3.5, and the differentials in the bigraded minimal free resolution are given by Proposition 3.2. For example, if U = Span{s2 u, s2 v, t2 u, t2 v + stv}, then by Corollary 3.5 and Theorem 3.3, the embedded primes of IU are hs, t, ui and hs, t, vi, and by Proposition 3.2 the bigraded Betti numbers of IU are: 0 ← IU ← R(−2, −1)4 ← R(−2, −2)⊕R(−3, −2)2 ⊕R(−4, −1)2 ← R(−4, −2)2 ← 0 Having the differentials in the free resolution allows us to use the method of approximation complexes to determine the implicit equation: it follows from Theorem 7.1 that the image of φU is the hypersurface XU = V(x0 x21 x2 − x21 x22 + 2x0 x1 x2 x3 − x20 x23 ). Theorem 7.3 shows that the reduced codimension one singular locus of XU is V(x0 , x2 ) ∪ V(x1 , x3 ) ∪ V(x0 , x1 ). The key feature of this example is that there is Figure 1. XU on the open set Ux0 a linear syzygy of bidegree (0, 1): v · (s2 u) − u · (s2 v) = 0. In Lemmas 3.1 and 4.1 we show that with an appropriate choice of generators for IU , any bigraded linear first syzygy has the form above. Existence of a bidegree (0, 1) syzygy implies that the pullbacks to P1 × P1 of the two linear forms defining P(U ) share a factor. Theorem 8.5 connects this to work of [19]. 1.1. Previous work on the (2, 1) case. For surfaces in P3 of bidegree (2, 1), in addition to the classical work of Edge, Salmon and others, more recently Degan [13] studied such surfaces with basepoints and Zube [30], [31] describes the possibilities for the singular locus. In [18], Elkadi-Galligo-Lê give a geometric description of the image and singular locus for a generic U and in [19], Galligo-Lê follow up with an analysis for the nongeneric case. A central part of their analysis is the geometry of a certain dual scroll which we connect to syzygies in §8. Cox, Dickenstein and Schenck study the bigraded commutative algebra of a three dimensional basepoint free subspace W ⊆ R2,1 in [11], showing that there are two numerical types of possible bigraded minimal free resolution of IW , determined by σ2,1 how P(W ) ⊆ P(R2,1 ) = P5 meets the image Σ2,1 of the Segre map P2 × P1 −→ P5 . If W is basepoint free, then there are two possibilities: either P(W ) ∩ Σ2,1 is a finite set of points, or a smooth conic. The current paper extends the work of [11] to the more complicated setting of a four dimensional space of sections. A key difference SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 3 is that for a basepoint free subspace W of dimension three, there can never be a linear syzygy on IW . As illustrated in the example above, this is not true for the four dimensional case. It turns out that the existence of a linear syzygy provides a very powerful tool for analyzing both the bigraded commutative algebra of IU , as well as for determining the implicit equation and singular locus of XU . In studying the bigraded commutative algebra of IU , we employ a wide range of tools • Approximation complexes [2], [3], [5], [6], [7]. • Bigraded generic initial ideals [1]. • Geometry of the Segre-Veronese variety [22]. • Fitting ideals and Mapping cones [15]. • Connection between associated primes and Ext modules [16]. • Buchsbaum-Eisenbud exactness criterion [4]. 1.2. Approximation complexes. The key tool in connecting the syzygies of IU to the implicit equation for XU is an approximation complex, introduced by HerzogSimis-Vasconcelos in [23],[24]. We give more details of the construction in §7. The basic idea is as follows: let RI = R⊕IU ⊕IU2 ⊕· · · . Then the graph Γ of the map φU is equal to BiP roj(RI ) and the embedding of Γ in (P1 × P1 ) × P(U ) corresponds s to the ring map S = R[x0 , . . . , x3 ] → RI given by xi 7→ pi . Let β denote the kernel of s, so β1 consists of the syzygies of IU and SI = SymR (I) = S/β1 . Then Γ ⊆ BiP roj(SI ) ⊆ BiP roj(S). The works [3], [5], [6], [2] show that if U is basepoint free, then the implicit equation for XU may be extracted from the differentials of a complex Z associated to the intermediate object SI and in particular the determinant of the complex is a power of the implicit equation. In bidegree (2, 1), a result of Botbol [2] shows that the implicit equation may be obtained from a 4 × 4 minor of d1 ; our work yields an explicit description of the relevant minor. 1.3. Main results. The following two tables describe our classification. Type refers to the graded Betti numbers of the bigraded minimal free resolution for IU : we prove there are six numerical types possible. Proposition 6.3 shows that the only possible embedded primes of IU are m = hs, t, u, vi or Pi = hli , s, ti, where li is a linear form of bidegree (0, 1). While Type 5a and 5b have the same bigraded Betti numbers, Proposition 3.2 and Corollary 3.5 show that both the embedded primes and the differentials in the minimal resolution differ. We also connect our classification to the reduced, codimension one singular locus of XU . In the table below T denotes a twisted cubic curve, C a smooth plane conic and Li a line. Type 1 2 3 4 5a 5b 6 Lin. Syz. Emb. Pri. Sing. Loc. Example none m T s2 u+stv, t2 u, s2 v+stu, t2 v+stv none m, P1 C ∪ L1 s2 u, t2 u, s2 v + stu, t2 v + stv 1 type (1, 0) m L1 s2 u + stv, t2 u, s2 v, t2 v + stu 1 type (1, 0) m, P1 L1 stv, t2 v, s2 v − t2 u, s2 u 1 type (0, 1) P1 , P2 L1 ∪ L2 ∪ L3 s2 u, s2 v, t2 u, t2 v + stv 1 type (0, 1) P1 L1 ∪ L2 s2 u, s2 v, t2 u, t2 v + stu 2 type (0, 1) none ∅ s2 u, s2 v, t2 u, t2 v Table 1. 4 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI The next table gives the possible numerical types for the bigraded minimal free resolutions, where we write (i, j) for the rank one free module R(i, j). We prove more: for Types 3, 4, 5 and 6, we determine all the differentials in the minimal free resolution. One striking feature of Table 1 is that if IU has a linear first syzygy (i.e. of bidegree (0, 1) or (1, 0)), then the codimension one singular locus of XU is either empty or a union of lines. We prove this in Theorem 7.3. Type 1 2 3 4 5 6 Bigraded Minimal Free Resolution of IU for U basepoint free (−2, −4) ⊕ (−3, −4)2 ⊕ 0 ← IU ← (−2, −1)4 ←− (−3, −2)4 ←− ←− (−4, −4) ← 0 ⊕ (−4, −2)3 (−4, −1)2 (−2, −3) ⊕ (−3, −3)2 ⊕ 0 ← IU ← (−2, −1)4 ←− (−3, −2)4 ←− ←− (−4, −3) ← 0 ⊕ (−4, −2)3 (−4, −1)2 (−2, −4) ⊕ (−3, −1) (−3, −4)2 ⊕ ⊕ (−4, −4) (−3, −2)2 ⊕ ⊕ ←− (−4, −3)2 ←− ←0 0 ← IU ← (−2, −1)4 ←− ⊕ (−5, −3) (−3, −3) (−5, −2)2 ⊕ (−4, −2) ⊕ (−5, −1) (−2, −3) ⊕ (−3, −1) (−3, −3) ⊕ ⊕ 0 ← IU ← (−2, −1)4 ←− (−3, −2)2 ←− (−4, −3) ←− (−5, −3) ← 0 ⊕ ⊕ (−4, −2) (−5, −2)2 ⊕ (−5, −1) (−2, −2) ⊕ 0 ← IU ← (−2, −1)4 ←− (−3, −2)2 ←− (−4, −2)2 ← 0 ⊕ (−4, −1)2 (−2, −2)2 4 ⊕ 0 ← IU ← (−2, −1) ←− ←− (−4, −2) ← 0 (−4, −1)2 Table 2. SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 5 2. Geometry and the Segre-Veronese variety Consider the composite maps (2.1) P1 × P1 / P(H 0 (OP1 (2))) × P(H 0 (OP1 (1))) / P(H 0 (OP1 ×P1 (2, 1))) φU π  , P(U ) The first horizontal map is ν2 × id, where ν2 is the 2-uple Veronese embedding and the second horizontal map is the Segre map σ2,1 : P2 × P1 → P5 . The image of σ2,1 is a smooth irreducible nondegenerate cubic threefold Σ2,1 . Any P2 ⊆ Σ2,1 is a fiber over a point of the P1 factor and any P1 ⊆ Σ2,1 is contained in the image of a fiber over P2 or P1 . For this see Chapter 2 of [21], which also points out that the Segre and Veronese maps have coordinate free descriptions ν d P(A) −→ σ P(A) × P(B) −→ P(Symd A) P(A ⊗ B) By dualizing we may interpret the image of νd as the variety of dth powers of linear forms on A and the image of σ as the variety of products of linear forms. The composition τ = σ2,1 ◦ (ν2 × id) is a Segre-Veronese map, with image consisting of polynomials which factor as l1 (s, t)2 · l2 (u, v). Note that Σ2,1 is also the locus of polynomials in R2,1 which factor as q(s, t) · l(u, v), with q ∈ R2,0 and l ∈ R0,1 . Since q ∈ R2,0 factors as l1 · l2 , this means Σ2,1 is the locus of polynomials in R2,1 which factor completely as products of linear forms. As in the introduction, U ⊆ H 0 (OP1 ×P1 (2, 1)) = V = Span{s2 u, stu, t2 u, s2 v, stv, t2 v}. The ideal of Σ2,1 is defined by the two by two minors of   x0 x1 x2 . x3 x4 x5 It will also be useful to understand the intersection of P(U ) with the locus of polynomials in V which factor as the product of a form q = a0 su+a1 sv+a2 tu+a3 tv of bidegree (1, 1) and l = b0 s + b1 t of bidegree (1, 0). This is the image of the map P(H 0 (OP1 ×P1 (1, 1))) × P(H 0 (OP1 ×P1 (1, 0))) = P3 × P1 −→ P5 , (a0 : a1 : a2 : a3 ) × (b0 : b1 ) 7→ (a0 b0 : a0 b1 + a2 b0 : a2 b1 : a1 b0 : a1 b1 + a3 b0 : a3 b1 ), which is a quartic hypersurface Q = V(x22 x23 − x1 x2 x3 x4 + x0 x2 x24 + x21 x3 x5 − 2x0 x2 x3 x5 − x0 x1 x4 x5 + x20 x25 ). As Table 1 shows, the key to classifying the minimal free resolutions is understanding the linear syzygies. In §3, we show that if IU has a first syzygy of bidegree (0, 1), then after a change of coordinates, IU = hpu, pv, p2 , p3 i and if IU has a first syzygy of bidegree (1, 0), then IU = hps, pt, p2 , p3 i. Proposition 2.1. If U is basepoint free, then the ideal IU (1) has a unique linear syzygy of bidegree (0, 1) iff F ⊆ P(U ) ∩ Σ2,1 , where F is a P1 fiber of Σ2,1 . (2) has a pair of linear syzygies of bidegree (0, 1) iff P(U ) ∩ Σ2,1 = Σ1,1 . (3) has a unique linear syzygy of bidegree (1, 0) iff F ⊆ P(U ) ∩ Q, where F is a P1 fiber of Q. 6 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI Proof. The ideal IU has a unique linear syzygy of bidegree (0, 1) iff qu, qv ∈ IU , with q ∈ R2,0 iff q · l(u, v) ∈ IU for all l(u, v) ∈ R0,1 iff P(U ) ∩ Σ2,1 contains the P1 fiber over the point q ∈ P(R2,0 ). For the second item, the reasoning above implies that P(U ) ∩ Σ2,1 contains two P1 fibers, over points q1 , q2 ∈ P(R2,0 ). But then IU also contains the line in P(R2,0 ) connecting q1 and q2 , as well as the P1 lying over any point on the line, yielding a P1 × P1 . For the third part, a linear syzygy of bidegree (1, 0) means that qs, qt ∈ IU , with q ∈ R1,1 iff q · l(s, t) ∈ IU for all l(s, t) ∈ R1,0 iff P(U ) ∩ Q contains the P1 fiber over the point q ∈ P(R1,1 ).  In Theorem 4.6, we show that Proposition 2.1 describes all possible linear syzygies. 3. First syzygies of bidegree (0, 1) Our main result in this section is a complete description of the minimal free resolution when IU has a first syzygy of bidegree (0, 1). As a consequence, if IU has a unique first syzygy of bidegree (0, 1), then the minimal free resolution has numerical Type 5 and if there are two linear first syzygies of bidegree (0, 1), the minimal free resolution has numerical Type 6. We begin with a simple observation Lemma 3.1. If IU has a linear first syzygy of bidegree (0, 1), then IU = hpu, pv, p2 , p3 i, where p is homogeneous of bidegree (2, 0). Proof. Rewrite the syzygy 3 X (ai u + bi v)pi = 0 = u · i=0 and let g0 = 3 P i=0 ai pi , g1 = 3 X ai pi + v · i=0 3 P 3 X bi pi , i=0 bi pi . The relation above implies that (g0 , g1 ) is a i=0 syzygy on (u, v). Since the syzygy module of (u, v) is generated by the Koszul syzygy, this means     g0 −v =p· g1 u  A similar argument applies if IU has a first syzygy of degree (1, 0). Lemma 3.1 has surprisingly strong consequences: Proposition 3.2. If U is basepoint free and IU has a unique linear first syzygy of bidegree (0, 1), then there is a complex of free R modules φ3 φ2 φ1 F1 : 0 −→ F3 −→ F2 −→ F1 −→ IU −→ 0,   where φ1 = p0 p1 p2 p3 , with ranks and shifts matching Type 5 in Table 2. Explicit formulas appear in the proof below. The differentials φi depend on whether p = L1 (s, t)L2 (s, t) of Lemma 3.1 has L1 = L2 . SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 7 Proof. Since IU has a syzygy of bidegree (0, 1), by Lemma 3.1, IU = hpu, pv, p2 , p3 i. Case 1: Suppose p = l(s, t)2 , then after a change of coordinates, p = s2 , so p0 = s2 u and p1 = s2 v. Eliminating terms from p2 and p3 , we may assume p2 p3 = stl1 (u, v) + t2 l2 (u, v) = t(sl1 (u, v) + tl2 (u, v)) = stl3 (u, v) + t2 l4 (u, v) = t(sl3 (u, v) + tl4 (u, v)). Let li (u, v) = ai u + bi v and define  A(u, v) = l1 l3 l2 l4  . Note that det A(u, v) = q(u, v) 6= 0. The rows cannot be dependent, since U spans a four dimensional subspace. If the columns are dependent, then {p2 , p3 } = {tl1 (s + kt), tl2 (s + kt)}, yielding another syzygy of bidegree (0, 1), contradicting our hypothesis. In the proof of Corollary 3.4, we show the hypothesis that U is basepoint free implies that A(u, v) is a 1-generic matrix, which means that A(u, v) cannot be made to have a zero entry using row and column operations. We obtain a first syzygy of bidegree (2, 0) as follows: s2 p2 = s3 tl1 + s2 t2 l2 = (a1 st + a2 t2 )s2 u + (b1 st + b2 t2 )s2 v = (a1 st + a2 t2 )p0 + (b1 st + b2 t2 )p1 A similar relation holds for stp3 , yielding two first syzygies of bidegree (2, 0). We next consider first syzygies of bidegree (1, 1). There is an obvious syzygy on p2 , p3 given by (sl1 (u, v) + tl2 (u, v))p3 = (sl3 (u, v) + tl4 (u, v))p2 Since detA(s, t) = q(u, v) 6= 0, from t2 q stq = l3 p2 − l1 p3 = l4 p2 − l2 p3 and the fact that q(u, v) = L1 (u, v)L2 (u, v) with Li (u, v) = αi u + βi v, so we obtain a pair of relations of bidegree (1, 1): sl4 p2 − sl2 p3 = s2 tL1 L2 = (α1 tL2 )s2 u + (β1 tL2 )s2 v. Case 2: p = l(s, t) · l0 (s, t) with l, l0 independent linear forms. Then after a change of coordinates, p = st, so p0 = stu and p1 = stv. Eliminating terms from p2 and p3 , we may assume p2 = s2 l1 (u, v) + t2 l2 (u, v) p3 = s2 l3 (u, v) + t2 l4 (u, v). Let li (u, v) = ai u + bi v. We obtain a first syzygy of bidegree (2, 0) as follows: stp2 = s3 tl1 + st3 l2 = s2 (stl1 ) + t2 (stl2 ) = (a1 s2 + a2 t2 )stu + (b1 s2 + b2 t2 )stv A similar relation holds for stp3 , yielding two first syzygies of bidegree (2, 0). We next consider first syzygies of bidegree (1, 1). Since q(u, v) 6= 0, from t2 q s2 q = l3 p2 − l1 p3 = l4 p2 − l2 p3 8 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI and the fact that q(u, v) = L1 (u, v)L2 (u, v) with Li (u, v) = αi u + βi v, we have relations sl3 p2 − sl1 p3 = st2 L1 L2 = (α1 tL2 )stu + (β1 tL2 )stv tl4 p2 − tl2 p3 = ts2 L1 L2 = (α1 sL2 )stu + (β1 sL2 )stv, which yield a pair of first syzygies of bidegree (1, 1). Putting everything together, we now have candidates for the differential φ2 in both cases. Computations exactly like those above yield similar candidates for φ3 in the two cases. In Case 1, we have     γ δ 2 2 v α1 tL2 0 a1 st + a2 t a3 st + a4 t  s t     −u β1 tL2 0 b1 st + b2 t2 b3 st + b4 t2   ,  0 s , φ = φ2 =  3 2    0  −sl4 sl3 + tl4 −s 0   −l −l 4 3 2 0 sl2 −sl1 − tl2 0 −s l2 l1 where δ = −α1 β2 t2 + (a1 st + a2 t2 )b3 − (a3 st + a4 t2 )b1 γ = −α1 β2 st + (a1 st + a2 t2 )b4 − (a3 st + a4 t2 )b2 For IU as in Case 2, let    v α1 tL2 α1 sL2 a1 s2 + a2 t2 a3 s2 + a4 t2    −u β1 tL2 β1 sL2 b1 s2 + b2 t2 b3 s2 + b4 t2     φ2 =   , φ3 =  0 −sl3 −tl4 −st 0  0 sl1 tl2 0 −st where γ δ We have already im(φ3 ) ⊆ ker(φ2 ), γ 0 s −l4 l2 δ t 0 −l3 l1    ,   = (−α1 β2 + a1 b4 − a3 b2 )s2 + (a2 b4 − a4 b2 )t2 = (a1 b3 − a3 b1 )s2 + (−α1 β2 + a2 b3 − a4 b1 )t2 shown that im(φ2 ) ⊆ ker(φ1 ), and an easy check shows that yielding a complex of free modules of numerical Type 5.  To prove that the complex above is actually exact, we use the following result of Buchsbaum and Eisenbud [4]: a complex of free modules φi+1 φi φi−1 φ1 F : · · · −→ Fi −→ Fi−1 −→ · · · F1 −→ F0 , is exact iff (1) rank(φi+1 ) + rank(φi ) = rank(Fi ). (2) depth(Irank(φi ) (φi )) ≥ i. Theorem 3.3. The complexes appearing in Proposition 3.2 are exact. Proof. Put F0 = R. An easy check shows that rank(φi+1 ) + rank(φi ) = rank(Fi ), so what remains is to show that depth(I2 (φ3 )) ≥ 3 and depth(I3 (φ2 )) ≥ 2. The fact that s - γ will be useful: to see this, note that both Case 1 and Case 2, s | γ iff a2 b4 − b2 a4 = 0, which implies l2 and l4 differ only by a scalar, contradicting the assumption that U is basepoint free. Case 1: We have us4 , vs4 ∈ I3 (φ2 ). Consider the minor λ = t2 L2 (sl1 + tl2 ) ((α1 b1 − β1 a1 )s + (α1 b2 − β1 a2 )t) obtained from the submatrix  α1 tL2  β1 tL2 sl2 0 0 −sl1 − tl2  a1 st + a2 t2 b1 st + b2 t2  0 SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 9 Note that s does not divide λ, for if s|λ then either l2 = 0 or α1 b2 − β1 a2 = 0. But none of the li can be zero because of the basepoint free assumption (see the proof of Corollary 3.4) and if α1 b2 − β1 a2 = 0, then L1 and l2 are the same up to a scalar multiple. Hence, since l1 l4 − l2 l3 = L1 L2 , we obtain l1 is equal to a scalar multiple of l2 or l3 , which again violates the basepoint free assumption. To conclude, note that u and v can not divide λ at the same time, therefore, λ and one of the us4 and vs4 form a regular sequence in I3 (φ2 ), showing that depth of I3 (φ2 ) is at least 2. To show that depth(I2 (φ3 )) ≥ 3, note that I2 (φ3 ) = hsl2 , sl4 , tl4 − sl3 , tl2 − sl1 , tγ − sδ, sγ, s2 , q(u, v)i Since l2 , l4 are independent, hsl2 , sl4 i = hsu, svi and using these we can reduce tl4 − sl3 , tl2 − sl1 to tu, tv. Since s - γ, modulo s2 , sγ reduces to st2 . Similarly, tγ − sδ reduces to t3 , so that in fact I2 (φ3 ) = hsu, sv, tu, tv, s2 , st2 , t3 , q(u, v)i, and {s2 , t3 , q(u, v)} is a regular sequence of length three. Case 2: We have us2 t2 , vs2 t2 ∈ I3 (φ2 ). Consider the minor λ = L2 (s2 l1 − t2 l2 ) (α1 b1 − β1 a1 )s2 + (α1 b2 − β1 a2 )t2 arising from the submatrix  α1 tL2  β1 tL2 sl1 α1 sL2 β1 sL2 tl2
 a1 s2 + a2 t2 b1 s2 + b2 t2  0 Note that s and t do not divide λ, for if s|λ then either l2 = 0 or α1 b2 − β1 a2 = 0. But none of the li can be zero because of the basepoint free assumption (see the proof of Corollary 3.4) and if α1 b2 − β1 a2 = 0, then L1 and l2 are the same up to a scalar multiple. Hence, since l1 l4 − l2 l3 = L1 L2 , we obtain l1 is equal to a scalar multiple of l2 or l3 , contradicting basepoint freeness. Furthermore, u and v cannot divide λ at the same time, so λ and one of the us2 t2 and vs2 t2 form a regular sequence in I3 (φ2 ). To show that depth(I2 (φ3 )) ≥ 3, note that I2 (φ3 ) = hsu, sv, tu, tv, st, tγ, sδ, q(u, v)i, where we have replaced sli , tlj as in Case 1. If t | δ, then a1 b3 − b1 a3 = 0, which would mean l1 = kl3 and contradict that U is basepoint free. Since s - γ and t - δ, {tγ, sδ, q(u, v)} is regular unless δ, γ share a common factor η = (as + bt). Multiplying out and comparing coefficients shows that this forces γ and δ to agree up to scalar. Combining this with the fact that t - δ, s - δ, we find that δ = as2 +bst+ct2 with a 6= 0 6= c. Reducing sδ and tδ by st then implies that t3 , s3 ∈ I2 (φ3 ).  Corollary 3.4. If U is basepoint free, then IU cannot have first syzygies of both bidegree (0, 1) and bidegree (1, 0). Proof. Suppose there is a first syzygy of bidegree (0, 1) and proceed as in the proof of Proposition 3.2. In the setting of Case 1, IU = hstu, stv, s2 l1 (u, v) + t2 l2 (u, v), s2 l3 (u, v) + t2 l4 (u, v)i. P If there is also a linear syzygy of bidegree (1, 0), expanding out (ai s+bi t)pi shows 3 3 that P the coefficient of s is a3 l1 + a4 l3 , and the coefficient of t is b3 l2 + b4 l4 . Since (ai s+bi t)pi = 0, both coefficients must vanish. In the proof of Proposition 3.2 we showed detA(u, v) = q(u, v) 6= 0. In fact, more is true: if any of the li is zero, then 10 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI U is not basepoint free. For example, if l1 = 0, then ht, l3 i is a minimal associated prime of IU . Since a3 l1 + a4 l3 = 0 iff a3 = a4 = 0 or l1 is a scalar multiple of l3 and the latter situation implies that U is not basepoint free, we must have a3 = a4 = 0. Reasoning similarly for b3 l2 + b4 l4 shows that a3 = a4 = b3 = b4 = 0. This implies the linear syzygy of bidegree (1, 0) can only involve stu, stv, which is impossible. This proves the result in Case 1 and similar reasoning works for Case 2.  Corollary 3.5. If IU has a unique linear first syzygy of bidegree (0, 1), then IU has either one or two embedded prime ideals of the form hs, t, Li (u, v)i. If q(u, v) = det A(u, v) = L1 (u, v)L2 (u, v) for A(u, v) as in Theorem 3.3, then: (1) If L1 = L2 , then the only embedded prime of IU is hs, t, L1 i. (2) If L1 6= L2 , then IU has two embedded primes hs, t, L1 i and hs, t, L2 i. Proof. In [16], Eisenbud, Huneke and Vasconcelos show that a prime P of codimension c is associated to R/I iff it is associated to Extc (R/I, R). If IU has a unique linear syzygy, then the free resolution is given by Proposition 3.2, and Ext3 (R/IU , R) = coker(φt3 ). By Proposition 20.6 of [15], if φ is a presentation matrix for a module M , then the radicals of ann(M ) and Irank(φ) (φ) are equal. Thus, if IU has a Type 5 resolution, the codimension three associated primes are the codimension three associated primes of I2 (φ3 ). The proof of Theorem 3.3 shows that in Case 1, I2 (φ3 ) = hsu, sv, tu, tv, s2 , st2 , t3 , q(u, v)i = hs2 , st2 , t3 , u, vi ∩ hs, t, L21 i = hs2 , st2 , t3 , u, vi ∩ hs, t, L1 i ∩ hs, t, L2 i if L1 = L2 if L1 6= L2 . The embedded prime associated to hs, t, u, vi is not an issue, since we are only interested in the codimension three associated primes. The proof for Case 2 works in the same way.  Next, we tackle the case where the syzygy of bidegree (0, 1) is not unique. Proposition 3.6. If U is basepoint free, then the following are equivalent (1) The ideal IU has two linear first syzygies of bidegree (0, 1). (2) The primary decomposition of IU is IU = hu, vi ∩ hq1 , q2 i, p where hq1 , q2 i = hs, ti and qi are of bidegree (2, 0). (3) The minimal free resolution of IU is of numerical Type 6. (4) XU ' Σ1,1 . Proof. By Lemma 3.1, since IU has a linear syzygy of bidegree (0, 1), IU = hq1 u, q1 v, p2 , p3 i. Proceed as in the proof of Proposition 3.2. In Case 1, the assumption that q1 = st means that p2 , p3 can be reduced to have no terms involving stu and stv, hence there cannot be a syzygy of bidegree (0, 1) involving pi and q1 u, q1 v. Therefore the second first syzygy of bidegree (0, 1) involves only p2 and p3 and the reasoning in the proof of Lemma 3.1 implies that hp2 , p3 i = hq2 u, q2 vi. Thus, we have the primary decomposition IU = hq1 , q2 i ∩ hu, vi, SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 11 √ with q1 , q2 of bidegree (2, 0). Since U is basepoint free, q1 , q2 = hs, ti, so q1 and q2 are a regular sequence in k[s, t]. Similar reasoning applies in the situation of Case 2. That the minimal free resolution is of numerical Type 6 follows from the primary decomposition above, which determines the differentials in the minimal free resolution:     q2  v 0 q2 0       −q1  −u 0 0 q2         −v  0 v −q1 0     2 (−2, −2) u 0 −u 0 −q1 4 ⊕ ←−−−−−− (−4, −2). IU ←− (−2, −1) ←−−−−−−−−−−−−−−−−−−−− (−4, −1)2        The last assertion follows since  det uq1 vq1 uq2 vq2  = 0. Hence, the image of φU is contained in V(xy − zw) = Σ1,1 . After a change of coordinates, q1 = s2 + ast and q2 = t2 + bst, with 0 6= a 6= b 6= 0. Therefore on the open set Us,u ⊆ P1 × P1 the map is defined by (at + 1, (at + 1)v, t2 + bt, (t2 + bt)u), so the image is a surface. Finally, if XU = Σ1,1 , then with a suitable choice of basis for U , p0 p3 − p1 p2 = 0, hence p0 |p1 p2 and p3 |p1 p2 . Since U is four dimensional, this means we must have p0 = αβ, p1 = αγ, p2 = βδ. Without loss of generality, suppose β is quadratic, so there is a linear first syzygy δp0 − αp2 = 0. Arguing similarly for p3 , we find that there are two independent linear first syzygies. Lemma 4.4 of the next section shows that if U is basepoint free, then there can be at most one first syzygy of bidegree (1, 0), so by Corollary 3.4, IU must have two first syzygies of bidegree (0, 1).  4. First syzygies of bidegree (1, 0) Recall that there is an analogue of Lemma 3.1 for syzygies of bidegree (1, 0): Lemma 4.1. If IU has a linear syzygy of bidegree (1, 0), then IU = hps, pt, p2 , p3 i, where p is homogeneous of bidegree (1, 1). Lemma 4.1 has strong consequences as well: we will prove that Proposition 4.2. If U is basepoint free and IU = hps, pt, p2 , p3 i, then (1) IU has numerical Type 4 if and only if p is decomposable. (2) IU has numerical Type 3 if and only if p is indecomposable. We begin with some preliminary lemmas: Lemma 4.3. If IU has a first syzygy of bidegree (1, 0), then IU has two minimal syzygies of bidegree (1, 1) and if p in Lemma 4.1 factors, then IU also has a minimal first syzygy of bidegree (0, 2). 12 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI Proof. First assume p is an irreducible bidegree (1, 1) form, then p = a0 su + a1 sv + a2 tu + a3 tv, with a0 a3 − a1 a2 6= 0. We may assume a0 6= 0 and scale so it is one. Then sp = s2 u + a1 s2 v + a2 stu + a3 stv tp = stu + a1 stv + a2 t2 u + a3 t2 v p2 = b0 t2 u + b1 s2 v + b2 stv + b3 t2 v p3 = c0 t2 u + c1 s2 v + c2 stv + c3 t2 v Here we have used tp and sp to remove all the terms involving s2 u and stu from p2 and p3 . A simple but tedious calculation then shows that p · p2 p · p3 = sp(b1 sv + b2 tv) + tp(b0 tu + b3 tv) = sp(c1 sv + c2 tv) + tp(c0 tu + c3 tv) Now suppose that p = L1 (s, t) · L2 (u, v) with L1 , L2 linear forms, then after a change of coordinates, p = su and a (possibly new) set of minimal generators for IU is hs2 u, stu, p2 , p3 i. Eliminating terms from p2 and p3 , we may assume p2 p3 = as2 v + bstv + t2 l1 = cs2 v + dstv + t2 l2 , where li = li (u, v) ∈ R(0,1) . There are two first syzygies of bidegree (1, 1): su(as2 v + bstv + t2 l1 ) = as3 uv + bs2 tuv + st2 ul1 (as + bt)v · s2 u + tl1 · stu = (as + bt)v · p0 + tl1 · p1 sup2 = = sup3 = su(cs2 v + dstv + t2 l2 ) = cs3 uv + ds2 tuv + st2 ul2 = (cs + dt)v · s2 u + tl2 · stu = (cs + dt)v · p0 + tl2 · p1 A syzygy of bidegree (0, 2) is obtained via: u(l2 p3 − l1 p2 ) = u(as2 vl2 + bstvl2 − cs2 vl1 − dstvl1 ) = (al2 − cl1 )v · s2 u + (bl2 − dl1 )v · stu = (al2 − cl1 )v · p0 + (bl2 − dl1 )v · p1  Lemma 4.4. If U is basepoint free, then there can be at most one linear syzygy of bidegree (1, 0). Proof. Suppose IU has a linear syzygy of bidegree (1, 0), so that sp, tp ∈ IU , with p = su+a1 sv+a2 tu+a3 tv. Note this takes care of both possible cases of Lemma 4.3: in Case 1, a3 − a1 a2 = 0 (e.g. for p = su, a1 = a2 = a3 = 0) and in Case 2, P a3 − a1 a2 6= 0. Now suppose another syzygy of bidegree (1, 0) exists: S = (di s + ei t)pi = 0. Expanding shows that S = d0 s3 u + (e0 + d1 )s2 tu + · · · So after reducing S by the Koszul syzygy on hsp, tpi, d0 = d1 = e0 = 0. In p2 and p3 , one of b1 or c1 must be non-zero. If not, then all of tp, p2 , p3 are divisible by t and since sp |t=0 = s2 (u + a1 v), this would mean ht, u + a1 vi is an associated prime of IU , contradicting basepoint freeness. WLOG b1 6= 0, scale it to one and use it to remove the term c1 s2 v from p3 . This means the coefficient of s3 v in S is d2 b1 = d2 , so d2 vanishes. At this point we have established that S does not involve sp, and involves only t on tp, p2 . SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 13 Now change generators so that p02 = e1 pt + e2 p2 . This modification does not affect tp, sp and IU , but now S involves only p02 and p3 : (t)p02 + (d3 s + e3 t)p3 = 0. As in the proof of Lemma 3.1, letting p002 = p02 + e3 p3 and p003 = d3 p3 , we see that S = tp002 + sp003 = 0, so that p002 = sq and p003 = −tq, hence IU = hs, ti ∩ hp, qi with p, q both of bidegree (1, 1). But on P1 × P1 , V(p, q) is always nonempty, which would mean IU has a basepoint, contradicting our hypothesis.  Remark 4.5. If the P1 fibers of Q did not intersect, Lemma 4.4 would follow easily from Lemma 4.1. However, because Q is a projection of Σ3,1 ⊆ P7 to P5 , the P1 fibers of Q do in fact intersect. Theorem 4.6. If U is basepoint free, then the only possibilities for linear first syzygies of IU are (1) IU has a unique first syzygy of bidegree (0, 1) and no other linear syzygies. (2) IU has a pair of first syzygies of bidegree (0, 1) and no other linear syzygies. (3) IU has a unique first syzygy of bidegree (1, 0) and no other linear syzygies. Proof. It follows from Proposition 3.2 and Proposition 3.6 that both of the first two items can occur. That there cannot be three or more linear syzygies of bidegree (0, 1) follows easily from the fact that if there are two syzygies of bidegree (0, 1) then IU has the form of Proposition 3.6 and the resolution is unique. Corollary 3.4 shows there cannot be linear syzygies of both bidegree (1, 0) and bidegree (0, 1) and Lemma 4.4 shows there can be at most one linear syzygy of bidegree (1, 0).  Our next theorem strengthens Lemma 4.3: there is a minimal first syzygy of bidegree (0, 2) iff the p in Lemma 4.1 factors. We need a pair of lemmas: Lemma 4.7. If P(U ) contains a P2 fiber of Σ2,1 , then U is not basepoint free. Proof. If P(U ) contains a P2 fiber of Σ2,1 over a point of P1 corresponding to a linear form l(u, v), after a change of basis l(u, v) = u and so IU = hs2 u, stu, t2 u, l1 (s, t)l2 (s, t)vi. This implies that hu, l1 (s, t)i ∈ Ass(IU ), so U is not basepoint free.  The next lemma is similar to a result of [11], but differs due to the fact that the subspaces P(W ) ⊆ P(V ) studied in [11] are always basepoint free. Lemma 4.8. If U is basepoint free, then there is a minimal first syzygy on IU of bidegree (0, 2) iff there exists P(W ) ' P2 ⊆ P(U ) such that P(W ) ∩ Σ2,1 is a smooth conic. P Proof. Suppose qi pi = 0 is a minimal P first syzygy P of bidegree P (0, 2), so that 2 2 2 qi = ai u2 + bP this as u a p + uv b p + v ci pi = 0 and i uv + ci v . Rewrite i i i i P P define f0 = ai pi , f1 = bi pi , f2 = ci pi . By construction, hf0 , f1 , f2 i ⊆ IU and [f0 , f1 , f2 ] is a syzygy on [u2 , uv, v 2 ], so         f0 v 0 αv  f1  = α ·  −u  + β ·  v  =  βv − αu  , for some α, β ∈ R2,0 . f2 0 −u −βu 14 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI If {f0 , f1 , f2 } are not linearly independent, there exist constants ci with c0 αv + c1 (βv − αu) − c2 βu = 0. This implies that (c0 α + c1 β)v = (c1 α − c2 β)u, so α = kβ. But then {αu, αv} ⊆ IU , which means there is a minimal first syzygy of bidegree (0, 1), contradicting the classification of §2. Letting W = Span{f0 , f1 , f2 }, we have that P2 ' P(W ) ⊆ P(U ). The actual bidegree (0, 2) syzygy is   v 0 f0 det  −u v f1  = 0. 0 −u f2 To see that the P(W ) meets Σ2,1 in a smooth conic, note that by Lemma 4.7, P(W ) cannot be equal to a P2 fiber of Σ2,1 , or P(U ) would have basepoints. The image of the map P1 → P(W ) defined by (x : y) 7→ x2 (αv) + xy(βv − αu) + y 2 (−βu) = (xα + yβ)(xv − yu) is a smooth conic C ⊆ P(W ) ∩ Σ2,1 . Since P(W ) ∩ Σ2,1 is a curve of degree at most three, if this is not the entire intersection, there would be a line L residual to C. If L ⊆ Fx , where Fx is a P2 fiber over x ∈ P1 , then for small , Fx+ also meets P(W ) in a line, which is impossible. If L is a P1 fiber of Σ2,1 , this would result in a bidegree (0, 1) syzygy, which is impossible by the classification of §2.  Definition 4.9. A line l ⊆ P(s2 , st, t2 ) with l = af + bg, f, g ∈ Span{s2 , st, t2 } is split if l has a fixed factor: for all a, b ∈ P1 , l = L(aL0 + bL00 ) with L ∈ R1,0 . Theorem 4.10. If U is basepoint free, then IU has minimal first syzygies of bidegree (1, 0) and (0, 2) iff P(U ) ∩ Σ2,1 = C ∪ L, 0 where L ' P(W ) is a split line in a P2 fiber of Σ2,1 and C is a smooth conic in P(W ), such that P(W 0 ) ∩ P(W ) = C ∩ L is a point and P(W 0 ) + P(W ) = P(U ). Proof. Suppose there are minimal first syzygies of bidegrees (1, 0) and (0, 2). By Lemma 4.8, the (0, 2) syzygy determines a conic C in a distinguished P(W ) ⊆ P(U ). Every point of C lies on both a P2 and P1 fiber of Σ2,1 . No P1 fiber of Σ2,1 is contained in P(U ), or there would be a first syzygy of bidegree (0, 1), which is impossible by Corollary 3.4. By Lemma 4.1, there exists W 0 = Span{ps, pt} ⊆ U , so we have a distinguished line P(W 0 ) ⊆ P(U ). We now consider two possibilities: Case 1: If p factors, then P(W 0 ) is a split line contained in Σ2,1 , which must therefore be contained in a P2 fiber and p = L(s, t)l(u, v), where l(u, v) corresponds to a point of a P1 fiber of Σ2,1 and P(W 0 )∩P(W ) is a point. In particular P(U )∩Σ2,1 is the union of a line and conic, which meet transversally at a point. Case 2: If p does not factor, then p = a0 su+a1 sv+a2 tu+a3 tv, a0 a3 −a1 a2 6= 0. The corresponding line L = P(ps, pt) meets P(W ) in a point and since p is irreducible L ∩ C = ∅. Since W = Span{αv, βv − αu, −βv}, we must have p · L0 (s, t) = aαv − bβu + c(βv − αu) for some {a, b, c}, where L0 (s, t) = b0 s + b1 t corresponds to L ∩ P(W ). Write p · L0 (s, t) as l1 uL0 (s, t) + l2 vL0 (s, t), where l1 (s, t) = a0 s + a2 t l2 (s, t) = a1 s + a3 t SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 15 Then p · L0 (s, t) = = = l1 (s, t)L0 (s, t)u + l2 (s, t)L0 (s, t)v aαv − bβu + c(βv − αu) (−bβ − cα)u + (aα + cβ)v a c c b In particular,  (4.1)      α l2 L0 · = β −l1 L0 If  det a c c b  6= 0, then applying Cramer’s rule to Equation 4.1 shows that α and β share a common factor L0 (s, t). But then W = Span{L0 γ, L0 δ, L0 }, which contradicts the basepoint freeness of U : change coordinates so L0 = s, so U = sγ, sδ, s, p3 . Since p3 |s=0 = t2 l(u, v), hs, l(u, v)i is an associated prime of IU , a contradiction. To conclude, consider the case ab − c2 = 0. Then there is a constant k such that       a c α l2 L0 · = , ka kc β −l1 L0 which forces kl1 (s, t) = −l2 (s, t). Recalling that l1 (s, t) = a0 s + a2 t and l2 (s, t) = a1 s + a3 t, this implies that   a0 a2 det = 0, a1 a3 contradicting the irreducibility of p. This shows that if IU has minimal first syzygies of bidegree (1, 0) and (0, 2), then P(U ) ∩ Σ2,1 = C ∪ L meeting transversally at a point. The remaining implication follows from Lemma 4.3 and Lemma 4.8.  For W ⊆ H 0 (OP1 ×P1 (2, 1)) a basepoint free subspace of dimension three, the minimal free resolution of IW is determined in [11]: there are two possible minimal free resolutions, which depend only whether P(W ) meets Σ2,1 in a finite set of points, or a smooth conic C. By Theorem 4.10, if there are minimal first syzygies of bidegrees (1, 0) and (0, 2), then U contains a W with P(W ) ∩ Σ2,1 = C, which suggests building the Type 4 resolution by choosing W = Span{p0 , p1 , p2 } to satisfy P(W ) ∩ Σ2,1 = C and constructing a mapping cone. There are two problems with this approach. First, there does not seem to be an easy description for hp0 , p1 , p2 i : p3 . Second, recall that the mapping cone resolution need not be minimal. A computation shows that the shifts in the resolutions of hp0 , p1 , p2 i : p3 and hp0 , p1 , p2 i overlap and there are many cancellations. However, choosing W to consist of two points on L and one on C solves both of these problems at once. Lemma 4.11. In the setting of Theorem 4.10, let p0 correspond to L ∩ C and let p1 , p2 correspond to points on L and C (respectively) distinct from p0 . If W = Span{p0 , p1 , p2 }, then IW has a Hilbert Burch resolution. Choosing coordinates so W = Span{sLv, tLv, βu}, the primary decomposition of IW is ∩4i=1 Ii = hs, ti2 ∩ hu, vi ∩ hβ, vi ∩ hL, ui. 16 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI Proof. After a suitable change of coordinates, W = Span{sLv, tLv, βu}, and writing β = (a0 s + a1 t)L0 (s, t), IW consists of the two by two minors of   t a0 uL0 φ2 =  −s a1 uL0  . 0 Lv Hence, the minimal free resolution of IW is φ2 0 ←− IW ←− (−2, −1)3 ←− (−3, −1) ⊕ (−3, −2) ←− 0 For the primary decomposition, since L does not divide β, hβ, vi ∩ hL, ui = hβL, vL, βu, vui, and intersecting this with hs, ti2 ∩ hu, vi gives IW .  Lemma 4.12. If U is basepoint free, W = Span{sLv, tLv, βu} and p3 = αu−βv = tLu − βv, then IW : p3 = hβL, vL, βu, vui. Proof. First, our choice of p0 to correspond to C ∩ L in Theorem 4.10 means we may write α = tL. Since hs, ti2 : p3 = 1 = hu, vi : p3 , IW : p3 = (∩4i=1 Ii ) : p3 = (hβ, vi : p3 ) ∩ (hL, ui : p3 ). Since p3 = tLu − βv, f p3 ∈ hβ, vi iff f tLu ∈ hβ, vi. Since tL = α and α, β and u, v are relatively prime, this implies f ∈ hβ, vi. The same argument shows that hL, ui : p3 must equal hL, ui.  Theorem 4.13. In the situation of Theorem 4.10, the minimal free resolution is of Type 4. If p0 corresponds to L ∩ C, p1 6= p0 to another point on L and p2 6= p0 to a point on C and W = Span{p0 , p1 , p2 }, then the minimal free resolution is given by the mapping cone of IW and IW : p3 . Proof. We construct a mapping cone resolution from the short exact sequence ·p3 0 ←− R/IU ←− R/IW ←− R(−2, −1)/IW : p3 ←− 0. By Lemma 4.12, IW : p3 = hβL, Lv, βu, uvi, which by the reasoning in the proof of Proposition 3.6 has minimal free resolution:     v 0 u 0  u    (−3, −1)  (−3, −0)   −L  −β 0 0 u        ⊕ ⊕  −v  0 v −L 0     (−1, −1) (−2, −2) 0 −β 0 −L β ⊕ ⊕ : p3 ←− ←−−−−−−−−−−−−−−−−−−− ←−−−−−− (−3, −2). (−2, −1) (−3, −1) ⊕ ⊕ (0, −2) (−1, −2)        IW A check shows that there are no overlaps in the mapping cone shifts, hence the mapping cone resolution is actually minimal.  SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 17 4.1. Type 3 resolution. Finally, suppose IU = hp0 , p1 , p2 , p3 i with p0 = ps and p1 = pt, such that p = a0 su + a1 sv + a2 tu + a3 tv is irreducible, so a0 a3 − a1 a2 6= 0. As in the case of Type 4, the minimal free resolution will be given by a mapping cone. However, in Type 3 the construction is more complicated: we will need two mapping cones to compute the resolution. What is surprising is that by a judicious change of coordinates, the bigrading allows us to reduce IU so that the equations have a very simple form. Theorem 4.14. If U is basepoint free and IU = hps, pt, p2 , p3 i with p irreducible, then the IU has a mapping cone resolution, and is of numerical Type 3. Proof. Without loss of generality, assume a0 = 1. Reducing p2 and p3 mod ps and pt, we have p2 = b0 t2 u + b1 s2 v + b2 stv + b3 t2 v p3 = c0 t2 u + c1 s2 v + c2 stv + c3 t2 v Since U is basepoint free, either b0 or c0 is nonzero, so after rescaling and reducing p3 mod p2 p2 = t2 u + b1 s2 v + b2 stv + b3 t2 v p3 = c1 s2 v + c2 stv + c3 t2 v = (c1 s2 + c2 st + c3 t2 )v = L1 L2 v for some Li ∈ R1,0 . If the Li ’s are linearly independent, then a change of variable replaces L1 and L2 with s and t. This transforms p0 , p1 to p0 l1 , p0 l2 , but since the li ’s are linearly independent linear forms in s and t, hp0 l1 , p0 l2 i = hp0 s, p0 ti, with p0 irreducible. So we may assume IU = hps, pt, p2 , p3 i, where p3 = stv or s2 v. With this change of variables, p2 = l2 u + (b01 s2 + b02 st + b03 t2 )v where l = as + bt and b 6= 0. We now analyze the two possible situations. First, suppose p3 = stv. Reducing p2 modulo hps, pt, stvi yields p2 = = αt2 u + b001 s2 v + b003 t2 v (αu + b003 v)t2 + b001 s2 v By basepoint freeness, α 6= 0, so changing variables via αu + b003 v 7→ u yields p2 = t2 u + b001 s2 v. Notice this change of variables does not change the form of the other pi . Now b001 6= 0 by basepoint freeness, so rescaling t (which again preserves the form of the other pi ) shows that IU = hps, pt, t2 u + s2 v, stvi. Since p is irreducible, p = sl1 + tl2 with li are linearly independent elements of R0,1 . Changing variables once again via l1 7→ u and l2 7→ v, we have IU = hs(su + tv), t(su + tv), stQ1 , s2 Q1 + t2 Q2 i, where Q1 = au + bv, Q2 = cu + dv are linearly independent with b 6= 0. Rescaling v and s we may assume b = 1. Now let IW = hs(su + tv), t(su + tv), stQ1 i. 18 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI The minimal free resolution of IW is     0 ←− IW ←− (−2, −1)3 t −s 0 0 sQ1 p ←−     (−3, −1) ⊕ (−3, −2) ←− 0, To obtain a mapping cone resolution, we need to compute IW : p2 . As in the Type 4 setting, we first find the primary decomposition for IW . (1) If a = 0, then = hs(su + tv), t(su + tv), stvi = hu, vi ∩ hs, ti2 ∩ hu, ti ∩ hsu + tv, (s, v)2 i = ∩4i=1 Ii IW (2) If a 6= 0, rescale u and t by a so Q1 = u + v. Then IW = = hs(su + tv), t(su + tv), st(u + v)i hu, vi ∩ hs, ti2 ∩ hv, si ∩ hu, ti ∩ hu + v, s − ti = ∩5i=1 Ii Since s2 Q1 + t2 Q2 ∈ I1 ∩ I2 in both cases, IW : s2 Q1 + t2 Q2 = ∩ni=2 Ii . So if a = 0, IW : s2 Q1 + t2 Q2 = = = = hu, ti : s2 Q1 + t2 Q2 ∩ hsu + tv, (s, v)2 i : s2 Q1 + t2 Q2 hu, ti ∩ hsu + tv, (s, v)2 i hsu + tv, uv 2 , tv 2 , stv, s2 ti hsu + tv, I3 (φ)i, while if a 6= 0, IW : s2 Q1 + t2 Q2 = = = = hv, si : s2 Q1 + t2 Q2 ∩ hu, ti : s2 Q1 + t2 Q2 ∩ hu + v, s − ti : s2 Q1 + t2 Q2 hv, si ∩ hu, ti ∩ hu + v, s − ti hsu + tv, uv(u + v), tv(u + v), tv(s − t), st(s − t)i hsu + tv, I3 (φ)i where  t 0  u s φ=  0 v 0 0   0  0   if a = 0, and φ =   s  v t 0 u s−t 0 u+v 0 0  0 0   if a 6= 0. s  v Since I3 (φ) has a Hilbert-Burch resolution, a resolution of IW : p2 = hsu+tv, I3 (φ)i can be obtained as the mapping cone of I3 (φ) with I3 (φ) : p. There are no overlaps, so the result is a minimal resolution. However, there is no need to do this, because the change of variables allows us to do the computation directly and we find (−1, −1) ⊕ (0, −3) (−1, −3)2 ⊕ ⊕ (−2, −3) ⊕ 0 ← Iw : p3 ←− (−1, −2) ←− (−2, −2)2 ←− ←0 ⊕ ⊕ (−3, −2) (−2, −1) (−3, −1)2 ⊕ (−3, −0) This concludes the proof if p3 = stv. When p3 = s2 v, the argument proceeds in similar, but simpler, fashion.  SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 19 5. No linear first syzygies 5.1. Hilbert function. Proposition 5.1. If U is basepoint free, then there are six types of bigraded Hilbert function in one-to-one correspondence with the resolutions of Table 2. The tables below contain the values of hi,j = HF ((i, j), R/IU ), for i < 5, j < 6 listed in the order corresponding to the six numerical types in Table 2. 0 1 2 3 4 5 0 1 2 3 4 5 6 1 2 4 2 0 0 0 2 3 6 1 0 0 0 3 4 8 0 0 0 0 4 5 10 0 0 0 0 0 1 2 3 4 5 0 1 2 3 4 5 6 1 2 4 2 0 0 0 2 3 6 1 0 0 0 3 4 8 1 0 0 0 4 5 10 1 0 0 0 0 1 2 3 4 5 0 1 2 3 4 5 6 1 2 4 2 1 0 0 2 3 6 1 0 0 0 3 4 8 0 0 0 0 4 5 10 0 0 0 0 0 1 2 3 4 5 0 1 2 3 4 5 6 1 2 4 2 1 0 0 2 3 6 1 0 0 0 3 4 8 1 0 0 0 4 5 10 1 0 0 0 0 1 2 3 4 5 0 1 2 3 4 5 6 1 2 4 2 0 0 0 2 3 6 2 0 0 0 3 4 8 2 0 0 0 4 5 10 2 0 0 0 0 1 2 3 4 5 0 1 2 3 4 5 6 1 2 4 2 0 0 0 2 3 6 3 0 0 0 3 4 8 4 0 0 0 4 5 10 5 0 0 0 The entries of the first two rows and the first column are clear: h0,j = HF ((0, j), R) = j + 1, h1,j = HF ((1, j), R) = 2j + 2 and hi,0 = HF ((i, 0), R) = i + 1. Furthermore h2,1 = 2 by the linear independence of the minimal generators of IU . The proof of the proposition is based on the following lemmas concerning hi,j , i ≥ 2, j ≥ 1. Lemma 5.2. For IU an ideal generated by four independent forms of bidegree (2, 1) (1) h3,1 is the number of bidegree (1, 0) first syzygies (2) h2,2 − 1 is the number of bidegree (0, 1) first syzygies Proof. From the free resolution R(−2, −3)B ⊕ 0 ← R/IU ← R ← R(−2, −1)4 ←− R(−3, −1)A ←− F2 ←− F3 ← 0, ⊕ F1 we find h3,1 = HF ((3, 1), R) − 4HF ((1, 0), R) + AHF ((0, 0), R) = 8 − 8 + A = A h2,2 = HF ((2, 2), R) − 4HF ((0, 1), R) + BHF ((0, 0), R) = 9 − 8 + B = B + 1, since Fi are free R-modules generated in degree (i, j) with i > 3 or j > 2.  Lemma 5.3. If U is basepoint free, then h3,2 = 0 for every numerical type. Proof. If there are no bidegree (1, 0) syzygies then HF ((3, 1), R/IU ) = 0 and consequently HF ((3, 2), R/IU ) = 0. If there are bidegree (1, 0) syzygies then we are in 20 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI Type 3 or 4 where by Proposition 4.2 we know the relevant part of the resolution is R(−3, −1) ⊕ 0 ← R/IU ← R ← R(−2, −1)4 ←− R(−3, −2)2 ←− F2 ←− F3 ← 0 ⊕ F1 Then h3,2 = HF ((3, 2), R) − 4HF ((1, 1), R) + HF ((0, 1), R) + 2HF ((0, 0), R) = 12 − 16 + 2 + 2 = 0.  So far we have determined the following shape of the Hilbert function of R/IU : 0 1 2 3 4 5 0 1 2 3 4 5 6 1 2 4 2 h3,1 h4,1 h5,1 2 3 6 3 4 8 h2,2 0 0 0 h2,3 0 0 0 4 5 10 h2,4 0 0 0 If linear syzygies are present we know from the previous section the exact description of the possible minimal resolutions of IU and it is an easy check that they agree with the last four Hilbert functions in Proposition 5.1. Next we focus on the case when no linear syzygies are present. By Lemma 5.2 this yields h2,2 = 1 and h3,1 = 0, hence hi,1 = 0 for i ≥ 3. We show that in the absence of linear syzygies only the first two Hilbert functions in Proposition 5.1 may occur: 5.2. Types 1 and 2. In the following we assume that the basepoint free ideal IU has no linear syzygies. We first determine the maximal numerical types which correspond to the Hilbert functions found in §4.1 and then we show that only the Betti numbers corresponding to linear syzygies cancel. Proposition 5.4. If U is basepoint free and IU has no linear syzygies, then (1) IU cannot have two or more linearly independent bidegree (0, 2) first syzygies (2) IU cannot have two minimal first syzygies of bidegrees (0, 2), (0, j), j > 2 (3) IU has a single bidegree (0, 2) minimal syzygy iff h2,j = 1 for j ≥ 3 (4) IU has no bidegree (0, 2) minimal syzygy iff h2,j = 0 for j ≥ 3 Proof. (1) Suppose IU has two linearly independent bidegree (0, 2) first syzygies which can be written down by a similar procedure to the one used in Lemma 3.1 as u2 p + uvq + v 2 r = 0 u2 p0 + uvq 0 + v 2 r0 = 0 with p, q, r, p0 , q 0 , r0 ∈ U . Write p = p1 u + p2 v with p1 , p2 ∈ R2,0 and similarly for p0 , q, q 0 , r, r0 . Substituting in the equations above one obtains p1 p2 + q1 q2 + r1 r2 = 0 = 0 = 0 = 0 p02 q20 p01 + q10 + r10 r20 = 0 = 0 = 0 = 0 SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 21 hence p q r = p2 v, = −(p2 u + r1 v), = r1 v, p0 q0 r0 = = = p02 v 0 −(p2 u + r10 v) r10 v are elements of IU . If both of the pairs p2 v, p02 v or r1 u, r10 u consists of linearly independent elements of R2,1 , then U ∩ Σ2,1 contains a P1 inside each of the P2 fibers over the points corresponding to u, v in the P1 factor of the map P2 × P1 7→ P5 . Pulling back the two lines from Σ2,1 to the domain of its defining map, one obtains two lines in P2 which must meet (or be identical). Taking the image of the intersection point we get two elements of the form αu, αv ∈ IU which yield a (0, 1) syzygy, thus contradicting our assumption. Therefore it must be the case that p02 = ap2 or r10 = br1 with a, b ∈ k. The reasoning being identical, we shall only analyze the case p02 = ap2 . A linear combination of the elements q = −(p2 u + r1 v), q 0 = −(p02 u + r10 v) ∈ IU produces (r10 − ar1 )v ∈ IU and a linear combination of the elements r1 u, r10 u ∈ IU produces (r10 − ar1 )u ∈ IU , hence again we obtain a (0, 1) syzygy unless r10 = ar1 . But then (p0 , q 0 , r0 ) = a(p, q, r) and these triples yield linearly dependent bidegree (0, 2) syzygies. (2) The assertion that IU cannot have a bidegree (0, 2) and a (distinct) bidegree (0, j), j ≥ 2 minimal first syzygies is proved by induction on j. The base case j = 2 has already been solved. Assume IU has a degree (0, 2) syzygy u2 p + uvq + v 2 r = 0 with p = p1 u + p2 v, q, r ∈ IU and a bidegree (0, j) syzygy uj w1 + uj−1 vw2 + . . . + v j wj+1 = 0 with wi = yi u + zi v ∈ IU . Then as before p1 = 0, z1 = 0, r2 = 0, zj+1 = 0 and the same reasoning shows one must have z1 = ap2 or yj+1 = br1 . Again we handle the case z1 = ap2 where a linear combination of the two syzygies produces the new syzygy uj−1 v(w2 − aq) + uj−2 v 2 (w3 − ar) + uj−3 v 3 (w3 ) . . . + v j wj+1 = 0. Dividing by v: uj−1 (w2 − aq) + uj−2 v(w3 − ar) + uj−3 v 2 (w3 ) . . . + v j−1 wj+1 = 0, which is a minimal bidegree (0, j − 1) syzygy iff the original (0, j) syzygy was minimal. This contradicts the induction hypothesis. (3) An argument similar to Lemma 5.2 shows that in the absence of (0, 1) syzygies h2,3 is equal to the number of bidegree (0, 2) syzygies on IU . Note that the absence of (0, 1) syzygies implies there can be no bidegree (0, 2) second syzygies of IU to cancel the effect of bidegree (0, 2) first syzygies on the Hilbert function. This covers the converse implications of both (3) and (4) as well as the case j = 3 of the direct implications. The computation of h2,j , j ≥ 3 is completed as follows h2,j = HF ((2, j), R) − HF ((2, j), R(−2, −1)4 ) + HF ((2, j), R(−2, −3)) = 3(j + 1) − 4j + (j − 2) =1 (4) In this case we compute h2,3 = HF ((2, 3), R) − HF ((2, 3), R(−2, −1)4 ) = 12 − 12 = 0 HF ((2, 4), R) − HF ((2, 4), R(−2, −1)4 ) = 15 − 16 = −1 The fact that h2,j = 0 for higher values of j follows from h2,3 = 0. In fact even more is true: IU is forced to have a single bidegree (0, 3) first syzygy to ensure that h2,j = 0 for j ≥ 4.  22 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI Corollary 5.5. There are only two possible Hilbert functions for basepoint free ideals IU without linear syzygies, depending on whether there is no (0, 2) syzygy or exactly one (0, 2) syzygy. The two possible Hilbert functions are 0 1 2 3 4 5 0 1 2 3 4 5 6 1 2 4 2 0 0 0 2 3 6 1 0 0 0 3 4 8 0 0 0 0 4 5 10 0 0 0 0 0 1 2 3 4 5 0 1 2 3 4 5 6 1 2 4 2 0 0 0 2 3 6 1 0 0 0 3 4 8 1 0 0 0 4 5 10 1 0 0 0 Proposition 5.6. If U is basepoint free and IU has no linear syzygies, then IU has (1) exactly 4 bidegree (1, 1) first syzygies (2) exactly 2 bidegree (2, 0) first syzygies Proof. Note that there cannot be any second syzygies in bidegrees (1, 1) and (2, 0) 1 1 of bidebecause of the absence of linear first syzygies. Thus the numbers β3,2 , β4,1 gree (1, 1) and (2, 0) first syzygies are determined by the Hilbert function: 1 β3,2 = h3,2 − HF ((3, 2), R) + HF ((3, 2), R(−2, −1)4 ) = 0 − 12 + 16 = 4 1 β4,1 = h4,1 − HF ((4, 1), R) + HF ((4, 1), R(−2, −1)4 ) = 0 − 10 + 12 = 2  Next we obtain upper bounds on the bigraded Betti numbers of IU by using bigraded initial ideals. The concept of initial ideal with respect to any fixed term order is well known and so is the cancellation principle asserting that the resolution of an ideal can be obtained from that of its initial ideal by cancellation of some consecutive syzygies of the same bidegree. In general the problem of determining which cancellations occur is very difficult. In the following we exploit the cancellation principle by using the bigraded setting to our advantage. For the initial ideal computations we use the revlex order induced by s > t > u > v. In [1], Aramova, Crona and de Negri introduce bigeneric initial ideals as follows (we adapt the definition to our setting): let G = GL(2, 2) × GL(2, 2) with an element g = (dij , ekl ) ∈ G acting on the variables in R by g : s 7→ d11 s + d12 t, t 7→ d21 s + d22 t, u 7→ e11 u + e12 v, v 7→ e21 u + e22 v We shall make use of the following results of [1]. Theorem 5.7. [[1] Theorem 1.4] Let I ⊂ R be a bigraded ideal. There is a Zariski open set U in G and an ideal J such that for all g ∈ U we have in(g(I)) = J. Definition 5.8. The ideal J in Theorem 5.7 is defined to be the bigeneric initial ideal of I, denoted by bigin(I). Definition 5.9. A monomial ideal I ⊂ R is bi-Borel fixed if g(I) = I for any upper triangular matrix g ∈ G. Definition 5.10. A monomial ideal I ⊂ R = k[s, t, u, v] is strongly bistable if for every monomial m ∈ I the following conditions are satisfied: (1) if m is divisible by t, then sm/t ∈ I. (2) if m is divisible by v, then um/v ∈ I . SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 23 As in the Z-graded case, the ideal bigin(I) has the same bigraded Hilbert function as I. Propositions 1.5 and 1.6 of [1] show that bigin(I) is bi-Borel fixed, and in characteristic zero, bigin(I) is strongly bistable. Proposition 5.11. For each of the Hilbert functions in Proposition 5.4 there are exactly two strongly bistable monomial ideals realizing it. These ideals and their respective bigraded resolutions are: (1) G1 = hs2 u, s2 v, stu, stv, t2 u2 , t2 uv, t3 u, t3 v, t2 v 3 i with minimal resolution (5.1) (−2, −1)4 ⊕ (−2, −2)2 ⊕ 0 ← G1 ← (−2, −3) ⊕ (−3, −1)2 (−2, −2)2 ⊕ (−3, −2) (−2, −3) ⊕ ⊕ (−3, −3)2 (−2, −4) (−4, −3) ⊕ ⊕ ⊕ ←0 ←− (−3, −1)2 ←− (−3, −4)2 ←− (−4, −4) ⊕ ⊕ (−4, −2)3 (−3, −2)5 ⊕ ⊕ (−4, −3) (−3, −3)2 ⊕ (−4, −1)2 G01 = hs2 u, s2 v, stu, t2 u, stv 2 , st2 v, t3 v, t2 v 3 i with minimal resolution (5.2) (−2, −1)4 ⊕ (−2, −2) ⊕ 0 ← G01 ← (−2, −3) ⊕ (−3, −1)2 (−2, −2) ⊕ (−2, −3) (−3, −3)2 ⊕ ⊕ (−2, −4) (−4, −3) (−3, −4)2 ⊕ ⊕ ⊕ ←− ←0 ←− (−3, −1)2 ←− (−4, −2)3 (−4, −4) ⊕ ⊕ (−3, −2)4 ⊕ (−4, −3) (−3, −3)2 ⊕ (−4, −1)2 (2) G2 = hs2 u, s2 v, stu, stv, t2 u2 , t2 uv, t3 u, t3 vi with minimal resolution (5.3) (−2, −1)4 ⊕ 0 ← G2 ← (−2, −2)2 ⊕ (−3, −1)2 (−2, −2)2 ⊕ (−2, −3) (−3, −2) ⊕ ⊕ ←− (−3, −1)2 ←− (−3, −3)2 ←− (−4, −3) ← 0 ⊕ ⊕ (−3, −2)5 (−4, −2)3 ⊕ (−4, −1)2 24 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI G02 = hs2 u, s2 v, stu, t2 u, stv 2 , st2 v, t3 vi with minimal resolution (5.4) (−2, −1)4 ⊕ 0 ← G02 ← (−2, −2) ⊕ (−3, −1)2 (−2, −2) ⊕ (−2, −3) ⊕ (−3, −3)2 2 ⊕ ←− (−4, −3) ← 0 ←− (−3, −1) ←− ⊕ (−4, −2)3 (−3, −2)4 ⊕ (−4, −1)2 Proof. There are only two strongly bistable sets of four monomials in R2,1 : {s2 u, s2 v, stu, stv} and {s2 , s2 v, stu, t2 u}. To complete {s2 u, s2 v, stu, stv} to an ideal realizing one of the Hilbert functions in Proposition 5.4 we need two additional monomials in R2,2 , which must be t2 u2 , t2 uv in order to preserve bistability. Then we must add the two remaining monomials t3 u, t3 v in R3,1 , which yields the second Hilbert function. To realize the first Hilbert function we must also include the remaining monomial t2 v 3 ∈ R2,3 . To complete {s2 , s2 v, stu, t2 u} to an ideal realizing one of the Hilbert functions in Proposition 5.4, we need one additional monomial in R2,2 which must be stv 2 in order to preserve bistability. Then we must add the two remaining monomials st2 v, t3 v ∈ R3,1 . Then to realize the first Hilbert function, we must add the remaining monomial t2 v 3 ∈ R2,3 .  Theorem 5.12. There are two numerical types for the minimal Betti numbers of basepoint free ideals IU without linear syzygies. (1) If there is a bidegree (0, 2) first syzygy then IU has numerical Type 2. (2) If there is no bidegree (0, 2) first syzygy then IU has numerical Type 1. Proof. Proposition 5.4 establishes that the two situations above are the only possibilities in the absence of linear syzygies and gives the Hilbert function corresponding to each of the two cases. Proposition 5.11 identifies the possible bigeneric initial ideals for each case. Since these bigeneric initial ideals are initial ideals obtained following a change of coordinates, the cancellation principle applies. We now show the resolutions (5.1), (5.2) must cancel to the Type 1 resolution and the resolutions (5.3), (5.4) must cancel to the Type 2 resolution. Since IU is assumed to have no linear syzygies, all linear syzygies appearing in the resolution of its bigeneric initial ideal must cancel. Combined with Proposition 5.6, this establishes that in (5.3) or (5.4) the linear cancellations are the only ones that occur. In (5.1), the cancellations of generators and first syzygies in bidegrees (2, 2), (2, 3), (3, 1) are obvious. The second syzygy in bidegree (3, 2) depends on the cancelled first syzygies, therefore it must also be cancelled. This is natural, since by Proposition 5.6, there are exactly four bidegree (3, 2) first syzygies. An examination of the maps in the resolution (5.1) shows that the bidegree (3, 3) second syzygies depend on the cancelled first syzygies, so they too must cancel. Finally the bidegree (4, 3) last syzygy depends on the previous cancelled second syzygies and so must also cancel. In (5.2), the cancellations of generators and first syzygies in bidegrees (2, 2), (2, 3), (3, 1) are obvious. The second syzygies of bidegree (3, 3) depend only on the cancelled first syzygies, so they too cancel. Finally the bidegree (4, 3) last syzygy depends on the previous cancelled second syzygies and so it must also cancel.  SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 25 6. Primary decomposition Lemma 6.1. If U is basepoint free, all embedded primes of IU are of the form (1) hs, t, l(u, v)i (2) hu, v, l(s, t)i (3) m = hs, t, u, vi √ Proof. Since IU = hs, ti ∩ hu, vi, an embedded prime must contain either hs, ti or hu, vi and modulo these ideals any remaining minimal generators can be considered as irreducible polynomials in k[u, v] or k[s, t] (respectively). But the only prime ideals here are hli i with li a linear form, or the irrelevant ideal.  Lemma 6.2. If U is basepoint free, then the primary components corresponding to minimal associated primes of IU are (1) Q1 = hu, vi √ (2) Q2 = hs, ti2 or Q2 = hp, qi, with p, q ∈ R2,0 and p, q = hs, ti. Proof. Let Q1 , Q2 be the primary components associated to hu, vi and hs, ti respectively. Since IU ⊂ Q1 ⊂ hu, vim and IU is generated in bidegree (2, 1), Q1 must contain at least one element of bidegree (0, 1). If Q1 contains exactly one element p(u, v) of bidegree (0, 1), then V is contained in the fiber of Σ2,1 over the point V (p(u, v)), which contradicts the basepoint free assumption. Therefore Q1 must contain two independent linear forms in u, v and hence Q1 = hu, vi. Since IU ⊂ Q2 and IU contains elements of bidegree (2, 1), Q2 must contain at least one element of bidegree (2, 0). If Q2 contains exactly one element q(s, t) of bidegree (2, 0), then V is contained in the fiber of Σ2,1 over the point V (q(s, t)), which contradicts the basepoint free assumption. If Q2 contains exactly two elements of bidegree (2, 0) which share a common linear factor l(s, t), then IU is contained in the ideal hl(s, t)i, which contradicts the basepoint free assumption as well. Since the bidegree (2, 0) part of Q2 is contained in the linear span of s2 , t2 , st, it follows that the only possibilities consistent with the conditions above √ are Q2 = hp, qi with p, q = hs, ti or Q2 = hs2 , t2 , sti.  Proposition 6.3. For each type of minimal free resolution of IU with U basepoint free, the embedded primes of IU are as in Table 1. Proof. First observe that m = hs, t, u, vi is an embedded prime for each of Type 1 to Type 4. This follows since the respective free resolutions have length four, so Ext4R (R/IU , R) 6= 0. 0 By local duality, this is true iff Hm (R/IU ) 6= 0 iff m ∈ Ass(IU ). Since the resolutions for Type 5 and Type 6 have projective dimension less than four, this also shows that in Type 5 and Type 6, m 6∈ Ass(IU ). Corollary 3.5 and Proposition 3.6 show the embedded primes for Type 5 and Type 6 are as in Table 1. Thus, by Lemma 6.1, all that remains is to study primes of the form hs, t, L(u, v)i and hu, v, L(s, t)i for Type 1 through 4. For this, suppose IU = I1 ∩ I2 ∩ I3 , where (1) I1 is the intersection of primary components corresponding to the two minimal associated primes identified in Lemma 6.2. (2) I2 is the intersection of embedded primary components not primary to m. (3) I3 is primary to m. 26 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI √ By Lemma 6.2, if I1 = hu, vi∩hp, qi with p, q = hs, ti, then I1 is basepoint free and consists of four elements of bidegree (2, 1), thus IU = I1 and has Type 6 primary decomposition. So we may assume I1 = hu, vi ∩ hs, ti2 . Now we switch gears and consider all ideals in the Z–grading where the variables have degree one. In the Z–grading HP (R/I1 , t) = 4t + 2. Since the Hilbert polynomials of R/(I1 ∩I2 ) and R/IU are identical, we can compute the Hilbert polynomials of R/(I1 ∩ I2 ) for Type 1 through 4 using Theorems 5.12, 4.13 and 4.14. For example, in Type 1, the bigraded minimal free resolution is (−2, −4) ⊕ (−3, −4)2 4 4 ⊕ 0 ← IU ← (−2, −1) ←− (−3, −2) ←− ←− (−4, −4) ← 0. ⊕ (−4, −2)3 (−4, −1)2 Therefore, the Z–graded minimal free resolution is (−7)2 (−6) ⊕ ⊕ ←− (−8) ← 0. ←− 0 ← IU ← (−3) ←− (−6)3 (−5)6 4 Carrying this out for the other types shows that the Z–graded Hilbert polynomial of R/(I1 ∩ I2 ) is (1) In Type 1 and Type 3: HP (R/(I1 ∩ I2 ), t) = 4t + 2. (2) In Type 2 and Type 4: HP (R/(I1 ∩ I2 ), t) = 4t + 3. In particular, for Type 1 and Type 3, HP (R/I1 , t) = HP (R/(I1 ∩ I2 ), t), and in Type 2 and Type 4, the Hilbert polynomials differ by one: HP (I1 /(I1 ∩ I2 ), t) = 1. Now consider the short exact sequence 0 −→ I1 ∩ I2 −→ I1 −→ I1 /(I1 ∩ I2 ) −→ 0. Since I1 ∩ I2 ⊆ I1 , in Type 1 and Type 3 where the Hilbert Polynomials are equal, there can be no embedded primes save m. In Type 2 and Type 4, since HP (I1 /(I1 ∩ I2 ), t) = 1, I1 /(I1 ∩ I2 ) is supported at a point of P3 which corresponds to a codimension three prime ideal of the form hl1 , l2 , l3 i. Switching back to the fine grading, by Lemma 6.1, this prime must be either hs, t, l(u, v)i or hu, v, l(s, t)i. Considering the multidegrees in which the Hilbert function of I1 /(I1 ∩I2 ) is nonzero shows that the embedded prime is of type hs, t, l(u, v)i.  SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 27 7. The Approximation complex and Implicit equation of XU The method of using moving lines and moving quadrics to obtain the implicit equation of a curve or surface was developed by Sederberg and collaborators in [27], [28], [29]. In [10], Cox gives a nice overview of this method and makes explicit the connection to syzygies. In the case of tensor product surfaces these methods were first applied by Cox-Goldman-Zhang in [12]. The approximation complex was introduced by Herzog-Simis-Vasconcelos in [23],[24]. From a mathematical perspective, the relation between the implicit equation and syzygies comes from work of Busé-Jouanolou [5] and Busé-Chardin [6] on approximation complexes and the Rees algebra; their work was extended to the multigraded setting in [2], [3]. The next theorem follows from work of Botbol-Dickenstein-Dohm [3] on toric surface parameterizations, and also from a more general result of Botbol [2]. The novelty of our approach is that by obtaining an explicit description of the syzygies, we obtain both the implicit equation for the surface and a description of the singular locus. Theorem 7.3 gives a particularly interesting connection between syzygies of Iu and singularities of XU . Theorem 7.1. If U is basepoint free, then the implicit equation for XU is determinantal, obtained from the 4 × 4 minor of the first map of the approximation complex Z in bidegree (1, 1), except for Type 6, where φU is not birational. 7.1. Background on approximation complexes. We give a brief overview of approximation complexes, for an extended survey see [7]. For I = hf1 , . . . , fn i ⊆ R = k[x1 , . . . xm ], let Ki ⊆ Λi (Rn ) be the kernel of the ith Koszul differential on {f1 , . . . , fn }, and S = R[y1 , . . . , yn ]. Then the approximation complex Z has ith term Z i = S ⊗R K i . The differential is the Koszul differential on {y1 , . . . , yn }. It turns out that H0 (Z) is SI and the higher homology depends (up to isomorphism) only on I. For µ a bidegree in R, define d di−1 i Z µ : · · · −→ k[y1 , . . . , yn ] ⊗k (Ki )µ −→ k[y1 , . . . , yn ] ⊗k (Ki−1 )µ −→ · · · If the bidegree µ and base locus of I satisfy certain conditions, then the determinant of Z µ is a power of the implicit equation of the image. This was first proved in [5]. In Corollary 14 of [3], Botbol-Dickenstein-Dohm give a specific bound for µ in the case of a toric surface and map with zero-dimensional base locus and show that in this case the gcd of the maximal minors of dµ1 is the determinant of the complex. For four sections of bidegree (2, 1), the bound in [2] shows that µ = (1, 1). To make things concrete, we work this out for Example 1.1. Example 7.2. Our running example is U = Span{s2 u, s2 v, t2 u, t2 v + stv}. Since K1 is the module of syzygies on IU , which is generated by the columns of   −v −t2 0 0 −tv u 0 −st − t2 0 0    2 0 s 0 −sv − tv −sv  0 0 s2 tu su 28 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI The first column encodes the relation ux1 − vx0 = relations s2 x2 − t2 x0 = 2 s x3 − (st + t2 )x1 = tux3 − (sv + tv)x2 = sux3 − svx2 − tvx0 = 0, then next four columns the 0 0 0 0 If we were in the singly graded case, we would need to use µ = 2, and a basis for Z12 consists of {s, t, u, v} · ux1 − vx0 , and the remaining four relations. With respect to the ordered basis {s2 , st, t2 , su, sv, tu, tv, u2 , uv, v 2 } for R2 and writing · for 0, the matrix for d21 : Z12 −→ Z02 is  ·  ·   ·   x1  −x0   ·   ·   ·   · · · · · · · x1 −x0 · · · · · · · · · · x1 −x0 · · · · · · · · · x1 −x0 x2 · −x0 · · · · · · · x3 −x1 −x1 · · · · · · · · · · · −x2 x3 −x2 · · ·  · ·   ·   x3   −x2   ·   −x0   ·   ·  · However, this matrix represents all the first syzygies of total degree two. Restricting to the submatrix of bidegree (1, 1) syzygies corresponds to choosing rows indexed by {su, sv, tu, tv}, yielding  x1 −x0   · · · · x1 −x0 · −x2 x3 −x2  x3 −x2   ·  −x0 We now study the observation made in the introduction, that linear syzygies manifest in a linear singular locus. Example 7.2 again provides the key intuition: (1,1) a linear first syzygy gives rise to two columns of d1 . Theorem 7.3. If U is basepoint free and IU has a unique linear syzygy, then the codimension one singular locus of XU is a union of lines. Proof. Without loss of generality we assume the linear syzygy involves the first two generators p0 , p1 of IU , so that in the two remaining columns corresponding to the linear syzygy the only nonzero entries are x0 and x1 , which appear exactly as in (1,1) Example 7.2. Thus, in bidegree (1, 1), the matrix for d1 has the form   x1 · ∗ ∗ −x0 · ∗ ∗    · x1 ∗ ∗ · −x0 ∗ ∗ Computing the determinant using two by two minors in the two left most columns shows the implicit equation of F is of the form x21 · f + x0 x1 g + x20 h, SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 29 which is singular along the line V(x0 , x1 ). To show that the entire singular locus is a union of lines when IU has resolution Type ∈ {3, 4, 5}, we must analyze the (1,1) structure of d1 . For Type 3 and 4, Theorems 4.14 and 4.13 give the first syzygies, and show that the implicit equation for XU is given by the determinant of   −x1 · x2 x3  ·  −x1 a1 x2 − b1 x0 a1 x3 − c1 x0  .  x0  · a2 x2 − b0 x1 a2 x3 − c0 x1 · x0 a3 x2 − b2 x0 − b3 x1 a3 x3 − c2 x0 − c3 x1 We showed above that V(x0 , x1 ) ⊆ Sing(XU ). Since XU \ V(x0 , x1 ) ⊆ Ux0 ∪ Ux1 , it suffices to check that XU ∩ Ux0 and XU ∩ Ux1 are smooth in codimension one. XU ∩ Ux0 is defined by −c1 y2 + b1 y3 = + + + (−b3 c0 + b0 c3 )y14 + (a3 c0 − a2 c3 )y13 y2 (−a3 b0 + a2 b3 )y13 y3 + +(−b2 c0 + b0 c2 )y13 (a1 c0 − a2 c2 − c3 )y12 y2 + (−a1 b0 + a2 b2 + b3 )y12 y3 (−b1 c0 + b0 c1 )y12 + (−a2 c1 − c2 )y1 y2 + (a2 b1 + b2 )y1 y3 . By basepoint freeness, b1 or c1 is nonzero, as is (−b3 c0 + b0 c3 ), so in fact XU ∩ Ux0 is smooth. A similar calculation shows that XU ∩ Ux1 is also smooth, so for Type 3 and Type 4, Sing(XU ) is a line. In Type 5 the computation is more cumbersome: with notation as in Proposition 3.2, the relevant 4 × 4 submatrix is   x1 · a1 x3 − a3 x2 β1 α2 x1 + α1 α2 x0 −x0 · b1 x3 − b3 x2 β1 β2 x1 + α1 β2 x0   .  · x1 β1 α2 x1 + α1 α2 x0 a2 x3 − a4 x2  · −x0 β1 β2 x1 + α1 β2 x0 b2 x 3 − b4 x 2 A tedious but straightforward calculation shows that Sing(XU ) consists of three lines in Type 5a and a pair of lines in Type 5b.  Theorem 7.4. If U is basepoint free, then the codimension one singular locus of XU is as described in Table 1. Proof. For a resolution of Type 3,4,5, the result follows from Theorem 7.3, and for Type 6 from Proposition 3.6. For the generic case (Type 1), the result is obtained by Elkadi-Galligo-Le [18], so it remains to analyze Type 2. By Lemma 4.8, the (0, 2) first syzygy implies that we can write p0 , p1 , p2 as αu, βv, αv + βu for some α, β ∈ R2,0 . Factor α as product of two linear forms in s and t, so after a linear change of variables we may assume α = s2 or st. If α = s2 , write β = (ms + nt)(m0 s + n0 t), and note that n and n0 cannot both vanish, because then β is a scalar multiple of α, violating linear independence of the pi . Thus, after a linear change of variables, we may assume β is of the form t(ks + lt), so the pi ’s are of the form {s2 u, (ks + lt)tv, s2 v + (ks + lt)tu, p3 } If l = 0, then IU = hs2 u, kstv, s2 v + kstu, p3 i, which is not basepoint free: if s = 0, thus the first 3 polynomials vanish and p3 becomes t2 (au + bv) which vanishes for some (u : v) ∈ P1 . So l 6= 0, and after a linear change of variables t 7→ tl and u 7→ lu, we may assume l = 1 and hence IU = hs2 u, t2 v + kstv, s2 v + t2 u + kstu, p3 i. 30 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI Now consider two cases, k = 0 and k 6= 0. In the latter, we may assume k = 1: first replace ks by s and then replace k 2 u by u. In either case, reducing p3 by the other generators of IU shows we can assume p3 = astu + bs2 v + cstv = s(atu + bsv + ctv). By Theorem 5.12 there is one first syzygy of bidegree (0, 2), two first syzygies of bidegree (2, 0), and four first syzygies of bidegree (1, 1). A direct calculation shows that two of the bidegree (1, 1) syzygies are   atu + bsv + ctv 0  0 asu − btu + csv   .   0 btv −su −tv The remaining syzygies depend on k and a. For example, if k = a = 0, then the remaining bidegree (1, 1) syzygies are   0 −b3 tv   b2 su + c2 sv c3 sv ,  2 2   −b sv −b csv 2 2 bsv − ctv b tu + bcsv − c tv and the bidegree (2, 0) syzygies are  0 b3 t2 2  bs + cst −c3 st   0 −b3 s2 2 2 2 −t b s − bcst + c2 t2     Thus, if k = a = 0, using the basis {su, tu, sv, tv} for R1,1 , the matrix whose determinant gives the implicit equation for XU is   −x3 0 b2 x 1 0  0  −bx1 0 b2 x 3    bx0 cx1 c2 x1 − b2 x2 + bx3 c3 x1 − b2 cx2 + bcx3  cx0 bx2 − x3 −cx3 −b3 x0 − c2 x3 Since k = a = 0, if both b and c vanish there will be a linear syzygy on IU , contradicting our assumption. So suppose b 6= 0 and scale the generator p3 so b = 1:   −x3 0 x1 0  0  −x1 0 x3   2 3  x0 cx1 c x1 − x2 + x3 c x1 − cx2 + cx3  cx0 x2 − x3 −cx3 −x0 − c2 x3 Expanding along the top two rows by 2 × 2 minors as in the proof of Theorem 7.3 shows that XU is singular along V(x1 , x3 ), and evaluating the Jacobian matrix with x1 = 0 shows this is the only component of the codimension one singular locus with x1 = 0. Next we consider the affine patch Ux1 . On this patch, the Jacobian ideal is h(4x3 +c2 )(x2 −2x3 −c2 ), (x2 −2x3 −c2 )(x2 +2x3 ), 2x0 −2x2 x3 −x2 c2 +2x23 +4x3 c2 +c4 i which has codimension one component given by V(x2 − 2x3 − c2 , x0 − x23 ), a plane conic. Similar calculations work for the other cases.  SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 31 8. Connection to the dual scroll We close by connecting our work to the results of Galligo-Lê in [19]. First, recall that the ideal of Σ2,1 is defined by the two by two minors of   x0 x1 x2 . x3 x4 x5 Combining this with the relations x21 − x0 x2 and x24 − x3 x5 arising from ν2 shows that the image of the map τ defined in Equation 2.1 is the vanishing locus of the two by two minors of   x0 x1 x3 x4 (8.1) . x1 x2 x4 x5 Let A denote the 4 × 6 matrix of coefficients of the polynomials defining U in the monomial basis above. We regard P(U ) ,→ P(V ) via a 7→ a · A. Note that IU and the implicit equation of XU are independent of the choice of generators pi ([7]). The dual projective space of P(V ) is P(V ∗ ) where V ∗ = Homk (V, k) and the projective subspace of P(V ∗ ) orthogonal to U is defined to be P((V /U )∗ ) = P(U ⊥ ), where U ⊥ = {f ∈ V ∗ |f (u) = 0, ∀u ∈ U } is algebraically described as the kernel of A. The elements of U ⊥ define the space of linear forms in xi which vanish on P(U ). In Example 1.1 U = Span{s2 u, s2 v, t2 u, t2 v + stv}, so A is the matrix   1 0 0 0 0 0 0 0 0 1 0 0   0 0 1 0 0 0 , 0 0 0 0 1 1 and P(U ) = V(x1 , x4 − x5 ) ⊆ P5 . The conormal variety N (X) is the incidence variety defined as the closure of the set of pairs {(x, π) ∈ P(V ) × P(V ∗ )} such that x is a smooth point of X and π is an element of the linear subspace orthogonal to the tangent space TX,x in the sense described above. N (X) is irreducible and for a varieties X embedded in P(V ) ' P5 the dimension of N (X) is 4. The image of the projection of N (X) onto the factor P(V ∗ ) is by definition the dual variety of X denoted X ∗ . g Denote by X U the variety XU re-embedded as a hypersurface of P(U ). Proposition 1.4 (ii) of [8] applied to our situation reveals ∗ ∗ 3 g Proposition 8.1. If X is a hypersurface which is swept by a one dimenU ⊂P ∗ g sional family of lines, then X is either a 2-dimensional scroll or else a curve. U The cited reference includes precise conditions that allow the two possibilities to be distinguished. Proposition 1.2 of [8] reveals the relation between XU∗ ⊂ P(V ∗ ) ∗ ∗ g and X U ⊂ P(U ), namely ∗ ∗ g Proposition 8.2. In the above setup, XU∗ ⊂ P(V ∗ ) is a cone over X U ⊂ P(U ) ⊥ ∗ with vertex U = P((V /U ) ). It will be useful for this reason to consider the map π : P(V ) → P(U ∗ ) defined by π(p) = (`1 (p) : . . . : `4 (p)), where `1 , . . . , `4 are the defining equations of U ⊥ . ∗ g The map π is projection from P(U ⊥ ) and π(XU∗ ) = X U . Using a direct approach ∗ ∗ g Galligo-Lê obtain in[19] that π −1 (X U ) is a (2,2)-scroll in P(V ) which they denote ∗ g by F∗2,2 . For brevity we write F for π −1 (X U ). 32 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI Galligo-Lê classify possibilities for the implicit equation of XU by considering the pullback φ∗U of P(U ∗ ) ∩ F to (P1 × P1 )∗ . The two linear forms Li defining P(U ∗ ) pull back to give a pair of bidegree (2, 1) forms on (P1 × P1 )∗ . Proposition 8.3 ([19] §6.5). If φ∗U (L1 ) ∩ φ∗U (L2 ) is infinite, then φ∗U (L1 ) and φ∗U (L2 ) share a common factor g, for which the possibilities are: (1) deg(g) = (0, 1). (2) deg(g) = (1, 1) (g possibly not reduced). (3) deg(g) = (1, 0) (residual system may have double or distinct roots). (4) deg(g) = (2, 0) (g can have a double root). Example 8.4. In the following we use capital letters to denote elements of the various dual spaces. Note that the elements of the basis of (P1 × P1 )∗ that pair dually to {s2 u, stu, t2 u, s2 v, stv, t2 v} are respectively { 21 S 2 U, ST U, 12 T 2 U, 21 S 2 V, ST V, 12 T 2 V }. Recall we have the following dual maps of linear spaces: φU P1 × P1 −→ A P(U ) −→ P(V ) φ∗ π U (P1 × P1 )∗ ←− P(U ∗ ) ←− P(V ∗ ) ∗ In Example 1.1, P(U ∗ ) = V(X1 , X4 − X5 ) ⊆ P5 , φ∗U (X1 ) = ST U and φ∗U (X4 − X5 ) = ST V − 21 T 2 V , so there is a shared common factor T of degree (1, 0) and the residual system {SU, 12 T V − SV } has distinct roots (0 : 1) × (1 : 0), (1 : 2) × (0 : 1). Taking points (1 : 0) × (1 : 0) and (1 : 0) × (0 : 1) on the line T = 0 and the points above shows that the forms below are in (φ∗U P (U ∗ ))⊥ = φU −1 (P(U )) φU ((0 : 1) × (1 : 0)) = t2 u φU ((1 : 2) × (0 : 1)) = (s + 2t)2 v φU ((1 : 0) × (1 : 0)) = s2 u φU ((1 : 0) × (0 : 1)) = s2 v and in terms of our chosen basis the  0 0 0 0  1 0 0 0 corresponding matrix A is  1 0 0 0 0 1 4 4 , 0 0 0 0 0 1 0 0 whose rows span the expected linear space U with basis hx0 , x2 , x3 , x4 + x5 i. 8.1. Connecting φ∗U (L1 ) ∩ φ∗U (L2 ) to syzygies. There is a pleasant relation between Proposition 8.3 and bigraded commutative algebra. This is hinted at by the result of [11] relating the minimal free resolution of IW to P(W ) ∩ Σ2,1 . Theorem 8.5. If φ∗U (L1 ) ∩ φ∗U (L2 ) is infinite, then: (1) If deg(g) = (0, 1) then U is not basepoint free. (2) If deg(g) = (1, 1) then (a) if g is reducible then U is not basepoint free. (b) if g is irreducible then IU is of Type 3. (3) If deg(g) = (1, 0) then IU is of Type 5. Furthermore (a) The residual scheme is reduced iff IU is of Type 5a. (b) The residual scheme has a double root iff IU is of Type 5b. (4) If deg(g) = (2, 0) then IU is of Type 6. Proof. We use the notational conventions of 8.4. If deg(g) = (0, 1) then we may assume after a change of coordinates that φ∗U (L1 ) = P U and φ∗U (L2 ) = QU , SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 33 with P, Q of bidegree (2, 0) and independent. In particular, denoting by p, q ∈ R the dual elements of P, Q a basis for R2,1 is {pu, qu, ru, pv, qv, rv}. Hence U = Span{ru, pv, qv, rv}, so if r = l1 (s, t)l2 (s, t) then hv, l1 (s, t)i is an associated prime of IU and U is not basepoint free. Next, suppose deg(g) = (1, 1). If g factors, then after a change of variable we may assume g = SU and U ⊥ = Span{S 2 U, ST U }. This implies that hv, ti is an associated prime of IU , so U is not basepoint free. If g is irreducible, then g = a0 SU + a1 SV + a2 T U + a3 T V, with a0 a3 − a1 a2 6= 0. Since U ⊥ = Span{gS, gT }, U is the kernel of  a0 a2 0 a1 a3 0 a0 a2 0 a1 so U contains the columns of  a3  a1   0  −a2  −a0 0  0 , a3  0 a3   a1  . 0   −a2  −a0 In particular, a3 s2 u + a1 stu − a2 s2 v − a0 stv a3 stu + a1 t2 u − a2 stv − a0 t2 v = s(a3 su + a1 tu − a2 sv − a0 tv) = sp = t(a3 su + a1 tu − a2 sv − a0 tv) = tp are both in IU , yielding a linear syzygy of bidegree (1, 0). Since p is irreducible, the result follows from Proposition 4.2. The proofs for the remaining two cases are similar and omitted.  There is an analog of Proposition 8.3 when the intersection of the pullbacks is finite and a corresponding connection to the minimal free resolutions, which we leave for the interested reader. Concluding remarks Our work raises a number of questions: (1) How much generalizes to other line bundles OP1 ×P1 (a, b) on P1 × P1 ? We are at work extending the results of §7 to a more general setting. (2) What can be said about the minimal free resolution if IU has basepoints? (3) Is there a direct connection between embedded primes and the implicit equation? (4) If U ⊆ H 0 (OX (D)) is four dimensional and has base locus of dimension at most zero and X is a toric surface, then the results of [3] give a bound on the degree µ needed to determine the implicit equation. What can be said about the syzygies in this case? (5) More generally, what can be said about the multigraded free resolution of IU , when IU is graded by Pic(X)? Acknowledgments Evidence for this work was provided by many computations done using Macaulay2, by Dan Grayson and Mike Stillman. Macaulay2 is freely available at http://www.math.uiuc.edu/Macaulay2/ and scripts to perform the computations are available at http://www.math.uiuc.edu/~schenck/O21script 34 HAL SCHENCK, ALEXANDRA SECELEANU, AND JAVID VALIDASHTI We thank Nicolás Botbol, Marc Chardin and Claudia Polini for useful conversations, and an anonymous referee for a careful reading of the paper. References 1. A. Aramova, K. Crona, E. De Negri, Bigeneric initial ideals, diagonal subalgebras and bigraded Hilbert functions, J. Pure Appl. Algebra 150 (2000), 312–335. 2. N. Botbol, The implicit equation of a multigraded hypersurface, Journal of Algebra, J. Algebra 348 (2011), 381-401. 3. N. Botbol, A. Dickenstein, M. Dohm, Matrix representations for toric parametrizations, Comput. Aided Geom. Design 26 (2009), 757–771. 4. D. Buchsbaum, D. Eisenbud, What makes a complex exact?, J. Algebra 25 (1973), 259–268. 5. L. Busé, J.-P. Jouanolou, On the closed image of a rational map and the implicitization problem, J. Algebra 265 (2003), 312–357. 6. L. Busé, M. Chardin, Implicitizing rational hypersurfaces using approximation complexes, J. Symbolic Computation 40 (2005), 1150–1168. 7. M. Chardin, Implicitization using approximation complexes, in “Algebraic geometry and geometric modeling”, Math. Vis., Springer, Berlin (2006), 23–35. 8. C. Ciliberto, F. Russo, A. Simis, Homaloidal hypersurfaces and hypersurfaces with vanishing Hessian, Adv. Mat 218 (2008), 1759–1805. 9. D. Cox, The moving curve ideal and the Rees algebra, Theoret. Comput. Sci. 392 (2008), 23–36. 10. D. Cox, Curves, surfaces and syzygies, in “Topics in algebraic geometry and geometric modeling”, Contemp. Math. 334 (2003) 131–150. 11. D. Cox, A. Dickenstein, H. Schenck, A case study in bigraded commutative algebra, in “Syzygies and Hilbert Functions”, edited by Irena Peeva, Lecture notes in Pure and Applied Mathematics 254, (2007), 67–112. 12. D. Cox, R. Goldman, M. Zhang, On the validity of implicitization by moving quadrics for rational surfaces with no basepoints, J. Symbolic Computation 29 (2000), 419–440. 13. W.L.F. Degen, The types of rational (2, 1)-Bézier surfaces. Comput. Aided Geom. Design 16 (1999), 639–648. 14. A. Dickenstein, I. Emiris, Multihomogeneous resultant formulae by means of complexes, J. Symbolic Computation 34 (2003), 317–342. 15. D. Eisenbud, Commutative Algebra with a view towards Algebraic Geometry, Springer-Verlag, Berlin-Heidelberg-New York, 1995. 16. D. Eisenbud, C. Huneke, W. Vasconcelos, Direct methods for primary decomposition, Invent. Math. 110 (1992), 207–235. 17. W. Edge, The theory of ruled surfaces, Cambridge University Press, 1931. 18. M. Elkadi, A. Galligo and T. H. Lê, Parametrized surfaces in P3 of bidegree (1, 2), Proceedings of the 2004 International Symposium on Symbolic and Algebraic Computation, ACM, New York, 2004, 141–148. 19. A. Galligo, T. H. Lê, General classification of (1, 2) parametric surfaces in P3 , in “Geometric modeling and algebraic geometry”, Springer, Berlin, (2008) 93–113. 20. I. Gelfand, M. Kapranov, A. Zelevinsky, Discriminants, Resultants and Multidimensional Determinants, Birkhauser, Boston, 1984. 21. J. Harris, Algebraic Geometry, A First Course, Springer-Verlag, Berlin-Heidelberg-New York, 1992. 22. R. Hartshorne, Algebraic Geometry, Springer-Verlag, Berlin-Heidelberg-New York, 1977. 23. J. Herzog, A. Simis, W. Vasconcelos Approximation complexes of blowing-up rings, J. Algebra 74 (1982), 466–493. 24. J. Herzog, A. Simis, W. Vasconcelos Approximation complexes of blowing-up rings II, J. Algebra 82 (1983), 53–83. 25. J. Migliore, Introduction to Liason Theory and Deficiency Modules, Progress in Math. vol. 165, Birkhäuser, Boston Basel, Berlin, 1998. 26. G. Salmon, Traité de Géométrie analytique a trois dimensiones, Paris, Gauthier-Villars, 1882. 27. T. W. Sederberg, F. Chen, Implicitization using moving curves and surfaces, in Proceedings of SIGGRAPH, 1995, 301–308. SYZYGIES AND SINGULARITIES OF TENSOR PRODUCT SURFACES 35 28. T. W. Sederberg, R. N. Goldman and H. Du, Implicitizing rational curves by the method of moving algebraic curves, J. Symb. Comput. 23 (1997), 153–175. 29. T. W. Sederberg, T. Saito, D. Qi and K. S. Klimaszewksi, Curve implicitization using moving lines, Comput. Aided Geom. Des. 11 (1994), 687–706. 30. S. Zube, Correspondence and (2, 1)-Bézier surfaces, Lithuanian Math. J. 43 (2003), 83–102. 31. S. Zube, Bidegree (2, 1) parametrizable surfaces in P3 , Lithuanian Math. J. 38 (1998), 291– 308. Department of Mathematics, University of Illinois, Urbana, IL 61801 E-mail address: schenck@math.uiuc.edu Department of Mathematics, University of Nebraska, Lincoln, NE 68588 E-mail address: aseceleanu2@math.unl.edu Department of Mathematics, University of Illinois, Urbana, IL 61801 E-mail address: jvalidas@illinois.edu 

Logical Methods in Computer Science Vol. 11(4:13)2015, pp. 1–23 www.lmcs-online.org Submitted Published Sep. 16, 2014 Dec. 22, 2015 TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS WITH COMPOSITE REGULAR TYPES ∗ LUCA PADOVANI Dipartimento di Informatica, Università di Torino, Italy e-mail address: luca.padovani@di.unito.it Abstract. We extend the linear π-calculus with composite regular types in such a way that data containing linear values can be shared among several processes, if there is no overlapping access to such values. We describe a type reconstruction algorithm for the extended type system and discuss some practical aspects of its implementation. 1. Introduction The linear π-calculus [15] is a formal model of communicating processes that distinguishes between unlimited and linear channels. Unlimited channels can be used without restrictions, whereas linear channels can be used for one communication only. Despite this seemingly severe restriction, there is evidence that a significant portion of communications in actual systems take place on linear channels [15]. It has also been shown that structured communications can be encoded using linear channels and a continuation-passing style [13, 3]. The interest in linear channels has solid motivations: linear channels are efficient to implement, they enable important optimizations [9, 8, 15], and communications on linear channels enjoy important properties such as interference freedom and partial confluence [18, 15]. It follows that understanding whether a channel is used linearly or not has a primary impact in the analysis of systems of communicating processes. Type reconstruction is the problem of inferring the type of entities used in an unannotated (i.e., untyped) program. In the case of the linear π-calculus, the problem translates into understanding whether a channel is linear or unlimited, and determining the type of messages sent over the channel. This problem has been addressed and solved in [10]. The goal of our work is the definition of a type reconstruction algorithm for the linear 2012 ACM CCS: [Theory of computation]: Models of computation; Semantics and reasoning— Program constructs; [Software and its engineering]: Software notations and tools—General programming languages—Language features. Key words and phrases: linear pi-calculus, composite regular types, shared access to data structures with linear values, type reconstruction. ∗ A preliminary version of this paper [20] appears in the proceedings of the 17th International Conference on Foundations of Software Science and Computation Structures (FoSSaCS’14). This work has been supported by ICT COST Action IC1201 BETTY, MIUR project CINA, Ateneo/CSP project SALT, and the bilateral project RS13MO12 DART. l LOGICAL METHODS IN COMPUTER SCIENCE c DOI:10.2168/LMCS-11(4:13)2015 CC L. Padovani Creative Commons 2 L. PADOVANI π-calculus extended with pairs, disjoint sums, and possibly infinite types. These features, albeit standard, gain relevance and combine in non-trivial ways with the features of the linear π-calculus. We explain why this is the case in the rest of this section. The term below *succ?(x,y).y!(x + 1) | new a in (succ!(39,a) | a?(z).print!z) (1.1) models a program made of a persistent service (the *-prefixed process waiting for messages on channel succ) that computes the successor of a number and a client (the new-scoped process) that invokes the service and prints the result of the invocation. Each message sent to the service is a pair made of the number x and a continuation channel y on which the service sends the result of the computation back to the client. There are three channels in this program, succ for invoking the service, print for printing numbers, and a private channel a which is used by the client for receiving the result of the invocation. In the linear π-calculus, types keep track of how each occurrence of a channel is being used. For example, the above program is well typed in the environment print : [int]0,1 , succ : [int × [int]0,1 ]ω,1 where the type of print indicates not only the type of messages sent over the channel (int in this case), but also that print is never used for input operations (the 0 annotation) and is used once for one output operation (the 1 annotation). The type of succ indicates that messages sent over succ are pairs of type int×[int]0,1 – the service performs exactly one output operation on the channel y which is the second component of the pair – and that succ is used for an unspecified number of input operations (the ω annotation) and exactly one output operation (the 1 annotation). Interestingly, the overall type of succ can be expressed as the combination of two slightly different types describing how each occurrence of succ is being used by the program: the leftmost occurrence of succ is used according to the type [int × [int]0,1 ]ω,0 (arbitrary inputs, no outputs), while the rightmost occurrence of succ is used according to the type [int × [int]0,1 ]0,1 (no inputs, one output). Following [15], we capture the overall use of a channel by means of a combination operator + on types such that, for example, [int × [int]0,1 ]ω,0 + [int × [int]0,1 ]0,1 = [int × [int]0,1 ]ω,1 Concerning the restricted channel a, its rightmost occurrence is used according to the type [int]1,0 , since there a is used for one input of an integer number; the occurrence of a in (39,a) is in a message sent on succ, and we have already argued that the service uses this channel according to the type [int]0,1 ; the type of the leftmost, binding occurrence of a is the combination of these two types, namely: [int]0,1 + [int]1,0 = [int]1,1 The type of a indicates that the program performs exactly one input and exactly one output on a, hence a is a linear channel. Since a is restricted in the program, even if the program is extended with more processes, it is not possible to perform operations on a other than the ones we have tracked in its type. The key ingredient in the discussion above is the notion of type combination [15, 10, 24], which allows us to gather the overall number of input/output operations performed on a channel. We now discuss how type combination extends to composite and possibly infinite types, which is the main novelty of the present work. TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 3 So far we have taken for granted the ability to perform pattern matching on the message received by the service on succ and to assign distinct names, x and y, to the components of the pair being analyzed. Pattern matching is usually compiled using more basic operations. For example, in the case of pairs these operations are the fst and snd projections that respectively extract the first and the second component of the pair. So, a low-level modeling of the successor service that uses fst and snd could look like this: *succ?(p).snd(p)!(fst(p) + 1) (1.2) This version of the service is operationally equivalent to the previous one, but from the viewpoint of typing there is an interesting difference: in (1.1) the two components of the pair are given distinct names x and y and each name is used once in the body of the service; in (1.2) there is only one name p for the whole pair which is projected twice in the body of the service. Given that each projection accesses only one of the two components of the pair and ignores the other, we can argue that the occurrence of p in snd(p) is used according to the type int × [int]0,1 (the 1 annotation reflects the fact that the second component of p is a channel used for an output operation) whereas the occurrence of p in fst(p) is used according to the type int × [int]0,0 (the second component of p is not used). The key idea, then, is that we can extend the type combination operator + component-wise to product types to express the overall type of p as the combination of these two types: (int × [int]0,1 ) + (int × [int]0,0 ) = (int + int) × ([int]0,1 + [int]0,0 ) = int × [int]0,1 According to the result of such combination, the second component of p is effectively used only once despite the multiple syntactic occurrences of p. The extension of type combination to products carries over to disjoint sums and also to infinite types as well. To illustrate, consider the type tlist satisfying the equality tlist = Nil ⊕ Cons([int]1,0 × tlist ) which is the disjoint sum between Nil, the type of empty lists, and Cons([int]1,0 × tlist ), the type of non-empty lists with head of type [int]1,0 and tail of type tlist (we will see shortly that there is a unique type tlist satisfying the above equality relation). Now, tlist can be expressed as the combination todd + teven , where todd and teven are the types that satisfy the equalities todd = Nil ⊕ Cons([int]1,0 × teven ) and teven = Nil ⊕ Cons([int]0,0 × todd ) (1.3) (again, there are unique todd and teven that satisfy these equalities, see Section 3). In words, todd is the type of lists of channels in which each channel in an odd-indexed position is used for one input, while teven is the type of lists of channel in which each channel in an even-indexed position is used for one input. The reason why this particular decomposition of tlist could be interesting is that it enables the sharing of a list containing linear channels among two processes, if we know that one process uses the list according to the type todd and the other process uses the same list according to the type teven . For 4 L. PADOVANI example, the process R defined below P def = *odd?(l,acc,r).case l of Nil ⇒ r!acc Cons(x,l′ ) ⇒ x?(y).even!(l′ ,(acc + y),r) def Q = *even?(l,acc,r).case l of Nil ⇒ r!acc Cons(x,l′ ) ⇒ odd!(l′ ,acc,r) def R = P | Q | new a,b in (odd!(l,0,a) | even!(l,0,b) | a?(x).b?(y).r!(x + y)) uses each channel in a list l for receiving a number, sums all such numbers together, and sends the result on another channel r. However, instead of scanning the list l sequentially in a single thread, R spawns two parallel threads (defined by P and Q) that share the very same list l: the first thread uses only the odd-indexed channels in l, whereas the second thread uses only the even-indexed channels in l; the (partial) results obtained by these two threads are collected by R on two locally created channels a and b; the overall result is eventually sent on r. We are then able to deduce that R makes full use of the channels in l, namely that l has type tlist , even though the list as a whole is simultaneously accessed by two parallel threads. In general, we can see that the extension of type combination to composite, potentially infinite types is an effective tool that fosters the parallelization of programs and allows composite data structures containing linear values to be safely shared by a pool of multiple processes, if there is enough information to conclude that each linear value is accessed by exactly one of the processes in the pool. Such detailed reasoning on the behavior of programs comes at the price of a more sophisticated definition of type combination. This brings us back to the problem of type reconstruction. The reconstruction algorithm described in this article is able to infer the types todd and teven of the messages accepted by P and Q by looking at the structure of these two processes and of understanding that the overall type of l in R is tlist , namely that every channel in l is used exactly once. Related work. Linear type systems with composite types have been discussed in [8, 9] for the linear π-calculus and in [25] for a functional language. In these works, however, every structure that contains linear values becomes linear itself (there are a few exceptions for specific types [14] or relaxed notions of linearity [11]). The original type reconstruction algorithm for the linear π-calculus is described in [10]. Our work extends [10] to composite and infinite types. Unlike [10], however, we do not deal with structural subtyping, whose integration into our type reconstruction algorithm is left for future work. The type reconstruction algorithm in [10] and the one we present share a common structure in that they both comprise constraint generation and constraint resolution phases. The main difference concerns the fact that we have to deal with constraints expressing the combination of yet-to-be-determined types, whereas in [10] non-trivial type combinations only apply to channel types. This allows [10] to use an efficient constraint resolution algorithm based on unification. In our setting, the presence of infinite types hinders the use of unification, and in some cases the resolution algorithm may conservatively approximate the outcome in order to ensure proper termination. TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS P, Q ::= | | | | | | e, f idle e?(x).P e!f P |Q *P new a in P case e {i(xi ) ⇒ Pi }i=inl,inr Process (idle process) (input) (output) (parallel composition) (process replication) (channel restriction) (pattern matching) n u (e,f) fst(e) snd(e) inl(e) inr(e) Expression (integer constant) (name) (pair) (first projection) (second projection) (left injection) (right injection) ::= | | | | | | 5 Table 1: Syntax of processes and expressions. Session types [6, 7] describe linearized channels, namely channels that can be used for multiple communications, but only in a sequential way. There is a tight connection between linear and linearized channels: as shown in [13, 4, 3, 2], linearized channels can be encoded in the linear π-calculus. A consequence of this encoding is that the type reconstruction algorithm we present in this article can be used for inferring possibly infinite session types (we will see an example of this feature in Section 7). The task of reconstructing session types directly has been explored in [17], but for finite types only. Structure of the paper. We present the calculus in Section 2 and the type system in Section 3. The type reconstruction algorithm consists of a constraint generation phase (Section 4) and a constraint resolution phase (Section 5). We discuss some important issues related to the implementation of the algorithm in Section 6 and a few more elaborate examples in Section 7. Section 8 concludes and hints at some ongoing and future work. Proofs of the results in Sections 3 and 4 are in Appendixes A and B, respectively. Appendix C illustrates a few typing derivations of examples discussed in Section 5. A proof-of-concept implementation of the algorithm is available on the author’s home page. 2. The π-calculus with data types In this section we define the syntax and operational semantics of the formal language we work with, which is an extension of the π-calculus featuring base and composite data types and a pattern matching construct. 6 L. PADOVANI [s-par 1] [s-par 2] [s-par 3] [s-rep] idle | P ≡ P P |Q≡Q|P P | (Q | R) ≡ (P | Q) | R ∗P 4 ∗P | P [s-res 2] [s-res 1] new a in new b in P ≡ new b in new a in P a 6∈ fn(Q) (new a in P ) | Q ≡ new a in (P | Q) Table 2: Structural pre-congruence for processes. 2.1. Syntax. Let us introduce some notation first. We use integer numbers m, n, . . . , a countable set of channels a, b, . . . , and a countable set of variables x, y, . . . which is disjoint from the set of channels; names u, v, . . . are either channels or variables. The syntax of expressions and processes is given in Table 1. Expressions e, f, . . . are either integers, names, pairs (e,f) of expressions, the i-th projection of an expression i(e) where i ∈ {fst, snd}, or the injection i(e) of an expression e using the constructor i ∈ {inl, inr}. Using projections fst and snd instead of a pair splitting construct, as found for instance in [24, 23], is somewhat unconventional, but helps us highlighting some features of our type system. We will discuss some practical aspects of this choice in Section 6.3. Values v, w, . . . are expressions without variables and occurrences of the projections fst and snd. Processes P , Q, . . . comprise and extend the standard constructs of the asynchronous π-calculus. The idle process performs no action; the input process e?(x).P waits for a message v from the channel denoted by e and continues as P where x has been replaced by v; the output process e!f sends the value resulting from the evaluation of f on the channel resulting from the evaluation of e; the composition P | Q executes P and Q in parallel; the replication *P denotes infinitely many copies of P executing in parallel; the restriction new a in P creates a new channel a with scope P . In addition to these, we include a pattern matching construct case e {i(xi ) ⇒ Pi }i=inl,inr which evaluates e to a value of the form i(v) for some i ∈ {inl, inr}, binds v to xi and continues as Pi . The notions of free names fn(P ) and bound names bn(P ) of P are as expected, recalling that case e {i(xi ) ⇒ Pi }i=inl,inr binds xi in Pi . We identify processes modulo renaming of bound names and we write e{v/x} and P {v/x} for the capture-avoiding substitutions of v for the free occurrences of x in e and P , respectively. Occasionally, we omit idle when it is guarded by a prefix. 2.2. Operational semantics. The operational semantics of the language is defined in terms of a structural pre-congruence relation for processes, an evaluation relation for expressions, and a reduction relation for processes. Structural pre-congruence 4 is meant to rearrange process terms which should not be distinguished. The relation is defined in Table 2, where we write P ≡ Q in place of the two inequalities P 4 Q and Q 4 P . Overall ≡ coincides with the conventional structural congruence of the π-calculus, except that, as in [12], we omit the relation *P | P 4 *P (the reason will be explained in Remark 3.10). ℓ Evaluation e ↓ v and reduction P −→ Q are defined in Table 3. Both relation are fairly standard. As in [15], reduction is decorated with a label ℓ that is either a channel or the special symbol τ : in [r-comm] the label is the channel a on which a message is exchanged; in [r-case] it is τ since pattern matching is an internal computation not involving TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS [e-int] [e-chan] n↓n a↓a 7 [e-pair] [e-fst] [e-snd] [e-inr], [e-inl] ei ↓ vi (i=1,2) (e1 ,e2 ) ↓ (v1 ,v2 ) e ↓ (v,w) fst(e) ↓ v e ↓ (v,w) snd(e) ↓ w e↓v [r-comm] ei ↓ a [r-case] (i=1,2) f↓v e ↓ k(v) case e {i(xi ) ⇒ Pi }i=inl,inr −→ Pk {v/xk } e1 !f | e2 ?(x).Q −→ Q{v/x} [r-par] [r-new 2] [r-new 1] P −→ P ℓ ℓ a ′ P | Q −→ P ′ | Q k ∈ {inl, inr} τ a ℓ k ∈ {inl, inr} k(e) ↓ k(v) P −→ Q P −→ Q τ new a in P −→ new a in Q ℓ 6= a ℓ new a in P −→ new a in Q [r-struct] P 4 P′ ℓ P ′ −→ Q′ Q′ 4 Q ℓ P −→ Q Table 3: Evaluation of expressions and reduction of processes. communications. Note that, as we allow expressions in input and output processes for both the subject and the object of a communication, rule [r-comm] provides suitable premises to evaluate them. Rules [r-par], [r-new 1], and [r-new 2] propagate labels through parallel compositions and restrictions. In [r-new 1], the label a becomes τ when it escapes the scope of a. Rule [r-struct] closes reduction under structural congruence. Example 2.1 (list sharing). Below are the desugared representations of P and Q discussed in Section 1: P′ Q′ def = *odd?(z). case fst(z) of inl(_) ⇒ snd(snd(z))!fst(snd(z)) inr(x) ⇒ fst(x)?(y).even!(snd(x),(fst(snd(z)) + y,snd(snd(z)))) def = *even?(z). case fst(z) of inl(_) ⇒ snd(snd(z))!fst(snd(z)) inr(x) ⇒ odd!(snd(x),(fst(snd(z)),snd(snd(z)))) where the constructors inl and inr respectively replace Nil and Cons, inl has an (unused) argument denoted by the anonymous variable _, and tuple components are accessed using (possibly repeated) applications of fst and snd.  8 L. PADOVANI 3. Type system In this section we define a type system for the language presented in Section 2. The type system extends the one for the linear π-calculus [15] with composite and possibly infinite, regular types. The key feature of the linear π-calculus is that channel types are enriched with information about the number of times the channels they denote are used for input/output operations. Such number is abstracted into a use κ, . . . , which is an element of the set {0, 1, ω} where 0 and 1 obviously stand for no use and one use only, while ω stands for any number of uses. Definition 3.1 (types). Types, ranged over by t, s, . . . , are the possibly infinite regular trees built using the nullary constructor int, the unary constructors [ · ]κ1 ,κ2 for every combination of κ1 and κ2 , the binary constructors · × · (product) and · ⊕ · (disjoint sum). The type [t]κ1 ,κ2 denotes channels for exchanging messages of type t. The uses κ1 and κ2 respectively denote how many input and output operations are allowed on the channel. For example: a channel with type [t]0,1 cannot be used for input and must be used once for sending a message of type t; a channel with type [t]0,0 cannot be used at all; a channel with type [t]ω,ω can be used any number of times for sending and/or receiving messages of type t. A product t1 × t2 describes pairs (v1 ,v2 ) where vi has type ti for i = 1, 2. A disjoint sum t1 ⊕ t2 describes values of the form inl(v) where v has type t1 or of the form inr(v) where v has type t2 . Throughout the paper we let ⊙ stand for either × or ⊕. We do not provide a concrete, finite syntax for denoting infinite types and work directly with regular trees instead. Recall that a regular tree is a partial function from paths to type constructors (see e.g. [22, Chapter 21]), it consists of finitely many distinct subtrees, and admits finite representations using either the well-known µ notation or finite systems of equations [1] (our implementation internally uses both). Working directly with regular trees gives us the coarsest possible notion of type equality (t = s means that t and s are the same partial function) and it allows us to reuse some key results on regular trees that will be essential in the following. In particular, throughout the paper we will implicitly use the next result to define types as solutions of particular systems of equations: Theorem 3.2. Let {αi = Ti | 1 ≤ i ≤ n} be a finite system of equations where each Ti is a finite term built using the constructors in Definition 3.1 and the pairwise distinct unknowns {α1 , . . . , αn }. If none of the Ti is an unknown, then there exists a unique substitution σ = {αi 7→ ti | 1 ≤ i ≤ n} such that ti = σTi and ti is a regular tree for each 1 ≤ i ≤ n. Proof. All the right hand sides of the equations are finite – hence regular – and different from an unknown, therefore this result is just a particular case of [1, Theorem 4.3.1]. Example 3.3 (integer stream). The type of integer streams int × (int × (int × · · · )) is the unique regular tree t such that t = int × t. To make sense out of this statement we have to be sure that such t does exist and is indeed unique. Consider the equation α = int × α obtained from the above equality by turning each occurrence of the metavariable t into the unknown α and observe that the right hand side of such equation is not an unknown. By Theorem 3.2, there exists a unique regular tree t such that t = int × t. Note that t consists of two distinct subtrees, int and t itself.  Example 3.4 (lists). To verify the existence of the types todd and teven informally introduced in Section 1, consider the system of equations {α1 = int ⊕ ([int]1,0 × α2 ), α2 = int ⊕ ([int]0,0 × α1 )} TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 9 obtained by turning the metavariables todd and teven in (1.3) respectively into the unknowns α1 and α2 and by using basic types and disjoint sums in place of the list constructors Nil and Cons. Theorem 3.2 says that there exist two unique regular trees todd and teven such that todd = int ⊕ ([int]1,0 × teven ) and teven = int ⊕ ([int]0,0 × todd ). Similarly, tlist is  the unique type such that tlist = int ⊕ ([int]1,0 × tlist ). We now define some key notions on uses and types. To begin with, we define a binary operation + on uses that allows us to express the combined use κ1 + κ2 of a channel that is used both as denoted by κ1 and as denoted by κ2 . Formally:   κ1 if κ2 = 0 def (3.1) κ1 + κ2 = κ2 if κ1 = 0   ω otherwise Note that 0 is neutral and ω is absorbing for + and that 1 + 1 = ω, since ω is the only use allowing us to express the fact that a channel is used twice. In a few places we will write 2κ as an abbreviation for κ + κ. We now lift the notion of combination from uses to types. Since types may be infinite, we resort to a coinductive definition. Definition 3.5 (type combination). Let Ctype be the largest relation between pairs of types and types such that ((t1 , t2 ), s) ∈ Ctype implies either: • t1 = t2 = s = int, or • t1 = [t]κ1 ,κ2 and t2 = [t]κ3 ,κ4 and s = [t]κ1 +κ3 ,κ2 +κ4 , or • t1 = t11 ⊙ t12 and t2 = t21 ⊙ t22 and s = s1 ⊙ s2 and ((t1i , t2i ), si ) ∈ Ctype for i = 1, 2. Observe that Ctype is a partial binary function on types, that is ((t1 , t2 ), s1 ) ∈ Ctype and ((t1 , t2 ), s2 ) ∈ Ctype implies s1 = s2 . When (t, s) ∈ dom(Ctype ), we write t + s for Ctype (t, s), that is the combination of t and s. Occasionally we also write 2t in place of t + t. Intuitively, basic types combine with themselves and the combination of channel types with equal message types is obtained by combining corresponding uses. For example, we have [int]0,1 + [int]1,0 = [int]1,1 and [[int]1,0 ]0,1 + [[int]1,0 ]1,1 = [[int]1,0 ]1,ω . In the latter example, note that the uses of channel types within the top-most ones are not combined together. Type combination propagates component-wise on composite types. For instance, we have ([int]0,1 × [int]0,0 ) + ([int]0,0 × [int]1,0 ) = ([int]0,1 + [int]0,0 ) × ([int]0,0 + [int]1,0 ) = [int]0,1 × [int]1,0 . Unlike use combination, type combination is a partial operation: it is undefined to combine two types having different structures, or to combine two channel types carrying messages of different types. For example, int+[int]0,0 is undefined and so is [[int]0,0 ]0,1 + [[int]0,1 ]0,1 , because [int]0,0 and [int]0,1 differ. Types that can be combined together play a central role, so we name a relation that characterizes them: Definition 3.6 (coherent types). We say that t and s are structurally coherent or simply coherent, notation t ∼ s, if t + s is defined, namely there exists t′ such that ((t, s), t′ ) ∈ Ctype . Observe that ∼ is an equivalence relation, implying that a type can always be combined with itself (i.e., 2t is always defined). Type combination is also handy for characterizing a fundamental partitioning of types: Definition 3.7 (unlimited and linear types). We say that t is unlimited, notation un(t), if 2t = t. We say that it is linear otherwise. 10 L. PADOVANI Channel types are either linear or unlimited depending on their uses. For example, [t]0,0 is unlimited because [t]0,0 +[t]0,0 = [t]0,0 , whereas [t]1,0 is linear because [t]1,0 +[t]1,0 = [t]ω,0 6= [t]1,0 . Similarly, [t]ω,ω is unlimited while [t]0,1 and [t]1,1 are linear. Other types are linear or unlimited depending on the channel types occurring in them. For instance, [t]0,0 × [t]1,0 is linear while [t]0,0 × [t]ω,0 is unlimited. Note that only the topmost channel types of a type matter. For example, [[t]1,1 ]0,0 is unlimited despite of the fact that it contains the subterm [t]1,1 which is itself linear, because such subterm is found within an unlimited channel type. We use type environments to track the type of free names occurring in expressions and processes. Type environments Γ , . . . are finite maps from names to types that we write as u1 : t1 , . . . , un : tn . We identify type environments modulo the order of their associations, write ∅ for the empty environment, dom(Γ ) for the domain of Γ , namely the set of names for which there is an association in Γ , and Γ1 , Γ2 for the union of Γ1 and Γ2 when dom(Γ1 ) ∩ dom(Γ2 ) = ∅. We also extend the partial combination operation + on types to a partial combination operation on type environments, thus: ( Γ1 , Γ2 if dom(Γ1 ) ∩ dom(Γ2 ) = ∅ def Γ1 + Γ2 = (3.2) ′ ′ (Γ1 + Γ2 ), u : t1 + t2 if Γi = Γi′ , u : ti for i = 1, 2 The operation + extends type combination in [15] and the ⊎ operator in [24]. Note that Γ1 + Γ2 is undefined if there is u ∈ dom(Γ1 ) ∩ dom(Γ2 ) such that Γ1 (u) + Γ2 (u) is undefined. Note also that dom(Γ1 + Γ2 ) = dom(Γ1 ) ∪ dom(Γ2 ). Thinking of type environments as of specifications of the resources used by expressions/processes, Γ1 + Γ2 expresses the combined use of the resources specified in Γ1 and Γ2 . Any resource occurring in only one of these environments occurs in Γ1 + Γ2 ; any resource occurring in both Γ1 and Γ2 is used according to the combination of its types in Γ1 + Γ2 . For example, if a process sends an integer over a channel a, it will be typed in an environment that contains the association a : [int]0,1 ; if another process uses the same channel a for receiving an integer, it will be typed in an environment that contains the association a : [int]1,0 . Overall, the parallel composition of the two processes uses channel a according to the type [int]0,1 + [int]1,0 = [int]1,1 and therefore it will be typed in an environment that contains the association a : [int]1,1 . The last notion we need before presenting the type rules is that of an unlimited type environment. This is a plain generalization of the notion of unlimited type, extended to the range of a type environment. We say that Γ is unlimited, notation un(Γ ), if un(Γ (u)) for every u ∈ dom(Γ ). A process typed in an unlimited type environment need not use any of the resources described therein. Type rules for expressions and processes are presented in Table 4. These rules are basically the same as those found in the literature [15, 10]. The possibility of sharing data structures among several processes, which we have exemplified in Section 1, is a consequence of our notion of type combination extended to composite regular types. Type rules for expressions are unremarkable. Just observe that unused type environments must be unlimited. Also, the projections fst and snd discard one component of a pair, so the discarded component must have an unlimited type. Let us move on to the type rules for processes. The idle process does nothing, so it is well typed only in an unlimited environment. Rule [t-in] types an input process e?(x).P . The subject e must evaluate to a channel whose input use is either 1 or ω and whose output use is either 0 or ω. We capture the first condition saying that the input use of the channel has the form 1 + κ1 for some κ1 , and the second condition saying that the output use of TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 11 Expressions [t-int] [t-name] [t-inl] [t-inr] un(Γ ) Γ ⊢ n : int un(Γ ) Γ, u : t ⊢ u : t Γ ⊢e:t Γ ⊢ inl(e) : t ⊕ s Γ ⊢e:s Γ ⊢ inr(e) : t ⊕ s [t-pair] [t-fst] [t-snd] Γi ⊢ ei : ti (i=1,2) Γ1 + Γ2 ⊢ (e1 ,e2 ) : t1 × t2 Γ ⊢e:t×s un(s) Γ ⊢ fst(e) : t Γ ⊢e:t×s un(t) Γ ⊢ snd(e) : s Processes [t-idle] un(Γ ) Γ ⊢ idle [t-out] [t-in] 1+κ1 ,2κ2 Γ1 ⊢ e : [t] Γ2 , x : t ⊢ P Γ1 + Γ2 ⊢ e?(x).P Γ1 ⊢ e : [t]2κ1 ,1+κ2 Γ2 ⊢ f : t Γ1 + Γ2 ⊢ e!f [t-rep] [t-par] [t-new] Γ ⊢P un(Γ ) Γ ⊢ *P Γi ⊢ Pi (i=1,2) Γ1 + Γ2 ⊢ P1 | P2 Γ , a : [t]κ,κ ⊢ P Γ ⊢ new a in P [t-case] Γ1 ⊢ e : t ⊕ s Γ2 , xi : t ⊢ Pi (i=inl,inr) Γ1 + Γ2 ⊢ case e {i(xi ) ⇒ Pi }i=inl,inr Table 4: Type rules for expressions and processes. the channel has the form 2κ2 for some κ2 . The continuation P is typed in an environment enriched with the association for the received message x. Note the combination Γ1 + Γ2 in the conclusion of rule [t-in]. In particular, if e evaluates to a linear channel, its input capability is consumed by the operation and such channel can no longer be used for inputs in the continuation. Rule [t-out] types an output process e!f. The rule is dual to [t-in] in that it requires the channel to which e evaluates to have a positive output use. Rule [t-rep] states that a replicated process *P is well typed in the environment Γ provided that P is well typed in an unlimited Γ . The rationale is that *P stands for an unbounded number of copies of P composed in parallel, hence P cannot contain (free) linear channels. The rules [t-par] and [t-case] are conventional, with the by now familiar use of environment combination for properly distributing linear resources to the various subterms of a process. The rule [t-new] is also conventional. We require the restricted channel to have the same input and output uses. While this is not necessary for the soundness of the type system, in practice it is a reasonable requirement. We also argue that this condition is important for the modular application of the type reconstruction algorithm; we will discuss this aspect more in detail in Section 6. As in many behavioral type systems, the type environment in which the reducing process is typed may change as a consequence of the reduction. More specifically, reductions involving a communication on channels consume 1 unit from both the input and output uses 12 L. PADOVANI of the channel’s type. In order to properly state subject reduction, we define a reduction ℓ relation over type environments. In particular, we write −→ for the least relation between type environments such that τ a Γ + a : [t]1,1 −→ Γ Γ −→ Γ τ In words, −→ denotes an internal computation (pattern matching) or a communication on some restricted channel which does not consume any resource from the type environment, a while −→ denotes a communication on channel a which consumes 1 use from both the input and output slots in a’s type. For example, we have a a : [int]1,1 −→ a : [int]0,0 a def by taking Γ = a : [int]0,0 in the definition of −→ above, since Γ +a : [int]1,1 = a : [int]1,1 . The residual environment denotes the fact that the (linear) channel a can no longer be used for communication. Now we have: ℓ ℓ Theorem 3.8. Let Γ ⊢ P and P −→ Q. Then Γ ′ ⊢ Q for some Γ ′ such that Γ −→ Γ ′ . Theorem 3.8 establishes not only a subject reduction result, but also a soundness result because it implies that a channel is used no more than its type allows. It is possible to establish more properties of the linear π-calculus, such as the fact that communications involving linear channels enjoy partial confluence. In this work we focus on the issue of type reconstruction. The interested reader may refer to [15] for further results. Example 3.9. We consider again the processes P ′ and Q′ in Example 2.1 and sketch a few key derivation steps to argue that they are well typed. To this aim, consider the types todd , teven , and tzero that satisfy the equalities below todd teven tzero = int ⊕ ([int]0,1 × teven ) = int ⊕ ([int]0,0 × todd ) = int ⊕ ([int]0,0 × tzero ) and also consider the types of the messages respectively carried by odd and even: sodd seven def = todd × (int × [int]0,1 ) def = teven × (int × [int]0,1 ) Now, in the inl branch of P ′ we derive (D1) z : tzero × (int × [int]0,1 ) ⊢ z : tzero × (int × [int]0,1 ) z : tzero × (int × [int]0,1 ) ⊢ snd(z) : int × [int]0,1 z : tzero × (int × [int]0,1 ) ⊢ snd(snd(z)) : [int]0,1 using the fact that un(tzero ) and un(int). We also derive (D2) z : tzero × (int × [int]0,0 ) ⊢ fst(snd(z)) : int [t-snd] [t-snd] z : tzero × (int × [int]0,0 ) ⊢ z : tzero × (int × [int]0,0 ) z : tzero × (int × [int]0,0 ) ⊢ snd(z) : int × [int]0,0 [t-name] [t-name] [t-snd] [t-fst] TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 13 using the fact that un(tzero ) and un([int]0,0 ), therefore we derive (D3) (D1) (D2) z : tzero × (int × [int]0,1 ), _ : int ⊢ snd(snd(z))!fst(snd(z)) using the combination [t-out] (tzero × (int × [int]0,1 )) + (tzero × (int × [int]0,0 )) = tzero × (int × [int]0,1 ) Already in this sub-derivation we appreciate that although the pair z is accessed twice, its type in the conclusion of (D3) correctly tracks the fact that the channel contained in z is only used once, for an output. For the inr branch in P ′ there exists another derivation (D4) concluding .. .. . . even : [seven ]0,ω , x : [int]1,0 × teven , z : tzero × int × [int]1,0 ⊢ fst(x)?(y). · · · Now we conclude z : sodd ⊢ z : sodd [t-name] (D3) (D4) even : [seven ]0,ω , z : sodd ⊢ case z of · · · [t-in] [t-case] odd : [sodd ]ω,0 , even : [seven ]0,ω ⊢ odd?(z).case z of · · · [t-in] [t-rep] odd : [sodd ]ω,0 , even : [seven ]0,ω ⊢ P ′ Note that odd and even must be unlimited channels because they occur free in a replicated process, for which rule [t-rep] requires an unlimited environment. A similar derivation shows that Q′ is well typed in an environment where the types of odd and even have swapped uses .. . [t-rep] odd : [sodd ]0,ω , even : [seven ]ω,0 ⊢ Q′ so the combined types of odd and even are [sodd ]ω,ω and [seven ]ω,ω , respectively. Using these, we find a typing derivation for the process R in Section 1. Proceeding bottom-up we have .. . odd : [sodd ]ω,ω , l : todd , a : [int]0,1 ⊢ odd!(l,0,a) and [t-out] .. . even : [seven ]ω,ω , l : teven , b : [int]0,1 ⊢ even!(l,0,b) [t-out] as well as .. . [t-in] a : [int]1,0 , b : [int]1,0 , r : [int]0,1 ⊢ a?(x).b?(y).r!(x + y) from which we conclude .. . ==========ω,ω ==============ω,ω ================0,1 =============== [t-new] (twice) odd : [sodd ] , even : [seven ] , l : tlist , r : [int] ⊢ new a,b in · · · using the property todd + teven = tlist .  14 L. PADOVANI We conclude this section with a technical remark to justify the use of a structural precongruence relation in place of a more familiar symmetric one. Remark 3.10. Let us show why the relation *P | P 4 *P would invalidate Theorem 3.8 (more specifically, Lemma A.3) in our setting (a similar phenomenon is described in [12]). To this aim, consider the process def P = a?(x).new c in (*c?(y).c!y | c!b) def and the type environment Γκ = a : [int]ω,0 , b : [int]0,κ for an arbitrary κ. We can derive .. .. . . [t-out] .. c : [[int]0,κ ]ω,ω , y : [int]0,κ ⊢ c!y . [t-in] .. .. . . c : [[int]0,κ ]ω,ω ⊢ c?(y).c!y c : [[int]0,κ ]ω,ω ⊢ *c?(y).c!y .. . [t-rep] Γκ , c : [[int]0,κ ]0,1 ⊢ c!b Γκ , c : [[int]0,κ ]ω,ω ⊢ *c?(y).c!y | c!b Γκ ⊢ new c in · · · [t-out] [t-par] [t-new] [t-in] Γκ ⊢ P where we have elided a few obvious typing derivations for expressions. In particular, we can find a derivation where b has an unlimited type (κ = 0) and another one where b has a linear type (κ = 1). This is possible because channel c, which is restricted within P , can be given different types – respectively, [[int]0,0 ]ω,ω and [[int]0,1 ]ω,ω – in the two derivations. We can now obtain .. . .. . Γ ⊢P 0 Γ0 ⊢ *P [t-rep] Γ1 ⊢ P [r-par] Γ1 ⊢ *P | P because un(Γ0 ) and Γ0 + Γ1 = Γ1 . If we allowed the structural congruence rule *P | P 4 *P , then Γ1 ⊢ *P would not be derivable because Γ1 is linear, hence typing would not be preserved by structural pre-congruence. This problem is avoided in [15, 10] by limiting replication to input prefixes, omitting any structural congruence rule for replications, and adding a dedicated synchronization rule for them. In [15] it is stated that “the full picalculus replication operator poses no problems for the linear type system”, but this holds because there the calculus is typed, so multiple typing derivations for the same process P above would assign the same type to c and, in turn, the same type to b.  4. Constraint Generation We formalize the problem of type reconstruction as follows: given a process P , find a type environment Γ such that Γ ⊢ P , provided there is one. In general, in the derivation for Γ ⊢ P we also want to identify as many linear channels as possible. We will address this latter aspect in Section 5. TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS T, S ::= | | | | α int [T]U,V T×S T⊕S Type expression (type variable) (integer) (channel) (product) (disjoint sum) U, V ::= | | ̺ κ U+V 15 Use expression (use variable) (use constant) (use combination) Table 5: Syntax of use and type expressions. 4.1. Syntax-directed generation algorithm. The type rules shown in Table 4 rely on a fair amount of guessing that concerns the structure of types in the type environment, how they are split/combined using +, and the uses occurring in them. So, these rules cannot be easily interpreted as a type reconstruction algorithm. The way we follow to define one is conventional: first, we give an alternative set of (almost) syntax-directed rules that generate constraints on types; then, we search for a solution of such constraints. The main technical challenge is that we cannot base our type reconstruction algorithm on conventional unification because we have to deal with constraints expressing not only the equality between types and uses, but also the combination of types and uses. In addition, we work with possibly infinite types. To get started, we introduce use and type expressions, which share the same structure as uses/types but they differ from them in two fundamental ways: (1) We allow use/type variables to stand for unknown uses/types. (2) We can express symbolically the combination of use expressions. We therefore introduce a countable set of use variables ̺, . . . as well as a countable set of type variables α, β, . . . ; the syntax of use expressions U, V, . . . and of type expressions T, S, . . . is given in Table 5. Observe that every use is also a use expression and every finite type is also a type expression. We say that T is proper if it is different from a type variable. Constraints ϕ, . . . are defined by the grammar below: ϕ ::= | | | T= ˆ S T= ˆ S1 + S2 T∼ ˆ S U= ˆ V Constraint (type equality) (type combination) (type coherence) (use equality) Constraints express relations between types/uses that must be satisfied in order for a given process to be well typed. In particular, we need to express equality constraints between types (T = ˆ S) and uses (U = ˆ V), coherence constraints (T ∼ ˆ S), and combination constraints between types (T = ˆ S1 + S2 ). We will write un(T) as an abbreviation for the constraint T = ˆ T + T. This notation is motivated by Definition 3.7, according to which a type is unlimited if and only if it is equal to its own combination. We let C, . . . range over finite constraint sets. The set of expressions of a constraint set C, written expr(C), is the (finite) set of use and type expressions occurring in the constraints in C. The type reconstruction algorithm generates type environments for the expressions and processes being analyzed. Unlike the environments in Section 3, these environments associate names with type expressions. For this reason we will let ∆, . . . range over the 16 L. PADOVANI [c-env 1] [c-env 2] dom(∆1 ) ∩ dom(∆2 ) = ∅ ∆1 ⊔ ∆2 ∆1 , ∆2 ; ∅ ∆1 ⊔ ∆2 ∆; C α fresh (∆1 , u : T) ⊔ (∆2 , u : S) ∆, u : α; C ∪ {α = ˆ T + S} [m-env 1] ∅⊓∅ ∅; ∅ [m-env 2] ∆1 ⊓ ∆2 (∆1 , u : T) ⊓ (∆2 , u : S) ∆; C ∆, u : T; C ∪ {T = ˆ S} Table 6: Combining and merging operators for type environments. environments generated by the reconstruction algorithm, although we will refer to them as type environments. The algorithm also uses two auxiliary operators ⊔ and ⊓ defined in Table 6. The relation ∆1 ⊔ ∆2 ∆; C combines two type environments ∆1 and ∆2 into ∆ when the names in dom(∆1 ) ∪ dom(∆2 ) are used both as specified in ∆1 and also as specified in ∆2 and, in doing so, generates a set of constraints C. So ⊔ is analogous to + in (3.2). When ∆1 and ∆2 have disjoint domains, ∆ is just the union of ∆1 and ∆2 and no constraints are generated. Any name u that occurs in dom(∆1 )∩dom(∆2 ) is used according to the combination of ∆1 (u) and ∆2 (u). In general, ∆1 (u) and ∆2 (u) are type expressions with free type variables, hence this combination cannot be “computed” or “checked” right away. Instead, it is recorded as the constraint α = ˆ ∆1 (u) + ∆2 (u) where α is a fresh type variable. The relation ∆1 ⊓ ∆2 ∆; C merges two type environments ∆1 and ∆2 into ∆ when the names in dom(∆1 ) ∪ dom(∆2 ) are used either as specified in ∆1 or as specified in ∆2 and, in doing so, generates a constraint set C. This merging is necessary when typing the alternative branches of a case: recall that rule [t-case] in Table 4 requires the same type environment Γ for typing the two branches of a case. Consequently, ∆1 ⊓ ∆2 is defined only when ∆1 and ∆2 have the same domain, and produces a set of constraints C saying that the corresponding types of the names in ∆1 and ∆2 must be equal. The rules of the type reconstruction algorithm are presented in Table 7 and derive judgments e : T ◮ ∆; C for expressions and P ◮ ∆; C for processes. In both cases, ∆ is the generated environment that contains associations for all the free names in e and P , while C is the set of constraints that must hold in order for e or P to be well typed in ∆. In a judgment e : T ◮ ∆; C, the type expression T denotes the type of the expression e. There is a close correspondence between the type system (Table 4) and the reconstruction algorithm (Table 7). In a nutshell, unknown uses and types become fresh use and type variables (all use/type variables introduced by the rules are assumed to be fresh), every application of + in Table 4 becomes an application of ⊔ in Table 7, and every assumption on the form of types becomes a constraint. Constraints accumulate from the premises to the conclusion of each rule of the reconstruction algorithm, which we now review briefly. Rule [i-int] deals with integer constants. Their type is obviously int, they contain no free names and therefore they generate the empty environment and the empty set of constraints. Rule [i-name] deals with the free occurrence of a name u. A fresh type variable standing for the type of this occurrence of u is created and used in the resulting type environment u : α. Again, no constraints are generated. In general, different occurrences of the same name may have different types which are eventually combined with α later on in the reconstruction process. In rules [i-inl] and [i-inr] the type of the summand that TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 17 Expressions [i-int] [i-name] n : int ◮ ∅; ∅ u : α ◮ u : α; ∅ [i-inl] [i-inr] e : T ◮ ∆; C inl(e) : T ⊕ α ◮ ∆; C e : T ◮ ∆; C inr(e) : α ⊕ T ◮ ∆; C [i-pair] ei : Ti ◮ ∆i ; Ci (i=1,2) ∆1 ⊔ ∆2 ∆; C3 (e1 ,e2 ) : T1 × T2 ◮ ∆; C1 ∪ C2 ∪ C3 [i-fst] [i-snd] e : T ◮ ∆; C fst(e) : α ◮ ∆; C ∪ {T = ˆ α × β, un(β)} e : T ◮ ∆; C snd(e) : β ◮ ∆; C ∪ {T = ˆ α × β, un(α)} Processes [i-idle] idle ◮ ∅; ∅ [i-in] e : T ◮ ∆1 ; C1 P ◮ ∆2 , x : S; C2 ∆1 ⊔ ∆2 e?(x).P ◮ ∆; C1 ∪ C2 ∪ C3 ∪ {T = ˆ [S] 1+̺1 ,2̺2 ∆; C3 } [i-rep] [i-out] e : T ◮ ∆1 ; C1 f : S ◮ ∆2 ; C2 ∆1 ⊔ ∆2 ∆; C3 e!f ◮ ∆; C1 ∪ C2 ∪ C3 ∪ {T = ˆ [S]2̺1 ,1+̺2 } P ◮ ∆; C ∆⊔∆ ∆′ ; C ′ *P ◮ ∆′ ; C ∪ C ′ [i-par] [i-new] Pi ◮ ∆i ; Ci (i=1,2) ∆1 ⊔ ∆2 ∆; C3 P1 | P2 ◮ ∆; C1 ∪ C2 ∪ C3 P ◮ ∆, a : T; C new a in P ◮ ∆; C ∪ {T = ˆ [α]̺,̺ } [i-case] e : T ◮ ∆1 ; C1 ∆inl ⊓ ∆inr Pi ◮ ∆i , xi : Ti ; Ci (i=inl,inr) ∆2 ; C2 ∆1 ⊔ ∆2 ∆; C3 case e {i(xi ) ⇒ Pi }i=inl,inr ◮ ∆; C1 ∪ C2 ∪ C3 ∪ Cinl ∪ Cinr ∪ {T = ˆ Tinl ⊕ Tinr } [i-weak] P ◮ ∆; C P ◮ ∆, u : α; C ∪ {un(α)} Table 7: Constraint generation for expressions and processes. was guessed in [t-inl] and [t-inr] becomes a fresh type variable. Rule [t-pair] creates a product type from the type of the components of the pairs, combines the corresponding environments and joins all the constraints generated in the process. Rules [i-fst] and [i-snd] deal with pair projections. The type T of the projected expression must be a product of the form α × β. Since the first projection discards the second component of a pair, β must be unlimited in [i-fst]. Symmetrically for [i-snd]. 18 L. PADOVANI Continuing on with the rules for processes, let us consider [i-in] and [i-out]. The main difference between these rules and the corresponding ones [t-in] and [t-out] is that the use information of the channel on which the communication occurs is unknown, hence it is represented using fresh use variables. The 1 + ̺i part accounts for the fact that the channel is being used at least once, for an input or an output. The 2̺j part accounts for the fact that the use information concerning the capability (either input or output) that is not exercised must be unlimited (note that we extend the notation 2κ to use expressions). Rule [i-rep] deals with a replicated process *P . In the type system, *P is well typed in an unlimited environment. Here, we are building up the type environment for *P and we do so by combining the environment ∆ generated by P with itself. The rationale is that ∆ ⊔ ∆ yields an unlimited type environment that grants at least all the capabilities granted by ∆. By now most of the main ingredients of the constraint generation algorithm have been revealed, and the remaining rules contain no further novelties but the expected use of the merging operator ⊓ in [i-case]. There is, however, a rule [i-weak] that has no correspondence in Table 4. This rule is necessary because [i-in], [i-new], and [i-case], which correspond to the binding constructs of the calculus, assume that the names they bind do occur in the premises on these rules. But since type environments are generated by the algorithm as it works through an expression or a process, this may not be the case if a bound name is never used and therefore never occurs in that expression or process. Furthermore, the ⊓ operator is defined only on type environments having the same domain. This may not be the case if a name occurs in only one branch of a pattern matching, and not in the other one. With rule [i-weak] we can introduce missing names in type environments wherever this is necessary. Naturally, an unused name has an unknown type α that must be unlimited, whence the constraint un(α) (see Example 4.4 for an instance where [i-weak] is necessary). Strictly speaking, with [i-weak] this set of rules is not syntax directed, which in principle is a problem if we want to consider this as an algorithm. In practice, the places where [i-weak] may be necessary are easy to spot (in the premises of all the aforementioned rules for the binding constructs). What we gain with [i-weak] is a simpler presentation of the rules for constraint generation. 4.2. Correctness and completeness. If the constraint set generated from P is satisfiable, then it corresponds to a typing for P . To formalize this property, we must first define what “satisfiability” means for a constraint set. A substitution σ is a finite map from type variables to types and from use variables to uses. We write dom(σ) for the set of type and use variables for which there is an association in σ. The application of a substitution σ to a use/type expression U/T, respectively denoted by σU and σT, replaces use variables ̺ and type variables α in U/T with the corresponding uses σ(̺) and types σ(α) and computes use combinations whenever possible:   σ(α) if T = α ∈ dom(σ)     if U = ̺ ∈ dom(σ) σ(̺) [σS]σU,σV if T = [S]U,V def def σU = σU1 + σU2 if U = U1 + U2 σT =   σT1 ⊙ σT2 if T = T1 ⊙ T2    U otherwise  T otherwise We will make sure that the application of a substitution σ to a type expression T is always well defined: either dom(σ) contains no type variables, in which case σT is a type expression, or dom(σ) includes all use/type variables occurring in T, in which case we say that σ covers TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 19 T and σT is a type. We extend application pointwise to type environments, namely def σ∆ = {u : σ∆(u) | u ∈ dom(∆)}, and we say that σ covers ∆ if it covers all the type expressions in the range of ∆. Definition 4.1 (solution, satisfiability, equivalence). A substitution σ is a solution for a constraint set C if it covers all the T ∈ expr(C) and the following conditions hold: • T= ˆ S ∈ C implies σT = σS, and • T= ˆ S1 + S2 ∈ C implies σT = σS1 + σS2 , and • T∼ ˆ S ∈ C implies σT ∼ σS, and • U= ˆ V ∈ C implies σU = σV. We say that C is satisfiable if it has a solution and unsatisfiable otherwise. We say that C1 and C2 are equivalent if they have the same solutions. We can now state the correctness result for the type reconstruction algorithm: Theorem 4.2. If P ◮ ∆; C and σ is a solution for C that covers ∆, then σ∆ ⊢ P . Note that Theorem 4.2 not only requires σ to be a solution for C, but also that σ must include suitable substitutions for all use and type variables occurring in ∆. Indeed, it may happen that ∆ contains use/type variables not involved in any constraint in C, therefore a solution for C does not necessarily cover ∆. The reconstruction algorithm is also complete, in the sense that each type environment Γ such that Γ ⊢ P can be obtained by applying a solution for C to ∆. Theorem 4.3. If Γ ⊢ P , then there exist ∆, C, and σ such that P ◮ ∆; C and σ is a solution for C that covers ∆ and Γ = σ∆. Example 4.4. Below we illustrate the reconstruction algorithm at work on the process new a in (a!3 | a?(x).idle) which will be instrumental also in the following section: idle ◮ ∅; ∅ a : α1 ◮ a : α1 ; ∅ 3 : int ◮ ∅; ∅ a!3 ◮ a : α1 ; {α1 = ˆ [int]2̺1 ,1+̺2 } [i-out] a : α2 ◮ a : α2 ; ∅ [i-idle] idle ◮ x : γ; {un(γ)} [i-weak] a?(x).idle ◮ a : α2 ; {α2 = ˆ [γ]1+̺3 ,2̺4 , un(γ)} a!3 | a?(x).idle ◮ a : α; {α = ˆ α1 + α2 , α1 = ˆ [int]2̺1 ,1+̺2 , α2 = ˆ [γ]1+̺3 ,2̺4 , un(γ)} [i-par] [i-new] new a in (a!3 | a?(x).idle) ◮ ∅; {α = ˆ [δ]̺5 ,̺5 , α = ˆ α1 + α2 , . . . } The synthesized environment is empty, since the process has no free names, and the resulting constraint set is {α = ˆ [δ]̺5 ,̺5 , α = ˆ α1 + α2 , α1 = ˆ [int]2̺1 ,1+̺2 , α2 = ˆ [γ]1+̺3 ,2̺4 , un(γ)} Observe that a is used twice and each occurrence is assigned a distinct type variable αi . Eventually, the reconstruction algorithm finds out that the same channel a is used simultaneously in different parts of the process, so it records the fact that the overall type α of a must be the combination of α1 and α2 in the constraint α = ˆ α1 + α2 . A solution for the obtained constraint set is the substitution {α 7→ [int]1,1 , α1 7→ [int]0,1 , α2 7→ [int]1,0 , γ 7→ int, δ 7→ int, ̺1..4 7→ 0, ̺5 7→ 1} [i-in] 20 L. PADOVANI confirming that a is a linear channel. This is not the only solution of the constraint set: another one can be obtained by setting all the use variables to ω, although in this case a is not recognized as a linear channel. Note also that the application of [i-in] is possible only if the name x of the received message occurs in the environment synthesized for the continuation process idle. Since the continuation process contains no occurrence of x, this name can only be introduced using [i-weak]. In general, [i-weak] is necessary to prove the completeness of the reconstruction algorithm as stated in Theorem 4.3. For example, x : int ⊢ idle is derivable according to the rules in Table 4, but as we have seen in the above derivation the reconstruction algorithm without [i-weak] would synthesize for idle an empty environment, not containing an association for x.  Example 4.5. We compute the constraint set of a simple process that accesses the same composite structure containing linear values. The process in Example 2.1 is too large to be discussed in full, so we consider the following, simpler process fst(x)?(y).snd(x)!(y + 1) which uses a pair x of channels and sends on the second channel in the pair the successor of the number received from the first channel (we assume that the language and the type reconstruction algorithm have been extended in the obvious way to support operations on numbers such as addition). We derive x : α1 ◮ x : α1 ; ∅ [i-name] fst(x) : β1 ◮ x : α1 ; {α1 = ˆ β1 × β2 , un(β2 )} for the first projection of x and x : α2 ◮ x : α2 ; ∅ [i-fst] [i-name] [i-snd] snd(x) : γ2 ◮ x : α2 ; {α2 = ˆ γ1 × γ2 , un(γ1 )} for the second projection of x. For the output operation we derive .. . y : δ ◮ y : δ; ∅ [i-name] 1 : int ◮ ∅; [i-int] y + 1 : int ◮ y : δ; {δ = ˆ int} snd(x)!(y + 1) ◮ x : α2 , y : δ; {α2 = ˆ γ1 × γ2 , un(γ1 ), γ2 = ˆ [int]2̺3 ,1+̺4 , δ = ˆ int} so for the whole process we obtain .. . fst(x)?(y).snd(x)!(y + 1) ◮ x : α; {α = ˆ α1 + α2 , α1 = ˆ β1 × β2 , α2 = ˆ γ1 × γ2 , β1 = ˆ [δ]1+̺1 ,2̺2 , γ2 = ˆ [int]2̺3 ,1+̺4 , un(β2 ), un(γ1 ), δ = ˆ int} [i-out] [i-in] Like in Example 4.4, here too the variable x is used multiple times and each occurrence is assigned a distinct type variable αi , but this time such type variables must be assigned with a pair type in order for the constraint set to be solved.  TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 21 5. Constraint Solving In this section we describe an algorithm that determines whether a given constraint set C is satisfiable and, if this is the case, computes a solution for C. Among all possible solutions for C, we strive to find one that allows us to identify as many linear channels as possible. To this aim, it is convenient to recall the notion of solution preciseness from [10]. Definition 5.1 (solution preciseness). Let ≤ be the total order on uses such that 0 ≤ 1 ≤ ω. Given two solutions σ1 and σ2 for a constraint set C, we say that σ1 is more precise than σ2 if σ1 (̺) ≤ σ2 (̺) for every ̺ ∈ expr(C). Roughly speaking, the preciseness of a solution is measured in terms of the numbers of unused and linear channels it identifies, which are related to the number of use variables assigned to 0 and 1. We will use Definition 5.1 as a guideline for developing our algorithm, although the algorithm may be unable to find the most precise solution. There are two reasons for this. First, there can be solutions with minimal use assignments that are incomparable according to Definition 5.1. This is related to the fact that the type system presented in Section 3 lacks the principal typing property. Second, to ensure termination when constraints concern infinite types, our algorithm makes some simplifying assumptions that may – in principle – imply a loss of precision of the resulting solution (see Example 5.11). Despite this, experience with the implementation suggests that the algorithm is indeed capable of identifying as many unused and linear channels as possible in practical situations, even when infinite types are involved. Before embarking in the technical description of the algorithm, we survey the key issues that we have to address and how they are addressed. 5.1. Overview. We begin by considering again the simple process below new a in (a!3 | a?(x).idle) (5.1) for which we have shown the reconstruction algorithm at work in Example 4.4. The process contains three occurrences of the channel a, two of them in subject position for input/output operations and one binding occurrence in the new construct. We have seen that the constraint generation algorithm associates the two rightmost occurrences of a with two type variables α1 and α2 that must respectively satisfy the constraints α1 = ˆ [int]2̺1 ,1+̺2 α2 = ˆ [γ] 1+̺3 ,2̺4 (5.2) (5.3) whereas the leftmost occurrence of a has a type α which must satisfy the constraints α = ˆ α1 + α2 α = ˆ [δ]̺5 ,̺5 (5.4) (5.5) Even if none of these constraints concerns use variables directly, use variables are subject to implicit constraints that should be taken into account for finding a precise solution. To expose such implicit constraints, observe that in this first example we are in the fortunate situation where the type variables α, α1 , and α2 occur on the left-hand side of a constraint of the form β = ˆ T where T is different from a type variable. In this case we say that β is defined and we call T its definition. If we substitute each type variable in (5.4) with its definition we obtain [δ]̺5 ,̺5 = ˆ [int]2̺1 ,1+̺2 + [γ]1+̺3 ,2̺4 22 L. PADOVANI that reveals the relationships between the use variables. Knowing how type combination operates (Definition 3.5), we can derive two constraints concerning use variables ̺5 = ˆ 2̺1 + 1 + ̺3 ̺5 = ˆ 1 + ̺2 + 2̺4 for which it is easy to figure out a solution that includes the substitutions {̺1..4 7→ 0, ̺5 7→ 1} (see Example 4.4). No substitution can be more precise than this one hence such solution, which identifies a as a linear channel, is in fact optimal. Let us now consider the following variation of (5.1) a!3 | a?(x).idle where we have removed the restriction. In this case the generated constraints are the same (5.2), (5.3), and (5.4) as above, except that there is no constraint (5.5) that provides a definition for α. In a sense, α is defined because we know that it must be the combination of α1 and α2 for which we do have definitions. However, in order to come up with a general strategy for solving constraint sets, it is convenient to complete the constraint set with a defining equation for α: we know that α must be a channel type with messages of type int, because that is the shape of the definition for α1 , but we do not know precisely the overall uses of α. Therefore, we generate a new constraint defining the structure of the type α, but with fresh use variables ̺5 and ̺6 in place of the unknown uses: α= ˆ [int]̺5 ,̺6 We can now proceed as before, by substituting all type variables in (5.4) with their definition and deriving the use constraints below: ̺5 = ˆ 2̺1 + 1 + ̺3 ̺6 = ˆ 1 + ̺2 + 2̺4 Note that, unlike in (5.1), we do not know whether ̺5 and ̺6 are required to be equal or not. Here we are typing an open process which, in principle, may be composed in parallel with other uses of the same channel a. Nonetheless, we can easily find a solution analogous to the previous one but with the use assignments {̺1..5 7→ 0, ̺5,6 7→ 1}. The idea of completing constraints with missing definitions is a fundamental ingredient of our constraint solving technique. In the previous example, completion was somehow superfluous because we could have obtained a definition for α by combining the definitions of α1 and α2 , which were available. However, completion allowed us to patch the constraint set so that it could be handled as in the previous case of process (5.1). In fact, it is easy to find processes for which completion becomes essential. Consider for example new a in (a!3 | b!a) (5.6) where the bound channel a is used once for an output and then extruded through a free channel b. For this process, the reconstruction algorithm infers the type environment b : β and the constraints below: α1 β α α = ˆ = ˆ = ˆ = ˆ [int]2̺1 ,1+̺2 [α2 ]2̺3 ,1+̺4 α1 + α2 [δ]̺5 ,̺5 TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 23 where the three occurrences of a are associated from left to right with the type variables α, α1 , and α2 (Section C gives the derivation for (5.6)). Note that there is no constraint that defines α2 . In fact, there is just no constraint with α2 on the left hand side at all. The only hint that we have concerning α2 is that it must yield α when combined with α1 . Therefore, according to the definition of type combination, we can once more deduce that α2 shares the same structure as α and α1 and we can complete the set of constraints with α2 = ˆ [int]̺6 ,̺7 where ̺6 and ̺7 are fresh use variables. After performing the usual substitutions, we can finally derive the use constraints ̺5 = ˆ 2̺1 + ̺6 ̺5 = ˆ 1 + ̺2 + ̺7 for which we find a solution including the assignments {̺1..4,7 7→ 0, ̺5,6 7→ 1}. The interesting fact about this solution is the substitution ̺6 7→ 1, meaning that the constraint solver has inferred an input operation for the rightmost occurrence of a in (5.6), even though there is no explicit evidence of this operation in the process itself. The input operation is deduced “by subtraction”, seeing that a is used once in (5.6) for an output operation and knowing that a restricted (linear) channel like a must also be used for a matching input operation. Note also that this is not the only possible solution for the use constraints. If, for example, it turns out that the extruded occurrence of a is never used (or is used twice) for an input, it is possible to obtain various solutions that include the assignments {̺5,6 7→ ω}. However, the solution we have found above is the most precise according to Definition 5.1. It is not always possible to find the most precise solution. This can be seen in the following variation of (5.6) new a in (a!3 | b!a | c!a) (5.7) where a is extruded twice, on b and on c (Section C gives the derivation). Here, as in (5.6), an input use for a is deduced “by subtraction”, but there is an ambiguity as to whether such input capability is transmitted through b or through c. Hence, there exist two incomparable solutions for the constraint set generated for (5.6). The lack of an optimal solution in general (hence of a principal typing) is a consequence of the condition imposing equal uses for restricted channels (see [t-new] and [i-new]). Without this condition, it would be possible to find the most precise solution for the constraints generated by (5.7) noticing that a is never explicitly used for an input operation, and therefore its input use could be 0. We think that this approach hinders the applicability of the reconstruction algorithm in practice, where separate compilation and type reconstruction of large programs are real concerns. We will elaborate more on this in Example 7.3. For the time being, let us analyze one last example showing a feature that we do not handle in our type system, namely polymorphism. The process a?(x).b!x (5.8) models a forwarder that receives a message x from a and sends it on b. For this process the constraint generation algorithm yields the environment a : α, b : β and the constraints α = ˆ [γ]1+̺1 ,2̺2 β = ˆ [γ]2̺3 ,1+̺4 (Section C gives the complete derivation). In particular, there is no constraint concerning the type variable γ and for good reasons: since the message x is only passed around in (5.8) 24 L. PADOVANI [c-axiom] C ∪ {ϕ} ϕ [c-refl] [c-symm] [c-trans] T ∈ expr(C) C T R̂ S C C C S R̂ T T R̂ T C [c-coh 1] [c-coh 2] [c-cong 1] C C C C T= ˆ S T∼ ˆ S T= ˆ S1 + S2 C T∼ ˆ Si i ∈ {1, 2} [c-cong 2] C T R̂ T′ T′ R̂ S T R̂ S [T]U1 ,U2 ∼ ˆ [S]V1 ,V2 C T= ˆ S [c-cong 3] T1 ⊙ T2 R̂ S1 ⊙ S2 C C Ti R̂ Si C i ∈ {1, 2} T1 ⊙ T2 = ˆ S1 ⊙ S2 + S3 ⊙ S4 C Ti = ˆ Si + Si+2 i ∈ {1, 2} [c-subst] C T1 = ˆ T2 + T3 C Ti = ˆ Si C S1 = ˆ S2 + S3 [c-use 1] C (1≤i≤3) [c-use 2] [T]U1 ,U2 = ˆ [S]V1 ,V2 i ∈ {1, 2} C Ui = ˆ Vi C [T]U1 ,U2 = ˆ [S1 ]V1 ,V2 + [S2 ]V3 ,V4 i ∈ {1, 2} C Ui = ˆ Vi + Vi+2 Table 8: Constraint deduction system. but never actually used, the channels a and b should be considered polymorphic. Note that in this case we know nothing about the structure of γ hence completion of the constraint set is not applicable. In this work we do not deal with polymorphism and will refrain from solving sets of constraints where there is no (structural) information for unconstrained type variables. Just observe that handling polymorphism is not simply a matter of allowing (universally quantified) type variables in types. For example, a type variable involved in a constraint α = ˆ α+ α does not have any structural information and therefore is polymorphic, but can only be instantiated with unlimited types. The implementation has a defaulting mechanism that forces unconstrained type variables to a base type. We now formalize the ideas presented so far into an algorithm, for which we have already identified the key phases: the ability to recognize types that “share the same structure”, which we call structurally coherent (Definition 3.6); the completion of a set of constraints with “missing definitions” so that each type variable has a proper definition; the derivation and solution of use constraints. Let us proceed in order. 5.2. Verification. In Section 5.1 we have seen that some constraints can be derived from the ones produced during the constraint generation phase (Section 4). We now define a deduction system that, starting from a given constraint set C, computes all the “derivable facts” about the types in expr(C). Such deduction system is presented as a set of inference rules in Table 8, where R ranges over the symbols = and ∼. Each rule derives a judgment of the form C ϕ meaning that the constraint ϕ is derivable from those in C (Proposition 5.2 below formalizes this property). Rule [c-axiom] simply takes each constraint in C as an axiom. Rules [c-refl], [c-symm], and [c-trans] state the obvious reflexivity, symmetry, and TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 25 transitivity of = and ∼. Rules [c-coh 1] and [c-coh 2] deduce coherence relations: equality implies coherence, for = ⊆ ∼, and each component of a combination is coherent to the combination itself (and therefore, by transitivity, to the other component). Rules [c-cong 1] through [c-cong 3] state congruence properties of = and ∼ which follow directly from Definition 3.5: when two channel types are coherent, their message types must be equal; corresponding components of R-related composite types are R-related. Rule [c-subst] allows the substitution of equal types in combinations. Finally, [c-use 1] and [c-use 2] allow us to deduce use constraints of the form U = ˆ V involving use variables. Both rules are self-explanatory and follow directly from Definition 3.5. We state two important properties of this deduction system: Proposition 5.2. Let C ϕ. The following properties hold: (1) C and C ∪ {ϕ} are equivalent (see Definition 4.1). (2) expr(C) = expr(C ∪ {ϕ}). Proof. A simple induction on the derivation of C ϕ. The first property confirms that all the derivable relations are already encoded in the original constraint set, in a possibly implicit form. The deduction system makes them explicit. The second property assures us that no new type expressions are introduced by the deduction system. Since the inference rules in Section 4 always generate finite constraint sets, this implies that the set of all derivable constraints is also finite and can be computed in finite time. This is important because the presence or absence of particular constraints determines the (un)satisfiability of a constraint set: Proposition 5.3. If C T ∼ ˆ S where T and S are proper type expressions with different topmost constructors, then C has no solution. Proof. Suppose by contradiction that σ is a solution for C. By Proposition 5.2(1) we have that σ is also a solution for C ∪ {T ∼ ˆ S}. This is absurd, for if T and S have different topmost constructors, then so do σT and σS, hence σT 6∼ σS. The converse of Proposition 5.3 is not true in general. For example, the constraint set ˆ [int]1,0 } has no solution because of the implicit constraints on corresponding {[int]0,1 = uses and yet it satisfies the premises of Proposition 5.3. However, when C is a constraint set generated by the inference rules in Section 4, the converse of Proposition 5.3 holds. This means that we can use structural coherence as a necessary and sufficient condition for establishing the satisfiability of constraint sets generated by the reconstruction algorithm. Before proving this fact we introduce some useful notation. For R ∈ {=, ∼} let def RC = {(T, S) | C T R̂ S} and observe that RC is an equivalence relation on expr(C) by construction, because of the rules [c-refl], [c-symm], and [c-trans]. Therefore, it partitions the type expressions in C into R-equivalence classes. Now, we need some way to choose, from each R-equivalence class, one representative element of the class. To this aim, we fix a total order ⊑ between type expressions such that T ⊑ α for every proper T and every α and we define:1 1In a Haskell or OCaml implementation such total order could be, for instance, the one automatically defined for the algebraic data type that represents type expressions and where the value constructor representing type variables is the last one in the data type definition. 26 L. PADOVANI Definition 5.4 (canonical representative). Let crepR (C, T) be the ⊑-least type expression S such that T RC S. We say that crepR (C, T) is the canonical representative of T with respect to the relation RC . Note that, depending on ⊑, we may have different definitions of crepR (C, T). The exact choice of the canonical representative does not affect the ability of the algorithm to compute a solution for a constraint set (Theorem 5.12) although – in principle – it may affect the precision of the solution (Example 5.11). Note also that, because of the assumption we have made on the total order ⊑, crepR (C, T) is proper whenever T is proper or when T is some type variable α such that there is a “definition” for α in C. In fact, it is now time to define precisely the notion of defined and undefined type variables: Definition 5.5 (defined and undefined type variables). Let def def R (C) = {α ∈ expr(C) | crepR (C, α) is proper} def undef R (C) = {α ∈ expr(C) \ def R (C)} We say that α is R-defined or R-undefined in C according to α ∈ def R (C) or α ∈ undef R (C). We can now prove that the coherence check is also a sufficient condition for satisfiability. Proposition 5.6. Let P ◮ ∆; C. If C T ∼ ˆ S where T and S are proper type expressions implies that T and S have the same topmost constructor, then C has a solution. Proof. We only sketch the proof, since we will prove a more general result later on (see def Theorem 5.12). Consider the use substitution σuse = {̺ 7→ ω | ̺ ∈ expr(C)} mapping all use variables in C to ω, let Σ be the system of equations {αi = Ti | 1 ≤ i ≤ n} defined by def Σ = {α = σuse crep∼ (C, α) | α ∈ def ∼ (C)} ∪ {α = int | α ∈ undef ∼ (C)} and observe that every Ti is a proper type expression. From Theorem 3.2 we know that Σ has a unique solution σtype = {αi 7→ ti | 1 ≤ i ≤ n} such that ti = σtype Ti for every 1 ≤ i ≤ n. It only remains to show that σuse ∪ σtype is a solution for C. This follows from two facts: (1) from the hypothesis P ◮ ∆; C we know that all channel types in C have one use variable in each of their use slots, hence the substitution σuse forces all uses to ω; (2) from the hypothesis and the rules [c-cong *] we know that all proper type expressions in the same (∼)-equivalence class have the same topmost constructor. In Proposition 5.6, for finding a substitution for all the type variables in C, we default each type variable in undef ∼ (C) to int. This substitution is necessary in order to satisfy the constraints un(α), namely those of the form α = ˆ α + α, when α ∈ undef ∼ (C). These α’s are the “polymorphic type variables” that we have already discussed earlier. Since we leave polymorphism for future work, in the rest of this section we make the assumption that undef ∼ (C) = ∅, namely that all type variables are (∼)-defined. Example 5.7. Below is a summary of the constraint set C generated in Example 4.5: α= ˆ α1 + α2 δ= ˆ int α1 = ˆ β1 × β2 α2 = ˆ γ1 × γ2 β1 = ˆ [δ]1+̺1 ,2̺2 β2 = ˆ β2 + β2 γ1 = ˆ γ1 + γ1 γ2 = ˆ [int]2̺3 ,1+̺4 TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 27 Note that {α, β2 , γ1 } ⊆ def ∼ (C) \ def = (C). In particular, they all have a proper canonical representative, which we may assume to be the following ones: crep∼ (C, α) = crep∼ (C, α1 ) = crep∼ (C, α2 ) crep∼ (C, β1 ) = crep∼ (C, γ1 ) crep∼ (C, β2 ) = crep∼ (C, γ2 ) crep∼ (C, δ) = = = = β1 × β2 [δ]1+̺1 ,2̺2 [int]2̺3 ,1+̺4 int It is immediate to verify that the condition of Proposition 5.6 holds, hence we conclude that C is satisfiable. Indeed, a solution for C is {α, α1,2 7→ [int]ω,ω × [int]ω,ω , β1,2 , γ1,2 7→ [int]ω,ω , δ 7→ int, ̺1..4 7→ ω} even though we will find a more precise solution in Example 5.13.  5.3. Constraint set completion. If the satisfiability of the constraint set is established (Proposition 5.6), the subsequent step is its completion in such a way that every type variable α has a definition in the form of a constraint α = ˆ T where T is proper. Recall that this step is instrumental for discovering all the (implicit) use constraints. In Example 5.7 we have seen that some type variables may be (∼)-defined but (=)undefined. The ∼ relation provides information about the structure of the type that should be assigned to the type variable, but says nothing about the uses in them. Hence, the main task of completion is the creation of fresh use variables for those channel types of which only the structure is known. In the process, fresh type variables need to be created as well, and we should make sure that all such type variables are (=)-defined to guarantee that completion eventually terminates. We will be able to do this, possibly at the cost of some precision of the resulting solution. We begin the formalization of completion by introducing an injective function t that, given a pair of type variables α and β, creates a new type variable t(α, β). We assume that t(α, β) is different from any type variable generated by the algorithm in Section 4 so that the type variables obtained through t are effectively fresh. Then we define an instantiation function instance that, given a type variable α and a type expression T, produces a new type expression that is structurally coherent to T, but where all use expressions and type variables have been respectively replaced by fresh use and type variables. The first argument α of instance records the fact that such instantiation is necessary for completing α. Formally:  t(α, β) if T = β    int if T = int def (5.9) instance(α, T) = ̺ ,̺ 1 2  [S] if T = [S]U,V , ̺i fresh    instance(α, T1 ) ⊙ instance(α, T2 ) if T = T1 ⊙ T2 All the equations but the first one are easily explained: the instance of int cannot be anything but int itself; the instance of a channel type [S]U,V is the type expression [S]̺1 ,̺2 where we generate two fresh use variables corresponding to U and V; the instance of a composite type T ⊙ S is the composition of the instances of T and S. For example, we have instance(α, β × [[int]U1 ,U2 ]V1 ,V2 ) = t(α, β) × [[int]U1 ,U2 ]̺1 ,̺2 28 L. PADOVANI where ̺1 and ̺2 are fresh. Note that, while instantiating a channel type [S]U,V , there is no need to instantiate S because [t]κ1 ,κ2 ∼ [s]κ3 ,κ4 implies t = s so S is exactly the message type we must use in the instance of [S]U,V . Concerning the first equation in (5.9), in principle we want instance(α, β) to be the same as instance(α, crep∼ (C, β)), but doing so directly would lead to an ill-founded definition for instance, since nothing prevents β from occurring in crep∼ (C, β) (types can be infinite). We therefore instantiate β to a new type variable t(α, β) which will in turn be defined by a new constraint t(α, β) = ˆ instance(α, crep∼ (C, β)). There are a couple of subtleties concerning the definition of instance. The first one is that, strictly speaking, instance is a relation rather than a function because the fresh use variables in (5.9) are not uniquely determined. In practice, instance can be turned into a proper function by devising a deterministic mechanism that picks fresh use variables in a way similar to the t function that we have defined above. The formal details are tedious but well understood, so we consider the definition of instance above satisfactory as is. The second subtlety is way more serious and has to do with the instantiation of type variables (first equation in (5.9)) which hides a potential approximation due to this completion phase. To illustrate the issue, suppose that α∼ ˆ [int]U,V × α (5.10) is the only constraint concerning α in some constraint set C so that we need to provide a (=)-definition for α. According to (5.9) we have instance(α, [int]U,V × α) = [int]̺1 ,̺2 × t(α, α) so by adding the constraints α= ˆ t(α, α) and t(α, α) = ˆ [int]̺1 ,̺2 × t(α, α) (5.11) we complete the definition for α. There is a fundamental difference between the constraint (5.10) and those in (5.11) in that the former admits far more solutions than those admitted by (5.11). For example, the constraint (5.10) can be satisfied by a solution that contains the assignment α 7→ t where t = [int]1,0 × [int]0,1 × t, but the constraint (5.11) cannot. The problem of a constraint like (5.10) is that, when we only have structural information about a type variable, we have no clue about the uses in its definition, if they follow a pattern, and what the pattern is. In principle, in order to account for all the possibilities, we should generate fresh use variables in place of any use slot in the possibly infinite type. In practice, however, we want completion to eventually terminate, and the definition of instance given by (5.9) is one easy way to ensure this: what we are saying there is that each type variable β that contributes to the definition of a (=)-undefined type variable α is instantiated only once. This trivially guarantees completion termination, for there is only a finite number of distinct variables to be instantiated. The price we pay with this definition of instance is a potential loss of precision in the solution of use constraints. We say “potential” because we have been unable to identify a concrete example that exhibits such loss of precision. Part of the difficulty of this exercise is due to the fact that the effects of the approximation on the solution of use constraints may depend on the particular choice of canonical representatives, which is an implementation detail of the constraint solver (see Definition 5.4). In part, the effects of the approximation are limited to peculiar situations: (1) There is only a fraction of constraint sets where the same type variable occurring in several different positions must be instantiated, namely those having as solution types TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 29 with infinite branches containing only finitely many channel type constructors. The constraint (5.10) is one such example. In all the other cases, the given definition of instance does not involve any approximation. (2) A significant fraction of the type variables for which only structural information is known are those generated by the rules [i-fst], [i-snd], and [i-weak]. These type variables stand for unlimited types, namely for types whose uses are either 0 or ω. In fact, in most cases all the uses in these unlimited types are 0. Therefore, the fact that only a handful of fresh use variables is created, instead of infinitely many, does not cause any approximation at all, since the use variables in these type expressions would all be instantiated to 0 anyway. We define the completion of a constraint set C as the least superset of C where all the (=)-undefined type variables in C have been properly instantiated: Definition 5.8 (completion). The completion of C, written C, is the least set such that: (1) C ⊆ C; (2) α ∈ undef = (C) implies α = ˆ t(α, α) ∈ C; (3) t(α, β) ∈ expr(C) implies t(α, β) = ˆ instance(α, crep∼ (C, β)) ∈ C. The completion C of a finite constraint set C can always be computed in finite time as the number of necessary instantiations is bound by the square of the cardinality of undef = (C). Because of the approximation of instances for undefined variables, C and C are not equivalent in general (see Example 5.11 below). However, the introduction of instances does not affect the satisfiability of the set of constraints. Proposition 5.9. The following properties hold: (1) If C is satisfiable, then C is satisfiable. (2) If σ is a solution for C, then σ is also a solution for C. Proof. Each (∼)-equivalence class in C contains exactly one (∼)-equivalence class in C, for each new type expression that has been introduced in C is structurally coherent to an existing type expression in C. Then item (1) is a consequence of Proposition 5.6, while item (2) follows from the fact that C ⊆ C. Example 5.10. Considering the constraint set C in Example 5.7, we have three type variables requiring instantiation, namely α, β2 , and γ1 . According to Definition 5.8, and using the same canonical representatives mentioned in Example 5.7, we augment the constraint set with the constraints α= ˆ t(α, α) β2 = ˆ t(β2 , β2 ) γ1 = ˆ t(γ1 , γ1 ) t(α, α) t(α, β1 ) t(α, β2 ) t(β2 , β2 ) t(γ1 , γ1 ) = ˆ = ˆ = ˆ = ˆ = ˆ instance(α, β1 × β2 ) instance(α, [δ]1+̺1 ,2̺2 ) instance(α, [int]2̺3 ,1+̺4 ) instance(β2 , [int]2̺3 ,1+̺4 ) instance(γ1 , [δ]1+̺1 ,2̺2 ) = = = = = t(α, β1 ) × t(α, β2 ) [δ]̺5 ,̺6 [int]̺7 ,̺8 [int]̺9 ,̺10 [δ]̺11 ,̺12 where the ̺i with i ≥ 5 are all fresh. Observe that the canonical (∼)-representative of β2 is instantiated twice, once for defining α and once for defining β2 itself. We will see in Example 5.13 that this double instanti ation is key for inferring that snd(x) in Example 4.5 is used linearly. 30 L. PADOVANI Example 5.11. In this example we show the potential effects of instantiation on the ability of the type reconstruction algorithm to identify linear channels. To this aim, consider the following constraint set ∼ ˆ = ˆ = ˆ = ˆ α β γ α [int]U,V × α [int]0,1+̺1 × [int]0,2̺2 × β [int]0,0 × [int]0,0 × γ β+γ where, to limit the number of use variables without defeating the purpose of the example, we write the constant use 0 in a few use slots. Observe that this constraint set admits the solution {α 7→ t, β 7→ t, γ 7→ s, ̺1,2 7→ 0} where t and s are the types that satisfy the equalities t = [int]0,1 × [int]0,0 × t and s = [int]0,0 × s. Yet, if we instantiate α following the procedure outlined above we obtain the constraints α= ˆ t(α, α) and t(α, α) = ˆ [int]̺3 ,̺4 × t(α, α) and now the two constraints below follow by the congruence rule [c-cong *]: [int]̺3 ,̺4 = ˆ [int]0,1+̺1 + [int]0,0 [int]̺3 ,̺4 = ˆ [int]0,2̺2 + [int]0,0 This implies that the use variable ̺4 must simultaneously satisfy the constraints ̺4 = ˆ 1 + ̺1 and ̺4 = ˆ 2̺2 which is only possible if we assign ̺1 and ̺2 to a use other than 0 and ̺4 to ω. In other words, after completion the only feasible solutions for the constraint set above have the form {α 7→ t′ , β 7→ t′ , γ 7→ s, ̺1,2 7→ κ, ̺3 7→ 0, ̺4 7→ ω} for 1 ≤ κ where t′ = [int]0,ω × t′ , which are less precise than the one that we could figure out before the instantiation: t denotes an infinite tuple of channels in which those in odd-indexed positions are used for performing exactly one output operation; t′ denotes an infinite tuple of channels, each being used for an unspecified number of output operations.  5.4. Solution synthesis. In this phase, substitutions are found for all the use and type variables that occur in a (completed) constraint set. We have already seen that it is always possible to consider a trivial use substitution that assigns each use variable to ω. In this phase, however, we have all the information for finding a use substitution that, albeit not necessarily optimal because of the approximation during the completion phase, is minimal according to the ≤ precision order on uses of Definition 5.1. The first step for computing a use substitution is to collect the whole set of constraints concerning use expressions. This is done by repeatedly applying the rules [c-use 1] and [c-use 2] shown in Table 8. Note that the set of derivable use constraints is finite and can be computed in finite time because C is finite. Also, we are sure to derive all possible use constraints if we apply these two rules to a completed constraint set. Once use constraints have been determined, any particular substitution for use variables can be found by means of an exhaustive search over all the possible substitutions: the number of such substitutions is finite because the number of use variables is finite and so is the domain {0, 1, ω} on which they range. Clearly this brute force approach is not practical in general and in Section 6 we will discuss two techniques that reduce the search space TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 31 for use substitutions. The main result of this section is independent of the particular use substitution σuse that has been identified. Theorem 5.12 (correctness of the constraint solving algorithm). Let P ◮ ∆; C. If (1) C T∼ ˆ S where T and S are proper type expressions implies that T and S have the same topmost constructor, and (2) σuse is a solution of the use constraints of C, and def (3) σtype is the solution of the system Σ = {α = σuse crep= (C, α) | α ∈ expr(C)}, then σuse ∪ σtype is a solution for C. def Proof. Let σ = σuse ∪ σtype . We have to prove the implications of Definition 4.1 for C. We focus on constraints of the form T = ˆ S1 + S2 , the other constraints being simpler and/or handled in a similar way. def ˆ S1 + S2 }. It is enough to show that R satisfies Let R = {((σS1 , σS2 ), σT) | C T = the conditions of Definition 3.5, since type combination is the largest relation that satisfies those same conditions. Suppose ((s1 , s2 ), t) ∈ R. Then there exist T, S1 , and S2 such that C T= ˆ S1 +S2 and t = σT and si = σSi for i = 1, 2. Without loss of generality, we may also assume that T, S1 , and S2 are proper type expressions. Indeed, suppose that this is not the ˆ S1 + S2 case and, for instance, T = α. Then, from [c-subst] we have that C crep= (C, α) = and, since σ is a solution of Σ, we know that σ(α) = σcrep= (C, α). Therefore, the same pair ((s1 , s2 ), t) ∈ R can also be obtained from the triple (crep= (C, α), S1 , S2 ) whose first component is proper. The same argument applies for S1 and S2 . Now we reason by cases on the structure of T, S1 , and S2 , knowing that all these type expressions have the same topmost constructor from hypothesis (1) and [c-coh 2]: • If T = S1 = S2 = int, then condition (1) of Definition 3.5 is satisfied. • If T = [T′ ]U1 ,U2 and Si = [S′i ]V2i−1 ,V2i for i = 1, 2, then from [c-coh 2] and [c-cong 1] we deduce C T′ = ˆ S′i and from [c-use 2] we deduce C Ui = ˆ Vi + Vi+2 for i = 1, 2. Since σ is a solution for the equality constraints in C, we deduce σT′ = σS1 = σS2 . Since σ is a solution for the use constraints in C, we conclude σUi = σVi + σVi+2 for i = 1, 2. Hence, condition (2) of Definition 3.5 is satisfied. • If T = T1 ⊙ T2 and Si = Si1 ⊙ Si2 , then from [c-cong 3] we deduce C Ti = ˆ Si1 + Si2 for i = 1, 2. We conclude ((σSi1 , σSi2 ), σTi ) ∈ R by definition of R, hence condition (3) of Definition 3.5 is satisfied. Note that the statement of Theorem 5.12 embeds the constraint solving algorithm, which includes a verification phase (item (1)), a constraint completion phase along with an (unspecified, but effective) computation of a solution for the use constraints (item (2)), and the computation of a solution for the original constraint set in the form of a finite system of equations (item (3)). The conclusion of the theorem states that the algorithm is correct. Example 5.13. There are three combination constraints in the set C obtained in Example 5.10, namely α = ˆ α1 + α2 , β2 = ˆ β2 + β2 , and γ1 = ˆ γ1 + γ1 . By performing suitable substitutions with [c-subst] we obtain [c-axiom] C α= ˆ α1 + α2 ================================= [c-subst] (multiple applications) C t(α, β1 ) × t(α, β2 ) = ˆ β1 × β2 + γ1 × γ2 C t(α, βi ) = ˆ βi + γi [c-cong 3] 32 L. PADOVANI from which we can further derive .. . C t(α, β1 ) = ˆ β1 + γ1 ======= ========1+̺ ==== =========== [c-subst] (multiple applications) ̺5 ,̺6 C [δ] = ˆ [δ] 1 ,2̺2 + [δ]̺11 ,̺12 as well as .. . C t(α, β2 ) = ˆ β2 + γ2 ======= ========̺== ============ [c-subst] (multiple applications) ̺7 ,̺8 C [δ] = ˆ [δ] 9 ,̺10 + [δ]2̺3 ,1+̺4 Analogous derivations can be found starting from β2 = ˆ β2 + β2 and γ1 = ˆ γ1 + γ1 . At this point, using [c-use 2], we derive the following set of use constraints: ̺5 ̺6 ̺7 ̺8 = ˆ = ˆ = ˆ = ˆ 1 + ̺1 + ̺11 2̺2 + ̺12 ̺9 + 2̺3 ̺10 + 1 + ̺4 ̺11 ̺12 ̺9 ̺10 = ˆ = ˆ = ˆ = ˆ 2̺11 2̺12 2̺9 2̺10 for which we find the most precise solution {̺1..4,6,7,9..12 7→ 0, ̺5,8 7→ 1}. From this set of use constraints we can also appreciate the increased accuracy deriving from distinguishing the instance t(α, β2 ) of the type variable β2 used for defining α from the instance t(β2 , β2 ) of the same type variable β2 for defining β2 itself. Had we chosen to generate a unique instance of β2 , which is equivalent to saying that ̺8 and ̺10 are the same use variable, we would be required to satisfy the use constraint ̺10 + 1 + ̺4 = ˆ 2̺10 which is only possible if we take ̺8 = ̺10 = ω. But this assignment fails to recognize that snd(x) is used linearly in the process of Example 4.5.  6. Implementation In this section we cover a few practical aspects concerning the implementation of the type reconstruction algorithm. 6.1. Derived constraints. The verification phase of the solver algorithm requires finding all the constraints of the form T ∼ ˆ S that are derivable from a given constraint set C. Doing so allows the algorithm to determine whether C is satisfiable or not (Proposition 5.6). In principle, then, one should compute the whole set of constraints derivable from C. The particular nature of the ∼ relation enables a more efficient way of handling this phase. The key observation is that there is no need to ever perform substitutions (with the rule [c-subst]) in order to find all the ∼ ˆ constraints. This is because [c-coh 2] allows one to relate the type expressions in a combination, since they must all be structurally coherent and ∼ is insensitive to the actual content of the use slots in channel types. This means that all ∼ ˆ constraints can be computed efficiently using conventional unification techniques (ignoring the content of use slots). In fact, the implementation uses unification also for the constraints of the form T = ˆ S. Once all the = ˆ constraints have been found and the TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 33 constraint set has been completed, substitutions in constraints expressing combinations can be performed efficiently by mapping each type variable to its canonical representative. 6.2. Use constraints resolution. In Section 5 we have refrained from providing any detail about how use constraints are solved and argued that a particular use substitution can always be found given that both the set of constraints and the domain of use variables are finite. While this argument suffices for establishing the decidability of this crucial phase of the reconstruction algorithm, a naı̈ve solver based on an exhaustive search of all the use substitutions would be unusable, since the number of use variables is typically large, even in small processes. Incidentally, note that completion contributes significantly to this number, since it generates fresh use variables for all the instantiated channel types. There are two simple yet effective strategies that can be used for speeding up the search of a particular use substitution (both have been implemented in the prototype). The first strategy is based on the observation that, although the set of use variables can be large, it can often be partitioned into many independent subsets. Finding partitions is easy: two variables ̺1 and ̺2 are related in C if C U = ˆ V and ̺1 , ̺2 occur in U = ˆ V (regardless of where ̺1 and ̺2 occur exactly). The dependencies between variables induce a partitioning of the use constraints such that the use variables occurring in the constraints of a partition are all related among them, and are not related with any other use variable occurring in a use constraint outside the partition. Once the partitioning of use constraints has been determined, each partition can be solved independently of the others. The second strategy is based on the observation that many use constraints have the form ̺= ˆ U where ̺ does not occur in U. In this case, the value of ̺ is in fact determined by U. So, U can be substituted in place of all the occurrences of ̺ in a given set of use constraints and, once a substitution is found for the use variables in the set of use constraints with the substitution, the substitution for ̺ can be determined by simply evaluating U under such substitution. 6.3. Pair splitting versus pair projection. It is usually the case that linearly typed languages provide a dedicated construct for splitting pairs (a notable exception is [14]). The language introduced in [23, Chapter 1], for example, has an expression form split e as x,y in f that evaluates e to a pair, binds the first and second component of the pair respectively to the variables x and y, and then evaluates f. At the same time, no pair projection primitives are usually provided. This is because in most linear type systems linear values “contaminate” with linearity the composite data structures in which they occur: for example, a pair containing linear values is itself a linear value and can only be used once, whereas for extracting both components of a pair using the projections one would have to project the pair twice, once using fst and one more time using snd. For this reason, the split construct becomes the only way to use linear pairs without violating linearity, as it grants access to both components of a pair but accessing the pair only once. The process language we used in an early version of this article [20] provided a split construct for splitting pairs and did not have the projections fst and snd. In fact, the ability to use fst and snd without violating linearity constraints in our type system was pointed out by a reviewer of [20] and in this article we have decided to promote projections as the sole mechanism for accessing pair components. Notwithstanding this, there is a 34 L. PADOVANI practical point in favor of split when considering an actual implementation of the type system. Indeed, the pair projection rules [i-fst] and [i-snd] are among the few that generate constraints of the form un(α) for some type variable α. In the case of [i-fst] and [i-snd], the unlimited type variable stands for the component of the pair that is discarded by the projection. For instance, we can derive x : β1 ◮ x : β1 ; ∅ x : β2 ◮ x : β2 ; ∅ fst(x) : α1 ◮ x : β1 ; {β1 = ˆ α1 × γ1 , un(γ1 )} snd(x) : α2 ◮ x : β2 ; {β2 = ˆ γ2 × α2 , un(γ2 )} (fst(x),snd(x)) : α1 × α2 ◮ x : α; {α = ˆ β1 + β2 , β1 = ˆ α1 × γ1 , β2 = ˆ γ2 × α2 , un(γ1 ), un(γ2 )} and we observe that γ1 and γ2 are examples of those type variables for which only structural information is known, but no definition is present in the constraint set. Compare this with a hypothetical derivation concerning a splitting construct (for expressions) x1 : α1 ◮ x1 : α1 ; ∅ x : α ◮ x : α; ∅ x2 : α2 ◮ x2 : α2 ; ∅ (x1 ,x2 ) : α1 × α2 ◮ x1 : α1 , x2 : α2 ; ∅ split x as x1 ,x2 in (x1 ,x2 ) : α1 × α2 ◮ x : α; {α = ˆ α1 × α2 } producing a much smaller constraint set which, in addition, is free from un(·) constraints and includes a definition for α. The constraint set obtained from the second derivation is somewhat easier to solve, if only because it requires no completion, meaning fewer use variables to generate and fewer chances of stumbling on the approximated solution of use constraints (Example 5.11). Incidentally we observe, somehow surprisingly, that the two constraint sets are not exactly equivalent. In particular, the constraint set obtained from the first derivation admits a solution containing the substitutions {α 7→ [int]ω,0 × [int]0,ω , α1 7→ [int]1,0 , α2 7→ [int]0,1 } whereas in the second derivation, if we fix α as in the substitution above, we can only have {α 7→ [int]ω,0 × [int]0,ω , α1 7→ [int]ω,0 , α2 7→ [int]0,ω } meaning that, using projections, it is possible to extract from a pair only the needed capabilities, provided that what remains unused has an unlimited type. On the contrary, split always extracts the full set of capabilities from each component of the pair. In conclusion, in spite of the features of the type system we argue that it is a good idea to provide both pair projections and pair splitting, and that pair splitting should be preferred whenever convenient to use. 7. Examples In this section we discuss three more elaborate examples that highlight the features of our type reconstruction algorithm. For better clarity, in these examples we extend the language with triples, boolean values, conditional branching, arithmetic and relational operators, OCaml-like polymorphic variants [16, Chapter 4], and a more general form of pattern matching. All these extensions can be easily accommodated or encoded in the language presented in Section 2 and are supported by the prototype implementation of the reconstruction algorithm. TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 35 • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • • Figure 1: Regions of a complete binary tree used by take. Example 7.1. The purpose of this example is to show the reconstruction algorithm at work on a fairly complex traversal of a binary tree. The traversal is realized by the two processes take and skip below *take?(x).case x of Leaf ⇒ idle Node(c,y,z) ⇒ c!3 | take!y | skip!z | *skip?(x).case x of Leaf ⇒ idle Node(_,y,z) ⇒ skip!y | take!z where, as customary, we identify the name of a process with the replicated channel on which the process waits for invocations. Both take and skip receive as argument a binary tree x and analyze its structure by means of pattern matching. If the tree is empty, no further operation is performed. When take receives a non-empty tree, it uses the channel c found at the root of the tree, it recursively visits the left branch y and passes the right branch z to skip. The process skip does not use the channel found at the root of the tree, but visits the left branch recursively and passes the right branch to take. The types inferred for take and skip are take : [t]ω,ω and skip : [s]ω,ω where t and s are the types that satisfy the equalities t = Leaf ⊕ Node([int]0,1 × t × s) s = Leaf ⊕ Node([int]0,0 × s × t) In words, take uses every channel that is found after an even number of right traversals, whereas skip uses every channel that is found after an odd number of right traversals. Figure 1 depicts the regions of a (complete) binary tree of depth 4 that are used by take, while the unmarked regions are those used by skip. Overall, the invocation take!tree | skip!tree allows the reconstruction algorithm to infer that all the channels in tree are used, namely  that tree has type ttree = Leaf ⊕ Node([int]0,1 × ttree × ttree ) = t + s. Example 7.2. In this example we show how our type reconstruction algorithm can be used for inferring session types. Some familiarity with the related literature and particularly with [3, 2] is assumed. Session types [6, 7, 5] are protocol specifications describing the sequence of input/output operations that are meant to be performed on a (private) communication channel. In most presentations, session types T , . . . include constructs like 36 L. PADOVANI ?t.T (input a message of type t, then use the channel according to T ) or !t.T (output a message of type t, then use the channel according to T ) and possibly others for describing terminated protocols and protocols with branching structure. By considering also recursive session types (as done, e.g., in [5]), or by taking the regular trees over such constructors (as we have done for our type language in this paper), it is possible to describe potentially infinite protocols. For instance, the infinite regular tree T satisfying the equality T = !int.?bool.T describes the protocol followed by a process that alternates outputs of integers and inputs of booleans on a session channel, whereas the infinite regular tree S satisfying the equality S = ?int.!bool.S describes the protocol followed by a process that alternates inputs of integers and outputs of booleans. According to the conventional terminology, T and S above are dual of each other: each action described in T (like the output of a message of type int) is matched by a corresponding co-action in S (like the input of a message of type int). This implies that two processes that respectively follow the protocols T and S when using the same session channel can interact without errors: when one process sends a message of type t on the channel, the other process is ready to receive a message of the same type from the channel. Two such processes are those yielded by the outputs foo!c and bar!c below: *foo?(x).x!random.x?(_).foo!x | *bar?(y).y?(n).y!(n mod 2).bar!y | new c in (foo!c | bar!c) It is easy to trace a correspondence of the actions described by T with the operations performed on x, and of the actions described by S with the operations performed on y. Given that x and y are instantiated with the same channel c, and given the duality that relates T and S, this process exhibits no communication errors even if the same channel c is exchanging messages of different types (int or bool). For this reason, c is not a linear channel and the above process is ill typed according to our typing discipline. However, as discussed in [13, 3, 2], binary sessions and binary session types can be encoded in the linear π-calculus using a continuation passing style. The key idea of the encoding is that each communication in a session is performed on a distinct linear channel, and the exchanged message carries, along with the actual payload, a continuation channel on which the rest of the conversation takes place. According to this intuition, the process above is encoded in the linear π-calculus as the term: *foo?(x).new a in (x!(random,a) | a?(_,x′ ).foo!x′ ) | *bar?(y).y?(n,y ′ ).new b in (y ′ !(n mod 2,b) | bar!b) | new c in (foo!c | bar!c) where a, b, and c are all linear channels (with possibly different types) used for exactly one communication. The encoding of processes using (binary) sessions into the linear π-calculus induces a corresponding encoding of session types into linear channel types. In particular, input and output session types are encoded according to the laws J?t.T K = [t × JT K]1,0 J!t.T K = [t × JT K]0,1 (7.1) TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 37 where we use T to denote the dual protocol of T . Such encoding is nothing but the coinductive extension of the one described in [3] to infinite protocols. Note that in J!t.T K, the type of the continuation channel is the encoding of the dual of T . This is because the transmitted continuation will be used by the receiver process in a complementary fashion with respect to T , which instead describes the continuation of the protocol from the viewpoint of the sender. As an example, the protocols T and S above can be respectively encoded as the types t and s that satisfy the equalities t = [int × [bool × s]0,1 ]0,1 s = [int × [bool × s]0,1 ]1,0 It turns out that these are the types that our type reconstruction algorithm associates with x and y. This is not a coincidence, for essentially three reasons: (1) the encoding of a well-typed process making use of binary sessions is always a well-typed process in the linear π-calculus [3, 2], (2) our type reconstruction algorithm is complete (Theorem 4.3), and (3) it can identify a channel as linear when it is used for one communication only (Section 5.4). The upshot is that, once the types t and s have been reconstructed, the protocols T and S can be obtained by a straightforward procedure that “decodes” t and s using the inverse of the transformation sketched by the equations (7.1). There is a technical advantage of such rather indirect way of performing session type reconstruction. Duality accounts for a good share of the complexity of algorithms that reconstruct session types directly [17]. However, as the authors of [3] point out, the notion of duality that relates T and S – and that globally affects their structure – boils down to a local swapping of uses in the topmost channel types in t and s. This is a general property of the encoding that has important practical implications: the hard work is carried over during type reconstruction for the linear π-calculus, where there is no duality to worry about; once such phase is completed, session types can be obtained from linear channel types with little effort. We have equipped the prototype implementation of the type reconstruction algorithm with a flag that decodes linear channel types into session types (the decoding procedure accounts for a handful of lines of code). In this way, the tool can be used for inferring the communication protocol of processes encoded in the linear π-calculus. Since the type reconstruction algorithm supports infinite and disjoint sum types, both infinite protocols and protocols with branches can be inferred. Examples of such processes, like for instance the server for mathematical operations described in [5], are illustrated on the home page of the tool and in its source archive.  Example 7.3. In this example we motivate the requirement expressed in the rules [t-new] and [i-new] imposing that the type of restricted channels should have the same use in its input/output use slots. To this aim, consider the process below *filter?(a,b).a?(n,c).if n ≥ 0 then new d in (b!(n,d1 ) | filter!(c,d2 )) else filter!(c,b) which filters numbers received from channel a and forwards the non-negative ones on channel b. Each number n comes along with a continuation channel c from which the next number in the stream will be received. Symmetrically, any message sent on b includes a continuation d on which the next non-negative number will be sent. For convenience, we distinguish d bound by new from the two rightmost occurrences d1 and d2 of d. For this process the reconstruction algorithm infers the type filter : [t × [int × t]0,1 ]ω,ω (7.2) 38 L. PADOVANI where t is the type that satisfies the equality t = [int × t]1,0 meaning that d1 and d2 are respectively assigned the types t and [int × t]0,1 and overall d has type t + [int × t]0,1 = [int × t]1,0 + [int × t]0,1 = [int × t]1,1 . The reason why d2 has type [int × t]0,1 , namely that d2 is used for an output operation, is clear, since d2 must have the same type as b and b is indeed used for an output operation in the body of filter. However, in the whole process there is no explicit evidence that d1 will be used for an input operation, and the input use 1 in its type t = [int × t]1,0 is deduced “by subtraction”, as we have discussed in the informal overview at the beginning of Section 5. If we do not impose the constraint that restricted (linear) channel should have the same input/output use, we can find a more precise solution that determines for filter the type filter : [t × [int × s]0,1 ]ω,ω (7.3) 0,0 where s is the type that satisfies the equality s = [int × s] . According to (7.3), d1 is assigned the type s saying that no operation will ever be performed on it. This phenomenon is a consequence of the fact that, when we apply the type reconstruction algorithm on an isolated process, like filter above, which is never invoked, the reconstruction algorithm has only a partial view of the behavior of the process on the channel it creates. For extruded channels like d, in particular, the algorithm is unable to infer any direct use. We argue that the typing (7.3) renders filter a useless process from which it is not possible to receive any message, unless filter is typed along with the rest of the program that invokes it. But this latter strategy prevents de facto the modular application of the reconstruction algorithm to the separate constituents of a program. The typing (7.2) is made possible by the completion phase (Section 5), which is an original feature of our type reconstruction algorithm. The prototype implementation of the algorithm provides a flag that disables the constraint on equal uses in [i-new] allowing experimentation of the behavior of the algorithm on examples like this one.  8. Concluding Remarks Previous works on the linear π-calculus either do not treat composite types [15, 10] or are based on an interpretation of linearity that limits data sharing and parallelism [8, 9]. Type reconstruction for recursive or, somewhat equivalently, infinite types has also been neglected, despite the key role played by these types for describing structured data (lists, trees, etc.) and structured interactions [2]. In this work we have extended the linear πcalculus with both composite and infinite types and have adopted a more relaxed attitude towards linearity that fosters data sharing and parallelism while maintaining the availability of a type reconstruction algorithm. The extension is a very natural one, as witnessed by the fact that our type system uses essentially the same rules of previous works, the main novelty being a different type combination operator. This small change has nonetheless nontrivial consequences on the reconstruction algorithm, which must reconcile the propagation of constraints across composite types and the impossibility to rely on plain type unification: different occurrences of the same identifier may be assigned different types and types may be infinite. Our extension also gives renewed relevance to types like [t]0,0 . In previous works these types were admitted but essentially useless: channels with such types could only be passed around in messages without actually ever being used. That is, they could be erased without affecting processes. In our type system, it is the existence of these types TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 39 that enables the sharing of structured data (see the decomposition of tlist into teven and todd in Section 1). Binary sessions [6, 7] can be encoded into the linear π-calculus [13, 3]. Thus, we indirectly provide a complete reconstruction algorithm for possibly infinite, higher-order, binary session types. As shown in [17], direct session type reconstruction poses two major technical challenges: on the one hand, the necessity to deal with dual types; on the other hand, the fact that subtyping must be taken into account for that is the only way to properly handle selections in conditionals. Interestingly, both complications disappear when session types are encoded in the linear π-calculus: duality simply turns into swapping the input/output use annotations in channel types [3], whereas selections become outputs of variant data types which can be dealt with using conventional techniques based on unification [16]. To assess the feasibility of the approach, we have implemented the type reconstruction algorithm in a tool for the static analysis of π-calculus processes. Given that even simple processes generate large constraint sets, the prototype has been invaluable for testing the algorithm at work on non-trivial examples. The reconstruction described in this article is only the first step for more advanced forms of analysis, such as those for reasoning on deadlocks and locks [19]. We have extended the tool in such a way that subsequent analyses can be plugged on top of the reconstruction algorithm for linear channels [21]. Structural subtyping and polymorphism are two natural developments of our work. The former has already been considered in [9], but it is necessary to understand how it integrates with our notion of type combination and how it affects constraint generation and resolution. Polymorphism makes sense for unlimited channels only (there is little point in having polymorphic linear channels, since they can only be used once anyway). Nevertheless, support for polymorphism is not entirely trivial, since some type variables may need to be restricted to unlimited types. For example, the channel first in the process *first?(x,y).y!fst(x) would have type ∀α.∀β.un(β) ⇒ [(α × β) × [α]0,1 ]ω,0 . Acknowledgements. The author is grateful to the anonymous reviewers whose numerous questions, detailed comments and suggestions have significantly contributed to improving both content and presentation of this article. The author is also grateful to Naoki Kobayashi for his comments on an earlier version of the article. References [1] B. Courcelle. Fundamental properties of infinite trees. Theor. Comp. Sci., 25:95–169, 1983. [2] O. Dardha. Recursive session types revisited. In BEAT’14, 2014. [3] O. Dardha, E. Giachino, and D. Sangiorgi. Session types revisited. In PPDP’12, pages 139–150. ACM, 2012. [4] R. Demangeon and K. Honda. Full abstraction in a subtyped pi-calculus with linear types. In CONCUR’11, LNCS 6901, pages 280–296. Springer, 2011. [5] S. J. Gay and M. Hole. Subtyping for session types in the pi calculus. Acta Informatica, 42(2-3):191–225, 2005. [6] K. Honda. Types for dyadic interaction. In CONCUR’93, LNCS 715, pages 509–523. Springer, 1993. [7] K. Honda, V. T. Vasconcelos, and M. Kubo. Language primitives and type disciplines for structured communication-based programming. In ESOP’98, LNCS 1381, pages 122–138. Springer, 1998. [8] A. Igarashi. Type-based analysis of usage of values for concurrent programming languages, 1997. Available at http://www.sato.kuis.kyoto-u.ac.jp/~ igarashi/papers/. [9] A. Igarashi and N. Kobayashi. Type-based analysis of communication for concurrent programming languages. In SAS’97, LNCS 1302, pages 187–201. Springer, 1997. 40 L. PADOVANI [10] A. Igarashi and N. Kobayashi. Type Reconstruction for Linear π-Calculus with I/O Subtyping. Inf. and Comp., 161(1):1–44, 2000. [11] N. Kobayashi. Quasi-linear types. In POPL’99, pages 29–42. ACM, 1999. [12] N. Kobayashi. A type system for lock-free processes. Inf. and Comp., 177(2):122–159, 2002. [13] N. Kobayashi. Type systems for concurrent programs. In 10th Anniversary Colloquium of UNU/IIST, LNCS 2757, pages 439–453. Springer, 2002. Extended version at http://www.kb.ecei.tohoku.ac.jp/~ koba/papers/tutorial-type-extended.pdf. [14] N. Kobayashi. A new type system for deadlock-free processes. In CONCUR’06, LNCS 4137, pages 233–247. Springer, 2006. [15] N. Kobayashi, B. C. Pierce, and D. N. Turner. Linearity and the pi-calculus. ACM Trans. Program. Lang. Syst., 21(5):914–947, 1999. [16] X. Leroy, D. Doligez, A. Frisch, J. Garrigue, D. Rémy, and J. Vouillon. The OCaml system release 4.01, 2013. Available at http://caml.inria.fr/pub/docs/manual-ocaml-4.01/index.html. [17] L. G. Mezzina. How to infer finite session types in a calculus of services and sessions. In COORDINATION’08, LNCS 5052, pages 216–231. Springer, 2008. [18] U. Nestmann and M. Steffen. Typing confluence. In FMICS’97, pages 77–101, 1997. Also available as report ERCIM-10/97-R052, European Research Consortium for Informatics and Mathematics, 1997. [19] L. Padovani. Deadlock and Lock Freedom in the Linear π-Calculus. In CSL-LICS’14, pages 72:1–72:10. ACM, 2014. [20] L. Padovani. Type reconstruction for the linear π-calculus with composite and equi-recursive types. In FoSSaCS’14, LNCS 8412, pages 88–102. Springer, 2014. [21] L. Padovani, T.-C. Chen, and A. Tosatto. Type Reconstruction Algorithms for Deadlock-Free and Lock-Free Linear π-Calculi. In COORDINATION’15, LNCS 9037, pages 83–98. Springer, 2015. [22] B. C. Pierce. Types and Programming Languages. The MIT Press, 2002. [23] B. C. Pierce. Advanced Topics in Types and Programming Languages. The MIT Press, 2004. [24] D. Sangiorgi and D. Walker. The Pi-Calculus - A theory of mobile processes. Cambridge University Press, 2001. [25] D. N. Turner, P. Wadler, and C. Mossin. Once upon a type. In FPCA’95, pages 1–11, 1995. Appendix A. Supplement to Section 3 To prove Theorem 3.8 we need a series of standard auxiliary results, including weakening (Lemma A.1) and substitution (Lemma A.2) for both expressions and processes. Lemma A.1 (weakening). The following properties hold: (1) If Γ ⊢ e : t and un(Γ ′ ) and Γ + Γ ′ is defined, then Γ + Γ ′ ⊢ e : t. (2) If Γ ⊢ P and un(Γ ′ ) and Γ + Γ ′ is defined, then Γ + Γ ′ ⊢ P . Proof. Both items are proved by a standard induction on the typing derivation. In case (2) we assume, without loss of generality, that bn(P ) ∩ dom(Γ ) = ∅ (recall that we identify processes modulo renaming of bound names). Lemma A.2 (substitution). Let Γ1 ⊢ v : t. The following properties hold: (1) If Γ2 , x : t ⊢ e : s and Γ1 + Γ2 is defined, then Γ1 + Γ2 ⊢ e{v/x} : s. (2) If Γ2 , x : t ⊢ P and Γ1 + Γ2 is defined, then Γ1 + Γ2 ⊢ P {v/x}. Proof. The proofs are standard, except for the following property of the type system: un(t) implies un(Γ1 ), which can be easily proved by induction on the derivation of Γ1 ⊢ v : t. TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 41 Next is type preservation under structural pre-congruence. Lemma A.3. If Γ ⊢ P and P 4 Q, then Γ ⊢ Q. Proof. We only show the case in which a replicated process is expanded. Assume P = *P ′ 4 *P ′ | P ′ = Q. From the hypothesis Γ ⊢ P and [t-rep] we deduce Γ ⊢ P ′ and un(Γ ). By definition of unlimited environment (see Definition 3.7) we have Γ = Γ + Γ . We conclude Γ ⊢ Q with an application of [t-par]. ℓ ℓ Lemma A.4. If Γ −→ Γ ′ and Γ + Γ ′′ is defined, then Γ + Γ ′′ −→ Γ ′ + Γ ′′ . ℓ Proof. Easy consequences of the definition of −→ on type environments. We conclude with type preservation for expressions and subject reduction for processes. Lemma A.5. Let Γ ⊢ e : t and e ↓ v. Then Γ ⊢ v : t. Proof. By induction on e ↓ v using the hypothesis that e is well typed. ℓ ℓ Theorem 3.8. Let Γ ⊢ P and P −→ Q. Then Γ ′ ⊢ Q for some Γ ′ such that Γ −→ Γ ′ . ℓ Proof. By induction on the derivation of P −→ Q and by cases on the last rule applied. We only show a few interesting cases; the others are either similar or simpler. [r-comm] Then P = e1 !f | e2 ?(x).R and ei ↓ a for every i = 1, 2 and f ↓ v and ℓ = a and Q = R{v/x}. From [t-par] we deduce Γ = Γ1 + Γ2 where Γ1 ⊢ e1 !f and Γ2 ⊢ e2 ?(x).R. From [t-out] we deduce Γ1 = Γ11 + Γ12 and Γ11 ⊢ e1 : [t]2κ1 ,1+κ2 and Γ12 ⊢ f : t. From [t-in] we deduce Γ2 = Γ21 +Γ22 and Γ21 ⊢ e2 : [s]1+κ3 ,2κ4 and Γ22 , x : s ⊢ R. From Lemma A.5 we have Γ11 ⊢ a : [t]2κ1 ,1+κ2 and Γ12 ⊢ v : t and Γ21 ⊢ a : [s]1+κ3 ,2κ4 . Also, since Γ11 + Γ21 is defined, it must be the case that t = s. Note that 1 + κ2 = 1 + 2κ2 and 1 + κ3 = 1 + 2κ3 . Hence, ′ , a : [t]2κ1 ,1+κ2 = (Γ ′ , a : [t]2κ1 ,2κ2 ) + a : [t]0,1 and from [t-name] we deduce that Γ11 = Γ11 11 ′ , a : [t]1+κ3 ,2κ4 = (Γ ′ , a : [t]2κ3 ,2κ4 ) + a : [t]1,0 for some unlimited Γ ′ and Γ ′ . Γ21 = Γ21 21 21 11 def ′ def ′ 2κ3 ,2κ4 2κ1 ,2κ2 ′′ ′′ ′′ ′′ and observe that Γ11 and Γ21 are also and Γ12 = Γ21 , a : [t] Let Γ11 = Γ11 , a : [t] ′′ + Γ + Γ ′′ + Γ . unlimited. From Lemma A.2 we deduce Γ12 + Γ22 ⊢ R{v/x}. Take Γ ′ = Γ11 22 12 21 a ′ ′ From Lemma A.1 we deduce Γ ⊢ Q and we conclude by observing that Γ −→ Γ thanks to Lemma A.4. [r-case] Then P = case e {i(xi ) ⇒ Pi }i=inl,inr and e ↓ k(v) for some k ∈ {inl, inr} and ℓ = τ and Q = Pk {v/xk }. From [t-case] we deduce that Γ = Γ1 + Γ2 and Γ1 ⊢ e : tinl ⊕ tinr and Γ2 , x : tk ⊢ Pk . From Lemma A.5 and either [t-inl] or [t-inr] we deduce Γ1 ⊢ v : tk . We conclude Γ ⊢ Pk {v/xk } by Lemma A.2. ℓ [r-par] Then P = P1 | P2 and P1 −→ P1′ and Q = P1′ | P2 . From [t-par] we deduce Γ = Γ1 + Γ2 and Γi ⊢ Pi . By induction hypothesis we deduce Γ1′ ⊢ P1′ for some Γ1′ such that ℓ ℓ Γ1 −→ Γ1′ . By Proposition A.4 we deduce that Γ −→ Γ1′ + Γ2 . We conclude Γ ′ ⊢ Q by taking Γ ′ = Γ1′ + Γ2 . 42 L. PADOVANI Appendix B. Supplement to Section 4 First of all we prove two technical lemmas that explain the relationship between the operators ⊔ and ⊓ used by the constraint generation rules (Table 7) and type environment combination + and equality used in the type rules (Table 4). Lemma B.1. If ∆1 ⊔ ∆2 ∆; C and σ is a solution for C covering ∆, then σ∆ = σ∆1 + σ∆2 . Proof. By induction on the derivation of ∆1 ⊔ ∆2 applied. We have two cases: ∆; C and by cases on the last rule dom(∆1 ) ∩ dom(∆2 ) = ∅ Then ∆ = ∆1 , ∆2 and we conclude σ∆ = σ∆1 , σ∆2 = σ∆1 + σ∆2 . ∆1 = ∆′1 , u : T and ∆2 = ∆′2 , u : S Then ∆′1 ⊔ ∆′2 ∆′ ; C ′ and ∆ = ∆′ , u : α and C = C ′ ∪ {α = ˆ T + S} for some α. Since σ is a solution for C, we deduce σ(α) = σT + σS. By induction hypothesis we deduce σ∆′ = σ∆′1 + σ∆′2 . We conclude σ∆ = σ∆′ , u : σ(α) = σ∆′ , u : σT + σS = (σ∆′1 + σ∆′2 ), u : σT + σS = σ∆1 + σ∆2 . Lemma B.2. If ∆1 ⊓∆2 ∆; C and σ is a solution for C covering ∆, then σ∆ = σ∆1 = σ∆2 . Proof. Straightforward consequence of the definition of ∆1 ⊓ ∆2 ∆; C. The correctness of constraint generation is proved by the next two results. Lemma B.3. If e : T ◮ ∆; C and σ is a solution for C covering ∆, then σ∆ ⊢ e : σT. Proof. By induction on the derivation of e : T ◮ ∆; C and by cases on the last rule applied. We only show two significant cases. [i-name] Then e = u and T = α fresh and ∆ = u : α and C = ∅. We have σ∆ = u : σ(α) and σT = σ(α), hence we conclude σ∆ ⊢ e : σT. [i-pair] Then e = (e1 ,e2 ) and T = T1 × T2 and C = C1 ∪ C2 ∪ C3 where ∆1 ⊔ ∆2 ∆; C3 and ei : Ti ◮ ∆i ; Ci for i = 1, 2. We know that σ is a solution for Ci for all i = 1, 2, 3. By induction hypothesis we deduce σ∆i ⊢ e : σTi for i = 1, 2. From Lemma B.1 we obtain σ∆ = σ∆1 + σ∆2 . We conclude with an application of [t-pair]. Theorem 4.2. If P ◮ ∆; C and σ is a solution for C that covers ∆, then σ∆ ⊢ P . Proof. By induction on the derivation of P ◮ ∆; C and by cases on the last rule applied. [i-idle] Then P = idle and ∆ = ∅ and C = ∅. We conclude with an application of [t-idle]. [i-in] Then P = e?(x).Q and e : T ◮ ∆1 ; C1 and Q ◮ ∆2 , x : S; C2 and ∆1 ⊔ ∆2 ∆; C3 and 1+̺1 ,2̺2 C = C1 ∪ C2 ∪ C3 ∪ {T = ˆ [S] }. By Lemma B.3 we deduce σ∆1 ⊢ e : σT. By induction hypothesis we deduce σ∆2 , x : σS ⊢ Q. By Lemma B.1 we deduce σ∆ = σ∆1 + σ∆2 . From the hypothesis that σ is a solution for C we know σT = [σS]1+σ(̺1 ),2σ(̺2 ) . We conclude with an application of [t-in]. [i-out] Then P = e!f and e : T ◮ ∆1 ; C1 and f : S ◮ ∆2 ; C2 and ∆1 ⊔ ∆2 2̺1 ,1+̺2 ∆; C3 and C = C1 ∪ C2 ∪ C3 ∪ {T = ˆ [S] }. By Lemma B.3 we deduce σ∆1 ⊢ e : σT and σ∆2 ⊢ f : σS. By Lemma B.1 we deduce σ∆ = σ∆1 + σ∆2 . From the hypothesis that σ is a solution for C we know σT = [σS]2σ(̺1 ),1+σ(̺2 ) . We conclude with an application of [t-out]. TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 43 [i-par] Then P = P1 |P2 and Pi ◮ ∆i ; Ci for i = 1, 2 and ∆1 ⊔∆2 ∆; C3 and C = C1 ∪C2 ∪C3 . By induction hypothesis we deduce σ∆i ⊢ Pi for i = 1, 2. By Lemma B.1 we deduce σ∆ = σ∆1 + σ∆2 . We conclude with an application of [t-par]. [i-rep] Then P = *Q and Q ◮ ∆′ ; C1 and ∆′ ⊔ ∆′ ∆; C2 and C = C1 ∪ C2 . By induction hypothesis we deduce σ∆′ ⊢ Q. By Lemma B.1 we deduce σ∆ = σ∆′ +σ∆′ . By Definition 3.7 we know that un(σ∆) holds. Furthermore, σ∆′ + σ∆ is defined. By Lemma A.1 and Definition 3.5 we deduce σ∆ ⊢ Q. We conclude with an application of [t-rep]. ˆ [α]̺,̺ }. By [i-new] Then P = new a in Q and Q ◮ ∆, a : T; C ′ and C = C ′ ∪ {T = induction hypothesis we deduce σ∆, a : σT ⊢ Q. Since σ is a solution for C ′ we know that σT = [σ(α)]σ(̺),σ(̺) . We conclude with an application of [t-new]. [i-case] Then P = case e {i(xi ) ⇒ Pi }i=inl,inr and e : t ◮ ∆1 ; C1 and Pi ◮ ∆i , xi : Ti ; Ci for i = inl, inr and ∆inl ⊓ ∆inr ∆2 ; C2 and ∆1 ⊔ ∆2 ∆; C3 and C = C1 ∪ C2 ∪ ˆ Tinl ⊕ Tinr }. By Lemma B.3 we deduce σ∆1 ⊢ e : σT. By C3 ∪ Cinl ∪ Cinr ∪ {T = induction hypothesis we deduce σ∆i ⊢ Pi for i = inl, inr. By Lemma B.2 we deduce σ∆inl = σ∆inr = σ∆2 . By Lemma B.1 we deduce σ∆ = σ∆1 + σ∆2 . Since σ is a solution for C, we have σT = σTinl ⊕ σTinr . We conclude with an application of [t-case]. [i-weak] Then ∆ = ∆′ , u : α and C = C ′ ∪ {un(α)} where α is fresh and P ◮ ∆′ ; C ′ . By induction hypothesis we deduce σ∆′ ⊢ P . Since σ is a solution for C ′ we know that un(σ(α)) holds. Since u 6∈ dom(∆′ ) we know that ∆′ σ + u : σ(α) is defined. By Lemma A.1(2) we conclude σ∆′ , u : σ(α) ⊢ P . The next lemma relates once more ⊔ and type environment combination +. It is, in a sense, the inverse of Lemma B.1. Lemma B.4. If σ∆1 +σ∆2 is defined, then there exist ∆, C, and σ ′ ⊇ σ such that ∆1 ⊔∆2 ∆; C and σ ′ is a solution for C that covers ∆. Proof. By induction on the maximum size of ∆1 and ∆2 . We distinguish two cases. def def def dom(∆1 ) ∩ dom(∆2 ) = ∅ We conclude by taking ∆ = ∆1 , ∆2 and C = ∅ and σ ′ = σ and observing that ∆1 ⊔ ∆2 ∆; ∅. ∆1 = ∆′1 , u : T and ∆2 = ∆′2 , u : S Since σ∆1 + σ∆2 is defined, we know that σ∆′1 + σ∆′2 is defined as well and furthermore that (σ∆1 + σ∆2 )(u) = σT + σS. By induction hypothesis ∆′ ; C ′ and σ ′′ is a we deduce that there exist ∆′ , C ′ , and σ ′′ ⊇ σ such that ∆′1 ⊔ ∆′2 def def solution for C ′ that covers ∆′ . Take ∆ = ∆′ , u : α where α is fresh, C = C ′ ∪ {α = ˆ T + S} def ′′ ′ ′ and σ = σ ∪ {α 7→ σT + σS}. We conclude observing that σ is a solution for C that covers ∆. In order to prove the completeness of type reconstruction for expressions, we extend the reconstruction algorithm with one more weakening rule for expressions: [i-weak expr] e : T ◮ ∆; C e : T ◮ ∆, u : α; C ∪ un(α) 44 L. PADOVANI This rule is unnecessary as far as completeness is concerned, because there is already a weakening rule [i-weak] for processes that can be used to subsume it. However, [i-weak expr] simplifies both the proofs and the statements of the results that follow. Lemma B.5. If Γ ⊢ e : t, then there exist T, ∆, C, and σ such that e : T ◮ ∆; C and σ is a solution for C and Γ = σ∆ and t = σT. Proof. By induction on the derivation of Γ ⊢ e : t and by cases on the last rule applied. We only show two representative cases. def [t-name] Then e = u and Γ = Γ ′ , u : t and un(Γ ′ ). Let Γ ′ = {ui : ti }i∈I . Take T = α def def def and ∆ = {ui : αi }i∈I , u : α and C = {un(αi ) | i ∈ I} and σ = {αi 7→ ti }i∈I ∪ {α 7→ t} where α and the αi ’s are all fresh type variables. Observe that e : T ◮ ∆; C by means of one application of [i-name] and as many applications of [i-weak expr] as the cardinality of I. We conclude observing that σ is a solution for C and Γ = σ∆ and t = σT by definition of σ. [t-pair] Then e = (e1 ,e2 ) and Γ = Γ1 + Γ2 and t = t1 × t2 and Γi ⊢ ei : ti for i = 1, 2. By induction hypothesis we deduce that there exist Ti , ∆i , Ci , and σi solution for Ci such that ei : Ti ◮ ∆i ; Ci and Γi = σi ∆i and ti = σi Ti for i = 1, 2. Since the reconstruction algorithm always chooses fresh type variables, we also know that dom(σ1 ) ∩ dom(σ2 ) = ∅. def Take σ ′ = σ1 ∪ σ2 . We have that σ ′ ∆1 + σ ′ ∆2 = Γ1 + Γ2 is defined. Therefore, by Lemma B.4, we deduce that there exist ∆, C3 , and σ ⊇ σ ′ such that ∆1 ⊔ ∆2 ∆; C3 and σ is a solution def for C that covers ∆. We conclude with an application of [i-pair] and taking T = T1 × T2 def and C = C1 ∪ C2 ∪ C3 . Theorem 4.3. If Γ ⊢ P , then there exist ∆, C, and σ such that P ◮ ∆; C and σ is a solution for C that covers ∆ and Γ = σ∆. Proof. By induction on the derivation of Γ ⊢ P and by cases on the last rule applied. We only show a few cases, the others being analogous. def [t-idle] Then P = idle and un(Γ ). Let Γ = {ui : ti }i∈I . Take ∆ = {ui : αi }i∈I and def def C = {un(αi )}i∈I and σ = {αi 7→ ti }i∈I where the αi ’s are all fresh type variables. By repeated applications of [i-weak] and one application of [i-idle] we derive idle ◮ ∆; C. We conclude observing that σ is a solution for C and Γ = σ∆. [t-in] Then P = e?(x).Q and Γ = Γ1 + Γ2 and Γ1 ⊢ e : [t]1+κ1 ,2κ2 and Γ2 , x : t ⊢ Q. By Lemma B.5 we deduce that there exist T, ∆1 , C1 , and σ1 solution for C1 such that e : T ◮ ∆1 ; C1 and Γ1 = σ1 ∆1 and [t]1+κ1 ,2κ2 = σ1 T. By induction hypothesis we deduce that there exist ∆′2 , C2 , and σ2 solution for C2 such that Γ2 , x : t = σ2 ∆′2 . Then it must be the case that ∆′2 = ∆2 , x : S for some ∆2 and S such that Γ2 = σ2 ∆2 and t = σ2 S. Since all type variables chosen by the type reconstruction algorithm are fresh, we know that dom(σ1 )∩ dom(σ2 ) = ∅. def Take σ ′ = σ1 ∪σ2 ∪{̺1 7→ κ1 , ̺2 7→ κ2 }. Observe that σ ′ ∆1 +σ ′ ∆2 = Γ1 +Γ2 which is defined. By Lemma B.4 we deduce that there exist ∆, C3 , and σ ⊇ σ ′ such that ∆1 ⊔ ∆2 ∆; C3 def 1+̺1 ,2̺2 and σ is a solution for C3 that covers ∆. Take C = C1 ∪ C2 ∪ C3 ∪ {T = ˆ [S] }. Then σ is a solution for C, because σT = [t]1+κ1 ,2κ2 = [σS]1+σ(̺1 ),2σ(̺2 ) = σ[S]1+̺1 ,2̺2 . Also, by Lemma B.1 we have σ∆ = σ∆1 + σ∆2 = Γ1 + Γ2 = Γ . We conclude P ◮ ∆; C with an application of [i-in]. TYPE RECONSTRUCTION FOR THE LINEAR π-CALCULUS 45 [t-par] Then P = P1 | P2 and Γ = Γ1 + Γ2 and Γi ⊢ Pi for i = 1, 2. By induction hypothesis we deduce that, for every i = 1, 2, there exist ∆i , Ci , and σi solution for Ci such that Pi ◮ ∆i ; Ci and Γi = σi ∆i . We also know that dom(σ1 ) ∩ dom(σ2 ) = ∅ because type/use def variables are always chosen fresh. Take σ ′ = σ1 ∪ σ2 . By Lemma B.4 we deduce that there exist ∆, C3 , and σ ⊇ σ ′ such that ∆1 ⊔ ∆2 ∆; C3 and σ is a solution for C3 that covers ∆. By Lemma B.1 we also deduce that σ∆ = σ∆1 + σ∆2 = Γ1 + Γ2 = Γ . We conclude by taking def C = C1 ∪ C2 ∪ C3 with an application of [i-par]. [t-rep] Then P = *Q and Γ ⊢ Q and un(Γ ). By induction hypothesis we deduce that there exist ∆′ , C ′ , and σ ′ solution for C ′ such that Q ◮ ∆′ ; C ′ and Γ = σ ′ ∆′ . Obviously σ ′ ∆′ + σ ′ ∆′ is defined, hence by Lemma B.4 we deduce that there exist ∆, C ′′ , and σ ⊇ σ ′ such that ∆′ ⊔ ∆′ ∆; C ′′ and σ is a solution for C ′′ . By Lemma B.1 we deduce σ∆ = σ∆′ + σ∆′ = Γ + Γ = Γ , where the last equality follows from the hypothesis un(Γ ) and Definition 3.7. We def conclude P ◮ ∆; C with an application of [i-rep] by taking C = C ′ ∪ C ′′ . Appendix C. Supplement to Section 5 Below is the derivation showing the reconstruction algorithm at work on the process (5.6). a : α1 ◮ a : α1 ; ∅ 3 : int ◮ ∅; ∅ 2̺1 ,1+̺2 a!3 ◮ a : α1 ; {α1 = ˆ [int] } [i-out] b : β ◮ b : β; ∅ a : α2 ◮ a : α2 ; ∅ b!a ◮ a : α2 , b : β; {β = ˆ [α2 ]2̺3 ,1+̺4 } ˆ [α2 ]2̺3 ,1+̺4 } a!3 | b!a ◮ a : α, b : β; {α = ˆ α1 + α2 , α1 = ˆ [int]2̺1 ,1+̺2 , β = new a in (a!3 | b!a) ◮ b : β; {α = ˆ [δ]̺5 ,̺5 , α = ˆ α1 + α2 , . . . } [i-out] [i-par] [i-new] Below is the derivation showing the reconstruction algorithm at work on the process (5.7). Only the relevant differences with respect to the derivation above are shown. .. .. .. .. . . . . .. . b!a ◮ a : α2 , b : β; {β = ˆ [α2 ]2̺3 ,1+̺4 } c!a ◮ a : α3 , c : γ; {γ = ˆ [α3 ]2̺5 ,1+̺6 } b!a | c!a ◮ a : α23 , b : β, c : γ; {α23 = ˆ α2 + α3 , . . . } a!3 | b!a | c!a ◮ a : α, b : β, c : γ; {α = ˆ α1 + α23 , . . . } [i-par] [i-par] new a in (a!3 | b!a | c!a) ◮ b : β, c : γ; {α = ˆ [δ]̺5 ,̺5 , α = ˆ α1 + α23 , . . . } [i-new] Below is the derivation showing the reconstruction algorithm at work on the process (5.8). b : β ◮ b : β; ∅ x : γ ◮ x : γ; ∅ b!x ◮ b : β, x : γ; {β = ˆ [γ]2̺1 ,1+̺2 } [i-out] a?(x).b!x ◮ a : α, b : β; {α = ˆ [γ]1+̺3 ,2̺4 , β = ˆ [γ]2̺1 ,1+̺2 } [i-in] 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 

Ensemble of Heterogeneous Flexible Neural Trees Using Multiobjective Genetic Programming Varun Kumar Ojhaa,∗, Ajith Abrahamb , Václav Snášela a IT4Innovations, arXiv:1705.05592v1 [cs.NE] 16 May 2017 b Machine VŠB-Technical University of Ostrava, Ostrava, Czech Republic Intelligence Research Labs (MIR Labs), Auburn, WA, USA Abstract Machine learning algorithms are inherently multiobjective in nature, where approximation error minimization and model’s complexity simplification are two conflicting objectives. We proposed a multiobjective genetic programming (MOGP) for creating a heterogeneous flexible neural tree (HFNT), tree-like flexible feedforward neural network model. The functional heterogeneity in neural tree nodes was introduced to capture a better insight of data during learning because each input in a dataset possess different features. MOGP guided an initial HFNT population towards Pareto-optimal solutions, where the final population was used for making an ensemble system. A diversity index measure along with approximation error and complexity was introduced to maintain diversity among the candidates in the population. Hence, the ensemble was created by using accurate, structurally simple, and diverse candidates from MOGP final population. Differential evolution algorithm was applied to fine-tune the underlying parameters of the selected candidates. A comprehensive test over classification, regression, and time-series datasets proved the efficiency of the proposed algorithm over other available prediction methods. Moreover, the heterogeneous creation of HFNT proved to be efficient in making ensemble system from the final population. Keywords: Pareto-based multiobjectives, flexible neural tree, ensemble, approximation, feature selection; 1. Introduction Structure optimization of a feedforward neural network (FNN) and its impact on FNN’s generalization ability inspired the flexible neural tree (FNT) [1]. FNN components such as weights, structure, and activation function are the potential candidates for the optimization, which improves FNN’s generalization ability to a great extent [2]. These efforts are notable because of FNN’s ability to solve a large range of realworld problems [3, 4, 5, 6]. Followings are the significance structure optimization methods: constructive and pruning algorithms [7, 8], EPNet [2], NeuroEvolution of Augmenting Topologies [9], sparse neural ∗ Varun Kumar Ojha Email addresses: varun.kumar.ojha@vsb.cz (Varun Kumar Ojha), ajith.abraham@ieee.org (Ajith Abraham), vaclav.snasel@vsb.cz (Václav Snášel) Preprint submitted to Applied Soft Computing, Volume 52 Pages 909 to 924 May 17, 2017 trees [10], Cooperative co-evolution approach [11], etc. Similarly, many efforts focus on the optimization of hybrid training of FNN such as [12, 13, 14]. FNT was an additional step into this series of efforts, which was proposed to evolve as a tree-like feed-forward neural network model, where the probabilistic incremental program evolution (PIPE) [15] was applied optimize the tree structure [1]. The underlying parameter vector of the developed FNT (weights associated with the edges and arguments of the activation functions) was optimized by metaheuristic algorithms, which are nature-inspired parameter optimization algorithms [16]. The evolutionary process allowed FNT to select significant input features from an input feature set. In the design of FNT, the non-leaf nodes are the computational node, which takes an activation function. Hence, rather than relying on a fixed activation function, if the selection of activation function at the computational nodes is allowed to be selected by the evolutionary process. Then, it produces heterogeneous FNTs (HFNT) with the heterogeneity in its structure, computational nodes, and input set. In addition, heterogeneous function allowed HFNT to capture different feature of the datasets efficiently since each input in the datasets posses different features. The evolutionary process provides adaptation in structure, weights, activation functions, and input features. Therefore, an optimum HFNT is the one that offers the lowest approximation error with the simplest tree structure and the smallest input feature set. However, approximation error minimization and structure simplification are two conflicting objectives [17]. Hence, a multiobjective evolutionary approach [18] may offer an optimal solution(s) by maintaining a balance between these objectives. Moreover, in the proposed work, an evolutionary process guides a population of HFNTs towards Pareto-optimum solutions. Hence, the final population may contain several solutions that are close to the best solution. Therefore, an ensemble system was constructed by exploiting many candidates of the population (candidate, solution, and model are synonymous in this article). Such ensemble system takes advantage of many solutions including the best solution [19]. Diversity among the chosen candidates holds the key in making a good ensemble system [20]. Therefore, the solutions in a final population should fulfill the following objectives: low approximation error, structural simplicity, and high diversity. However, these objectives are conflicting to each other. A fast elitist nondominated sorting genetic algorithm (NSGA-II)based multiobjective genetic programming (MOGP) was employed to guide a population of HFNTs [21]. The underlying parameters of selected models were further optimized by using differential evaluation (DE) algorithm [22]. Therefore, we may summarize the key contributions of this work are as follows: 1) A heterogeneous flexible neural tree (HFNT) for function approximation and feature selection was proposed. 2) HFNT was studied under an NSGA-II-based multiobjective genetic programming framework. Thus, it was termed HFNTM . 3) Alongside approximation error and tree size (complexity), a diversity index was introduced to maintain 2 diversity among the candidates in the population. 4) HFNTM was found competitive with other algorithms when compared and cross-validated over classification, regression, and time-series datasets. 5) The proposed evolutionary weighted ensemble of HFNTs final population further improved its performance. A detailed literature review provides an overview of FNT usage over the past few years (Section 2). Conclusions derived from literature survey supports our HFNTM approach, where a Pareto-based multiobjective genetic programming was used for HFNT optimization (Section 3.1). Section 3.2 provides a detailed discussion on the basics of HFNT: MOGP for HFNT structure optimization, and DE for HFNT parameter optimization. The efficiency of the above-mentioned hybrid and complex multiobjective FNT algorithm (HFNTM ) was tested over various prediction problems using a comprehensive experimental setup (Section 4). The experimental results support the merits of proposed approach (Section 5). Finally, we provide a discussion of experimental outcomes in Section 6 followed by conclusions in Section 7. 2. Literature Review The literature survey describes the following points: basics of FNT, approaches that improvised FNT, and FNTs successful application to various real-life problems. Subsequently, the shortcomings of basic FNT version are concluded that inspired us to propose HFNTM . FNT was first proposed by Chen et al. [1], where a tree-like-structure was optimized by using PIPE. Then, its approximation ability was tested for time-series forecasting [1] and intrusion detection [23], where a variant of simulated annealing (called degraded ceiling) [24], and particle swarm optimization (PSO) [25], respectively, were used for FNT parameter optimization. Since FNT is capable of input feature selection, in [26], FNT was applied for selecting input features in several classification tasks, in which FNT structure was optimized by using genetic programming (GP) [27], and the parameter optimization was accomplished by using memetic algorithm [28]. Additionally, they defined five different mutation operators, namely, changing one terminal node, all terminal nodes, growing a randomly selected sub-tree, pruning a randomly selected sub-tree, and pruning redundant terminals. Li et al. [29] proposed FNTbased construction of decision trees whose nodes were conditionally replaced by neural node (activation node) to deal with continuous attributes when solving classification tasks. In many other FNT based approaches, like in [30], GP was applied to evolve hierarchical radial-basis-function network model, and in [31] a multi-input-multi-output FNT model was evolved. Wu et al. [32] proposed to use grammar guided GP [33] for FNT structure optimization. Similarly, in [34], authors proposed to apply multi-expression programming (MEP) [35] for FNT structure optimization and immune programming algorithm [36] for the parameter vector optimization. To improve classification accuracy of FNT, Yang et al. [37] proposed a hybridization of FNT with a further-division-of-partition-space method. In [38], authors illustrated 3 crossover and mutation operators for evolving FNT using GP and optimized the tree parameters using PSO algorithm. A model is considered efficient if it has generalization ability. We know that a consensus decision is better than an individual decision. Hence, an ensemble of FNTs may lead to a better-generalized performance than a single FNT. To address this, in [39], authors proposed to make an ensemble of FNTs to predict the chaotic behavior of stock market indices. Similarly, in [40], the proposed FNTs ensemble predicted the breast cancer and network traffic better than individual FNT. In [41], protein dissolution prediction was easier using ensemble than the individual FNT. To improve the efficiency in terms of computation, Peng et al. [42] proposed a parallel evolving algorithm for FNT, where the parallelization took place in both tree-structure and parameter vector populations. In another parallel approach, Wang et al. [43] used gene expression programming (GEP) [44] for evolving FNT and used PSO for parameter optimization. A multi-agent system [45] based FNT (MAS-FNT) algorithm was proposed in [46], which used GEP and PSO for the structure and parameter optimization, respectively. The MAS-FNT algorithm relied on the division of the main population into sub-population, where each sub-population offered local solutions and the best local solution was picked-up by analyzing tree complexity and accuracy. Chen et al. [1, 26] referred the arbitrary choice of activation function at non-leaf nodes. However, they were restricted to use only Gaussian functions. A performance analysis of various activation function is available in [47]. Bouaziz et al. [48, 49] proposed to use beta-basis function at non-leaf nodes of an FNT. Since beta-basis function has several controlling parameters such as shape, size, and center, they claimed that the beta-basis function has advantages over other two parametric activation functions. Similarly, many other forms of neural tree formation such as balanced neural tree [50], generalized neural tree [51], and convex objective function neural tree [52], were focused on the tree improvement of neural nodes. FNT was chosen over the conventional neural network based models for various real-world applications related to prediction modeling, pattern recognition, feature selection, etc. Some examples of such applications are cement-decomposing-furnace production-process modeling [53], time-series prediction from gene expression profiling [54]. stock-index modeling [39], anomaly detection in peer-to-peer traffic [55], intrusion detection [56], face identification [57], gesture recognition [58], shareholder’s management risk prediction [59], cancer classification [60], somatic mutation, risk prediction in grid computing [61], etc. The following conclusions can be drawn from the literature survey. First, FNT was successfully used in various real-world applications with better performance than other existing function approximation models. However, it was mostly used in time-series analysis. Second, the lowest approximation error obtained by an individual FNT during an evolutionary phase was considered as the best structure that propagated to the parameter optimization phase. Hence, there was no consideration as far as structural simplicity and generalization ability are concerned. Third, the computational nodes of the FNT were 4 fixed initially, and little efforts were made to allow for its automatic adaptation. Fourth, little attention was paid to the statistical validation of FNT model, e.g., mostly the single best model was presented as the experimental outcome. However, the evolutionary process and the meta-heuristics being stochastic in nature, statistical validation is inevitably crucial for performance comparisons. Finally, to create a generalized model, an ensemble of FNTs were used. However, FNTs were created separately for making the ensemble. Due to stochastic nature of the evolutionary process, FNT can be structurally distinct when created at different instances. Therefore, no explicit attention was paid to create diverse FNTs within a population itself for making ensemble. In this article, a heterogeneous FNT called HFNT was proposed to improve the basic FNT model and its performance by addressing above mentioned shortcomings. 3. Multi-objectives and Flexible Neural Tree In this section, first, Pareto-based multiobjective is discussed. Second, we offer a detailed discussion on FNT and its structure and parameter optimization using NSGA-II-based MOGP and DE, respectively. Followed by a discussion on making an evolutionary weighted ensemble of the candidates from the final population. 3.1. Pareto-Based Multi-objectives Usually, learning algorithms owns a single objective, i.e., the approximation error minimization, which is often achieved by minimizing mean squared error (MSE) on the learning data. MSE E on a learning data is computed as: E= N 1 X (di − yi )2 , N i=1 (1) where di and yi are the desired output and the model’s output, respectively and N indicates total data pairs in the learning set. Additionally, a statistical goodness measure, called, correlation coefficient r that tells the relationship between two variables (i.e., between the desired output d and the model’s output y) may also be used as an objective. Correlation coefficient r is computed as:
PN i=1 di − d̄i (yi − ȳi ) , r = qP
2 PN N 2 (y − ȳ ) d − d̄ i i i i i=1 i=1 (2) where d̄ and ȳ are means of the desired output d and the model’s output y, respectively. However, single objective comes at the expense of model’s complexity or generalization ability on unseen data, where generalization ability broadly depends on the model’s complexity [62]. A common model complexity indicator is the number of free parameters in the model. The approximation error (1) and the number of free parameters minimization are two conflicting objectives. One approach is to combine these two objectives as: f = αE + (1 − α)D, 5 (3) where 0 ≤ α ≤ 1 is a constant, E is the MSE (1) and D is the total free parameter in a model. The scalarized objective f in (3), however, has two disadvantages. First, determining an appropriate α that controls the conflicting objectives. Hence, generalization ability of the produced model will be a mystery [63]. Second, the scalarized objective f in (3) leads to a single best model that tells nothing about how the conflicting objectives were achieved. In other words, no single solution exists that may satisfy both objectives, simultaneously. We study a multiobjective optimization problem of the form: minimize {f1 (w), f2 (w), . . . , fm (w)} subject to w ∈ W where we have m ≥ 2 objective functions fi : Rn → R. We denote the vector of objective functions by f (w) = hf1 (w), f2 (w), . . . , fm (w)i. The decision (variable) vectors w = hw1 , w2 , . . . , wn i belong to the set W ⊂ Rn , which is a subset of the decision variable space Rn . The word ‘minimize’ means that we want to minimize all the objective functions simultaneously. A nondominated solution is one in which no one objective function can be improved without a simultaneous detriment to at least one of the other objectives of the solution [21]. The nondominated solution is also known as a Pareto-optimal solution. Definition 1. Pareto-dominance - A solution w1 is said to dominate a solution w2 if ∀i = 1, 2, . . . , m, fi (w1 ) ≤ fi (w2 ), and there exists j ∈ {1, 2, . . . , m} such that fj (w1 ) < fj (w2 ) holds. Definition 2. Pareto-optimal - A solution w1 is called Pareto-optimal if there does not exist any other solution that dominates it. A set Pareto-optimal solution is called Pareto-front. Algorithm 1 is a basic framework of NSGA-II based MOGP, which was used for computing Paretooptimal solutions from an initial HFNT population. The individuals in MOGP were sorted according to their dominance in population. Note that the function size(·) returns total number of rows (population size) for a 2-D matrix and returns total number of elements for a vector. The Moreover, individuals were sorted according to the rank/Pareto-front. MOGP is an elitist algorithm that allowed the best individuals to propagate into next generation. Diversity in the population was maintained by measuring crowding distance among the individuals [21]. 6 Data: Problem and Objectives Result: A bag M of solutions selected from Pareto-fronts initialization: HFNT population P ; evaluation: nondominated sorting of P ; while termination criteria not satisfied do selection: binary tournament selection; generation: a new population Q; recombination: R = P + Q; evaluation: nondominated sorting of R; elitism: P = size(P ) best individuals from R; end Algorithm 1: NSGA-II based multiobjective genetic programming 3.2. Heterogeneous Flexible Neural Tree HFNT is analogous to a multi-layer feedforward neural network that has over-layer connections and activation function at the nodes. HFNT construction has two phases [1]: 1) the tree construction phase, in which evolutionary algorithms are applied to construct tree-like structure; and 2) the parameter-tuning phase, in which genotype of HFNT (underlying parameters of tree-structure) is optimized by using parameter optimization algorithms. To create a near-optimum model, phase one starts with random tree-like structures (population of initial solutions), where parameters of each tree are fixed by a random guess. Once a near-optimum tree structure is obtained, parameter-tuning phase optimizes its parameter. The phases are repeated until a satisfactory solution is obtained. Figure 1 is a lucid illustration of these two phases that work in some co-evolutionary manner. From Figure 1, it may be observed that two global search algorithms MOGP (for structure optimization) and DE (for parameter optimization) works in a nested manner to obtain a near optimum tree that may have less complex tree structure and better parameter. Moreover, evolutionary algorithm allowed HFNT to select activation functions and input feature at the nodes from sets of activation functions and input features, respectively. Thus, HFNT possesses automatic feature selection ability. 3.2.1. Basic Idea of HFNT An HFNT S is a collection of function set F and instruction set T : n o U (k) U (k) U (k) S = F ∪ T = +2 , +3 , · · · , +tn ∪ {x1 , x2 , . . . , xd } (4) where +kj (j = 2, 3, . . . , tn) denotes a non-leaf instruction (a computational node). It receives 2 ≤ j ≤ tn arguments and U (k) is a function that randomly takes an activation function from a set of k activation functions. Maximum arguments tn to a computational node are predefined. A set of seven activation 7 Input: Training data and parameter settings MOGP/SOGP: Initialization of HFNT Population and objective function setting Yes No If MOGP ? NSGA-II-based nondominated sorting Fitness based sorting New population using selection, crossover, and mutation Fitness Evaluation No max iteration? Yes DE: Initialization of the population for parameter tuning for a selected fixed HFNT structure New population using selection, crossover, and mutation No max iteration? Yes No Satisfactory solution found ? Yes STOP Figure 1: Co-evolutionary construction of the heterogeneous flexible neural tree. 8 Table 1: Set of activation function used in neural tree construction Activation-function k Expression for ϕki (a, b, x)
Gaussian Function 1 f (x, a, b) = exp −((x − a)2 )/(b2 ) Tangent-Hyperbolic 2 f (x) = (ex − e−x )/(ex + e−x ) Fermi Function 3 f (x) = 1/(1 + e−x ) Linear Fermi 4 f (x, a, b) = a × 1/((1 + e−x )) + b Linear Tangent-hyperbolic 5 f (x, a, b) = a × (ex − e−x )/(ex + e−x ) + b Bipolar Sigmoid 6 f (x, a) = (1 − e−2xa )/(a(1 + e−2xa )) Unipolar Sigmoid 7 f (x, a) = (2|a|)/(1 + e−2|a|x ) functions is shown in Table 1. Leaf node’s instruction x1 , x2 , . . . , xd denotes input variables. Figure 2 is an illustration of a typical HFNT. Similarly, Figure 3 is an illustration of a typical node in an HFNT. The i-th computational node (Figure 3) of a tree (say i-th node in Figure 2) receives ni inputs (denoted as zji ) through ni connection-weights (denoted as wji ) and takes two adjustable parameters ai and bi that represents the arguments of the activation function ϕki (.) at that node. The purpose of using an activation function at a computational node is to limit the output of the computational node within a certain range. For example, if the i-th node contains a Gaussian function k = 1 (Table 1). Then, its output yi is computed as:    oi − ai yi = ϕki (ai , bi , oi ) = exp − bi (5) where oi is the weighted summation of the inputs zji and weights wji (j = 1 to ni ) at the i-th computational node (Figure 3), also known as excitation of the node. The net excitation oi of the i-th node is computed as: i oi = n X wji zji (6) j=1 where zji ∈ {x1 , x2 , . . . , xd } or, zji ∈ {y1 , y2 , . . . , ym }, i.e., zji can be either an input feature (leaf node value) or the output of another node (a computational node output) in the tree. Weight wji is the connection weight of real value in the range [wl , wu ]. Similarly, the output of a tree y is computed from the root node of the tree, which is recursively computed by computing each node’s output using (5) from right to left in a depth-first method. The fitness of a tree depends on the problem. Usually, learning algorithm uses approximation error, i.e., MSE (1). Other fitness measures associated with the tree are tree size and diversity index. The tree size is the number of nodes (excluding root node) in a tree, e.g., the number of computational nodes and leaf nodes in the tree in Figure 2 is 11 (three computational nodes and eight leaf-nodes). The number of distinct activation functions (including root node function) randomly selected from a set of activation functions gives the diversity index of a tree. Total activation functions (denoted as k in +kj ) selected by the tree in Figure 2 is three (+13 , +43 , and +53 ). Hence, its diversity index is three. 9 +13 Root node x1 +52 x2 x3 +13 Leaf nodes x2 x1 x3 x1 +43 Function nodes x3 Depth-first search computation y  Figure 2: Typical representation of a neural tree S = F ∪ T whose function instruction set F = +13 , +42 , +53 and terminal instruction set T = {x1 , x2 , x3 , x4 }. z1i w1i w2i z2i yi i n P j=1 wji zji wni i zni i Figure 3: Illustration of a computational node. The variable ni indicates the number of inputs zji and weights wji received at the i-th node and the variable y i is the output of the i-th node. 10 3.3. Structure and Parameter Learning (Near optimal Tree) A tree that offers the lowest approximation error and the simplest structure is a near optimal tree, which can be obtained by using an evolutionary algorithm such as GP [27], PIPE [15], GEP [44], MEP [35], and so on. To optimize tree parameters, algorithms such as genetic algorithm [64], evolution strategy [64], artificial bee colony [65], PSO [25, 66], DE [22], and any hybrid algorithm such as GA and PSO [67] can be used. 3.3.1. Tree-construction The proposed multiobjective optimization of FNT has three fitness measures: approximation error (1) minimization, tree size minimization, and diversity index maximization. These objectives are simultaneously optimized during the tree construction phase using MOGP, which guides an initial population P of random tree-structures according to Algorithm 1. The detailed description of the components of Algorithm 1 are as follows: Selection. In selection operation, a mating pool of size size(P )r is created using binary tournament selection, where two candidates are randomly selected from a population and the best (according to rank and crowding distance) among them is placed into the mating pool. This process is continued until the mating pool is full. An offspring population Q is generated by using the individuals of mating pool. Two distinct individuals (parents) are randomly selected from the mating pool to create new individuals using genetic operators: crossover and mutation. The crossover and mutation operators are applied with probabilities pc and pm, respectively. Crossover. In crossover operation, randomly selected sub-trees of two parent trees were swapped. The swapping includes the exchange of activation-nodes, weights, and inputs as it is described in [38, 64, 68]. Mutation. The mutation of a selected individual from mating pool took place in the following manner [38, 64, 68]: 1) A randomly selected terminal node is replaced by a newly generated terminal node. 2) All terminal nodes of the selected tree were replaced by randomly generated new terminal nodes. 3) A randomly selected terminal node or a computational node is replaced by a randomly generated sub-tree. 4) A randomly selected terminal node is replaced by a randomly generated computational node. In the proposed MOGP, during the each mutation operation event, one of the above-mentioned four mutation operators was randomly selected for mutation of the tree. Recombination. The offspring population Q and the main population P , are merged to make a combined population R. 11 Elitism. In this step, size(Q) worst individuals are weeded out. In other words, size(P ) best individuals are propagated to a new generation as main population P . 3.3.2. Parameter-tuning In parameter-tuning phase, a single objective, i.e., approximation error was used in optimization of HFNT parameter by DE. The tree parameters such as weights of tree edges and arguments of activation functions were encoded into a vector w = hw1 , w2 , . . . , wn i for the optimization. In addition, a crossvalidation (CV) phase was used for statistical validation of HFNTs. The basics of DE is as follows. For an initial population H of parameter vectors wi for i = 1 to size(H), DE repeats its steps mutation, recombination, and selection until an optimum parameter vector w∗ is obtained. DE updates each parameter vector wi ∈ H by selecting the best vector wgi and three random vectors r0i , r1i , and r2i from H such that r0i 6= r1i 6= r2i holds. The random vector r0 is considered as a t trial vector wit . Hence, for all i = 1, 2, . . . , size(H), and j = 1, 2, . . . , n, the j-th variable wij of i-th trail-vectors wit is generated by using crossover, mutation, and recombination as:   r(0) + F (wg − r0 ) + F (r1 − r2 ) u < cr k j = k ij ij ij ij ij ij t wij =  r(0) uij ≥ cr & j 6= k ij (7) where k is a random index in [1, n], uij is within [0, 1], k is in {1, 2, . . . , n}, cr is crossover probability, and F ∈ [0, 2] is mutation factor. The trail vector wit is selected if   wt i wi =  w i f (wit ) < f (wi ) f (wit ) ≥ f (wi ) (8) where f (.) returns fitness of a vector as per (1). Hence, the process of crossover, mutation, recombination, and selection are repeated until an optimal parameter vector solution w∗ is found. 3.4. Ensemble: Making use of MOGP Final Population In tree construction phase, MOGP provides a population from which we can select tree models for making the ensemble. Three conflicting objectives such as approximation error, tree size, and diversity index allows the creation of Pareto-optimal solutions, where solutions are distributed on various Paretooptimal fronts according to their rank in population. Ensemble candidates can be selected from the first line of solutions (Front 1), or they can be chosen by examining the three objectives depending on the user’s need and preference. Accuracy and diversity among the ensemble candidate are important [20]. Hence, in this work, approximation error, and diversity among the candidates were given preference over tree size. Not to confuse “diversity index ” with “diversity”. The diversity index is an objective in MOGP, and the diversity is the number of distinct candidates in an ensemble. A collection M of the diverse candidate is called a bag of candidates [69]. In this work, any two trees were considered diverse (distinct) if the 12 followings hold: 1) Two trees were of different size. 2) The number of function nodes/or leaf nodes in two trees were dissimilar. 3) Two models used a different set of input features. 4) Two models used a different set of activation functions. Hence, diversity div of ensemble M (a bag of solutions) was computed as: div = distinct(M ) , size(M ) (9) where distinct(M ) is a function that returns total distinct models in an ensemble M and size(M ) is a total number of models in the bag. Now, for a classification problem, to compute combined vote of respective candidate’s outputs m1 , m2 , . . ., msize(M ) of bag M and classes ω1 , ω2 , . . . , ωC , we used an indicator function I (.) which takes 1 if ‘.’ is true, and takes 0 if ‘.’ is false. Thus, ensemble decisions by weighted majority voting is computed as [70, 71]: size(M ) X C y = arg max j=1 wt I (mt = ωj ) , (10) t=1 where wt is weight associated with the t-th candidate mt in an ensemble M and y is set to class ωj if the total weighted vote received by ωj is higher than the total vote received by any other class. Similarly, the ensemble of regression methods was computed by weighted arithmetic mean as [70]: size(M ) y= X wt mt , (11) t=1 where wt and mt are weight and output of t-th candidate in a bag M , respectively, and y is the ensemble output, which is then used for computing MSE (1) and correlation coefficient (2). The weights may be computed according to fitness of the models, or by using a metaheuristic algorithm. In this work, DE was applied to compute the ensemble weights wt , where population size was set to 100 and number of function evaluation was set to 300,000. 3.5. Multiobjective: A General Optimization Strategy A summary of general HFNT learning algorithm is as follows: Step 1. Initializing HFNT training parameters. Step 2. Apply tree construction phase to guide initial HFNT population towards Pareto-optimal solutions. Step 3. Select tree-model(s) from MOGP final population according to their approximation error, tree size, and diversity index from the Pareto front. Step 4. Apply parameter-tuning phase to optimize the selected tree-model(s). Step 5. Go to Step 2, if no satisfactory solution found. Else go to Step 6. Step 6. Using a cross-validation (CV) method to validate the chosen model(s). Step 7. Use the chosen tree-model(s) for making ensemble (recommended). Step 8. Compute ensemble results of the ensemble model (recommended). 13 Table 2: Multiobjective flexible neural tree parameter set-up for the experiments Parameter Definition Default Rang Value Scaling Input-features scaling range. [0,1] Tree height Maximum depth (layers) of a tree model. 4 Tree arity Maximum arguments of a node +ktn . [dl, du], dl ∈ R, du ∈ R  td ∈ Z+ |td > 1  tn ∈ Z+ |n ≥ 2 5 Node range Search space of functions arguments. [nl, nu], nl ∈ R, nu ∈ R [0,1] Edge range Search space for edges (weights) of tree. [wl , wu ], wl ∈ R, wu ∈ R [-1,1] P MOGP population. size(P ) > 20 30 Mutation Mutation probability pm 0.3 Crossover Crossover probability pc = 1 − pm 0.7 Mating pool Size of the pool of selected candidates. size(P )r, 0 ≤ r ≤ 1 0.5 Tournament Tournament selection size. 2 ≤ bt ≤ size(P ) 2 H DE population. 50 General ig Maximum number of trails. Structure is MOGP iterations Parameter ip DE iterations size(H) ≥ 50  ig ∈ Z+ |ig > 1  is ∈ Z+ |is ≥ 50  ip ∈ Z+ |ip ≥ 100 3 30 1000 4. Experimental Set-Up Several experiments were designed for evaluating the proposed HFNTM . A careful parameter-setting was used for testing its efficiency. A detailed description of the parameter-setting is given in Table 2, which includes: definitions, default range, and selected value. The phases of the algorithm were repeated until the stopping criteria met, i.e., either the lowest predefined approximation error was achieved, or the maximum function evaluations were reached. The repetition holds the key to obtaining a good solution. A carefully designed repetition of these two phases may offer a good solution in fewer of function evaluations. In this experiment, three general repetitions ig were used with 30 tree construction iterations is , and 1000 parameter-tuning iterations ip (Figure 1). Hence, the maximum function evaluation1 [size(P ) + ig {is (size(P ) + size(P )r) + ip size(H)}] was 154, 080. The DE version DE/rand − to − best/1/bin [22] with cr equal to 0.9 and F equal to 0.7 was used in the parameter-tuning phase. The experiments were conducted over classification, regression, and time-series datasets. A detailed description of the chosen dataset from the UCI machine learning [72] and KEEL [73] repository is available in Table A.17. The parameter-setting mentioned in Table 2 was used for the experiments over each dataset. Since the stochastic algorithms depend on random initialization, a pseudorandom number generator called, Mersenne Twister algorithm that draws random values using probability distribution in a pseudo-random manner was used for initialization of HFNTs [74]. Hence, each run of the experiment was conducted with a random seed drawn from the system. We compared HFNTM performance with various other 1 Initial GP population + three repetition ((GP population + mating pool size) × MOGP iterations + MH population × MH iterations) = 30 + 3 × [(30 + 15) × 30 + 50 × 1000] = 154, 080. 14 High Size & Diversity-index 12 5 4.5 4 3.5 3 2.5 2 1.5 1 14 Pareto-front surface tofro nt l ine 8 6 Par e to-f 4 ron t li 2 10 8 Siz e 6 4 Lo w 0.08 0.1 0.2 0.22 0.16 0.18 High 0.12 0.14 Error Low 0 0.08 Low (a) Error versus tree size versus diversity index Figure 4: Pa re 10 ne 12 h Low High Diversity Low Hig 14 0.1 0.12 0.14 Error 0.16 0.18 0.2 0.22 High (b) Error versus tree size and diversity index Pareto-front of a final population of 50 individuals generated from the training dataset of time-series problem MGS. (a) 3-D plot of solutions and a Pareto-front is a surface. (b) 2-D plot of Error versus complexity (in blue dots) and Error versus diversity (in red squares). approximation models collected from literature. A list of such models is provided in Table B.18. A developed software tool based on the proposed HFNTM algorithm for predictive modeling is available in [75]. To construct good ensemble systems, highly diverse and accurate candidates were selected in the ensemble bag M . To increase diversity (9) among the candidates, the Pareto-optimal solutions were examined by giving preference to the candidates with low approximation error, small tree size and distinct from others selected candidates. Hence, size(M ) candidates were selected from a population P . An illustration of such selection method is shown in Figure 4, which represents an MOGP final population of 50 candidate solutions computed over dataset MGS. MOGP simultaneously optimized three objectives. Hence, the solutions were arranged on the threedimensional map (Figure 4(a)), in which along the x-axis, error was plotted; along the y-axis, tree size was plotted; and along z-axis, diversity index (diversity) was plotted. However, for the simplicity, we have arranged solutions also in 2-D plots (Figure 4(b)), in which along the x-axis, computed error was plotted; and along the y-axis, tree size (indicated by blue dots) and diversity index (indicated by red squares) were plotted. From Figure 4(b), it is evident that a clear choice is difficult since decreasing approximation error increases models tree size (blue dots in Figure 4(b)). Similarly, decreasing approximation error increases models tree size and diversity (red squares in Figure 4(b)). Hence, solutions along the Paretofront (rank-1), i.e., Pareto surface indicated in the 3-D map of the solutions in Figure 4(a) were chosen for the ensemble. For all datasets, ensemble candidates were selected by examining Pareto-fronts in a similar fashion as described for the dataset MGS in Figure 4. The purpose of our experiment was to obtain sufficiently good prediction models by enhancing pre15 50 140 1 2 3 4 100 4 5 7 8 40 Tree size Tree size 180 60 20 0 1 5 30 2 6 3 7 4 8 20 10 1 250 500 750 Single objective optimization course 0 1000 (a) Single objective optimization 1 250 500 750 Multiobjective optimization course 1000 (b) Multiobjective objective optimization Figure 5: Comparison of single and multiobjective optimization course. dictability and lowering complexity. We used MOGP for optimization of HFNTs. Hence, we were compromising fitness by lowering models complexity. In single objective optimization, we only looked for models fitness. Therefore, we did not possess control over model’s complexity. Figure 5 illustrates eight runs of both single and multiobjective optimization course of HFNT, where models tree size (complexity) is indicated along y-axis and x-axis indicates fitness value of the HFNT models. The results shown in Figure 5 was conducted over MGS dataset. For each single objective GP and multiobjective GP, optimization course was noted, i.e., successive fitness reduction and tree size were noted for 1000 iterations. It is evident from Figure 5 that the HFNTM approach leads HFNT optimization by lowering model’s complexity. Whereas, in the single objective, model’s complexity was unbounded and was abruptly increased. The average tree size of eight runs of single and eight runs of multiobjective were 39.265 and 10.25, respectively; whereas, the average fitness were 0.1423 and 0.1393, respectively. However, in single objective optimization, given the fact that the tree size is unbounded, the fitness of a model may improve at the expense of model’s complexity. Hence, the experiments were set-up for multiobjective optimization that provides a balance between both objectives as described in Figure 4. 5. Results Experimental results were classified into three categories: classification, regression, and time-series. Each category has two parts: 1) First part describes the best and average results obtained from the experiments; 2) Second part describes ensemble results using tabular and graphical form. 5.1. Classification dataset We chose five classification datasets for evaluating HFNTM , and the classification accuracy was computed as: fa = tp + tn , tp + f n + f p + tn (12) where tp is the total positive samples correctly classified as positive samples, tn is the total negative samples correctly classified as negative samples, f p is the total negative samples incorrectly classified as 16 positive samples, and f n is the total positive samples incorrectly classified as negative samples. Here, for a binary class classification problem, the positive sample indicates the class labeled with ‘1’ and negative sample indicates class labeled with ‘0’. Similarly, for a three-class ( ω1 , ω2 , and ω3 ) classification problem, the samples which are labeled as a class ω1 are set to 1, 0, 0, i.e., set to positive for class ω1 and negative for ω2 , and ω3 . The samples which are labeled as a class ω2 are set to 0, 1, 0, and the samples which are labeled as a class ω3 are set to 0, 0, 1. 5.1.1. 10-Fold CV The experiments on classification dataset were conducted in three batches that produced 30 models, and each model was cross-validated using 10-fold CV, in which a dataset is equally divided into 10 sets and the training of a model was repeated 10 times. Each time a distinct set was picked for the testing the models, and the rest of nine set was picked for the training of the model. Accordingly, the obtained results are summarized in Table 3. Each batch of experiment produced an ensemble system of 10 models whose results are shown in Table 7. The obtained results presented in Table 3 describes the best and mean results of 30 models. We present a comparative study of the best 10-fold CV models results of HFNTM and the results reported in the literature in Table 4. In Table 4, the results of HDT and FNT [29] were of 10 fold CV results on the test dataset. Whereas, the results of FNT [76] was the best test accuracy and not the CV results. The results summarized in Table 4 suggests a comparatively better performance of the proposed HFNTM over the previous approaches. For the illustration of a model created by HFNTM approach, we chose the best model of dataset WDB that has a test accuracy of 97.02% (shown in Table 3). A pictorial representation of the WDB model is shown in Figure 6, where the model’s tree size is 7, total input features are 5, (x3 , x4 , x12 , x17 , and x22 ) and the selected activation function is tangent hyperbolic (k = 2) at both the non-leaf nodes. Similarly, we may represent models of all other datasets. Table 3: Best and mean results of 30 10-fold CV models (300 runs) of HFNTM Best of 30 models Mean of 30 models Data train fa test fa tree size Features train fa test fa avg. tree size diversity AUS 87.41% 87.39% 4 3 86.59% 85.73% 5.07 0.73 HRT 87.41% 87.04% 8 5 82.40% 80.28% 7.50 0.70 ION 90.92% 90.29% 5 3 87.54% 86.14% 6.70 0.83 PIM 78.67% 78.03% 10 5 71.12% 70.30% 6.33 8.67 WDB 97.02% 96.96% 6 5 94.51% 93.67% 7.97 0.73 In this work, Friedman test was conducted to examine the significance of the algorithms. For this purpose, the classification accuracy (test results) was considered (Table 4). The average ranks obtained by each method in the Friedman test is shown in Table 5. The Friedman statistic at α = 0.05 (distributed according to chi-square with 2 degrees of freedom) is 5.991, i.e., χ2(α,2) = 5.991. The obtained test value Q 17 Table 4: Comparative results: 10-fold CV test accuracy fa and variance σ of algorithms Algorithms AUS test fa HRT σ test fa ION σ test fa PIM σ test fa WDB σ test fa HDT [29] 86.96% 2.058 76.86% 2.086 89.65% 1.624 73.95% 2.374 FNT [29] 83.88% 4.083 83.82% 3.934 88.03% 0.953 77.05% 2.747 FNT [76] HFNTM 93.66% σ n/a 87.39% 0.029 87.04% 0.053 90.29% 0.044 78.03% 0.013 96.96% 0.005 according to Friedman statistic is 6. Since Q > χ2(α,2) , then the null hypothesis that “there is no difference between the algorithms” is rejected. In other words, the computed p-value by Friedman test is 0.049787 which is less than or equal to 0.05, i.e., p-value ≤ α-value. Hence, we reject the null hypothesis. Table 5 describes the significance of differences between the algorithms. To compare the differences between the best rank algorithm in Friedman test, i.e., between the proposed algorithm HFNTM and the other two algorithms, Holm’s method [77] was used. Holm’s method rejects the hypothesis of equality between the best algorithm (HFNTM ) and other algorithms if the p-value is less than α/i, where i is the position of an algorithm in a list sorted in ascending order of z-value (Table 6). From the post hoc analysis, it was observed that the proposed algorithm HFNTM outperformed both HDT [29] and FNT [29] algorithms. Table 5: Average rankings of the algorithms Algorithm Ranking HFNTM 1.0 HDT 2.5 FNT 2.5 Table 6: Post Hoc comparison between HFNTM and other algorithms for α = 0.1 i algorithm z p α/i Hypothesis 2 HDT 2.12132 0.033895 0.05 rejected 1 FNT 2.12132 0.033895 0.1 rejected 5.1.2. Ensembles The best accuracy and the average accuracy of 30 models presented in Table 3 are the evidence of HFNTM efficiency. However, as mentioned earlier, a generalized solution may be obtained by using an ensemble. All 30 models were created in three batches. Hence, three ensemble systems were obtained. The results of those ensemble systems are presented in Table 7, where ensemble results are the accuracies fa obtained by weighted majority voting (10). In Table 7, the classification accuracies fa were computed over 18 y +22 03 -0 .98 0.288 0.8 x3 1 x4 +22 -0 .60 -0.529 27 0.5 3 x17 x12 x22 Figure 6: HFNT model of classification dataset WDB (test fa = 97.02%). CV test dataset. From Table 7, it may be observed that high diversity among the ensemble candidates offered comparatively higher accuracy. Hence, an ensemble model may be adopted by examining the performance of an ensemble system, i.e., average tree size (complexity) of the candidates within the ensemble and the selected input features. An ensemble system created from a genetic evolution and adaptation is crucial for feature selection and analysis. Summarized ensemble results in Table 7 gives the following useful information about the HFNTM feature selection ability: 1) TSF - total selected features; 2) MSF - most significant (frequently selected) features; and 3) MIF - most infrequently selected features. Table 7 illustrates feature selection results. Table 7: Ensemble results (10-fold CV) of each classification dataset Data Batch AUS HRT test fa avg. D div (9) TSF 1 86.96% 5 0.7 4 2 85.51% 6 0.7 5 3 86.81% 1 77.41% 2 70.37% 3 ION PIM WDB 87.04% 4.2 0.8 5 6.8 0.5 6 7.6 0.6 9 8.1 1 10 1 82.86% 7.2 0.9 15 2 90.29% 7.3 1 16 3 86.57% 5.6 0.6 6 1 76.32% 6.9 1 8 2 64.74% 5.6 0.7 7 3 64.21% 7.4 0.9 8 1 94.29% 8.2 0.7 15 2 93.75% 5 1 15 3 94.29% 10.7 0.5 19 15 MSF MIF x6 , x8 , x10 , x1 , x2 , x3 , x12 x11 , x14 x3 , x4 , x12 , x13 x6 x15 , x16 , x18 , x2 , x4 , x5 , x27 x19 , x21 , x23 , x25 , x30 , x32 x1 , x3 , x4 , x5 , x6 , x7 x2 x21 , x22 , x24 , x1 , x5 , x6 , x8 , x25 x14 , x20 , x30 5.2. Regression dataset 5.2.1. 5-Fold CV For regression dataset, the performance of HFNTM was examined by using 5-fold CV method, in which the dataset was divided into 5 sets, each was 20% in size, and the process was repeated five times. Each time, four set was used to training and one set for testing. Hence, a total 5 runs were used for each model. As described in [78], MSE E = 0.5 × E was used for evaluating HFNTM , where E was computed as per (1). The training MSE is represented as En and test MSE is represented as Et . Such setting of MSE computation and cross-validation was taken for comparing the results collected from [78]. Table 8 presents results of 5-fold CV of each dataset for 30 models. Hence, each presented result is averaged over a total 150 runs of experiments. Similarly, in Table 9, a comparison between HFNTM and other collected algorithms from literature is shown. It is evident from comparative results that HFNTM performs very competitive to other algorithms. The literature results were averaged over 30 runs of experiments; whereas, HFNTM results were averaged of 150 runs of experiments. Hence, a competitive result of HFNTM is evidence of its efficiency. Moreover, HFNTM is distinct from the other algorithm mentioned in Table 9 because it performs feature selection and models complexity minimization, simultaneously. On the other hand, the other algorithms used entire available features. Therefore, the result’s comparisons were limited to assessing average MSE, where HFNTM , which gives simple models in comparison to others, stands firmly competitive with the others. An illustration of the best model of regression dataset DEE is provided in Figure 7, where the model offered a test MSE Et of 0.077, tree size equal to 10, and four selected input features (x1 , x3 , x4 , and x5 ). The selected activation functions were unipolar sigmoid (+72 ), bipolar sigmoid (+62 ), tangent hyperbolic (+22 ), and Gaussian (+12 ). Note that while creating HFNT models, the datasets were normalized as described in Table 2 and the output of models were denormalized accordingly. Therefore, normalized inputs should be presented to the tree (Figure 7), and the output y of the tree (Figure 7) should be denormalized. Table 8: Best and mean results of 30 5-fold CV models (150 runs) of HFNTM . Best of 30 models Mean of 30 models Data train En test Et tree size #Features train En test Et tree size diversity ABL 2.228 2.256 14 5 2.578 2.511 11.23 0.7 BAS 198250 209582 11 5 261811 288688.6 7.69 0.6 DEE 0.076 0.077 10 4 0.0807 0.086 11.7 0.7 ∗ ELV 8.33 8.36 11 7 1.35 1.35 7.63 0.5 FRD 2.342 2.425 6 5 3.218 3.293 6.98 0.34 Note: ∗ Results of ELV should be multiplied with 10-5 20 Table 9: Comparative results: 5-fold CV training MSE En and test MSE Et of algorithms. Algorithms ABL MLP BAS En Et En ELV∗ DEE Et En Et En FRD Et - 2.694 - 540302 - 0.101 - 2.04 ANFIS-SUB 2.008 2.733 119561 1089824 3087 2083 61.417 61.35 TSK-IRL 2.581 2.642 0.545 882.016 LINEAR-LMS 2.413 2.472 224684 269123 0.081 0.085 4.254 4.288 En Et 0.085 3.158 0.433 1.419 3.612 3.653 3.194 2.04 2.412 9607 461402 0.662 0.682 0.322 1.07 METSK-HDe 2.205 2.392 47900 368820 0.03 0.103 6.75 7.02 1.075 1.887 ∗∗ 2.578 2.511 261811 288688.6 0.0807 0.086 1.35 1.35 3.218 3.293 LEL-TSK HFNTM Note: ∗ ELV results should be multiplied with 10-5 , ∗∗ HFNTM results were averaged over 150 runs compared to MLP, ANFIS-SUB, TSK-IRL, LINEAR-LMS, LEL-TSK, and METSK-HDe , which were averaged over 30 runs. For regression datasets, Friedman test was conducted to examine the significance of the algorithms. For this purpose, the best test MSE was considered of the algorithms MLP, ANFIS-SUB, TSK-IRL, LINEAR-LMS, LEL-TSK, and METSK-HDe from Table 9 and the best test MSE of algorithm HFNTM was considered from Table 8. The average ranks obtained by each method in the Friedman test is shown in Table 10. The Friedman statistic at α = 0.05 (distributed according to chi-square with 5 degrees of freedom) is 11, i.e., χ2(α,5) = 11. The obtained test value Q according to Friedman statistic is 11. Since Q > χ2(α,5) , then the null hypothesis that “there is no difference between the algorithms” is rejected. In other words, the computed p-value by Friedman test is 0.05 which is less than or equal to 0.05, i.e., p-value ≤ α-value. Hence, we reject the null hypothesis. Table 10: Average rankings of the algorithms Algorithm Ranking HFNTM 1.5 e METSK-HD 2.75 LEL-TSK 3.25 LINEAR-LSM 3.5 MLP 4.5 ANFIS-SUB 5.5 From the Friedman test, it is clear that the proposed algorithm HFNTM performed best among all the other algorithms. However, in the post-hoc analysis presented in Table 11 describes the significance of difference between the algorithms. For this purpose, we apply Holm’s method [77], which rejects the hypothesis of equality between the best algorithm (HFNTM ) and other algorithms if the p-value is less than α/i, where i is the position of an algorithm in a list sorted ascending order of z-value (Table 11). In the obtained result, the equality between ANFIS-SUB, MLP and HFNTM was rejected, whereas 21 the HFNTM equality with other algorithms can not be rejected with α = 0.1, i.e., with 90% confidence. However, the p-value shown in Table 11 indicates the quality of their performance and the statistical closeness to the algorithm HFNTM . It can be observed that the algorithm METSK-HDe performed closer to algorithm HFNTM , followed by LEL-TSK, and LINEAR-LSM. Table 11: Post Hoc comparison between HFNTM and other algorithms for α = 0.1. i algorithm z p α/i Hypothesis 5 ANFIS-SUB 3.023716 0.002497 0.02 rejected 4 MLP 2.267787 0.023342 0.025 rejected 3 LINEAR-LSM 1.511858 0.13057 0.033 2 LEL-TSK 1.322876 0.185877 0.05 1 METSK-HDe 0.944911 0.344704 0.1 y +72 54 0.35 0.979 -0.2 0.444 x1 6 x5 +62 -1.0 0.7 0.5 51 37 1.0 4 x5 57 x4 -0. 0.8 +22 26 x5 0.6 09 +12 0.999 6.68e−5 x3 Figure 7: HFNT model of regression dataset DEE (test MSE Et = 0.077). 5.2.2. Ensembles For each dataset, we constructed five ensemble systems by using 10 models in each batch. In each batch, 10 models were created and cross-validated using 5 × 2-fold CV. In 5 × 2-fold CV, a dataset is randomly divided into two equal sets: A and B. Such partition of the dataset was repeated five times and each time when the set A was presented for training, the set B was presented for testing, and vice versa. Hence, total 10 runs of experiments for each model was performed. The collected ensemble results are presented in Table 12, where ensemble outputs were obtained by using weighted arithmetic mean as mentioned in (11). The weights of models were computed by using DE algorithm, where the parameter setting was similar to the one mentioned in classification dataset. Ensemble results shown in Table 12 are MSE 22 Table 12: Ensemble test MSE Et computed for 5 × 2-fold CV of 10 model in each batch Data batch MSE Et ABL BAS∗ DEE EVL∗∗ FRD Note: rt avg. D div (9) TSF MSF 1 3.004 0.65 5 0.1 3 2 2.537 0.72 3 3.042 0.65 8.3 1 7 8.5 0.5 5 4 2.294 0.75 10.7 1 7 5 1 2.412 0.73 11.2 0.7 7 2.932 0.79 5.6 0.3 5 2 3.275 0.76 8.2 0.3 6 3 3.178 0.77 5 0.2 7 4 3.051 0.78 5.7 0.3 5 5 2.707 0.81 7.3 0.7 9 4 1 0.112 0.88 4.3 0.2 2 0.115 0.88 8.9 0.6 6 3 0.108 0.88 5.4 0.5 3 4 0.123 0.87 10.8 0.9 5 5 0.111 0.88 5.2 0.6 4 1 1.126 0.71 9.3 0.1 12 2 1.265 0.67 9.6 0.1 12 3 1.124 0.71 10.4 0.1 15 4 1.097 0.72 9.2 0.2 10 5 2.047 0.31 3.8 0.4 3 1 3.987 0.86 6.2 0.2 4 2 4.154 0.83 8 0.2 4 3 4.306 0.83 5.2 0.4 5 4 3.809 0.86 7.8 0.5 4 5 2.395 0.91 7.7 0.4 5 x2 , x3 , x5 , x6 MIF x1 x3 , x7 , x8 , x1 , x2 , x5 , x9 , x11 , x13 x6 , x10 x1 , x3 , x4 , x5 , x6 x1 , x3 , x4 , x6 , x17 x1 , x2 , x4 , x5 x2 x7 , x8 , x12 x3 ∗ BAS results should be multiplied with 105 , ∗∗ ELV results should be multiplied with 10-5 . and correlation coefficient computed on CV test dataset. From ensemble results, it can be said that the ensemble with higher diversity offered better results than the ensemble with lower diversity. The models of the ensemble were examined to evaluate MSF and MIF presented in Table 12. A graphical illustration of ensemble results is shown in Figure 8 using scattered (regression) plots, where a scatter plots show how much one variable is affected by another (in this case model’s and desired outputs). Moreover, it tells the relationship between two variables, i.e., their correlation. Plots shown in Figure 8 represents the best ensemble batch (numbers indicated bold in Table 12) four, five, three, four and five where MSEs are 2.2938, 270706, 0.1085, 1.10E−05 and 2.3956, respectively. The values of r2 in plots tell about the regression curve fitting over CV test datasets. In other words, it can be said that the ensemble models were obtained with generalization ability. 5.3. Time-series dataset 5.3.1. 2-Fold CV In literature survey, it was found that efficiency of most of the FNT-based models was evaluated over time-series dataset. Mostly, Macky-Glass (MGS) dataset was used for this purpose. However, only the best-obtained results were reported. For time-series prediction problems, the performances were computed 23 20 4500 R² = 0.573 18 3500 14 prediction prediction R² = 0.6272 4000 16 12 10 8 3000 2500 2000 1500 6 4 1000 2 500 0 0 0 5 10 15 20 target 25 0 30 1000 2000 3000 4000 5000 6000 7000 target (a) Dataset ABL. rt = 0.75 (b) Dataset BAS. Et = 0.81 0.07 5 R² = 0.7682 4.5 0.06 4 0.05 prediction 3 2.5 2 1.5 R² = 0.5481 0.04 0.03 0.02 1 0.01 0.5 0 0 0 1 2 3 4 target 5 0 6 0.01 0.02 0.03 0.04 0.05 0.06 target (c) Dataset DEE. rt = 0.88 (d) Dataset EVL. rt = 0.72 30 R² = 0.8228 25 prediction prediction 3.5 20 15 10 5 0 0 5 10 15 target 20 25 30 35 (e) Dataset FRD. rt = 0.91 Figure 8: Regression plots of the best ensemble batches on datasets R1, R2, R3, R4, and R5. 24 0.07 using the root of mean squared error (RMSE), i.e., we took the square root of E given in (1). Additionally, correlation coefficient (2) was also used for evaluating algorithms performance. For the experiments, first 50% of the dataset was taken for training and the rest of 50% was used for testing. Table 13 describes the results obtained by HFNTM , where En is RMSE for training set and Et is RMSE for test-set. The best test RMSE obtained by HFNTM was Et = 0.00859 and Et = 0.06349 on datasets MGS and WWR, respectively. HFNTM results are competitive with most of the algorithms listed in Table 14. Only a few algorithms such as LNF and FWNN-M reported better results than the one obtained by HFNTM . FNT based algorithms such as FNT [1] and FBBFNT-EGP&PSO reported RMSEs close to the results obtained by HFNTM . The average RMSEs and its variance over test-set of 70 models were 0.10568 and 0.00283, and 0.097783 and 0.00015 on dataset MGS and WWR, respectively. The low variance indicates that most models were able to produce results around the average RMSE value. The results reported by other function approximation algorithms (Table 13) were merely the best RMSEs. Hence, the robustness of other reported algorithm cannot be compared with the HFNTM . However, the advantage of using HFNTM over other algorithms is evident from the fact that the average complexity of the predictive models were 8.15 and 8.05 for datasets MGA and WWR, respectively. The best model obtained for dataset WWR is shown in Figure 9, where the tree size is equal to 17 and followings are the selected activation functions: tangent hyperbolic, Gaussian, unipolar sigmoid, bipolar sigmoid and linear tangent hyperbolic. The selected input features in the tree (Figure 9) are x1 , x2 , x3 and x4 . Since in time series category experiment, we have only two datasets and for each dataset HFNTM was compared with different models from literature. Hence, the statistical test was not conducted in this category because differences between algorithms are easy to determine from Table 14. Table 13: Best and mean results 2-fold CV training RMSE En and test RMSE Et . Best of 70 models Data En Et D Features Mean of 70 models En Et D MGS 0.00859 0.00798 21 4 0.10385 0.10568 8.15 WWR 0.06437 0.06349 17 4 0.10246 0.09778 8.05 5.3.2. Ensembles The ensemble results of time-series datasets are presented in Table 15, where the best ensemble system of dataset MGS (marked bold in Table 15) offered a test RMSE Et = 0.018151 with a test correlation coefficient rt = 0.99. Similarly, the best ensemble system of dataset WWR (marked bold in Table 15) offered a test RMSE Et = 0.063286 with a test correlation coefficient rt = 0.953. However, apart from the best results, most of the ensemble produced low RMSEs, i.e., high correlation coefficients. The best ensemble batches (marked bold in Table 15) of dataset MGS and WWR were used for graphical plots in Figure 10. A one-to-one fitting of target and prediction values is the evidence of a high correlation between model’s output and desired output, which is a significant indicator of model’s efficient performance. 25 Table 14: Comparative results: training RMSE En and test RMSE Et for 2-fold CV. Algorithms MGS CPSO WWR En Et 0.0199 0.0322 PSO-BBFN - 0.027 HCMSPSO 0.0095 0.0208 HMDDE-BBFNN 0.0094 0.017 G-BBFNN - 0.013 Classical RBF 0.0096 0.0114 FNT [1] 0.0071 0.0069 FBBFNT-EGP&PSO 0.0053 0.0054 FWNN-M 0.0013 0.00114 LNF 0.0007 0.00079 - - BPNN EFuNNs HFNTM En Et - 0.200 - - 0.1063 0.0824 0.00859 0.00798 0.064377 0.063489 y +23 0.99 +62 0.99 +73 99 -1.0 0.9 -0.819 -0.9 +52 0.11364 8.95e−5 84 -0.205 -0.6 49 03 -0.8 0.1 28 x1 -0 .96 2 x2 x4 x3 x1 x4 9 22 x3 0.9999 +1 0.1501 3 0.99 0.01462 1 0.11364 +3 0.3 x3 2 x1 -0. 54 99 0.9 34 x1 -0.906 0.792 0.1 -0.513 x4 x3 Figure 9: HFNT model of time-series dataset WWR (RMSE = 0.063489). 6. Discussions HFNTM was examined over three categories of datasets: classification, regression, and time-series. The results presented in Section 5, clearly suggests a superior performance of HFNTM approach. In HFNTM approach, MOGP guided an initial HFNT population towards Pareto-optimal solutions, where HFNT final population was a mixture of heterogeneous HFNTs. Alongside, accuracy and simplicity, a Pareto-based multiobjective approach ensured diversity among the candidates in final population. Hence, HFNTs in the final population were fairly accurate, simple, and diverse. Moreover, HFNTs in the final 26 Table 15: Ensemble results computed for 50% test samples of time-series datasets Data batch MGS 1 WWR Et rt avg. tree size div (9) TSF MSF MIF 0.018 0.99 9.4 0.6 4 x1 , x3 , x4 - 2 0.045 0.98 5.8 0.2 3 3 0.026 0.99 15.2 0.5 3 4 0.109 0.92 5.1 0.4 3 5 0.156 0.89 7 0.2 3 6 0.059 0.97 8.2 0.5 3 7 0.054 0.98 6.4 0.4 4 1 0.073 0.94 5 0.1 3 x1 , x2 - 2 0.112 0.85 6 0.2 2 3 0.097 0.91 10.6 0.3 4 4 0.113 0.84 5 0.1 2 5 0.063 0.96 14.4 0.9 4 6 0.099 0.89 8.5 0.7 3 7 0.101 0.88 6.9 0.4 3 Note: Et , rt , and div indicate test RMSE, test correlation coefficient, and diversity, respectively population were diverse according to structure, parameters, activation function, and input feature. Hence, the model’s selection from Pareto-fronts, as indicated in Section 4, led to a good ensemble system. Table 16: Performance of activation functions during the best performing ensembles activation function (k) Data 1 2 3 4 5 6 7 AUS 10 - - 2 - - - HRT 10 - 9 4 - 5 3 ION 6 5 - - 2 4 4 PIM 3 8 2 5 2 1 - WDB - 3 - 7 8 10 8 ABL 2 10 - - - 10 - BAS 2 5 - - 2 10 - DEE - 6 6 4 4 10 - EVL 10 5 - 3 - - 6 FRD 10 10 - - - - - MGS 4 1 - 2 1 10 10 WWR 10 - 4 - 4 7 - Total 67 53 21 27 23 67 31 Note: 67 is the best and 21 is the worst HFNTM was applied to solve classification, regression, and time-series problems. Since HFNTM is stochastic in nature, its performance was affected by several factors: random generator algorithm, random seed, the efficiency of the meta-heuristic algorithm used in parameter-tuning phase, the activation function 27 1.4 1.2 outputs 1 0.8 0.6 0.4 Target 0.2 Prediction 1 18 35 52 69 86 103 120 137 154 171 188 205 222 239 256 273 290 307 324 341 358 375 392 409 426 443 460 477 494 0 time steps (a) Dataset MGS Et = 01815 1.2 Target 1 Prediction outputs 0.8 0.6 0.4 0.2 1 10 19 28 37 46 55 64 73 82 91 100 109 118 127 136 145 154 163 172 181 190 199 208 217 226 235 0 time steps (b) Dataset WWR Et = 0.06328 Figure 10: Target versus prediction plot obtained for time-series datasets MGS and WWR. selected at the nodes, etc. Therefore, to examine the performance of HFNTM , several HFNT-models were created using different random seeds and the best and average approximation error of all created models were examined. In Section 5, as far as the best model is concerned, the performance of HFNTM surpass other approximation models mentioned from literature. Additionally, in the case of each dataset, a very low average value (high accuracy in the case of classification and low approximation errors in case of regression and time-series) were obtained, which significantly suggests that HFNTM often led to good solutions. Similarly, in the case of the ensembles, it is clear from the result that combined output of diverse and accurate candidates offered high quality (in terms of generalization ability and accuracy) approximation/prediction model. From the results, it is clear that the final population of HFNTM offered the best ensemble when the models were carefully examined based on approximation error, average complexity (tree size), and selected features. Moreover, the performances of the best performing activation functions were examined. For this purpose, the best ensemble system obtained for each dataset were considered. Accordingly, the performance of activation functions was evaluated as follows. The best ensemble system of each dataset had 10 models; 28 therefore, in how many models (among 10) an activation function k appeared, was counted. Hence, for a dataset, if an activation function appeared in all models of an ensemble system, then the total count was 10. Subsequently, counting was performed for all the activation functions for the best ensemble systems of all the datasets. Table 16, shows the performance of the activation functions. It can be observed that the activation function Gaussian (k = 1) and Bipolar Sigmoid (k = 6) performed the best among all the other activation functions followed by Tangent-hyperbolic (k = 2) function. Hence, no one activation function performed exceptionally well. Therefore, the efforts of selecting activation function, adaptively, by MOGP was essential in HFNTs performance. In this work, we were limited to examine the performance of our approach to only benchmark problems. Therefore, in presences of no free lunch theorem [79, 80] and the algorithm’s dependencies on random number generator, which are platforms, programming language, and implementation sensitive [81], it is clear that performance of the mentioned approach is subjected to careful choice of training condition and parameter-setting when it comes to deal with other real-world problems. 7. Conclusion Effective use of the final population of the heterogeneous flexible neural trees (HFNTs) evolved using Pareto-based multiobjective genetic programming (MOGP) and the subsequent parameter tuning by differential evolution led to the formation of high-quality ensemble systems. The simultaneous optimization of accuracy, complexity, and diversity solved the problem of structural complexity that was inevitably imposed when a single objective was used. MOGP used in the tree construction phase often guided an initial HFNT population towards a population in which the candidates were highly accurate, structurally simple, and diverse. Therefore, the selected candidates helped in the formation of a good ensemble system. The result obtained by HFNTM approach supports its superior performance over the algorithms collected for the comparison. In addition, HFNTM provides adaptation in structure, computational nodes, and input feature space. Hence, HFNT is an effective algorithm for automatic feature selection, data analysis, and modeling. Acknowledgment This work was supported by the IPROCOM Marie Curie Initial Training Network, funded through the People Programme (Marie Curie Actions) of the European Unions Seventh Framework Programme FP7/20072013/, under REA grant agreement number 316555. 29 References References [1] Y. Chen, B. Yang, J. Dong, A. Abraham, Time-series forecasting using flexible neural tree model, Information Sciences 174 (3) (2005) 219–235. [2] X. Yao, Y. Liu, A new evolutionary system for evolving artificial neural networks, IEEE Transactions on Neural Networks 8 (3) (1997) 694–713. [3] I. Basheer, M. Hajmeer, Artificial neural networks: Fundamentals, computing, design, and application, Journal of Microbiological Methods 43 (1) (2000) 3–31. [4] A. J. Maren, C. T. Harston, R. M. Pap, Handbook of neural computing applications, Academic Press, 2014. [5] I. K. Sethi, A. K. Jain, Artificial neural networks and statistical pattern recognition: Old and new connections, Vol. 1, Elsevier, 2014. [6] M. Tkáč, R. Verner, Artificial neural networks in business: Two decades of research, Applied Soft Computing 38 (2016) 788–804. [7] S. E. Fahlman, C. Lebière, The cascade-correlation learning architecture, in: D. S. Touretzky (Ed.), Advances in Neural Information Processing Systems 2, Morgan Kaufmann Publishers Inc., 1990, pp. 524–532. [8] J.-P. Nadal, Study of a growth algorithm for a feedforward network, International Journal of Neural Systems 1 (1) (1989) 55–59. [9] K. O. Stanley, R. Miikkulainen, Evolving neural networks through augmenting topologies, Evolutionary Computation 10 (2) (2002) 99–127. [10] B.-T. Zhang, P. Ohm, H. Mühlenbein, Evolutionary induction of sparse neural trees, Evolutionary Computation 5 (2) (1997) 213–236. [11] M. A. Potter, K. A. De Jong, Cooperative coevolution: An architecture for evolving coadapted subcomponents, Evolutionary computation 8 (1) (2000) 1–29. [12] M. Yaghini, M. M. Khoshraftar, M. Fallahi, A hybrid algorithm for artificial neural network training, Engineering Applications of Artificial Intelligence 26 (1) (2013) 293–301. [13] S. Wang, Y. Zhang, Z. Dong, S. Du, G. Ji, J. Yan, J. Yang, Q. Wang, C. Feng, P. Phillips, Feedforward neural network optimized by hybridization of PSO and ABC for abnormal brain detection, International Journal of Imaging Systems and Technology 25 (2) (2015) 153–164. 30 [14] S. Wang, Y. Zhang, G. Ji, J. Yang, J. Wu, L. Wei, Fruit classification by wavelet-entropy and feedforward neural network trained by fitness-scaled chaotic abc and biogeography-based optimization, Entropy 17 (8) (2015) 5711–5728. [15] R. Salustowicz, J. Schmidhuber, Probabilistic incremental program evolution, Evolutionary Computation 5 (2) (1997) 123–141. [16] A. K. Kar, Bio inspired computing–a review of algorithms and scope of applications, Expert Systems with Applications 59 (2016) 20–32. [17] Y. Jin, B. Sendhoff, Pareto-based multiobjective machine learning: An overview and case studies, IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews 38 (3) (2008) 397–415. [18] K. Deb, Multi-objective optimization using evolutionary algorithms, Vol. 16, John Wiley & Sons, 2001. [19] X. Yao, Y. Liu, Making use of population information in evolutionary artificial neural networks, IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics 28 (3) (1998) 417–425. [20] L. I. Kuncheva, C. J. Whitaker, Measures of diversity in classifier ensembles and their relationship with the ensemble accuracy, Machine Learning 51 (2) (2003) 181–207. [21] K. Deb, S. Agrawal, A. Pratap, T. Meyarivan, A fast elitist non-dominated sorting genetic algorithm for multi-objective optimization: NSGA-II, in: Parallel Problem Solving from Nature PPSN VI, Vol. 1917 of Lecture Notes in Computer Science, Springer, 2000, pp. 849–858. [22] S. Das, S. S. Mullick, P. Suganthan, Recent advances in differential evolution–an updated survey, Swarm and Evolutionary Computation 27 (2016) 1–30. [23] Y. Chen, A. Abraham, J. Yang, Feature selection and intrusion detection using hybrid flexible neural tree, in: Advances in Neural Networks–ISNN, Vol. 3498 of Lecture Notes in Computer Science, Springer, 2005, pp. 439–444. [24] L. Sánchez, I. Couso, J. A. Corrales, Combining GP operators with SA search to evolve fuzzy rule based classifiers, Information Sciences 136 (1) (2001) 175–191. [25] J. Kennedy, R. C. Eberhart, Y. Shi, Swarm Intelligence, Morgan Kaufmann, 2001. [26] Y. Chen, A. Abraham, B. Yang, Feature selection and classification using flexible neural tree, Neurocomputing 70 (1) (2006) 305–313. [27] R. Riolo, J. H. Moore, M. Kotanchek, Genetic programming theory and practice XI, Springer, 2014. 31 [28] X. Chen, Y.-S. Ong, M.-H. Lim, K. C. Tan, A multi-facet survey on memetic computation, IEEE Transactions on Evolutionary Computation 15 (5) (2011) 591–607. [29] H.-J. Li, Z.-X. Wang, L.-M. Wang, S.-M. Yuan, Flexible neural tree for pattern recognition, in: Advances in Neural Networks–ISNN, Vol. 3971 of Lecture Notes in Computer Science, Springer Berlin Heidelberg, 2006, pp. 903–908. [30] Y. Chen, Y. Wang, B. Yang, Evolving hierarchical RBF neural networks for breast cancer detection, in: Neural Information Processing, Vol. 4234 of Lecture Notes in Computer Science, Springer, 2006, pp. 137–144. [31] Y. Chen, F. Chen, J. Yang, Evolving MIMO flexible neural trees for nonlinear system identification, in: International Conference on Artificial Intelligence, Vol. 1, 2007, pp. 373–377. [32] P. Wu, Y. Chen, Grammar guided genetic programming for flexible neural trees optimization, in: Advances in Knowledge Discovery and Data Mining, Springer, 2007, pp. 964–971. [33] Y. Shan, R. McKay, R. Baxter, H. Abbass, D. Essam, H. Nguyen, Grammar model-based program evolution, in: Congress on Evolutionary Computation, Vol. 1, 2004, pp. 478–485. [34] G. Jia, Y. Chen, Q. Wu, A MEP and IP based flexible neural tree model for exchange rate forecasting, in: Fourth International Conference on Natural Computation, Vol. 5, IEEE, 2008, pp. 299–303. [35] M. Oltean, C. Groşan, Evolving evolutionary algorithms using multi expression programming, in: Advances in Artificial Life, Springer, 2003, pp. 651–658. [36] P. Musilek, A. Lau, M. Reformat, L. Wyard-Scott, Immune programming, Information Sciences 176 (8) (2006) 972–1002. [37] B. Yang, L. Wang, Z. Chen, Y. Chen, R. Sun, A novel classification method using the combination of FDPS and flexible neural tree, Neurocomputing 73 (46) (2010) 690 – 699. [38] S. Bouaziz, H. Dhahri, A. M. Alimi, A. Abraham, Evolving flexible beta basis function neural tree using extended genetic programming & hybrid artificial bee colony, Applied Soft Computing. [39] Y. Chen, B. Yang, A. Abraham, Flexible neural trees ensemble for stock index modeling, Neurocomputing 70 (46) (2007) 697 – 703. [40] B. Yang, M. Jiang, Y. Chen, Q. Meng, A. Abraham, Ensemble of flexible neural tree and ordinary differential equations for small-time scale network traffic prediction, Journal of Computers 8 (12) (2013) 3039–3046. 32 [41] V. K. Ojha, A. Abraham, V. Snasel, Ensemble of heterogeneous flexible neural tree for the approximation and feature-selection of Poly (Lactic-co-glycolic Acid) micro-and nanoparticle, in: Proceedings of the Second International Afro-European Conference for Industrial Advancement AECIA 2015, Springer, 2016, pp. 155–165. [42] L. Peng, B. Yang, L. Zhang, Y. Chen, A parallel evolving algorithm for flexible neural tree, Parallel Computing 37 (10–11) (2011) 653–666. [43] L. Wang, B. Yang, Y. Chen, X. Zhao, J. Chang, H. Wang, Modeling early-age hydration kinetics of portland cement using flexible neural tree, Neural Computing and Applications 21 (5) (2012) 877–889. [44] C. Ferreira, Gene expression programming: mathematical modeling by an artificial intelligence, Vol. 21, Springer, 2006. [45] G. Weiss, Multiagent systems: A modern approach to distributed artificial intelligence, MIT Press, 1999. [46] M. Ammar, S. Bouaziz, A. M. Alimi, A. Abraham, Negotiation process for bi-objective multi-agent flexible neural tree model, in: International Joint Conference on Neural Networks (IJCNN), 2015, IEEE, 2015, pp. 1–9. [47] T. Burianek, S. Basterrech, Performance analysis of the activation neuron function in the flexible neural tree model, in: Proceedings of the Dateso 2014 Annual International Workshop on DAtabases, TExts, Specifications and Objects, 2014, pp. 35–46. [48] S. Bouaziz, H. Dhahri, A. M. Alimi, A. Abraham, A hybrid learning algorithm for evolving flexible beta basis function neural tree model, Neurocomputing 117 (2013) 107–117. [49] S. Bouaziz, A. M. Alimi, A. Abraham, Universal approximation propriety of flexible beta basis function neural tree, in: International Joint Conference on Neural Networks, IEEE, 2014, pp. 573– 580. [50] C. Micheloni, A. Rani, S. Kumar, G. L. Foresti, A balanced neural tree for pattern classification, Neural Networks 27 (2012) 81–90. [51] G. L. Foresti, C. Micheloni, Generalized neural trees for pattern classification, IEEE Transactions on Neural Networks 13 (6) (2002) 1540–1547. [52] A. Rani, G. L. Foresti, C. Micheloni, A neural tree for classification using convex objective function, Pattern Recognition Letters 68 (2015) 41–47. 33 [53] Q. Shou-ning, L. Zhao-lian, C. Guang-qiang, Z. Bing, W. Su-juan, Modeling of cement decomposing furnace production process based on flexible neural tree, in: Information Management, Innovation Management and Industrial Engineering, Vol. 3, IEEE, 2008, pp. 128–133. [54] B. Yang, Y. Chen, M. Jiang, Reverse engineering of gene regulatory networks using flexible neural tree models, Neurocomputing 99 (2013) 458–466. [55] Z. Chen, B. Yang, Y. Chen, A. Abraham, C. Grosan, L. Peng, Online hybrid traffic classifier for peer-to-peer systems based on network processors, Applied Soft Computing 9 (2) (2009) 685–694. [56] T. Novosad, J. Platos, V. Snásel, A. Abraham, Fast intrusion detection system based on flexible neural tree, in: International Conference on Information Assurance and Security, IEEE, 2010, pp. 106–111. [57] Y.-Q. Pan, Y. Liu, Y.-W. Zheng, Face recognition using kernel PCA and hybrid flexible neural tree, in: International Conference on Wavelet Analysis and Pattern Recognition, 2007. ICWAPR’07, Vol. 3, IEEE, 2007, pp. 1361–1366. [58] Y. Guo, Q. Wang, S. Huang, A. Abraham, Flexible neural trees for online hand gesture recognition using surface electromyography, Journal of Computers 7 (5) (2012) 1099–1103. [59] S. Qu, A. Fu, W. Xu, Controlling shareholders management risk warning based on flexible neural tree, Journal of Computers 6 (11) (2011) 2440–2445. [60] A. Rajini, V. K. David, Swarm optimization and flexible neural tree for microarray data classification, in: International Conference on Computational Science, Engineering and Information Technology, ACM, 2012, pp. 261–268. [61] S. Abdelwahab, V. K. Ojha, A. Abraham, Ensemble of flexible neural trees for predicting risk in grid computing environment, in: Innovations in Bio-Inspired Computing and Applications, Springer, 2016, pp. 151–161. [62] Y. Jin, B. Sendhoff, E. Körner, Evolutionary multi-objective optimization for simultaneous generation of signal-type and symbol-type representations, in: Evolutionary Multi-Criterion Optimization, Vol. 3410 of Lecture Notes in Computer Science, Springer, 2005, pp. 752–766. [63] I. Das, J. E. Dennis, A closer look at drawbacks of minimizing weighted sums of objectives for pareto set generation in multicriteria optimization problems, Structural optimization 14 (1) (1997) 63–69. [64] A. E. Eiben, J. E. Smith, Introduction to Evolutionary Computing, Springer, 2015. [65] D. Karaboga, B. Basturk, A powerful and efficient algorithm for numerical function optimization: Artificial bee colony (ABC) algorithm, Journal of Global Optimization 39 (3) (2007) 459–471. 34 [66] Y. Zhang, S. Wang, G. Ji, A comprehensive survey on particle swarm optimization algorithm and its applications, Mathematical Problems in Engineering 2015 (2015) 1–38. [67] C.-F. Juang, A hybrid of genetic algorithm and particle swarm optimization for recurrent network design, IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics) 34 (2) (2004) 997–1006. [68] W. Wongseree, N. Chaiyaratana, K. Vichittumaros, P. Winichagoon, S. Fucharoen, Thalassaemia classification by neural networks and genetic programming, Information Sciences 177 (3) (2007) 771 – 786. [69] T. Hastie, R. Tibshirani, J. Friedman, T. Hastie, J. Friedman, R. Tibshirani, The elements of statistical learning, Vol. 2, Springer, 2009. [70] R. Polikar, Ensemble based systems in decision making, IEEE Circuits and Systems Magazine 6 (3) (2006) 21–45. [71] Z.-H. Zhou, Ensemble methods: Foundations and algorithms, CRC Press, 2012. [72] M. Lichman, UCI machine learning repository, http://archive.ics.uci.edu/ml Accessed on: 01.05.2016 (2013). [73] J. Alcala-Fdez, L. Sanchez, S. Garcia, M. J. del Jesus, S. Ventura, J. Garrell, J. Otero, C. Romero, J. Bacardit, V. M. Rivas, et al., Keel: a software tool to assess evolutionary algorithms for data mining problems, Soft Computing 13 (3) (2009) 307–318. [74] M. Matsumoto, T. Nishimura, Mersenne twister: A 623-dimensionally equidistributed uniform pseudo-random number generator, ACM Transactions on Modeling and Computer Simulation 8 (1) (1998) 3–30. [75] V. K. Ojha, MOGP-FNT multiobjective flexible neural tree tool, http://dap.vsb.cz/aat/ Accessed on: 01.05.2016 (May 2016). [76] Y. Chen, A. Abraham, Y. Zhang, et al., Ensemble of flexible neural trees for breast cancer detection, The International Journal of Information Technology and Intelligent Computing 1 (1) (2006) 187–201. [77] S. Holm, A simple sequentially rejective multiple test procedure, Scandinavian Journal of Statistics (1979) 65–70. [78] M. J. Gacto, M. Galende, R. Alcalá, F. Herrera, METSK-HDe : A multiobjective evolutionary algorithm to learn accurate TSK-fuzzy systems in high-dimensional and large-scale regression problems, Information Sciences 276 (2014) 63–79. 35 [79] D. H. Wolpert, W. G. Macready, No free lunch theorems for optimization, IEEE Transactions on Evolutionary Computation 1 (1) (1997) 67–82. [80] M. Koppen, D. H. Wolpert, W. G. Macready, Remarks on a recent paper on the” no free lunch” theorems, IEEE Transactions on Evolutionary Computation 5 (3) (2001) 295–296. [81] P. L’Ecuyer, F. Panneton, Fast random number generators based on linear recurrences modulo 2: Overview and comparison, in: Proceedings of the 2005 Winter Simulation Conference, IEEE, 2005, pp. 10–pp. [82] S. Haykin, Neural networks and learning machines, Vol. 3, Pearson Education Upper Saddle River, 2009. [83] Z.-H. Zhou, Z.-Q. Chen, Hybrid decision tree, Knowledge-Based Systems 15 (8) (2002) 515–528. [84] J.-S. R. Jang, ANFIS: adaptive-network-based fuzzy inference system, IEEE Transactions on Systems, Man and Cybernetics 23 (3) (1993) 665–685. [85] O. Cordón, F. Herrera, A two-stage evolutionary process for designing TSK fuzzy rule-based systems, IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics 29 (6) (1999) 703–715. [86] J. S. Rustagi, Optimization techniques in statistics, Academic Press, 1994. [87] R. Alcalá, J. Alcalá-Fdez, J. Casillas, O. Cordón, F. Herrera, Local identification of prototypes for genetic learning of accurate tsk fuzzy rule-based systems, International Journal of Intelligent Systems 22 (9) (2007) 909–941. [88] K. B. Cho, B. H. Wang, Radial basis function based adaptive fuzzy systems and their applications to system identification and prediction, Fuzzy Sets and Systems 83 (3) (1996) 325–339. [89] F. Van den Bergh, A. P. Engelbrecht, A cooperative approach to particle swarm optimization, IEEE Transactions on Evolutionary Computation 8 (3) (2004) 225–239. [90] A. M. A. H. Dhahri, F. Karray, Designing beta basis function neural network for optimization using particle swarm optimization, in: IEEE Joint Conference on Neural Network, 2008, pp. 2564–2571. [91] C. Aouiti, A. M. Alimi, A. Maalej, A genetic designed beta basis function neural network for approximating multi-variables functions, in: International Conference on Artificial Neural Nets and Genetic Algorithms, Springer, 2001, pp. 383–386. [92] C.-F. Juang, C.-M. Hsiao, C.-H. Hsu, Hierarchical cluster-based multispecies particle-swarm optimization for fuzzy-system optimization, IEEE Transactions on Fuzzy Systems 18 (1) (2010) 14–26. 36 [93] S. Yilmaz, Y. Oysal, Fuzzy wavelet neural network models for prediction and identification of dynamical systems, IEEE Transactions on Neural Networks 21 (10) (2010) 1599–1609. [94] H. Dhahri, A. M. Alimi, A. Abraham, Hierarchical multi-dimensional differential evolution for the design of beta basis function neural network, Neurocomputing 97 (2012) 131–140. [95] A. Miranian, M. Abdollahzade, Developing a local least-squares support vector machines-based neurofuzzy model for nonlinear and chaotic time series prediction, IEEE Transactions on Neural Networks and Learning Systems 24 (2) (2013) 207–218. [96] N. K. Kasabov, Foundations of neural networks, fuzzy systems, and knowledge engineering, Marcel Alencar, 1996. [97] N. Kasabov, Evolving fuzzy neural networks for adaptive, on-line intelligent agents and systems, in: Recent Advances in Mechatronics, Springer, Berlin, 1999. [98] S. Bouaziz, A. M. Alimi, A. Abraham, Extended immune programming and opposite-based PSO for evolving flexible beta basis function neural tree, in: IEEE International Conference on Cybernetics, IEEE, 2013, pp. 13–18. 37 Appendix A. Dataset Description Table A.17: Collected datasets for testing HFNTM Index Name Features Samples Output AUS Australia 14 691 2 HRT Heart 13 270 2 ION Ionshpere 33 351 2 PIM Pima 8 768 2 WDB Wdbc 30 569 2 ABL Abalone 8 4177 1 16 337 1 6 365 1 BAS Baseball DEE DEE EVL Elevators 18 16599 1 FRD Fridman 5 1200 1 MGS Mackey-Glass 4 1000 1 WWR Waste Water 4 475 1 Type Classification Regression Time-series Appendix B. Algorithms from literature Table B.18: Algorithms from literature for the comparative study with HFNTM Ref. Algorithms Definition [82] MLP Multi-layer Perceptron [83] HDT Hybrid Decision Tree [76] FNT Flexible Neural Tree [84] ANFIS-SUB Adaptive Neuro-Fuzzy Inference System Using Subtractive Clustering [85] TSK-IRL Genetic Learning of TSK-rules Under Iterative Rule Learning [86] LINEAR-LMS Least Mean Squares Linear Regression [87] LEL-TSK Local Evolutionary Learning of TSK-rules [88] RBF Classical Radial Basis Function [89] CPSO Cooperative Particle Swarm Optimization (PSO) [90] PSO-BBFN PSO-based Beta Basis Function Neural Network [91] G-BBFNN GA-based BBFNN [92] HCMSPSO Hierarchical Cluster-Based Multispecies PSO [93] FWNN-M Fuzzy Wavelet Neural Network Models [94] HMDDE-BBFNN Hierarchical Multidimensional DE-Based BBFNN [95] LNF Local Least-Squares Support Vector Machines-Based Neuro-Fuzzy Mode [96] BPNN Back-propagation Neural Network [97] EFuNNs Evolving Fuzzy Neural Networks [98] FBBFNT-EGP&PSO Extended Immune Programming and Opposite-PSO for Flexible BBFNN [78] METSK-HDe Multiobjective Evolutionary Learning of TSK-rules for High-Dimensional Problems 38 

Feedback Control of Real-Time Display Advertising Weinan Zhang1 , Yifei Rong2,1 , Jun Wang1 , Tianchi Zhu3 , Xiaofan Wang4 1 University College London, 2 YOYI Inc., 3 Big Tree Times Co., 4 Shanghai Jiao Tong University arXiv:1603.01055v1 [cs.GT] 3 Mar 2016 1 {w.zhang, j.wang}@cs.ucl.ac.uk, 2 yifei.rong@yoyi.com.cn 3 tc@bigtree.mobi, 4 xfwang@sjtu.edu.cn ABSTRACT Real-Time Bidding (RTB) is revolutionising display advertising by facilitating per-impression auctions to buy ad impressions as they are being generated. Being able to use impression-level data, such as user cookies, encourages user behaviour targeting, and hence has significantly improved the effectiveness of ad campaigns. However, a fundamental drawback of RTB is its instability because the bid decision is made per impression and there are enormous fluctuations in campaigns’ key performance indicators (KPIs). As such, advertisers face great difficulty in controlling their campaign performance against the associated costs. In this paper, we propose a feedback control mechanism for RTB which helps advertisers dynamically adjust the bids to effectively control the KPIs, e.g., the auction winning ratio and the effective cost per click. We further formulate an optimisation framework to show that the proposed feedback control mechanism also has the ability of optimising campaign performance. By settling the effective cost per click at an optimal reference value, the number of campaign’s ad clicks can be maximised with the budget constraint. Our empirical study based on real-world data verifies the effectiveness and robustness of our RTB control system in various situations. The proposed feedback control mechanism has also been deployed on a commercial RTB platform and the online test has shown its success in generating controllable advertising performance. Keywords Feedback Control, Demand-Side Platform, Real-Time Bidding, Display Advertising 1. INTRODUCTION Emerged in 2009, Real-Time Bidding (RTB) has become a new paradigm in display advertising [22, 12]. Different from the conventional human negotiation or pre-setting a fixed price for impressions, RTB creates an impression-level auction and enables advertisers to bid for individual impression through computer algorithms served by demand-side platforms (DSPs) [33]. The bid decision could depend on the evaluation of both the utility (e.g., the likelihood and economic value of an impression for generating click or conversion) and the cost (e.g., the actual paid price) of each ad WSDM 2016, February 22-25, 2016, San Francisco, CA, USA. arXiv version. impression. More importantly, real-time information such as the specific user demographics, interest segments and various context information is leveraged to help the bidding algorithms evaluate each ad impression. With the real-time decision making mechanism, it is reported that RTB yields significantly higher return-on-investment (ROI) than other online advertising forms [31]. Despite the ability of delivering performance-driven advertising, RTB, unfortunately, results in high volatilities, measured by major Key Performance Indicators (KPIs), such as CPM (cost per mille), AWR (auction winning ratio), eCPC (effective cost per click) and CTR (click-through rate). To illustrate this, Figure 1 plots the four major KPIs over time for two example campaigns in a real-world RTB dataset. All four KPIs fluctuate heavily across the time under a widelyused bidding strategy [25]. Such instability causes advertisers ample difficulty in optimising and controlling the KPIs against their cost. In this paper, we propose to employ feedback control theory [2] to solve the instability problem in RTB. Feedback controllers are widely used in various applications for maintaining dynamically changing variables at the predefined reference values. The application scenarios range from the plane direction control [23] to the robot artificial intelligence [26]. In our RTB scenario, the specific KPI value, depending on the requirements from the advertisers, is regarded as the variable we want to control with a pre-specified reference value. Our study focuses on two use cases. (i) For performance-driven advertising, we concern with the feedback control of the average cost on acquiring a click, measured by effective cost per click (eCPC). (ii) For branding based advertising, to ensure a certain high exposure of a campaign, we focus on the control of the ratio of winning the auctions for the targeted impressions, measured by auction winning ratio (AWR). More specifically, we take each of them as the control input signal and consider the gain (the adjustment value) of bid price as the control output signal for each incoming ad display opportunity (the bid request). We develop two controllers to test: the widely used proportional-integral-derivative (PID) controller [6] and the waterlevel-based (WL) controller [10]. We conduct largescale experiments to test the feedback control performance with different settings of reference value and reference dynamics. Through the empirical study, we find that the PID and WL controllers are capable of controlling eCPC and AWR, while PID further provides a better control accuracy and robustness than WL. Furthermore, we investigate whether the proposed feedback control can be employed for controllable bid optimisation. It is common that the performance of an ad campaign (e.g., eCPC) varies from different channels (e.g., ad exchanges, user geographic regions and PC/mobile devices) 100 ● 0.4 ● ● ● ● 90 80 ● ● 70 ● ● ● ● AWR CPM ● ● ● ● ● 3476 ● ● ● ● ● ● ● ● ● ● 0 5 ● 0.1 ● 5 10 15 20 Hour ● ● ● ● 50 3358 ● ● ● ● ● 0 campaign id ● 0.2 ● 60 0.3 ● ● ● ● ● ● ● ● ● 10 ● 15 20 Hour ● ● 75 ● ● ● ● campaign id ● ● CTR eCPC 0.010 ● 50 ● ● ● ● ● ● ● ● ● ● ● ● ● ● 5 ● ● ● 10 ● ● ● ● ● 3358 3476 ● ● ● ● ● ● 15 ● ● ● ● ● ● ● 0.000 0 ● ● 0.005 25 ● ● ● 20 Hour 0 5 10 15 20 Hour Figure 1: The instability of CPM (cost per mille), AWR (auction winning ratio), eCPC (effective cost per click), and CTR (click-through rate) for two sample campaigns without a controller. Dataset: iPinYou. [34]. If one can reallocate some budget from less cost-effective channels to more cost-effective ones, the campaign-level performance would improve [35]. In this paper, we formulate the multi-channel bid optimisation problem and propose a model to calculate the optimal reference eCPC for each channel. Our experiments show that the campaign-level click number and eCPC achieve significant improvements with the same budget. Moreover, the proposed feedback control mechanism has been implemented and integrated in a commercial DSP. The conducted live test shows that in a real and noisy setting the proposed feedback mechanism has the ability to produce controllable advertising performance. To sum up, the contributions of our work are as follows. (i) We study the instability problem in RTB and investigate its solution by leveraging the feedback control mechanism. (ii) Comprehensive offline and online experiments show that PID controller is better than other alternatives and finds the optimal way to settle the variable in almost all studied cases. (iii) We further discover that feedback controllers are of great potential to perform bid optimisation through settling the eCPC at the reference value calculated by our proposed mathematical click maximisation framework. The rest of this paper is organised as follows. Section 2 provides preliminaries for RTB and feedback control. Our solution is formally presented in Section 3. The empirical study is reported in Section 4, while the online deployment and live test are given in Section 5. Section 6 discusses the related work and we finally conclude this paper in Section 7. 2. PRELIMINARIES To make the paper self-contained, in this section, we take a brief review on the RTB eco-system, bidding strategies, and some basics of feedback control theory. and each of its qualified ads, the bidding agent calculates a bid price. Then the bid response (ad, bid price) is sent back to the exchange. (3) Having received the bid responses from the advertisers, the ad exchange hosts an auction and picks the ad with the highest bid as the auction winner. (4) Then the winner is notified of the auction winning from the ad exchange. (5) Finally, the winner’s ad will be shown to the visitor along with the regular content of the publisher’s site. It is commonly known that a long time page-loading would greatly reduce users’ satisfactory [22]. Thus, advertiser bidding agents are usually required to return a bid in a very short time frame (e.g., 100 ms). (6) The user’s feedback (e.g., click and conversion) on the displayed ad is tracked and finally sent back to the winner advertiser. For a detailed discussion about RTB eco-systems, we refer to [34, 31]. The above interaction steps have the corresponding positions in Figure 2, as we will discuss later. 2.2 Bidding Strategies A basic problem for DSP bidding agents is to figure out how much to bid for an incoming bid request. The bid decision depends on two factors for each ad impression: the utility (e.g., CTR, expected revenue) and cost (i.e., expected charged price) [33]. In a widely adopted bidding strategy [25], the utility is evaluated by CTR estimation while the base bid price is tuned based on the bid landscape [9] for the cost evaluation. The generalised bidding strategy in [25] is b(t) = b0 θt , θ0 (1) where θt is the estimated CTR for the bid request at moment t; θ0 is the average CTR under a target condition (e.g., a user interest segment); and b0 is the tuned base bid price for the target condition. In this work, we adopt this widely used bidding strategy and adopt a logistic CTR estimator [27]. 2.3 Feedback Control Theory Feedback control theory deals with the reaction and control of dynamic systems from feedback and outside noise [2]. The usual objective of feedback control theory is to control a dynamic system so that the system output follows a desired control signal, called the reference, which may be a fixed or changing value. To attain this objective, a controller is designed to monitors the output and compares it with the reference. The difference between actual and desired output, called the error factor, is applied as feedback from the dynamic system to the control system. With the specific control function, the controller outputs the control signal, which is then transformed by the actuator into the system input signal sent back to the dynamic system. These processes form a feedback control loop between the dynamic system and the controller. Control techniques are widely used in various engineering applications for maintaining some signals at the predefined or changing reference values, such as plane navigation [23] and water distribution control [10]. 2.1 RTB Flow Steps 3. RTB FEEDBACK CONTROL SYSTEM The interaction process among the main components of the RTB eco-system is summarised into the following steps: (0) when a user visits an ad-supported site (e.g., web pages, streaming videos and mobile apps), each ad placement will trigger a call for ad (ad request) to the ad exchange. (1) The ad exchange sends the bid requests for this particular ad impression to each advertiser’s DSP bidding agent, along with the available information such as the user and context information. (2) With the information of the bid request Figure 2 presents the diagram of the proposed RTB feedback control system. The traditional bidding strategy is represented as the bid calculator module in the DSP bidding agent. The controller plays as a role which adjusts the bid price from the bid calculator. Specifically, the monitor receives the auction win notice from the ad exchange and the user click feedback from the ad tracking system, which as a whole we regard as the dynamic system. Then the current KPI values, such as AWR and Control Function Error factors Measured KPI value Actuator Adjusted Bid price 2. Bid Response 4. Win Notice RTB Ad Exchange 5. Ad Reference KPI 150 100 100 50 25 Page 50 0 0 1 eCPC can be calculated. If the task is to control the eCPC with the reference value, the error factor between the reference eCPC and the measured eCPC is calculated then sent into the control function. The output control signal is sent to the actuator, which uses the control signal to adjust the original bid price from the bid calculator. The adjusted bid price is packaged with the qualified ad into the bid response and sent back to the ad exchange for auction. 3.1 Actuator For the bid request at the moment t, the actuator takes into the current control signal φ(t) to adjust the bid price from b(t) (Eq. (1)) to a new value ba (t). In our model, the control signal, which will be mathematically defined in the next subsections, is a gain on the bid price. Generally, when the control signal φ(t) is zero, there should be no bid adjustment. There could be different actuator models, and in our work we choose to use ba (t) = b(t) exp{φ(t)}, (2) where the model satisfies ba (t) = b(t) when φ(t) = 0. Other models such as the linear model ba (t) ≡ b(t)(1 + φ(t)) are also investigated in our study but it performs poorly in the situations when a big negative control signal is sent to the actuator, where the linear actuator will usually respond a negative or a zero bid, which is meaningless in our scenario. By contrast, the exponential model is a suitable solution to addressing the above drawback because it naturally avoids generating a negative bid. In the later empirical study we mainly report the analysis based on the exponential-form actuator model. 3.2 PID Controller The first controller we investigate is the classic PID controller [6]. As its name implies, a PID controller produces the control signal from a linear combination of the proportional factor, the integral factor and the derivative factor based on the error factor: e(tk ) = xr − x(tk ), (3) e(tj )△tj + λD △e(tk ) , △tk 2 3 Ad Exchange 0 1 2 3 Ad Exchange 1 2 3 Ad Exchange User Figure 2: Feedback controller integrated in the RTB system. j=1 50 3. Auction 6. User Feedback φ(tk+1 ) ← λP e(tk ) + λI Campaign 3476 200 75 0. Ad Request Monitor k X Campaign 3427 150 eCPC Bid Calculator Bid Price Control Signal Campaign 1458 100 eCPC Ad Controller Dynamic System 1. Bid Request eCPC DSP Bidding Agent (4) where the error factor e(tk ) is the reference value xr minus the current controlled variable value x(tk ), the update time interval is given as △tj = tj − tj−1 , the change of error factors is △e(tk ) = e(tk ) − e(tk−1 ), and λP , λI , λD are the weight parameters for each control factor. Note that here the control factors are all in discrete time (t1 , t2 , . . .) because bidding events are discrete and it is practical to periodically update the control factors. All control factors (φ(t), e(tk ), λP , λI , λD ) remain the same between two updates. Thus for all time t between tk and tk+1 , the control signal φ(t) in Eq. (2) equals φ(tk ).We see that P factor tends to push the current variable value to the reference value; I Figure 3: Different eCPCs across different ad exchanges. Dataset: iPinYou. factor reduces the accumulative error from the beginning to the current time; D factor controls the fluctuation of the variable. 3.3 Waterlevel-based Controller The Waterlevel-based (WL) controller is another feedback control model which was originally used to switching devices controlled by water level [10]: φ(tk+1 ) ← φ(tk ) + γ(xr − x(tk )), (5) where γ is the step size parameter for φ(tk ) update in exponential scale. Compared to PID, the WL controller only takes the difference between the variable value and the reference value into consideration. Moreover, it provides a sequential control signal. That is, the next control signal is an adjustment based on the previous one. 3.4 Setting References for Click Maximisation Given that the feedback controller is an effective tool to deliver advertisers’ KPI goal, in this subsection, we demonstrate that the feedback control mechanism can be leveraged as a model-free click maximisation framework embedded with any bidding strategies [25, 33] and performs automatic budget allocation [17] across different channels via setting smart reference values. When an advertiser specifies the targeted audience (usually also combined with ad impression contextual categories) for their specific campaign, the impressions that fit the target rules may come from separate channels such as different ad exchanges, user regions, users’ PC/mobile devices etc. It is common that the DSP integrates with several ad exchanges and delivers the required ad impressions from all those ad exchanges (as long as the impressions fit the target rule), although the market prices [1] may be significantly different. Figure 3 illustrates that, for the same campaign, there is a difference in terms of eCPC across different ad exchanges. As pointed out in [34], the differences are also found in other channels such as user regions and devices. The cost differences provide advertisers a further opportunity to optimise their campaign performance based on eCPCs. To see this, suppose a DSP is integrated to two ad exchanges A and B. For a campaign in this DSP, if its eCPC from exchange A is higher than that from exchange B, which means the inventories from exchange B is more cost effective than those from exchange A, then by reallocating some budget from exchange A to B will potentially reduce the overall eCPC of this campaign. Practically the budget reallocation can be done by reducing the bids for exchange A while increasing the bids for exchange B. Here we formally propose a model of calculating the equilibrium eCPC of each ad exchange, which will be used as the optimal reference eCPC for the feedback control that leads to a maximum number of clicks given the budget constraint. Mathematically, suppose for a given ad campaign, there are n ad exchanges (could be other channels), i.e., 1, 2, . . . , n, that have the ad volume for a target rule. In our formula- Campaign 1458 Campaign 3427 900 700 800 600 Campaign 3476 400 700 Clicks Clicks Clicks exchange 500 1 300 2 3 200 400 600 100 300 500 0 5 10 15 20 0 10 20 eCPC 30 40 0 30 eCPC 60 90 120 eCPC Figure 4: Number of Clicks against eCPC. Clicks and eCPC are calculated across the whole iPinYou training dataset of each campaign by tuning b0 in Eq. (1). tion we focus on optimising clicks, while the formulation of conversions can be obtained similarly. Let ξi be the eCPC on ad exchange i, and ci (ξi ) be the click number that the campaign acquires in the campaign’s lifetime if we tune the bid price to make its eCPC be ξi for ad exchange i. For advertisers, they want to maximise the campaign-level click number given the campaign budget B [33]: max ξ1 ,...,ξn s.t. X ci (ξi ) (6) X ci (ξi )ξi = B. (7) i i Its Lagrangian is L(ξ1 , . . . , ξn , α) = X ci (ξi ) − α( X ci (ξi )ξi − B), (8) i i where α is the Lagrangian multiplier. Then we take its gradient on ξi and let it be 0: ∂L(ξ1 , . . . , ξn , α) = c′i (ξi ) − α(c′i (ξi )ξi + ci (ξi )) = 0, (9) ∂ξi c′ (ξi )ξi + ci (ξi ) ci (ξi ) 1 = i = ξi + ′ , (10) α c′i (ξi ) ci (ξi ) where the equation holds for each ad exchange i. As such, we can use α to bridge the equations for any two ad exchanges i and j: 1 ci (ξi ) cj (ξj ) = ξi + ′ = ξj + ′ . α ci (ξi ) cj (ξj ) (11) So the optimal solution condition is given as follows: 1 c1 (ξ1 ) c2 (ξ2 ) cn (ξn ) = ξ1 + ′ = ξ2 + ′ = · · · = ξn + ′ , (12) α c1 (ξ1 ) c2 (ξ2 ) cn (ξn ) X ci (ξi )ξi = B. (13) i With sufficient data instances, we find that ci (ξi ) is usually a concave and smooth function. Some examples are given in Figure 4. Based on the observation, it is reasonable to define a general polynomial form of the ci (ξi ) functions:  ξ bi i ci (ξi ) = c∗i ai ∗ , (14) ξi where ξi∗ is the campaign’s historic average eCPC on the ad inventories from ad exchange i during the training data period, and c∗i is the corresponding click number. These two factors are directly obtained from the training data. Parameters ai and bi are to be tuned to fit the training data. Substituting Eq. (14) into Eq. (12) gives ci (ξi ) 1 = ξi + ′ = ξi + α ci (ξi ) c∗ i ai bi ξ ξi∗ bi i c∗ i ai bi ξibi −1 ξ∗ bi i  1 ξi . = 1+ bi We can then rewrite Eq. (12) as    1 1 1 1 = 1+ ξ1 = 1 + ξ2 = · · · = 1 + ξn . (16) α b1 b2 bn bi . (17) Thus ξi = α(bi + 1) (15) Interestingly, from Eq. (17) we find that the equilibrium is not in the state that the eCPCs from the exchanges are the same. Instead, it is when any amount of budget reallocated among the exchanges does not make any more total clicks; for instance, in a two-exchange case, the equilibrium reaches when the increase of the clicks from one exchange equals the decrease from the other (Eq. (9)). More specifically, from Eq. (17) we observe that for ad exchange i, if its click function ci (ξi ) is quite flat, i.e., the click number increases much slowly as its eCPC increases in a certain area, then i its learned bi should be small. This means the factor bib+1 is small as well; then from Eq. (17) we can see the optimal eCPC in ad exchange i should be relatively small. Substituting Eqs. (14) and (17) into Eq. (7) gives X c∗i ai  bi bi +1  1 bi +1 = B, (18) α ξi∗ bi bi + 1 i where for simplicity, we denote for each ad exchange i, its  bi +1 c∗ a i as δi . This give us a simpler form parameter ξi∗ bii bib+1 i as: X  1 bi +1 = B. (19) δi α i There is no closed form to P solve Eq. (19) for α. However, as bi cannot be negative and i δi ( α1 )bi +1 monotonically increases against α1 , one can easily obtain the solution for α by using a numeric solution such as the stochastic gradient decent or the Newton method [5]. Finally, based on the solved α, we can find the optimal eCPC ξi for each ad exchange i using Eq. (17). In fact, these eCPCs are the reference value we want the campaign to achieve for the corresponding ad exchanges. We can use PID controllers, by setting xr in Eq. (3) as ξi for each ad exchange i, to achieve these reference eCPCs so as to achieve the maximum number of clicks on the campaign level. As a special case, if we regard the whole volume of the campaign as one channel, this method can be directly used as a general bid optimisation tool. It makes use of the campaign’s historic data to decide the optimal eCPC and then the click optimisation is performed by control the eCPC to settle at the optimal eCPC as reference. Note that this multi-channel click maximisation framework is flexible to incorporate any bidding strategies. 4. EMPIRICAL STUDY We conduct comprehensive experiments to study the proposed RTB feedback control mechanism. Our focus in this section is on offline evaluation using a publicly-available realworld dataset. To make our experiment repeatable, we have published the experiment code1 . The online deployment and test on a commercial DSP will be reported in Section 5. 4.1 Evaluation Setup Dataset. We test our system on a publicly available dataset collected from iPinYou DSP [19]. It contains the ad log data from 9 campaigns during 10 days in 2013, which consists of 64.75M bid records, 19.50M impressions, 14.79K 1 https://github.com/wnzhang/rtbcontrol clicks and 16K Chinese Yuan (CNY) expense. According to the data publisher [19], the last three-day data of each campaign is split as the test data and the rest as the training data. The dataset disk size is 35GB. More statistics and analysis of the dataset is available in [34]. The dataset is in a record-per-row format, where each row consists of three parts: (i) The features for this auction, e.g., the time, location, IP address, the URL/domain of the publisher, ad slot size, user interest segments etc. The features of each record are indexed as a 0.7M-dimension sparse binary vector which is fed into a logistic regression CTR estimator of the bidding strategy in Eq. (1); (ii) The auction winning price, which is the threshold of the bid to win this auction; (iii) The user feedback on the ad impression, i.e., click or not. Evaluation Protocol. We follow the evaluation protocol from previous studies on bid optimisation [33, 34] and an RTB contest [19] to run our experiment. Specifically, for each data record, we pass the feature information to our bidding agent. Our bidding agent generates a new bid based on the CTR prediction and other parameters in Eq. (1). We then compare the generated bid with the logged actual auction winning price. If the bid is higher than the auction winning price, we know the bidding agent has won this auction, paid the winning price, and obtained the ad impression. If from the ad impression record there is a click, then the placement has generated a positive outcome (one click) with a cost equal to the winning price. If there is no click, the placement has resulted in a negative outcome and wasted the money. The control parameters are updated every 2 hours (as one round). It is worth mentioning that historical user feedback has been widely used for evaluating information retrieval systems [29] and recommender systems [13]. All of them used historic clicks as a proxy for relevancy to train the prediction model as well as to form the ground truth. Similarly, our evaluation protocol keeps the user contexts, displayed ads (creatives etc.), bid requests, and auction environment unchanged. We intend to answer that under the same context if the advertiser were given a different or better bidding strategy or employed a feedback loop, whether they would be able to get more clicks with the budget limitation. The click would stay the same as nothing has been changed for the users. This methodology works well for evaluating bid optimisation [1, 33] and has been adopted in the display advertising industry [19]. Evaluation Measures. We adopt several commonly used measures in feedback control systems [3]. We define the error band as the ±10% interval around the reference value. If the controlled variable settles within this area, we consider that the variable is successfully controlled. The speed of convergence (to the reference value) is also important. Specifically, we evaluate the rise time to check how fast the controlled variable will get into the error band. We also use the settling time to evaluate how fast the controlled variable will be successfully restricted into the error band. However, fast convergence may bring the problem of inaccurate control. Thus, two control accuracy measures are introduced. We use the overshoot to measure the percentage of value that the controlled variable passes over the reference value. After the settling (called the steady state), we use the RMSE-SS to evaluate the root mean square error between the controlled variable value and the reference value. At last, we measure the control stability by calculating the standard deviation of the variable value after the settling, named as SD-SS. For bid optimisation performance, we use the campaign’s total achieved click number and eCPC as the prime evaluation measures. We also monitor the impression related performance such as impression number, AWR and CPM. Table 1: Overall control performance on eCPC. Cpg. 1458 2259 2261 2821 2997 3358 3386 3427 3476 Cntr PID WL PID WL PID WL PID WL PID WL PID WL PID WL PID WL PID WL Rise 1 6 7 6 3 5 17 17 3 9 1 1 1 Settling 5 36 7 23 22 17 7 13 12 5 - Overshoot 7.73 0 8.03 0 17.66 0 14.47 0 0.75 0 23.89 0 7.90 0 29.03 0 7.64 17.11 RMSE-SS 0.0325 0.0845 0.0449 0.0299 0.0242 0.0361 0.0337 0.0341 0.0396 0.0327 - SD-SS 0.0313 0.0103 0.0411 0.0294 0.0216 0.026 0.0287 0.0341 0.0332 0.031 - Table 2: Overall control performance on AWR. Cpg. 1458 2259 2261 2821 2997 3358 3386 3427 3476 Cntr PID WL PID WL PID WL PID WL PID WL PID WL PID WL PID WL PID WL Rise 4 3 4 1 1 1 6 1 1 1 2 1 4 1 2 3 2 1 Settling 10 7 6 13 4 8 3 8 8 7 8 5 6 13 6 7 Overshoot 16.86 0.00 17.08 3.91 16.39 2.02 16.44 5.77 13.68 0.00 22.08 0.13 18.85 2.95 26.63 0.24 27.15 1.49 RMSE-SS 0.0153 0.0448 0.0076 0.0833 0.0205 0.0086 0.0501 0.0151 0.0250 0.0332 0.0133 0.0300 0.0200 0.0482 0.0175 0.0308 SD-SS 0.0093 0.0231 0.0072 0.0113 0.0203 0.0086 0.0332 0.0151 0.0213 0.0211 0.0118 0.0291 0.0179 0.0257 0.0161 0.0271 Empirical Study Organisation. Our empirical study consists of five parts with the focus on controlling two KPIs: eCPC and AWR. (i) In Section 4.2, we answer whether the proposed feedback control systems are practically capable of controlling the KPIs. (ii) In Section 4.3, we study the control difficulty with different reference value settings. (iii) In Section 4.4, we focus on the PID controller and investigate its attributes on settling the target variable. (iv) In Section 4.5, we leverage the PID controllers as a bid optimisation tool and study their performance on optimising the campaign’s clicks and eCPC across multiple ad exchanges. (v) Finally, more discussions about PID parameter tuning and online update will be given in Section 4.6. 4.2 Control Capability For each campaign, we check the performance of the two controllers on two KPIs. We first tune the control parameters on the training data to minimise the settling time. Then we adopt the controllers over the test data and observe the performance. The detailed control performance on each campaign is provided in Table 1 for eCPC2 and Table 2 for AWR. Figure 5 shows the controlled KPI curves against the timesteps (i.e., round). The dashed horizontal line means the reference. We see from the results that (i) all the PID controllers can settle both KPIs within the error band (with the settling time less than 40 rounds), which indicates that the PID control is capable of settling both KPIs at the given reference value. (ii) The WL controller on eCPC does not work that well on test data, even though we could find good parameters on training data. This is due to the fact that WL controller tries to affect the average system behaviour through transient performance feedbacks while facing the 2 “-” cells mean invalid because of the failure to rise or settle. Campaign 1458 ● ● 0 ● ● ● ● ● ●● ● ● 10 ● ●● ● ● ● ● ●● ● 20 ● ●●● ● ● ●●●●● WL ● 30 ● ● 0.5 40 ● PID ● ● ● ●● ●●●●●●● ●●●●●●●●●●●●●●●●● ●●●● ●● ● 50 10 ●●●● 20 30 controller ● ● 0 ●● ● ● ● ● ● ● ● ● ● ●●●●●●● WL ● ●● ● ●●●●●●●●●●● ● ● ● 10 20 30 40 PID ● ● ●● ●● ●●● ●● ●●●●●● ●●●●●●●● ●●●●● 10 20 30 40 round 0.05 30 0.04 SD−SS RMSE−SS Settling Time ● 0.03 0.03 0.02 ● 10 ● ● 0.02 0.01 low mid high Reference low mid high Reference low mid high Reference (a) PID on eCPC 0.030 0.025 0.025 0.020 ● 0.020 SD−SS RMSE−SS Settling Time 10.0 7.5 0.015 0.015 5.0 0.010 0.010 low mid high Reference ●●●●●●●● low mid high Reference high 30 ● 0.4 ●●●● 40 ● ● ● ● ● ● 0 ●●● ● ● ●● ● 10 ●● ●● ● ●● ● 0 10 ● ● ● ● ● ● ● ● ● ●●●●●●●●●●●● 20 30 40 ● reference 0.8 ● low middle 0.6 ●●● ● 40 ● WL on AWR ● ● ● round 1.0 60 ● high ● ● ●●●● ●● ●●● 20 ● ●● ●●●●●●●●● 30 ● ●●● 40 round 0.4 0 ● ● ● ● ● ● ● 10 ● ● ● ● ● ● ● ● ● ● ● ● ●●● 20 ● ● ●●●●●●●● ● ● ● 30 40 round Figure 7: Control performance for campaign 3386 on eCPC and AWR with different reference values. ● 0.04 ● ● ●● 20 20 ● 0 Figure 5: Control performance on eCPC and AWR. 20 10 ●● round 40 eCPC AWR 40 ● ●● WL on eCPC 0.8 0.6 ●● ●●● ● ● 100 ● 80 ● ●●● low middle ● ●● round 80 120 ● ●● 0 40 ● 0.6 ● 30 Campaign 3358 1.0 ● ● round Campaign 3358 ● ●●● ● 0 ● ● ● ● round reference 0.8 ● AWR ● ● ● ● ● 0.7 0.6 ● 40 ● AWR ● 70 controller eCPC 0.8 AWR eCPC 0.9 60 50 eCPC PID on AWR 1.0 ● ● 70 160 PID on eCPC Campaign 1458 1.0 ● low mid high Reference (b) PID on AWR Figure 6: Control difficulty comparison with PID. huge dynamics of RTB. (iii) For WL on AWR, most campaigns are controllable while there are still two campaigns that fail to settle at the reference value. (iv) Compared to PID on AWR, WL always results in higher RMSE-SS and SD-SS values but lower overshoot percentage. Those control settings with a fairly short rise time usually face a higher overshoot. (v) In addition, we observe that the campaigns with higher CTR estimator AUC performance (referring [34]) normally get shorter settling time. According to above results, PID controller outperforms the WL controller in the tested RTB cases. We believe this is due to the fact that the integral factor in PID controller helps reduce the accumulative error (i.e., RMSE-SS) and the derivative factor helps reduce the variable fluctuation (i.e., SD-SS). And it is easier to settle the AWR than the eCPC. This is mainly because AWR only depends on the market price distribution while eCPC additionally involves the user feedback, i.e., CTR, where the prediction is associated with significant uncertainty. 4.3 Control Difficulty In this section, we extend our control capability experiments further by adding higher and lower reference values in comparison. Our goal is to investigate the impact of different levels of reference values on control difficulty. We follow the same scheme to train and test the controllers as Section 4.2. However, instead of showing the exact performance value, our focus here is on the performance comparison with different reference settings. The distribution of achieved settling time, RMSE-SS and SD-SS, with the setting of three reference levels, i.e., low, middle and high, are shown in the form of box plot [20] in the Figure 6(a) and 6(b) for the eCPC and AWR control with PID. We observe that the average settling time, RMSE-SS and SD-SS, are reduced as the reference values get higher. This shows that generally the control tasks with higher reference eCPC and AWR are easier to achieve because one can simply bid higher to win more and spend more. Also as the higher reference is closer to the initial performance value, the control signal does not bring serious bias or volatility, which leads to the lower RMSE-SS and SD-SS. For the page limit, the control performance with WL is not presented here. The results are similar with PID. Figure 7 gives the specific control curves of the two controllers with three reference levels on a sample campaign 3386. We find that the reference value which is farthest away from the initial value of the controlled variable brings the largest difficulty for settling, both on eCPC and AWR. This suggests that advertisers setting an ambitious control target will introduce the risk of unsettling or large volatility. The advertisers should try to find a best trade-off between the target value and the practical control performance. 4.4 PID Settling: Static vs. Dynamic References The combination of proportional, integral and derivative factors enables the PID feedback to automatically adjust the settling progress during the control lifetime with high efficiency [7]. Alternatively, one can empirically adjust the reference value in order to achieve the desired reference value. For example of eCPC control, if the campaign’s achieved eCPC is higher than the initial reference value right after exhausting the first half budget, the advertiser might want to lower the reference value in order to accelerate the downward adjustment and finally reach its initial eCPC target before running out of the budget. PID feedback controller implicitly handles such problem via its integration factor [7, 28]. In this section, we investigate with our RTB feedback control mechanism whether it is still necessary for advertisers to intentionally adjust the reference value according to the campaign’s real-time performance. Dynamic Reference Adjustment Model. To simulate the advertisers’ strategies to adaptively change the reference value of eCPC and AWR under the budget constraint, we propose a dynamic reference adjustment model to calculate the new reference xr (tk+1 ) after tk : xr (tk+1 ) = (B − s(tk ))xr x(tk ) , Bx(tk ) − s(tk )xr (20) Campaign 2259 350 ●● ● Campaign 2259 1.0 ● ● 300 ● AWR eCPC 0.9 ● 250 ● ●● ●● ● ●● ● ●● ●● ●● 10 ●● ● ● 20 30 ● 0.5 40 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 5 10 15 round Campaign 3427 Campaign 3427 20 2259 ● variable AWR 70 60 ● ● ● ● ● ● ● ● ● ● ● ● 50 0.8 ● ● ● 40 0 ● ● 5 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● original−ref ● ● ● 10 15 20 0 5 10 round Figure 8: Dynamic reference control with PID. 6e+06 0.10 0.05 ● ● ● ● 2e+06 0.08 SD−SS 20 4e+06 RMSE−SS Settling Cost Settling Time 0.04 30 0.06 0.03 ● 0.02 ● 0.01 0.04 10 0e+00 dyn st Reference 0.00 ● dyn st Reference dyn st Reference ● dyn st Reference (a) PID on eCPC 10 6e+06 6 2e+06 0.03 0.04 0.03 ● ● dyn st Reference Rise 13 15 13 10 3 3 3 7 0 6 3 15 4 Settling 26 18 13 38 14 29 30 38 35 17 10 15 38 Cpg. 3358 3386 3427 3476 AdEx 1 2 3 1 2 3 1 2 3 1 2 3 Rise 9 14 26 6 12 1 16 35 23 18 22 19 Settling 20 39 26 18 12 1 16 35 23 29 28 22 0.02 Furthermore, we directly compare the quantitative control performance between dynamic-reference controllers (dyn) with the standard static-reference ones (st) using PID. Besides the settling time, we also compare the settling cost, which is the spent budget before settling. The overall performance across all the campaigns is shown in Figure 9(a) for eCPC control and Figure 9(b) for AWR control. The results show that (i) for eCPC control, the dynamic-reference controllers do not perform better than the static-reference ones; (ii) for AWR control, the dynamic-reference controllers could reduce the settling time and cost, but the accuracy (RMSE-SS) and stability (SD-SS) is much worse than the static-reference controllers. This is because the dynamic reference itself brings volatility (see Figure 8). These results demonstrate that PID controller does perform a good enough way to settling the variable towards the pre-specified reference without the need of dynamically adjusting the reference to accelerate using our methods. Other dynamic reference models might be somewhat effective but this is not the focus of this paper. 0.02 ● 4 SD−SS 4e+06 RMSE−SS Settling Cost Settling Time 0.05 8 2821 AdEx 1 2 3 1 2 3 1 2 3 1 2 3 4 15 round 40 2261 eCPC/AWR dynamic−ref 0.6 ● ● ● ● ● Cpg. 1458 ● ● 0 1.0 ● ● ● round ● ● 80 eCPC eCPC/AWR dynamic−ref 0.7 0.6 ● ●● ● ● 0 ● 0.8 Table 3: Control performance on multi-exchanges with the reference eCPC set for click maximisation. original−ref ● ●● 200 90 ●● ● ●● variable dyn st Reference 4.5 Reference Setting for Click Maximisation 0.01 0.01 0e+00 dyn st Reference dyn st Reference (b) PID on AWR Figure 9: Dynamic v.s. static reference with PID. where xr is the initial reference value, x(tk ) is the achieved KPI (eCPC or AWR) at timestep tk , B is the campaign budget, s(tk ) is the cost so far. We can see from Eq. (20) that when xr (tk ) = xr , xr (tk+1 ) will be set the same as xr ; when xr (tk ) > xr , xr (tk+1 ) will be set lower than xr and vice versa. For readability, we leave the detailed derivation in appendix. Using Eq. (20) we calculate the new reference eCPC/AWR xr (tk+1 ) and use it to substitute xr in Eq. (3) to calculated the error factor so as to make the dynamicreference control. Results and Discussions. Figure 8 shows the PID control performance with dynamic reference calculated based on Eq. (20). The campaign performance gets stopped at the point where the budget is exhausted. From the figure, we see that for both eCPC and AWR control, the dynamic reference takes an aggressive approach and pushes the eCPC or AWR across the original reference value (dashed line). This actually simulates some advertisers’ strategy: when the performance is lower than the reference, then higher the dynamic reference to push the total performance to the initial reference more quickly. Furthermore, for AWR control, we can see the dynamic reference fluctuates seriously when the budget is to be exhausted soon. This is because when there is insufficient budget left, the reference value will be set much high or low by Eq. (20) in order to push the performance back to the initial target. Apparently this is an ineffective solution. We now study how the proposed feedback control could be used for click optimisation purpose. As we have discussed in Section 3.4, bid requests usually come from different ad exchanges where the market power and thus the CPM prices are disparate. We have shown that given a budget constraint, the number of clicks is maximised if one can control the eCPC in each ad exchange by settling it at an optimal eCPC reference for each of them, respectively. In this experiment, we build a PID feedback controller for each of its integrated ad exchanges, where their reference eCPCs are calculated via Eqs. (17) and (19). We train the PID parameters on the training data of each campaign, and then test the bidding performance on the test data. As shown in Table 3, the eCPC on all the ad exchanges for all tested campaigns get settled at the reference values3 (settling time less than 40). We denote our multi-exchange eCPC feedback control method as multiple. Besides multiple, we also test a baseline method which assigns a single optimal uniform eCPC reference across all the ad exchanges, denoted as uniform. We also use the linear bidding strategy without feedback control [25] as a baseline, denoted as none4 . The comparisons over various evaluation measures are reported in Figure 10. We observe that (i) the feedbackcontrol-enabled bidding strategies uniform and multiple significantly outperform the non-controlled bidding strategy none in terms of the number of achieved clicks and eCPC. This suggests that properly controlling eCPCs would lead to an optimal solution for maximising clicks. (ii) By reallo3 Campaign 2997 is only integrated with one ad exchange, thus not compared here. 4 Other bidding strategies [33, 17] are also investigated. Producing similar results, they are omitted here for clarity. 80 50 1500 Campaign 1458 Campaign 2997 30 ● 40 ●● 25 0.10% 20 0 none uniform multiple 0.00% none uniform multiple Multiple v.s. None Multiple v.s. Uniform 8% Click Improvement 150% 100% 50% 6% 4% 2% 0% 0% 1458 2259 2261 2821 3358 3386 3427 3476 total 1458 2259 2261 2821 3358 3386 3427 3476 total 60 1.0e+06 0.4 7.5e+05 40 AWR 0.3 CPM 5.0e+05 0.2 20 2.5e+05 0.1 0.0e+00 0 none uniform multiple 0.0 none uniform multiple none uniform multiple Figure 10: Bid optimisation performance. Campaign 1458 6 control multiple exchange ● ● 4 1 eCPC 2 ● 3 none ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● uniform ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 2 0 10 20 30 40 round Figure 11: Settlement of multi-exchange feedback control. cating the budget via setting different reference eCPCs on different ad exchanges, multiple further outperforms uniform on 7 out of 8 tested campaigns. (iii) On the impression related measures, the feedback-control-enabled bidding strategies earn more impressions than the non-controlled bidding strategy by actively lowering their bids (CPM) and thus AWR, but achieving more bid volumes. This suggests that by allocating more budget to the lower valued impressions, one could potentially generate more clicks. As a by-product, this confirms the theoretical finding reported in [33]. As a case study, Figure 11 plots the settling performance of the three methods on campaign 1458. The three dashed horizontal lines are the reference eCPCs on three ad exchanges. We see that the eCPCs on the three ad exchanges successfully settle at the reference eCPCs. At the same time, the campaign-level eCPC (multiple) settles at a lower value than uniform and none. 4.6 PID Parameter Tuning In this subsection, we share some lessons learned about PID controller parameter tuning and online update. Parameter Search. Empirically, λD does not change the control performance significantly. Just a small valued λD , e.g., 1 × 10−5 , will reduce the overshoot and slightly shorten the settling time. Thus the parameter search is focused on ● ● ● ● ● ● ● ● ●● ● ●● ●● update ● ● ● ●●● ●●●●●● ●● offline ●● online ● ● ●● ● ● ● ●●●●●●●●●● ● ●● ● ● ● ● ● ●● ● ● ● ● ●●●●● ●● ● 10 10 20 30 40 round uniform multiple ● 20 15 ● 0 none ● ●● ● 10 0 Click Improvement ● 0.05% 500 Impressions 40 eCPC 20 eCPC 30 CTR eCPC Clicks 60 1000 ● 0.15% 0 10 20 30 40 round Figure 12: Control with online/offline parameter updating. λP and λI . Instead of using the computationally expensive grid search, we perform an adaptive coordinate search. For every update, we fix one parameter and shoot another one to seek for the optimal value leading shortest settling time, and the line searching step length shrinks exponentially for each shooting. Normally after 3 or 4 iterations, the local optima is reached and we find such solution is highly comparable with the expensive grid search. Setting φ(t) Bounds. We also find that setting up upper/lower bounds of control signal φ(t) is important to make KPIs controllable. Due to the dynamics in RTB, it is common that user CTR drops during a period, which makes eCPC much higher. The corresponding feedback would probably result in a large negative gain on the bids, leading extremely low bid price and thus no win, no click and no additional cost at all for remaining rounds. In such case, a proper lower bound (-2) of φ(t) aims to eliminate above extreme effects by preventing from a seriously negative control signal. In addition, an upper bound (5) is used in order to avoid excessive variable growth beyond the reference value. Online Parameter Updating. As the DSP running with feedback control, the collected data can be immediately utilised for training a new PID controller and updating the older one. We investigate the possibility of the online updating of PID parameters with the recent data. Specifically, after initialising the PID parameters using training data, we re-train the controller for every 10 rounds (i.e., before round 10, 20 and 30) in the test stage using all previous data with the same parameter searching method as in the training stage. The parameter searching in re-training takes about 10 minutes for each controller, which is far shorter than the round period (2 hours). Figure 12 shows the control performance with PID parameters tuned online and offline respectively. As we can see after the 10th round (i.e., the first online tuning point), the online-tuned PIDs manage to control the eCPC around the reference value more effectively than the offline-tuned one, resulting shorter settling time and lower overshoot. In addition, no obvious disturbance or instability occurs when we switch parameters. With the online parameter updating, we can start to train the controllers based on several-hour training data and adaptively update the parameters from the new data to improve the control performance. 5. ONLINE DEPLOYMENT AND TEST The proposed RTB feedback control system has been deployed and tested in live on BigTree DSP5 , a performancedriven mobile advertising DSP in China. BigTree DSP focuses on the programmatic buying for optimal advertising performance on mobile devices, which makes it an ideal place to test our proposed solution. The deployment environment is based on Aliyun elastic cloud computing servers. A three-node cluster is deployed for the DSP bidding agent, where each node is in Ubuntu 12.04, with 8 core Intel Xeon CPU E5-2630 (2.30GHz) and 5 http://www.bigtree.mobi/ Online eCPC control 90 ding strategy. In sum, the successful eCPC control on an online commercial DSP verifies the effectiveness of our proposed feedback control RTB system. 2500 80 2000 70 60 clicks eCPC 1500 50 40 variable reference eCPC clicks 1000 30 20 500 10 0 0 0 12 24 36 48 60 72 84 Hour Figure 13: The online eCPC control performance and the accumulative click numbers of a mobile game campaign on BigTree DSP. Bids Impressions Clicks CTR 1.000 1.000 1.000 1.000 0.975 0.975 0.975 0.975 0.950 0.950 0.950 0.950 0.925 0.925 0.925 0.925 0.900 0.900 0.900 none PID none PID 0.900 none PID none PID Figure 14: Relative performance for online test. 8GB RAM. The controller module is implemented in Python with uWSGI and Nginx. For BigTree DSP controller module, we deploy the PID control function and tune its parameters. Specifically, we use the last 6-week bidding log data in 2014 as the training data for tuning PID parameters. A three-fold validation process is performed to evaluate the generalisation of the PID control performance, where the previous week data is used as the training data while the later week data is used for validation. The control factors (φ(t), e(tk ) in Eq. (4)) are updated for every 90 minutes. After acquiring a set of robust and effective PID parameters, we launch the controller module, including the monitor and actuator submodules, on BigTree DSP. Figure 13 shows the online eCPC control performance on one of the iOS mobile game campaigns during 84 hours from 7 Jan. 2015 to 10 Jan. 2015. The reference eCPC is set as 28 RMB cent by the advertiser, which is about 0.8 times the average eCPC value of the previous week where there was no control. Following the same training process described in the previous section, we update the online control factors for every 90 minutes. From the result we can see the eCPC value dropped from the beginning 79 cent to 30 during the first day and then settled closed to the reference afterwards. In the meantime, A/B testing is used to compare with the non-controlled bidding agent (with the same sampling rate but disjoint bid requests). Figure 14 shows the corresponding advertising performance comparison between a non-controlled bidding agent and the PID-control bidding agent during the test period with the same budget. As we can see, by settling the eCPC value around the lower reference eCPC, the PID-control bidding agent acquires more bid volume and win more (higher-CTR) impressions and clicks, which demonstrates its ability of optimising the performance. Compared with the offline empirical study, the online running is more challenging: (i) all pipeline steps including the update of the CTR estimator, the KPI monitor linked to the database and the PID controller should operate smoothly against the market turbulence; (ii) the real market competition is highly dynamic during the new year period when we launched our test; (iii) other competitors might tune their bidding strategies independently or according to any changes of their performance after we employed the controlled bid- 6. RELATED WORK Enabling the impression-level evaluation and bidding, much research work has been done on RTB display advertising, including bidding strategy optimisation [25, 33], reserve price optimisation [30], ad exchange auction design [4], and ad tracking [11]. In order to perform the optimal bidding, the DSP bidding agent should estimate both utility and cost of a given ad impression. The impression-level utility evaluation, including CTR and conversion rate (CVR) estimation, is the essential part for each bidding agent in DSPs. In [18] the sparsity problem of CVR estimation is handled by modelling the conversions at different hierarchical levels. The user click behaviour on mobile RTB ads is studied in [24]. On the cost evaluation side, bid landscape modelling and forecasting is much important to inform the bidding agent about the competitiveness of the market. The authors in [9] break down the campaign-level bid landscape forecasting problem into “samples” by targeting rules and then employ a mixture model of log-normal distributions to build the campaign-level bid landscape. The authors in [16] try to reduce the bid landscape forecasting error through frequently re-building the landscape models. Based on the utility and cost evaluation of the ad inventory, bid optimisation is performed to improve the advertising performance under the campaign budget constraint. Given the estimated CTR/CVR, the authors in [25, 18] employ linear bidding functions based on truth-telling attributes of second price auctions. However, given the budget constraint, the advertisers’ bidding behaviour is not truth-telling. The authors in [33] propose a general bid optimisation framework to maximise the desired advertising KPI (e.g., total click number) under the budget constraint. Besides the general bid optimisation, the explicit bidding rules such as frequency and recency capping are studied in [31]. Moreover, the budget pacing [17] which refers to smoothly delivering the campaign budget is another important problem for DSPs. There are a few research papers on recommender systems leveraging feedback controllers for performance maintenance and improvement. In [21], a rating updating algorithm based on the PID controller is developed to exclude unfair ratings in order to build a robust reputation system. The authors in [32] apply a self-monitoring and self-adaptive approach to perform a dynamic update of the training data fed into the recommender system to automatically balance the computational cost and the prediction accuracy. Furthermore, the authors in [14] adopt the more effective and well-studied PID controller to the data-feeding scheme of recommender systems, which is proved to be practically effective in their studied training task. Compared to the work of controlling the recommender system performance by changing the number of training cases, our control task in RTB is more challenging, with much various dynamics from advertising environment such as the fluctuation of market price, auction volume and user behaviour. In [8], the authors discuss multiple aspects in a performancedriven RTB system, where the impression volume control is one of discussed aspects. Specifically, WL and a modelbased controller are implemented to control the impression volume during each time interval. In [15], feedback control is used to perform budget pacing in order to stablise the conversion volume. Compared to [8, 15], our work is a more comprehensive study focused on the feedback control techniques to address the instability problem in RTB. Besides WL, we intensively investigate the PID controller, which takes more factors into consideration than WL. For the controlled KPIs, we look into the control tasks on both eCPC and AWR, which are crucial KPIs for performance-driven campaigns and branding-based campaigns, respectively. In addition, we proposed an effective model to calculate the optimal eCPC reference to maximise the campaign’s clicks using feedback controllers. 7. CONCLUSIONS In this paper, we have proposed a feedback control mechanism for RTB display advertising, with the aim of improving its robustness of achieving the advertiser’s KPI goal. We mainly studied PID and WL controllers for controlling the eCPC and AWR KPIs. Through our comprehensive empirical study, we have the following discoveries. (i) Despite of the high dynamics in RTB, the KPI variables are controllable using our feedback control mechanism. (ii) Different reference values bring different control difficulties, which are reflected in the control speed, accuracy and stability. (iii) PID controller naturally finds its best way to settle the variable, and there is no necessity to adjust the reference value for accelerating the PID settling. (iv) By settling the eCPCs to the optimised reference values, the feedback controller is capable of making bid optimisation. Deployed on a commercial DSP, the online test demonstrates the effectiveness of the feedback control mechanism in generating controllable advertising performance. In the future work, we will further study the applications based on feedback controllers in RTB, such as budget pacing and retargeting frequency capping. 8. REFERENCES [1] K. Amin, M. Kearns, P. Key, and A. Schwaighofer. Budget optimization for sponsored search: Censored learning in mdps. UAI, 2012. [2] K. J. Åström and P. Kumar. Control: A perspective. Automatica, 50(1):3–43, 2014. [3] K. J. Åström and R. M. Murray. Feedback systems: an introduction for scientists and engineers. Princeton university press, 2010. [4] S. Balseiro, O. Besbes, and G. Y. Weintraub. Repeated auctions with budgets in ad exchanges: Approximations and design. Columbia Business School Research Paper, (12/55), 2014. [5] R. Battiti. First-and second-order methods for learning: between steepest descent and newton’s method. Neural computation, 4(2):141–166, 1992. [6] S. Bennett. Development of the pid controller. Control Systems, 13(6):58–62, 1993. [7] S. P. Bhattacharyya, H. Chapellat, and L. H. Keel. Robust control. The Parametric Approach, by Prentice Hall PTR, 1995. [8] Y. Chen, P. Berkhin, B. Anderson, and N. R. Devanur. Real-time bidding algorithms for performance-based display ad allocation. KDD, 2011. [9] Y. Cui, R. Zhang, W. Li, and J. Mao. Bid landscape forecasting in online ad exchange marketplace. In KDD, pages 265–273. ACM, 2011. [10] B. W. Dezotell. Water level controller, June 9 1936. US Patent 2,043,530. [11] R. Gomer, E. M. Rodrigues, N. Milic-Frayling, and M. Schraefel. Network analysis of third party tracking: User exposure to tracking cookies through search. In WI, 2013. [12] Google. The arrival of real-time bidding. Technical report, 2011. [13] Y. Hu, Y. Koren, and C. Volinsky. Collaborative filtering for implicit feedback datasets. In ICDM, 2008. [14] T. Jambor, J. Wang, and N. Lathia. Using control theory for stable and efficient recommender systems. In WWW, 2012. [15] N. Karlsson and J. Zhang. Applications of feedback control in online advertising. In American Control Conference, 2013. [16] K. J. Lang, B. Moseley, and S. Vassilvitskii. Handling forecast errors while bidding for display advertising. In WWW, 2012. [17] K.-C. Lee, A. Jalali, and A. Dasdan. Real time bid optimization with smooth budget delivery in online advertising. In ADKDD, 2013. [18] K.-C. Lee, B. Orten, A. Dasdan, and W. Li. Estimating conversion rate in display advertising from past performance data. In KDD, 2012. [19] H. Liao, L. Peng, Z. Liu, and X. Shen. ipinyou global rtb bidding algorithm competition dataset. In ADKDD, 2014. [20] R. McGill, J. W. Tukey, and W. A. Larsen. Variations of box plots. The American Statistician, 32(1):12–16, 1978. [21] K. Meng, Y. Wang, X. Zhang, X.-c. Xiao, et al. Control theory based rating recommendation for reputation systems. In ICNSC, 2006. [22] S. Muthukrishnan. Ad exchanges: Research issues. In Internet and network economics, pages 1–12. Springer, 2009. [23] R. C. Nelson. Flight stability and automatic control, volume 2. WCB/McGraw Hill, 1998. [24] R. J. Oentaryo, E. P. Lim, D. J. W. Low, D. Lo, and M. Finegold. Predicting response in mobile advertising with hierarchical importance-aware factorization machine. In WSDM, 2014. [25] C. Perlich, B. Dalessandro, R. Hook, O. Stitelman, T. Raeder, and F. Provost. Bid optimizing and inventory scoring in targeted online advertising. In KDD, pages 804–812, 2012. [26] G. Peterson and D. J. Cook. Decision-theoretic layered robotic control architecture. In AAAI/IAAI, page 976, 1999. [27] M. Richardson, E. Dominowska, and R. Ragno. Predicting clicks: estimating the click-through rate for new ads. In WWW, pages 521–530. ACM, 2007. [28] D. E. Rivera, M. Morari, and S. Skogestad. Internal model control: Pid controller design. Industrial & engineering chemistry process design and development, 25(1), 1986. [29] J. Xu, C. Chen, G. Xu, H. Li, and E. R. T. Abib. Improving quality of training data for learning to rank using click-through data. In WSDM, 2010. [30] S. Yuan, B. Chen, J. Wang, P. Mason, and S. Seljan. An Empirical Study of Reserve Price Optimisation in Real-Time Bidding. In KDD, 2014. [31] S. Yuan, J. Wang, and X. Zhao. Real-time bidding for online advertising: measurement and analysis. In ADKDD, 2013. [32] V. Zanardi and L. Capra. Dynamic updating of online recommender systems via feed-forward controllers. In SEAMS, 2011. [33] W. Zhang, S. Yuan, and J. Wang. Optimal real-time bidding for display advertising. In KDD, 2014. [34] W. Zhang, S. Yuan, J. Wang, and X. Shen. Real-time bidding benchmarking with ipinyou dataset. arXiv, 2014. [35] W. Zhang, Y. Zhang, B. Gao, Y. Yu, X. Yuan, and T.-Y. Liu. Joint optimization of bid and budget allocation in sponsored search. In KDD, 2012. APPENDIX Reference Adjust Models. Here we provide the detailed derivation of the proposed dynamic-reference model Eq. (20) in Section 4.4. We mainly introduce the derivation of reference eCPC adjustment, while the derivation of reference AWR adjustment can be obtained similarly. Let ξr be the initial eCPC target, ξ(tk ) be the achieved eCPC before the moment tk , s(tk ) be the total cost so far, and B be the campaign budget. In such setting, the current achieved click number is s(tk )/ξ(tk ) and the target click number is B/ξr . In order to push the overall eCPC to ξr , i.e., push the total click number B/ξr with the budget B, the reference eCPC for the remaining time ξr (tk+1 ) should satisfy B − s(tk ) B s(tk ) + = . ξ(tk ) ξr (tk+1 ) ξr (21) Solving the equation we have ξr (tk+1 ) = (B − s(tk ))ξr ξ(tk ) . Bξ(tk ) − s(tk )ξr (22) The derivation of reference AWR adjustment is much similar but with an extra winning function which links between the bid price and winning probability [33]. The result formula is just the same as Eq. (22). Using x as a general notation for eCPC and AWR variables results in Eq. (20) in Section 4.4. 

Dual Recurrent Attention Units for Visual Question Answering Ahmed Osman 1,2 , Wojciech Samek 1 1 Fraunhofer Heinrich Hertz Institute, Berlin, Germany 2 University of Freiburg, Freiburg, Germany arXiv:1802.00209v1 [cs.AI] 1 Feb 2018 {ahmed.osman, wojciech.samek}@hhi.fraunhofer.de Abstract We propose an architecture for VQA which utilizes recurrent layers to generate visual and textual attention. The memory characteristic of the proposed recurrent attention units offers a rich joint embedding of visual and textual features and enables the model to reason relations between several parts of the image and question. Our single model outperforms the first place winner on the VQA 1.0 dataset, performs within margin to the current state-of-the-art ensemble model. We also experiment with replacing attention mechanisms in other state-of-the-art models with our implementation and show increased accuracy. In both cases, our recurrent attention mechanism improves performance in tasks requiring sequential or relational reasoning on the VQA dataset. Figure 1. Diagram of the DRAU network. tial manner which recurrent layers are better suited due to their ability to capture relevant information over an input sequence. In this paper, we propose a RNN-based joint representation to generate visual and textual attention. We argue that embedding a RNN in the joint representation helps the model process information in a sequential manner and determine what is relevant to solve the task. We refer to the combination of RNN embedding and attention as Recurrent Textual Attention Unit (RTAU) and Recurrent Visual Attention Unit (RVAU) respective of their purpose. Furthermore, we employ these units in a fairly simple network, referred to as Dual Recurrent Attention Units (DRAU) network, and show improved results over several baselines. Finally, we enhance state-of-the-art models by replacing the model’s default attention mechanism with RVAU. Our main contributions are the following: 1. Introduction Although convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been successfully applied to various image and natural language processing tasks (cf. [1, 2, 3, 4]), these breakthroughs only slowly translate to multimodal tasks such as visual question answering (VQA) where the model needs to create a joint understanding of the image and question. Such multimodal tasks require joint visual and textual representations. Since global features can hardly answer questions about certain local parts of the input, attention mechanisms have been extensively used in VQA recently [5, 6, 7, 8, 9, 10, 11, 12]. It attempts to make the model predict based on spatial or lingual context. However, most attention mechanisms used in VQA models are rather simple, consisting of two convolutional layers followed by a softmax to generate the attention weights which are summed over the image features. These shallow attention mechanisms may fail to select the relevant information from the joint representation of the question and image. Creating attention for complex questions, particularly sequential or relational reasoning questions, requires processing information in a sequen- • We introduce a novel approach to generate soft attention. To the best of our knowledge, this is the first attempt to generate attention maps using recurrent neural networks. We provide quantitative and qualitative results showing performance improvements over the de1 fault attention used in most VQA models. • Our attention modules are modular, thus, they can substitute existing attention mechanisms in most models fairly easily. We show that state-of-the-art models with RVAU “plugged-in” perform consistently better than their vanilla counterparts. 2. Related Work This section discusses common methods that have been explored in the past for VQA. Bilinear representations Fukui et al. [7] use compact bilinear pooling to attend over the image features and combine it with the language representation. The basic concept behind compact bilinear pooling is approximating the outer product by randomly projecting the embeddings to a higher dimensional space using Count Sketch projection [13] and then exploiting Fast Fourier Transforms to compute an efficient convolution. An ensemble model using MCB won first place in VQA (1.0) 2016 challenge. Kim et al. [5] argues that compact bilinear pooling is still expensive to compute and shows that it can be replaced by element-wise product (Hadamard product) and a linear mapping (i.e. fullyconnected layer) which gives a lower dimensional representation and also improves the model accuracy. Recently, Ben-younes et al. [14] proposed using Tucker decomposition [15] with a low-rank matrix constraint as a bilinear representation. They propose this fusion scheme in an architecture they refer to as MUTAN which as of this writing is the current state-of-the-art on the VQA 1.0 dataset. Attention-based Closely related to our work, Lu et al. [9] were the first to feature a co-attention mechanism that applies attention to both the question and image. Nam et al. [6] use a Dual Attention Network (DAN) that employs attention on both text and visual features iteratively to predict the result. The goal behind this is to allow the image and question attentions to iteratively guide each other in a synergistic manner. RNNs for VQA Using recurrent neural networks (RNNs) for VQA has been explored in the past. Xiong et al. [16] build upon the dynamic memory network from Kumar and Varaiya [17] and proposes DMN+. DMN+ uses episodic modules which contain attention-based Gated Recurrent Units (GRUs). Note that this is not the same as what we propose; Xiong et al. generate soft attention using convolutional layers and then uses it to substitute the update gate of the GRU. In contrast, our approach uses the recurrent layers to generate the attention. Noh and Han [8] propose recurrent answering units in which each unit is a complete module that can answer a question about an image. They use joint loss minimization to train the units. However during testing, they use the first answering unit which was trained from other units through backpropagation. Notable mentions Kazemi and Elqursh [18] show that a simple model can get state-of-the-art results with proper training parameters. Wu et al. [19] construct a textual representation of the semantic content of an image and merges it with textual information sourced from a knowledge base. Ray et al. [20] introduce a task of identifying relevant questions for VQA. Kim et al. [21] apply residual learning techniques to VQA and propose a novel attention image attention visualization method using backpropagation. 3. Dual Recurrent Attention in VQA We propose our method in this section. Figure 1 illustrates the flow of information in the DRAU model. Given an image and question, we create the input representations v and q. Next, these features are combined by 1 × 1 convolutions into two separate branches. Then, the branches are passed to an RTAU and RVAU. Finally, the branches are combined using a fusion operation and fed to the final classifier. The full architecture of the network is depicted in Figure 2. 3.1. Input Representation Image representation We use the 152-layer “ResNet” pretrained CNN from He et al. [1] to extract image features. Similar to [7, 6], we resize the images to 448 × 448 and extract the last layer before the final pooling layer (res5c) with size 2048 × 14 × 14. Finally, we use l2 normalization on all dimensions. Recently, Anderson et al. [22] have shown that object-level features can provide a significant performance uplift compared to global-level features from pretrained CNNs. Therefore, we experiment with replacing the ResNet features with FRCNN [23] features with a fixed number of proposals per image (K = 36). Question representation We use a fairly similar representation as [7]. In short, the question is tokenized and encoded using an embedding layer followed by a tanh activation. We also exploit pretrained GloVe vectors [24] and concatenate them with the output of the embedding layer. The concatenated vector is fed to a two-layer unidirectional LSTM that contains 1024 hidden states each. In contrast to Fukui et al., we use all the hidden states of both LSTMs rather than concatenating the final states to represent the final question representation. 3.2. 1 × 1 Convolution and PReLU We apply multiple 1 × 1 convolution layers in the network for mainly two reasons. First, they learn weights from Figure 2. The proposed network. ⊕ denotes concatenation. the image and question representations in the early layers. This is important especially for the image representation, since it was originally trained for a different task. Second, they are used to generate a common representation size. To obtain a joint representation, we apply 1 × 1 convolutions followed by PReLU activations [1] on both the image and question representations. Through empirical evidence, PReLU activations were found to reduce training time significantly and improve performance compared to ReLU and tanh activations. We provide these results in Section 4. 3.3. Recurrent Attention Units The result from the above-mentioned layers is concatenated and fed to two separate recurrent attention units (RAU). Each RAU starts with another 1 × 1 convolution and PReLU activation:
ca = PReLU Wa x (1) where Wa is the 1 × 1 convolution weights, x is the input to the RAU, and ca is the output of the first PReLU. Furthermore, we feed the previous output into an unidirectional LSTM:
ha,n = LSTM ca,n (2) Next, we use the attention weights to compute a weighted average of the image and question features. atta,n = N X Watt,n fn (4) n=1 where fn is the input representation and atta,n is the attention applied on the input. Finally, the attention maps are fed into a fully-connected layer followed by a PReLU activation. Figure 3 illustrates the structure of a RAU.
yatt,n = PReLU Wout atta,n (5) where Wout is a weight vector of the fully connected layer and yatt,n is the output of each RAU. 3.4. Reasoning layer A fusion operation is used to merge the textual and visual branches. For DRAU, we experiment with using elementwise multiplication (Hadamard product) and MCB [7, 25]. The result of the fusion is given to a many-class classifier using the top 3000 frequent answers. We use a single-layer softmax with cross-entropy loss. This can be written as:  
Pa = softmax fusion op ytext , yvis Wans (6) where ha,n is the hidden state at time n. To generate the attention weights, we feed all the hidden states of the previous LSTM to a 1 × 1 convolution layer followed by a softmax function. The 1×1 convolution layer could be interpreted as the number of glimpses the model sees. 
 Watt,n = softmax PReLU Wg ha,n (3) where ytext and yvis are the outputs of the RAUs, Wans represents the weights of the multi-way classifier, and Pa is the probability of the top 3000 frequent answers. The final answer â is chosen according to the following: where Wg is the glimpses’ weights and Watt,n is the attention weight vector. Experiments are performed on the VQA 1.0 and 2.0 datasets [26, 27]. These datasets use images from the â = argmax Pa (7) 4. Experiments and Results VQA 1.0 Validation Split Open Ended Task Baselines Y/N Num. Other Language only Simple MCB Joint LSTM Figure 3. Recurrent Attention Unit. MS-COCO dataset [28] and generate questions and labels (10 labels per question) using Amazon’s Mechanical Turk (AMT). Compared to VQA 1.0, VQA 2.0 adds more imagequestion pairs to balance the language prior present in the VQA 1.0 dataset. The ground truth answers in the VQA dataset are evaluated using human consensus.  P a is in human annotation ,1 (8) Acc(a) = min 3 We evaluate our results on the validation, test-dev, teststd splits of each dataset. Models evaluated on the validation set use train and Visual Genome for training. For the other splits, we include the validation set in the training data. However, the models using FRCNN features do not use data augmentation with Visual Genome. To train our model, we use Adam [29] for optimization with β1 = 0.9, β2 = 0.999, and an initial learning rate of  = 7 × 10−4 . The final model is trained with a small batch size of 32 for 400K iterations. We did not fully explore tuning the batch size which explains the relatively high number of training iterations. Dropout (p = 0.3) is applied after each LSTM and after the fusion operation. All weights are initialized as described in [30] except LSTM layers which use an uniform weight distribution. The pretrained ResNet was fixed during training due to the massive computational overhead of fine-tuning the network for the VQA task. While VQA datasets provide 10 answers per image-question pair, we sample one answer randomly for each training iteration. 4.1. VQA 1.0 Experiments During early experiments, the VQA 2.0 dataset was not yet released. Thus, the baselines and early models were evaluated on the VQA 1.0 dataset. While building the final model, several parameters were changed, mainly, the learning rate, activation functions, dropout value, and other modifications which we discuss in this section. Baselines We started by designing three baseline architectures. The first baseline produced predictions solely from the question while totally ignoring the image. The 78.56 78.64 79.90 27.98 32.98 36.96 30.76 39.79 49.58 All 48.3 54.82 59.34 Table 1. Evaluation of the baseline models on the VQA 1.0 validation split. model used the same question representation described in [7] and passed the output to a softmax 3000-way classification layer. The goal of this architecture was to assess the extent of the language bias present in VQA. The second baseline is a simple joint representation of the image features and the language representation. The representations were combined using the compact bilinear pooling from [25]. We chose this method specifically because it was shown to be effective by Fukui et al. [7]. The main objective of this model is to measure how a robust pooling method of multimodal features would perform on its own without a deep architecture or attention. We refer to this model as Simple MCB. For the last baseline, we substituted the compact bilinear pooling from Simple MCB with an LSTM consisting of hidden states equal to the image size. A 1 × 1 convolutional layer followed by a tanh activation were used on the image features prior to the LSTM, while the question representation was replicated to have a common embedding size for both representations This model is referred to as Joint LSTM. We begin by testing our baseline models on the VQA 1.0 validation set. As shown in Table 1, the language-only baseline model managed to get 48.3% overall. More impressively, it scored 78.56% on Yes/No questions. The Simple MCB model further improves the overall performance, although little improvement is gained in the binary Yes/No tasks. Replacing MCB with our basic Joint LSTM embedding improves performance across the board. Modifications to the Joint LSTM Model We test several variations of the Joint LSTM baseline which are highlighted in Table 2. Using PReLU activations has helped in two ways. First, it reduced time for convergence from 240K iterations to 120K. Second, the overall accuracy has improved, especially in the Other category. The next modifications were inspired by the results from [18]. We experimented with appending positional features which can be described as the coordinates of each pixel to the depth/feature dimension of the image representation. When unnormalized with respect to the other features, it worsened results significantly, dropping the overall accuracy by over 2 points. Model VQA 1.0 Validation Split Open Ended Task Y/N Num. Other Joint LSTM baseline PReLU Pos. features Pos. features (norm.) High dropout Extra FC 79.90 79.61 79.68 79.69 79.03 78.86 36.96 36.21 36.52 36.36 34.84 33.51 49.58 50.77 46.59 50.69 47.25 45.57 All 59.34 59.74 57.71 59.75 57.59 56.51 Table 2. Evaluation of the Joint LSTM model and its modifications on the VQA 1.0 validation split. Normalizing positional features did not have enough of a noticeable improvement (0.01 points overall) to warrant its effectiveness. Next, all dropout values are increased from 0.3 to 0.5 deteriorated the network’s accuracy, particularly in the Number and Other categories. The final modification was inserting a fully connected layer with 1024 hidden units before the classifier, which surprisingly dropped the accuracy massively. 4.2. VQA 2.0 Experiments After the release of VQA 2.0, we shifted our empirical evaluation towards the newer dataset. First, we retrain and retest our best performing VQA 1.0 model Joint LSTM as well as several improvements and modifications. Since VQA 2.0 was built to reduce the language prior and bias inherent in VQA, the accuracy of Joint LSTM drops significantly as shown in Table 3. Note that all the models that were trained so far do not have explicit visual or textual attention implemented. Our first network with explicit visual attention, RVAU, shows an accuracy jump by almost 3 points compared to the Joint LSTM model. This result highlights the importance of attention for good performance in VQA. Training the RVAU network as a multilabel task (RVAUmultilabel ), i.e. using all available annotations at each training iteration, drops the accuracy horribly. This is the biggest drop in performance so far. This might be caused by the variety of annotations in VQA for each question which makes the task for optimizing all answers at once much harder. DRAU Evaluation The addition of RTAU marks the creation of our DRAU network. The DRAU model shows favorable improvements over the RVAU model. Adding textual attention improves overall accuracy by 0.56 points. Substituting the PReLU activations with ReLU (DRAUReLU ) massively drops performance. While further training might have helped the model improve, PReLU offers much faster 1 Concurrent Work Model VQA 2.0 Validation Split Open Ended Task Y/N Num. Other All Joint LSTM w/PReLU 72.04 37.95 48.58 56.00 RVAU RVAUmultilabel 74.59 77.53 37.75 36.05 52.81 40.18 59.02 53.67 DRAUHadamard fusion DRAUanswer vocab = 5k DRAUReLU DRAUno final dropout DRAUhigh final dropout 76.62 76.33 72.69 77.02 76.47 38.92 38.21 34.92 38.26 38.71 52.09 51.85 45.05 50.17 52.52 59.58 59.27 54.11 58.69 59.71 - - - 59.14 59.67 MCB [26] Kazemi and Elqursh [18]1 Table 3. Evaluation of RVAU and DRAU-based models on the VQA 2.0 validation split. convergence. Increasing the value of the dropout layer after the fusion operation (DRAUhigh final dropout ) improves performance by 0.13 points, in contrast to the results of the Joint LSTM model on VQA 1.0. Note that on the VQA 1.0 tests, we changed the values of all layers that we apply dropout on, but here we only change the last one after the fusion operation. Totally removing this dropout layer worsens accuracy. This suggests that the optimal dropout value should be tuned per-layer. We test a few variations of DRAU on the test-dev set. We can observe that VQA benefits from more training data; the same DRAU network performs better (62.24% vs. 59.58%) thanks to the additional data. Most of the literature resize the original ResNet features from 224 × 224 to 448 × 448. To test the effect of this scaling, we train a DRAU variant with the original ResNet size (DRAUsmall ). Reducing the image feature size from 2048 × 14 × 14 to 2048 × 7 × 7 adversely affects accuracy as shown in Table 4. Adding more glimpses significantly reduces the model’s accuracy (DRAUglimpses = 4 ). A cause of this performance drop could be related to the fact that LSTMs process the input in a onedimensional fashion and thus decide that each input is either relevant or non-relevant. This might explain why the attention maps of DRAU separate the objects from the background in two glimpses as we will mention in Section 5. 2D Grid LSTMs [31] might help remove this limitation. Removing the extra data from Visual Genome hurts the model’s accuracy. That supports the fact that VQA is very diverse and that extra data helps the model perform better. Finally, substituting Hadamard product of MCB in the final fusion operation boosts the network’s accuracy significantly by 1.17 points (DRAUMCB fusion ). As mentioned in Section 3.1, we experiment replacing the global ResNet features with object-level features as sug- VQA 2.0 Test-Dev Split Open Ended Task Y/N Num. Other Model DRAUHadamard fusion DRAUsmall DRAUglimpses = 4 DRAUno genome DRAUMCB fusion DRAUFRCNN features 78.27 77.53 76.82 79.63 78.97 82.85 40.31 38.78 39.15 39.55 40.06 44.78 53.57 49.93 51.07 51.81 55.47 57.4 4.4. DRAU versus the state-of-the-art All 62.24 60.03 60.32 61.88 63.41 66.45 Table 4. Evaluation of later DRAU-based models on the VQA 2.0 test-dev split. VQA 2.0 Test-dev Split Open Ended Task Y/N Num. Other Model All 3 MCB [7] MCB w/RVAU 78.41 77.31 38.81 40.12 53.23 54.64 61.96 62.33 MUTAN [14] MUTAN w/RVAU 79.06 79.33 38.95 39.48 53.46 53.28 62.36 62.45 Table 5. Results of state-of-the-art models with RVAU. gested by [22]. This change provides a significant performance increase of 3.04 points (DRAUFRCNN features ). 4.3. Transplanting RVAU in other models To verify the effectiveness of the recurrent attention units, we replace the attention layers in MCB and MUTAN [14] with RVAU. For MCB [7] we remove all the layers after the first MCB operation until the first 2048-d output and replace them with RVAU. Due to GPU memory constraints, we reduced the size of each hidden unit in RVAU’s LSTM from 2048 to 1024. In the same setting, RVAU significantly helps improve the original MCB model’s accuracy as shown in Table 5. The most noticeable performance boost can be seen in the number category, which supports our hypothesis that recurrent layers are more suited for sequential reasoning. Furthermore, we test RVAU in the MUTAN model [14]. The authors use a multimodal vector with dimension size of 510 for the joint representations. For coherence, we change the usual dimension size in RVAU to 510. At the time of this writing, the authors have not released results on VQA 2.0 using a single model rather than a model ensemble. Therefore, we train a single-model MUTAN using the authors’ implementation.2 The story does not change here, RVAU improves the model’s overall accuracy. 2 https://github.com/Cadene/vqa.pytorch 3 http://www.visualqa.org/roe_2017.html VQA 1.0 Table 6 shows a comparison between DRAU and other state-of-the-art models. Excluding model ensembles, DRAU performs favorably against other models. To the best of our knowledge, [5] has the best single model performance of 65.07% on the test-std split which is very close our best model (65.03%). Small modifications or hyperparameter tuning could push our model further. Finally, the FRCNN image features boosts the model’s performance close to the state-of-the-art ensemble model. VQA 2.0 Our model DRAUMCB fusion landed the 8th place in the VQA 2.0 Test-standard task.4 . Currently, all reported submissions that outperform our single model use model ensembles. Using FRCNN features boosted the model’s performance to outperform some of the ensemble models (66.85%). The first place submission [22] reports using an ensemble of 30 models. In their report, the best single model that uses FRCNN features achieves 65.67% on the test-standard split which is outperformed by our best single model DRAUFRCNN features . 5. DRAU versus MCB In this section, we provide qualitative results that highlight the effect of the recurrent layers compared to the MCB model. The strength of RAUs is notable in tasks that require sequentially processing the image or relational/multi-step reasoning. In the same setting, DRAU outperforms MCB in counting questions. This is validated in a subset of the validation split questions in the VQA 2.0 dataset as shown in Figure 4. Figure 5 shows some qualitative results between DRAU and MCB. For fair comparison we compare the first attention map of MCB with the second attention map of our model. We do so because the authors of MCB [7] visualize the first map in their work5 . Furthermore, the first glimpse of our model seems to be the complement of the second attention, i.e. the model separates the background and the target object(s) into separate attention maps. We have not tested the visual effect of more than two glimpses on our model. In Figure 5, it is clear that the recurrence helps the model attend to multiple targets as apparent in the difference of the attention maps between the two models. DRAU seems to also know how to count the right object(s). The top right example in Figure 5 illustrates that DRAU is not easily fooled by counting whatever object is present in the image but rather the object that is needed to answer the question. This 4 https://evalai.cloudcv.org/web/challenges/ challenge-page/1/leaderboard 5 https://github.com/akirafukui/vqa-mcb/blob/ master/server/server.py#L185 Model VQA 1.0 Open Ended Task Test-dev Y/N Num. Other All Y/N Test-standard Num. Other SAN [10] DMN+ [16] MRN [21] HieCoAtt [9] RAU [8] DAN [6] MCB [7] (e = 7) MLB [5] (1 model) MLB [5] (e = 7) MUTAN [14] (e = 5) 79.3 80.5 82.28 79.7 81.9 83.0 83.4 84.57 85.14 36.6 36.8 38.82 38.7 39.0 39.1 39.8 39.21 39.81 46.1 48.3 49.25 51.7 53.0 53.9 58.5 57.81 58.52 58.7 60.3 61.68 61.8 63.3 64.3 66.7 66.77 67.42 82.39 81.7 82.8 83.24 84.02 84.61 84.91 38.23 38.2 38.1 39.47 37.90 39.07 39.79 49.41 52.8 54.0 58.00 54.77 57.79 58.35 58.9 60.4 61.84 62.1 63.2 64.2 66.47 65.07 66.89 67.36 DRAUHadamard fusion DRAUMCB fusion DRAUFRCNN features 82.73 82.44 84.92 38.18 38.22 39.16 54.43 56.30 57.70 64.3 65.1 66.86 82.41 84.87 38.33 40.02 55.97 57.91 65.03 67.16 All Table 6. DRAU compared to the state-of-the-art on the VQA 1.0 dataset.e = n corresponds to a model ensemble of size n. Model VQA 2.0 Open Ended Task Test-dev Y/N Num. Other All Y/N Test-standard Num. Other UPC MIC TJ neural-vqa-attention[10] CRCV REU VQATeam MCB [26] DCD ZJU[32] VQAMachine [33] POSTECH UPMC-LIP6[14] LV NUS[34] DLAIT HDU-USYD-UNCC[35] Adelaide-Teney ACRV MSR[36] 67.1 69.02 70.1 73.91 78.41 79.84 79.4 78.98 81.96 81.95 82.94 84.39 85.24 31.54 34.52 35.39 36.82 38.81 38.72 40.95 40.9 41.62 48.31 47.08 45.76 48.19 25.46 35.76 47.32 54.85 53.23 53.08 53.24 55.35 57.07 59.99 59.94 59.14 59.88 43.3 49.32 55.35 60.65 61.96 62.47 62.62 63.45 65.57 67.71 67.95 68.02 69.00 66.97 69.22 69.77 74.08 78.82 79.85 79.82 79.32 82.07 81.92 83.17 84.5 85.54 31.38 34.16 35.65 36.43 38.28 38.64 40.91 40.67 41.06 48.38 46.66 45.39 47.45 25.81 35.97 47.18 54.84 53.36 52.95 53.35 55.3 57.12 59.63 60.15 59.01 59.82 43.48 49.56 55.28 60.81 62.27 62.54 62.97 63.66 65.71 67.64 68.22 68.09 69.13 DRAUHadamard fusion DRAUMCB fusion DRAUFRCNN features 78.27 78.97 82.85 40.31 40.06 44.78 53.58 55.48 57.4 62.24 63.41 66.45 78.86 79.27 83.35 39.91 40.15 44.37 53.76 55.55 57.63 62.66 63.71 66.85 All Table 7. DRAU compared to the current submissions on the VQA 2.0 dataset. property also translates to questions that require relational reasoning. The second column in Figure 5 demonstrates how DRAU can attend the location required to answer the question based on the textual and visual attention maps. 6. Conclusion We proposed an architecture for VQA with a novel attention unit, termed the Recurrent Attention Unit (RAU). The recurrent layers help guide the textual and visual atten- tion since the network can reason relations between several parts of the image and question. We provided quantitative and qualitative results indicating the usefulness of a recurrent attention mechanism. Our DRAU model showed improved performance in tasks requiring sequential/complex reasoning such as counting or relational reasoning over [7], the winners of the VQA 2016 challenge. In VQA 1.0, we achieved near state-of-the-art results for single model performance with our DRAU network (65.03 vs. 65.07 [5]). Adding the FRCNN features gets the model within margin Question Type (Occurrence) Accuracy(%) 47.28 46.31 43.98 42.06 DRAU MCB 44.30 42.57 Visual Question Answering and Visual Grounding,” in Empirical Methods in Natural Language Processing (EMNLP), 2016, pp. 457–468. [8] H. Noh and B. Han, “Training Recurrent Answering Units with Joint Loss Minimization for VQA,” arXiv:1606.03647, 2016. [9] J. Lu, J. Yang, D. Batra, and D. Parikh, “Hierarchical Question-Image Co-Attention for Visual Question Answering,” in Advances in Neural Information Processing Systems (NIPS), 2016, pp. 289–297. how many people are in (905) how many people are (2,005) how many (20,462) Figure 4. Results on questions that require counting in the VQA 2.0 validation set. of the state-of-the-art 5-model ensemble MUTAN [14]. Finally, we demonstrated that substituting the visual attention mechanism in other networks, MCB [7] and MUTAN [14], consistently improves their performance. In future work we will investigate implicit recurrent attention mechanism using recently proposed explanation methods [37, 38]. References [1] K. He, X. Zhang, S. Ren, and J. Sun, “Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification,” in IEEE International Conference on Computer Vision (ICCV), 2015, pp. 1026–1034. [2] S. Bosse, D. Maniry, K.-R. Müller, T. Wiegand, and W. Samek, “Deep neural networks for no-reference and fullreference image quality assessment,” IEEE Transactions on Image Processing, vol. 27, no. 1, pp. 206–219, 2018. [10] Z. Yang, X. He, J. Gao, L. Deng, and A. Smola, “Stacked Attention Networks for Image Question Answering,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 21–29. [11] K. Chen, J. Wang, L.-C. Chen, H. Gao, W. Xu, and R. Nevatia, “ABC-CNN: An Attention Based Convolutional Neural Network for Visual Question Answering,” arXiv:1511.05960, 2015. [12] H. Xu and K. Saenko, “Ask, Attend and Answer: Exploring Question-Guided Spatial Attention for Visual Question Answering,” in European Conference on Computer Vision (ECCV), 2016, pp. 451–466. [13] M. Charikar, K. Chen, and M. Farach-Colton, “Finding frequent items in data streams,” Theoretical Computer Science, vol. 312, no. 1, pp. 3–15, 2004. [14] H. Ben-younes, R. Cadene, M. Cord, and N. Thome, “MUTAN: Multimodal Tucker Fusion for Visual Question Answering,” arXiv:1705.06676, 2017. [15] L. R. Tucker, “Some mathematical notes on three-mode factor analysis,” Psychometrika, vol. 31, no. 3, pp. 279–311, 1966. [3] D. Bahdanau, K. Cho, and Y. Bengio, “Neural Machine Translation by Jointly Learning to Align and Translate,” in International Conference on Representation Learning (ICLR), 2015. [16] C. Xiong, S. Merity, and R. Socher, “Dynamic Memory Networks for Visual and Textual Question Answering,” in International Conference on Machine Learning (ICML), 2016, pp. 2397–2406. [4] R. Nallapati, B. Zhou, C. Gulcehre, B. Xiang et al., “Abstractive text summarization using sequence-to-sequence rnns and beyond,” arXiv:1602.06023, 2016. [17] P. R. Kumar and P. Varaiya, Stochastic systems: Estimation, identification, and adaptive control. SIAM, 2015. [5] J.-H. Kim, K.-W. On, J. Kim, J.-W. Ha, and B.-T. Zhang, “Hadamard Product for Low-rank Bilinear Pooling,” in International Conference on Representation Learning (ICLR), 2017. [6] H. Nam, J.-W. Ha, and J. Kim, “Dual Attention Networks for Multimodal Reasoning and Matching,” in IEEE International Conference on Computer Vision (CVPR), 2017, pp. 299–307. [7] A. Fukui, D. H. Park, D. Yang, A. Rohrbach, T. Darrell, and M. Rohrbach, “Multimodal Compact Bilinear Pooling for [18] V. Kazemi and A. Elqursh, “Show, Ask, Attend, and Answer: A Strong Baseline For Visual Question Answering,” arXiv:1704.03162, 2017. [19] Q. Wu, P. Wang, C. Shen, A. Dick, and A. van den Hengel, “Ask Me Anything: Free-form Visual Question Answering Based on Knowledge from External Sources,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 4622–4630. [20] A. Ray, G. Christie, M. Bansal, D. Batra, and D. Parikh, “Question Relevance in VQA: Identifying Non-Visual And False-Premise Questions,” in Empirical Methods in Natural Language Processing (EMNLP), 2016, pp. 919–924. How many How many camels are in lanterns hang the photo? DRAU: 0 off the clock tower ? DRAU: 4 How many lanterns hang off the clock tower? GT: 4 MCB: 1 How many camels are in the photo? GT: 0 What is on the floor leaning on the bench in How many people are between the shown ? DRAU: 3 How many people are shown? GT: 3 MCB: 2 people ? DRAU: racket What is on the floor leaning on the bench in between the people? GT: racket does the vase animals are depict ? DRAU: owl there ? DRAU: 2 MCB: one MCB: backpack What animal How many How many animals are there? GT: 2 MCB: 1 What animal does the vase depict? GT: owl MCB: elephant Figure 5. DRAU vs. MCB Qualitative examples. Attention maps for both models shown, only DRAU has textual attention. [21] J.-H. Kim, S.-W. Lee, D. Kwak, M.-O. Heo, J. Kim, J.-W. Ha, and B.-T. Zhang, “Multimodal Residual Learning for Visual QA,” in Advances in Neural Information Processing Systems 29, 2016, pp. 361–369. [22] P. Anderson, X. He, C. Buehler, D. Teney, M. Johnson, S. Gould, and L. Zhang, “Bottom-Up and Top-Down At- tention for Image Captioning and VQA,” arXiv:1707.07998, 2017. [23] S. Ren, K. He, R. Girshick, and J. Sun, “Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks,” arXiv:1506.01497, Jun. 2015. [24] J. Pennington, R. Socher, and C. D. Manning, “GloVe: Global Vectors for Word Representation,” in Empirical Methods in Natural Language Processing (EMNLP), 2014, pp. 1532–1543. [25] Y. Gao, O. Beijbom, N. Zhang, and T. Darrell, “Compact bilinear pooling,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 317–326. [26] Y. Goyal, T. Khot, D. Summers-Stay, D. Batra, and D. Parikh, “Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering,” in IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 6904–6913. [27] S. Antol, A. Agrawal, J. Lu, M. Mitchell, C. L. Zitnick, D. Batra, and D. Parikh, “VQA: Visual Question Answering,” in IEEE International Conference on Computer Vision (CVPR), 2015, pp. 2425–2433. [28] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick, “Microsoft coco: Common objects in context,” in European Conference on Computer Vision – ECCV 2014, 2014, pp. 740–755. [29] D. P. Kingma and J. Ba, “Adam: A Method for Stochastic Optimization,” arXiv:1412.6980, 2014. [30] X. Glorot and Y. Bengio, “Understanding the difficulty of training deep feedforward neural networks,” in AISTATS, 2010, pp. 249–256. [31] N. Kalchbrenner, I. Danihelka, and A. Graves, “Grid long short-term memory,” arXiv:1507.01526, 2015. [32] Y. Lin, Z. Pang, D. Wang, and Y. Zhuang, “Task-driven Visual Saliency and Attention-based Visual Question Answering,” arXiv:1702.06700, 2017. [33] P. Wang, Q. Wu, C. Shen, and A. van den Hengel, “The VQA-Machine: Learning How to Use Existing Vision Algorithms to Answer New Questions,” arXiv:1612.05386, 2016. [34] I. Ilievski and J. Feng, “A Simple Loss Function for Improving the Convergence and Accuracy of Visual Question Answering Models,” arXiv:1708.00584, 2017. [35] Z. Yu, J. Yu, C. Xiang, J. Fan, and D. Tao, “Beyond Bilinear: Generalized Multi-modal Factorized High-order Pooling for Visual Question Answering,” arXiv:1708.03619, 2017. [36] D. Teney, P. Anderson, X. He, and A. van den Hengel, “Tips and Tricks for Visual Question Answering: Learnings from the 2017 Challenge,” arXiv:1708.02711, 2017. [37] G. Montavon, W. Samek, and K.-R. Müller, “Methods for interpreting and understanding deep neural networks,” Digital Signal Processing, vol. 73, pp. 1–15, 2018. [38] L. Arras, G. Montavon, K.-R. Müller, and W. Samek, “Explaining recurrent neural network predictions in sentiment analysis,” in EMNLP’17 Workshop on Computational Approaches to Subjectivity, Sentiment & Social Media Analysis (WASSA), 2017, pp. 159–168. 

Learning of Human-like Algebraic Reasoning Using Deep Feedforward Neural Networks arXiv:1704.07503v1 [cs.AI] 25 Apr 2017 Cheng-Hao Cai1 Dengfeng Ke1∗ Yanyan Xu2 Kaile Su3 1 National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences 2 School of Information Science and Technology, Beijing Forestry University 3 Institute for Integrated and Intelligent Systems, Griffith University chenghao.cai@outlook.com, dengfeng.ke@nlpr.ia.ac.cn, xuyanyan@bjfu.edu.cn, k.su@griffith.edu.au Abstract There is a wide gap between symbolic reasoning and deep learning. In this research, we explore the possibility of using deep learning to improve symbolic reasoning. Briefly, in a reasoning system, a deep feedforward neural network is used to guide rewriting processes after learning from algebraic reasoning examples produced by humans. To enable the neural network to recognise patterns of algebraic expressions with non-deterministic sizes, reduced partial trees are used to represent the expressions. Also, to represent both top-down and bottom-up information of the expressions, a centralisation technique is used to improve the reduced partial trees. Besides, symbolic association vectors and rule application records are used to improve the rewriting processes. Experimental results reveal that the algebraic reasoning examples can be accurately learnt only if the feedforward neural network has enough hidden layers. Also, the centralisation technique, the symbolic association vectors and the rule application records can reduce error rates of reasoning. In particular, the above approaches have led to 4.6% error rate of reasoning on a dataset of linear equations, differentials and integrals. 1 Introduction It is challenging to integrate symbolic reasoning and deep learning in effective ways [Garcez et al., 2015]. In the field of symbolic reasoning, much work has been done on using formal methods to model reliable reasoning processes [Chang and Lee, 1973]. For instance, algebraic reasoning can be modelled by using first-order predicate logics or even higher-order logics, but these logics are usually designed by experienced experts, because it is challenging for machines to learn these logics from data automatically [Bundy and Welham, 1981; Nipkow et al., 2002]. On the other hand, recent approaches on deep learning have revealed that deep neural networks are powerful tools for learning from data [Lecun et al., 2015], especially for learning speech features ∗ Corresponding Author. [Mohamed et al., 2012] and image features [Sun et al., 2015]. However, not much work has been done on using deep neural networks to learn formal symbolic logics. To close the gap between symbolic reasoning and deep learning, this research explores the possibility of using deep feedforward neural networks to learn logics of rewriting in algebraic reasoning. In other words, we try to teach neural networks to solve mathematical problems, such as finding the solution of an equation and calculating the differential or integral of an expression, by using a rewriting system. Rewriting is an important technique in symbolic reasoning. Its core concept is to simply reasoning process by using equivalence relations between different expressions [Bundy, 1983]. Usually, rewriting is based on a tree-manipulating system, as many algebraic expressions can be represented by using tree structures, and the manipulation of symbols in the expressions is equivalent to the manipulation of nodes, leaves and sub-trees on the trees [Rosen, 1973]. To manipulate symbols, a rewriting system usually uses one way matching, which is a restricted application of unification, to find a desired pattern from an expression and then replaces the pattern with another equivalent pattern [Bundy, 1983]. In order to reduce the search space, rewriting systems are expected to be Church-Rosser, which means that they should be terminating and locally confluent [Rosen, 1973; Huet, 1980]. Thus, very careful designs and analyses are needed: A design can start from small systems, because proving termination and local confluence of a smaller system is usually easier than proving those of a larger system [Bundy and Welham, 1981]. Some previous work has focused on this aspect: The Knuth-Bendix completion algorithm can be used to solve the problem of local confluence [Knuth and Bendix, 1983], and Huet [1981] has provided a proof of correctness for this algorithm. Also, dependency pairs [Arts and Giesl, 2000] and semantic labelling [Zantema, 1995] can solve the problem of termination for some systems. After multiple small systems have been designed, they can be combined into a whole system, because the direct sum of two Church-Rosser systems holds the same property [Toyama, 1987]. Deep neural networks have been used in many fields of artificial intelligence, including speech recognition [Mohamed et al., 2012], human face recognition [Sun et al., 2015], natural language understanding [Sarikaya et al., 2014], reinforcement learning for playing video games [Mnih et al., 2015] and Monte Carlo tree search for playing Go [Silver et al., 2016]. Recently, some researchers are trying to extend them to reasoning tasks. For instance, Irving et al. [2016] have proposed DeepMath for automated theorem proving with deep neural networks. Also, Serafini and Garcez [2016] have proposed logic tensor networks to combine deep learning with logical reasoning. In addition, Garnelo et al. [2016] have explored deep symbolic reinforcement learning. In this research, we use deep feedforward neural networks [Lecun et al., 2015] to guide rewriting processes. This technique is called human-like rewriting, as it is adapted from standard rewriting and can simulate human’s behaviours of using rewrite rules after learning from algebraic reasoning schemes. The following sections provide detailed discussions about this technique: Section 2 introduces the core method of human-like rewriting. Section 3 discusses algebraic reasoning schemes briefly. Section 4 provides three methods for system improvement. Section 5 provides experimental results of the core method and the improvement methods. Section 6 is for conclusions. 2 Rewriting is an inference technique for replacing expressions or subexpressions with equivalent ones [Bundy, 1983]. For instance1 , given two rules of the Peano axioms: x+0⇒x (1) x + S(y) ⇒ S(x + y) (2) S(0) + S(S(0)) can be rewritten via: S(0) + S(S(0)) | {z } by (2) ⇒ S( S(0) + S(0) ) {z } | by (2) (3) ⇒ S(S( S(0) + 0 )) | {z } by (1) ⇒ S(S(S(0))) More detailed discussions about the Peano axioms can be found from [Pillay, 1981]. Generally, rewriting requires a source expression s and a set of rewrite rules τ . Let l ⇒ r denote a rewrite rule in τ , t a subexpression of s, and θ the most general unifier of one way matching from l and t. A single rewriting step of inference can be formed as: (l ⇒ r) ∈ τ s(r[θ]) l[θ] ≡ t (4) It is noticeable that θ is only applied to l, but not to t. The reason is that one way matching, which is a restricted application of unification, requires that all substitutions in a unifier are only applied to the left-hand side of a unification pair. Standard rewriting is to repeat the above step until 1 ⇒ ⇒ ⇒ ⇒ Human-like Rewriting s(t) no rule can be applied to the expression further. It requires the set of rewrite rules τ to be Church-Rosser, which means that τ should be terminating and locally confluent. This requirement restricts the application of rewriting in many D(f ) fields. For instance, the chain rule in calculus ⇒ D(x) D(f ) D(u) · , which is very important for computing D(u) D(x) derivatives, will result in non-termination: We use the mathematical convention that a word is a constant if its first letter is in upper case, and it is a variable if its first letter is in lower case. D(Sin(X)) D(X) D(Sin(X)) · D(u1 ) D(Sin(X)) · D(u2 ) D(Sin(X)) · D(u3 ) ··· D(u1 ) D(X) D(u2 ) · D(u1 ) D(u3 ) · D(u2 ) D(u1 ) D(X) D(u2 ) D(u1 ) · D(u1 ) D(X) (5) The above process means that it is challenging to use the chain rule in standard rewriting. Similarly, a commutativity rule x ◦ y ⇒ y ◦ x, where ◦ is an addition, a multiplication, a logical conjunction, a logical disjunction or another binary operation satisfying commutativity, is difficult to be used in standard rewriting. If termination is not guaranteed, it will be difficult to check local confluence, as local confluence requires a completely developed search tree, but non-termination means that the search tree is infinite and cannot be completely developed. More detailed discussion about standard rewriting and Church-Rosser can be found from [Bundy, 1983]. Human-like rewriting is adapted from standard rewriting. It uses a deep feedforward neural network [Lecun et al., 2015] to guide rewriting processes. The neural network has learnt from some rewriting examples produced by humans, so that it can, to some extent, simulate human’s ways of using rewrite rules: Firstly, non-terminating rules are used to rewrite expressions. Secondly, local confluence is not checked. Lastly, experiences of rewriting can be learnt and can guide future rewriting processes. To train the feedforward neural network, input data and target data are required. An input can be generated via the following steps: Firstly, an expression is transformed to a parsing tree [Huth and Ryan, 2004] with position annotations. A position annotation is a unique label < p1 , p2 , · · · , pN > indicating a position on a tree, where each pi is the order of a branch. Then the tree is reduced to a set of partial trees with a predefined maximum depth d. Next, the partial trees are expanded to perfect k-ary trees with the depth d and a predefined breadth k. In particular, empty positions on the prefect k-ary trees are filled by Empty. After that, the perfect k-ary trees are transformed to lists via in-order traversal. Detailed discussions about perfect k-ary trees and in-order traversal can be found from [Cormen et al., 2001]. Finally, the lists with their position annotations are transformed to a set of one-hot representations [Turian et vector, Sof tmax the standard Softmax function [Bishop, 2006], and y the output vector. The averaged Softmax layer is defined as: P 1X ui = xj,i (6) P j=1 6 × Y = Y + Ln(3) <1> = <1,1> <1,2> × <1,1,1> + <1,1,2> 6 <1,2,1> Y y = Sof tmax(W · u + b) (7) It is noticeable that the output is a single vector regardless of the number of the input vectors. The feedforward neural network is trained by using the back-propagation algorithm with the cross-entropy error function [Hecht-Nielsen, 1988; Bishop, 2006]. After training, the neural network can be used to guide a rewriting procedure: Given the RPT representation of an expression, the neural network uses forward computation to get an output vector, and the position of the maximum element indicates the name of a rewrite rule and a possible position for the application of the rule. <1,2,2> Y Ln <1,2,2,1> 3 <1> <1,2> = + × 6 + Y Y 3 Ln <1,2,2> × Ln 3 Y <1> <1,2> = + × Y Ln E E × Ln E 3 E E E 3 <1,2,2> Y E Ln <1,1> 6 3 Y + Y E Ln <1,1> 6 6 Y E E E E <1> =,×,6,Y,+,Y,Ln <1,2> +,Y,E,E,Ln,3,E <1,1> ×,6,E,E,Y,E,E <1,2,2> Ln,3,E,E,E,E,E [1, 0, 0, 0, …, 0, 0, 1, 0, 0, 0, 0, 0, …, 0, 0, 1, 0] [0, 1, 0, 0, …, 0, 0, 0, 1, 0, 0, 0, 0, …, 0, 0, 0, 0] [0, 0, 1, 0, …, 0, 0, 0, 0, 1, 0, 0, 0, …, 0, 0, 0, 0] [0, 0, 0, 1, …, 0, 0, 0, 0, 0, 0, 0, 0, …, 0, 0, 0, 0] Figure 1: An expression 6×Y = Y +Ln(3) is transformed to the RPT representation. The maximum depth and the breadth of a tree are defined as 2. “E” is an abbreviation of “Empty”. al., 2010]. In particular, Empty is transformed to a zero block. Figure 1 provides an example for the above procedure. This representation is called a reduced partial tree (RPT) representation of the expression. A target is the one-hot representation [Turian et al., 2010] of a rewrite rule name with a position annotation for applying the rule. It is noticeable that the input of the neural network is a set of vectors, and the number of vectors is non-deterministic, as it depends on the structure of the expression. However, the target is a single vector. Thus, the dimension of the input will disagree with the dimension of the target if a conventional feedforward neural network structure is used. To solve this problem, we replace its Softmax layer with an averaged Softmax layer. Let xj,i denote the ith element of the jth input vector, P the number of the input vectors, u an averaged input vector, ui the ith element of u, W a weight matrix, b a bias Algebraic Reasoning Schemes The learning of the neural network is based on a set of algebraic reasoning schemes. Generally, an algebraic reasoning scheme consists of a question, an answer and some intermediate reasoning steps. The question is an expression indicating the starting point of reasoning. The answer is an expression indicating the goal of reasoning. Each intermediate reasoning step is a record consisting of: • A source expression; • The name of a rewrite rule; • A position annotation for applying the rewrite rule; • A target expression. In particular, the source expression of the first reasoning step is the question, and the target expression of the final reasoning step is the answer. Also, for each reasoning step, the target expression will be the source expression of the next step if the “next step” exists. By applying all intermediate reasoning steps, the question can be rewritten to the answer deterministically. In this research, algebraic reasoning schemes are developed via a rewriting system in SWI-Prolog [Wielemaker et al., 2012]. The rewriting system is based on Rule (4), and it uses breadth-first search to find intermediate reasoning steps from a question to an answer. Like most rewriting systems and automated theorem proving systems2 , its ability of reasoning is restricted by the problem of combinatorial explosion: The number of possible ways of reasoning can grow rapidly when the question becomes more complex [Bundy, 1983]. Therefore, a full algebraic reasoning scheme of a complex question is usually difficult to be generated automatically, and guidance from humans is required. In other words, if the system fails to develop the scheme, we will apply rewrite rules manually until the remaining part of the scheme can be developed automatically, or we will 2 A practical example is the “by auto” function of Isabelle/HOL [Nipkow et al., 2002]. It is often difficult to prove a complex theorem automatically, so that experts’ guidance is often required. provide some subgoals for the system to reduce the search space. After algebraic reasoning schemes are developed, their intermediate reasoning steps are used to train the neural network: For each step, the RPT representation of the source expression is the input of the neural network, and the onehot representation of the rewrite rule name and the position annotation is the target of the neural network, as discussed by Section 2. 4 Methods for System Improvement 4.1 Centralised RPT Representation The RPT representation discussed before is a top-down representation of an expression: A functor in the expression is a node, and arguments dominated by the functor are child nodes or leaves of the node. However, it does not record bottom-up information about the expression. For instance, in Figure 1, the partial tree labelled < 1, 1 > does not record any information about its parent node “=”. = <1,2> × <1,1,1> + <1,1,2> 6 <1,2,1> Y <1,2,2> Y Ln <1,2,2,1> 3 <1> = × 6 Y + = E E E <1,1> × 6 E E = Y × E E × × + E <1,2> Consider the following rewrite rule: x × x ⇒ x2 4.3 E E + 3 E + × + E <1,2,2> Ln E 3 E E Ln E E Rule Application Record Qi ≡ < rulei , posi > = Ln + E Y (8) Previous applications of rewrite rules can provide hints for current and future applications. In this research, we use rule application records (RAR) to record the previous applications of rewrite rules: Let Qi denote the ith element of an RAR Q, rulei the name of the previous ith rewrite rule, and posi the position annotation for applying the rule. Qi is defined as: + Y Symbolic Association Vector After the matrix is produced, it can be reshaped to a vector and be a part of an input vector of the neural network. E Y Ln = 4.2 The application of this rule requires that two arguments of “×” are the same. If this pattern exists in an expression, it will be a useful hint for selecting rules. In such case, the use of a symbolic association vector (SAV) can provide useful information for the neural network: Assume that H is the list representation of a perfect k-ary tree (which has been discussed by Section 2) with a length L. S is defined as an L × L matrix which satisfies:  1, if Hi = Hj and i 6= j; Si,j = (9) 0, otherwise. <1> <1,1> both top-down and bottom-up information of an expression: Firstly, every node on a tree considers itself as the centre of the tree and grows an additional branch to its parent node (if it exists), so that the tree becomes a directed graph. This step is called “centralisation”. Then the graph is reduced to a set of partial trees and expanded to a set of perfect k-ary trees. In particular, each additional branch is defined as the kth branch of its parent node, and all empty positions dominated by the parent node are filled by Empty. Detailed discussions about perfect k-ary trees and directed graphs can be found from [Cormen et al., 2001]. Figure 2 provides an example for the above steps. Finally, these perfect k-ary trees are transformed to lists and further represented as a set of vectors, as discussed by Section 2. Ln = Figure 2: The parsing tree of 6 × Y = Y + Ln(3) is centralised, reduced and expanded to a set of perfect k-ary trees. Dashed arrows denote additional branchs generated by centralisation. A centralised RPT (C-RPT) representation can represent (10) Usually, the RAR only records the last N applications of rewrite rules, where N is a predefined length of Q. To enable the neural network to read the RAR, it needs to be transformed to a one-hot representation [Turian et al., 2010]. A drawback of RARs is that they cannot be used in the first N steps of rewriting, as they record exactly N previous applications of rewrite rules. 5 5.1 Experiments Datasets and Evaluation Metrics A dataset of algebraic reasoning schemes is used to train and test models. This dataset contains 400 schemes about linear equations, differentials and integrals and 80 rewrite rules, and these schemes consist of 6,067 intermediate reasoning steps totally. We shuffle the intermediate steps and then divide them into a training set and a test set randomly: The training set contains 5,067 examples, and the test set contains 1,000 examples. After training a model with the training set, an error rate of reasoning on the test set is used to evaluate the model, and it can be computed by: Error Rate = NError × 100% NT otal (11) where NError is the number of cases when the model fails to indicate an expected application of rewrite rules, and NT otal is the number of examples in the test set. 5.2 layers can bring about significantly better performance of learning. On the other hand, if the neural network only has a single hidden layer, the learning will stop early, while the cross-entropy loss is very high. Also, by comparing the curves with the same type of line, it is noticeable that a deeper RPT often brings about better performance of learning, but an exception is the curve of the “FNN5 + RPT2” model. Table 1: Error Rates (%) of Reasoning on the Test Set. FNNn n=1 n=3 n=5 Using RPT Representations and Neural Networks In this part, we evaluate the core method of human-like rewriting: All expressions in the dataset are represented by using the RPT representations. The breadth of an RPT is set to 2, because the expressions in the dataset are unary or binary. The depth of an RPT is set to 1, 2 or 3. Also, feedforward neural networks [Lecun et al., 2015] with 1, 3 and 5 hidden layers are used to learn from the dataset. The number of units in each hidden layer is set to 1,024, and their activation functions are rectified linear units (ReLU) [Glorot et al., 2011]. The output layer of each neural network is an averaged Softmax layer. The neural networks are trained via the back-propagation algorithm with the crossentropy error function [Hecht-Nielsen, 1988; Bishop, 2006]. When training models, learning rates are decided by the Newbob+/Train strategy [Wiesler et al., 2014]: The initial learning rate is set to 0.01, and the learning rate is halved when the average improvement of the cross-entropy loss on the training set is smaller than 0.1. The training process stops when the improvement is smaller than 0.01. 5 FNN1 + RPT1 FNN1 + RPT2 FNN1 + RPT3 FNN3 + RPT1 FNN3 + RPT2 FNN3 + RPT3 FNN5 + RPT1 FNN5 + RPT2 FNN5 + RPT3 4 Cross-Entropy Loss 3 2 1 0 1 2 3 4 5 6 7 Epoch 8 9 10 11 12 Figure 3: Learning Curves of FNN + RPT Models. Figure 3 provides learning curves of the models, where “FNNn” means that the neural network has n hidden layers, and “RPTm” means that the depth of RPTs is m. To aid the readability, the curves of “FNN1”, “FNN3” and “FNN5” are in blue, red and green respectively, and the curves of “RPT1”, “RPT2” and “RPT3” are displayed by using dotted lines, dashed lines and solid lines respectively. By comparing the curves with the same colour, it is noticeable that more hidden m=1 80.4 27.4 16.5 RPTm m=2 79.9 20.4 16.6 m=3 76.1 18.8 12.9 Table 1 reveals performance of the trained models on the test set. In this table, results in “FNNn” rows and “RPTm” columns correspond to the “FNNn+RPTm” models in Figure 3. It is noticeable that the error rates of reasoning decrease significantly when the numbers of hidden layers increase. Also, the error rates of reasoning often decrease when the depths of RPTs increase, but an exception occurs in the case of “FNN5 + RPT2”. We believe that the reason why the exception occurs is that the learning rate strategy results in early stop of training. In addition, the error rate of the FNN5 + RPT3 model is the best among all results. 5.3 Using Improvement Methods In Section 5.2, we have found that the neural networks with 5 hidden layers have better performance than those with 1 or 3 hidden layers on the task of human-like rewriting. Based on the neural networks with 5 hidden layers, we apply the three improvement methods to these models. Figure 4 shows learning curves of models improved by CRPTs, SAVs and RARs. Also, learning curves of the baseline RPTm models are displayed by using dashed lines, where m is the depth of RPTs. Learning curves of the C-RPT models are displayed by Figure 4(a). A comparison between two lines in the same colour reveals that the C-RPT representation can improve the model when m is fixed. Also, the C-RPT2 curve is very close to the RPT3 curve during the last 6 epochs, which reveals that there might be a trade-off between using C-RPTs and increasing the depth of RPTs. The best learning curve is the C-RPT3 curve, as its cross-entropy loss is always the lowest during all epochs. Figure 4(b) provides learning curves of the RPT models with the SAV method. It is noticeable that SAVs have two effects: The first is that they can bring about lower cross-entropy losses. The second is that they can reduce the costs of learning time, as each RPTm + SAV model uses fewer epochs to finish learning than its counterpart. Figure 4(c) shows learning curves of the RPT models with the RAR method. This figure reveals that RARs always improve the models. In particular, even the RPT1 + RAR model has better learning performance than the RPT3 model. Also, the RPT1 + RAR model and the RPT3 + RAR model use less epochs to be trained, which means that RARs may reduce the time consumption of learning. Figure 4(d) provides learning curves of the models with all 3 3 RPT1 RPT1 C-RPT1 RPT1 + SAV RPT2 C-RPT2 RPT3 C-RPT3 1 RPT2 2 Cross-Entropy Loss Cross-Entropy Loss 2 RPT2 + SAV RPT3 RPT3 + SAV 1 0 0 1 2 3 4 5 6 7 Epoch 8 9 10 11 1 2 (a) C-RPT Models. 3 3 4 5 6 7 Epoch 3 C-RPT1 + SAV + RAR RPT2 RPT2 + RAR RPT3 RPT3 + RAR RPT2 2 Cross-Entropy Loss Cross-Entropy Loss 10 11 RPT1 RPT1 + RAR 1 9 (b) SAV Models. RPT1 2 8 C-RPT2 + SAV + RAR RPT3 C-RPT3 + SAV + RAR 1 0 0 1 2 3 4 5 6 7 Epoch 8 9 10 11 1 (c) RAR Models. 2 3 4 5 6 7 Epoch 8 9 10 11 (d) C-RPT + SAV + RAR Models. Figure 4: Learning Curves of Models Improved by C-RPTs, SAVs and RARs. improvement methods. A glance at the figure reveals that these models have better performance of learning than the baseline models. Also, they require less epochs to be trained than their counterparts. In addition, the final cross-entropy loss of the C-RPT2 + SAV + RAR model is the lowest among all results. Table 2: Error Rates (%) After Improvement. Model RPTm (Baseline) RPTm + SAV RPTm + RAR RPTm + SAV + RAR C-RPTm C-RPTm + SAV C-RPTm + RAR C-RPTm + SAV + RAR m=1 16.5 16.8 6.7 6.8 16.1 11.9 6.3 5.4 Value of m m=2 16.6 15.8 6.0 5.2 12.9 15.1 5.1 4.6 m=3 12.9 11.5 5.4 5.4 11.6 11.8 5.1 5.3 particular, the C-RPT2 + SAV + RAR model reaches the best error rate (4.6%) among all models. 6 Conclusions and Future Work Deep feedforward neural networks are able to guide rewriting processes after learning from algebraic reasoning schemes. The use of deep structures is necessary, because the behaviours of rewriting can be accurately modelled only if the neural networks have enough hidden layers. Also, it has been shown that the RPT representation is effective for the neural networks to model algebraic expressions, and it can be improved by using the C-RPT representation, the SAV method and the RAR method. Based on these techniques, human-like rewriting can solve many problems about linear equations, differentials and integrals. In the future, we will try to use human-like rewriting to deal with more complex tasks of mathematical reasoning and extend it to more general first-order logics and higher-order logics. Acknowledgments Table 2 shows error rates of reasoning on the test set after using the improvement methods. It is noticeable that: Firstly, the C-RPTm models have lower error rates than the baseline RPTm models, especially when m = 2. Secondly, the RPTm + SAV models have lower error rates than the baseline RPTm model when m is 2 or 3, but this is not the case for the RPT1 + SAV model. Thirdly, the RARs can reduce the error rates significantly. Finally, the error rates can be reduced further when the three improvement methods are used together. In This work is supported by the Fundamental Research Funds for the Central Universities (No. 2016JX06) and the National Natural Science Foundation of China (No. 61472369). References [Arts and Giesl, 2000] Thomas Arts and Jürgen Giesl. Termination of term rewriting using dependency pairs. Theor. Comput. Sci., 236(1-2):133–178, 2000. [Bishop, 2006] Christopher M Bishop. Pattern Recognition and Machine Learning (Information Science and Statistics). Springer-Verlag New York, Inc., 2006. [Bundy and Welham, 1981] Alan Bundy and Bob Welham. Using meta-level inference for selective application of multiple rewrite rule sets in algebraic manipulation. Artif. Intell., 16(2):189–212, 1981. [Bundy, 1983] Alan Bundy. The computer modelling of mathematical reasoning. Academic Press, 1983. [Chang and Lee, 1973] Chin-Liang Chang and Richard C. T. Lee. Symbolic logic and mechanical theorem proving. Computer science classics. Academic Press, 1973. [Cormen et al., 2001] T H Cormen, C E Leiserson, R L Rivest, and C. Stein. Introduction to algorithms (second edition). page 1297C1305, 2001. [Garcez et al., 2015] Artur D ’Avila Garcez, Tarek R Besold, Luc De Raedt, Peter Fldiak, Pascal Hitzler, Thomas Icard, Kai Uwe Khnberger, Luis C Lamb, Risto Miikkulainen, and Daniel L Silver. Neural-symbolic learning and reasoning: Contributions and challenges. In AAAI Spring Symposium - Knowledge Representation and Reasoning: Integrating Symbolic and Neural Approaches, 2015. [Garnelo et al., 2016] Marta Garnelo, Kai Arulkumaran, and Murray Shanahan. Towards deep symbolic reinforcement learning. CoRR, abs/1609.05518, 2016. [Glorot et al., 2011] Xavier Glorot, Antoine Bordes, and Yoshua Bengio. Deep sparse rectifier neural networks. In Proceedings of the Fourteenth International Conference on Artificial Intelligence and Statistics, AISTATS 2011, Fort Lauderdale, USA, April 11-13, 2011, pages 315–323, 2011. [Hecht-Nielsen, 1988] Robert Hecht-Nielsen. Theory of the backpropagation neural network. Neural Networks, 1(Supplement-1):445–448, 1988. [Huet, 1980] Grard Huet. Confluent reductions: Abstract properties and applications to term rewriting systems: Abstract properties and applications to term rewriting systems. Journal of the Acm, 27(4):797–821, 1980. [Huet, 1981] Gérard P. Huet. A complete proof of correctness of the knuth-bendix completion algorithm. J. Comput. Syst. Sci., 23(1):11–21, 1981. [Huth and Ryan, 2004] Michael Huth and Mark Dermot Ryan. Logic in computer science - modelling and reasoning about systems (2. ed.). Cambridge University Press, 2004. [Irving et al., 2016] Geoffrey Irving, Christian Szegedy, Alexander A. Alemi, Niklas Eén, François Chollet, and Josef Urban. Deepmath - deep sequence models for premise selection. In Advances in Neural Information Processing Systems 29: Annual Conference on Neural Information Processing Systems 2016, December 5-10, 2016, Barcelona, Spain, pages 2235–2243, 2016. [Knuth and Bendix, 1983] Donald E. Knuth and Peter B. Bendix. Simple word problems in universal algebras. Computational Problems in Abstract Algebra, pages 263– 297, 1983. [Lecun et al., 2015] Yann Lecun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436– 444, 2015. [Mnih et al., 2015] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. Nature, 518(7540):529–533, 2015. [Mohamed et al., 2012] Abdel-rahman Mohamed, George E. Dahl, and Geoffrey E. Hinton. Acoustic modeling using deep belief networks. IEEE Trans. Audio, Speech & Language Processing, 20(1):14–22, 2012. [Nipkow et al., 2002] Tobias Nipkow, Lawrence C. Paulson, and Markus Wenzel. Isabelle/HOL - A Proof Assistant for Higher-Order Logic, volume 2283 of Lecture Notes in Computer Science. Springer, 2002. [Pillay, 1981] Anand Pillay. Models of peano arithmetic. Journal of Symbolic Logic, 67(3):1265–1273, 1981. [Rosen, 1973] Barry K. Rosen. Tree-manipulating systems and church-rosser theorems. In Acm Symposium on Theory of Computing, pages 117–127, 1973. [Sarikaya et al., 2014] Ruhi Sarikaya, Geoffrey E. Hinton, and Anoop Deoras. Application of deep belief networks for natural language understanding. IEEE/ACM Trans. Audio, Speech & Language Processing, 22(4):778–784, 2014. [Serafini and d’Avila Garcez, 2016] Luciano Serafini and Artur S. d’Avila Garcez. Logic tensor networks: Deep learning and logical reasoning from data and knowledge. CoRR, abs/1606.04422, 2016. [Silver et al., 2016] David Silver, Aja Huang, Chris J. Maddison, Arthur Guez, Laurent Sifre, George van den Driessche, Julian Schrittwieser, Ioannis Antonoglou, Veda Panneershelvam, Marc Lanctot, Sander Dieleman, Dominik Grewe, John Nham, Nal Kalchbrenner, Ilya Sutskever, Timothy Lillicrap, Madeleine Leach, Koray Kavukcuoglu, Thore Graepel, and Demis Hassabis. Mastering the game of Go with deep neural networks and tree search. Nature, 529(7587):484–489, January 2016. [Sun et al., 2015] Yi Sun, Ding Liang, Xiaogang Wang, and Xiaoou Tang. Deepid3: Face recognition with very deep neural networks. CoRR, abs/1502.00873, 2015. [Toyama, 1987] Yoshihito Toyama. On the church-rosser property for the direct sum of term rewriting systems. J. ACM, 34(1):128–143, 1987. [Turian et al., 2010] Joseph P. Turian, Lev-Arie Ratinov, and Yoshua Bengio. Word representations: A simple and general method for semi-supervised learning. In ACL 2010, Proceedings of the 48th Annual Meeting of the Association for Computational Linguistics, July 11-16, 2010, Uppsala, Sweden, pages 384–394, 2010. [Wielemaker et al., 2012] Jan Wielemaker, Tom Schrijvers, Markus Triska, and Torbjörn Lager. Swi-prolog. TPLP, 12(1-2):67–96, 2012. [Wiesler et al., 2014] Simon Wiesler, Alexander Richard, Ralf Schlüter, and Hermann Ney. Mean-normalized stochastic gradient for large-scale deep learning. In IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP 2014, Florence, Italy, May 4-9, 2014, pages 180–184, 2014. [Zantema, 1995] Hans Zantema. Termination of term rewriting by semantic labelling. Fundam. Inform., 24(1/2):89–105, 1995. 

arXiv:1603.08578v2 [math.ST] 21 Jul 2016 Analysis of k-Nearest Neighbor Distances with Application to Entropy Estimation Shashank Singh Barnabás Póczos Carnegie Mellon University, 5000 Forbes Ave., Pittsburgh, PA 15213 USA Abstract Estimating entropy and mutual information consistently is important for many machine learning applications. The Kozachenko-Leonenko (KL) estimator (Kozachenko & Leonenko, 1987) is a widely used nonparametric estimator for the entropy of multivariate continuous random variables, as well as the basis of the mutual information estimator of Kraskov et al. (2004), perhaps the most widely used estimator of mutual information in this setting. Despite the practical importance of these estimators, major theoretical questions regarding their finite-sample behavior remain open. This paper proves finite-sample bounds on the bias and variance of the KL estimator, showing that it achieves the minimax convergence rate for certain classes of smooth functions. In proving these bounds, we analyze finitesample behavior of k-nearest neighbors (k-NN) distance statistics (on which the KL estimator is based). We derive concentration inequalities for k-NN distances and a general expectation bound for statistics of k-NN distances, which may be useful for other analyses of k-NN methods. 1. Introduction Estimating entropy and mutual information in a consistent manner is of importance in a number problems in machine learning. For example, entropy estimators have applications in goodness-of-fit testing (Goria et al., 2005), parameter estimation in semi-parametric models (Wolsztynski et al., 2005), studying fractal random walks (Alemany & Zanette, 1994), and texture classification (Hero et al., 2002a;b). Mutual information estimators have applications in feature selection (Peng & Dind, BAPOCZOS @ CS . CMU . EDU 2005), clustering (Aghagolzadeh et al., 2007), causality detection (Hlaváckova-Schindler et al., 2007), optimal experimental design (Lewi et al., 2007; Póczos & Lőrincz, 2009), fMRI data processing (Chai et al., 2009), prediction of protein structures (Adami, 2004), and boosting and facial expression recognition (Shan et al., 2005). Both entropy estimators and mutual information estimators have been used for independent component and subspace analysis (Learned-Miller & Fisher, 2003; Szabó et al., 2007; Póczos & Lőrincz, 2005; Hulle, 2008), as well as for image registration (Kybic, 2006; Hero et al., 2002a;b). For further applications, see (Leonenko et al., 2008). In this paper, we focus on the problem of estimating the Shannon entropy of a continuous random variable given samples from its distribution. All of our results extend to the estimation of mutual information, since the latter can be written as a sum of entropies. 1 In our setting, we assume we are given n IID samples from an unknown probability measure P . Under nonparametric assumptions (on the smoothness and tail behavior of P ), our task is then to estimate the differential Shannon entropy of P . Estimators of entropy and mutual information come in many forms (as reviewed in Section 2), but one common approach is based on statistics of k-nearest neighbor (kNN) distances (i.e., the distance from a sample to its k th nearest neighbor amongst the samples, in some metric on the space). These nearest-neighbor estimates are largely based on initial work by Kozachenko & Leonenko (1987), who proposed an estimate for differential Shannon entropy and showed its weak consistency. Henceforth, we refer to this historic estimator as the ‘KL estimator’, after its discoverers. Although there has been much work on the problem of entropy estimation in the nearly three decades since the KL estimator was proposed, there are still major open questions about the finite-sample behavior of the KL estimator. The goal of this paper is to address some of these questions in the form of finite-sample bounds on the bias 1 Preprint, Copyright 2016 by the author(s). SSS 1@ ANDREW. CMU . EDU Specifically, for random variables X and Y , I(X; Y ) = H(X) + H(Y ) − H(X, Y ). Analysis of k-Nearest Neighbor Statistics with Application to Entropy Estimation and variance of the estimator. Organization Specifically, our main contributions are the following: Section 2 discusses related work. Section 3 gives theoretical context and assumptions underlying our work. In Section 4, we prove concentration boundss for k-NN distances, and we use these in Section 5 to derive bounds on the expectations of k-NN distance statistics. Section 6 describes the KL estimator, for which we prove bounds on the bias and variance in Sections 7 and 8, respectively.   β/D bounds on the bias of the 1. We derive O (k/n) KL estimate, where β is a measure of the smoothness (i.e., Hölder continuity) of the sampling density, D is the intrinsic dimension of the support of the distribution, and n is the sample size.
2. We derive O n−1 bounds on the variance of the KL estimator. 3. We derive concentration inequalities for k-NN distances, as well as general bounds on expectations of k-NN distance statistics, with important special cases: 2. Related Work Here, we review previous work on the analysis of k-nearest neighbor statistics and their role in estimating information theoretic functionals, as well as other approaches to estimating information theoretic functionals. 2.1. The Kozachenko-Leonenko Estimator of Entropy (a) We bound the moments of k-NN distances, which play a role in analysis of many applications of k-NN methods, including both the bias and variance of the KL estimator. In particular, we significantly relax strong assumptions underlying previous results by Evans et al. (2002), such as compact support and smoothness of the sampling density. Our results are also the first which apply to negative moments (i.e., E [X α ] with α < 0); these are important for bounding the variance of the KL estimator. (b) We give upper and lower bounds on the logarithms of k-NN distances. These are important for bounding the variance of the KL estimator, as well as k-NN estimators for divergences and mutual informations. We present our results in the general setting of a set equipped with a metric, a base measure, a probability density, and an appropriate definition of dimension. This setting subsumes Euclidean spaces, in which k-NN methods have traditionally been analyzed, 2 but also includes, for instance, Riemannian manifolds, and perhaps other spaces of interest. We also strive to weaken some of the restrictive assumptions, such as compact support and boundedness of the density, on which most related work depends. We anticipate that the some of the tools developed here may be used to derive error bounds for k-NN estimators of mutual information, divergences (Wang et al., 2009), their generalizations (e.g., Rényi and Tsallis quantities (Leonenko et al., 2008)), norms, and other functionals of probability densities. We leave such bounds to future work. 2 A recent exception in the context of classification, is Chaudhuri & Dasgupta (2014) which considers general metric spaces. In general contexts, only weak consistency of the KL estimator is known (Kozachenko & Leonenko, 1987). Biau & Devroye (2015) recently reviewed finite-sample results known for the KL estimator. They show (Theorem 7.1) that, if the density p has compact support, then the variance of the KL estimator decays as O(n−1 ). They also claim (Theorem 7.2) to bound the bias of the KL estimator by O(n−β ), under the assumptions that p is βHölder continuous (β ∈ (0, 1]), bounded away from 0, and supported on the interval [0, 1]. However, in their proof Biau & Devroye (2015) neglect the additional bias incurred at the boundaries of [0, 1], where the density cannot simultaneously be bounded away from 0 and continuous. In fact, because the KL estimator does not attempt to correct for boundary bias, for densities bounded away from 0, the estimator may suffer bias worse than O(n−β ). The KL estimator is also important for its role in the mutual information estimator proposed by Kraskov et al. (2004), which we refer to as the KSG estimator. The KSG estimator expands the mutual information as a sum of entropies, which it estimates via the KL estimator with a particular random (i.e., data-dependent) choice of the nearestneighbor parameter k. The KSG estimator is perhaps the most widely used estimator for the mutual information between continuous random variables, despite the fact that it currently appears to have no theoretical guarantees, even asymptotically. In fact, one of the few theoretical results, due to Gao et al. (2015b), concerning the KSG estimator is a negative result: when estimating the mutual information between strongly dependent variables, the KSG estimator tends to systematically underestimate mutual information, due to increased boundary bias. 3 Nevertheless, the 3 To alleviate this, Gao et al. (2015b) provide a heuristic correction based on using local PCA to estimate the support of the distribution. Gao et al. (2015a) provide and prove asymptotic un- Analysis of k-Nearest Neighbor Statistics with Application to Entropy Estimation widespread use of the KSG estimator motivates study of its behavior. We hope that our analysis of the KL estimator, in terms of which the KSG estimator can be written, will lead to a better understanding of the latter. 2.2. Analysis of nearest-neighbor distance statistics Evans (2008) derives a law of large numbers for k-NN statistics with uniformly bounded (central) kurtosis as the sample size n → ∞. Although it is not obvious that the kurtosis of log-k-NN distances is uniformly bounded (indeed, each log-k-NN distance approaches −∞ almost surely), we show in Section 8 that this is indeed the case, and we apply the results of Evans (2008) to bound the variance of the KL estimator. Evans et al. (2002) derives asymptotic limits and convergence rates for moments of k-NN distances, for sampling densities with bounded derivatives and compact domain. In contrast, we use weaker assumptions to simply prove bounds on the moments of k-NN distances. Importantly, whereas the results of Evans et al. (2002) apply only to non-negative moments (i.e., E [|X|α ] with α ≥ 0), our results also hold for certain negative moments, which is crucial for our bounds on the variance of the KL estimator. 2.3. Other Approaches to Estimating Information Theoretic Functionals Analysis of convergence rates: For densities over RD satisfying a Hölder smoothness condition parametrized by β ∈ (0, ∞), the minimax rate for estimating entropy known  has been  since Birge & Massart (1995) to be 8β − min{ 4β+D ,1} O n in mean squared error, where n is the sample size. Quite recently, there has been much work on analyzing new estimators for entropy, mutual information, divergences, and other functionals of densities. Most of this work has been along one of three approaches. One series of papers (Liu et al., 2012; Singh & Poczos, 2014b;a) studied boundary-corrected plug-in approach based on undersmoothed kernel density estimation. This approach has strong finite sample guarantees, but requires prior knowledge of the support of the density and can necessitate computationally demanding numerical integration. A second approach (Krishnamurthy et al., 2014; Kandasamy et al., 2015) uses von Mises expansion to correct the bias of optimally smoothed density estimates. This approach shares the difficulties of the previous approach, but is statistically more efficient. Finally, a long line of work (Pérez-Cruz, 2008; Pál et al., 2010; Sricharan et al., 2012; Sricharan et al., 2010; Moon & Hero, 2014) has studied enbiasedness of another estimator, based on local Gaussian density estimation, that directly adapts to the boundary. tropy estimation based on continuum limits of certain properties of graphs (including k-NN graphs, spanning trees, and other sample-based graphs). Most of 2βthese estimators achieve  rates of  4β Only O n− min{ β+D ,1} or O n− min{ 2β+D ,1} . the von Mises approach of Krishnamurthy et al. (2014) is known to achieve the minimax rate for general β and D, but due to its high computational demand (O(2D n3 )), the authors suggest the use of other statistically less efficient estimators for moderately sized datasets. In this paper, we prove that,for β ∈ (0, 2], the  KL estimator converges 4β at the rate O n− min{ 2β+D ,1} . It is also worth noting the relative of the KL estimator
computational efficiency
(O Dn2 , or O 2D n log n using k-d trees for small D). Boundedness of the density: For all of the above approaches, theoretical finite-sample results known so far assume that the sampling density is lower and upper bounded by positive constants. This also excludes most distributions with unbounded support, and hence, many distributions of practical relevance. A distinctive feature of our results is that they hold for a variety of densities that approach 0 and ∞ on their domain, which may be unbounded. Our bias bounds apply, for example, to densities that decay exponentially, such as Gaussian distributions. To our knowledge, the only previous results that apply to unbounded densities are √those of Tsybakov & van der Meulen (1996), who show n-consistency of a truncated modification of the KL estimate for a class of functions with exponentially decaying tails. In fact, components of our analysis are inspired by Tsybakov & van der Meulen (1996), and some of our assumptions are closely related. Their analysis only applies to the √ case β = 2 and D = 1, for which our results also imply n-consistency, so our results can be seen in some respects as a generalization of this work. 3. Setup and Assumptions While most prior work on k-NN estimators has been restricted to RD , we present our results in a more general setting. This includes, for example, Riemannian manifolds embedded in higher dimensional spaces, in which case we note that our results depend on the intrinsic, rather than extrinsic, dimension. Such data can be better behaved in their native space than when embedded in a lower dimensional Euclidean space (e.g., working directly on the unit circle avoids boundary bias caused by mapping data to the interval [0, 2π]). Definition 1. (Metric Measure Space): A quadruple (X, d, Σ, µ) is called a metric measure space if X is a set, d : X × X → [0, ∞) is a metric on X, Σ is a σ-algebra on X containing the Borel σ-algebra induced by d, and µ : Σ → [0, ∞] is a σ-finite measure on the measurable Analysis of k-Nearest Neighbor Statistics with Application to Entropy Estimation space (X, Σ). Definition 2. (Dimension): A metric measure space (X, d, Σ, µ) is said to have dimension D ∈ [0, ∞) if there exist constants cD , ρ > 0 such that, ∀r ∈ [0, ρ], x ∈ X , µ(B(x, r)) = cD rD . 4 Definition 3. (Full Dimension): Given a metric measure space (X, d, Σ, µ) of dimension D, a measure P on (X, Σ) is said to have full dimension on a set X ⊆ X if there exist functions γ∗ , γ ∗ : X → (0, ∞) such that, for all r ∈ [0, ρ] and µ-almost all x ∈ X , γ∗ (x)rD ≤ P (B(x, r)) ≤ γ ∗ (x)rD . Remark 4. If X = RD , d is the Euclidean metric, and µ is the Lebesgue measure, then the dimension of the metric measure space is D. However, if X is a lower dimensional subspace of RD , then the dimension may be less than D. For example, if X = SD−1 := {x ∈ RD : kxk2 = 1}), d is the geodesic distance on SD−1 , and µ is the (D − 1)dimensional surface measure, then the dimension is D − 1. Remark 5. In previous work on k-NN statistics (Evans et al., 2002; Biau & Devroye, 2015) and estimation of information theoretic functionals (Sricharan et al., 2010; Krishnamurthy et al., 2014; Singh & Poczos, 2014b; Moon & Hero, 2014), it has been common to make the assumption that the sampling distribution has full dimension with constant γ∗ and γ ∗ (or, equivalently, that the density is lower and upper bounded by positive constants). This excludes distributions with densities approaching 0 or ∞ on their domain, and hence also densities with unbounded support. By letting γ∗ and γ ∗ be functions, our results extend to unbounded densities that instead satisfy certain tail bounds. In order to ensure that entropy is well defined, we assume that P is a probability measure absolutely continuous with respect to µ, and that its probability density function p : X → [0, ∞) satisfies 5 Z H(p) := E [log p(X)] = p(x) log p(x) dµ(x) ∈ R. X∼P X (1) Finally, we assume we have n + 1 samples X, X1 , ..., Xn drawn IID from P . We would like to use these samples to estimate the entropy H(p) as defined in Equation (1). Our analysis and methods relate to the k-nearest neighbor distance εk (x), defined for any x ∈ X by εk (x) = d(x, Xi ), where Xi is the k th -nearest neighbor of x in 4 Here and in what follows, B(x, r) := {y ∈ X : d(x, y) < r} denotes the open ball of radius r centered at x. 5 See (Baccetti & Visser, 2013) for discussion of sufficient conditions for H(p) < ∞. the set {X1 , ..., Xn }. Note that, since the definition of dimension used precludes the existence of atoms (i.e., for all x ∈ X , p(x) = µ({x}) = 0), εk (x) > 0, µ-almost everywhere. This is important, since we will study log εk (x). Initially (i.e., in Sections 4 and 5), we will study log εk (x) with fixed x ∈ X , for which we will derive bounds in terms of γ∗ (x) and γ ∗ (x). When we apply these results to analyze the KL estimator in Section 7 and 8, we will need to take expectations such as E [log εk (X)] (for which we reserve the extra sample X), leading to ‘tail bounds’ on p in terms of the functions γ∗ and γ ∗ . 4. Concentration of k-NN Distances We begin with a consequence of the multiplicative Chernoff bound, asserting a sort of concentration of the distance of any point in X from its k th -nearest neighbor in {X1 , . . . , Xn }. Since the results of this section are concerned with fixed x ∈ X , for notational simplicity, we suppress the dependence of γ∗ and γ ∗ on x. Lemma 6. Let (X, d, Σ, µ) be a metric measure space of dimension D. Suppose P is an absolutely continuous probability measure with full dimension on X ⊆ X and density function p : X → [0, ∞). For x ∈ X , if 1/D   k r∈ , ρ , then γ∗ n P [εk (x) > r] ≤ e−γ∗ r  and, if r ∈ 0, min  k γ∗n D 1/D P [εk (x) ≤ r] ≤  n  k γ∗ r D n e . k ,ρ  eγ ∗ rD n k , then kγ∗ /γ ∗ . 5. Bounds on Expectations of KNN Statistics Here, we use the concentration bounds of Section 4 to bound expectations of functions of k-nearest neighbor distances. Specifically, we give a simple formula for deriving bounds that applies to many functions of interest, including logarithms and (positive and negative) moments. As in the previous section, the results apply to a fixed x ∈ X , and we continue to suppress the dependence of γ∗ and γ ∗ on x. Theorem 7. Let (X , d, Σ, µ) be a metric measure space of dimension D. Suppose P is an absolutely continuous probability measure with full dimension and density func- Analysis of k-Nearest Neighbor Statistics with Application to Entropy Estimation tion p : X → [0, ∞) that satisfies the tail condition 6  Z ∞  n CT −1 ≤ 1 − P (B(X, f (r))) E n X∼P ρ f (x) = xα for certain α, as we will use these bounds when analyzing the KL estimator. (2) for some constant CT > 0. Suppose f : (0, ∞) → R is continuously differentiable, with f ′ > 0. Fix x ∈ X . Then, we have the upper bound   D1 ! CT k + (3) E [f+ (εk (x))] ≤ f+ γ∗ n n   D1 ! Z ∞ 1 y (e/k)k dy e−y y k+ D −1 f ′ + 1 nγ∗ D(nγ∗ ) D k and the lower bound E [f− (εk (x))] ≤ f− +  enγ ∗ k  k γ∗n 1 ∗ Z  kγ ( γ ∗kn ) D γ∗ y 1/D ! Dkγ∗ /γ ∗ + ′ CT n f (y) dy When f (x) = log(x), (3) gives       e k Γ(k, k) k 1 + log+ E log+ (εk (x)) ≤ D γ∗ n k D    k 1 1 + log+ (5) ≤ D γ∗ n R∞ (where Γ(s, x) := x ts−1 e−t dt denotes the upper incomplete Gamma function, and we used the bound Γ(s, x) ≤ xs−1 e−x ), and (4) gives     k 1 + C1 , (6) log log (ε (x)) ≤ E − − k D γ∗n for C1 = α E [εk (x)] (4) 0 Remark 9. The tail condition (2) is difficult to validate directly for many distributions. Clearly, it is satisfied when the support of p is bounded. However, (Tsybakov & van der Meulen, 1996) show that, for the functions f we are interested in (i.e., logarithms and power functions), when X = RD , d is the Euclidean metric, and µ is the Lebesgue measure, (2) is also satisfied by upper-bounded densities with exponentially decreasing tails. More precisely, that is when there exist a, b, α, δ > 0 and β > 1 such that, whenever kxk2 > δ, β β ae−αkxk ≤ p(x) ≤ be−αkxk , which permits, for example, Gaussian distributions. It should be noted that the constant CT depends only on the metric measure space, the distribution P , and the function f , and, in particular, not on k. 5.1. Applications of Theorem 7 We can apply Theorem 7 to several functions f of interest. Here, we demonstrate the cases f (x) = log x and  . For α > 0, f (x) = xα , (3) gives k γ∗ n   Dα k γ∗ n +  Dα  e k αΓ (k + α/D, k) k D(nγ∗ )α/D , (7) α where C2 = 1 + 2 D . For any α ∈ [−Dkγ∗ /γ ∗ , 0], when α f (x) = −x , (4) gives α E [εk (x)] ≤ C3  k γ∗n  Dα , (8) ∗ where C3 = 1 + αγ ∗ ekγ∗ /γ Dkγ∗ +αγ ∗ . 6. The KL Estimator for Entropy Recall that, for a random variable X sampled from a probability density p with respect to a base measure µ, the Shannon entropy is defined as Z H(X) = − p(x) log p(x) dx. X As discussed in Section 1, many applications call for estimate of H(X) given n IID samples X1 , . . . , Xn ∼ p. For a positive integer k, the KL estimator is typically written as Ĥk (X) = ψ(n) − ψ(k) + log cD + n DX log εk (Xi ), n i=1 where ψ : N → R denotes the digamma function. The motivating insight is the observation that, independent of the sampling distribution, 7 6 Since f need not be surjective, we use the generalized inverse f : R → [0, ∞] defined by f −1 (ε) := inf{x ∈ (0, ∞) : f (x) ≥ ε}. ≤ ∗ ≤ C2 (f+ (x) = max{0, f (x)} and f− (x) = − min{0, f (x)} denote the positive and negative parts of f , respectively). Remark 8. If f : (0, ∞) → R is continuously differentiable with f ′ < 0, we can apply Theorem 7 to −f . Also, similar techniques can be used to prove analogous lower bounds (i.e., lower bounds on the positive part and upper bounds on the negative part). γ ∗ ekγ∗ /γ Dkγ∗ E [log P (B(Xi , εk (Xi )))] = ψ(k) − ψ(n), −1 7 See (Kraskov et al., 2004) for a concise proof of this fact. Analysis of k-Nearest Neighbor Statistics with Application to Entropy Estimation Hence, h i E Ĥk (X) " = E − log P (B(Xi , εk (Xi ))) + log cD + Definition 11. Given a constant β > 0 and an open set X ⊆ RD , a function f : X → R is called β-Hölder continuous if f is ℓ times differentiable and there exists L > 0 #such that, for any multi-index α ∈ ND with |α| < β, n DX log εk (Xi ) n i=1  n 1X P (B(xi , εk (Xi ))) = −E log n i=1 cD ε D k (Xi ) # " n   1X log pεk (i) (Xi ) = − E log pεk (X1 ) (X1 ) , = −E n i=1 " # where, for any x ∈ X , ε > 0, Z 1 P (B(x, ε)) pε (x) = p(y) dµ(y) = D cD ε cD ε D B(x,ε) sup x6=y∈X where ℓ := ⌊β⌋ is the greatest integer strictly less than β. Definition 12. Given an open set X ⊆ RD and a function f : X → R, f is said to vanish on the boundary ∂X of X if, ′ for any sequence {xi }∞ i=1 in X with inf x′ ∈∂X kx − x k2 → 0 as i → ∞, f (x) → 0 as i → ∞. Here, ∂X := {x ∈ RD : ∀δ > 0, B(x, δ) 6⊆ X and B(x, δ) 6⊆ X c }, denotes the local average of p in a ball of radius ε around x. Since pε is a smoothed approximation of p (with smoothness increasing with ε), the KL estimate can be intuitively thought of as a plug-in estimator for H(X), using a density estimate with an adaptive smoothing parameter. In the next two sections, we utilize the bounds derived in Section 5 to bound the bias and variance of the KL estimator. We note that, for densities in the β-Hölder smoothness class (β ∈ (0, 2]), our results imply a mean-squared error of O(n−2β/D ) when β < D/2 and O(n−1 ) when β ≥ D/2. 7. Bias Bound In this section, we prove bounds on the bias of the KL estimator, first in a relatively general setting, and then, as a corollary, in a more specific but better understood setting. Theorem 10. Suppose (X, d, Σ, µ) and P satisfy the conditions of Theorem 7, and there exist C, β ∈ (0, ∞) with denotes the boundary of X . Corollary 13. Consider the metric measure space (RD , d, Σ, µ), where d is Euclidean and µ is the Lebesgue measure. Let P be an absolute continuous probability measure with full dimension and density p supported on an open set X ⊆ RD . Suppose p satisfies (9) and the conditions of Theorem 7 and is β-Hölder continuous (β ∈ (0, 2]) with constant L. Assume p vanishes on ∂X . If β > 1, assume k∇pk2 vanishes on ∂X . Then, h  n − Dβ i , E Ĥk (X) − H(X) ≤ CH k LD where CH = (1 + cD )C2 Γ D+β . Remark 14. The assumption that p (and perhaps k∇pk) vanish on the boundary of X can be thought of as ensuring that the trivial continuation q : RD → [0, ∞) q(x) = sup |p(x) − pε (x)| ≤ Cβ εβ , x∈X and suppose p satisfies a ‘tail bound’ h i − β+D ΓB := E (γ∗ (X)) D < ∞. |Dα f (x) − Dα f (y)| ≤ L, kx − ykβ−ℓ  p(x) 0 x∈X x ∈ RD \X of p to RD is β-Hölder continuous. This reduces boundary bias, for which the KL estimator does not correct. 8 (9) X∼P Then,   Dβ h i k , H(X) − Ĥ (X) ≤ C E k B n where CB = (1 + cD )C2 Cβ ΓB . We now show that the conditions of Theorem 10 are satisfied by densities in the commonly used nonparametric class of β-Hölder continuous densities on RD . 8. Variance Bound We first use the bounds proven in Section 5 to prove uniform (in n) bounds on the moments of E [log εk (X)]. We the for any fixed x ∈ X , although log εk (x) → −∞ almost surely as n → ∞, V [log εk (x)], and indeed all higher central moments of log εk (x), are bounded, uniformly in n. In fact, there exist exponential bounds, independent of n, on the density of log εk (x) − E [log εk (x)]. 8 Several estimators controlling for boundary bias have been proposed (e.g., Sricharan et al. (2010) give a modified k-NN estimator that accomplishes this without prior knowledge of X . Analysis of k-Nearest Neighbor Statistics with Application to Entropy Estimation 8.1. Moment Bounds on Logarithmic k-NN distances Lemma 15. Suppose (X, d, Σ, µ) and P satisfy the conditions of Theorem 7. Suppose also that Γ0 :=   γ ∗ (x) Dk supx∈X γ∗ (x) < ∞. Let λ ∈ 0, Γ0 and assume the following expectations are finite:  ∗  γ (X) Γ := E < ∞. (10) X∼P γ∗ (X) i h −λ/D < ∞. (11) Γ∗ (λ) := E (γ∗ (X)) X∼P i h Γ∗ (λ) := E (γ ∗ (X))λ/D < ∞. (12) X∼P 9. Bounds on the Mean Squared Error The bias and variance bounds (Theorems 10 and 18) imply a bound on the mean squared error of the KL estimator: Corollary 20. Suppose p 1. is β-Hölder continuous with β ∈ (0, 2]. 2. vanishes on ∂X . If β > 1, then also suppose k∇pk2 vanishes on ∂X . Then, for any integer ℓ > 1, the ℓth central moment i h ℓ Mℓ := E (log εk (X) − E [log εk (X)]) 3. satisfies Mℓ ≤ CM ℓ!/λℓ , Remark 19. Nk depends only on k and the geometry of the metric space (X , d). For example, Corollary A.2 of Evans (2008) shows that, when X = RD and d is the Euclidean metric, then Nk ≤ kK(D), where K(D) is the kissing number of Rd . (13) where CM > 0 is a constant independent of n, ℓ, and λ. Remark 16. The conditions (10), (11), and (12) are mild. For example, when X = RD , d is the Euclidean metric, and µ is the Lebesgue measure, it suffices that p is LipsD2 chitz continuous 9 and there exist c, r > 0, p > D−α such −p that p(x) ≤ ckxk whenever kxk2 > r. The condition Γ0 < ∞ is more prohibitive, but still permits many (possibly unbounded) distributions of interest. Remark 17. If the terms log εk (Xi ) were independent, a Bernstein inequality, together with the moment bound (13) would imply a sub-Gaussian concentration bound on the KL estimator about its expectation. This may follow from one of several more refined concentration results relaxing the independence assumption that have been proposed. 8.2. Bound on the Variance of the KL Estimate Bounds on the variance of the KL estimator now follow from the law of large numbers in Evans (2008) (itself an application of the Efron-Stein inequality to k-NN statistics). Theorem 18. Suppose (X, d, Σ, µ) and P satisfy the conditions of Lemma 15, and that that there exists a constant Nk ∈ N such that, for any finite F ⊆ X , any x ∈ F can be among the k-NN hof at most i Nk other points in that set. Then, Ĥk (X) → E Ĥk (X) almost surely (as n → ∞), and, for n ≥ 16k and M4 satisfying (13).   i 5(3 + kN )(3 + 64k)M h 1 k 4 , ∈O V Ĥk (X) ≤ n nk 9 Significantly milder conditions than Lipschitz continuity suffice, but are difficult to state here due to space limitations. [TODO: Other assumptions.] satisfies the assumptions of Theorems 10 and 18. Then,  2β/D  2  k CV 2 ≤ CB + . (14) E Ĥk (X) − H(X) n nk If we let k scale as k ≍ nmax{0, 2β+D } this gives an overall convergence rate of  2β/D  2  k CV 2 + . (15) ≤ CB E Ĥk (X) − H(X) n nk 2β−D 10. Conclusions and Future Work This paper derives finite sample bounds on the bias and variance of the KL estimator under general conditions, including for certain classes of unbounded distributions. As intermediate results, we proved concentration inequalities for k-NN distances and bounds on the expectations of statistics of k-NN distances. We hope these results and methods may lead to convergence rates for the widely used KSG mutual information estimator, or to generalize convergence rates for other estimators of entropy and related functionals to unbounded distributions. Acknowledgements This material is based upon work supported by a National Science Foundation Graduate Research Fellowship to the first author under Grant No. DGE-1252522. References Adami, C. Information theory in molecular biology. Physics of Life Reviews, 1:3–22, 2004. Analysis of k-Nearest Neighbor Statistics with Application to Entropy Estimation Aghagolzadeh, M., Soltanian-Zadeh, H., Araabi, B., and Aghagolzadeh, A. A hierarchical clustering based on mutual information maximization. In in Proc. of IEEE International Conference on Image Processing, pp. 277– 280, 2007. Alemany, P. A. and Zanette, D. H. Fractal random walks from a variational formalism for Tsallis entropies. Phys. Rev. E, 49(2):R956–R958, Feb 1994. doi: 10.1103/ PhysRevE.49.R956. Baccetti, Valentina and Visser, Matt. Infinite shannon entropy. Journal of Statistical Mechanics: Theory and Experiment, 2013(04):P04010, 2013. Biau, Gérard and Devroye, Luc. Entropy estimation. In Lectures on the Nearest Neighbor Method, pp. 75–91. Springer, 2015. Birge, L. and Massart, P. Estimation of integral functions of a density. A. Statistics, 23:11–29, 1995. Chai, B., Walther, D. B., Beck, D. M., and Fei-Fei, L. Exploring functional connectivity of the human brain using multivariate information analysis. In NIPS, 2009. Chaudhuri, Kamalika and Dasgupta, Sanjoy. Rates of convergence for nearest neighbor classification. In Advances in Neural Information Processing Systems, pp. 3437– 3445, 2014. Evans, D. A law of large numbers for nearest neighbor statistics. In Proceedings of the Royal Society, volume 464, pp. 3175–3192, 2008. Evans, Dafydd, Jones, Antonia J, and Schmidt, Wolfgang M. Asymptotic moments of near–neighbour distance distributions. In Proceedings of the Royal Society of London A: Mathematical, Physical and Engineering Sciences, volume 458, pp. 2839–2849. The Royal Society, 2002. Gao, Shuyang, Steeg, Greg Ver, and Galstyan, Aram. Estimating mutual information by local gaussian approximation. arXiv preprint arXiv:1508.00536, 2015a. Gao, Shuyang, Ver Steeg, Greg, and Galstyan, Aram. Efficient estimation of mutual information for strongly dependent variables. In Proceedings of the Eighteenth International Conference on Artificial Intelligence and Statistics, pp. 277–286, 2015b. Goria, M. N., Leonenko, N. N., Mergel, V. V., and Inverardi, P. L. Novi. A new class of random vector entropy estimators and its applications in testing statistical hypotheses. J. Nonparametric Statistics, 17:277–297, 2005. Hero, A. O., Ma, B., Michel, O., and Gorman, J. Alphadivergence for classification, indexing and retrieval, 2002a. Communications and Signal Processing Laboratory Technical Report CSPL-328. Hero, A. O., Ma, B., Michel, O. J. J., and Gorman, J. Applications of entropic spanning graphs. IEEE Signal Processing Magazine, 19(5):85–95, 2002b. Hlaváckova-Schindler, K., Paluŝb, M., Vejmelkab, M., and Bhattacharya, J. Causality detection based on information-theoretic approaches in time series analysis. Physics Reports, 441:1–46, 2007. Hulle, M. M. Van. Constrained subspace ICA based on mutual information optimization directly. Neural Computation, 20:964–973, 2008. Kandasamy, Kirthevasan, Krishnamurthy, Akshay, Poczos, Barnabas, Wasserman, Larry, et al. Nonparametric von mises estimators for entropies, divergences and mutual informations. In Advances in Neural Information Processing Systems, pp. 397–405, 2015. Kozachenko, L. F. and Leonenko, N. N. A statistical estimate for the entropy of a random vector. Problems of Information Transmission, 23:9–16, 1987. Kraskov, A., Stögbauer, H., and Grassberger, P. Estimating mutual information. Phys. Rev. E, 69:066138, 2004. Krishnamurthy, A., Kandasamy, K., Poczos, B., and Wasserman, L. Nonparametric estimation of renyi divergence and friends. In International Conference on Machine Learning (ICML), 2014. Kybic, J. Incremental updating of nearest neighbor-based high-dimensional entropy estimation. In Proc. Acoustics, Speech and Signal Processing, 2006. Learned-Miller, E. G. and Fisher, J. W. ICA using spacings estimates of entropy. J. Machine Learning Research, 4: 1271–1295, 2003. Leonenko, N., Pronzato, L., and Savani, V. A class of Rényi information estimators for multidimensional densities. Annals of Statistics, 36(5):2153–2182, 2008. Lewi, J., Butera, R., and Paninski, L. Real-time adaptive information-theoretic optimization of neurophysiology experiments. In Advances in Neural Information Processing Systems, volume 19, 2007. Liu, H., Lafferty, J., and Wasserman, L. Exponential concentration inequality for mutual information estimation. In Neural Information Processing Systems (NIPS), 2012. Analysis of k-Nearest Neighbor Statistics with Application to Entropy Estimation Moon, Kevin R and Hero, Alfred O. Ensemble estimation of multivariate f-divergence. In Information Theory (ISIT), 2014 IEEE International Symposium on, pp. 356–360. IEEE, 2014. Pál, D., Póczos, B., and Szepesvári, Cs. Estimation of Rényi entropy and mutual information based on generalized nearest-neighbor graphs. In Proceedings of the Neural Information Processing Systems, 2010. Peng, H. and Dind, C. Feature selection based on mutual information: Criteria of max-dependency, maxrelevance, and min-redundancy. IEEE Trans On Pattern Analysis and Machine Intelligence, 27, 2005. Pérez-Cruz, F. Estimation of information theoretic measures for continuous random variables. In Advances in Neural Information Processing Systems 21, 2008. Póczos, B. and Lőrincz, A. Independent subspace analysis using geodesic spanning trees. In ICML, pp. 673–680, 2005. Póczos, B. and Lőrincz, A. Identification of recurrent neural networks by Bayesian interrogation techniques. J. Machine Learning Research, 10:515–554, 2009. Shan, C., Gong, S., and Mcowan, P. W. Conditional mutual information based boosting for facial expression recognition. In British Machine Vision Conference (BMVC), 2005. Singh, S. and Poczos, B. Exponential concentration of a density functional estimator. In Neural Information Processing Systems (NIPS), 2014a. Singh, S. and Poczos, B. Generalized exponential concentration inequality for Rényi divergence estimation. In International Conference on Machine Learning (ICML), 2014b. Sricharan, K., Raich, R., and Hero, A. Empirical estimation of entropy functionals with confidence. Technical Report, http://arxiv.org/abs/1012.4188, 2010. Sricharan, K., Wei, D., and Hero, A. Ensemble estimators for multivariate entropy estimation, 2012. http://arxiv.org/abs/1203.5829. Szabó, Z., Póczos, B., and Lőrincz, A. Undercomplete blind subspace deconvolution. J. Machine Learning Research, 8:1063–1095, 2007. Tsybakov, A. B. and van der Meulen, E. C. Root-n consistent estimators of entropy for densities with unbounded support. Scandinavian J. Statistics, 23:75–83, 1996. Tsybakov, A.B. Introduction to Nonparametric Estimation. Springer Publishing Company, Incorporated, 1st edition, 2008. ISBN 0387790519, 9780387790510. Wang, Q., Kulkarni, S.R., and Verdú, S. Divergence estimation for multidimensional densities via k-nearestneighbor distances. IEEE Transactions on Information Theory, 55(5), 2009. Wolsztynski, E., Thierry, E., and Pronzato, L. Minimumentropy estimation in semi-parametric models. Signal Process., 85(5):937–949, 2005. ISSN 0165-1684. doi: http://dx.doi.org/10.1016/j.sigpro.2004.11.028. 

arXiv:1405.4672v1 [math.AT] 19 May 2014 Homology of torus spaces with acyclic proper faces of the orbit space Anton Ayzenberg Abstract. Let X be 2n-dimensional compact manifold and T n ñ X be a locally standard action of a compact torus. The orbit space X{T is a manifold with corners. Suppose that all proper faces of X{T are acyclic. In the paper we study ˚ the homological spectral sequence XE ˚,˚ ñ H˚ pXq corresponding to the filtration of X by orbit types. When the free part of the action is not twisted, we describe ˚ the whole spectral sequence XE ˚,˚ in terms of homology and combinatorial structure of the orbit space X{T . In this case we describe the kernel and the cokernel of the natural map krX{T s{pl.s.o.p.q Ñ H ˚ pXq, where krX{T s is a face ring of X{T and pl.s.o.p.q is the ideal generated by a linear system of parameters (this ideal appears as the image of H ą0 pBT q in HT˚ pXq). There exists a natural double grading on H˚ pXq, which satisfies bigraded Poincare duality. This general theory is applied to compute homology groups of origami toric manifolds with acyclic proper faces of the orbit space. A number of natural generalizations is considered. These include Buchsbaum simplicial complexes and posets. h1 - and h2 -numbers of simplicial posets appear as the ranks of certain terms in the spectral sequence X ˚ E ˚,˚ . In particular, using topological argument we show that Buchsbaum posets have nonnegative h2 -vectors. The proofs of this paper rely on the theory of cellular sheaves. We associate to a torus space certain sheaves and cosheaves on the underlying simplicial poset, and observe an interesting duality between these objects. This duality seems to be a version of Poincare–Verdier duality between cellular sheaves and cosheaves. Contents 1. 2. 3. 4. 5. 6. Introduction Preliminary constructions Buchsbaum pseudo-cell complexes Torus spaces over Buchsbaum pseudo-cell complexes Main results Duality between certain cellular sheaves and cosheaves The author is supported by the JSPS postdoctoral fellowship program. 1 2 4 7 15 21 26 HOMOLOGY OF TORUS SPACES 7. Face vectors and ranks of border components 8. Geometry of equivariant cycles 9. Examples and calculations 10. Concluding remarks Acknowledgements References 2 30 35 40 43 44 45 1. Introduction Let M be 2n-dimensional compact manifold with a locally standard action of a compact torus T n . This means, by definition, that the action of T n on M 2n is locally modeled by a standard coordinate-wise action of T n on Cn . Since Cn {T n can be identified with a nonnegative cone Rně , the quotient space Q “ M{T has a natural structure of a compact manifold with corners. The general problem is the following: Problem 1. Describe the (co)homology of M in terms of combinatorics and topology of the orbit space Q and the local data of the action. The answer is known in the case when Q and all its faces are acyclic (so called homology polytope) [9]. In this case the equivariant cohomology ring of M coincides with the face ring of simplicial poset SQ dual to Q, and the ordinary cohomology has description similar to that of toric varieties or quasitoric manifolds: H ˚ pM; kq – krSQ s{pθ1 , . . . , θn q, deg θi “ 2. In this case krSQ s is Cohen–Macaulay and θ1 , . . . , θn is a linear regular sequence determined by the characteristic map on Q. In particular, cohomology vanishes in odd degree, and dim H 2i pMq “ hi pSQ q. In general, there is a topological model of a manifold M, called canonical model. The construction is the following. Start with a nice manifold with corners Q, consider a principal T n -bundle Y over Q, and then consider the quotient space X “ Y {„ determined by a characteristic map [17, def. 4.2]. It is known that X is a manifold with locally standard torus action, and every manifold with l.s.t.a. is equivariantly homeomorphic to such canonical model. Thus it is sufficient to work with canonical models to answer Problem 1. In this paper we study the case when all proper faces of Q are acyclic, but Q itself may be arbitrary. Homology of X can be described by the spectral sequence X r E ˚,˚ associated to the filtration of X by orbit types: (1.1) X0 Ă X1 Ă . . . Ă Xn´1 Ă Xn “ X, dim Xi “ 2i. This filtration is covered by the filtration of Y : (1.2) Y0 Ă Y1 Ă . . . Ă Yn´1 Ă Yn “ Y, Xi “ Yi {„ . HOMOLOGY OF TORUS SPACES 2 3 We prove that most entries of the second page XE ˚,˚ coincide with corresponding 2 entries of Y E ˚,˚ (Theorem 1). When Y is a trivial T n -bundle, Y “ Q ˆ T n , this ˚ observation allows to describe XE ˚,˚ completely in terms of topology and combinatorics of Q (Theorem 2, statement 5.3 and Theorem 3). This answers Problem 1 additively. From this description in particular follows that Betti numbers of X do not depend on the choice of characteristic map. We hope, that this technic will lead to the description of cohomology multiplication in H ˚ pXq as well. Another motivation for this paper comes from a theory of Buchsbaum simplicial complexes and posets. The notions of h1 - and h2 -vectors of simplicial poset S first appeared in combinatorial commutative algebra [15, 13]. These invariants emerge naturally in the description of homology of X (Theorems 3 and 4). The space X “ Y { „ can be constructed not only in the case when Q is a manifold with corners (“manifold case”), but also in the case when Q is a cone over geometric realization of simplicial poset S (“cone case”). In the cone case, surely, X may not be a manifold. But there still exists filtration (1.1), and homology groups of X can be calculated by the same method as for manifolds, when S is Buchsbaum. In 8 the cone case we prove that dim XE i,i “ h2 pSq (Theorem 5). Thus, in particular, h2 -vector of Buchsbaum simplicial poset is nonnegative. This result is proved in commutative algebra by completely different methods [13]. The exposition of the paper is built in such way that both situations: manifolds with acyclic faces, and cones over Buchsbaum posets are treated in a common context. In order to do this we introduce the notion of Buchsbaum pseudo-cell complex which is very natural and includes both motivating examples. A theory of cellular sheaves over simplicial posets is used to prove basic theorems. The coincidence of r r most parts of XE ˚,˚ and Y E ˚,˚ follows from the Key lemma (lemma 5.1) which is an instance of general duality between certain sheaves and cosheaves (Theorem 6). In the manifold case this duality can be deduced from Verdier duality for cellular sheaves, described in [6]. The paper is organized as follows. Section 2 contains preliminaries on simplicial posets and cellular sheaves. In section 3 we introduce the notion of simple pseudocell complex and describe spectral sequences associated to filtrations by pseudo-cell skeleta. Section 4 is devoted to torus spaces over pseudo-cell complexes. The main results (Theorems 1–5) are stated in section 5. The rest of section 5 contains the description of homology of X. There is an additional grading on homology groups, and in the manifold case there is a bigraded Poincare duality. Section 6 contains a sheaf-theoretic discussion of the subject. In this section we prove Theorem 6 which can be considered as a version of cellular Verdier duality. This proves the Key lemma, from which follow Theorems 1 and 2. Section 7 is devoted to the combinatorics of simplicial posets. In this section we recall combinatorial definitions of f -, h-, h1 and h2 -vectors and prove Theorems 3–5. The structure of equivariant cycles and cocycles of a manifold X with locally standard torus action is the subject of section HOMOLOGY OF TORUS SPACES 4 8. There exists a natural map krSs{pθ1 , . . . , θn q Ñ H ˚pXq, where pθ1 , . . . , θn q is a linear system of parameters, associated to a characteristic map. In general (i.e. when Q is not a homology polytope) this map may be neither injective nor surjective. The kernel of this map is described by corollary 8.8. The calculations for some particular examples are gathered in section 9. The main family of nontrivial examples is the family of origami toric manifolds with acyclic proper faces of the orbit space. 2. Preliminary constructions 2.1. Preliminaries on simplicial posets. First, recall several standard definitions. A partially ordered set (poset in the following) is called simplicial if it has a minimal element ∅ P S, and for any I P S, the lower order ideal SďI “ tJ | J ď Iu is isomorphic to the boolean lattice 2rks (the poset of faces of a simplex). The number k is called the rank of I P S and is denoted |I|. Also set dim I “ |I| ´ 1. A vertex is a simplex of rank 1 (i.e. the atom of the poset); the set of all vertices is denoted by VertpSq. A subset L Ă S, for which I ă J, J P L implies I P L is called a simplicial subposet. The notation I ăi J is used whenever I ă J and |J| ´ |I| “ i. If S is a simplicial poset, then for each I ă2 J P SQ there exist exactly two intermediate simplices J 1 , J 2: (2.1) I ă1 J 1 , J 2 ă1 J. For simplicial poset S a “sign convention” can be chosen. It means that we can associate an incidence number rJ : Is “ ˘1 to any I ă1 J P S in such way that for (2.1) holds (2.2) rJ : J 1 s ¨ rJ 1 : Is ` rJ : J 2 s ¨ rJ 2 : Is “ 0. The choice of a sign convention is equivalent to the choice of orientations of all simplices. For I P S consider the link: lkS I “ tJ P S | J ě Iu. It is a simplicial poset with minimal element I. On the other hand, lkS I can also be considered as a subset of S. It can be seen that Sz lkS I is a simplicial subposet. Note, that lkS ∅ “ S. Let S 1 be the barycentric subdivision of S. By definition, S 1 is a simplicial complex on the set Sz∅ whose simplices are the chains of elements of S. By definition, the geometric realization of S is the geometric realization of its barycentric subdividef sion |S| “ |S 1 |. One can also think about |S| as a CW-complex with simplicial cells [2]. A poset S is called pure if all its maximal elements have equal dimensions. A poset S is pure whenever S 1 is pure. HOMOLOGY OF TORUS SPACES 5 Definition 2.1. Simplicial complex K of dimension n ´ 1 is called Buchsbaum r if Hi plkK Iq “ 0 for all ∅ ‰ I P K and i ‰ n ´ 1 ´ |I|. If K is Buchsbaum and, r i pKq “ 0 for i ‰ n ´ 1 then K is called Cohen–Macaulay. Simplicial moreover, H poset S is called Buchsbaum (Cohen–Macaulay) if S 1 is a Buchsbaum (resp. Cohen– Macaulay) simplicial complex. Remark 2.2. Whenever the coefficient ring in the notation of (co)homology is omitted it is supposed to be the ground ring k, which is either a field or the ring of integers. r i plkS Iq “ 0 for all Remark 2.3. By [13, Sec.6], S is Buchsbaum whenever H r i plkS Iq “ 0 for ∅ ‰ I P S and i ‰ n ´ 1 ´ |I|. Similarly, S is Cohen–Macaulay if H all I P S and i ‰ n ´ 1 ´ |I|. A poset S is Buchsbaum whenever all its proper links are Cohen–Macaulay. One easily checks that Buchsbaum property implies purity. 2.2. Cellular sheaves. Let MODk be the category of k-modules. The notation dim V is used for the rank of a k-module V . Each simplicial poset S defines a small category CATpSq whose objects are the elements of S and morphisms — the inequalities I ď J. A cellular sheaf [6] (or a stack [12], or a local coefficient system elsewhere) is a covariant functor A : CATpSq Ñ MODk . We simply call A a sheaf on S and hope that this will not lead to a confusion, since different meanings of this word do not appear in the paper. The maps ApJ1 ď J2 q are called the restriction maps. The cochain complex pC ˚ pS; Aq, dq is defined as follows: à i à ApIq, C ˚ pS; Aq “ C pS; Aq, C i pS; Aq “ iě´1 dim I“i d : C i pS; Aq Ñ C i`1 pS; Aq, d“ Iă1 à rI 1 : IsApI ď I 1 q. I 1 ,dim I“i By the standard argument involving sign convention (2.2), d2 “ 0, thus pC ˚ pK; Aq, dq is a differential complex. Define the cohomology of A as the cohomology of this complex: (2.3) def H ˚ pS; Aq “ H ˚ pC ˚ pS; Aq, dq. Remark 2.4. Cohomology of A defined this way coincide with any other meaningful definition of cohomology. E.g. the derived functors of the functor of global sections are isomorphic to (2.3) (see [6] for the vast exposition of this subject). A sheaf A on S can be restricted to a simplicial subposet L Ă S. The complexes pC ˚ pL, Aq, dq and pC ˚ pS; Aq{C ˚pL; Aq, dq are defined in a usual manner. The latter complex gives rise to a relative version of sheaf cohomology: H ˚pS, L; Aq. HOMOLOGY OF TORUS SPACES 6 Remark 2.5. It is standard in topological literature to consider cellular sheaves which do not take values on ∅ P S, since in general this element has no geometrical meaning. However, this extra value Ap∅q is very important in the considerations of this paper. Thus the cohomology group may be nontrivial in degree dim ∅ “ ´1. If a sheaf A is defined on S, then we often consider its truncated version A which coincides with A on Szt∅u and vanishes on ∅. Example 2.6. Let W be a k-module. By abuse of notation let W denote the (globally) constant sheaf on S. It takes constant value W on ∅ ‰ I P S and vanishes on ∅; all nontrivial restriction maps are identity isomorphisms. In this case H ˚pS; W q – H ˚pSq b W . Example 2.7. A locally constant sheaf valued by W P MODk is a sheaf W which satisfies Wp∅q “ 0, WpIq – W for I ‰ ∅ and all nontrivial restriction maps are isomorphisms. Example 2.8. Let I P S and W P MODk . Consider the sheaf t I u # W, if J ě I W t u (2.4) I pJq “ 0, otherwise, W defined by W with the restriction maps t I u pJ1 ď J2 q either identity on W (when I ď J1 ), or 0 W k W k (otherwise). Then t I u “ t I u b W and H ˚ pS; t I u q – H ˚pS; t I u q b W . We have k H ˚pS; t I u q – H ˚´|I|plkS Iq, since corresponding differential complexes coincide. In the following if A and B are two sheaves on S we denote by A b B their componentwise tensor product: pA b BqpIq “ ApIq b BpIq with restriction maps defined in the obvious way. Example 2.9. As a generalization of the previous example consider the sheaf I b A. Then k H ˚ pS; t I u b Aq – H ˚´|I| plkS I; A|lkS I q. t uk Example 2.10. Following [12], define i-th local homology sheaf Ui on S by setting Ui p∅q “ 0 and (2.5) Ui pJq “ Hi pS, Sz lkS Jq for J ‰ ∅. The restriction maps Ui pJ1 ă J2 q are induced by inclusions lkS J2 ãÑ lkS J1 . A poset S is Buchsbaum if and only if Ui “ 0 for i ă n ´ 1. Definition 2.11. Buchsbaum poset S is called homology manifold (orientable over k) if its local homology sheaf Un´1 is isomorphic to the constant sheaf k. If |S| is a compact closed orientable topological manifold then S is a homology manifold. HOMOLOGY OF TORUS SPACES 7 2.3. Cosheaves. A cellular cosheaf (see [6]) is a contravariant functor Ap : CATop pSq Ñ MODk . The homology of a cosheaf is defined similar to cohomology of sheaves: à à p “ p Ci pS; Aq p “ C˚ pS; Aq Ci pS; Aq ApIq iě´1 p Ñ Ci´1 pS; Aq, p d : Ci pS; Aq d“ à Ią1 I 1 ,dim I“i dim I“i p ě I 1 q, rI : I 1 sApI p def p dq. H˚ pS; Aq “ H˚ pC˚ pS; Aq, Example 2.12. Each locally constant sheaf W on S defines the locally constant x by inverting arrows, i.e. WpIq x – WpIq and WpI x ą Jq “ pWpJ ă Iqq´1 . cosheaf W 3. Buchsbaum pseudo-cell complexes 3.1. Simple pseudo-cell complexes. Definition 3.1 (Pseudo-cell complex). A CW-pair pF, BF q will be called kdimensional pseudo-cell, if F is compact and connected, dim F “ k, dim BF ď k ´ 1. A (regular finite) pseudo-cell complex Q is a space which is a union of an expanding sequence of subspaces Qk such that Q´1 is empty and Qk is the pushout obtained from Qk´1 by attaching finite number of k-dimensional pseudo-cells pF, BF q along injective attaching maps BF Ñ Qk´1 . The images of pF, BF q in Q will be also called pseudo-cells and denoted by the same letters. Remark 3.2. In general, situations when BF “ ∅ or Q0 “ ∅ are allowed by this definition. Thus the construction of pseudo-cell complex may actually start not from Q0 but from higher dimensions. Let F ˝ “ F zBF denote open cells. In the following we assume that the boundary of each cell is a union of lower dimensional cells. Thus all pseudo-cells of Q are partially ordered by inclusion. We denote by SQ the poset of faces with the reversed order, i.e. F ăSQ G iff G Ď BF Ă F . To distinguish abstract elements of poset SQ from faces of Q the former are denoted by I, J, . . . P SQ , and corresponding faces — FI , FJ , . . . Ď Q. Definition 3.3. A pseudo-cell complex Q, dim Q “ n is called simple if SQ is a simplicial poset of dimension n ´ 1 and dim FI “ n ´ 1 ´ dim I for all I P SQ . Thus for every face F , the upper interval tG | G Ě F u is isomorphic to a boolean lattice 2rcodim F s . In particular, there exists a unique maximal pseudo-cell F∅ of dimension n, i.e. Q itself. In case of simple pseudo-cell complexes we adopt the following naming convention: pseudo-cells different from F∅ “ Q are called faces, and faces of codimension 1 — facets. Facets correspond to vertices (atoms) of SQ . Each face F is contained in exactly codim F facets. In this paper only simple pseudo-cell complexes are considered. HOMOLOGY OF TORUS SPACES 8 Example 3.4. Nice (compact connected) manifolds with corners as defined in [9] are examples of simple pseudo-cell complexes. Each face F is itself a manifold and BF is the boundary in a common sense. Example 3.5. Each pure simplicial poset S determines a simple pseudo-cell complex P pSq such that SP pSq “ S by the following standard construction. Consider the barycentric subdivision S 1 and construct the cone P pSq “ | Cone S 1 |. By definition, Cone S 1 is a simplicial complex on the set S and k-simplices of Cone S 1 have the form pI0 ă I1 ă . . . ă Ik q, where Ii P S. For each I P S consider the pseudo-cell: FI “ |tpI0 ă I1 ă . . .q P Cone S 1 such that I0 ě Iu| Ă | Cone S 1 | BFI “ |tpI0 ă I1 ă . . .q P Cone S 1 such that I0 ą Iu| Ă | Cone S 1 | Since S is pure, dim FI “ n ´ dim I ´ 1. These sets define a pseudo-cell structure on P pSq. One shows that FI Ă FJ whenever J ă I. Thus SP pSq “ S. Face FI is called dual to I P S. The filtration by pseudo-cell skeleta (3.1) ∅ “ Q´1 Ă Q0 Ă Q1 Ă . . . Ă Qn´1 “ BQ “ |S|, is called the coskeleton filtration of |S| (see [12]). The maximal pseudo-cell F∅ of P pSq is P pSq – Cone |S|, and BF∅ “ |S|. Note that BFI can be identified with the barycentric subdivision of lkS I. Face FI is the cone over BFI . If S is non-pure, this construction makes sense as well, but the dimension of FI may not be equal to n ´ dim I ´ 1. So P pSq is not a simple pseudo-cell complex if S is not pure. For a general pseudo-cell complex Q there is a skeleton filtration (3.2) Q0 Ă Q1 Ă . . . Ă Qn´1 “ BQ Ă Qn “ Q and the corresponding spectral sequences in homology and cohomology are: (3.3) (3.4) Q 1 E p,q “ Hp`q pQp , Qp´1 q ñ Hp`q pQq, Q p,q E 1 “ H p`q pQp , Qp´1 q ñ H p`q pQq r r drQ : QE ˚,˚ Ñ QE ˚´r,˚`r´1 ˚,˚ pdQ qr : QE r ˚`r,˚´r`1 Ñ QE r . In the following only homological case is considered; the cohomological case being completely parallel. Similar to ordinary cell complexes the first term of the spectral sequence is described as a sum: à Hp`q pF, BF q. Hp`q pQp , Qp´1q – dim F “p HOMOLOGY OF TORUS SPACES 9 The differential d1Q is the sum over all pairs I ă1 J P S of the maps: (3.5) mqI,J : Hq`dim FI pFI , BFI q Ñ Hq`dim FI ´1 pBFI q Ñ Ñ Hq`dim FI ´1 pBFI , BFI zFJ˝ q – Hq`dim FJ pFJ , BFJ q, where the last isomorphism is due to excision. Also consider the truncated spectral sequence BQ 1 E p,q “ Hp`q pQp , Qp´1q, p ă n ñ Hp`q pBQq. Construction 3.6. Given a sign convention on SQ , for each q consider the sheaf Hq on SQ given by Hq pIq “ Hq`dim FI pFI , BFI q with restriction maps Hq pI ă1 Jq “ rJ : IsmqI,J . For general I ăk J consider any saturated chain I ă1 J1 ă1 . . . ă1 Jk´1 ă1 J (3.6) and set Hq pI ăk Jq to be equal to the composition Hq pJk´1 ă1 Jq ˝ . . . ˝ Hq pI ă1 J1 q. Lemma 3.7. The map Hq pI ăk Jq does not depend on a saturated chain (3.6). Proof. The differential d1Q satisfies pd1Q q2 “ 0, thus mqJ 1 ,J ˝mqI,J 1 `mqJ 2 ,J ˝mqI,J 2 “ 0. By combining this with (2.2) we prove that Hq pI ă2 Jq is independent of a chain. In general, since tT | I ď T ď Ju is a boolean lattice, any two saturated chains between I and J are connected by a sequence of elementary flips rJk ă1 T1 ă1 Jk`2 s ù rJk ă1 T2 ă1 Jk`2s.  Thus the sheaves Hq are well defined. These sheaves will be called the structure sheaves of Q. Consider also the truncated structure sheaves # Hq pIq if I ‰ ∅, Hq pIq “ 0, if I “ ∅. 1 Corollary 3.8. The cochain complexes of structure sheaves coincide with QE ˚,˚ up to change of indices: 1 pQE ˚,q , d1Q q – pC n´1´˚ pHq q, dq, 1 pBQE ˚,q , d1Q q – pC n´1´˚ pHq q, dq. Proof. Follows from the definition of the cochain complex of a sheaf.  Remark 3.9. Let S be a pure simplicial poset of dimension n ´ 1 and P pSq — its dual simple pseudo-cell complex. In this case there exists an isomorphism of sheaves (3.7) Hq – Uq`n´1 , HOMOLOGY OF TORUS SPACES 10 where U˚ are the sheaves of local homology defined in example 2.10. Indeed, it can be shown that Hi pS, Sz lkS Iq – Hi´dim I pFI , BFI q and these isomorphisms can be chosen compatible with restriction maps. For simplicial complexes this fact is proved in [12, Sec.6.1]; the case of simplicial posets is rather similar. Note that Hq depends on the sign convention while U does not. There is a simple explanation: the isomorphism (3.7) itself depends on the orientations of simplices. 3.2. Buchsbaum pseudo-cell complexes. Definition 3.10. A simple pseudo-cell complex Q of dimension n is called Buchsbaum if for any face FI Ă Q, I ‰ ∅ the following conditions hold: (1) FI is acyclic over Z; (2) Hi pFI , BFI q “ 0 if i ‰ dim FI . Buchsbaum complex Q is called Cohen–Macaulay if these two conditions also hold for I “ ∅. The second condition in Buchsbaum case is equivalent to Hq “ 0 for q ‰ 0. Cohen–Macaulay case is equivalent to Hq “ 0 for q ‰ 0. Obviously, Q is Buchsbaum if and only if all its proper faces are Cohen–Macaulay. Thus any face of dimension p ě 1 has nonempty boundary of dimension p ´ 1. In particular, this implies SQ is pure. Definition 3.11. Buchsbaum pseudo-cell complex Q is called (k-orientable) Buchsbaum manifold if H0 is isomorphic to a constant sheaf k. Note that this definition actually describes only the property of BQ not Q itself. Example 3.12. If Q is a nice compact manifold with corners in which every proper face is acyclic and orientable, then Q is a Buchsbaum pseudo-cell complex. Indeed, the second condition of 3.10 follows by Poincare–Lefschetz duality. If, moreover, Q is orientable itself then Q is a Buchsbaum manifold (over all k). Indeed, the restriction maps H0 p∅ Ă Iq send the fundamental cycle rQs P Hn pQ, BQq – k to fundamental cycles of proper faces, thus identifying H0 with the constant sheaf k. The choice of orientations establishing this identification is described in details in section 8. Example 3.13. Simplicial poset S is Buchsbaum (resp. Cohen–Macaulay) whenever P pSq is a Buchsbaum (resp. Cohen–Macaulay) simple pseudo-cell complex. Indeed, any face of P pSq is a cone, thus contractible. On the other hand, Hi pFI , BFI q – r i´1 p| lkS I|q. Thus condition 2 in definition 3.10 is satisHi pCone | lkS I|, | lkS I|q – H r i plkS Iq “ 0 for i ‰ n ´ 1 ´ |I|. This is equivalent to Buchsbaumness fied whenever H (resp. Cohen–Macaulayness) of S by remark 2.3. Poset S is a homology manifold if and only if P pSq is a Buchsbaum manifold. This follows from remark 3.9. In particular, if |S| is a closed orientable manifold then P pSq is a Buchsbaum manifold. HOMOLOGY OF TORUS SPACES 11 In general, if Q is Buchsbaum, then its underlying poset SQ is also Buchsbaum, see lemma 3.14 below. In Buchsbaum (resp. Cohen–Macaulay) case the spectral sequence BQE (resp. Q E) collapses at the second page, thus 2 – H n´1´p pSQ ; H0 q – BQE p,0 ñ Hp pBQq, if Q is Buchsbaum – 2 H n´1´p pSQ ; H0 q – QE p,0 ñ Hp pQq, if Q is Cohen–Macaulay In particular, if Q is a Buchsbaum manifold, then H n´1´p pSQ q – Hp pBQq (3.8) Let Q be a simple pseudo-cell complex, and ∅ ‰ I P SQ . The face FI is a simple pseudo-cell complex itself, and SFI “ lkS I. The structure sheaves of FI are the restrictions of Hq to lkSQ I Ă SQ . If Q is Buchsbaum, then FI is Cohen–Macaulay, thus – H k plkS I; H0 q ñ Hdim FI ´1´k pFI q, (3.9) which is either k (in case k “ dim FI ´ 1) or 0 (otherwise), since FI is acyclic. 3.3. Universality of posets. The aim of this subsection is to show that Buchsbaum pseudo-cell complex coincides up to homology with the underlying simplicial poset away from maximal cells. This was proved for nice manifolds with corners in [9] and essentially we follow the proof given there. Lemma 3.14. p1qn Let Q be Buchsbaum pseudo-cell complex of dimension n, SQ — its underlying poset, and P “ P pSQ q — simple pseudo-cell complex associated to SQ (example 3.5), BP “ |SQ |. Then there exists a face-preserving map ϕ : Q Ñ P which induces the identity isomorphism of posets and the isomorphism of the truncated spectral r r – sequences ϕ˚ : BQE ˚,˚ Ñ BPE ˚,˚ for r ě 1. p2qn If Q is Cohen–Macaulay of dimension n, then ϕ induces the isomorphism r r – of non-truncated spectral sequences ϕ˚ : QE ˚,˚ Ñ PE ˚,˚ . Proof. The map ϕ is constructed inductively. 0-skeleta of Q and P are naturally identified. There always exists an extension of ϕ to higher-dimensional faces since all pseudo-cells of P are cones. The lemma is proved by the following scheme of induction: p2qďn´1 ñ p1qn ñ p2qn . The case n “ 0 is clear. Let us prove p1qn ñ p2qn . The map ϕ induces the homomorphism of the long exact sequences: r ˚ pBQq H r ˚ pQq H /  r ˚ pBP q H H˚ pQ, BQq /  / r ˚ pP q H /  H˚ pP, BP q r ˚´1 pBQq H / r ˚´1 pQq H /  / r ˚´1 pBP q H  / r ˚´1 pP q H HOMOLOGY OF TORUS SPACES 12 r ˚ pQq Ñ H r ˚ pP q are isomorphisms since both groups are trivial. The The maps H – r ˚ pBQq Ñ H r ˚ pBP q are isomorphisms by p1qn , since BQE ñ maps H H˚ pBQq and – BP Q 1 P 1 E ñ H˚ pBP q. Five lemma shows that ϕ˚ : E n,˚ Ñ E n,˚ is an isomorphism as well. This imply p2qn . Now we prove p2qďn´1 ñ p1qn . Let FI be faces of Q and FrI — faces of P . All proper faces of Q are Cohen–Macaulay of dimension ď n ´ 1. Thus p2qďn´1 implies isomorphisms H˚ pFI , BFI q Ñ H˚ pFrI , B FrI q which sum together to the isomorphism 1 1 – ϕ˚ : BQE ˚,˚ Ñ BPE ˚,˚ .  Corollary 3.15. If Q is a Buchsbaum (resp. Cohen–Macaulay) pseudo-cell complex, then SQ is a Buchsbaum (resp. Cohen–Macaulay) simplicial poset. If Q is a Buchsbaum manifold, then SQ is a homology manifold. In particular, according to lemma 3.14, if Q is Buchsbaum, then BQ is homologous to |SQ | “ BP pSQ q. So in the following we may not distinguish between their homology. If Q is Buchsbaum manifold, then (3.8) implies Poincare duality for BQ: H n´1´p pBQq – Hp pBQq. (3.10) r 3.4. Structure of QE ˚,˚ in Buchsbaum case. Let δi : Hi pQ, BQq Ñ Hi´1 pBQq be the connecting homomorphisms in the long exact sequence of the pair pQ, BQq. ˚ Lemma 3.16. The second term of QE ˚,˚ for Buchsbaum pseudo-cell complex Q is described as follows: $ Hp pBQq, if p ď n ´ 2, q “ 0, ’ ’ ’ ’ ’ &Coker δn , if p “ n ´ 1, q “ 0, 2 Q (3.11) E p,q – Ker δn , if p “ n, q “ 0, ’ ’ ’Hn`q pQ, BQq, if p “ n, q ă 0, ’ ’ % 0, otherwise. Proof. The first page of the non-truncated spectral sequence has the form Q 1 E p,q n´2 0 C 0 n´1 1 ... .. . 0 ´n q 0 pS; H0 q . . . C pS; H0 q C pS; H0 q C 0 ´1 n n´1 ... 0 0 .. . .. . 0 0 ´1 pS; H0 q “ Hn pQ, BQq C ´1 pS; H´1 q “ Hn´1 pQ, BQq .. . C ´1 pS; H´n q “ H0 pQ, BQq p HOMOLOGY OF TORUS SPACES 13 2 By the definition of Buchsbaum complex, QE p,q “ 0 if p ă n and q ‰ 0. Terms of the second page with p ď n ´ 2 coincide with their non-truncated versions: 2 Q 2 E p,0 “ BQE p,0 – H n´1´p pSQ ; H0 q – Hp pBQq. For p “ n, q ă 0 the first differential 2 1 vanishes, thus QE n,q “ QE n,q – Hn`q pQ, BQq. The only two cases that require further investigation are pp, qq “ pn ´ 1, 0q and pn, 0q. To describe these cases consider the short exact sequence of sheaves 0 Ñ H0 Ñ H0 Ñ H0 {H0 Ñ 0. (3.12) The quotient sheaf H0 {H0 is concentrated in degree ´1 and its value on ∅ is Hn pQ, BQq. Sequence (3.12) induces the long exact sequence in cohomology (middle row): δn Hn pQ, BQq / O Hn´1 pBQq O – 0 / H ´1 pSQ ; H0 q / – H ´1 pSQ ; H0 {H0 q / H 0 pSQ ; H0 q – / H 0 pSQ ; H0 q –  Q 2 E n,0 BQ –   2 E n´1,0 2 / / Q 2 E n´1,0 2 Thus QE n,0 – Ker δn and QE n´1,0 – Coker δn .  In the situations like this, we call a spectral sequence G-shaped. The only r r r non-vanishing differentials in QE for r ě 2 are dr : QE n,1´r Ñ QE n´r,0 . They have r pairwise different domains and targets, thus QE p,q ñ Hp`q pQq folds in a long exact sequence, which is isomorphic to the long exact sequence of the pair pQ, BQq: (3.13) ... / Hi pQq / dn`1´i Q n`1´i Q / QE n`1´i E n,i´n i´1,0 / O O Hi´1 pQq / ... Hi´1 pQq / ... – – Q 2 E i´1,0 Q 1 E n,i´n – ... / Hi pQq / Hi pQ, BQq δi /  Hi´1 pBQq / This gives a complete characterization of QE in terms of the homological long exact sequence of the pair pQ, BQq. 1` 3.5. Artificial page QE ˚,˚ . In this subsection we formally introduce an addi˚ tional term in the spectral sequence to make description of QE ˚,˚ more convenient 0 HOMOLOGY OF TORUS SPACES 14 and uniform. The goal is to carry away δn (which appears in (3.11)) from the description of the page and treat it as one of higher differentials. 1` Let XE ˚,˚ be the collection of k-modules defined by Q 1` E p,q $ BQ 2 ’ & E p,q , if p ď n ´ 1, def 1 “ QE p,q , if p “ n, ’ %0, otherwise. Let d1´ Q be the differential of degree p´1, 0q operating on d1´ Q ÀQ 1 E p,q by: # 1 1 d1Q : QE p,q Ñ QE p´1,q , if p ď n ´ 1, “ 0, otherwise 1 1` Q It is easily seen that HpQE ˚,˚ ; d1´ Q q is isomorphic to E ˚,˚ . Now consider the differÀ 1` Q ential d1` E p,q : Q of degree p´1, 0q operating on d1` Q “ 2 Then QE – HpQE 1` # 0, if p ď n ´ 1; Q 1 E n,q d1Q 2 1 ÝÑ QE n´1,q ÝÑ QE n´1,q , if p “ n. , d1` Q q. These considerations are shown on the diagram: Q 1` d1´ Q Q 1 E <E d1` Q " d1Q / Q 2 E d2Q / Q 3 E d3Q / ... in which the dotted arrows represent passing to homology. To summarize: 1` Claim 3.17. There is a spectral sequence whose first page is pQE , d1` Q q and Q r subsequent terms coincide with E for r ě 2. Its nontrivial differentials for r ě 1 are the maps r r drQ : QE n,1´r Ñ QE n´r,0 which coincide up to isomorphism with δn`1´r : Hn`1´r pQ, BQq Ñ Hn´r pBQq. HOMOLOGY OF TORUS SPACES 15 r Thus the spectral sequence QE ˚,˚ for r ě 1` up to isomorphism has the form Q 1` E p,q H0 pBQq ... Hn´2 pBQq Hn´1 pBQq d1` Q “ δn d2Q “ δn´1 Hn pQ, BQq Hn´1 pQ, BQq .. . dn Q “ δ1 H1 pQ, BQq 4. Torus spaces over Buchsbaum pseudo-cell complexes 4.1. Preliminaries on torus maps. Let N be a nonnegative integer. Consider a compact torus T N “ pS 1 qN . The homology algebra H˚ pT N ; kq is the exterior algebra Λ “ Λk rH1pT N qs. Let Λpqq denote the `N˘graded component of Λ of degree q, Λ “ À N pqq pqq N pqq – Hq pT q, dim Λ “ q . q“0 Λ , Λ N If T acts on a space Z, then H˚ pZq obtains the structure of Λ-module (i.e. two-sided Λ-module with property a ¨ x “ p´1qdeg a deg x x ¨ a for a P Λ, x P H˚ pZq); T N -equivariant maps f : Z1 Ñ Z2 induce module homomorphisms f˚ : H˚ pZ1 q Ñ H˚ pZ2 q; and equivariant filtrations induce spectral sequences with Λ-module structures. Construction 4.1. Let TN be the set of all 1-dimensional toric subgroups of T N . Let M be a finite subset of TN , i.e. a collection of subgroups M “ tTs1 , is : Ts1 ãÑ T N u. Consider the homomorphism ź ź def def iM : T M Ñ T N , T M “ Ts1 , iM “ is . M M Definition 4.2. We say that the collection M of 1-dimensional subgroups satisfies p˚k q-condition if the map piM q˚ : H1 pT M ; kq Ñ H1 pT N ; kq is injective and splits. If iM itself is injective, then M satisfies p˚k q for k “ Z and all fields. Moreover, p˚Z q is equivalent to injectivity of iM . Generally, p˚k q implies that Γ “ ker iM is a finite subgroup of T M . For a set M satisfying p˚k q consider the exact sequence i ρ M 0 Ñ Γ Ñ T M ÝÑ T N ÝÑ G Ñ 0, where G “ T N {iM pT M q is isomorphic to a torus T N´|M | . HOMOLOGY OF TORUS SPACES 16 Lemma 4.3. Let IM be the ideal of Λ generated by iM pH1 pT M qq. Then there exists a unique map β which encloses the diagram ρ˚ H˚ pT N q H˚ pGq / O – β  q Λ Λ{IM / and β is an isomorphism. Proof. We have ρ˚ : H˚ pT N q Ñ H˚ pGq – Λ˚ rH1 pGqs. Map ρ is T N -equivariant, thus ρ˚ is a map of Λ-modules. Since ρ˚ ppiM q˚ H1 pT M qq “ 0, we have ρ˚ pIM q “ 0, thus ρ˚ factors through the quotient module, ρ˚ “ β ˝ q. Since ρ˚ is surjective so is β. By p˚k q-condition we have a split exact sequence 0 Ñ H1 pT M q Ñ H1 pT N q Ñ H1 pGq Ñ 0, So far there is a section α : H1 pGq Ñ H1 pT N q of the map ρ˚ in degree 1. This section extends to α r : H˚ pGq “ Λ˚ rH1 pGqs Ñ Λ, which is a section of ρ˚ . Thus β is injective.  4.2. Principal torus bundles. Let ρ : Y Ñ Q be a principal T N -bundle over a simple pseudo-cell complex Q. Lemma 4.4. If Q is Cohen–Macaulay, then Y is trivial. More precisely, there exists an isomorphism ξ: ξ Y❃ ❃ ❃❃ ρ ❃❃ ❃❃  Q / Q ˆ TN ①① ①① ① ①① {① ① The induced isomorphism ξ˚ identifies H˚ pY, BY q with H˚ pQ, BQq b Λ and H˚ pY q with Λ. Proof. Principal T N -bundles are classified by their Euler classes, sitting in H pQ; ZN q “ 0 (recall that Q is acyclic over Z). The second statement follows from the Künneth isomorphism.  2 For a general principal T N -bundle ρ : Y Ñ Q consider the filtration ∅ “ Y´1 Ă Y0 Ă Y1 Ă . . . Ă Yn´1 Ă Yn “ Y, (4.1) where Yi “ ρ´1 pQi q. For each I P SQ consider the subsets YI “ ρ´1 pFI q and BYI “ ρ´1 pBFI q. In particular, Y∅ “ Y , BY “ Yn´1. ˚ Let Y E ˚,˚ be the spectral sequence associated with filtration (4.1), i.e.: Y 1 E p,q – Hp`q pYp , Yp´1q ñ Hp`q pY q, r r drY : Y E ˚,˚ Ñ Y E ˚´r,˚`r´1 HOMOLOGY OF TORUS SPACES 17 À and Hp`q pYp , Yp´1q – |I|“n´p Hp`q pYI , BYI q. Similar to construction 3.6 we define the sheaf HqY on SQ by setting HqY pIq “ Hq`n´|I|pYI , BYI q. (4.2) The restriction maps coincide with the differential d1Y up to incidence signs. Note def À Y that HY “ q Hq has a natural Λ-module structure induced by the torus action. ˚ The cochain complex of HY coincides with the first page of Y E ˚,˚ up to change of indices. As before, consider also the truncated spectral sequence: BY 1 E p,q – Hp`q pYp , Yp´1q, p ă n ñ Hp`q pBY q, and the truncated sheaf: HYq p∅q “ 0, HYq pIq “ HY pIq for I ‰ ∅. Lemma 4.5. If Q is Buchsbaum, then HYq – H0 b Lpqq , where Lpqq is a locally constant sheaf on SQ valued by Λpqq . Proof. All proper faces of Q are Cohen–Macaulay, thus lemma 4.4 applies. We have Hq pYI , BYI q – H0 pFI , BFI q b Λpqq . For any I ă J there are two trivializations of YJ : the restriction of ξI , and ξJ itself: FJ  YJ  q q ξJ qqq q ξI |YJ q q  x qq q / FJ ˆ T N   ˆ TN / YI / ξI  FI ˆ T N Transition maps ξI |YJ ˝ pξJ q´1 induce the isomorphisms in homology Λpqq “ Hq pFJ ˆ T N q Ñ Hq pFJ ˆ T N q “ Λpqq which determine the restriction maps Lpqq pI Ă Jq. The locally constant sheaf Lpqq is thus defined, and the statement follows.  À À Denote L “ q Lpqq — the graded sheaf on SQ valued by Λ “ q Λpqq Remark 4.6. Our main example is the trivial bundle: Y “ Q ˆ T N . In this ˚ ˚ case the whole spectral sequence Y E ˚,˚ is isomorphic to QE ˚,˚ b Λ. For the structure sheaves we also have H˚Y “ H˚ b Λ˚ . In particular the sheaf L constructed in lemma ˚ 4.5 is globally trivial. By results of subsection 3.4, all terms and differentials of Y E ˚,˚ are described explicitly. Nevertheless, several results of this paper remain valid in a general setting, thus are stated in full generality where it is possible. Remark 4.7. This construction is very similar to the construction of the sheaf of local fibers which appears in the Leray–Serre spectral sequence. But contrary to this general situation, here we construct not just a sheaf in a common topological sense, but a cellular sheaf supported on the given simplicial poset SQ . Thus we prefer to provide all the details, even if they seem obvious to the specialists. HOMOLOGY OF TORUS SPACES 18 4.3. Torus spaces over simple pseudo-cell complexes. Recall, that TN denotes the set of all 1-dimensional toric subgroups of T N . Let Q be a simple pseudo-cell complex of dimension n, SQ — its underlying simplicial poset and ρ : Y Ñ Q — a principal T N -bundle. There exists a general definition of a characteristic pair in the case of manifolds with locally standard actions, see [17, Def.4.2]. We do not review this definition here due to its complexity, but prefer to work in Buchsbaum setting, in which case many things simplify. If Q is Buchsbaum, then its proper faces FI are Cohen–Macaulay, and according to lemma 4.4, there exist trivializations ξI which identify orbits over x P FI with T N . If x belongs to several faces, then different trivializations give rise to the transition homeomorphisms trIăJ : T N Ñ T N , and at the global level some nontrivial twisting may occur. To give the definition of characteristic map, we need to distinguish between these different trivializations. Denote by T N pIq the torus sitting over the face FI (via trivialization of lemma 4.4) and let TN pIq be the set of 1-dimensional subtori of T N pIq. The map trIăJ sends elements of TN pIq to TN pJq in an obvious way. One can think of TN p´q as a locally constant sheaf of sets on SQ zt∅u. Definition 4.8. A characteristic map λ is a collection of elements λpiq P TN piq defined for each vertex i P VertpSQ q. This collection should satisfy the following condition: for any simplex I P SQ , I ‰ ∅ with vertices i1 , . . . , ik the set (4.3) ttri1 ăI λpi1 q, . . . , trik ăI λpik qu satisfies p˚k q condition in T N pIq. Clearly, a characteristic map exists only if N ě n. Let T λpIq denote the subtorus of T N pIq generated by 1-dimensional subgroups (4.3). Construction 4.9 (Quotient construction). Consider the identification space: (4.4) X “ Y {„, where y1 „ y2 if ρpy1 q “ ρpy2 q P FI˝ for some ∅ ‰ I P SQ , and y1 , y2 lie in the same T λpIq -orbit. There is a natural action of T N on X coming from Y . The map µ : X Ñ Q is a projection to the orbit space X{T N – Q. The orbit µ´1 pbq over the point b P FI˝ Ă BQ is identified (via the trivializing homeomorphism) with T N pIq{T λpIq . This orbit has dimension N ´ dim T λpIq “ N ´ |I| “ dim FI ` pN ´ nq. The preimages of points b P QzBQ are the full-dimensional orbits. Filtration (4.1) descends to the filtration on X: (4.5) ∅ “ X´1 Ă X0 Ă X1 Ă . . . Ă Xn´1 Ă Xn “ X, where Xi “ Yi { „ for i ď n. In other words, Xi is the union of pď i ` N ´ nqdimensional orbits of the T N -action. Thus dim Xi “ 2i ` N ´ n for i ď n. HOMOLOGY OF TORUS SPACES 19 ˚ Let XE ˚,˚ be the spectral sequence associated with filtration (4.5): X 1 E p,q “ Hp`q pXp , Xp´1 q ñ Hp`q pXq, r r dX : XE ˚,˚ Ñ XE ˚´r,˚`r´1 . r The quotient map f : Y Ñ X induces a morphism of spectral sequences f˚r : Y E ˚,˚ Ñ X r E ˚,˚ , which is a Λ-module homomorphism for each r ě 1. 1 4.4. Structure of XE ˚,˚ . For each I P SQ consider the subsets XI “ YI {„ and BXI “ BYI {„. As before, define the family of sheaves associated with filtration 4.5: HqX pIq “ Hq`n´|I|pXI , BXI q, with the restriction maps equal to d1Y up to incidence signs. These sheaves can be 1 considered as a single sheaf HX graded by q. We have pXE ˚,q , dX q – pC n´1´˚ pSQ , HqX q, dq. There are natural morphisms of sheaves f˚ : HqY Ñ HqX induced by the quotient map f : Y Ñ X, and the corresponding map of cochain complexes coincides with À 1 1 f˚1 : Y E ˚,q Ñ XE ˚,q . Also consider the truncated versions: HX “ q HX q for which X H p∅q “ 0. Remark 4.10. The map f˚1 : H˚ pY, BY q Ñ H˚ pX, BXq is an isomorphism by excision since X{BX – Y {BY . Now we describe the truncated part of the sheaf HY in algebraical terms. Let I P SQ be a simplex and i ď I its vertex. Consider the element of exterior algebra ωi P LpIqp1q – Λp1q which is the image of the fundamental cycle of λpiq – T 1 under the transition map triďI : (4.6) ωi “ ptriďI q˚ rλpiqs P LpIqp1q Consider the subsheaf I of L whose value on a simplex I with vertices ti1 , . . . , ik u ‰ ∅ is: (4.7) IpIq “ pωi1 , . . . , ωik q Ă LpIq, — the ideal of the exterior algebra LpIq – Λ generated by linear forms. Also set Ip∅q “ 0. It is easily checked that LpI ă JqIpIq Ă IpJq, so I is a well-defined subsheaf of L. Lemma 4.11. The map of sheaves f˚ : HYq Ñ HX q is isomorphic to the quotient pqq pqq map of sheaves H0 b L Ñ H0 b pL{Iq . Proof. By lemma 4.4, pYI , BYI q Ñ pFI , BFI q is equivalent to the trivial T N bundle ξI : pYI , BYI q – pFI , BFI q ˆ T N pIq. By construction of X, we have identifications “ ‰ ξI1 : pXI , BXI q – pFI , BFI q ˆ T N pIq {„ . By excision, the group H˚ prFI ˆ T N pIqs{„, rBFI ˆ T N pIqs{„q coincides with H˚ pFI ˆ T N pIq{T λpIq , BFI ˆ T N pIq{T λpIq q “ H˚ pFI , BFI q b H˚ pT N pIq{T λpIq q. HOMOLOGY OF TORUS SPACES 20 The rest follows from lemma 4.3.  There is a short exact sequence of graded sheaves 0 ÝÑ I ÝÑ L ÝÑ L{I ÝÑ 0 Tensoring it with H0 produces the short exact sequence 0 ÝÑ H0 b I pqq ÝÑ HYq ÝÑ HX q ÝÑ 0 according to lemma 4.11. The sheaf H0 b I can also be considered as a subsheaf of non-truncated sheaf HY . Lemma 4.12. There is a short exact sequence of graded sheaves 0 Ñ H0 b I Ñ HY Ñ HX Ñ 0. Proof. Follows from the diagram 0 0 0 0 / /  H0 b I  / H0 b I  /  Y // Y // H _  H  Y H {H  0 Y – / HX _  HX / 0 / 0  H {HX X  0 The lower sheaves are concentrated in ∅ P SQ and the graded isomorphism between them is due to remark 4.10.  4.5. Extra pages of Y E and XE. To simplify further discussion we briefly sketch the formalism of additional pages of spectral sequences Y E and XE, which extends considerations of subsection 3.5. Consider the following bigraded module: # BY 2 E , if p ă n; 1` Y E p,q “ Y 1p,q E n,q , if p “ n. 1 1` 1` Y Y and define the differentials d1´ by Y on E and dY on E # # 0, if p ă n; d1Y , if p ă n; 1` d1´ “ d “ d1Y Y 1 Y Y 2 Y 1 0, if p “ n. E n´1,q ÝÑ Y E n´1,q if p “ n. E n,q ÝÑ HOMOLOGY OF TORUS SPACES 1` 1 21 2 1` Y Y It is easily checked that Y E – HpY E , d1´ , d1` Y q and E – Hp E Y q. The page 1 1 1´ 1` X 1` E and the differentials dX , dX are defined similarly. The map f˚1 : Y E Ñ XE 1` 1` induces the map between the extra pages: f˚1` : Y E Ñ XE . 5. Main results r 5.1. Structure of XE ˚,˚ . The short exact sequence of lemma 4.12 generates the long exact sequence in sheaf cohomology: (5.1) f˚2 Ñ H i´1 pSQ ; H0 bI pqq q Ñ H i´1 pSQ ; HqY q ÝÑ H i´1 pSQ ; HqX q ÝÑ H i pSQ ; H0 bI pqq q Ñ The following lemma is the cornerstone of the whole work. Lemma 5.1 (Key Lemma). H i pSQ ; H0 b I pqq q “ 0 if i ď n ´ 1 ´ q. The proof follows from a more general sheaf-theoretical fact and is postponed to section 6. In the following we simply write S instead of SQ . By construction, 2 Y 2 E p,q – H n´1´p pS; HqY q and XE p,q – H n´1´p pS; HqX q. The Key lemma 5.1 and exact sequence (5.1) imply Lemma 5.2. 2 2 f˚2 : Y E p,q Ñ XE p,q is an isomorphism if p ą q, 2 2 f˚2 : Y E p,q Ñ XE p,q is injective if p “ q. 2 2 In case N “ n this observation immediately describes XE ˚,˚ in terms of Y E ˚,˚ . Under the notation r r pdY qrp,q : Y E p,q Ñ Y E p´r,q`r´1, r r pdX qrp,q : XE p,q Ñ XE p´r,q`r´1 there holds Theorem 1. Let Q be Buchsbaum pseudo-cell complex of dimension n, Y be a principal T n -bundle over Q, f : Y Ñ X “ Y {„ — the quotient construction, and r r f˚r : Y E ˚,˚ Ñ XE ˚,˚ — the induced map of homological spectral sequences associated with filtrations 4.1, 4.5. Then # an isomorphism if q ă p or q “ p “ n, 2 Y 2 X 2 f˚ : E p,q Ñ E p,q is injective if q “ p ă n, 2 ˚ and XE p,q “ 0 if q ą p. Higher differentials of XE ˚,˚ thus have the form # f˚r ˝ pdY qrp,q ˝ pf˚r q´1 , if p ´ r ě q ` r ´ 1, pdX qrp,q “ 0 otherwise, for r ě 2. HOMOLOGY OF TORUS SPACES ˚ 22 ˚ If Y is a trivial T n -bundle, then the structure of Y E ˚,˚ – QE ˚,˚ b Λ is described ˚ completely by subsection 3.4. In this case almost all the terms of XE ˚,˚ are described explicitly. Theorem 2. In the notation of Theorem 1 suppose Y “ Q ˆ T n . Let Λpqq “ Hq pT n q and δi : Hi pQ, BQq Ñ Hi´1 pBQq be the connecting homomorphisms. Then (5.2) X 2 E p,q $ Hp pBQq b Λpqq , if q ă p ď n ´ 2; ’ ’ ’ ’ ’ Coker δn b Λpqq ,¨if q ă p “ n ´ 1; ’ ’ ˛ ’ ’ & À Hq1 pQ, BQq b Λpq2 q ‚, if q ă p “ n; – Ker δn b Λpqq ‘ ˝ ’ q1 `q2 “n`q ’ ’ q1 ăn ’ ’ ’ pnq ’ Hn pQ, BQq b Λ , if q “ p “ n; ’ ’ % 0, if q ą p. 2 The maps f˚2 : Hq pBQqbΛpqq ãÑ XE q,q are injective for q ă n´1. Higher differentials for r ě 2 are the following: $ XE ˚ XE ˚ q1 ´1,q2 n,q1 `q2 ´n ’ ’ Y Y ’ ’ pq2 q ’ H pBQq b Λpq2 q , H pQ, BQq b Λ Ñ δ b id : ’ q1 ´1 q1 q1 Λ ’ ’ & if r “ n ´ q1 ` 1, q1 ´ 1 ą q2 ; drX – 2 2 pq q ’ f˚ ˝ pδq1 b idΛ q : Hq1 pQ, BQq b Λ 2 Ñ Hq1 ´1 pBQq b Λpq2 q ãÑ XE q1 ´1,q1 ´1 , ’ ’ ’ ’ ’ if r “ n ´ q1 ` 1, q1 ´ 1 “ q2 ; ’ ’ % 0, otherwise. Using the formalism of extra pages introduced in subsection 4.5, Theorem 2 can be restated in a more convenient and concise form Statement 5.3. There exists a spectral sequence whose first term is (5.3) X 1` E p,q $ ’ H pBQq b Λpqq , if q ă p ă n; ’ & pÀ Hq1 pQ, BQq b Λpq2 q , if p “ n; – q1 `q2 “q`n ’ ’ % 0, if q ą p; HOMOLOGY OF TORUS SPACES 23 r and subsequent terms coincide with XE ˚,˚ for r ě 2. There exist injective maps 2 f˚1` : Hq pBQq b Λpqq ãÑ XE q,q for q ă n. Differentials for r ě 1 have the form $ XE ˚ XE ˚ q1 ´1,q2 n,q1 `q2 ´n ’ ’ Y Y ’ ’ pq q ’ δq1 b idΛ : Hq1 pQ, BQq b Λ 2 Ñ Hq1 ´1 pBQq b Λpq2 q , ’ ’ ’ & if r “ n ´ q1 ` 1, q1 ´ 1 ą q2 ; r dX – 2 pq q 1` 2 ’ f˚ ˝ pδq1 b idΛ q : Hq1 pQ, BQq b Λ Ñ Hq1 ´1 pBQq b Λpq2 q ãÑ XE q1 ´1,q1 ´1 , ’ ’ ’ ’ if r “ n ´ q1 ` 1, q1 ´ 1 “ q2 ; ’ ’ ’ % 0, otherwise. ˚ Note that the terms XE q,q for q ă n are not mentioned in the lists (5.2), (5.3). À 1` ˚ Let us call qăn XE q,q the border of XE ˚,˚ . This name is due to the fact that all ˚ entries above the border vanish: XE p,q “ 0 for q ą p. r p pSq “ dimk H r p pBQq by rbp pSq for p ă n. The ranks of the border Denote dimk H components are described as follows: Theorem 3. In the notation of Theorem 2 and statement 5.3 ˆ ˙ÿ q n X 1` dim E q,q “ hq pSq ` p´1qp`qrbp pSq q p“0 for q ď n ´ 1, where hq pSq are the h-numbers of the simplicial poset S. Theorem 4. 1` (1) Let Q be a Buchsbaum manifold over k. Then dim XE q,q “ h1n´q pSq for 1` q ď n ´ 2 and XE n´1,n´1 “ h11 pSq ` n. (2) Let Q be Buchsbaum manifold such that Hn pQ, BQq – k and δn : Hn pQ, BQq Ñ 2 Hn´1 pBQq is injective. Then dim XE q,q “ h1n´q pSq for 0 ď q ď n. The definitions of h-, h1 - and h2 -vectors and the proof of Theorems 3,4 and 5 are gathered in section 7. Note that it is sufficient to prove Theorem 3 and the first 2 1` part of Theorem 4 in the case Q “ P pSq. Indeed, by definition, XE q,q “ BXE q,q for q ď n ´ 1, and there exists a map pQ ˆ T n q{„Ñ pP pSQ q ˆ T n q{„ which covers the map ϕ of lemma 3.14 and induces the isomorphism of corresponding truncated spectral sequences. In the cone case the border components can be described explicitly up to 8-term. Theorem 5. Let S be a Buchsbaum poset, Q “ P pSq, Y “ Q ˆ T n , X “ r Y {„, XE p,q ñ Hp`q pXq — the homological spectral sequence associated with filtration (4.5). Then 8 dim XE q,q “ h2q pSq for 0 ď q ď n. HOMOLOGY OF TORUS SPACES 24 Corollary 5.4. If S is Buchsbaum, then h2i pSq ě 0. Proof. For any S there exists a characteristic map over Q. Thus there exists a space X “ pP pSq ˆ T n q{„ and Theorem 5 applies.  5.2. Homology of X. Theorem 2 implies the additional grading on H˚ pSq — the one given by the degrees of exterior forms. It is convenient to work with this double grading. Construction 5.5. Suppose Y “ QˆT n . For j P r0, ns consider the G-shaped spectral sequence Q r pjq Y r j E ˚,˚ “ E ˚,˚ b Λ . Àn Y r r def pjq Y ˚ Clearly, Y E ˚,˚ “ j“0 j E ˚,˚ and j E p,q ñ Hp`q´j,j pY q “ Hp`q´j pQq b Λ . In particular, ¯ à´ ¯ à ´Q 1` pjq Q 1` pjq Y 1` “ ‘ E b Λ E b Λ E p,0 n,q j ˚,˚ păn q ˚ Consider the corresponding G-shaped spectral subsequences in XE ˚,˚ . Start with the k-modules: ¯ à X 1` à 1` ´Q 1` X 1` pqq “ E ‘ f . E E b Λ p,j ˚ j n,q ˚,˚ păn q ˚ ˚ By statement 5.3, all the differentials of XE ˚,˚ preserve X E , thus the spectral Àn X j r ˚,˚ X r X r subsequences j E ˚,˚ are well defined, and E ˚,˚ “ j“0 j E ˚,˚ . Let Hi,j pXq be the r family of subgroups of H˚ pXq such that X j E p,q ñ Hp`q´j,j pXq. Then à Hk pXq “ Hi,j pXq i`j“k and the map f˚ : H˚ pY q Ñ H˚ pXq sends Hi,j pY q – Hi pQqbΛpjq to Hi,j pXq. The map r r r r f˚r : Y E Ñ XE sends Yj E to X j E for each j P t0, . . . , nu and we have commutative squares: Y r j E p,q +3 Hp`q´j,j pY q f˚r f˚  X r j E p,q +3  Hp`q´j,j pXq Proposition 5.6. (1) If i ą j, then f˚ : Hi,j pY q Ñ Hi,j pXq is an isomorphism. In particular, Hi,j pXq – Hi pQq b Λpjq . (2) If i ă j, then there exists an isomorphism Hi,j pXq – Hi pQ, BQq b Λpjq . HOMOLOGY OF TORUS SPACES 25 (3) In case i “ j ă n, the module Hi,i pXq fits in the exact sequence 8 0 Ñ XE i,i Ñ Hi,i pXq Ñ Hi pQ, BQq b Λpiq Ñ 0, or, equivalently, 1` 0 Ñ Im δi`1 b Λpiq Ñ XE i,i Ñ Hi,i pXq Ñ Hi pQ, BQq b Λpiq Ñ 0 8 1 (4) If i “ j “ n, then Hn,n pXq “ XE n,n “ XE n,n . Proof. According to statement 5.3 # the isomorphism if i ą j or i “ j “ n; 1` 1` (5.4) f˚1` : Yj E i,q Ñ X j E i,q is injective if i “ j. For each j both spectral sequences Yj E and X j E are G-shaped, thus fold in the long exact sequences: (5.5) ... / n´i`1 Y 1` j E i,j / Hi,j pY q f˚ / X 1` j E i,j /  Hi,j pXq dY Y 1` E j n,i´n`j / Y 1` j E i´1,j – f˚ f˚  ... / /  ... / f˚ n´i`1 dX X 1` j E n,i´n`j /  X 1` j E i´1,j / ... 1` Applying five lemma in the case i ą j proves (1). For i ă j, the groups X j E i,j , X 1` j E i´1,j 1` 1` Y vanish by dimensional reasons thus Hi,j pXq – X j E n,i´n`j – j E n,i´n`j – Hi pQ, BQq b Λpjq . Case i “ j also follows from (5.5) by a simple diagram chase.  In the manifold case proposition 5.6 reveals a bigraded duality. If Q is a nice manifold with corners, Y “ Q ˆ T n and λ is a characteristic map over Z, then X is a manifold with locally standard torus action. In this case Poincare duality respects the double grading. Proposition 5.7. If X “ pQ ˆ T n q{„ is a manifold with locally standard torus action and k is a field, then Hi,j pXq – Hn´i,n´j pXq. Proof. If i ă j, then Hi,j pXq – Hi pQ, BQq b Λpjq – Hn´i pQq b Λpn´jq – Hn´i,n´j pXq, since Hi pQ, BQq – Hn´i pQq by the Poincare–Lefschetz duality and Hj pT n q – Hn´j pT n q by the Poincare duality for the torus. The remaining isomorphism Hi,i pXq – Hn´i,n´i pXq now follows from the ordinary Poincare duality for X.  Remark 5.8. If the space X “ pQ ˆ T n q{„ is constructed from a manifold with corners pQ, BQq using characteristic map over Q (i.e. X is a toric orbifold), then proposition 5.7 still holds over Q. HOMOLOGY OF TORUS SPACES 26 6. Duality between certain cellular sheaves and cosheaves 6.1. Proof of the Key lemma. In this section we prove lemma 5.1. First recall the setting. ‚ Q : a Buchsbaum pseudo-cell complex with the underlying simplicial poset S “ SQ (this poset is Buchsbaum itself by corollary 3.15). ‚ H0 : the structure sheaf on S; H0 pJq “ Hdim FJ pFJ , BFJ q. ‚ L : a locally constant graded sheaf on S valued by exterior algebra Λ “ H˚ pT N q. This sheaf is associated in a natural way to a principal T N -bundle over Q, LpJq “ H˚ pT N pJqq for J ‰ ∅ and Lp∅q “ 0. By inverting all p restriction maps we obtain the cosheaf L. ‚ λ : a characteristic map over S. It determines T 1 -subgroup triďJ pλpiqq Ă T N pJq for each simplex J with vertex i. The homology class of this subgroup is denoted ωi P LpIq (see (4.6)). Note that the restriction isomorphism LpJ1 ă J2 q sends ωi P LpJ1 q to ωi P LpJ2 q. Thus we simply write ωi for all such elements since the ambient exterior algebra will be clear from the context. ‚ I : the sheaf of ideals, associated to λ. The value of I on a simplex J ‰ ∅ with vertices ti1 , . . . , ik u is the ideal IpJq “ pωi1 , . . . , ωik q Ă LpJq. Clearly, I is a graded subsheaf of L. We now introduce another type of ideals. Construction 6.1. Let J “ ti1 , . . . , ik u be a nonempty subset of vertices of Ź simplex I P S. Consider the element πJ P LpIq, πJ “ iPJ ωi . By the definition of characteristic map, the elements ωi are linearly independent, thus πJ is a non-zero |J|-form. Let ΠJ Ă LpIq be the principal ideal generated by πJ . The restriction maps LpI ă I 1 q identify ΠJ Ă LpIq with ΠJ Ă LpI 1 q. In particular, when J is the whole set of vertices of a simplex I ‰ ∅ we define def p p p ą I 1 q “ LpI 1 ă Iq´1 ΠpIq “ ΠI Ă LpIq. If I 1 ă I, then the corestriction map LpI p is a well-defined p p 1 q, since LpI p ą I 1 qπI is divisible by πI 1 . Thus Π injects ΠpIq into ΠpI p Formally set Πp∅q p graded subcosheaf of L. “ 0. Theorem 6. For Buchsbaum pseudo-cell complex Q and S “ SQ there exists an p which respects the gradings of I and isomorphism H k pS; H0 b Iq – Hn´1´k pS; Πq p Π. Before giving a proof let us deduce the Key lemma. We need to show that H pS; H0 b I pqq q “ 0 for i ď n ´ 1 ´ q. According to Theorem 6 this is equivalent to i p pqq q “ 0 for i ě q. Lemma 6.2. Hi pS; Π p Proof. The ideal ΠpIq “ ΠI is generated by the element πI of degree |I| “ pqq dim I ` 1. Thus ΠI “ 0 for q ď dim I. Hence the corresponding part of the chain complex vanishes.  HOMOLOGY OF TORUS SPACES 27 Proof of Theorem 6. The idea of proof is the following. First we construct a resolution of sheaf H0 b I whose terms are “almost acyclic”. By passing to cochain complexes this resolution generates a bicomplex C‚‚ . By considering two standard spectral sequences for this bicomplex we prove that both H k pS; H0 b Iq p are isomorphic to the cohomology of the totalization C ‚ . and Hn´1´k pS; Πq Tot t u ΠI For each ∅ ‰ I P S consider the sheaf RI “ H0 b I (see examples 2.8 and 2.10), i.e.: # H0 pJq b ΠI , if I ď J; RI pJq “ 0 otherwise. À pqq The sheaf RI is graded by degrees of exterior forms: RI “ q RI . Since I ą I 1 implies ΠI Ă ΠI 1 , and i P VertpSq, i ď J implies Πi Ă IpJq, there exist natural injective maps of sheaves: θIąI 1 : RI ãÑ RI 1 , and ηi : Ri ãÑ H0 b I. For each k ě 0 consider the sheaf à R´k “ RI , dim I“k These sheaves can be arranged in the sequence d d η H H . . . ÝÑ R´2 ÝÑ R´1 ÝÑ R0 ÝÑ H0 b I ÝÑ 0, À 1 where dH “ Ią1 I 1 rI : I sθIąI 1 and η “ iPVertpSq ηi . By the standard argument 1 involving incidence numbers rI : I s one shows that (6.1) is a differential complex of sheaves. Moreover, (6.1) R‚ : À Lemma 6.3. The sequence R‚ is exact. Proof. We should prove that the value of R‚ at each J P S is exact. Since RI pJq ‰ 0 only if I ď J the complex R‚ pJq has the form à à (6.2) . . . ÝÑ ΠI ÝÑ ΠI ÝÑ IpJq ÝÑ 0, IďJ,|I|“2 IďJ,|I|“1 tensored with H0 pJq. Without lost of generality we forget about H0 pJq. Maps in (6.2) are given by inclusions of sub-ideals (rectified by incidence signs). This looks very similar to the Taylor resolution of monomial ideal in commutative polynomial ring, but our situation is a bit different, since ΠI are not free modules over Λ. Anyway, the proof is similar to commutative case: exactness of (6.2) follows from inclusion-exclusion principle. To make things precise (and also to tackle the case k “ Z) we proceed as follows. By p˚k q-condition, the subspace xωj | j P Jy is a direct summand in Lp1q pJq – kn . Let tν1 , . . . , νn u be such a basis of Lp1q pJq, that its first |J| vectors are identified HOMOLOGY OF TORUS SPACES 28 with ωj , j P J. We simply write J for t1, . . . , |J|u Ď rns Àby abuse of notation. The module Λ splits in the multidegree components: Λ “ AĎrns ΛA , where ΛA is a 1Ź dimensional k-module generated by iPA νi . All modules and maps in (6.2) respect this splitting. Thus (6.2) can be written as à à à à à . . . ÝÑ ΛA ÝÑ ΛA ÝÑ ΛA ÝÑ 0, IĎJ,|I|“2 AĚI à A,AXJ‰∅ ˜ IĎJ,|I|“1 AĚI à . . . ÝÑ ΛA ÝÑ IĎAXJ,|I|“2 AXJ‰∅ à ΛA ÝÑ ΛA ÝÑ 0 IĎAXJ,|I|“1 ¸ r ˚ p∆AXJ ; ΛA q – For each A the cohomology of the complex in brackets coincides with H r ˚ p∆AXJ ; kq, the reduced simplicial homology of a simplex on the set A X J ‰ ∅. H Thus homology vanishes.  By passing to cochains and forgetting the last term we get a complex of complexes (6.3) dH dH C‚‚ “ CpS, R‚ q : . . . ÝÑ C ‚ pS; R´2 q ÝÑ C ‚ pS; R´1 q ÝÑ C ‚ pS; R0 q ÝÑ 0, whose horizontal cohomology vanishes except for the upmost right position. Let pqq dV be the “vertical” cohomology differential operating in each C ‚ pS; Rk q. Then ‚ ‚ dH dV “ dV dH . Thus C‚ can be considered as a bicomplex pCTot , Dq: à i à l ‚ i Ck , D “ dH ` p´1qk dV . CTot “ CTot , CTot “ i k`l“i ‚ There are two standard spectral sequences converging to H ˚ pCTot , Dq [11, Ch.2.4]. The first one, horizontal: H ˚,˚ Er , k,l k`1´r,l`r H H dH r : Er Ñ Er computes horizontal cohomology first, then vertical cohomology. The second, vertical, k,l k`r,l`1´r V ˚,˚ Er , dVr : V E r Ñ VE r computes vertical cohomology first, then horizontal. ‚ Lemma 6.4. H k pCTot , Dq – H k pS; H0 b Iq. Proof. Consider the horizontal spectral sequence: # C l pS; H0 b Iq, k “ 0; k,l H E 1 “ H k pC l pS, Rk q, dH q “ 0, o.w. # H l pS; H0 b Iq, if k “ 0; H k,l E2 “ 0, o.w. ˚,˚ Spectral sequence HE ˚ collapses at the second term and the statement follows.  HOMOLOGY OF TORUS SPACES 29 ‚ p Lemma 6.5. H k pCTot , Dq – Hn´1´k pS; Πq. Proof. Consider the vertical spectral sequence. It starts with à V k,l E 1 – H l pS; Rk q “ H l pS; RI q. dim I“´k Similar to example 2.9 we get ΠI H l pS; RI q “ H l pS; H0 b t I u q “ H l´|I| plkS I; H0 |lk I b ΠI q The restriction of H0 to lkS I Ă S coincides with the structure sheaf of FI and by (3.9) we have a collapsing – H l´|I| plkS I; H0 |lk I b ΠI q ñ Hn´1´l pFI q b ΠI A proper face FI is acyclic, thus H l pS; RI q – # ΠI , if l “ n ´ 1, 0, if l ‰ n ´ 1. The maps θIąI 1 induce the isomorphisms H˚ pFI q Ñ H˚ pFI 1 q which assemble in commutative squares H n´1 pS; RI q – +3 ΠI _ ˚ θIąI 1  H n´1 pS; RI 1 q – +3  ΠI 1 Thus the first term of vertical spectral sequence is identified with the chain complex p of cosheaf Π: # # p if l “ n ´ 1; p if l “ n ´ 1; C pS; Πq, H´k pS; Πq, k,l k,l ´k V V E1 “ E2 “ 0, if l ‰ n ´ 1. 0, if l ‰ n ´ 1. k,l The spectral sequence VE ˚ ñ H k`l pq C ‚Tot , Dq collapses at the second page. Lemma proved.  Theorem 6 follows from lemmas 6.4 and 6.5. 6.2. Extending duality to exact sequences. Theorem 6 can be refined: Statement 6.6. The short exact sequence of sheaves (6.4) 0 Ñ H0 b I Ñ H0 b L Ñ H0 b pL{Iq Ñ 0 and the short exact sequence of cosheaves p Ñ Lp Ñ L{ pΠ pÑ0 0ÑΠ  HOMOLOGY OF TORUS SPACES 30 induce isomorphic long exact sequences in (co)homology: (6.5) / H i pS; H0 b pL{Iqq / H i pS; H0 b Lq H i pS; H0 b Iq O O – O –  p Hn´1´i pS; Πq H i`1 pS; H0 b Iq O –  / / –  p Hn´1´i pS; Lq / p Πqq p Hn´1´i pS; pL{ /  p H n´2´i pS; Πq Proof. The proof goes essentially the same as in Theorem 6. Denote sequence (6.4) by seqI. For each ∅ ‰ I P S consider the short exact sequence of sheaves: ¯ ´ t u ΠI Ñ0 R : 0 Ñ R Ñ H b L Ñ H b L{ I seq I I 0 0 and define seqR´k “ à seqRI dim I“k One can view seqI, seqRI and seqR´k as the objects in a category of complexes. As before, we can form the sequence (6.6) seqR‚ : d d η H H . . . ÝÑ seqR´2 ÝÑ seqR´1 ÝÑ seqR0 ÝÑ seqI ÝÑ 0, which happens to be exact in all positions. This long exact sequence (after forgetting the last term) generates the bicomplex of short exact sequences (or the short exact sequence of bicomplexes) seqC ‚‚ . By taking totalization and considering standard spectral sequences we check that both rows in (6.5) are isomorphic to the long exact sequence of cohomology associated to pseqC ‚Tot , Dq.  6.3. Remark on duality. In the manifold case (i.e. sheaf H0 is isomorphic to k), the proof of Theorem 6 can be restated in more conceptual terms. In this case the cellular version of Verdier duality for manifolds [6, Th.12.3] asserts: H i pS; Iq “ Hn´1´i pS; Iq, where the homology groups of a cellular sheaf are defined as homology of global Π sections of projective sheaf resolution [6, Def.11.29]. The sheaf RI “ H0 b t I u I – t u ΠI I is projective ([6, Sec.11.1.1]), thus (6.1) is actually a projective resolution. Due p to the specific structure of this resolution, we have H˚ pS; Iq – H˚ pS; Πq. 7. Face vectors and ranks of border components In this section we prove Theorems 3, 4 and 5. HOMOLOGY OF TORUS SPACES 31 7.1. Preliminaries on face vectors. First recall several standard definitions from combinatorial theory of simplicial complexes and posets. Construction 7.1. Let S be a pure simplicial poset, dim S “ n ´ 1. Let fi pSq “ tI P S | dim I “ iu, f´1 pSq “ 1. The array pf´1 , f0 , . . . , fn´1 q is called the f -vector of S. We write fi instead of fi pSq ř since S is clear from the context. Let fS ptq be the generating polynomial: fS ptq “ iě0 fi´1 ti . Define the h-numbers by the relation: ˙ ˆ n n ÿ ÿ t i i n´i n . (7.1) hi pSqt “ fi´1 t p1 ´ tq “ p1 ´ tq fS 1´t i“0 i“0 ř ř r i pSq, χpSq “ n´1 p´1qi bi pSq “ n´1 p´1qi fi pSq Let bi pSq “ dim Hi pSq, rbi pSq “ dim H i“0 i“0 řn´1 r and χ rpSq “ i“0 bi pSq “ χpSq ´ 1. Thus fS p´1q “ 1 ´ χpSq. Also note that hn pSq “ p´1qn´1 χ rpSq. 1 Define h - and h2 -vectors by ¸ ˆ ˙ ˜ i´1 ÿ n (7.2) h1i “ hi ` p´1qi´j´1rbj´1 pSq for 0 ď i ď n; i j“1 ¸ ˆ ˙ ˆ ˙ ˜ÿ i n n rbi´1 pSq “ hi ` (7.3) h2i “ h1i ´ p´1qi´j´1rbj´1 pSq for 0 ď i ď n ´ 1, i i j“1 and h2n “ h1n . Note that (7.4) h1n “ hn ` n´1 ÿ j“0 p´1qn´j´1rbj´1 pSq “ rbn´1 pSq. Statement 7.2 (Dehn–Sommerville relations). If S is Buchsbaum and dim H0 pIq “ 1 for each I ‰ ∅, then ˆ ˙ i n p1 ´ p´1qn ´ χpSqq, (7.5) hi “ hn´i ` p´1q i or, equivalently: (7.6) ˆ ˙ n p1 ` p´1qn χ rpSqq, hi “ hn´i ` p´1q i i If, moreover, S is a homology manifold, then h2i “ h2n´i . Proof. The first statement can be found e.g. in [16] or [4, Thm.3.8.2]. Also see remark 7.5 below. The second then follows from the definition of h2 -vector and Poincare duality (3.8) bi pSq “ bn´1´i pSq.  HOMOLOGY OF TORUS SPACES 32 Definition 7.3. Let S be Buchsbaum. For i ě 0 consider ÿ ÿ r n´1´|I| plkS Iq “ fpi pSq “ dim H dim H0 pIq. IPS,dim I“i IPS,dim I“i If S is a homology manifold, then fpi pSq “ fi pSq. For general Buchsbaum complexes there is another formula connecting these quantities. ř Proposition 7.4. fS ptq “ p1 ´ χpSqq ` p´1qn kě0 fpk pSq ¨ p´t ´ 1qk`1. Proof. This follows from the general statement [8, Th.9.1],[4, Th.3.8.1], but we provide an independent proof for completeness. As stated in [1, Lm.3.7,3.8] for simř d plicial complexes (and also not difficult to prove for posets) dt fS ptq “ vPVertpSq flk v ptq, and, more generally, ˆ ˙k ÿ d fS ptq “ k! flk I ptq. dt IPS,|I|“k Thus for k ě 1: pkq fS p´1q “ k! ÿ p1 ´ χplkS Iqq “ IPS,|I|“k “ k! ÿ IPS,|I|“k r n´|I|´1plk Iq “ p´1qn´k k!fpk´1 pSq. p´1qn´|I| dim H Considering the Taylor expansion of fS ptq at ´1: ÿ ÿ 1 pkq fS p´1qpt` 1qk “ p1 ´ χpSqq ` p´1qn´k´1fpk pSq ¨pt` 1qk`1, fS ptq “ fS p´1q ` k! kě0 kě1 finishes the proof.  Remark 7.5. If S is a manifold, then proposition 7.4 implies fS ptq “ p1 ´ p´1qn ´ χpSqq ` p´1qn fS p´t ´ 1q, which is an equivalent form of Dehn–Sommerville relations (7.5). Lemma 7.6. For Buchsbaum poset S there holds n ÿ ÿ fpk pSq ¨ pt ´ 1qn´k´1. hi ti “ p1 ´ tqn p1 ´ χpSqq ` i“0 kě0 Proof. Substitute t{p1 ´ tq in proposition 7.4 and use (7.1). The coefficients of ti in lemma 7.6 give the relations ˙ ˆ ˆ ˙ ÿ n´k´i´1 n ´ k ´ 1 p i n fk pSq. p´1q ` (7.7) hi pSq “ p1 ´ χpSqqp´1q i i kě0  HOMOLOGY OF TORUS SPACES 33 1 7.2. Ranks of XE ˚,˚ . Our goal is to compute the ranks of border groups 1` dim XE q,q . The idea is very straightforward: statement 5.3 describes the ranks 1` 1 of all groups XE p,q except for p “ q; and the terms XE p,q are known as well; thus 1` dim XE q,q can be found by comparing Euler characteristics. Note that the terms 1 1` with p “ n do not change when passing from XE to XE . Thus it is sufficient to perform calculations with the truncated sequence BXE. By construction, 1 X 1 E p,q – BXE p,q – C n´p´1 pS; HqX q for p ă n. Thus lemma 4.11 implies for p ă n: dim 1 E p,q X 1 E p,q BX “ dim ÿ “ dim H0 pIq ¨ dimpΛ{IpIqq pqq |I|“n´p Let χ1q be the Euler characteristic of q-th row of χ1q (7.8) “ ÿ p p´1q dim pďn´1 1 E p,q BX BX ˆ ˙ p ¨ fpn´p´1pSq. “ q 1 E ˚,˚ : ˆ ˙ p p “ p´1q fn´p´1 q pďn´1 ÿ p Lemma 7.7. For q ď n ´ 1 there holds χ1q “ pχpSq ´ 1q `n˘ q ` p´1qq hq pSq. Proof. Substitute i “ q and k “ n ´ p ´ 1 in (7.7) and combine with (7.8).  1` 1` 7.3. Ranks of XE ˚,˚ . By construction of the extra page, XE p,q – 2 p ă n. Let χ2q be the Euler characteristic of q-th row of BXE ˚,˚ : χ2q “ (7.9) BX 2 E p,q for ÿ 2 p´1qp dim BXE p,q . p Euler characteristics of first and second terms coincide: χ2q “ χ1q . By statement 5.3, ` ˘ 1` dim XE p,q “ nq bp pSq for q ă p ă n. Lemma 7.7 yields q p´1q dim 1` E q,q X ˆ ˙ ˆ ˙ n n ` p´1qq hq . bp pSq “ pχpSq ´ 1q p´1q ` q q p“q`1 n´1 ÿ p By taking into account obvious relations between reduced and non-reduced Betti řn´1 numbers and equality χpSq “ p“0 bp pSq, this proves Theorem 3. HOMOLOGY OF TORUS SPACES 34 7.4. Manifold case. If X is a homology manifold, then Poincare duality bi pSq “ bn´i pSq and Dehn–Sommerville relations (7.6) imply ˆ ˙ÿ q n X 1` dim E q,q “ hq ` p´1qp`qrbp “ q p“0 ˆ ˙ ˆ ˙ÿ q n q n ` p´1qp`q bp “ “ hq ´ p´1q q p“0 q ˆ ˙ ˆ ˙ n´1 ÿ n q n ` p´1qn´1´p`q bp “ “ hq ´ p´1q q p“n´1´q q ff ˆ ˙« n´1 ÿ n ´p´1qn ` p´1qn χ ` p´1qn´1´p bp “ “ hn´q ` p´1qq q p“n´1´q ff ˆ ˙« n´q´2 ÿ n p´1qp`n bp . ´p´1qn ` “ hn´q ` p´1qq q p“0 ř p`nr bp pSq whenever the The final expression in brackets coincides with n´q´2 p“´1 p´1q 1` summation is taken over nonempty set, i.e. for q ă n ´ 1. Thus dim XE q,q “ h1n´q ` ˘ 1` n for q ă n ´ 1. In the case q “ n ´ 1 we get dim XE n´1,n´1 “ h1 ` n´1 “ h11 ` n. This proves part (1) of Theorem 4. Part (2) follows easily. Indeed, for q “ n: ˆ ˙ n X 1 X 2 dim Hn pQ, BQq “ 1 “ h10 dim E n,n “ dim E n,n “ n For q “ n ´ 1: dim 2 E n´1,n´1 X 2 “ dim 1` E n´1,n´1 X ˙ n dim Im δn “ h11 . ´ n´1 ˆ 1` If q ď n ´ 2, then XE q,q “ XE q,q , and the statement follows from part (1). 7.5. Cone case. If Q “ P pSq – Cone |S|, then the map δi : Hi pQ, BQq Ñ r Hi´1 pBQq is an isomorphism. Thus for q ď n ´ 1 statement 5.3 implies ˆ ˙ ˆ ˙ n n r X 8 X 1` X 1` dim E q,q “ dim E q,q ´ dim Hq`1 pQ, BQq “ dim E q,q ´ bq pSq. q q By Theorem 3 this expression is equal to ff ˆ ˙ ˆ ˙ «ÿ ˆ ˙ q´1 q n n ÿ n r p`qr hq pSq` bq pSq “ hq pSq` p´1q bp pSq ´ p´1qp`qrbp pSq “ h2q pSq. q q q p“0 p“0 HOMOLOGY OF TORUS SPACES 35 1` The case q “ n follows from (7.4). Indeed, the term XE n,n survives, thus: ˆ ˙ n X 8 dim Hn pQ, BQq “ bn´1 pSq “ h1n pSq “ h2n pSq. dim E n,n “ n This proves Theorem 5. 8. Geometry of equivariant cycles 8.1. Orientations. In this section we restrict to the case when Q is a nice manifold with corners, X “ pQ ˆ T n q{„ is a manifold with locally standard torus action, λ — a characteristic map over Z defined on the poset S “ SQ . As before, suppose that all proper faces of Q are acyclic and orientable and Q itself is orientable. The subset XI , I ‰ ∅ is a submanifold of X, preserved by the torus action; XI is called a face manifold, codim XI “ 2|I|. Submanifolds Xtiu , corresponding to vertices i P VertpSq are called characteristic submanifolds, codim Xtiu “ 2. Fix arbitrary orientations of the orbit space Q and the torus T n . This defines the orientation of Y “ Q ˆ T n and X “ Y {„. Also choose an omniorientation, i.e. orientations of all characteristic submanifolds Xtiu . The choice of omniorientation defines characteristic values ωi P H1 pT n ; Zq without ambiguity of sign (recall that previously they were defined only up to units of k). To perform calculations with the spectral sequences XE and Y E we also need to orient faces of Q. Lemma 8.1 (Convention). The orientation of each simplex of S (i.e. the sign convention on S) defines the orientation of each face FI Ă Q. Proof. Suppose that I P S is oriented. Let i1 , . . . , in´q be the vertices of I, listed in a positive order (this is where the orientation of I comes in play). The corresponding face FI lies in the intersecion of facets Fi1 , . . . , Fin´q . The normal bundles νi to facets Fi have natural orientations, in which inward normal vectors are positive. Orient FI in such way that Tx FI ‘νii ‘. . .‘νin´q – Tx Q is positive.  Thus there are distinguished elements rFI s P Hdim FI pFI , BFI q. One checks that for I ă1 J the maps m0I,J : Hdim FI pFI , BFI q Ñ Hdim FI ´1 pBFI q Ñ Hdim FJ pFJ , BFJ q (see (3.5)) send rFI s to rJ : Is ¨ rFJ s. Thus the restriction maps H0 pI ă Jq send rFI s to rFJ s by the definition of H0 . The choice of omniorientation and orientations of I P S determines the orientation of each orbit T n {T λpIq by the following convention. Construction 8.2. Let i1 , . . . , in´q be the vertices of I, listed in a positive order. Recall that H1 pT n {T λpIq q is naturally identified with H1 pT n q{IpIqp1q . The basis rγ1s, . . . , rγq s P H1 pT n {T λpIq q, rγl s “ γl ` IpIqp1q is defined to be positive if the basis pωi1 , . . . , ωin´q , γ1, . . . , γq q is positive in H1 pT n q. The orientation of T n {T λpIq Ź determines a distinguished “volume form” ΩI “ l rγl s P Hq pT n {T λpIq ; Zq. HOMOLOGY OF TORUS SPACES 36 The omniorientation and the orientation of S also determine the orientation of each manifold XI in a similar way. All orientations are compatible: rXI s “ rFI sbΩI . 8.2. Arithmetics of torus quotients. Let us fix a positive basis e1 , . . . , en of the lattice H1 pT n ; Zq. Let pλi,1 , . . . , λi,n q be the coordinates of ωi in this basis for each i P VertpSq. The following technical lemma will be used in subsequent computations. Lemma 8.3. Let I P S, I ‰ ∅ be a simplex with vertices ti1 , . . . , in´q u listed in a positive order. Let A “ tj1 ă . . . ă jq u Ă rns be a subset of indices and eA “ ej1 ^. . .^ejq the corresponding element of Hq pT n q. Consider the map ρ : T n Ñ T n {T λpIq . Then ρ˚ peA q “ CA,I ΩI P Hq pT n {T λpIq q with the constant: CA,I “ sgnA det pλi,j qiPti1 ,...,in´q u jPrnszA where sgnA “ ˘1 depends only on A Ă rns. Proof. Let pbl q “ pωi1 , . . . , ωin´q , γ1, . . . , γq q be a positive basis of lattice H1 pT n , Zq. Thus bl “ Uel , where the matrix U has the form ˛ ¨ λi1 ,1 . . . λin´q ,1 ˚ ˚ ˚ λi1 ,2 . . . λin´q ,2 ˚ ˚‹ ‹ ˚ U “˚ . .. .. ‹ .. . . . ˝ . . . .‚ . λi1 ,n . . . λin´q ,n ˚ ˚ We have det U “ 1 since both bases are positive. Consider the inverse matrix V “ U ´1 . Thus ÿ det pVj,αqjPA bα1 ^ . . . ^ bαq . eA “ ej1 ^ . . . ^ ejq “ M “tα1 ă...ăαq uĂrns αPM After passing to quotient Λ Ñ Λ{IpIq all summands with M ‰ tn ´ q ` 1, . . . , nu vanish. When M “ tn ´ q ` 1, . . . , nu, the element bn´q´1 ^ . . . ^ bn “ γ1 ^ . . . ^ γq goes to ΩI . Thus . CA,I “ det pVj,αqjPA αPtn´q`1,...,nu Now apply Jacobi’s identity which states the following (see e.g. [3, Sect.4]). Let U be an invertible n ˆ n-matrix, V “ U ´1 , M, N Ă rns subsets of indices, |M| “ |N| “ q. Then sgnM,N det pVr,s qrPM “ det pUr,s qrPrnszN , sPN det U sPrnszM ř where sgnM,N “ p´1q rPrnszN r` sign depends only on A Ă rns. ř sPrnszM s . In our case N “ tn ´ q ` 1, . . . , nu; thus the  HOMOLOGY OF TORUS SPACES 37 8.3. Face ring and linear system of parameters. Recall the definition of a face ring of a simplicial poset S. For I1 , I2 P S let I1 _ I2 Ă S denote the set of least upper bounds, and I1 X I2 P S — the intersection of simplices (it is well-defined and unique if I1 _ I2 ‰ ∅). Definition 8.4. The face ring krSs is the quotient ring of krvI | I P Ss, deg vI “ 2|I| by the relations ÿ vJ , v∅ “ 1, vI1 ¨ vI2 “ vI1 XI2 ¨ JPI1 _I2 where the sum over an empty set is assumed to be 0. Characteristic map λ determines the set of linear forms tθ1 , . . . , θn u Ă krSs, ř θj “ iPVertpSq λi,j vi . If J P S is a maximal simplex, |J| “ n, then (8.1) the matrix pλi,j qiďJ is invertible over k jPrns by the p˚k q-condition. Thus the sequence tθ1 , . . . , θn u Ă krSs is a linear system of parameters in krSs (see, e.g.,[4, lemma 3.5.8]). It generates an ideal pθ1 , . . . , θn q Ă krSs which we denote by Θ. The face ring krSs is an algebra with straightening law (see, e.g. [4, §.3.5]). Additively it is freely generated by the elements Pσ “ vI1 ¨ vI2 ¨ . . . ¨ vIt , σ “ pI1 ď I2 ď . . . ď It q. Lemma 8.5. The elements rvI s “ vI ` Θ additively generate krSs{Θ. Proof. Consider an element Pσ with |σ| ě 2. Using relations in the face ring, we express Pσ “ vI1 ¨ . . . ¨ vIt asřvi ¨ vI1 zi ¨ . . . ¨ vIt , for some vertex i ď I1 . The element vi can be expressed as i1 ęIt ai1 vi1 modulo Θ according to (8.1) (we can exclude all vi corresponding to the vertices of some maximal simplex J Ě It ). Thus vi vIt is expressed as a combination of vIt1 for It1 ą1 It . Therefore, up to ideal Θ, the element Pσ is expressed as a linear combination of elements Pσ1 which have either smaller length t (in case |I1 | “ 1) or smaller I1 (in case |I1 | ą 1). By iterating this descending process, the element Pσ `Θ P krSs{Θ is expressed as a linear combination of rvI s.  Note that the proof works for k “ Z as well. 8.4. Linear relations on equivariant (co)cycles. Let HT˚ pXq be a T n -equivariant cohomology ring of X. Any proper face of Q is acyclic, thus has a vertex. Therefore, there is the injective homomorphism krSs ãÑ HT˚ pXq, which sends vI to the cohomology class, equivariant Poincare dual to rXI s (see [9, Lemma 6.4]). The inclusion of a fiber in the Borel construction, X Ñ X ˆT ET n , HOMOLOGY OF TORUS SPACES 38 induces the map HT˚ pXq Ñ H ˚ pXq. The subspace V of H˚ pXq, Poincare dual to the image of g : krSs ãÑ HT˚ pXq Ñ H ˚ pXq (8.2) À 8 is generated by the elements rXI s, thus coincides with the 8-border: V “ q XE q,q Ă H˚ pXq. Now let us describe explicitly the linear relations on rXI s in H˚ pXq. Note that the elements rXI s “ rFI s b ΩI can also be considered as the free generators of the k-module àX 1 à à E q,q “ Hq pFI , BFI q b Hq pT n {T λpIq q. q q |I|“n´q The free k-module on generators rXI s is denoted by xrXI sy. Proposition 8.6. Let CA,J be the constants defined in lemma 8.3. There are only two types of linear relations on classes rXI s in H˚ pXq: (1) For each J P S, |J| “ n ´ q ´ 1, and A Ă rns, |A| “ q there is a relation ÿ RJ,A “ rI : JsCA,I rXI s “ 0; Ią1 J 8 (2) Let β be a homology class from Impδq`1 : Hq`1 pQ, BQq Ñ Hq pBQqq Ď BQE q,0 ř 1 for q ď n ´ 2, and let |I|“n´q BI rFI s P BQE q,0 be a chain representing β. Then ÿ 1 Rβ,A “ BI CA,I rXI s “ 0. |I|“n´q ˚ ˚ Proof. This follows from the structure of the map f˚ : QE ˚,˚ ˆH˚ pT n q Ñ XE ˚,˚ , lemma 8.3 and Theorem 1. Relations on rXI s appear as the images of the differr entials hitting XE q,q , r ě 1. Relations of the first type, RJ,A , are the images of À 1 1 2 d1X : XE q`1,q Ñ XE q,q . In particular, q XE q,q is identified with xrXI sy{xRJ,A y. Relations of the second type are the images of higher differentials drX , r ě 2.  Now we check that relations of the first type are exactly the relations in the quotient ring krSs{Θ. Proposition 8.7. Let ϕ : xrXI sy Ñ krSs be the degree reversing linear map, which sends rXI s to vI . Then ϕ descends to the isomorphism ϕ̃ : xrXI sy{xRJ,A y Ñ krSs{Θ. Proof. (1) First we prove that ϕ̃ is well defined. The image of RJ,A is the element ÿ ϕpRJ,A q “ rI : JsCA,I vI P krSs. Ią1 J HOMOLOGY OF TORUS SPACES 39 Let us show that ϕpRJ,A q P Θ. Let s “ |J|, and consequently, |I| “ s ` 1, |A| “ n ´ s ´ 1. Let rnszA “ tα1 ă . . . ă αs`1 u and let tj1 , . . . , js u be the vertices of J listed in a positive order. Consider s ˆ ps ` 1q matrix: ˛ ¨ λj1 ,α1 . . . λj1 ,αs`1 ˚ .. ‹ .. D “ ˝ ... . . ‚ λjs ,α1 . . . λjs ,αs`1 Denote by Dl the square submatrix obtained from D by deleting i-th column and let al “ p´1ql`1 det Dl . We claim that ϕpRJ,A q “ ˘vJ ¨ pa1 θα1 ` . . . ` as`1 θαs`1 q ř Indeed, after expanding each θl as iPVertpSq λi,l vi , all elements of the form vJ vi with i ă J cancel; others give rI : JsCA,I vI for I ą1 J according to lemma 8.3 and cofactor expansions of determinants (the incidence sign arise from shuffling columns). Thus ϕ̃ is well defined. (2) ϕ̃ is surjective by lemma 8.5. (3) The dimensions of both spaces are equal. Indeed, dimxrXI s | |I| “ n ´ 2 qy{xRJ,A y “ dim XE q,q “ h1n´q pSq by Theorem 4. But dimpkrSs{Θqpn´qq “ h1n´q pSq by Schenzel’s theorem [15], [16, Ch.II,§8.2], (or [13, Prop.6.3] for simplicial posets) since S is Buchsbaum. (4) If k is a field, then we are done. This implies the case k “ Z as well.  In particular, this proposition describes the additive structure of krSs{Θ in terms of the natural additive generators vI . Poincare duality in X yields Corollary 8.8. The map g : krSs Ñ H ˚ pXq factors through krSs{Θ and the kernel of g̃ : krSs{Θ Ñ H ˚ pXq is additively generated by the elements ÿ L1β,A “ BI CA,I vI |I|“n´q where q ď n´2, β P Impδq`1 : Hq`1 pQ, BQq Ñ Hq pBQqq, chain representing β, and A Ă rns, |A| “ q. ř |I|“n´q BI rFI s is a cellular Remark 8.9. The ideal Θ Ă krSs coincides with the image of the natural map H ą0 pBT n q Ñ HT˚ pXq. So the fact that Θ vanishes in H ˚ pXq is not surprising. The interesting thing is that Θ vanishes by geometrical reasons already in the second term of the spectral sequence, while other relations in H ˚ pXq are the consequences of higher differentials. Remark 8.10. From the spectral sequence follows that the element L1β,A P krSs{Θ does not depend on the cellular chain, representing β. All such chains À 2 produce the same element in q XE q,q “ xrXI sy{xRJ,A y – krSs{Θ. Theorem 2 also implies that the relations tL1β,A u are linearly independent in krSs{Θ when β runs over some basis of Im δq`1 and A runs over all subsets of rns of cardinality q. HOMOLOGY OF TORUS SPACES 40 9. Examples and calculations 9.1. Quasitoric manifolds. Let Q be n-dimensional simple polytope. Then S “ SQ “ BQ˚ is the boundary of the polar dual polytope. In this case Q – Cone |S|. Given a characteristic map λ : VertpKq Ñ Tn we construct a space X “ pQ ˆ T n q{„ which is a model of quasitoric manifold [7]. Poset S is a sphere thus h2i pSq “ h1i pSq “ hi pSq. Since δn : Hn pQ, BQq Ñ Hn´1pBQq is an isomorphism, Theorem 2 implies 2 X 2 E p,q “ 0 for p ‰ q. By Theorems 3 and 5, dim XE q,q “ hq pSq “ hn´q pSq. Spectral ˚ sequence XE ˚,˚ collapses at its second term, thus dim H2q pXq “ hq pSq “ hn´q pSq for 0 ď q ď n and dim H2q`1 pXq “ 0 which is well known. For bigraded Betti numbers proposition 5.6 implies Hi,j pXq “ 0 if i ‰ j, and dim Hi,i pXq “ hi pSq. 9.2. Homology polytopes. Let Q be a manifold with corners such that all its proper faces as well as Q itself are acyclic. Such objects were called homology polytopes in [9]. In this case everything stated in the previous paragraph remains valid, thus dim H2q pXq “ hq pSq “ hn´q pSq for 0 ď q ď n, and dim H2q`1 pXq “ 0 (see [9]). 9.3. Origami toric manifolds. Origami toric manifolds appeared in differential geometry as generalizations of symplectic toric manifolds (see [5],[10]). The original definition contains a lot of subtle geometrical details and in most part is irrelevant to this paper. Here we prefer to work with the ad hoc model, which captures most essential topological properties of origami manifolds. Definition 9.1. Topological toric origami manifold X 2n is a manifold with locally standard action T n ñ X such that all faces of the orbit space including X{T itself are either contractible or homotopy equivalent to wedges of circles. As before consider the canonical model. Let Qn be a nice manifold with corners in which every face is contractible or homotopy equivalent to a wedge of b1 circles. Every principal T n -bundle Y over Q is trivial (because H 2 pQq “ 0), thus Y “ QˆT n . Consider the manifold X “ Y {„ associated to some characteristic map over Z. Then X is a topological origami toric manifold. To apply the theory developed in this paper we also assume that all proper faces of Q are acyclic (in origami case this implies contractible) and Q itself is orientable. Thus, in particular, Q is a Buchsbaum manifold. First, describe the exact sequence of the pair pQ, BQq. By Poincare–Lefchetz duality: $ ’ &k, ifŽq “ n; n´q Hq pQ, BQq – H pQq – H 1 p b1 S 1 q – kb1 , if q “ n ´ 1; ’ %0, otherwise. In the following let m denote the number of vertices of S (the number of facets of Q). Thus h11 pSq “ h1 pSq “ m ´ n. Consider separately three cases: HOMOLOGY OF TORUS SPACES 41 (1) n “ 2. In this case Q is an orientable 2-dimensional surface of genus 0 with b1 ` 1 boundary components. Thus BQ is a disjoint union of b1 ` 1 circles and long exact sequence in homology has the form: 0 k k k kb1 `1 k δ kb1 k 0 2 H2 pQq ÝÑ H2 pQ, BQq ÝÑ H1 pBQq ÝÑ H1 pQq ÝÑ δ 0 1 ÝÑ H1 pQ, BQq ÝÑ H0 pBQq ÝÑ H0 pQq ÝÑ H0 pQ, BQq k kb1 k kb1 `1 k k k 0 2 The second term XE ˚,˚ of spectral sequence for X is given by Theorem 2. It is shown on a figure below (only ranks are written to save space). 2 1 1 0 b1 ` 1 m´2 b1 b1 2b1 ´1 b1 0 1 2 2 2 The only nontrivial higher differential is d2 : XE 2,´1 Ñ XE 0,0 ; it coincides with 2 the composition of δ1 b idH0 pT 2 q and injective map f˚2 : H0 pP q b H0 pT 2 q Ñ XE 0,0 . 8 8 8 8 Thus d2 is injective, and dim XE 2,2 “ dim XE 0,0 “ 1; dim XE 2,1 “ dim XE 1,0 “ b1 ; 8 8 dim XE 1,1 “ m ´ 2; dim XE 2,0 “ 2b1 . Finally, $ ’ &1, if i “ 0, 4; dim Hi pXq “ b1 , if i “ 1, 3; ’ %m ´ 2 ` 2b , if i “ 2. 1 This coincides with the result of computations in [14], concerning the same object. This result can be obtained simply by proposition 5.6: dim H0,0 pXq “ dim H2,2 pXq “ 1, dim H1,0 pXq “ dim H1,2 pXq “ b1 , dim H2,2 pXq “ m ´ 2 ` 2b1 . HOMOLOGY OF TORUS SPACES 42 (2) n “ 3. In this case the exact sequence of pQ, BQq splits in three essential parts: δ 3 H3 pQq ÝÑ H3 pQ, BQq ÝÑ H2 pBQq ÝÑ H2 pQq k k k 0 k 0 δ 2 H1 pBQq ÝÑ H1 pQq ÝÑ H1 pQ, BQq H2 pQq ÝÑ H2 pQ, BQq ÝÑ k kb1 k 0 k kb1 k 0 δ 1 H0 pBQq ÝÑ H0 pQq ÝÑ H0 pQ, BQq H1 pQ, BQq ÝÑ k k k 0 k 0 2 By Theorems 2, 4, XE p,q has the form 1 3 h11 b1 h12 0 3b1 2b1 0 3b1 2 1 0 h13 ´1 b1 0 1 2 3 2 2 2 There are two nontrivial higher differentials: d2 : XE 3,0 Ñ XE 1,1 and d2 : XE 3,´1 Ñ 8 8 8 8 X 2 E 1,0 ; both are injective. Thus dim XE 3,3 “ dim XE 0,0 “ 1; dim XE 3,1 “ dim XE 1,0 “ 8 8 8 b1 ; dim XE 2,2 “ h11 ; dim XE 3,1 “ 3b1 ; dim XE 1,1 “ h12 ´ 3b1 . Therefore, $ 1, if i “ 0, 6; ’ ’ ’ ’ ’ &b1 , if i “ 1, 5; dim Hi pXq “ h11 ` 3b1 , if i “ 4; ’ ’ ’ h12 ´ 3b1 , if i “ 2; ’ ’ % 0, if i “ 3. (3) n ě 4. In this case lacunas in the exact sequence for pQ, BQq imply that δi : Hi pQ, BQq Ñ Hi´1 pBQq is an isomorphism for i “ n´1, n, and is trivil otherwise. HOMOLOGY OF TORUS SPACES We have 43 $ ’ ’Hn pQ, BQq – k, if i “ n ´ 1; ’ ’ b ’ &Hn´1 pQ, BQq – k 1 , if i “ n ´ 2; Hi pBQq – H1 pQq – kb1 if i “ 1; ’ ’ ’ H0 pQq – k, if i “ 0; ’ ’ % 0, o.w. 2 By Theorems 2 and 4, XE p,q has the form 1 h11 h12 h13 .. h1n ´1 . h1n´1 0 0 `n˘ 0 0 0 b1 ` `n˘ 1 `n˘ 0 n ` ` b1 0 .. . b1 0 `n˘ n n´2 ˘ b1 n n´3 1 `n˘ 0 0 ˘ b1 ˘ b1 0 ` b1 n n´1 ˘ b1 n n´3 .. . 0 `n˘ b1 b1 ` ˘ 8 8 Thus we get: dim XE q,q “ h1n´q , if q ‰ n ´ 2; dim XE n´2,n´2 “ h12 ´ n2 b1 8 8 8 if q “ n ´ 2; dim XE n,n´1 “ dim XE 1,0 “ b1 ; dim XE n,n´2 “ nb1 . Finally, by proposition 5.6, dim H1,0 pXq “ dim Hn´1,n pXq “ b1 , dim Hn´1,n´1 pXq “ h11 ` nb1 , ` ˘ dim Hn´2,n´2 pXq “ h12 ´ n2 b1 , and dim Hi,i pXq “ h1n´i for i ‰ n ´ 1, n ´ 2. The differential hitting the marked position produces additional relations (of the second type) on the cycles rXI s P H 2n´4 pXq. These relations are described explicitly by proposition 8.6. Dually, this consideration shows that the map krSs{Θ Ñ H ˚ pXq has a nontrivial kernel only in degree 4. The generators of this kernel are described by corollary 8.8. 10. Concluding remarks Several questions concerning the subject of this paper are yet to be answered. HOMOLOGY OF TORUS SPACES 44 (1) Of course, the main question which remains open is the structure À ofX multi8 ˚ plication in the cohomology ring H pXq. The border module q E q,q Ă H˚ pXq represents an essential part of homology; the structure of multiplication on the corresponding subspace in cohomology can be extracted from the ring homomorphism krSs{Θ Ñ H ˚ pXq. Still there are cocycles which do not come from krSs and their products should be described separately. Proposition 5.6 suggests, that some products can be described via the multiplication in H ˚ pQ ˆ T n q – H ˚ pQq b H ˚ pT n q. This requires further investigation. À X 8 (2) It is not clear yet, if there is a torsion in the border module q E q,q in case k “ Z. Theorems 3,4,5 describe only the rank of the free part of this group, but the structure (and existence) of torsion remains open. Note that the homology of X itself can have a torsion. Indeed, the groups H˚ pQq, H˚ pQ, BQq can contain arbitrary torsion, and these groups appear in the description of H˚ pXq by proposition 5.6. (3) Corollary 8.8 describes the kernel of the map krSs{Θ Ñ H ˚ pXq. It seems that the elements of this kernel lie in a socle of krSs{Θ, i.e. in a submodule tx P krSs{Θ | pkrSs{Θq` x “ 0u. The existence of such elements is guaranteed in general by the Novik–Swartz theorem [13]. If the relations L1β,A do not lie in a socle, their existence would give refined inequalities on h-numbers of Buchsbaum posets. (4) Theorem 6 establish certain connection between the sheaf of ideals generated by linear elements and the cosheaf of ideals generated by exterior products. This connection should be clarified and investigated further. In particular, statement 6.6 can probably lead to the description of homology for the analogues of moment-angle complexes, i.e. the spaces of the form X “ Y {„, where Y is an arbitrary principal T N -bundle over Q. (5) There is a hope, that the argument of section 6 involving two spectral sequences for a sheaf resolution can be generalized to non-Buchsbaum case. (6) The real case, when T N is replaced by ZN2 , can, probably, fit in the same framework. Acknowledgements I am grateful to prof. Mikiya Masuda for his hospitality and for the wonderful environment with which he provided me in Osaka City University. The problem of computing the cohomology ring of toric origami manifolds, which he posed in 2013, was a great motivation for this work (and served as a good setting to test working hypotheses). Also I thank Shintaro Kuroki from whom I knew about h1 and h2 -vectors and their possible connection to torus manifolds. HOMOLOGY OF TORUS SPACES 45 References [1] A. A. Ayzenberg, V. M. Buchstaber, Nerve complexes and moment-angle spaces of convex polytopes, Proc. of the Steklov Institute of Mathematics, Vol.275, Issue 1, pp. 15–46, 2011. [2] A. Björner, Posets, regular CW complexes and Bruhat order, European J. Combin. 5 (1984), 7–16. [3] Richard A. Brualdi, Hans Schneider, Determinantal Identities: Gauss, Schur, Cauchy, Sylvester, Kronecker, Jacobi, Binet, Laplace, Muir, and Cayley, Linear Algebra and its Applications, 52/53:769–791 (1983). [4] Victor Buchstaber, Taras Panov, Toric Topology, preprint arXiv:1210.2368 [5] A. Cannas da Silva, V. Guillemin and A. R. Pires, Symplectic Origami, IMRN 2011 (2011), 4252–4293, arXiv:0909.4065. [6] Justin M. Curry, Sheaves, Cosheaves and Applications, arXiv:1303.3255v1 [7] M. Davis, T. Januszkiewicz, Convex polytopes, Coxeter orbifolds and torus actions, Duke Math. J., 62:2 (1991), 417–451. [8] Hiroshi Maeda, Mikiya Masuda, Taras Panov, Torus graphs and simplicial posets, Adv. Math. 212 (2007), no. 2, 458–483. [9] Mikiya Masuda, Taras Panov, On the cohomology of torus manifolds, Osaka J. Math. 43 (2006), 711–746. [10] Mikiya Masuda, Seonjeong Park, Toric origami manifolds and multi-fans, Preprint arXiv:1305.6347 [11] John McCleary, A User’s Guide to Spectral Sequences, second edition, Cambridge studies in advanced mathematics; 58. [12] Clint McCrory, Zeeman’s filtration on homology, Trans.of the AMS, Vol.250, 1979. [13] Isabella Novik, Ed Swartz, Socles of Buchsbaum modules, complexes and posets, Adv. Math., 222 (2009), 2059–2084. [14] Mainak Poddar, Soumen Sarkar, A class of torus manifolds with nonconvex orbit space, Preprint arXiv:1109.0798 [15] Peter Schenzel, On the Number of Faces of Simplicial Complexes and the Purity of Frobenius, Math. Zeitschrift 178, 125–142 (1981). [16] R. Stanley, Combinatorics and Commutative Algebra. Boston, MA: Birkhäuser Boston Inc., 1996. (Progress in Mathematics V. 41). [17] Takahiko Yoshida, Local torus actions modeled on the standard representation, Advances in Mathematics 227 (2011), pp. 1914–1955. Osaka City University E-mail address: ayzenberga@gmail.com 

ZEILBERGER’S KOH THEOREM AND THE STRICT UNIMODALITY OF q-BINOMIAL COEFFICIENTS arXiv:1311.4480v2 [math.CO] 1 Apr 2014 FABRIZIO ZANELLO Abstract. A recent nice result due to I. Pak and G. Panova is the strict unimodality
of the q-binomial coefficients a+b b q (see [2] and also [3] for a slightly revised version of their theorem). Since their proof used representation theory and Kronecker coefficients, the authors also asked for an argument that would employ Zeilberger’s KOH theorem. In this note, we give such a proof. Then, as a further application of our method, we also provide a
short proof of their conjecture that the difference between consecutive coefficients of a+b b q can get arbitrarily large, when we assume that b is fixed and a is large enough. A sequence c1 , c2 , . . . , ct is unimodal if it does not increase strictly after a strict decrease. It is symmetric if ci = ct−i for all i. The unimodality of the q-binomial coefficient   (1 − q)(1 − q 2 ) · · · (1 − q a+b ) a+b = , (1 − q)(1 − q 2 ) · · · (1 − q a ) · (1 − q)(1 − q 2 ) · · · (1 − q b ) b q which is easily proven to be a symmetric polynomial in q, is a classical and highly nontrivial result in combinatorics. It was first shown in 1878 by J.J. Sylvester, and has since received a number of other interesting proofs (see e.g. [4, 5, 7]). In particular, a celebrated paper
of K. O’Hara [1] provided a combinatorial proof for the unimodality of a+b . O’Hara’s b q argument was subsequently expressed in algebraic terms by D. Zeilberger [8] by means of
the beautiful KOH identity. This identity decomposes a+b into a finite sum of polynomials b q with nonnegative integer coefficients, which are all unimodal and symmetric about ab/2. More precisely, fix integers a ≥ b ≥ 2. For any given partition λ = (λ1 , λ2 , . . . ) of b, set P Yi = ij=1 λj for all i ≥ 1, and Y0 = 0. Then the KOH theorem can be stated as follows:
P Lemma 1 (KOH). a+b = λ⊢b Fλ (q), where b q Y j(a + 2) − Yj−1 − Yj+1 P 2 i≥1 (λ2i ) . Fλ (q) = q λj − λj+1 q j≥1 A recent nice result shown by I. Pak and G. Panova is a characterization of the strict
unimodality of q-binomial coefficients; i.e., they determined when a+b strictly increases b q 2010 Mathematics Subject Classification. Primary: 05A15; Secondary: 05A17. Key words and phrases. q-binomial coefficient; Gaussian polynomial; unimodality. 1 2 FABRIZIO ZANELLO from degree 1 to degree ⌊ab/2⌋ (see [2], and also [3] for a slightly revised version of the theorem). Since their argument employed the algebraic machinery of Kronecker coefficients, the authors asked whether a proof could also be given that uses Zeilberger’s KOH identity. We do this in the present note. Then, as a further pithy application of this method, using the KOH theorem we also give a very short proof of a conjecture stated in the same papers,
on the unbounded growth of the difference between consecutive coefficients of a+b . b q The next lemma is a trivial and probably well-known fact of which we omit the proof. Lemma 2. Let c and d be positive integers such that the q-binomial coefficient c+d d q
is strictly unimodal. Then, for any positive integer t ≤ cd such that t 6= cd − 2, the product
t+1
c+d is strictly unimodal (in all nonnegative degrees). d q 1 q Theorem 3 ([2, 3]). The q-binomial coefficient a+b b q a = b = 2 or b ≥ 5, with the exception of
is strictly unimodal if and only if (a, b) = (6, 5), (10, 5), (14, 5), (6, 6), (7, 6), (9, 6), (11, 6), (13, 6), (10, 7). Proof. We can assume that b ≥ 5, otherwise, as it is also noted in [2, 3], the result is easy to
show. By Lemma 1, since all terms in the KOH decomposition of a+b are unimodal and b q
symmetric with respect to ab/2, in order to show that a+b is strictly unimodal, it clearly b q suffices to determine, for each positive degree up to ab/2, some suitable KOH term that is
strictly increasing in that degree. We begin by showing that, for any a ≥ b ≥ 2, a+b b q strictly increases up to degree ab/2 − a for b even, and up to degree ab/2 − a/2 for b odd. Let b = 2m be even. Then the KOH term contributed by the partition λ = (λ1 = 2, . . . , λm−1 = 2, λm = 1, λm+1 = 1) of b is given by: Fλ (q) = q 2(m−1)    (m + 1)(a + 2) − (2m − 1) − 2m (m − 1)(a + 2) − 2(m − 2) − (2m − 1) 1 1 q q =q b−2    ab/2 + a − b + 3 ab/2 − a − b + 3 . 1 1 q q Notice that the product of the last two q-binomial coefficients is strictly increasing (by
1) from degree 0 to degree ab/2 − a − b + 2. Also, a+b is clearly strictly increasing b q from degree 1 to degree b − 2, since so is the usual partition function p(n) (see e.g. [6],
Chapter 1). From this, we easily have that a+b strictly increases from degree 1 to degree b q (ab/2 − a − b + 2) + (b − 2) = ab/2 − a. The proof for b = 2m + 1 odd, giving us that a+b b q
is strictly increasing up to degree ab/2 − a/2, is similar (using λ = (λ1 = 2, . . . , λm = 2, λm+1 = 1)) and thus will be omitted. ZEILBERGER’S KOH THEOREM AND THE STRICT UNIMODALITY OF q-BINOMIAL COEFFICIENTS3 Now, in order to show that for the desired values of a and b, a+b b q
strictly increases in each of the remaining degrees up to ab/2, we consider three cases depending on the residue of b modulo 3. We start with b ≡ 0 modulo 3, and assume that b ≥ 15. The KOH term corresponding to the partition λ = (b/3, b/3, b/3) of b is given by: Fλ (q) = q 6(b/3 2 )     3(a + 2) − 2b/3 − b b(b−3)/3 (3a − 2b + 6) + b/3 . =q b/3 b/3 q q Notice that b(b−3)/3 < ab/2−a, and 3a−2b+6 ≥ 15. Thus, it easily follows by induction

a+b implies that of , as desired. that, for b ≥ 15, the strict unimodality of (3a−2b+6)+b/3 b q b/3 q Let now b ≡ 1 modulo 3, and assume b ≥ 19. By considering the partition λ = ((b − 1)/3, (b − 1)/3, (b − 1)/3, 1) of b, we get: Fλ (q) = q (b−1)(b−4)/3    4a − 2b + 9 3a − 2b + 8 + (b − 4)/3 . 1 (b − 4)/3 q q It is easy to check that, under the current assumptions on a and b, we have (b−1)(b−4)/3 < ab/2 − a and (3a − 2b + 8)(b − 4)/3 ≥ (4a − 2b + 8) + 3. In particular, we are under the
hypotheses of Lemma 2. Thus, since 3a−2b+ 8 ≥ 15, the strict unimodality of a+b follows b q
3a−2b+8+(b−4)/3 , for all b ≥ 19, as we wanted to show. by induction from that of (b−4)/3 q The treatment of the case b ≡ 2 modulo 3, b ≥ 20, is analogous so we will omit the details. We only remark here that one considers the partition λ = ((b − 2)/3, (b − 2)/3, (b − 2)/3, 1, 1)
of b, whose contribution to the KOH expansion of a+b is: b q Fλ (q) = q (b−2)(b−5)/3    5a − 2b + 11 3a − 2b + 10 + (b − 5)/3 . 1 (b − 5)/3 q q
The strict unimodality of a+b , for b ≥ 20, then follows in a similar fashion from that of b q
3a−2b+10+(b−5)/3 , by employing Lemma 2 and induction. (b−5)/3 q Therefore, it remains to show the theorem for 5 ≤ b ≤ 17 (b 6= 15). We will assume for simplicity that a ≥ 2b + 13, the result being easy to verify directly for the remaining values
of a. The KOH term contributed by the partition (b) of b in the expansion of a+b is: b q F(b) (q) = q 2(2b)     a+2−b b(b−1) (a − 2b + 2) + b . =q b b q q Clearly, since a ≥ 2b + 13, we have b(b − 1) < ab/2 − a and a − 2b + 2 ≥ 15. Thus, by

induction, the strict unimodality of (a−2b+2)+b implies that of a+b , as desired.  b b q q In Remark 3.6 of [2, 3], the authors also conjectured that, roughly speaking, the difference between consecutive coefficients of a q-binomial coefficient is eventually larger than any fixed integer. As a further nice, and very brief, application of our method, we answer this 4 FABRIZIO ZANELLO conjecture in the positive using the KOH identity. (Just notice that unlike in the original formulation of the conjecture, our proof will assume that b is fixed and only a is large enough.) Proposition 4. Fix any integer d ≥ 1. Then there exist integers a0 , b and L such that, if
Pab a+b i = i=0 ci q , then ci − ci−1 ≥ d, for all indices L ≤ i ≤ ab/2 and for all a ≥ a0 . b q [k] [k] [k] Proof. Consider the partition λ[k] = (λ1 = b − k, λ2 = 1, . . . , λk+1 = 1) of b, where k ≥ 1.
It is easy to see that its contribution to the KOH identity for a+b is given by: b q    (k + 1)(a + 2) − 2b + 1 (b−k)(b−k−1) a − 2b + 2k + 2 + (b − k − 1) . Fλ[k] (q) = q 1 b−k−1 q q Set for instance b = 2d + 4 and a0 = (d + 2)(d + 3) + 6, where we can assume d ≥ 2. A standard computation gives that, for any a ≥ a0 and k ≤ b/2 − 2 = d, we are under the hypotheses of Lemma 2. Hence, by Theorem 3, each polynomial Fλ[k] (q) is strictly unimodal from degree (b − k)(b − k − 1) on, and the theorem now immediately follows by choosing P L = (b − 1)(b − 2) + 1 = 4d2 + 10d + 7 and considering the coefficients of dk=1 Fλ[k] (q).  1. Acknowledgements The idea of this paper originated during a visit to UCLA in October 2013, for whose invitation we warmly thank Igor Pak. We wish to thank the referee for a careful reading of our manuscript and for helpful suggestions that improved the presentation, and Igor Pak, Greta Panova, and Richard Stanley for several helpful discussions. We are also very grateful to Doron Zeilberger for comments (and for calling this a “proof from the book”). This work was done while the author was partially supported by a Simons Foundation grant (#274577). References [1] K. O’Hara: Unimodality of Gaussian coefficients: a constructive proof, J. Combin. Theory Ser. A 53 (1990), no. 1, 29–52. [2] I. Pak and G. Panova: Strict unimodality of q-binomial coefficients, C. R. Math. Acad. Sci. Paris 351 (2013), no. 11-12, 415–418. [3] I. Pak and G. Panova: Strict unimodality of q-binomial coefficients (new version), preprint. Available on the arXiv. [4] R. Proctor: Solution of two difficult combinatorial problems using linear algebra, Amer. Math. Monthly 89 (1982), no. 10, 721–734. [5] R. Stanley: Weyl groups, the hard Lefschetz theorem, and the Sperner property, SIAM J. Algebraic Discrete Methods 1 (1980), no. 2, 168–184. [6] R. Stanley: “Enumerative Combinatorics”, Vol. I, Second Ed., Cambridge University Press, Cambridge, U.K. (2012). [7] J.J. Sylvester: Proof of the hitherto undemonstrated fundamental theorem of invariants, Collect. Math. papers, Vol. 3, Chelsea, New York (1973), 117–126. ZEILBERGER’S KOH THEOREM AND THE STRICT UNIMODALITY OF q-BINOMIAL COEFFICIENTS5 [8] D. Zeilberger: Kathy O’Hara’s constructive proof of the unimodality of the Gaussian polynomials, Amer. Math. Monthly 96 (1989), no. 7, 590–602. Department of Mathematics, MIT, Cambridge, MA 02139-4307 matical Sciences, Michigan Tech, Houghton, MI 49931-1295 E-mail address: zanello@math.mit.edu, zanello@mtu.edu and Department of Mathe- 

MNRAS 000, 1–4 (2017) Preprint 15 June 2017 Compiled using MNRAS LATEX style file v3.0 An unbiased estimator for the ellipticity from image moments Nicolas Tessore⋆ arXiv:1705.01109v2 [astro-ph.CO] 14 Jun 2017 Jodrell Bank Centre for Astrophysics, University of Manchester, Alan Turing Building, Oxford Road, Manchester, M13 9PL, UK Accepted 2017 June 14. Received 2017 May 31; in original form 2017 May 07 ABSTRACT An unbiased estimator for the ellipticity of an object in a noisy image is given in terms of the image moments. Three assumptions are made: i) the pixel noise is normally distributed, although with arbitrary covariance matrix, ii) the image moments are taken about a fixed centre, and iii) the point-spread function is known. The relevant combinations of image moments are then jointly normal and their covariance matrix can be computed. A particular estimator for the ratio of the means of jointly normal variates is constructed and used to provide the unbiased estimator for the ellipticity. Furthermore, an unbiased estimate of the covariance of the new estimator is also given. Key words: gravitational lensing: weak – methods: statistical – techniques: image processing 1 INTRODUCTION A number of applications in astronomy require the measurement of shapes of objects from observed images. A commonly used shape descriptor is the ellipticity χ, which is defined in terms of the central image moments m pq as the complex number χ = χ1 + i χ2 = m20 − m02 + 2 i m11 . m20 + m02 (1) For example, in weak gravitational lensing, the gravitational field distorts the observed shapes of background galaxies, and this shear can be detected in ellipticity measurements. This is possible because the observed ellipticity of a source affected by gravitational lensing is (see e.g. Bartelmann & Schneider 2001) χ= χs + 2g + g2 χs∗ , 1 + |g| 2 + 2ℜ[g χs∗ ] (2) where χs is the intrinsic ellipticity of the source, and g = g1 + i g2 is the so-called reduced shear of the gravitational lens. In the limit of weak lensing, this can be approximated to linear order as χ ≈ χs + 2g . (3) Averaging over an ensemble of randomly-oriented background sources, i.e. h χs i = 0, the weak lensing equation (3) yields h χi = 2hgi . (4) The observed ellipticities are thus a direct estimator for the shear g from gravitational lensing. Similar ideas are used in Cosmology (for a recent review, see Kilbinger 2015). Here, cosmic shear from the large-scale structure of the universe imprints a specific signature onto the ellipticity two-point correlation functions, ξij (r) = χi (x) χj (y) |x−y |=r , ⋆ Email: nicolas.tessore@manchester.ac.uk © 2017 The Authors (5) where the average is taken over pairs of sources with the given separation r on the sky. Note that both the one-point function (4) and the two-point function (5) depend on the mean ellipticity over a potentially large sample of sources. In practice, an estimator χ̂ is used to measure the ellipticities of observed sources. In order to not introduce systematic errors into applications such as the above, the ellipticity estimator χ̂ must be unbiased, i.e. E[ χ̂] = χ. One of the biggest problems for estimators that directly work on the data is the noise bias (Refregier et al. 2012) arising from pixel noise in the observations. For example, the standard approach to moment-based shape measurement is to obtain estimates m̂ pq of the second-order moments from the (noisy) data and use (1) directly as an ellipticity estimate, ê = m̂20 − m̂02 + 2 i m̂11 . m̂20 + m̂02 (6) The statistical properties of this estimator have been studied by Melchior & Viola (2012) and Viola et al. (2014), who assumed that the image moments are jointly normal with some given variance and correlation. The estimator (6) then follows the distribution of Marsaglia (1965, 2006) for the ratio of jointly normal variates. None of the moments of this distribution exist, and even for a finite sample, small values in the denominator can quickly cause significant biases and large variances. The estimator (6) is thus generally poorly behaved, unless the signal-to-noise ratio of the data is very high. Here, a new unbiased estimator for the ellipticity χ from the second-order image moments is proposed. First, it is shown that for normally-distributed pixel noise with known covariance matrix and a fixed centre, the relevant combinations of image moments are indeed jointly normal and that their covariance matrix can easily be computed. In the appendix, an unbiased estimator for the ratio of the means of jointly normal variates is constructed, which can subsequently be applied to the image moments. This produces the ellipticity estimate, as well as unbiased estimates of its variance and covariance. L2 N. Tessore 2 AN UNBIASED ESTIMATOR FOR THE ELLIPTICITY It is assumed that the data is a random vector d = (d1, d2, . . . ) of pixels following a multivariate normal distribution centred on the unknown true signal µ, d ∼ N (µ, Σ) , m20 − m02 − m00 (π20 − π02 ) + 2 i m11 − 2 i m00 π11 , m20 + m02 − m00 (π20 + π02 ) (8) from the central moments π pq of the (normalised) PSF. Fixing a centre ( x̄, ȳ), the relevant combinations of moments to compute the ellipticity from data d are thus Õ
αi di , αi = wi (xi − x̄)2 − (yi − ȳ)2 − π20 + π02 , u= i v= Õ βi di ,
βi = wi 2 (xi − x̄) (yi − ȳ) − 2 π11 , s= Õ γi di ,
γi = wi (xi − x̄)2 + (yi − ȳ)2 − π20 − π02 , i i (9) where wi is the window function of the observation. To obtain the true ellipticity estimate of the signal, and for the PSF correction (8) to remain valid, the window function must be unity over the support of the signal.1 Due to the linearity in the pixel values di , the vector (u, v, s) can be written in matrix form, (u, v, s) = M d , (10) where the three rows αi , βi , γi of matrix M are defined by (9). The random vector (u, v, s) is hence normally distributed, (u, v, s) ∼ N (µ uvs, Σuvs ) , (11) with unknown mean µ uvs = (µu , µv, µs ) = M µ and known 3 × 3 covariance matrix Σuvs = M Σ MT with entries Õ Õ αi β j Σij , αi α j Σij , Σuv = Σvu = Σuu = Σvv = Σss = i, j i, j Õ αi γ j Σij , i, j Õ Õ Õ βi γ j Σij , βi β j Σij , Σus = Σsu = i, j γi γ j Σij , Σvs = Σsv = (12) i, j i, j where Σij are the entries of the covariance matrix Σ of the pixel noise. Hence the covariance matrix Σuvs of the moments can be computed if the pixel noise statistics are known. The true ellipticity (8) of the signal can be written in terms of the mean values of the variates u, v and s defined in (9) as χ = χ1 + i χ2 = µu + i µ v . µs (13) The problem is to find an unbiased estimate of χ1 and χ2 from the 1 p = u−as, (7) where Σ is the covariance matrix for the noise, which is assumed to be known but not restricted to a particular diagonal shape. The observed signal usually involves a point-spread function (PSF), and it is further assumed that this effect can be approximated as a linear convolution of the true signal and the discretised PSF. In this case, definition (1) can be extended to obtain the true ellipticity of the object before convolution, χ= observed values of u, v and s. In appendix A, an unbiased estimator is given for the ratio of the means of two jointly normal random variables. It can be applied directly to the ellipticity (13). First, two new variates p and q are introduced, This is in contrast to the weight functions that are sometimes used in moment-based methods to reduce the influence of noise far from the centre. q = v − bs , a = Σus /Σss , b = Σvs /Σss , (14) where a, b are constants. This definition corresponds to (A2) in the univariate case, and therefore both (p, s) and (q, s) are pairs of independent normal variates. The desired estimator for the ellipticity is then χ̂1 = a + p ĝ(s) , χ̂2 = b + q ĝ(s) . (15) Because the mean µs is always positive for a realistic signal, the function ĝ(s) is given by (A3). From the expectation E[ χ̂1 ] = a + E[p] E[ĝ(s)] = a + (µu − a µs )/µs = χ1 , E[ χ̂2 ] = b + E[q] E[ĝ(s)] = b + (µv − b µs )/µs = χ2 , (16) it follows that χ̂ = χ̂1 + i χ̂2 is indeed an unbiased estimator for the true ellipticity χ of the signal. In addition, an unbiased estimate of the the variance of χ̂1 and χ̂2 is provided by (A9),
2

V̂ar[ χ̂1 ] = p2 ĝ(s) − (p2 − Σuu )/Σss + a2 1 − s ĝ(s) ,
2

V̂ar[ χ̂2 ] = q2 ĝ(s) − (q2 − Σvv )/Σss + b2 1 − s ĝ(s) . (17) Similarly, there is an unbiased estimator for the covariance,
2

Ĉov[ χ̂1, χ̂2 ] = pq ĝ(s) − (pq − Σuv )/Σss + ab 1 − s ĝ(s) . (18) It follows that the individual estimates of the ellipticity components are in general not independent. However, for realistic pixel noise, window functions and PSFs, the correlations between u, v and s, and hence χ̂1 and χ̂2 , can become very small. To demonstrate that the proposed estimator is in fact unbiased under the given assumptions of i) normal pixel noise with known covariance, ii) a fixed centre for the image moments, and iii) the discrete convolution with a known PSF, a Monte Carlo simulation was performed using mock observations of an astronomical object. The images are postage stamps of 49×49 pixels, containing a centred source that is truncated near the image borders. A circular aperture is used as window function. The source is elliptical and follows the light profile of de Vaucouleurs (1948). The ellipse containing half the total light has a 10 pixel semi-major axis and ellipticity as specified. Where indicated, the signal is convolved with a Gaussian PSF with 5 pixel FWHM. The pixel noise is uncorrelated with unit variance. The normalisation N of the object (i.e. the total number of counts) varies to show the effect of the signal-to-noise ratio on the variance of the estimator.2 The signal ellipticity χ is computed from the image before the PSF is applied and noise is added. The mean of the estimator χ̂ is computed from 106 realisations of noise for the same signal. Also computed are the square root of the sample variance Var[ χ̂] and the mean of the estimated variance V̂ar[ χ̂], respectively. The results shown in Table 1 indicate that the estimator performs as expected. 2 To compare the results to a given data set, it is then possible to scale the data so that the noise has unit variance, and compare the number of counts. For example, the control-ground-constant data set of the GREAT3 challenge (Mandelbaum et al. 2014) has mostly N = 500–1000. MNRAS 000, 1–4 (2017) An unbiased estimator for the ellipticity L3 Table 1. Monte Carlo results for the unbiased ellipticity estimator χ1 0.1107 χ2 0.0000 PSF counts χ̂1 χ̂2 Var[ χ̂1 ]1/2 Var[ χ̂2 ]1/2 no 500 1000 2000 500 1000 2000 500 1000 2000 500 1000 2000 500 1000 2000 500 1000 2000 0.1113 (08) 0.1105 (02) 0.1108 (01) 0.1115 (07) 0.1111 (02) 0.1108 (01) 0.1820 (07) 0.1817 (02) 0.1822 (01) 0.1802 (07) 0.1819 (02) 0.1819 (01) −0.0005 (08) −0.0001 (02) −0.0002 (01) 0.0001 (07) 0.0002 (02) 0.0001 (01) 0.0006 (11) −0.0002 (02) 0.0000 (01) 0.0006 (06) −0.0001 (02) 0.0000 (01) −0.1840 (09) −0.1842 (02) −0.1842 (01) −0.1840 (06) −0.1845 (02) −0.1841 (01) 0.5489 (15) 0.5494 (03) 0.5493 (01) 0.5473 (10) 0.5488 (03) 0.5491 (01) 0.762 0.216 0.104 0.717 0.216 0.104 0.704 0.225 0.108 0.658 0.223 0.108 0.793 0.248 0.117 0.720 0.244 0.116 1.112 0.214 0.103 0.598 0.214 0.103 0.853 0.226 0.108 0.616 0.224 0.108 1.451 0.323 0.149 0.957 0.315 0.147 yes 0.1820 −0.1842 no yes 0.0000 0.5492 no yes 3 CONCLUSION & DISCUSSION The unbiased estimator (15) provides a new method for ellipticity measurement from noisy images. Its simple and analytic form allows quick implementation and fast evaluation, and statistics for the results can be obtained directly with unbiased estimates of the variance (17) and covariance (18). Using an unbiased estimator for the ellipticity of the signal µ eliminates the influence of noise from a subsequent analysis of the results (the so-called “noise bias” in weak lensing, Refregier et al. 2012). However, depending on the application, other kinds of biases might exist even for a noise-free image. For example, due to the discretisation of the image, the signal ellipticity can differ from the intrinsic ellipticity of the observed object. This “pixellation bias” (Simon & Schneider 2016) remains an issue in applications such as weak lensing, where the relevant effects must be measured from the intrinsic ellipticity of the objects. Furthermore, a fixed centre ( x̄, ȳ) for the moments has been assumed throughout. For a correct ellipticity estimate, this must be the centroid of the signal, which is usually estimated from the data itself. Centroid errors (Melchior & Viola 2012) might therefore ultimately bias the ellipticity estimator or increase its variance, although this currently does not seem to be a significant effect. In practice, additional biases might arise when the assumed requirements for the window function and PSF are not fulfilled by the data. The estimator should therefore always be carefully tested for the application at hand. Lastly, the ellipticity estimate might be improved by suitable filtering of the observed image. A linear filter with matrix A can be applied to the image before estimating the ellipticity, since the transformed pixels remain multivariate normal with mean µ ′ = A µ and covariance matrix Σ ′ = A Σ AT . Examples of viable filters are nearest-neighbour or bilinear interpolation, as well as convolution. A combination of these filters could be used to perform PSF deconvolution on the observed image, as an alternative to the algebraic PSF correction (8). MNRAS 000, 1–4 (2017) V̂ar[ χ̂1 ] 1/2 0.762 0.216 0.104 0.717 0.215 0.104 0.704 0.225 0.108 0.658 0.224 0.108 0.794 0.248 0.117 0.720 0.244 0.116 V̂ar[ χ̂2 ] 1/2 1.111 0.214 0.103 0.599 0.214 0.103 0.853 0.226 0.108 0.616 0.225 0.108 1.450 0.322 0.149 0.960 0.316 0.147 ACKNOWLEDGEMENTS NT would like to thank S. Bridle for encouragement and many conversations about shape measurement. The author acknowledges support from the European Research Council in the form of a Consolidator Grant with number 681431. REFERENCES Bartelmann M., Schneider P., 2001, Phys. Rep., 340, 291 Kilbinger M., 2015, Rep. Prog. Phys., 78, 086901 Mandelbaum R., et al., 2014, ApJS, 212, 5 Marsaglia G., 1965, J. Amer. Statist. Assoc., 60, 193 Marsaglia G., 2006, J. Stat. Softw., 16, 1 Melchior P., Viola M., 2012, MNRAS, 424, 2757 Refregier A., Kacprzak T., Amara A., Bridle S., Rowe B., 2012, MNRAS, 425, 1951 Simon P., Schneider P., 2016, preprint, (arXiv:1609.07937) Viola M., Kitching T. D., Joachimi B., 2014, MNRAS, 439, 1909 Voinov V. G., 1985, Sankhya B, 47, 354 de Vaucouleurs G., 1948, Ann. Astrophys., 11, 247 APPENDIX A: A RATIO ESTIMATOR FOR NORMAL VARIATES Let x and y be jointly normal variates with unknown means µx and µy , and known variances σx2 and σy2 and correlation ρ. The goal here is to find an unbiased estimate of the ratio r of their means, r = µx /µy . (A1) Under the additional assumption that the sign of µy is known, an unbiased estimator for r can be found in two short steps. First, the transformation of Marsaglia (1965, 2006) is used to construct a new variate, w = x−cy, c = ρ σx /σy . (A2) The constant c is chosen so that E[wy] = E[w] E[y]. It is clear that w is normal, and that variates w and y are jointly normal, uncorrelated, L4 N. Tessore and thus independent. Note that c = 0 and w = x for independent x and y. Secondly, Voinov (1985) derived an unbiased estimator for the inverse mean of the normal variate y, i.e. a function ĝ(y) with E[ĝ(y)] = 1/µy . For the relevant case of an unknown but positive mean µy > 0, this function is given by  2    √ y y π ĝ(y) = √ exp erfc . (A3) √ 2 σy2 2 σy 2 σy It is then straightforward to construct an estimator for the ratio (A1), (A4) r̂ = c + w ĝ(y) . Since w and y are independent, the expectation is E[r̂] = c + E[w] E[ĝ(y)] = c + (µx − c µy )/µy = µx /µy , (A5) which shows that r̂ is in fact an unbiased estimator for r. The variance of the ratio estimator r̂ is formally given by
Var[r̂] = E[w]2 + Var[w] Var[ĝ(y)] + Var[w] E[ĝ(y)]2 . (A6) As pointed out by Voinov (1985), the variance Var[ĝ(y)] does not exist for function (A3) due to a divergence at infinity. The confidence interval h with probability p,
Pr | ĝ(y) − 1/µy | < h = p , (A7) however, is well-defined, and the variance of ĝ(y) remains finite in applications where infinite values of y are not observed. In this case, an unbiased estimator
2
V̂ar[ĝ(y)] = ĝ(y) − 1 − y ĝ(y) /σy2 (A8) exists and, together with Voinov’s estimator for 1/µ2y , yields
2

V̂ar[r̂] = w2 ĝ(y) − (w2 − σx2 )/σy2 + c2 1 − y ĝ(y) (A9) as an unbiased estimate of the variance of the estimator r̂. When y is significantly larger than its standard deviation, e.g. y/σy > 10, the function ĝ(y) given by (A3) is susceptible to numerical overflow. However, in this regime, it is also very well approximated by its series expansion, ĝ(y) ≈ 1/y − σy2 /y 3 + 3 σy4 /y 5 , y/σy > 10 . (A10) For even larger values y ≫ σy , this approaches 1/y, as expected. This paper has been typeset from a TEX/LATEX file prepared by the author. MNRAS 000, 1–4 (2017) 

Parallel Pricing Algorithms for Multi–Dimensional Bermudan/American Options using Monte Carlo methods Viet_Dung Doan — Abhijeet Gaikwad — Mireille Bossy — Françoise Baude — Ian Stokes-Rees N° 6530 Mai 2008 apport de recherche ISRN INRIA/RR--6530--FR+ENG Thèmes COM et NUM ISSN 0249-6399 arXiv:0805.1827v1 [cs.DC] 13 May 2008 INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE Parallel Pricing Algorithms for Multi–Dimensional Bermudan/American Options using Monte Carlo methods Viet Dung Doan∗ , Abhijeet Gaikwad∗ , Mireille Bossy† , Françoise Baude∗ , Ian Stokes-Rees‡ Thèmes COM et NUM — Systèmes communicants et Systèmes numériques Projets OASIS et TOSCA Rapport de recherche n° 6530 — Mai 2008 — 16 pages Abstract: In this paper we present two parallel Monte Carlo based algorithms for pricing multi–dimensional Bermudan/American options. First approach relies on computation of the optimal exercise boundary while the second relies on classification of continuation and exercise values. We also evaluate the performance of both the algorithms in a desktop grid environment. We show the effectiveness of the proposed approaches in a heterogeneous computing environment, and identify scalability constraints due to the algorithmic structure. Key-words: Multi–dimensional Bermudan/American option, Parallel Distributed Monte Carlo methods, Grid computing. ∗ † ‡ INRIA, OASIS INRIA, TOSCA Dept. Biological Chemistry & Molecular Pharmacology, Harvard Medical School Unité de recherche INRIA Sophia Antipolis 2004, route des Lucioles, BP 93, 06902 Sophia Antipolis Cedex (France) Téléphone : +33 4 92 38 77 77 — Télécopie : +33 4 92 38 77 65 Algorithmes de Pricing parallèles pour des Options Bermudiennes/Américaines multidimensionnelles par une méthode de Monte Carlo Résumé : Dans ce papier, nous présentons deux algorithmes de type Monte Carlo pour le pricing d’options Bermudiennes/Américaines multidimensionnelles. La premiere approche repose sur le calcul de la frontière d’exercice, tandis que la seconde repose sur la classification des valeurs d’exercice et de continuation. Nous évaluons les performances des algorithmes dans un environnement grille. Nous montrons l’efficacité des approches proposées dans un environnement hétérogène. Nous identifions les contraintes d’évolutivité dues à la structure algorithmique. Mots-clés : options Bermudiennes/Américaines multidimensionnelles, Méthodes de Monte Carlo paralléles, Grid computing. Parallel Pricing Algorithms for Multi–Dimensional Bermudan/American Options 1 3 Introduction Options are derivative financial products which allow buying and selling of risks related to future price variations. The option buyer has the right (but not obligation) to purchase (for a call option) or sell (for a put option) any asset in the future (at its exercise date) at a fixed price. Estimates of the option price are based on the well known arbitrage pricing theory: the option price is given by the expected value of the option payoff at its exercise date. For example, the price of a call option is the expected value of the positive part of the difference between the market value of the underlying asset and the asset fixed price at the exercise date. The main challenge in this situation is modelling the future asset price and then estimating the payoff expectation, which is typically done using statistical Monte Carlo (MC) simulations and careful selection of the static and dynamic parameters which describe the market and assets. Some of the widely used options include American option, where the exercise date is variable, and its slight variation Bermudan/American (BA) option, with the fairly discretized variable exercise date. Pricing these options with a large number of underlying assets is computationally intensive and requires several days of serial computational time (i.e. on a single processor system). For instance, Ibanez and Zapatero (2004) [10] state that pricing the option with five assets takes two days, which is not desirable in modern time critical financial markets. Typical approaches for pricing options include the binomial method [5] and MC simulations [6]. Since binomial methods are not suitable for high dimensional options, MC simulations have become the cornerstone for simulation of financial models in the industry. Such simulations have several advantages, including ease of implementation and applicability to multi–dimensional options. Although MC simulations are popular due to their “embarrassingly parallel” nature, for simple simulations, allows an almost arbitrary degree of near-perfect parallel speed-up, their applicability to pricing American options is complex[10], [4], [12]. Researchers have proposed several approximation methods to improve the tractability of MC simulations. Recent advances in parallel computing hardware such as multi-core processors, clusters, compute “clouds”, and large scale computational grids have also attracted the interest of the computational finance community. In literature, there exist a few parallel BA option pricing techniques. Examples include Huang (2005) [9] or Thulasiram (2002) [15] which are based on the binomial lattice model. However, a very few studies have focused on parallelizing MC methods for BA pricing [16]. In this paper, we aim to parallelize two American option pricing methods: the first approach proposed in Ibanez and Zapatero (2004) [10] (I&Z) which computes the optimal exercise boundary and the second proposed by Picazo (2002) [8] (CMC) which uses the classification of continuation values. These two methods in their sequential form are similar to recursive programming so that at a given exercise opportunity they trigger many small independent MC simulations to compute the continuation values. The optimal strategy of an American option is to compare the exercise value (intrinsic value) with the continuation value (the expected cash flow from continuing the option contract), then exercise if the exercise value is more valuable. In the case of I&Z Algorithm the continuation values are used to parameterize the exercise boundary whereas CMC Algorithm classifies them to provide a characterization of RR n° 6530 4 Doan, Gaikwad, Bossy, Baude & Stokes-Rees the optimal exercise boundary. Later, both approaches compute the option price using MC simulations based on the computed exercise boundaries. Our roadmap is to study both the algorithms to highlight their potential for parallelization: for the different phases, our aim is to identify where and how the computation could be split into independent parallel tasks. We assume a master-worker grid programming model, where the master node splits the computation in such tasks and assigns them to a set of worker nodes. Later, the master also collects the partial results produced by these workers. In particular, we investigate parallel BA options pricing to significantly reduce the pricing time by harnessing the computational power provided by the computational grid. The paper is organized as follows. Sections 2 and 3 focus on two pricing methods and are structured in a similar way: a brief introduction to present the method, sequential followed by parallel algorithm and performance evaluation concludes each section. In section 4 we present our conclusions. 2 2.1 Computing optimal exercise boundary algorithm Introduction In [10], the authors propose an option pricing method that builds a full exercise boundary as a polynomial curve whose dimension depends on the number of underlying assets. This algorithm consists of two phases. In the first phase the exercise boundary is parameterized. For parameterization, the algorithm uses linear interpolation or regression of a quadratic or cubic function at a given exercise opportunity. In the second phase, the option price is computed using MC simulations. These simulations are run until the price trajectory reaches the dynamic boundary computed in the first phase. The main advantage of this method is that it provides a full parameterization of the exercise boundary and the exercise rule. For American options, a buyer is mainly concerned in these values as he can decide at ease whether or not to exercise the option. At each exercise date t, the optimal exercise point St∗ is defined by the following implicit condition, Pt (St∗ ) = I(St∗ ) (1) where Pt (x) is the price of the American option on the period [t, T ], I(x) is the exercise value (intrinsic value) of the option and x is the asset value at opportunity date t. As explained in [10], these optimal values stem from the monotonicity and convexity of the price function P (·) in (1). These are general properties satisfied by most of the derivative securities such as maximum, minimum or geometric average basket options. However, for the problems where the monotonicity and convexity of the price function can not be easily established, this algorithm has to be revisited. In the following section we briefly discuss the sequential algorithm followed by a proposed parallel solution. INRIA Parallel Pricing Algorithms for Multi–Dimensional Bermudan/American Options 2.2 5 Sequential Boundary Computation The algorithm proposed in [10] is used to compute the exercise boundary. To illustrate this approach, we consider a call BA option on the maximum of d assets modeled by Geometric Brownian Motion (GBM). It is a standard example for the multi–dimensional BA option with maturity date T , constant interest rate r and the price of this option at t0 is given as Pt0 = E (exp (−rτ )Φ(Sτ , τ )|St0 ) where τ is the optimal stopping time ∈ {t1 , .., T }, defined as the first time ti such that the underlying value Sτ surpasses the optimal exercise boundary at the opportunity τ otherwise the option is held until τ = ∞. The payoff at time τ is defined as follows: Φ(Sτ , τ ) = (maxi (Sτi ) − K)+ , where i = 1,..,d, S is the underlying asset price vector and K is the strike price. The parameter d has a strong impact on the complexity of the algorithm, except in some cases as the average options where the number of dimensions d can be easily reduced to one. For an option on the maximum of d assets there are d separate exercise regions which are characterized by d boundaries [10]. These boundaries are monotonic and smooth curves in Rd−1 . The algorithm uses backward recursive time programming, with a finite number of exercise opportunities m = 1, ..., NT . Each boundary is computed by regression on J numbers of boundary points in Rd−1 . At each given opportunity, these optimal boundary points are computed using N1 MC simulations. Further in case of an option on the maximum of d underlying assets, for each asset the boundary points are computed. P It takes n iterations to converge eachindividual point. The complexity of this step is NT O m=1 d × J × m × N1 × (NT − m) × n . After estimating J optimal boundary points for each asset, d regressions are performed over these points to get d execution boundaries. Let us assume that the complexity of this step is O(NT × regression(J)), where the complexity of the regression is assumed to be constant. After computing the boundaries at all m exercise opportunities, in the second phase, the price of the option is computed using a standard MC simulation of N paths in Rd . The complexity of the pricing phase is O(d × NT × N ). Thus the total complexity of the algorithm is as follows, P  NT O m=1 d × J × m × N1 × (NT − m) × n + NT × regression(J) + d × NT × N
≈ O NT2 × J × d × N1 × n + NT × (J + d × N ) For the performance benchmarks, we use the same simulation parameters as given in [10]. Consider a call BA option on the maximum of d assets with the following configuration. K = 100, interest rate r = 0.05, volatility rate σ = 0.2, dividend δ = 0.1, J = 128, N1 = 5e3, N = 1e6, d = 3, NT = 9 and T = 3 years. (2) The sequential pricing of this option (2) takes 40 minutes. The distribution of the total time for the different phases is shown in Figure 1. As can be seen, the data generation phase, which simulates and calculates J optimal boundary points, consumes most of the computational RR n° 6530 6 Doan, Gaikwad, Bossy, Baude & Stokes-Rees time. We believe that a parallel approach to this and other phases could dramatically reduce the computational time. This inspires us to investigate a parallel approach for I&Z Algorithm which is presented in the following section. The numerical results that we shall provide indicate that the proposed parallel solution is more efficient compared with the serial algorithm. 2.3 Parallel approach In this section, a simple parallel approach for I&Z Algorithm is presented and the pseudocode for the same is given in Algorithm 13. This approach is inspired from the solution proposed by Muni Toke [16], though he presents it in the context of a low–order homogeneous parallel cluster. The algorithm consists of two phases. The first parallel phase is based on the following observation: for each of the d boundaries, the computation of J optimal boundary points at a given exercise date can be simulated independently. The optimal exerAlgorithm 1 Parallel Ibanez and Zapatero Algorithm 1: [glp] Generation of the J “Good Lattice Points” 2: for t = tNT to t1 do 3: for di = 1 to d do 4: for j = 1 to J in parallel do 5: [calc] Computation of a boundary point with N1 Monte Carlo simulations 6: end for 7: [reg] Regression of the exercise boundary . 8: end for 9: end for 10: for i = 1 to N in parallel do 11: [mc] Computation of the partial option price. 12: end for 13: Estimation of the final option price by merging the partial prices. cise boundaries from opportunity date m back to m − 1 are computed as follows. Note that at m = NT , the boundary is known (e.g. for a call option the boundary at NT is defined as the strike value). Backward to m = NT − 1, we have to estimate J optimal points from J initial good lattice points [10], [7] to regress the boundary to this time. The regression of Rd → Rd is difficult to achieve in a reasonable amount of time in case of large number of training points. To decrease the size of the training set we utilize “Good Lattice Points” (GLPs) as described in [7],[14], and [3]. In particular case of a call on the maximum of d assets, only a regression of Rd−1 → R is needed, but we repeat it d times. The Algorithm 13 computes GLPs using either SSJ library [11] or the quantification number sequences as presented in [13]. SSJ is a Java library for stochastic simulation and it computes GLPs as a Quasi Monte Carlo sequence such as Sobol or Hamilton sequences. The algorithm can also use the number sequences readily available at [1]. These sequences are INRIA Parallel Pricing Algorithms for Multi–Dimensional Bermudan/American Options 7 generated using an optimal quadratic quantizer of the Gaussian distribution in more than one dimension. The [calc] phase of the Algorithm 13 is embarrassingly parallel and the J boundary points are equally distributed among the computing nodes. At each node, the algorithm simulates N1 paths to compute the approximate points. Then Newton’s iterations method is used to converge an individual approximated point to the optimal boundary point. After computing J optimal boundary points, these points are collected by the master node, for sequential regression of the exercise boundary. This node then repeats the same procedure for every date t, in a recursive way, until t = t1 in the [reg] phase. The [calc] phase provides the exact optimal exercise boundary at every opportunity date. After computation of the boundary, in the last [mc] phase, the option is priced using parallel MC simulations as shown in the Algorithm 13. 2.4 Numerical results and performance In this section we present performance and accuracy results due to the parallel I&Z Algorithm described in Algorithm 13. We price a basket BA call option on the maximum of 3 assets as given in (2). The start prices for the assets are varied as 90, 100, and 110. The prices estimated by the algorithm are presented in Table 1. To validate our results we compare the estimated prices with the prices mentioned in Andersen and Broadies (1997) [2]. Their results are reproduced in the column labeled as “Binomial”. The last column of the table indicates the errors in the estimated prices. As we can see, the estimated option prices are close to the desired prices by acceptable marginal error and this error is represented by a desirable 95% confidence interval. As mentioned earlier, the algorithm relies on J number of GLPs to effectively compute the optimal boundary points. Later the parameterized boundary is regressed over these points. For the BA option on the maximum described in (2), Muni Toke [16] notes that J smaller than 128 is not sufficient and prejudices the option price. To observe the effect of the number of optimal boundary points on the accuracy of the estimated price, the number of GLPs is varied as shown in Table 2. For this experiment, we set the start price of the option as S0 = 90. The table indicates that increasing number of GLPs has negligible impact on the accuracy of the estimated price. However, we observe the linear increase in the computational time with the increase in the number of GLPs. [INSERT TABLE 1 HERE] To evaluate the accuracy of the computed prices by the parallel algorithm, we obtained the numerical results with 16 processors. First, let us observe the effect of N1 , the number of simulations required in the first phase of the algorithm, on the computed option price. In [16], the author comments that the large values of N1 do not affect the accuracy of option price. For these experiments, we set the number of GLPs, J, as 128 and vary N1 as shown in Table 3. We can clearly observe that N1 in fact has a strong impact on the accuracy of the computed option prices: the computational error decreases with the increased N1 . A large value of N1 results in more accurate boundary points, hence more accurate exercise boundary. Further, if the exercise boundary is accurately computed, the resulting option prices are much closer to the true price. However this, as we can see in the third column, RR n° 6530 8 Doan, Gaikwad, Bossy, Baude & Stokes-Rees S0i 90 100 110 Option Price 11.254 18.378 27.512 Variance (95% CI) 153.857 (0.024) 192.540 (0.031) 226.332 (0.035) Binomial 11.29 18.69 27.58 Error 0.036 0.312 0.068 Table 1: Price of the call BA on the maximum of three assets (d = 3, with the spot price S0i for i = 1, .., 3) using I&Z Algorithm. (r = 0.05, δ = 0.1, σ = 0.2, ρ = 0.0, T = 3 and N = 9). The binomial values are referred as the true values. J 128 256 1024 Price 11.254 11.258 11.263 Time (in minute) 4.6 8.1 29.5 Error 0.036 0.032 0.027 Table 2: Impact of the value of J on the results of the maximum on three assets option (S0 = 90). The binomial price is 11.29. Running time on 16 processors. N1 5000 10000 100000 Price 11.254 11.262 11.276 Time (in minute) 4.6 6.9 35.7 Error 0.036 0.028 0.014 Table 3: Impact of the value of N1 on the results of the maximum on three assets option (S0 = 90). The binomial price is 11.29. Running time on 16 processors. comes at a cost of increased computational time. The I&Z algorithm highly relies on the accuracy and the convergence rate of the optimal boundary points. While the former affects the accuracy of the option price, the later affects the speed up of the algorithm. In each iteration, to converge to the optimal boundary point, the algorithm starts with an arbitrary point with the strike price K often as its initial value. The algorithm then uses N1 random MC paths to simulate the approximated point. A convergence criterion is used to optimize this approximated point. The simulated point is assumed to be optimal when it satisfies the i,(initial) i,(initial) i,(simulated) is the initial | < ǫ = 0.01, where the Stn − Stn following condition, |Stn i,(simulated) is the point simulated by point at a given opportunity tn , i = 1..J and the Stn using N1 MC simulations. In case, the condition is not satisfied, this procedure is repeated i,(simulated) . Note that the and now with the initial point as the newly simulated point Stn number of iterations n, required to reach to the optimal value, varies depending on the fixed precision in the Newton procedure (for instance, with a precision ǫ = 0.01, n varies from 30 to 60). We observed that not all boundary points take the same time for the convergence. INRIA Parallel Pricing Algorithms for Multi–Dimensional Bermudan/American Options 9 Some points converge faster to the optimal boundary points while some take longer than usual. Since the algorithm has to wait until all the points are optimized, the slower points increase the computational time, thus reducing the efficiency of the parallel algorithm, see Figure 2. Monte Carlo simulation 22.09% Regression 0.01% Data generation 77.90% Figure 1: The time distribution for the sequential optimal exercise boundary computation algorithm. The total time is about 40 minutes. 3 3.1 The Classification and Monte Carlo algorithm Introduction The Monte Carlo approaches for BA option pricing are usually based on continuation value computation [12] or continuation region estimation [8], [10]. The option holder decides either to execute or to continue with the current option contract based on the computed asset value. If the asset value is in the exercise region, he executes the option otherwise he continues to hold the option. Denote that the asset values which belong to the exercise region will form the exercise values and rest will belong to the continuation region. In [8] Picazo et al. propose an algorithm based on the observation that at a given exercise opportunity the option holder makes his decision based on whether the sign of (exercise value−continuation value) is positive or negative. The author focuses on estimating the continuation region and the exercise region by characterizing the exercise boundary based on these signs. The classification algorithm is used to evaluate such sign values at each opportunity. In this section we briefly describe the sequential algorithm described in [8] and propose a parallel approach followed by performance benchmarks. RR n° 6530 10 Doan, Gaikwad, Bossy, Baude & Stokes-Rees Figure 2: Speedup of the parallel I&Z Algorithm. 3.2 Sequential algorithm For illustration let us consider a BA option on d underlying assets modeled by Geometric Brownian Motion (GBM). St = (Sti ) with i = 1, .., d. The option price at time t0 is defined as follows: Pt0 (St0 ) = E (exp (−rτ )Φ(Sτ , τ )|St0 ) where τ is the optimal stopping time ∈ {t1 , .., T }, T is the maturity date, r is the constant interest rate and Φ(Sτ , τ ) is the payoff value at time τ . In case of I&Z Algorithm, the optimal stopping time is defined when the underlying asset value crosses the exercise boundary. The CMC algorithm defines the stopping time whenever the underlying asset value makes the sign of (exercise value − continuation value) positive. Without loss of generality, at a given time t the BA option price on the period [t, T ] is given by: Pt (St ) = E (exp (−r(τ − t))Φ(Sτ , τ )|St ) where τ is the optimal stopping time ∈ {1, .., T }. Let us define the difference between the payoff value and the option price at time tm as, β(tm , Stm ) = Φ(Stm , tm ) − Ptm (Stm ) where m ∈ {1, .., T }. The option is exercised when Stm ∈ {x|β(tm , x) > 0} which is the exercise region, and x is the simulated underlying asset value, otherwise the option is continued. The goal of the algorithm is to determinate the function β(·) for every opportunity INRIA Parallel Pricing Algorithms for Multi–Dimensional Bermudan/American Options 11 date. However, we do not need to fully parameterize this function. It is enough to find a function Ft (·) such that signFt (·) = signβ(t, ·). The algorithm consists of two phases. In the first phase, it aims to find a function Ft (·) having the same sign as the function β(t, ·). The AdaBoost or LogitBoost algorithm is used to characterize these functions. In the second phase the option is priced by a standard MC simulation by taking the advantage of the characterization of Ftm (·), so for the (i)th MC (i) simulation we get the optimal stopping time τ(i) = min{tm ∈ {t1 , t2 , ..., T }|Ft (St ) > 0}. The (i) is not the index of the number of assets. Now, consider a call BA option on the maximum of d underlying assets where the payoff at time τ is defined as Φ(Sτ , τ ) = (maxi (Sτi ) − K)+ with i = 1, .., d. During the first phase of the algorithm, at a given opportunity date tm with m ∈ 1, ..., NT , N1 underlying price vectors sized d are simulated. The simulations are performed recursively in backward from m = T to m = 1. From each price point, another N2 paths are simulated from a given opportunity date to the maturity date to compute the “small” BA option price at this opportunity (i.e. Ptm (Stm )). At this step, N1 option prices related to the opportunity date are computed. The time step complexity of this step is O(N1 × d × m × N2 × (NT − m)). In the classification phase, we use a training set of N1 underlying price points and their corresponding option prices at a given opportunity date. In this step, a non–parametric regression is done on N1 points to characterize the exercise boundary. This first phase is repeated for each opportunity date. In the second phase, the option value is computed by simulating a large number, N , of standard MC simulations with NT exercise opportunities. The complexity of this phase is O(d × NT × N ). Thus, the total time steps required for the algorithm can be given by the following formula, P  NT O N × d × m × N × (N − m) + N × classif ication(N ) + d × N × N 2 T T 1 T m=1 1
≈ O NT2 × N1 × d × N2 + NT × (N1 + d × N ) where O(classif ication(·)) is the complexity of the classification phase and the details of which can be found in [8]. For the simulations, we use the same option parameters as described in (2), taken from [10], and the parameters for the classification can be found in [8]. K = 100, interest rate r = 0.05, volatility rate σ = 0.2, dividend δ = 0.1, N1 = 5e3, N2 = 500, N = 1e6, d = 3 NT = 9 and T = 3 years. (3) Each of the N1 points of the training set acts as a seed which is further used to simulate N2 simulation paths. From the exercise opportunity m backward to m − 1, a Brownian motion bridge is used to simulate the price of the underlying asset. The time distribution of each phase of the sequential algorithm for pricing the option (3) is shown in Figure 3. As we can see from the figure, the most computationally intensive part is the data generation phase which is used to compute the option prices required for classification. In the following section we present a parallel approach for this and rest of the phases of the algorithm. RR n° 6530 12 3.3 Doan, Gaikwad, Bossy, Baude & Stokes-Rees Parallel approach The Algorithm 10 illustrates the parallel approach based on CMC Algorithm. At tm = T (i) we generate N1 points of the price of the underlying assets, Stm , i = 1, .., N1 then apply the Brownian bridge simulation process to get the price at the backward date, tm−1 . For simplicity we assume a master–worker programming model for the parallel implementation: the master is responsible for allocating independent tasks to workers and for collecting the results. The master divides N1 simulations into nb tasks then distributes them to a number of workers. Thus each worker has N1 /nb points to simulate in the [calc] phase. (i) Each worker, further, simulates N2 paths for each point from tm to tNT starting at Stm to compute the option price related to the opportunity date. Next the worker calculates the value yj = (exercise value − continuation value), j = 1, .., N1 /nb. The master collects the yj of these nb tasks from the workers and then classifies them in order to return the characterization model of the associated exercise boundary in the [class] phase. For the Algorithm 2 Parallel Classification and Monte Carlo Algorithm 1: for t = tNT to t1 do 2: for i = 1 to N1 in parallel do 3: [calc] Computation of training points. 4: end for 5: [class] Classification using boosting. 6: end for 7: for i = 1 to N in parallel do 8: [mc] The partial option price computation. 9: end for 10: Estimation of the final option price by merging the partial prices. 1 classification phase, the master does a non-parametric regression with the set (x(i) , y(i) )N i=1 , (i) where x(i) = Stm , to get the function Ftm (x) described above in Section (3.2). The algorithm recursively repeats the same procedure for earlier time intervals [m − 1, 1]. As a result we obtain the characterization of the boundaries, Ftm (x), at every opportunity tm . Using these boundaries, a standard MC simulation, [mc], is used to estimate the option price. The MC simulations are distributed among workers such that each worker has the entire characterization boundary information (Ftm (x), m = 1, .., NT ) to compute the partial option price. The master later merges the partially computed prices and estimates the final option price. 3.4 Numerical results and performance In this section we present the numerical and performance results of the parallel CMC Algorithm. We focus on the standard example of a call option on the maximum of 3 assets as given in (3). As it can be seen, the estimated prices are equivalent to the reference prices INRIA Parallel Pricing Algorithms for Multi–Dimensional Bermudan/American Options S0 90 100 110 Price 11.295 18.706 27.604 Variance (95% CI) 190.786 (0.027) 286.679 (0.033) 378.713 (0.038) Binomial 11.290 18.690 27.580 13 Error 0.005 0.016 0.024 Table 4: Price of the call BA on the maximum of three assets using CMC Algorithm. (r = 0.05, δ = 0.1, σ = 0.2, ρ = 0.0, T = 3, N = 9 opportunities) presented in Andersen and Broadies [2], which are represented in the “Binomial” column in Table 4. For pricing this option, the sequential execution takes up to 30 minutes and the time distribution for the different phases can be seen in Figure 3. The speed up achieved by the parallel algorithm is presented in Figure 4. We can observe from the figure that the parallel algorithm achieves linear scalability with a fewer number of processors. The different phases of the algorithm scale differently. The MC phase being embarrassingly parallel scales linearly, while, the number of processors has no impact on the scalability of the classification phase. The classification phase is sequential and takes a constant amount of time for the same option. This affects the overall speedup of the algorithm as shown in Figure 4. Monte Carlo simulation 38.40% Data generation 60.49% Boosting classification 1.11% Figure 3: The time distribution for different phases of the sequential Classification–Monte Carlo algorithm. The total time is about 30 minutes. 4 Conclusion The aim of the study is to develop and implement parallel Monte Carlo based Bermudan/American option pricing algorithms. In this paper, we particularly focused on multi– RR n° 6530 14 Doan, Gaikwad, Bossy, Baude & Stokes-Rees Figure 4: Speedup of the parallel CMC Algorithm. dimensional options. We evaluated the scalability of the proposed parallel algorithms in a computational grid environment. We also analyzed the performance and the accuracy of both algorithms. While I&Z Algorithm computes the exact exercise boundary, CMC Algorithm estimates the characterization of the boundary. The results obtained clearly indicate that the scalability of I&Z Algorithm is limited by the boundary points computation. The Table 2 showed that there is no effective advantage to increase the number of such points as will, just to take advantage of a greater number of available CPUs. Moreover, the time required for computing individual boundary points varies and the points with slower convergence rate often haul the performance of the algorithm. However, in the case of CMC Algorithm, the sequential classification phase tends to dominate the total parallel computational time. Nevertheless, CMC Algorithm can be used for pricing different option types such as maximum, minimum or geometric average basket options using a generic classification configuration. While the optimal exercise boundary structure in I&Z Algorithm needs to be tailored as per the option type and requires. Parallelizing classification phase presents us several challenges due to its dependency on inherently sequential non–parametric regression. Hence, we direct our future research to investigate efficient parallel algorithms for computing exercise boundary points, in case of I&Z Algorithm, and the classification phase, in case of CMC Algorithm. INRIA Parallel Pricing Algorithms for Multi–Dimensional Bermudan/American Options 5 15 Acknowledgments This research is supported by the French “ANR-CIGC GCPMF” project and Grid5000 has been funded by ACI-GRID. References [1] http://perso-math.univ-mlv.fr/users/printems.jacques/. [2] L. Andersen and M. Broadie. Primal-Dual Simulation Algorithm for Pricing Multidimensional American Options. Management Science, 50(9):1222–1234, 2004. [3] P.P. Boyle, A. Kolkiewicz, and K.S. Tan. Pricing American style options using low discrepancy mesh methods. Forthcoming, Mathematics and Computers in Simulation. [4] M. Broadie and J. Detemple. The Valuation of American Options on Multiple Assets. Mathematical Finance, 7(3):241–286, 1997. [5] J.C. Cox, S.A. Ross, and M. Rubinstein. Option Pricing: A Simplified Approach. Journal of Financial Economics, 7(3):229–263, 1979. [6] P. Glasserman. Monte Carlo Methods in Financial Engineering. Springer, 2004. [7] S. Haber. Parameters for Integrating Periodic Functions of Several Variables. Mathematics of Computation, 41(163):115–129, 1983. [8] F.J. Hickernell, H. Niederreiter, and K. Fang. Monte Carlo and Quasi-Monte Carlo Methods 2000: Proceedings of a Conference Held at Hong Kong Baptist University, Hong Kong SAR, China. Springer, 2002. [9] K. Huang and R.K. Thulasiram. Parallel Algorithm for Pricing American Asian Options with Multi-Dimensional Assets. Proceedings of the 19th International Symposium on High Performance Computing Systems and Applications, pages 177–185, 2005. [10] A. Ibanez and F. Zapatero. Monte Carlo Valuation of American Options through Computation of the Optimal Exercise Frontier. Journal of Financial and Quantitative Analysis, 39(2):239–273, 2004. [11] P. L’Ecuyer, L. Meliani, and J. Vaucher. SSJ: SSJ: a framework for stochastic simulation in Java. Proceedings of the 34th conference on Winter simulation: exploring new frontiers, pages 234–242, 2002. [12] F.A. Longstaff and E.S. Schwartz. Valuing American options by simulation: a simple least-squares approach. Review of Financial Studies, 2001. [13] G. Pagès and J. Printems. Optimal quadratic quantization for numerics: the Gaussian case. Monte Carlo Methods and Applications, 9(2):135–165, 2003. RR n° 6530 16 Doan, Gaikwad, Bossy, Baude & Stokes-Rees [14] I.H. Sloan and S. Joe. Lattice Methods for Multiple Integration. Oxford University Press, 1994. [15] R.K. Thulasiram and D.A. Bondarenko. Performance Evaluation of Parallel Algorithms for Pricing Multidimensional Financial Derivatives. The International Conference on Parallel Processing Workshops (ICPPW 02), 1530, 2002. [16] I.M. Toke. Monte Carlo Valuation of Multidimensional American Options Through Grid Computing. Lecture notes in computer science, Springer-Verlag, Volume 3743:page 462, 2006. INRIA Unité de recherche INRIA Sophia Antipolis 2004, route des Lucioles - BP 93 - 06902 Sophia Antipolis Cedex (France) Unité de recherche INRIA Futurs : Parc Club Orsay Université - ZAC des Vignes 4, rue Jacques Monod - 91893 ORSAY Cedex (France) Unité de recherche INRIA Lorraine : LORIA, Technopôle de Nancy-Brabois - Campus scientifique 615, rue du Jardin Botanique - BP 101 - 54602 Villers-lès-Nancy Cedex (France) Unité de recherche INRIA Rennes : IRISA, Campus universitaire de Beaulieu - 35042 Rennes Cedex (France) Unité de recherche INRIA Rhône-Alpes : 655, avenue de l’Europe - 38334 Montbonnot Saint-Ismier (France) Unité de recherche INRIA Rocquencourt : Domaine de Voluceau - Rocquencourt - BP 105 - 78153 Le Chesnay Cedex (France) Éditeur INRIA - Domaine de Voluceau - Rocquencourt, BP 105 - 78153 Le Chesnay Cedex (France) http://www.inria.fr ISSN 0249-6399 

Reversible Computation in Term Rewriting✩ Naoki Nishidaa , Adrián Palaciosb , Germán Vidalb,∗ a arXiv:1710.02804v1 [cs.PL] 8 Oct 2017 Graduate School of Informatics, Nagoya University Furo-cho, Chikusa-ku, 4648603 Nagoya, Japan b MiST, DSIC, Universitat Politècnica de València Camino de Vera, s/n, 46022 Valencia, Spain Abstract Essentially, in a reversible programming language, for each forward computation from state S to state S ′ , there exists a constructive method to go backwards from state S ′ to state S. Besides its theoretical interest, reversible computation is a fundamental concept which is relevant in many different areas like cellular automata, bidirectional program transformation, or quantum computing, to name a few. In this work, we focus on term rewriting, a computation model that underlies most rule-based programming languages. In general, term rewriting is not reversible, even for injective functions; namely, given a rewrite step t1 → t2 , we do not always have a decidable method to get t1 from t2 . Here, we introduce a conservative extension of term rewriting that becomes reversible. Furthermore, we also define two transformations, injectivization and inversion, to make a rewrite system reversible using standard term rewriting. We illustrate the usefulness of our transformations in the context of bidirectional program transformation. This work has been partially supported by the EU (FEDER) and the Spanish Ministerio de Economı́a y Competitividad (MINECO) under grants TIN2013-44742-C4-1-R and TIN2016-76843-C4-1-R, by the Generalitat Valenciana under grant PROMETEOII/2015/013 (SmartLogic), and by the COST Action IC1405 on Reversible Computation - extending horizons of computing. Adrián Palacios was partially supported by the EU (FEDER) and the Spanish Ayudas para contratos predoctorales para la formación de doctores and Ayudas a la movilidad predoctoral para la realización de estancias breves en centros de I+D, MINECO (SEIDI), under FPI grants BES-2014-069749 and EEBB-I-1611469. Part of this research was done while the second and third authors were visiting Nagoya University; they gratefully acknowledge their hospitality. c 2017. This manuscript version is made available under the CC-BY-NC-ND 4.0 license http://creativecommons.org/licenses/by-nc-nd/4.0/ ∗ Corresponding author. Email addresses: nishida@i.nagoya-u.ac.jp (Naoki Nishida), apalacios@dsic.upv.es (Adrián Palacios), gvidal@dsic.upv.es (Germán Vidal) ✩ Preprint submitted to Elsevier February 20, 2018 To appear in the Journal of Logical and Algebraic Methods in Programming. Keywords: term rewriting, reversible computation, program transformation 1. Introduction The notion of reversible computation can be traced back to Landauer’s pioneering work [22]. Although Landauer was mainly concerned with the energy consumption of erasing data in irreversible computing, he also claimed that every computer can be made reversible by saving the history of the computation. However, as Landauer himself pointed out, this would only postpone the problem of erasing the tape of a reversible Turing machine before it could be reused. Bennett [6] improved the original proposal so that the computation now ends with a tape that only contains the output of a computation and the initial source, thus deleting all remaining “garbage” data, though it performs twice the usual computation steps. More recently, Bennett’s result is extended in [9] to nondeterministic Turing machines, where it is also proved that transforming an irreversible Turing machine into a reversible one can be done with a quadratic loss of space. We refer the interested reader to, e.g., [7, 14, 40] for a high level account of the principles of reversible computation. In the last decades, reversible computing and reversibilization (transforming an irreversible computation device into a reversible one) have been the subject of intense research, giving rise to successful applications in many different fields, e.g., cellular automata [28], where reversibility is an essential property, bidirectional program transformation [24], where reversibility helps to automate the generation of inverse functions (see Section 6), reversible debugging [17], where one can go both forward and backward when seeking the cause of an error, parallel discrete event simulation [34], where reversible computation is used to undo the effects of speculative computations made on a wrong assumption, quantum computing [39], where all computations should be reversible, and so forth. The interested reader can find detailed surveys in the state of the art reports of the different working groups of COST Action IC1405 on Reversible Computation [20]. In this work, we introduce reversibility in the context of term rewriting [4, 36], a computation model that underlies most rule-based programming languages. In contrast to other, more ad-hoc approaches, we consider that term rewriting is an excellent framework to rigorously define reversible computation in a functional context and formally prove its main properties. We 2 expect our work to be useful in different (sequential) contexts, like reversible debugging, parallel discrete event simulation or bidirectional program transformation, to name a few. In particular, Section 6 presents a first approach to formalize bidirectional program transformation in our setting. To be more precise, we present a general and intuitive notion of reversible term rewriting by defining a Landauer embedding. Given a rewrite system R and its associated (standard) rewrite relation →R , we define a reversible extension of rewriting with two components: a forward relation ⇀R and a backward relation ↽R , such that ⇀R is a conservative extension of →R and, moreover, (⇀R )−1 = ↽R. We note that the inverse rewrite relation, (→R )−1 , is not an appropriate basis for “reversible” rewriting since we aim at defining a technique to undo a particular reduction. In other words, given a rewriting reduction s →∗R t, our reversible relation aims at computing the term s from t and R in a decidable and deterministic way, which is not possible using (→R )−1 since it is generally non-deterministic, non-confluent, and nonterminating, even for systems defining injective functions (see Example 6). In contrast, our backward relation ↽R is deterministic (thus confluent) and terminating. Moreover, our relation proceeds backwards step by step, i.e., the number of reduction steps in s ⇀∗R t and t ↽∗R s are the same. In order to introduce a reversibilization transformation for rewrite systems, we use a flattening transformation so that the reduction at top positions of terms suffices to get a normal form in the transformed systems. For instance, given the following rewrite system: add(0, y) → y, add(s(x), y) → s(add(x, y)) defining the addition on natural numbers built from constructors 0 and s( ), we produce the following flattened (conditional) system: R={ add(0, y) → y, add(s(x), y) → s(z) ⇐ add(x, y) ։ z } (see Example 29 for more details). This allows us to provide an improved notion of reversible rewriting in which some information (namely, the positions where reduction takes place) is not required anymore. This opens the door to compile the reversible extension of rewriting into the system rules. Loosely speaking, given a system R, we produce new systems Rf and Rb such that standard rewriting in Rf , i.e., →Rf , coincides with the forward reversible extension ⇀R in the original system, and analogously →Rb is equivalent to 3 ↽R . E.g., for the system R above, we would produce Rf = { Rb = { addi (0, y) → hy, β1i, addi (s(x), y) → hs(z), β2 (w)i ⇐ addi (x, y) ։ hz, wi } add−1 (y, β1) → h0, yi, add−1 (s(z), β2 (w)) → hs(x), yi ⇐ add−1 (z, w) → hx, yi } where addi is an injective version of function add, add−1 is the inverse of addi , and β1 , β2 are fresh symbols introduced to label the rules of R. In this work, we will mostly consider conditional rewrite systems, not only to have a more general notion of reversible rewriting, but also to define a reversibilization technique for unconditional rewrite systems, since the application of flattening (cf. Section 4) may introduce conditions in a system that is originally unconditional, as illustrated above. This paper is an extended version of [31]. In contrast to [31], our current paper includes the proofs of technical results, the reversible extension of term rewriting is introduced first in the unconditional case (which is simpler and more intuitive), and presents an improved injectivization transformation when the system includes injective functions. Furthermore, a prototype implementation of the reversibilization technique is publicly available from http://kaz.dsic.upv.es/rev-rewriting.html. The paper is organized as follows. After introducing some preliminaries in Section 2, we present our approach to reversible term rewriting in Section 3. Section 4 introduces the class of pure constructor systems where all reductions take place at topmost positions, so that storing this information in reversible rewrite steps becomes unnecessary. Then, Section 5 presents injectivization and inversion transformations in order to make a rewrite system reversible with standard rewriting. Here, we also present an improvement of the transformation for injective functions. The usefulness of these transformations is illustrated in Section 6. Finally, Section 7 discusses some related work and Section 8 concludes and points out some ideas for future research. 2. Preliminaries We assume familiarity with basic concepts of term rewriting. We refer the reader to, e.g., [4] and [36] for further details. 2.1. Terms and Substitutions A signature F is a set of ranked function symbols. Given a set of variables V with F ∩V = ∅, we denote the domain of terms by T (F , V). We use f, g, . . . to denote functions and x, y, . . . to denote variables. Positions are used to 4 address the nodes of a term viewed as a tree. A position p in a term t, in symbols p ∈ Pos(t), is represented by a finite sequence of natural numbers, where ǫ denotes the root position. We let t|p denote the subterm of t at position p and t[s]p the result of replacing the subterm t|p by the term s. Var(t) denotes the set of variables appearing in t. We also let Var(t1 , . . . , tn ) denote Var(t1 ) ∪ · · · ∪ Var(tn ). A term t is ground if Var(t) = ∅. A substitution σ : V 7→ T (F , V) is a mapping from variables to terms such that Dom(σ) = {x ∈ V | x 6= σ(x)} is its domain. A substitution σ is ground if xσ is ground for all x ∈ Dom(σ). Substitutions are extended to morphisms from T (F , V) to T (F , V) in the natural way. We denote the application of a substitution σ to a term t by tσ rather than σ(t). The identity substitution is denoted by id. We let “◦” denote the composition of substitutions, i.e., σ ◦ θ(x) = (xθ)σ = xθσ. The restriction θ |`V of a substitution θ to a set of variables V is defined as follows: xθ |`V = xθ if x ∈ V and xθ |`V = x otherwise. 2.2. Term Rewriting Systems A set of rewrite rules l → r such that l is a nonvariable term and r is a term whose variables appear in l is called a term rewriting system (TRS for short); terms l and r are called the left-hand side and the right-hand side of the rule, respectively. We restrict ourselves to finite signatures and TRSs. Given a TRS R over a signature F , the defined symbols DR are the root symbols of the left-hand sides of the rules and the constructors are CR = F \ DR . Constructor terms of R are terms over CR and V, denoted by T (CR , V). We sometimes omit R from DR and CR if it is clear from the context. A substitution σ is a constructor substitution (of R) if xσ ∈ T (CR , V) for all variables x. For a TRS R, we define the associated rewrite relation →R as the smallest binary relation on terms satisfying the following: given terms s, t ∈ T (F , V), we have s →R t iff there exist a position p in s, a rewrite rule l → r ∈ R, and a substitution σ such that s|p = lσ and t = s[rσ]p ; the rewrite step is sometimes denoted by s →p,l→r t to make explicit the position and rule used in this step. The instantiated left-hand side lσ is called a redex. A term s is called irreducible or in normal form with respect to a TRS R if there is no term t with s →R t. A substitution is called normalized with respect to R if every variable in the domain is replaced by a normal form with respect to R. We sometimes omit “with respect to R” if it is clear from the context. A derivation is a (possibly empty) sequence of rewrite steps. Given a binary relation →, we denote by →∗ its reflexive and transitive closure, i.e., s →∗R t means that s can be reduced to t in R in zero or more steps; we also use s →nR t to denote that s can be reduced to t in exactly n steps. 5 We further assume that rewrite rules are labelled, i.e., given a TRS R, we denote by β : l → r a rewrite rule with label β. Labels are unique in a TRS. Also, to relate label β to fixed variables, we consider that the variables of the rewrite rules are not renamed1 and that the reduced terms are always ground. Equivalently, one could require terms to be variable disjoint with the variables of the rewrite system, but we require groundness for simplicity. We often write s →p,β t instead of s →p,l→r t if rule l → r is labeled with β. 2.3. Conditional Term Rewrite Systems In this paper, we also consider conditional term rewrite systems (CTRSs); namely oriented 3-CTRSs, i.e., CTRSs where extra variables are allowed as long as Var(r) ⊆ Var(l) ∪ Var(C) for any rule l → r ⇐ C [26]. In oriented CTRSs, a conditional rule l → r ⇐ C has the form l → r ⇐ s1 ։ t1 , . . . , sn ։ tn , where each oriented equation si ։ ti is interpreted as reachability (→∗R ). In the following, we denote by on a sequence of elements o1 , . . . , on for some n. We also write oi,j for the sequence oi , . . . , oj when i ≤ j (and the empty sequence otherwise). We write o when the number of elements is not relevant. In addition, we denote a condition o1 ։ o′1 , . . . , on ։ o′n by on ։ o′n . As in the unconditional case, we consider that rules are labelled and that labels are unique in a CTRS. And, again, to relate label β to fixed variables, we consider that the variables of the conditional rewrite rules are not renamed and that the reduced terms are always ground. For a CTRS R, the associated rewrite relation →R is defined as the smallest binary relation satisfying the following: given ground terms s, t ∈ T (F ), we have s →R t iff there exist a position p in s, a rewrite rule l → r ⇐ sn ։ tn ∈ R, and a ground substitution σ such that s|p = lσ, si σ →∗R ti σ for all i = 1, . . . , n, and t = s[rσ]p . In order to simplify the presentation, we only consider deterministic CTRSs (DCTRSs), i.e., oriented 3-CTRSs where, for each rule l → r ⇐ sn ։ tn , we have Var(si ) ⊆ Var(l, ti−1 ) for all i = 1, . . . , n (see Section 3.2 for a justification of this requirement and how it could be relaxed to arbitrary 3-CTRSs). Intuitively speaking, the use of DCTRs allows us to compute the bindings for the variables in the condition of a rule in a deterministic way. E.g., given a ground term s and a rule β : l → r ⇐ sn ։ tn with s|p = lθ, we have that s1 θ is ground. Therefore, one can reduce s1 θ to some term s′1 such that s′1 is an instance of t1 θ with some ground substitution θ1 . Now, we have 1 This will become useful in the next section where the reversible extension of rewriting keeps a “history” of a computation in the form of a list of terms β(p, σ), and we want the domain of σ to be a subset of the left-hand side of the rule labelled with β. 6 that s2 θθ1 is ground and we can reduce s2 θθ1 to some term s′2 such that s′2 is an instance of t2 θθ1 with some ground substitution θ2 , and so forth. If all equations in the condition hold using θ1 , . . . , θn , we have that s →p,β s[rσ]p with σ = θθ1 . . . θn . Example 1. Consider the following DCTRS R that defines the function double that doubles the value of its argument when it is an even natural number: β1 : add(0, y) → y β4 : even(0) → true β2 : add(s(x), y) → s(add(x, y)) β5 : even(s(s(x))) → even(x) β3 : double(x) → add(x, x) ⇐ even(x) ։ true Given the term double(s(s(0))) we have, for instance, the following derivation: double(s(s(0)))→ǫ,β3 add(s(s(0)), s(s(0))) since even(s(s(0))) →∗R true with σ = {x 7→ s(s(0))} →ǫ,β2 s(add(s(0), s(s(0)))) with σ = {x 7→ s(0), y 7→ s(s(0))} →1,β2 s(s(add(0, s(s(0))))) with σ = {x 7→ 0, y 7→ s(s(0))} →1.1,β1 s(s(s(s(0)))) with σ = {y 7→ s(s(0))} 3. Reversible Term Rewriting In this section, we present a conservative extension of the rewrite relation which becomes reversible. In the following, we use ⇀R to denote our reversible (forward) term rewrite relation, and ↽R to denote its application in the reverse (backward) direction. Note that, in principle, we do not require ↽R = ⇀−1 R , i.e., we provide independent (constructive) definitions for each relation. Nonetheless, we will prove that ↽R = ⇀−1 R indeed holds (cf. Theorems 9 and 20). In some approaches to reversible computing, both forward and backward relations should be deterministic. Here, we will only require deterministic backward steps, while forward steps might be non-deterministic, as it is often the case in term rewriting. 3.1. Unconditional Term Rewrite Systems We start with unconditional TRSs since it is conceptually simpler and thus will help the reader to better understand the key ingredients of our approach. In the next section, we will consider the more general case of DCTRSs. Given a TRS R, reversible rewriting is defined on pairs ht, πi, where t is a ground term and π is a trace (the “history” of the computation so far). Here, a trace in R is a list of trace terms of the form β(p, σ) such that β is 7 a label for some rule l → r ∈ R, p is a position, and σ is a substitution with Dom(σ) = Var(l)\Var(r) which will record the bindings of erased variables when Var(l)\Var(r) 6= ∅ (and σ = id if Var(l)\Var(r) = ∅).2 Our trace terms have some similarities with proof terms [36]. However, proof terms do not store the bindings of erased variables (and, to the best of our knowledge, they are only defined for unconditional TRSs, while we use trace terms both for unconditional and conditional TRSs). Our reversible term rewriting relation is only defined on safe pairs: Definition 2. Let R be a TRS. The pair hs, πi is safe in R iff, for all β(p, σ) in π, σ is a ground substitution with Dom(σ) = Var(l)\Var(r) and β : l → r ∈ R. In the following, we often omit R when referring to traces and safe pairs if the underlying TRS is clear from the context. Safety is not necessary when applying a forward reduction step, but will become essential for the backward relation ↽R to be correct. E.g., all traces that come from the forward reduction of some initial pair with an empty trace will be safe (see below). Reversible rewriting is then introduced as follows: Definition 3. Let R be a TRS. A reversible rewrite relation ⇀R is defined on safe pairs ht, πi, where t is a ground term and π is a trace in R. The reversible rewrite relation extends standard rewriting as follows:3 hs, πi ⇀R ht, β(p, σ ′ ) : πi iff there exist a position p ∈ Pos(s), a rewrite rule β : l → r ∈ R, and a ground substitution σ such that s|p = lσ, t = s[rσ]p , and σ ′ = σ|`Var(l)\Var(r) . The reverse relation, ↽R , is then defined as follows: ht, β(p, σ ′) : πi ↽R hs, πi iff ht, β(p, σ ′) : πi is a safe pair in R and there exist a ground substitution θ and a rule β : l → r ∈ R such that Dom(θ) = Var(r), t|p = rθ and s = t[lθσ ′ ]p . Note that θσ ′ = σ ′ θ = θ ∪ σ ′ , where ∪ is the union of substitutions, since Dom(θ) = Var(r), Dom(σ ′ ) = (Var(l)\Var(r)) and both substitutions are ground, so Dom(θ) ∩ Dom(σ ′ ) = ∅. 2 Note that if a rule l → r is non-erasing, i.e., Var(l) = Var(r), then σ = id. In the following, we consider the usual infix notation for lists where [ ] is the empty list and x : xs is a list with head x and tail xs. 3 8 We denote the union of both relations ⇀R ∪ ↽R by ⇋R . Example 4. Let us consider the following TRS R defining the addition on natural numbers built from 0 and s( ), and the function fst that returns its first argument: β1 : add(0, y) → y β2 : add(s(x), y) → s(add(x, y)) β3 : fst(x, y) → x Given the term fst(add(s(0), 0), 0), we have, for instance, the following reversible (forward) derivation: hfst(add(s(0), 0), 0), [ ]i ⇀R hfst(s(add(0, 0)), 0), [β2(1, id)]i ⇀R hs(add(0, 0)), [β3(ǫ, {y 7→ 0}), β2 (1, id)]i ⇀R hs(0), [β1 (1, id), β3 (ǫ, {y 7→ 0}), β2(1, id)]i The reader can easily check that hs(0), [β1 (1, id), β3(ǫ, {y 7→ 0}), β2 (1, id)]i is reducible to hfst(add(s(0), 0), 0), [ ]i using the backward relation ↽R by performing exactly the same steps but in the backward direction. An easy but essential property of ⇀R is that it is a conservative extension of standard rewriting in the following sense (we omit its proof since it is straightforward): Theorem 5. Let R be a TRS. Given terms s, t, if s →∗R t, then for any trace π there exists a trace π ′ such that hs, πi ⇀∗R ht, π ′ i. Here, and in the following, we assume that ←R = (→R )−1 , i.e., s →−1 R t is denoted by s ←R t. Observe that the backward relation is not a conservative extension of ←R : in general, t ←R s does not imply ht, π ′ i ↽R hs, πi for any arbitrary trace π ′ . This is actually the purpose of our notion of reversible rewriting: ↽R should not extend ←R but is only aimed at performing exactly the same steps of the forward computation whose trace was stored, but in the reverse order. Nevertheless, one can still ensure that for all steps t ←R s, there exists some trace π ′ such that ht, π ′i ↽R hs, πi (which is an easy consequence of the above result and Theorem 9 below). Example 6. Consider again the following TRS R = {β : snd(x, y) → y}. Given the reduction snd(1, 2) →R 2, there are infinitely many reductions for 2 using ←R , e.g., 2 ←R snd(1, 2), 2 ←R snd(2, 2), 2 ←R snd(3, 2), etc. The relation is also non-terminating: 2 ←R snd(1, 2) ←R snd(1, snd(1, 2)) ←R · · · . In contrast, given a pair h2, πi, we can only perform a single deterministic and finite reduction (as proved below). For instance, if π = [β(ǫ, {x 7→ 1}), β(2, {x 7→ 1})], then the only possible reduction is h2, πi ↽R hsnd(1, 2), [β(2, {x 7→ 1})]i ↽R hsnd(1, snd(1, 2)), [ ]i 6↽R . 9 Now, we state a lemma which shows that safe pairs are preserved through reversible term rewriting (both in the forward and backward directions): Lemma 7. Let R be a TRS. Let hs, πi be a safe pair. If hs, πi ⇋∗R ht, π ′ i, then ht, π ′ i is also safe. Proof. We prove the claim by induction on the length k of the derivation. Since the base case k = 0 is trivial, consider the inductive case k > 0. Assume a derivation of the form hs, πi ⇋∗R hs0 , π0 i ⇋R ht, π ′ i. By the induction hypothesis, we have that hs0 , π0 i is a safe pair. Now, we distinguish two cases depending on the last step. If we have hs0 , π0 i ⇀R ht, π ′i, then there exist a position p ∈ Pos(s0 ), a rewrite rule β : l → r ∈ R, and a ground substitution σ such that s0 |p = lσ, t = s0 [rσ]p , σ ′ = σ |`Var(l)\Var(r) , and π ′ = β(p, σ ′) : π0 . Then, since σ ′ is ground and Dom(σ ′ ) = Var(l)\Var(r) by construction, the claim follows straightforwardly. If the last step has the form hs0 , π0 i ↽R ht, π ′ i, then the claim follows trivially since each step with ↽R only removes trace terms from π0 . ✷ Hence, since any pair with an empty trace is safe the following result, which states that every pair that is reachable from an initial pair with an empty trace is safe, straightforwardly follows from Lemma 7: Proposition 8. Let R be a TRS. If hs, [ ]i ⇋∗R ht, πi, then ht, πi is safe. Now, we state the reversibility of ⇀R , i.e., the fact that (⇀R)−1 = ↽R (and thus the reversibility of ↽R and ⇋R , too). Theorem 9. Let R be a TRS. Given the safe pairs hs, πi and ht, π ′ i, for all n ≥ 0, hs, πi ⇀nR ht, π ′ i iff ht, π ′ i ↽nR hs, πi. Proof. (⇒) We prove the claim by induction on the length n of the derivation hs, πi ⇀nR ht, π ′i. Since the base case n = 0 is trivial, let us consider the n−1 inductive case n > 0. Consider a derivation hs, πi ⇀R hs0 , π0 i ⇀R ht, π ′i. ′ By Lemma 7, both hs0 , π0 i and ht, π i are safe. By the induction hypothen−1 sis, we have hs0 , π0 i ↽R hs, πi. Consider now the step hs0 , π0 i ⇀R ht, π ′i. Then, there is a position p ∈ Pos(s0 ), a rule β : l → r ∈ R and a ground substitution σ such that s0 |p = lσ, t = s0 [rσ]p , σ ′ = σ |`Var(l)\Var(r) , and π ′ = β(p, σ ′) : π0 . Let θ = σ|`Var(r) . Then, we have ht, π ′ i ↽R hs′0 , π0 i with t|p = rθ, β : l → r ∈ R and s′0 = t[lθσ ′ ]p . Moreover, since σ = θσ ′ , we have s′0 = t[lθσ ′ ]p = t[lσ]p = s0 , and the claim follows. (⇐) This direction proceeds in a similar way. We prove the claim by induction on the length n of the derivation ht, π ′i ↽nR hs, πi. As before, 10 we only consider the inductive case n > 0. Let us consider a derivation n−1 ht, π ′ i ↽R hs0 , π0 i ↽R hs, πi. By Lemma 7, both hs0 , π0 i and hs, πi are n−1 safe. By the induction hypothesis, we have hs0 , π0 i ⇀R ht, π ′ i. Consider now the reduction step hs0 , π0 i ↽R hs, πi. Then, we have π0 = β(p, σ ′ ) : π, β : l → r ∈ R, and there exists a ground substitution θ with Dom(θ) = Var(r) such that s0 |p = rθ and s = s0 [lθσ ′ ]p . Moreover, since hs0 , π0 i is safe, we have that Dom(σ ′ ) = Var(l)\Var(r) and, thus, Dom(θ) ∩ Dom(σ ′ ) = ∅. Let σ = θσ ′ . Then, since s|p = lσ and Dom(σ ′ ) = Var(l)\Var(r), we can perform the step hs, πi ⇀R hs′0 , β(p, σ ′) : πi with s′0 = s[rσ]p = s[rθσ ′ ]p = s[rθ]p = s0 [rθ]p = s0 , and the claim follows. ✷ The next corollary is then immediate: Corollary 10. Let R be a TRS. Given the safe pairs hs, πi and ht, π ′ i, for all n ≥ 0, hs, πi ⇋nR ht, π ′ i iff ht, π ′ i ⇋nR hs, πi. A key issue of our notion of reversible rewriting is that the backward rewrite relation ↽R is deterministic (thus confluent), terminating, and has a constructive definition: Theorem 11. Let R be a TRS. Given a safe pair ht, π ′ i, there exists at most one pair hs, πi such that ht, π ′ i ↽R hs, πi. Proof. First, if there is no step using ↽R from ht, π ′ i, the claim follows trivially. Now, assume there is at least one step ht, π ′ i ↽R hs, πi. We prove that this is the only possible step. By definition, we have π ′ = β(p, σ ′) : π, p ∈ Pos(t), β : l → r ∈ R, and there exists a ground substitution θ with Dom(θ) = Var(r) such that t|p = rθ and s = t[lθσ ′ ]p . The only source of nondeterminism may come from choosing a rule labeled with β and from the computation of the substitution θ. The claim follows trivially from the fact that labels are unique in R and that, if there is some ground substitution θ′ with θ′ = Var(r) and t|p = rθ′ , then θ = θ′ . ✷ Therefore, ↽R is clearly deterministic and confluent. Termination holds straightforwardly for pairs with finite traces since its length strictly decreases with every backward step. Note however that even when ⇀R and ↽R are terminating, the relation ⇋R is always non-terminating since one can keep going back and forth. 11 3.2. Conditional Term Rewrite Systems In this section, we extend the previous notions and results to DCTRSs. We note that considering DCTRSs is not enough to make conditional rewriting deterministic. In general, given a rewrite step s →p,β t with p a position of s, β : l → r ⇐ sn → tn a rule, and σ a substitution such that s|p = lσ and si σ →∗R ti σ for all i = 1, . . . , n, there are three potential sources of nondeterminism: the selected position p, the selected rule β, and the substitution σ. The use of DCTRSs can only make deterministic the last one, but the choice of a position and the selection of a rule may still be non-deterministic. For DCTRSs, the notion of a trace term used for TRSs is not sufficient since we also need to store the traces of the subderivations associated to the condition of the applied rule (if any). Therefore, we generalize the notion of a trace as follows: Definition 12 (trace). Given a CTRS R, a trace in R is recursively defined as follows: • the empty list is a trace; • if π, π1 , . . . , πn are traces in R, n ≥ 0, β : l → r ⇐ sn ։ tn ∈ R is a rule, p is a position, and σ is a ground substitution, then β(p, σ, π1 , . . . , πn ) : π is a trace in R. We refer to each component β(p, σ, π1 , . . . , πn ) in a trace as a trace term. Intuitively speaking, a trace term β(p, σ, π1 , . . . , πn ) stores the position of a reduction step, a substitution with the bindings that are required for the step to be reversible (e.g., the bindings for the erased variables, but not only; see below) and the traces associated to the subcomputations in the condition. The notion of a safe pair is now more involved in order to deal with conditional rules. The motivation for this definition will be explained below, after introducing reversible rewriting for DCTRSs. Definition 13 (safe pair). Let R be a DCTRS. A trace π is safe in R iff, for all trace terms β(p, σ, Sπn ) in π, σ is a ground substitution with Dom(σ) = (Var(l)\Var(r, sn , tn )) ∪ ni=1 Var(ti )\Var(r, si+1,n ), β : l → r ⇐ sn ։ tn ∈ R, and πn are safe too. The pair hs, πi is safe in R iff π is safe. Reversible (conditional) rewriting can now be introduced as follows: Definition 14 (reversible rewriting). Let R be a DCTRS. The reversible rewrite relation ⇀R is defined on safe pairs ht, πi, where t is a ground term 12 and π is a trace in R. The reversible rewrite relation extends standard conditional rewriting as follows: hs, πi ⇀R ht, β(p, σ ′, π1 , . . . , πn ) : πi iff there exist a position p ∈ Pos(s), a rewrite rule β : l → r ⇐ sn ։ tn ∈ R, and a ground substitution σ such that s|p = lσ, hsi σ, [ ]i ⇀∗R hti σ, πi i for all i = 1, . . . , n, t = s[rσ]p , and σ ′ = σ|`(Var(l)\Var(r,sn ,tn ))∪Sn Var(ti )\Var(r,si+1,n ) . i=1 The reverse relation, ↽R , is then defined as follows: ht, β(p, σ ′ , π1 , . . . , πn ) : πi ↽R hs, πi iff ht, β(p, σ ′, πn ) : πi is a safe pair in R, β : l → r ⇐ sn ։ tn ∈ R and there is a ground substitution θ such that Dom(θ) = Var(r, sn )\Dom(σ ′ ), t|p = rθ, hti θσ ′ , πi i ↽∗R hsi θσ ′ , [ ]i for all i = 1, . . . , n, and s = t[lθσ ′ ]p . Note that θσ ′ = σ ′ θ = θ ∪ σ ′ since Dom(θ) ∩ Dom(σ ′ ) = ∅ and both substitutions are ground. As in the unconditional case, we denote the union of both relations ⇀R ∪ ↽R by ⇋R . Example 15. Consider again the DCTRS R from Example 1: β1 : add(0, y) → y β4 : even(0) → true β2 : add(s(x), y) → s(add(x, y)) β5 : even(s(s(x))) → even(x) β3 : double(x) → add(x, x) ⇐ even(x) ։ true Given the term double(s(s(0))), we have, for instance, the following forward derivation: hdouble(s(s(0))), [ ]i ⇀R hadd(s(s(0)), s(s(0))), [β3 (ǫ, id, π)]i ⇀R · · · ⇀R hs(s(s(s(0)))), [β1 (1.1, id), β2(1, id), β2 (ǫ, id), β3 (ǫ, id, π)]i where π = [β4 (ǫ, id), β5 (ǫ, id)] since we have the following reduction: heven(s(s(0))), [ ]i ⇀R heven(0), [β5 (ǫ, id)]i ⇀R htrue, [β4 (ǫ, id), β5 (ǫ, id)]i The reader can easily construct the associated backward derivation: hadd(s(s(0)), s(s(0))), [β1 (1.1, id), β2(1, id), . . .]i ↽∗R hdouble(s(s(0))), [ ]i Let us now explain why we need to store σ ′ in a step of the form hs, πi ⇀R ht, β(p, σ ′ , πn ) : πi. Given a DCTRS, for each rule l → r ⇐ sn ։ tn , the following conditions hold: 13 • 3-CTRS: Var(r) ⊆ Var(l, sn , tn ). • Determinism: for all i = 1, . . . , n, we have Var(si ) ⊆ Var(l, ti−1 ). Intuitively, the backward relation ↽R can be seen as equivalent to the forward relation ⇀R but using a reverse rule of the form r → l ⇐ tn ։ sn , . . . , t1 ։ s1 . Therefore, in order to ensure that backward reduction is deterministic, we need the same conditions as above but on the reverse rule:4 • 3-CTRS: Var(l) ⊆ Var(r, sn , tn ). • Determinism: for all i = 1, . . . , n, Var(ti ) ⊆ Var(r, si+1,n ). Since these conditions cannot be guaranteed in general, we store σ ′ = σ|`(Var(l)\Var(r,sn ,tn ))∪Sn i=1 Var(ti )\Var(r,si+1,n ) in the trace term so that (r → l ⇐ tn ։ sn , . . . , t1 ։ s1 )σ ′ is deterministic and fulfills the conditions of a 3-CTRS by construction, i.e., Var(lσ ′ ) ⊆ Var(rσ ′ , sn σ ′ , tn σ ′ ) and for all i = 1, . . . , n, Var(ti σ ′ ) ⊆ Var(rσ ′ , si+1,n σ ′ ); see the proof of Theorem 21 for more details. Example 16. Consider the following DCTRS: β1 : f(x, y, m) → s(w) ⇐ h(x) ։ x, g(y, 4) ։ w β2 : h(0) → 0 β3 : h(1) → 1 β4 : g(x, y) → x and the step hf(0, 2, 4), [ ]i ⇀R hs(2), [β1 (ǫ, σ ′ , π1 , π2 )]i with σ ′ = {m 7→ 4, x 7→ 0}, π1 = [β2 (ǫ, id)] and π2 = [β4 (ǫ, {y 7→ 4})]. The binding of variable m is required to recover the value of the erased variable m, but the binding of variable x is also needed to perform the subderivation hx, π1 i ↽R hh(x), [ ]i when applying a backward step from hs(2), [β1 (ǫ, σ ′ , π1 , π2 )]i. If the binding for x were unknown, this step would not be deterministic. As mentioned above, an instantiated reverse rule (s(w) → f(x, y, m) ⇐ w ։ g(y, 4), x ։ h(x))σ ′ = s(w) → f(0, y, 4) ⇐ w ։ g(y, 4), 0 ։ h(0) would be a legal DCTRS rule thanks to σ ′ . We note that similar conditions could be defined for arbitrary 3-CTRSs. However, the conditions would be much more involved; e.g., one had to compute first the variable dependencies between the equations in the conditions. 4 We note that the notion of a non-erasing rule is extended to the DCTRSs in [32], which results in a similar condition. 14 Therefore, we prefer to keep the simpler conditions for DCTRSs (where these dependencies are fixed), which is still quite a general class of CTRSs. Reversible rewriting is also a conservative extension of rewriting for DCTRSs (we omit the proof since it is straightforward): Theorem 17. Let R be a DCTRS. Given ground terms s, t, if s →∗R t, then for any trace π there exists a trace π ′ such that hs, πi ⇀∗R ht, π ′ i. For the following result, we need some preliminary notions (see, e.g., [36]). For every oriented CTRS R, we inductively define the TRSs Rk , k ≥ 0, as follows: R0 = ∅ Rk+1 = {lσ → rσ | l → r ⇐ sn ։ tn ∈ R, si σ →∗Rk ti σ for all i = 1, . . . , n} S Observe that Rk ⊆ Rk+1 for all k ≥ 0. We have →R = i≥0 →Ri . We also have s →R t iff s →Rk t for some k ≥ 0. The minimum such k is called the depth of s →R t, and the maximum depth k of s = s0 →Rk1 · · · →Rkm sm = t (i.e., k is the maximum of depths k1 , . . . , km ) is called the depth of the derivation. If a derivation has depth k and length m, we write s →m Rk t. Analogous notions can naturally be defined for ⇀R , ↽R , and ⇋R . The next result shows that safe pairs are also preserved through reversible rewriting with DCTRSs: Lemma 18. Let R be a DCTRS and hs, πi a safe pair. If hs, πi ⇋∗R ht, π ′ i, then ht, π ′ i is also safe. Proof. We prove the claim by induction on the lexicographic product (k, m) ′ of the depth k and the length m of the derivation hs, πi ⇋m Rk ht, π i. Since the base case is trivial, we consider the inductive case (k, m) > (0, 0). Consider a derivation hs, πi ⇋m−1 hs0 , π0 i ⇋Rk ht, π ′ i. By the induction hypotheRk sis, we have that hs0 , π0 i is safe. Now, we distinguish two cases depending on the last step. If the last step is hs0 , π0 i ⇀Rk ht, π ′ i, then there exist a position p ∈ Pos(s0 ), a rewrite rule β : l → r ⇐ sn ։ tn ∈ R, and a ground substitution σ such that s0 |p = lσ, hsi σ, [ ]i ⇀∗Rk hti σ, πi i for all i i = 1, . . . , n, t = s0 [rσ]p , σ ′ = σ |`(Var(l)\Var(r,sn ,tn ))∪Sn Var(ti )\Var(r,si+1,n ) , and i=1 π ′ = β(p, σ ′ , π1 , . . . , πn ). Then, sinceSki < k, i = 1, . . . , n, σ ′ is ground and Dom(σ ′ ) = (Var(l)\Var(r, sn , tn )) ∪ ni=1 Var(ti )\Var(r, si+1,n ) by construction, the claim follows by induction. Finally, if the last step has the form hs0 , π0 i ↽Rk ht, π ′ i, then the claim follows trivially since a step with ↽R only removes trace terms from π0 . ✷ 15 As in the unconditional case, the following proposition follows straightforwardly from the previous lemma since any pair with an empty trace is safe. Proposition 19. Let R be a DCTRS. If hs, [ ]i ⇋∗R ht, πi, then ht, πi is safe in R. Now, we can already state the reversibility of ⇀R for DCTRSs: Theorem 20. Let R be a DCTRS. Given the safe pairs hs, πi and ht, π ′ i, ′ ′ m for all k, m ≥ 0, hs, πi ⇀m Rk ht, π i iff ht, π i ↽Rk hs, πi. Proof. (⇒) We prove the claim by induction on the lexicographic product ′ (k, m) of the depth k and the length m of the derivation hs, πi ⇀m Rk ht, π i. Since the base case is trivial, we consider the inductive case (k, m) > (0, 0). ′ Consider a derivation hs, πi ⇀m−1 Rk hs0 , π0 i ⇀Rk ht, π i whose associated product is (k, m). By Proposition 19, both hs0 , π0 i and ht, π ′ i are safe. By the induction hypothesis, since (k, m − 1) < (k, m), we have hs0 , π0 i ↽m−1 Rk hs, πi. Consider now the step hs0 , π0 i ⇀Rk ht, π ′ i. Thus, there exist a position p ∈ Pos(s0 ), a rule β : l → r ⇐ sn ։ tn ∈ R, and a ground substitution σ such that s0 |p = lσ, hsi σ, [ ]i ⇀∗Rk hti σ, πi i for all i = 1, . . . , n, t = s0 [rσ]p , i σ ′ = σ|`(Var(l)\Var(r,sn ,tn ))∪Sn Var(ti )\Var(r,si+1,n ) , and π ′ = β(p, σ ′ , π1 , . . . , πn ) : π0 . i=1 By definition of ⇀Rk , we have that ki < k and, thus, (ki , m1 ) < (k, m2 ) for all i = 1, . . . , n and for all m1 , m2 . Hence, by the induction hypothesis, we have hti σ, πi i ↽∗Rk hsi σ, [ ]i for all i = 1, . . . , n. Let θ = σ |`Var(r,sn )\Dom(σ′ ) , so that i σ = θσ ′ and Dom(θ) ∩ Dom(σ ′ ) = ∅. Therefore, we have ht, π ′ i ↽Rk hs′0 , π0 i with t|p = rθ, β : l → r ⇐ sn ։ tn ∈ R and s′0 = t[lθσ ′ ]p = t[lσ]p = s0 , and the claim follows. (⇐) This direction proceeds in a similar way. We prove the claim by induction on the lexicographic product (k, m) of the depth k and the length m of the considered derivation. Since the base case is trivial, let us consider the inductive case (k, m) > (0, 0). Consider a derivation ht, π ′i ↽m−1 Rk hs0 , π0 i ↽Rk hs, πi whose associated product is (k, m). By Proposition 19, both hs0 , π0 i and hs, πi are safe. By the induction hypothesis, since (k, m − ′ 1) < (k, m), we have hs0 , π0 i ⇀m−1 Rk ht, π i. Consider now the step hs0 , π0 i ↽Rk hs, πi. Then, we have π0 = β(p, σ ′ , π1 , . . . , πn ) : π, β : l → r ⇐ sn ։ tn ∈ R, and there exists a ground substitution θ with Dom(θ) = Var(r, sn )\Dom(σ ′ ) such that s0 |p = rθ, hti θσ ′ , πi i ↽∗Rk hsi θσ ′ , [ ]i for all i = 1, . . . , n, and i ′ s = s0 [lθσ ′ ]p . Moreover, since hs , π 0 0 i is safe, we have that Dom(σ ) = Sn (Var(l)\Var(r, sn , tn )) ∪ i=1 Var(ti )\Var(r, si+1,n ). By definition of ↽Rk , we have that ki < k and, thus, (ki , m1 ) < (k, m2 ) for all i = 1, . . . , n and for all m1 , m2 . By the induction hypothesis, we have hsi θσ ′ , [ ]i ⇀∗Rk hti θσ ′ , πi i i 16 for all i = 1, . . . , n. Let σ = θσ ′ , with Dom(θ) ∩ Dom(σ ′ ) = ∅. Then, since s|p = lσ, we can perform the step hs, πi ⇀Rk hs′0 , β(p, σ ′, π1 , . . . , πn ) : πi with s′0 = s[rσ]p = s[rθσ ′ ]p ; moreover, s[rθσ ′ ]p = s[rθ]p = s0 [rθ]p = s0 since Dom(σ ′ ) ∩ Var(r) = ∅, which concludes the proof. ✷ In the following, we say that ht, π ′ i ↽R hs, πi is a deterministic step if there is no other, different pair hs′′ , π ′′ i with ht, π ′ i ↽R hs′′ , π ′′ i and, moreover, the subderivations for the equations in the condition of the applied rule (if any) are deterministic, too. We say that a derivation ht, π ′ i ↽∗R hs, πi is deterministic if each reduction step in the derivation is deterministic. Now, we can already prove that backward reversible rewriting is also deterministic, as in the unconditional case: Theorem 21. Let R be a DCTRS. Let ht, π ′ i be a safe pair with ht, π ′ i ↽∗R hs, πi for some term s and trace π. Then ht, π ′i ↽∗R hs, πi is deterministic. Proof. We prove the claim by induction on the lexicographic product (k, m) of the depth k and the length m of the steps. The case m = 0 is trivial, and thus we let m > 0. Assume ht, π ′ i ↽m−1 hu, π ′′ i ↽Rk hs, πi. For the base Rk case k = 1, the applied rule is unconditional and the proof is analogous to that of Theorem 11. Let us now consider k > 1. By definition, if hu, π ′′ i ↽Rk hs, πi, we have π ′′ = β(p, σ ′, π1 , . . . , πn ) : π, β : l → r ⇐ sn ։ tn ∈ R and there exists a ground substitution θ with Dom(θ) = Var(r) such that u|p = rθ, hti θσ ′ , πi i ↽∗Rj hsi θσ ′ , [ ]i, j < k, for all i = 1, . . . , n, and s = t[lθσ ′ ]p . By the induction hypothesis, the subderivations hti θσ ′ , πi i ↽∗Rj hsi θσ ′ , [ ]i are deterministic, i.e., hsi θσ ′ , [ ]i is a unique resulting term obtained by reducing hti θσ ′ , πi i. Therefore, the only remaining source of nondeterminism can come from choosing a rule labeled with β and from the computed substitution θ. On the one hand, the labels are unique in R. As for θ, we prove that this is indeed the only possible substitution for the reduction step. Consider the instance of rule l → r ⇐ sn ։ tn with σ ′ : lσ ′ → rσ ′ ⇐ sn σ ′ ։ tn σ ′ . Since hu, π ′′i is safe, we S have that σ ′ is a ground substitution and Dom(σ ′ ) = (Var(l)\Var(r, sn , tn )) ∪ ni=1 Var(ti )\Var(r, si+1,n ). Then, the following properties hold: • Var(lσ ′ ) ⊆ Var(rσ ′ , sn σ ′ , tn σ ′ ), since σ ′ is ground and it covers all the variables in Var(l)\Var(r, sn , tn ). • Var(ti σ ′ ) ⊆ Var(rσ ′ , si+1,nSσ ′ ) for all i = 1, . . . , n, since σ ′ is ground and it covers all variables in ni=1 Var(ti )\Var(r, si+1,n ). 17 The above properties guarantee that a rule of the form rσ ′ → lσ ′ ⇐ tn σ ′ ։ sn σ ′ , . . . , t1 σ ′ ։ s1 σ ′ can be seen as a rule of a DCTRS and, thus, there exists a deterministic procedure to compute θ, which completes the proof. ✷ Therefore, ↽R is deterministic and confluent. Termination is trivially guaranteed for pairs with a finite trace since the trace’s length strictly decreases with every backward step. 4. Removing Positions from Traces Once we have a feasible definition of reversible rewriting, there are two refinements that can be considered: i) reducing the size of the traces and ii) defining a reversibilization transformation so that standard rewriting becomes reversible in the transformed system. In this section, we consider the first problem, leaving the second one for the next section. In principle, one could remove information from the traces by requiring certain conditions on the considered systems. For instance, requiring injective functions may help to remove rule labels from trace terms. Also, requiring non-erasing rules may help to remove the second component of trace terms (i.e., the substitutions). In this section, however, we deal with a more challenging topic: removing positions from traces. This is useful not only to reduce the size of the traces but it is also essential to define a reversibilization technique for DCTRSs in the next section.5 In particular, we aim at transforming a given DCTRS into one that fulfills some conditions that make storing positions unnecessary. In the following, given a CTRS R, we say that a term t is basic [18] if it has the form f (tn ) with f ∈ DR a defined function symbol and tn ∈ T (CR , V) constructor terms. Furthermore, in the remainder of this paper, we assume that the right-hand sides of the equations in the conditions of the rules of a DCTRS are constructor terms. This is not a significant restriction since these terms cannot be reduced anyway (since we consider oriented equations in this paper), and still covers most practical examples. Now, we introduce the following subclass of DCTRSs: Definition 22 (pcDCTRS [30]). We say that a DCTRS R is a pcDCTRS (“pc” stands for pure constructor ) if, for each rule l → r ⇐ sn ։ tn ∈ R, we have that l and sn are basic terms and r and tn are constructor terms. 5 We note that defining a transformation with traces that include positions would be a rather difficult task because positions are dynamic (i.e., they depend on the term being reduced) and thus would require a complex (and inefficient) system instrumentation. 18 Pure constructor systems are called normalized systems in [3]. Also, they are mostly equivalent to the class IIIn of conditional systems in [8], where t1 , . . . , tn are required to be ground unconditional normal forms instead.6 In principle, any DCTRS with basic terms in the left-hand sides (i.e., a constructor DCTRS) and constructor terms in the right-hand sides of the equations of the rules can be transformed into a pcDCTRS by applying a few simple transformations: flattening and simplification of constructor conditions. Let us now consider each of these transformations separately. Roughly speaking, flattening involves transforming a term (occurring, e.g., in the right-hand side of a DCTRS or in the condition) with nested defined functions like f(g(x)) into a term f(y) and an (oriented) equation g(x) ։ y, where y is a fresh variable. Formally, Definition 23 (flattening). Let R be a CTRS, R = (l → r ⇐ sn ։ tn ) ∈ R be a rule and R′ be a new rule either of the form l → r ⇐ s1 ։ t1 , . . . , si |p ։ w, si [w]p ։ ti , . . . , sn ։ tn , for some p ∈ Pos(si ), 1 6 i 6 n, or l → r[w]q ⇐ sn ։ tn , r|q ։ w, for some q ∈ Pos(r), where w is a fresh variable.7 Then, a CTRS R′ is obtained from R by a flattening step if R′ = (R\{R}) ∪ {R′ }. Note that, if an unconditional rule is non-erasing (i.e., Var(l) ⊆ Var(r) for a rule l → r), any conditional rule obtained by flattening is trivially nonerasing too, according to the notion of non-erasingness for DCTRSs in [32].8 Flattening is trivially complete since any flattening step can be undone by binding the fresh variable again to the selected subterm and, then, proceeding as in the original system. Soundness is more subtle though. In this work, we prove the correctness of flattening for arbitrary DCTRSs with respect to innermost rewriting. As usual, the innermost rewrite relation, in symbols, i →R , is defined as the smallest binary relation satisfying the following: given i ground terms s, t ∈ T (F ), we have s →R t iff there exist a position p in s such that no proper subterms of s|p are reducible, a rewrite rule l → r ⇐ sn ։ tn ∈ R, and a normalized ground substitution σ such that s|p = lσ, i si σ →∗R ti σ, for all i = 1, . . . , n, and t = s[rσ]p . In order to prove the correctness of flattening, we state the following auxiliary lemma: 6 Given a CTRS R, we define Ru = {l → r | l → r ⇐ sn ։ tn ∈ R}. A term is an unconditional normal form in R, if it is a normal form in Ru . 7 The positions p, q can be required to be different from ǫ, but this is not strictly necessary. 8 Roughly, a DCTRS is considered non-erasing in [32] if its transformation into an unconditional TRS by an unraveling transformation gives rise to a non-erasing TRS. 19 Lemma 24. Let R be a DCTRS. Given terms s and t, with t a normal form, i i i and a position p ∈ Pos(s), we have s →∗R t iff s|p →∗R wσ and s[wσ]p →∗R t, for some fresh variable w and normalized substitution σ. Proof. (⇒) Let us consider an arbitrary position p ∈ Pos(s). If s|p is normalized, the proof is straightforward. Otherwise, since we use innermost reduction (leftmost innermost, for simplicity), we can represent the derivation i s →∗R t as follows: i i i s[s|p ]p →∗R s′ [s|p ]p →∗R s′ [s′′ ]p →∗R t i where s′′ is a normal form and the subderivation s[s|p ]p → ∗R s′ [s|p ]p reduces the leftmost innermost subterms that are to the left of s|p (if any). i Then, by choosing σ = {w 7→ s′′ } we have s|p → ∗R wσ (by mimicking the i i steps of s′ [s|p ]p → ∗R s′ [s′′ ]p ), s[wσ]p →∗R s′ [wσ]p (by mimicking the steps of i i i s[s|p ]p →∗R s′ [s|p ]p ), and s′ [wσ]p →∗R t (by mimicking the steps of s′ [s′′ ]p →∗R t), which concludes the proof. (⇐) This direction is perfectly analogous to the previous case. We consider an arbitrary position p ∈ Pos(s) such that s|p is not normalized (otherwise, the proof is trivial). Now, since derivations are innermost, we can i i i consider that s[wσ]p → ∗R t is as follows: s[wσ]p → ∗R s′ [wσ]p → ∗R t, where i s[wσ]p →∗R s′ [wσ]p reduces the innermost subterms to the left of s|p . Therei i fore, we have s[s|p ]p →∗R s′ [s|p ]p (by mimicking the steps of s[wσ]p →∗R s′ [wσ]p ), i i s′ [s|p ]p →∗R s′ [s′′ ]p (by mimicking the steps of s|p →∗R wσ, with σ = {w 7→ s′′ }), i i and s′ [s′′ ]p →∗R t (by mimicking the steps of s′ [wσ]p →∗R t). ✷ The following theorem is an easy consequence of the previous lemma: Theorem 25. Let R be a DCTRS. If R′ is obtained from R by a flattening step, then R′ is a DCTRS and, for all ground terms s, t, with t a normal i i form, we have s →∗R t iff s →∗R′ t. Proof. (⇒) We prove the claim by induction on the lexicographic product i (k, m) of the depth k and the length m of the derivation s →∗Rk t. Since the base case is trivial, we consider the inductive case (k, m) > (0, 0). Assume i i i that s →∗Rk t has the form s[lσ]u →Rk s[rσ]u →∗Rk t with l → r ⇐ sn ։ tn ∈ R i and si σ →∗Rk ti σ, ki < k, i = 1, . . . , n. If l → r ⇐ sn ։ tn ∈ R′ , the claim i follows directly by induction. Otherwise, we have that either l → r ⇐ s1 ։ 20 t1 , . . . , si |p ։ w, si [w]p ։ ti , . . . , sn ։ tn ∈ R′ , for some p ∈ Pos(si ), 1 6 i 6 n, or l → r[w]q ⇐ sn ։ tn , r|q ։ w ∈ R′ , for some q ∈ Pos(r), where w is a fresh variable. Consider first the case l → r ⇐ s1 ։ t1 , . . . , si |p ։ w, si [w]p ։ ti , . . . , sn ։ tn ∈ R′ , for some p ∈ Pos(si ), 1 6 i 6 n. Since i si σ → ∗Rk ti σ, ki < k, i = 1, . . . , n, by the induction hypothesis, we have i i si σ → ∗R′ ti σ, i = 1, . . . , n. By Lemma 24, there exists σ ′ = {w 7→ s′ } i for some normal form s′ such that si |p σ = si |p σσ ′ → ∗Rk wσσ ′ = wσ ′ and i ′ si [w]p σσ = si σ[wσ ′ i ]p →∗Rk ti . i Moreover, since w is an extra variable, we also i →∗R′ tj σ = tj σσ ′ for j = 1, . . . , i − 1, i + 1, . . . , n. Therefore, have sj σσ ′ = sj σ i since lσσ ′ = lσ and rσσ ′ = rσ, we have s[lσ]u →R s[rσ]u , and the claim follows by induction. Consider the second case. By the induction hypothesis, i i we have s[rσ]u → ∗R′ t and si σ → ∗R′ ti σ for all i = 1, . . . , n. By Lemma 24, there exists a substitution σ ′ = {w 7→ s′ } such that s′ is the normal form of i i r|q σ and we have r|q σ →∗R′ wσ ′ and s[rσ[wσ ′ ]q ]u →∗R′ t. Moreover, since w is i a fresh variable, we have si σσ ′ →∗R′ ti σσ ′ for all i = 1, . . . , n. Therefore, we i have s[lσσ ′ ]u = s[lσ]u →R′ s[rσ[wσ ′ ]q ]u , which concludes the proof. (⇐) This direction is perfectly analogous to the previous one, and follows easily by Lemma 24 too. ✷ Let us now consider the second kind of transformations: the simplification of constructor conditions. Basically, we can drop an equation s ։ t when the terms s and t are constructor, called a constructor condition, by either applying the most general unifier (mgu) of s and t (if it exists) to the remaining part of the rule, or by deleting entirely the rule if they do not unify because (under innermost rewriting) the equation will never be satisfied by any normalized substitution. Similar transformations can be found in [33]. In order to justify these transformations, we state and prove the following results. In the following, we let mgu(s, t) denote the most general unifier of terms s and t if it exists, and fail otherwise. Theorem 26 (removal of unifiable constructor conditions). Let R be a DCTRS and let R = (l → r ⇐ sn ։ tn ) ∈ R be a rule with mgu(si , ti ) = θ, for some i ∈ {1, . . . , n}, where si and ti are constructor terms. Let R′ be a new rule of the form lθ → rθ ⇐ s1 θ ։ t1 θ, . . . , si−1 θ ։ ti−1 θ, si+1 θ ։ ti+1 θ, . . . , sn θ ։ tn θ.9 Then R′ = (R\{R}) ∪ {R′ } is a DCTRS and, for all 9 In [33], the condition Dom(θ) ∩ Var(l, r, s1 , t1 , . . . , sn , tn ) = ∅ is required, but this condition is not really necessary. 21 i i ground terms s and t, we have s →∗R t iff s →∗R′ t. Proof. (⇒) First, we prove the following claim by induction on the lexicographic product (k, m) of the depth k and the length m of the steps: if i ∗ i s →m Rk t, then s →R′ t. It suffices to consider the case where R is applied, i.e., i i s = s[lσ]p →{R} s[rσ]p with sj σ →∗Rk tj σ for all j ∈ {1, . . . , n}. By definition, j σ is normalized. Hence, since si and ti are constructor terms, we have that si σ and ti σ are trivially normal forms since the normalized subterms introduced by σ cannot become reducible in a constructor context. Therefore, we have si σ = ti σ. Thus, σ is a unifier of si and ti and, hence, θ is more general than σ. Let δ be a substitution such that σ = θδ. Since σ is normalized, so is δ. Since kj < k for all j = 1, . . . , n, by the induction hypothesis, we have i that sj σ = sj θδ →∗R′ tj θδ = tj σ for j ∈ {1, . . . , i − 1, i + 1, . . . , n}. Therefore, i we have that s[lσ]p = s[lθδ]p →{R′ } s[rθδ]p = s[rσ]p . (⇐) Now, we prove the following claim by induction on the lexicographic i product (k, m) of the depth k and the length m of the steps: if s → m R′ t, k i then s → ∗R t. It suffices i s[lθδ]p →{R} s[rθδ]p with ′ to consider the case where R is applied, i.e., s = i sj θδ →∗R′ tj θδ for all j ∈ {1, . . . , i − 1, i + 1, . . . , n}. kj By the assumption and the definition, θ and δ are normalized, and thus, si θδ and ti θδ are normal forms (as in the previous case, because the normalized subterms introduced by θδ cannot become reducible in a constructor context), i.e., si θδ = ti θδ. Since kj < k for all j ∈ {1, . . . , i − 1, i + 1, . . . , n}, by the i induction hypothesis, we have that sj θδ → ∗R tj θδ for j ∈ {1, . . . , i − 1, i + i 1, . . . , n}. Therefore, we have that s[lσ]p = s[lθδ]p →{R} s[rθδ]p = s[rσ] with σ = θδ. ✷ Now we consider the case when the terms in the constructor condition do not unify: Theorem 27 (removal of infeasible rules). Let R be a DCTRS and let R = (l → r ⇐ sn ։ tn ) ∈ R be a rule with mgu(si , ti ) = fail , for some i ∈ {1, . . . , n}. Then R′ = R\{R} is a DCTRS and, for all ground terms s i i and t, we have s →∗R t iff s →∗R′ t. Proof. Since R ⊇ R′ , the if part is trivial, and thus, we consider the onlyif part. To apply R to a term, there must exist a normalized substitution σ i such that si σ →∗R ti σ. Since si , ti are constructor terms and σ is normalized, si σ is a normal form (because the normalized subterms introduced by σ 22 i cannot become reducible in a constructor context). If si σ →∗R ti σ is satisfied (i.e., si σ = ti σ), then si and ti are unifiable, and thus, this contradicts the i assumption. Therefore, R is never applied to any term, and hence, s →∗R t i iff s →∗R′ t. ✷ Using flattening and the simplification of constructor conditions, any constructor DCTRS with constructor terms in the right-hand sides of the equations of the rules can be transformed into a pcDCTRS. One can use, for instance, the following simple algorithm. Let R be such a constructor DCTRS. We apply the following transformations as much as possible: (flattening-rhs) Assume that R contains a rule of the form R = (l → r ⇐ sn ։ tn ) where r is not a constructor term. Let r|q , q ∈ Pos(r), be a basic subterm of r. Then, we replace rule R by a new rule of the form l → r[w]q ⇐ sn ։ tn , r|q ։ w, where w is a fresh variable. (flattening-condition) Assume that R contains a rule of the form R = (l → r ⇐ sn ։ tn ) where si is neither a constructor term nor a basic term, i ∈ {1, . . . , n}. Let si |q , q ∈ Pos(s1 ), be a basic subterm of si . Then, we replace rule R by a new rule of the form l → r ⇐ s1 ։ t1 , . . . , si |q ։ w, si [w]q ։ ti , . . . , sn ։ tn , where w is a fresh variable. (removal-unify) Assume that R contains a rule of the form R = (l → r ⇐ sn ։ tn ) where si is a constructor term, i ∈ {1, . . . , n}. If mgu(si , ti ) = θ 6= fail , then we replace rule R by a new rule of the form lθ → rθ ⇐ s1 θ ։ t1 θ, . . . , si−1 θ ։ ti−1 θ, si+1 θ ։ ti+1 θ, . . . , sn θ ։ tn θ. (removal-fail) Assume that R contains a rule of the form R = (l → r ⇐ sn ։ tn ) where si is a constructor term, i ∈ {1, . . . , n}. If mgu(si , ti ) = fail , then we remove rule R from R. Trivially, by applying rule flattening-rhs as much as possible, we end up with a DCTRS where all the right-hand sides are constructor terms; analogously, the exhaustive application of rule flattening-condition allows us to ensure that the left-hand sides of all equations in the conditions of the rules are either constructor or basic; finally, the application of rules removal-unify and removal-fail produces a pcDCTRS by removing those equations in which the left-hand side is a constructor term. Therefore, in the remainder of this paper, we only consider pcDCTRSs. A nice property of pcDCTRSs is that one can consider reductions only at topmost positions. Formally, given a pcDCTRS R, we say that s →p,l→r⇐sn։tn t is a top reduction step if p = ǫ, there is a ground substitution σ with s = lσ, 23 si σ →∗R ti σ for all i = 1, . . . , n, t = rσ, and all the steps in si σ →∗R ti σ for ǫ i = 1, . . . , n are also top reduction steps. We denote top reductions with → ǫ ǫ for standard rewriting, and ⇀R , ↽R for our reversible rewrite relations. i ǫ The following result basically states that → and → are equivalent for pcDCTRSs: Theorem 28. Let R be a constructor DCTRS with constructor terms in the right-hand sides of the equations and R′ be a pcDCTRS obtained from R by a sequence of transformations of flattening and simplification of constructor conditions. Given ground terms s and t such that s is basic and t is i ǫ normalized, we have s →∗R t iff s →∗R′ t. Proof. First, it is straightforward to see that an innermost reduction in R′ can only reduce the topmost positions of terms since defined functions can only occur at the root of terms and the terms introduced by instantiation are, by definition, irreducible. Therefore, the claim is a consequence of Theorems 25, 26 and 27, together with the above fact. ✷ Therefore, when considering pcDCTRSs and top reductions, storing the reduced positions in the trace terms becomes redundant since they are always ǫ. Thus, in practice, one can consider simpler trace terms without positions, β(σ, π1 , . . . , πn ), that implicitly represent the trace term β(ǫ, σ, π1 , . . . , πn ). Example 29. Consider the following TRS R defining addition and multiplication on natural numbers, and its associated pcDCTRS R′ : R={ add(0, y) add(s(x), y) mult(0, y) mult(s(x), y) R′ = { add(0, y) add(s(x), y) mult(0, y) mult(s(x), y) → → → → → → → → y, s(add(x, y)), 0, add(mult(x, y), y)} y, s(z) ⇐ add(x, y) ։ z, 0, w ⇐ mult(x, y) ։ z, add(z, y) ։ w} For instance, given the following reduction in R: i i i mult(s(0), s(0)) →R add(mult(0, s(0)), s(0)) →R add(0, s(0)) →R s(0) we have the following counterpart in R′ : ǫ ǫ mult(s(0), s(0)) →R′ s(0) with mult(0, s(0)) →R′ 0 ǫ and add(0, s(0)) →R′ s(0) 24 Trivially, all results in Section 3 hold for pcDCTRSs and top reductions starting from basic terms. The simpler trace terms without positions will allow us to introduce appropriate injectivization and inversion transformations in the next section. 5. Reversibilization In this section, we aim at compiling the reversible extension of rewriting into the system rules. Intuitively speaking, given a pure constructor system R, we aim at producing new systems Rf and Rb such that standard rewriting in Rf , i.e., →Rf , coincides with the forward reversible extension ⇀R in the original system, and analogously →Rb is equivalent to ↽R . Therefore, Rf can be seen as an injectivization of R, and Rb as the inversion of Rf . In principle, we could easily introduce a transformation for pcDCTRSs that mimicks the behavior of the reversible extension of rewriting. For instance, given the pcDCTRS R of Example 16, we could produce the following injectivized version Rf :10 hf(x, y, m), wsi → hs(w), β1 (m, x, w1 , w2 ) : wsi ⇐ hh(x), [ ]i ։ hx, w1 i, hg(y, 4), [ ]i ։ hw, w2i hh(0), wsi → h0, β2 : wsi hh(1), wsi → h1, β3 : wsi hg(x, y), wsi → hx, β4 (y) : wsi ǫ For instance, the reversible step hf(0, 2, 4), [ ]i ⇀R hs(2), [β1 (σ ′ , π1 , π2 )]i with σ ′ = {m 7→ 4, x 7→ 0}, π1 = [β2 (id)] and π2 = [β4 ({y 7→ 4})], has the following counterpart in Rf : ǫ hf(0, 2, 4), [ ]i →Rf hs(2), [β1 (4, 0, [β2 ], [β4 (4)])]i ǫ ǫ with hh(0), [ ]i →Rf h0, [β2 ]i and hg(2, 4), [ ]i →Rf h2, [β4(4)]i The only subtle difference here is that a trace term like β1 ({m 7→ 4, x 7→ 0}, [β2 (id)], [β4 ({y 7→ 4})]) is now stored in the transformed system as β1 (4, 0, [β2 ], [β4 (4)]) 10 We will write just β instead of β() when no argument is required. 25 Furthermore, we could produce an inverse Rb of the above system as follows: hs(w), β1 (m, x, w1 , w2 ) : wsi−1 → hf(x, y, m), wsi−1 ⇐ hw, w2i−1 ։ hg(y, 4), [ ]i−1, hx, w1 i−1 ։ hh(x), [ ]i−1 h0, β2 : wsi−1 → hh(0), wsi−1 h1, β3 : wsi−1 → hh(1), wsi−1 hx, β4 (y) : wsi−1 → hg(x, y), wsi−1 mainly by switching the left- and right-hand sides of each rule and condition. The correctness of these injectivization and inversion transformations would be straightforward. These transformations are only aimed at mimicking, step by step, the reversible relations ⇀R and ↽R . Roughly speaking, for each step hs, πi ⇀R ht, π ′ i in a system R, we have hs, πi →Rf ht, π ′i, where Rf is the injectivized version of R, and for each step hs, πi ↽R ht, π ′ i in R, we have hs, πi →Rb ht, π ′ i, where Rb is the inverse of Rf . More details on this approach can be found in [31]. Unfortunately, it might be much more useful to produce injective and inverse versions of each function defined in a system R. Note that, in the above approach, the system Rf only defines a single function h , i and Rb only defines h , i−1 , i.e., we are computing systems that define the relations ⇀R and ↽R rather than the injectivized and inverse versions of the functions in R. In the following, we introduce more refined transformations that can actually produce injective and inverse versions of the original functions. 5.1. Injectivization In principle, given a function f, one can consider that the injectivization of a rule of the form11 β : f(s0 ) → r ⇐ f1 (s1 ) ։ t1 , . . . , fn (sn ) ։ tn produces the following rule f i (s0 ) → hr, β(y, wn )i ⇐ f1i (s1 ) ։ ht1 , w1 i . . . , fni (sn ) ։ htn , wn i S where {y} = (Var(l)\Var(r, sn , tn )) ∪ ni=1 Var(ti )\Var(r, si+1,n ) and wn are fresh variables. The following example, though, illustrates that this is not correct in general. 11 By abuse of notation, here we let s0 , . . . , sn denote sequences of terms of arbitrary length, i.e., s0 = s0,1 , . . . , s0,l0 , s1 = s1,1 , . . . , s1,l1 , etc. 26 Example 30. Consider the following pcDCTRS R: β1 : f(x, y) → z ⇐ h(y) ։ w, first(x, w) ։ z β2 : h(0) → 0 β3 : first(x, y) → x together with the following top reduction: ǫ f(2, 1) →R 2 with σ = {x 7→ 2, y 7→ 1, w 7→ h(1), z 7→ 2} ǫ where h(y)σ = h(1) →∗R h(1) = wσ ǫ and first(x, w)σ = first(2, h(1)) →R 2 = zσ Following the scheme above, we would produce the following pcDCTRS f i (x, y) → hz, β1 (w1 , w2 )i ⇐ hi (y) ։ hw, w1i, firsti (x, w) ։ hz, w2 i hi (0) → h0, β2 i i first (x, y) → hx, β3 (y)i Unfortunately, the corresponding reduction for f i (2, 1) above cannot be done in this system since hi (1) cannot be reduced to hhi (1), [ ]i. In order to overcome this drawback, one could complete the function definitions with rules that reduce each irreducible term t to a tuple of the form ht, [ ]i. Although we find it a promising idea for future work, in this paper we propose a simpler approach. In the following, we consider a refinement of innermost reduction where only constructor substitutions are c computed. Formally, the constructor reduction relation, →, is defined as c follows: given ground terms s, t ∈ T (F ), we have s →R t iff there exist a position p in s such that no proper subterms of s|p are reducible, a rewrite rule l → r ⇐ sn ։ tn ∈ R, and a ground constructor substitution σ such c that s|p = lσ, si σ →∗R ti σ for all i = 1, . . . , n, and t = s[rσ]p . Note that the c results in the previous section also hold for →. In the following, given a basic term t = f(s), we denote by ti the term i f (s). Now, we introduce our injectivization transformation as follows: Definition 31 (injectivization). Let R be a pcDCTRS. We produce a new CTRS I(R) by replacing each rule β : l → r ⇐ sn ։ tn of R by a new rule of the form li → hr, β(y, wn )i ⇐ sin ։ htn , wn i S in I(R), where {y} = (Var(l)\Var(r, sn , tn )) ∪ ni=1 Var(ti )\Var(r, si+1,n ) and wn are fresh variables. Here, we assume that the variables of y are in lexicographic order. 27 Observe that now we do not need to keep a trace in each term, but only a single trace term since all reductions finish in one step in a pcDCTRS. The relation between the original trace terms and the information stored in the injectivized system is formalized as follows: Definition 32. Given a trace term π = β({ym 7→ tm }, π1 , . . . , πn ), we define π b recursively as follows: π b = β(tm , πb1 , . . . , π cn ), where we assume that the variables ym are in lexicographic order. Moreover, in order to simplify the notation, we consider that a a trace term π and a singleton list of the form [π] denote the same object. The correctness of the injectivization transformation is stated as follows: Theorem 33. Let R be a pcDCTRS and Rf = I(R) be its injectivization. c Then Rf is a pcDCTRS and, given a basic ground term s, we have hs, [ ]i ⇀R c ht, πi iff si →Rf ht, π bi. Proof. The fact that Rf is a pcDCTRS is trivial. Regarding the second part, we proceed as follows: c (⇒) We proceed by induction on the depth k of the step hs, [ ]i ⇀Rk ht, πi. Since the depth k = 0 is trivial, we consider the inductive case k > 0. Thus, there is a rule β : l → r ⇐ sn ։ tn ∈ R, and a subc stitution σ such that s = lσ, hsi σ, [ ]i⇀Rki hti σ, πi i, i = 1, . . . , n, t = rσ, σ ′ = σ |`(Var(l)\Var(r,sn ,tn ))∪Sn Var(ti )\Var(r,si+1,n ) , and π = β(σ ′ , π1 , . . . , πn ). By i=1 definition of ⇀Rk , we have that ki < k for all i = 1, . . . , n and, thus, by the inc duction hypothesis, we have (si σ)i →Rf hti σ, π bi i for all i = 1, . . . , n. Consider i now the equivalent rule in Rf : l → hr, β(y, wn )i ⇐ si1 ։ ht1 , w1 i, . . . , sin ։ i c htn , wn i. Therefore, we b1 , . . . , π bn )i where {y} = Snhave s →Rf ht, β(yσ, π (Var(l)\Var(r, sn , tn )) ∪ i=1 Var(ti )\Var(r, si+1,n ) and, thus, we can conclude that π b = β(yσ, π b1 , . . . , π bn ). (⇐) This direction is analogous. We proceed by induction on the depth c bi. Since the depth k = 0 is trivial, we consider k of the step si →Rfk ht, π the inductive case k > 0. Thus, there is a rule li → hr, β(y, wn )i ⇐ si1 ։ ht1 , w1 i, . . . , sin ։ htn , wn i in Rf and a substitution θ such that li θ = si , c bi. Assume that σ sii θ →Rfk hti , wi iθ, i = 1, . . . , n, and hr, β(y, wn )iθ = ht, π i is the restriction of θ to the variables of the rule, excluding the fresh variables wn , and that wi θ = π bi for all i = 1, . . . , n. Therefore, hsi , [ ]iθ = hsi σ, [ ]i and hti , wi iθ = hti σ, π bi i, i = 1, . . . , n. Then, by definition of Rfki , we have that ki < k for all i = 1, . . . , n and, thus, by the induction hypothesis, c we have hsi σ, [ ]i⇀R hti σ, πi i, i = 1, . . . , n. Consider now the equivalent rule c in R: β : l → r ⇐ sn ։ tn ∈ R. Therefore, we have hs, [ ]i ⇀R ht, πi, 28 σ ′ = σ |`(Var(l)\Var(r,sn ,tn ))∪Sn Var(ti )\Var(r,si+1,n ) , and π = β(σ ′ , π1 , . . . , πn ). Fii=1 S nally, since {y} = (Var(l)\Var(r, sn , tn )) ∪ ni=1 Var(ti )\Var(r, si+1,n ), we can conclude that π b = π. ✷ 5.2. Inversion Given an injectivized system, inversion basically amounts to switching the left- and right-hand sides of the rule and of every equation in the condition, as follows: Definition 34 (inversion). Let R be a pcDCTRS and Rf = I(R) be its injectivization. The inverse system Rb = I−1 (Rf ) is obtained from Rf by replacing each rule12 f i (s0 ) → hr, β(y, wn )i ⇐ f1i (s1 ) ։ ht1 , w1i, . . . , fni (sn ) ։ htn , wn i of Rf by a new rule of the form f −1 (r, β(y, wn )) → hs0 i ⇐ fn−1 (tn , wn ) ։ hsn i, . . . , f1−1 (t1 , w1 ) ։ hs1 i in I−1 (Rf ), where the variables of y are in lexicographic order. Example 35. Consider again the pcDCTRS of Example 16. Here, injectivization returns the following pcDCTRS I(R) = Rf : f i (x, y, m) → hs(w), β1(m, x, w1 , w2 )i ⇐ hi (x) ։ hx, w1 i, gi (y, 4) ։ hw, w2i hi (0) → h0, β2 i hi (1) → h1, β3 i gi (x, y) → hx, β4 (y)i Then, inversion with I−1 produces the following pcDCTRS I−1 (I(R)) = Rb : f −1 (s(w), β1 (m, x, w1 , w2 )) → hx, y, mi ⇐ g−1 (w, w2) ։ hy, 4i, h−1(x, w1 ) ։ hxi −1 h (0, β2 ) → h0i h−1 (1, β3 ) → h1i −1 g (x, β4 (y)) → hx, yi Finally, the correctness of the inversion transformation is stated as follows: 12 Here, we assume that s0 , s1 ,. . . , sn denote arbitrary sequences of terms, i.e., s0 = s0,1 , . . . , s0,l0 , s1 = s1,1 , . . . , s1,l1 , etc. We use this notation for clarity. 29 Theorem 36. Let R be a pcDCTRS, Rf = I(R) its injectivization, and Rb = I−1 (Rf ) the inversion of Rf . Then, Rb is a basic pcDCTRS and, given a basic ground term f(s) and a constructor ground term t with ht, πi a c c b) →Rb hsi. safe pair, we have ht, πi ↽R hf(s), [ ]i iff f −1 (t, π Proof. The fact that Rf is a pcDCTRS is trivial. Regarding the second part, we proceed as follows. c (⇒) We proceed by induction on the depth k of the step ht, πi ↽Rk hf(s), [ ]i. Since the depth k = 0 is trivial, we consider the inductive case k > 0. Let π = β(σ ′ , πn ). Thus, we have that ht, β(σ ′ , πn )i is a safe pair, there is a rule β : f(s0 ) → r ⇐ f1 (s1 ) ։ t1 , . . . , fn (sn ) ։ tn and substitution θ with Dom(θ) = (Var(r, s1 , . . . , sn )\Dom(σ ′ )) such that t = rθ, c hti θσ ′ , πi i→Rki hf(si )θσ ′ , [ ]i for all i = 1, . . . , n, and f(s) = f(s0 )θσ ′ . Note that s0 , . . . , sn denote sequences of terms of arbitrary length, i.e., s0 = s0,1 , . . . , s0,l0 , s1 = s1,1 , . . . , s1,l1 , etc. Since ht, Sπi is a safe pair, we have that Dom(σ ′ ) = (Var(s0 )\Var(r, s1 , . . . , sn , tn )) ∪ ni=1 Var(ti )\Var(r, si+1 , . . . , sn ). By definition of ↽Rk , we have that ki < k for all i = 1, . . . , n and, by the c induction hypothesis, we have f −1 (ti σ, πbi )→Rb hsi σi for all i = 1, . . . , n. Let us now consider the equivalent rule in Rb : f −1 (r, β(y, wn ))) → hs0 i ⇐ fn−1 (tn , wn ) ։ hsn i, . . . , f1−1 (t1 , w1) ։ hs1 i Hence, we have f −1 (t, β(yσ, π b1 , . . . , π b1 )) →Rb hs0 σi = hsi, where {y} = (Var(s0 )\Var(r, s1 , . . . , sn , tn )) ∪ n [ Var(ti )\Var(r, si+1 , . . . , sn ) i=1 and, thus, we can conclude that π b = β(yσ, π b1 , . . . , π bn ). (⇐) This direction is analogous. We proceed by induction on the depth c k of the step f −1 (t, π b) →Rbk hsi. Since the depth k = 0 is trivial, we consider the inductive case k > 0. Thus, there is a rule f −1 (r, β(y, wn ))) → hs0 i ⇐ fn−1 (tn , wn ) ։ hsn i, . . . , f1−1 (t1 , w1 ) ։ hs1 i in Rb and a substitution θ c such that f −1 (r, β(y, wn ))θ = f −1 (t, π b), fi−1 (ti , wi )θ →Rbk hsi iθ, i = n, . . . , 1, i and f −1 (r, ws)θ = hsi. Assume that σ is the restriction of θ to the variables of the rule, excluding the fresh variables wn , and that wi θ = π bi for −1 −1 all i = 1, . . . , n. Therefore, f (r, β(y, wn ))θ = f (rσ, β(yσ, π b1 , . . . , π bn ), −1 −1 fi (ti , wi )θ = fi (ti σ, π bi ) and hsi iθ = hsi σi, i = 1, . . . , n. Then, by definition of Rbki , we have that ki < k for all i = 1, . . . , n and, thus, by the c induction hypothesis, we have hti σ, πi i↽R hfi (si σ), [ ]i, i = 1, . . . , n. Consider now the equivalent rule in R: β : f(s0 ) → r ⇐ f1 (s1 ) ։ t1 , . . . , fn (sn ) ։ tn c in R. Therefore, we have ht, πi ↽R hf(s), [ ]i, σ ′ = σ|`(Var(s0 )\Var(r,s1 ,...,sn ,tn ))∪Sn i=1 30 Var(ti )\Var(r,si+1 ,...,sn ) and π = β(σ ′ , π1 , . . . , πn ). Finally, since {y} = (Var(s0 )\Var(r, s1 , . . . , sn , tn ))∪ S n b = π. ✷ i=1 Var(ti )\Var(r, si+1 , . . . , sn ), we can conclude that π 5.3. Improving the transformation for injective functions When a function is injective, one can expect the injectivization transformation to be unnecessary. This is not generally true, since some additional syntactic conditions might also be required. Furthermore, depending on the considered setting, it can be necessary to have an injective system, rather than an injective function. Consider, e.g., the following simple TRS: R = { f1 → f2 , f2 → 0, g1 → g2 , g2 → 0 } Here, all functions are clearly injective. However, given a reduction like f1 →R f2 →R 0, we do not know which rule should be applied to 0 in order to go backwards until the initial term (actually, both the second and the fourth rules are applicable in the reverse direction). Luckily, in our context, the injectivity of a function suffices since reductions in pcDCTRSs are performed in a single step. Therefore, given a reduction of the form f i (sn ) →R t, a backward computation will have the form f −1 (t) →R hsn i, so that we know that only the inverse rules of f are applicable. Now, we present an improvement of the injectivization transformation presented in Section 5.1 which has some similarities with that in [24]. Here, we consider that the initial system is a TRS R since, to the best of our knowledge, there is no reachability analysis defined for DCTRSs. In the following, given a term s, we let range(s) = {t | sσ →∗R t, σ : V 7→ T (C), and t ∈ T (C)} i.e., range(s) returns a set with the constructor normal forms of all possible ground constructor instances of s. Although computing this set is generally undecidable, there are some overapproximations based on the use of tree automata (see, e.g., [15] and the most recent approach for innermost rewriting [16]). Let us consider that rangeα (s) is such an approximation, with rangeα (s) ⊇ range(s) for all terms s. Here, we are interested in determining when the right-hand sides, r1 and r2 , of two rules do not overlap, i.e., range(r1 ) ∩ range(r2 ) = ∅. For this purpose, we will check whether rangeα (r1 ) ∩ rangeα (r2 ) = ∅. Since finite tree automata are closed under intersection and the emptiness of a finite tree automata is decidable, checking the emptiness of rangeα (r1 )∩rangeα (r2 ) is decidable and can be used to safely identify non-overlapping right-hand sides, i.e., if rangeα (r1 ) ∩ rangeα (r2 ) = ∅, 31 then r1 and r2 are definitely non-overlapping; otherwise, they may be overlapping or non-overlapping. Now, we summarize our method to simplify some trace terms. Given a constructor TRS R and a rule β : l → r ∈ R, we check the following conditions: 1. the right-hand side r of the rule does not overlap with the right-hand side of any other rule defining the same function; 2. the rule is non-erasing, i.e., Var(l) = Var(r); 3. the right-hand side r contains a single occurrence of a defined function symbol, say f ∈ D. If these conditions hold, then the rule has the form l → r[f(s)]p with l and f(s) basic terms,13 and r[x]p and s constructor terms, where x is a fresh variable. In this case, we can safely produce the following injective version:14 li → hr[x]p , wi ⇐ f i (s) ։ hx, wi instead of li → hr[x]p , β(w)i ⇐ f i (s) ։ hx, wi Let us illustrate this improved transformation with a couple of examples. Example 37. Consider the following TRS: R = { f(s(x)) → g(x), f(c(x)) → h(x), g(x) → s(x), h(x) → c(x)} Here, it can easily be shown that rangeα (g(x)) ∩ rangeα (h(x)) = ∅, the two rules defining f are non-erasing, and both contain a single occurrence of a defined function symbol in the righ-hand sides. Therefore, our improved injectivization applies and we get the following pcDCTRS Rf : f i (s(x)) → hy, wi ⇐ gi (x) ։ hy, wi f i (c(x)) → hy, wi ⇐ hi (x) ։ hy, wi gi (x) → hs(x), β3 i hi (x) → hc(x), β4 i In contrast, the original injectivization transformation would return the following system: f i (s(x)) → hy, β1(w)i ⇐ gi (x) ։ hy, wi f i (c(x)) → hy, β2(w)i ⇐ hi (x) ։ hy, wi 13 gi (x) → hs(x), β3 i hi (x) → hc(x), β4 i Note that l is a basic term since we initially consider a constructor TRS and, thus, all left-hand sides are basic terms by definition. 14 Since l → r is non-erasing, the pcDCTRS rule l → r[x]p ⇐ f(s) ։ x is trivially nonerasing too (according to [32], i.e., (Var(l)\Var(r[x]p , f(s), x)) ∪ Var(x)\Var(r[x]p ) = ∅) and, thus, no binding should be stored during the injectivization process. 32 Finally, the inverse system Rb obtained from Rf using the original transformation has the following form: f −1 (y, w) → hs(x)i ⇐ g−1 (y, w) ։ hxi f −1 (y, w) → hc(x)i ⇐ h−1 (y, w) ։ hxi g−1 (s(x), β3 ) → hxi h−1 (c(x), β4 ) → hxi For instance, given the forward reduction f i (s(0)) →Rf hs(0), β3 i, we can build the corresponding backward reduction: f −1 (s(0), β3 ) →Rb hs(0)i. Note, however, that the left-hand sides of f −1 overlap and we should reduce the conditions in order to determine which rule to apply. Therefore, in some cases, there is a trade-off between the size of the trace terms and the complexity of the reduction steps. The example above, though, only produces a rather limited improvement since the considered functions are not recursive. Our next example shows a much significant improvement. Here, we consider the function zip (also used in [24] to illustrate the benefits of an injectivity analysis). Example 38. Consider the following TRS R defining the function zip: zip([ ], ys) → [ ] zip(xs, [ ]) → [ ] zip(x : xs, y : ys) → pair(x, y) : zip(xs, ys) Here, since the third rule is non-erasing, its right-hand side contains a single occurrence of a defined function, zip, and it does not overlap with any other right-hand side, our improved injectivization applies and we get the following pcDCTRS Rf : zipi ([ ], ys) → h[ ], β1(ys)i zipi (xs, [ ]) → h[ ], β2(xs)i zipi (x : xs, y : ys) → hpair(x, y) : zs, wi ⇐ zipi (xs, ys) ։ hzs, wi In contrast, the original injectivization transformation would return the following system R′f : zipi ([ ], ys) → h[ ], β1 (ys)i zipi (xs, [ ]) → h[ ], β2 (xs)i zipi (x : xs, y : ys) → hpair(x, y) : zs, β3 (w)i ⇐ zipi (xs, ys) ։ hzs, wi It might seem a small difference, but if we call zipi with two lists of n elements, the system R′f would build a trace term of the form β3 (. . . β3 (β1 (. . .)) . . .) with n nested constructors β3 , while Rf would just build the trace term β1 (. . .). For large values of n, this is a significant improvement in memory usage. 33 6. Bidirectional Program Transformation We illustrate a practical application of our reversibilization technique in the context of bidirectional program transformation (see [10] for a survey). In particular, we consider the so-called view-update problem. Here, we have a data structure (e.g., a database) called the source, which is transformed to another data structure, called the view. Typically, we have a view function, view: Source → View that takes the source and returns the corresponding view, together with an update function, upd: View × Source → Source that propagates the changes in a modified view to the original source. Two basic properties that these functions should satisfy in order to be well-behaved are the following [13]: ∀s ∈ Source, ∀v ∈ View : view(upd(v, s)) = v ∀s ∈ Source: upd(view(s), s) = s Bidirectionalization (first proposed in the database community [5]) basically consists in, given a view function, “bidirectionalize” it in order to derive an appropriate update function. For this purpose, first, a view complement function is usually defined, say viewc , so that the tupled function view △ viewc: Source → View × Comp becomes injective. Therefore, the update function can be defined as follows: upd(v, s) = (view △ viewc )−1 (v, viewc (s)) This approach has been applied to bidirectionalize view functions in a functional language in [24]. In the following, we apply our injectivization and inversion transformations in order to produce a bidirectionalization transformation that may be useful in the context of the view-update problem (with some limitations). Let us assume that we have a view function, view, that takes a source and returns the corresponding view, and which is defined by means of a pcDCTRS. Following our approach, given the original program R, we produce an injectivized version Rf and the corresponding inverse Rb . Therefore, in principle, one can use Rf ∪ Rb , which will include the functions viewi and view−1, to define an update function as follows: upd(v, s) → s′ ⇐ viewi (s) ։ hv ′ , πi, view−1(v, π) ։ hs′ i where s is the original source, v is the updated view, and s′ , the returned value, is the corresponding updated source. Note that, in our context, the function viewi is somehow equivalent to view △ viewc above. 34 Let us now illustrate the bidirectionalization process with an example. Consider a particular data structure, a list of records of the form r(t, v) where t is the type of the record (e.g., book, dvd, pen, etc.) and v is its price tag. The following system defines a view function that takes a type and a list of records, and returns a list with the price tags of the records of the given type:15 view(t, nil) view(t, r(t′ , v) : rs) view(t, r(t′ , v) : rs) eq(book, book) eq(book, dvd) val(r(t, v)) → → → → → → nil val(r(t′ , v)) : view(t, rs) ⇐ eq(t, t′ ) ։ true view(t, rs) ⇐ eq(t, t′ ) ։ false true eq(dvd, dvd) → true false eq(dvd, book) → false v However, this system is not a pcDCTRS. Here, we use a flattening transformation to produce the following (labeled) pcDCTRS R which is equivalent for constructor derivations: β1 : view(t, nil) → nil β2 : view(t, r(t′ , v) : rs) → p : r ⇐ eq(t, t′ ) ։ true, val(r(t′ , v)) ։ p, view(t, rs) ։ r ′ β3 : view(t, r(t , v) : rs) → r ⇐ eq(t, t′ ) ։ false, view(t, rs) ։ r β4 : β6 : eq(book, book) → true eq(book, dvd) → false β8 : val(r(t, v)) → v β5 : eq(dvd, dvd) → true β7 : eq(dvd, book) → false Now, we can apply our injectivization transformation which returns the following pcDCTRS Rf = I(R): viewi (t, nil) → hnil, β1 (t)i viewi (t, r(t′ , v) : rs) → hp : r, β2 (w1 , w2 , w3 )i ⇐ eqi (t, t′ ) ։ htrue, w1 i, vali (r(t′ , v)) ։ hp, w2i, viewi (t, rs) ։ hr, w3 i viewi (t, r(t′ , v) : rs) → hr, β3 (v, w1, w2 )i ⇐ eqi (t, t′ ) ։ hfalse, w1 i, viewi (t, rs) ։ hr, w2i eqi (book, book) → htrue, β4 i eqi (book, dvd) → hfalse, β6 i eqi (dvd, dvd) → htrue, β5 i eqi (dvd, book) → hfalse, β7 i vali (r(t, v)) → hv, β8(t)i 15 For simplicity, we restrict the record types to only book and dvd. 35 Finally, inversion returns the following pcDCTRS Rb = I(Rf ): view−1(nil, β1 (t)) → ht, nili view (p : r, β2 (w1 , w2, w3 )) → ht, r(t′ , v) : rsi ⇐ eq−1(true, w1 ) ։ ht, t′ i, val−1(p, w2 ) ։ hr(t′ , v)i, view−1(r, w3 ) ։ ht, rsi view−1(r, β3 (v, w1, w2 )) → ht, r(t′ , v) : rsi ⇐ eq−1(false, w1 ) ։ ht, t′ i, view−1(r, w2 ) ։ ht, rsi −1 eq−1(true, β4 ) → hbook, booki eq−1(true, β5 ) → hdvd, dvdi eq−1 (false, β6 ) → hbook, dvdi eq−1(false, β7 ) → hdvd, booki val−1(v, β8 (t)) → hr(t, v)i For instance, the term view(book, [r(book, 12), r(dvd, 24)]), reduces to [12] in the original system R. Given a modified view, e.g., [15], we can compute the modified source using function upd above: upd([r(book, 12), r(dvd, 24)], [15]) Here, we have the following subcomputations:16 viewi (book, [r(book, 12), r(dvd, 24)]) →Rf h[12], β2 (β4 , β8 (book), β3 (24, β6 , β1 (book)))i view−1([15], β2 (β4 , β8 (book), β3 (24, β6 , β1 (book)))) →Rb hbook, [r(book, 15), r(dvd, 24)]i Thus upd returns the updated source [r(book, 15), r(dvd, 24)], as expected. We note that the considered example cannot be transformed using the technique in [24], the closer to our approach, since the right-hand sides of some rules contain functions which are not treeless.17 Nevertheless, one could consider a transformation from pcDCTRS to functional programs with treeless functions so that the technique in [24] becomes applicable. Our approach can solve a view-update problem as long as the view function can be encoded in a pcDCTRS. When this is the case, the results from Section 5 guarantee that function upd is well defined. Formally analyzing the class of view functions that can be represented with a pcDCTRS is an interesting topic for further research. 7. Related Work There is no widely accepted notion of reversible computing. In this work, we have considered one of its most popular definitions, according to which a 16 Note that, in this case, the function view requires not only the source but also the additional parameter book. 17 A call is treeless if it has the form f(x1 , . . . , xn ) and x1 , . . . , xn are different variables. 36 computation principle is reversible if there is a method to undo a (forward) computation. Moreover, we expect to get back to an exact past state of the computation. This is often referred to as full reversibility. As we have mentioned in the introduction, some of the most promising applications of reversibility include cellular automata [28], bidirectional program transformation [24], already discussed in Section 6, reversible debugging [17], where the ability to go both forward and backward when seeking the cause of an error can be very useful for the programmer, parallel discrete event simulation [34], where reversibility is used to undo the effects of speculative computations made on a wrong assumption, quantum computing [39], where all computations should be reversible, and so forth. The interested reader can find detailed surveys in the state of the art reports of the different working groups of COST Action IC1405 on Reversible Computation [20]. Intuitively speaking, there are two broad approaches to reversibility from a programming language perspective: Reversible programming languages. In this case, all constructs of the programming language are reversible. One of the most popular languages within the first approach is the reversible (imperative) language Janus [23]. The language was recently rediscovered [42, 41, 43] and has since been formalized and further developed. Irreversible programming languages and Landauer’s embedding. Alternatively, one can consider an irreversible programming language, and enhance the states with some additional information (typically, the history of the computation so far) so that computations become reversible. This is called Landauer’s embedding. In this work, we consider reversibility in the context of term rewriting. To the best of our knowledge, we have presented the first approach to reversibility in term rewriting. A closest approach was introduced by Abramsky in the context of pattern matching automata [2], though his developments could easily be applied to rewrite systems as well. In Abramsky’s approach, biorthogonality was required to ensure reversibility, which would be a very significant restriction for term rewriting systems. Basically, biorthogonality requires that, for every pair of (different) rewrite rules l → r and l′ → r ′ , l and l′ do not overlap (roughly, they do not unify) and r and r ′ do not overlap too. Trivially, the functions of a biorthogonal system are injective and, thus, computations are reversible without the need of a Landauer embedding. Therefore, Abramsky’s work is aimed at defining a reversible language, in contrast to our approach that is based on defining a Landauer embedding for standard term rewriting and a general class of rewrite systems. 37 Defining a Landauer embedding in order to make a computation mechanism reversible has been applied in different contexts and computational models, e.g., a probabilistic guarded command language [44], a low level virtual machine [35], the call-by-name lambda calculus [19, 21], cellular automata [38, 27], combinatory logic [11], a flowchart language [41], etc. In the context of declarative languages, we find the work by Mu et al. [29], where a relational reversible language is presented (in the context of bidirectional programming). A similar approach was then introduced by Matsuda et al. [24, 25] in the context of functional programs and bidirectional transformation. The functional programs considered in [24] can be seen as linear and right-treeless 18 constructor TRSs. The class of functional programs is more general in [25], which would correspond to left-linear, right-treeless TRSs. The reversibilization technique of [24, 25] includes both an injectivization stage (by introducing a view complement function) and an inversion stage. These methods are closely related to the transformations of injectivization and inversion that we have presented in Section 5, although we developed them from a rather different starting point. Moreover, their methods for injectivization and inversion consider a more restricted class of systems than those considered in this paper. On the other hand, they apply a number of analyses to improve the result, which explains the smaller traces in their approach. All in all, we consider that our approach gives better insights to understand the need for some of the requirements of the program transformations and the class of considered programs. For instance, most of our requirements come from the need to remove programs positions from the traces, as shown in Section 4. Finally, [37] considers the reversible language RFUN. Similarly to Janus, computations in RFUN are reversible without the need of a Landauer embedding. The paper also presents a transformation from a simple (irreversible) functional language, FUN, to RFUN, in order to highlight how irreversibilities are handled in RFUN. The transformation has some similarities with both the approach of [24] and our improved transformation in Section 5.3; on the other hand, though, [37] also applies the Bennett trick [6] in order to avoid some unnecessary information. 18 There are no nested defined symbols in the right-hand sides, and, moreover, any term rooted by a defined function in the right-hand sides can only take different variables as its proper subterms. 38 8. Discussion and Future Work In this paper, we have introduced a reversible extension of term rewriting. In order to keep our approach as general as possible, we have initially considered DCTRSs as input systems, and proved the soundness and reversibility of our extension of rewriting. Then, in order to introduce a reversibilization transformation for these systems, we have also presented a transformation from DCTRSs to pure constructor systems (pcDCTRSs) which is correct for constructor reduction. A further improvement is presented for injective functions, which may have a significant impact in memory usage in some cases. Finally, we have successfully applied our approach in the context of bidirectional program transformation. We have developed a prototype implementation of the reversibilization transformations introduced in Section 5. The tool can read an input TRS file (format .trs [1]) and then it applies in a sequential way the following transformations: flattening, simplification of constructor conditions, injectivization, and inversion. The tool prints out the CTRSs obtained at each transformation step. It is publicly available through a web interface from http://kaz.dsic.upv.es/rev-rewriting.html, where we have included a number of examples to easily test the tool. As for future work, we plan to investigate new methods to further reduce the size of the traces. In particular, we find it interesting to define a reachability analysis for DCTRSs. A reachability analysis for CTRSs without extra-variables (1-CTRSs) can be found in [12], but the extension to deal with extra-variables in DCTRSs (since a DCTRS is a particular case of 3-CTRS) seems challenging. Furthermore, as mentioned in the paper, a completion procedure to add default cases to some functions (as suggested in Section 5.1) may help to broaden the applicability of the technique and avoid the restriction to constructor reduction. Finally, our injectivization and inversion transformations are correct w.r.t. innermost reduction. Extending our results to a lazy strategy is also an interesting topic for further research. Acknowledgements We thank the anonymous reviewers for their useful comments and suggestions to improve this paper. References [1] Annual international termination competition. Available from URL: http://www.termination-portal.org/wiki/Termination Competition. 39 [2] S. Abramsky. A structural approach to reversible computation. Theoretical Computer Science, 347(3):441–464, 2005. [3] J. M. Almendros-Jiménez and G. Vidal. Automatic partial inversion of inductively sequential functions. In Z. Horváth, V. Zsók, and A. Butterfield, editors, Implementation and Application of Functional Languages, 18th International Symposium (IFL 2006), Revised Selected Papers, volume 4449 of Lecture Notes in Computer Science, pages 253–270. Springer, 2007. [4] F. Baader and T. Nipkow. Term Rewriting and All That. Cambridge University Press, 1998. [5] F. Bancilhon and N. Spyratos. Update semantics of relational views. ACM Transactions on Database Systems, 6(4):557–575, 1981. [6] C. H. Bennett. Logical reversibility of computation. IBM Journal of Research and Development, 17:525–532, 1973. [7] C. H. Bennett. Notes on the history of reversible computation. IBM Journal of Research and Development, 44(1):270–278, 2000. [8] J. Bergstra and J. Klop. Conditional Rewrite Rules: confluence and termination. Journal of Computer and System Sciences, 32:323–362, 1986. [9] P. Crescenzi and C. H. Papadimitriou. Reversible simulation of spacebounded computations. Theoretical Computer Science, 143(1):159–165, 1995. [10] K. Czarnecki, J. N. Foster, Z. Hu, R. Lämmel, A. Schürr, and J. F. Terwilliger. Bidirectional transformations: A cross-discipline perspective. In R. F. Paige, editor, Proc. of the 2nd Int’l Conf. on Theory and Practice of Model Transformations (ICMT 2009), volume 5563 of Lecture Notes in Computer Science, pages 260–283. Springer, 2009. [11] A. Di Pierro, C. Hankin, and H. Wiklicky. Reversible combinatory logic. Mathematical Structures in Computer Science, 16(4):621–637, 2006. [12] G. Feuillade and T. Genet. Reachability in Conditional Term Rewriting Systems. Electronic Notes in Theoretical Computer Science, 86(1):133–146, 2003. [13] J. N. Foster, M. B. Greenwald, J. T. Moore, B. C. Pierce, and A. Schmitt. Combinators for bidirectional tree transformations: A linguistic approach to the view-update problem. ACM Transactions on Programming Languages and Systems, 29(3):17, 2007. [14] M. P. Frank. Introduction to reversible computing: motivation, progress, and challenges. In N. Bagherzadeh, M. Valero, and A. Ramı́rez, editors, Proceedings of the Second Conference on Computing Frontiers, pages 385– 390. ACM, 2005. 40 [15] T. Genet. Decidable Approximations of Sets of Descendants and Sets of Normal Forms. In T. Nipkow, editor, Proc. of the 9th International Conference on Rewriting Techniques and Applications (RTA’98), volume 1379 of Lecture Notes in Computer Science, pages 151–165. Springer, 1998. [16] T. Genet and Y. Salmon. Reachability Analysis of Innermost Rewriting. In M. Fernández, editor, Proc. of the 26th International Conference on Rewriting Techniques and Applications (RTA’15), volume 36 of LIPIcs, pages 177–193. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2015. [17] E. Giachino, I. Lanese, and C. A. Mezzina. Causal-consistent reversible debugging. In S. Gnesi and A. Rensink, editors, Proc. of the 17th International Conference on Fundamental Approaches to Software Engineering (FASE 2014), volume 8411 of Lecture Notes in Computer Science, pages 370– 384. Springer, 2014. [18] N. Hirokawa and G. Moser. Automated Complexity Analysis Based on the Dependency Pair Method. In A. Armando, P. Baumgartner, and G. Dowek, editors, Proc. of IJCAR 2008, volume 5195 of Lecture Notes in Computer Science, pages 364–379. Springer, 2008. [19] L. Huelsbergen. A logically reversible evaluator for the call-by-name lambda calculus. In T. Toffoli and M. Biafore, editors, Proc. of PhysComp96, pages 159–167. New England Complex Systems Institute, 1996. [20] COST Action IC1405 on Reversible Computation - extending horizons of computing. URL: http://revcomp.eu/. [21] W. E. Kluge. A reversible SE(M)CD machine. In P. W. M. Koopman and C. Clack, editors, Proc. of the 11th International Workshop on the Implementation of Functional Languages, IFL’99. Selected Papers, volume 1868 of Lecture Notes in Computer Science, pages 95–113. Springer, 2000. [22] R. Landauer. Irreversibility and heat generation in the computing process. IBM Journal of Research and Development, 5:183–191, 1961. [23] C. Lutz and H. Derby. Janus: A time-reversible language, 1986. A letter to R. Landauer. Available from URL http://tetsuo.jp/ref/janus.pdf. [24] K. Matsuda, Z. Hu, K. Nakano, M. Hamana, and M. Takeichi. Bidirectionalization transformation based on automatic derivation of view complement functions. In R. Hinze and N. Ramsey, editors, Proc. of the 12th ACM SIGPLAN International Conference on Functional Programming, ICFP 2007, pages 47–58. ACM, 2007. 41 [25] K. Matsuda, Z. Hu, K. Nakano, M. Hamana, and M. Takeichi. Bidirectionalizing programs with duplication through complementary function derivation. Computer Software, 26(2):56–75, 2009. In Japanese. [26] A. Middeldorp and E. Hamoen. Completeness results for basic narrowing. Applicable Algebra in Engineering, Communication and Computing, 5:213– 253, 1994. [27] K. Morita. Reversible simulation of one-dimensional irreversible cellular automata. Theoretical Computer Science, 148(1):157–163, 1995. [28] K. Morita. Computation in reversible cellular automata. International Journal of General Systems, 41(6):569–581, 2012. [29] S. Mu, Z. Hu, and M. Takeichi. An injective language for reversible computation. In D. Kozen and C. Shankland, editors, Proc. of the 7th International Conference on Mathematics of Program Construction (MPC 2004), volume 3125 of Lecture Notes in Computer Science, pages 289–313. Springer, 2004. [30] M. Nagashima, M. Sakai, and T. Sakabe. Determinization of conditional term rewriting systems. Theoretical Computer Science, 464:72–89, 2012. [31] N. Nishida, A. Palacios, and G. Vidal. Reversible term rewriting. In D. Kesner and B. Pientka, editors, Proc. of the 1st International Conference on Formal Structures for Computation and Deduction (FSCD’16), volume 52 of LIPIcs, pages 28:1–28:18. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2016. [32] N. Nishida, M. Sakai, and T. Sakabe. Soundness of unravelings for conditional term rewriting systems via ultra-properties related to linearity. Logical Methods in Computer Science, 8(3-4):1–49, Aug. 2012. [33] N. Nishida and G. Vidal. Program inversion for tail recursive functions. In M. Schmidt-Schauß, editor, Proceedings of the 22nd International Conference on Rewriting Techniques and Applications (RTA 2011), volume 10 of LIPIcs, pages 283–298. Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2011. [34] M. Schordan, D. R. Jefferson, P. D. B. Jr., T. Oppelstrup, and D. J. Quinlan. Reverse code generation for parallel discrete event simulation. In J. Krivine and J. Stefani, editors, Proc. of the 7th International Conference on Reversible Computation (RC 2015), volume 9138 of Lecture Notes in Computer Science, pages 95–110. Springer, 2015. [35] B. Stoddart, R. Lynas, and F. Zeyda. A virtual machine for supporting reversible probabilistic guarded command languages. Electronic Notes in Theoretical Computer Science, 253(6):33–56, 2010. 42 [36] Terese. Term Rewriting Systems, volume 55 of Cambridge Tracts in Theoretical Computer Science. Cambridge University Press, 2003. [37] M. K. Thomsen and H. B. Axelsen. Interpretation and programming of the reversible functional language RFUN. In R. Lämmel, editor, Proc. of the 27th Symposium on the Implementation and Application of Functional Programming Languages (IFL’15), pages 8:1–8:13. ACM, 2015. [38] T. Toffoli. Computation and construction universality of reversible cellular automata. Journal of Computer and System Sciences, 15(2):213–231, 1977. [39] T. Yamakami. One-way reversible and quantum finite automata with advice. Information and Computation, 239:122–148, 2014. [40] T. Yokoyama. Reversible computation and reversible programming languages. Electronic Notes in Theoretical Computer Science, 253(6):71–81, 2010. [41] T. Yokoyama, H. Axelsen, and R. Glück. Fundamentals of reversible flowchart languages. Theoretical Computer Science, 611:87–115, 2016. [42] T. Yokoyama, H. B. Axelsen, and R. Glück. Reversible flowchart languages and the structured reversible program theorem. In Proc. of the 35th International Colloquium on Automata, Languages and Programming (ICALP 2008), volume 5126 of Lecture Notes in Computer Science, pages 258–270. Springer, 2008. [43] T. Yokoyama and R. Glück. A reversible programming language and its invertible self-interpreter. In G. Ramalingam and E. Visser, editors, Proc. of the 2007 ACM SIGPLAN Workshop on Partial Evaluation and Semanticsbased Program Manipulation (PEPM 2007), pages 144–153. ACM, 2007. [44] P. Zuliani. Logical reversibility. IBM Journal of Research and Development, 45(6):807–818, 2001. 43 

Transivity of Commutativity for Second-Order Linear Time-Varying Analog Systems Mehmet Emir KOKSAL Department of Mathematics, Ondokuz Mayis University, 55139 Atakum, Samsun, Turkey emir_koksal@hotmail.com Abstract: After reviewing commutativity of second-order linear time-varying analog systems, the inverse commutativity conditions are derived for these systems by considering non-zero initial conditions. On the base of these conditions, the transitivity property is studied for second order linear time-varying unrelaxed analog systems. It is proven that this property is always valid for such systems when their initial states are zero; when non-zero initial states are present, it is shown that the validity of transitivity does not require any more conditions and it is still valid. Throughout the study it is assumed that the subsystems considered can not be obtained from each other by any feed-forword and feed-back structure. The results are well validated by MATLAB simulations. Keywords: Differential equations, Initial conditions, Linear time-varying systems, Commutativity, Transitivity AMS Subject Classification: 93C05, 93C15, 93A30 I. Introduction Second-order differential equations originate in electromagnetic, electrodynamics, transmission lines and communication, circuit and system theory, wave motion and distribution, and in many fields of electrics-electronics engineering. They play a prominent 1 role for modelling problems occurring in electrical systems, fluid systems, thermal systems and control systems. Especially, they are used as a powerful tool for modelling, analyzing and solving problems in classical control theory, modern control theory, robust control theory and automatic control, which is essential in any field of engineering and sciences, and for discussing the results turned up at the end of analyzing for resolution of naturel problems. For example, they are used in cascade connected and feedback systems to design higher order composite systems for achieving several beneficial properties such as controllability, sensitivity, robustness, and design flexibility. When the cascade connection which is an old but still an up to date trend in system design [1-4] is considered, the commutativity concept places an important role to improve different system performances. On the other hand, since the commutativity of linear time-invariant relaxed systems is straightforward and time-varying systems have found a great deal of applications recently [5-10], the scope of this paper is focused on commutativity of linear time-varying systems only. When two systems 𝐴 and 𝐵 are are interconnected one after the other so that the output of the former acts as the input of the later, it is said that these systems are connected in cascade [11]. If the order of connection in the sequence is not effective on the inputoutput relation of the cascade connection, then these systems are commutative [12]. Figure 1: Cascade connection of the differential system 𝐴 and 𝐵 The tutorial papers [13, 14] cover almost all the subjects scattered in a great deal of 2 literature about commutativity. Some of the important results about the commutativity are summarized in the sequel superficially. J. E. Marshall has proven that “for commutativity, either both systems are timeinvariant or both systems are time-varying” [12]. After many contributions appeared as conference presentations and a few short papers focusing to special cases such as first, second, third, and forth order systems, the exhaustive journal paper of M. Koksal introduced the basic fundamentals of the subject [13]. Another work joint by the same author has presented explicit commutativity conditions of fifth order systems in addition to reviews of commutativity of systems with non-zero initial conditions, commutativity and system disturbance, commutativity of Euler systems [14]. There is some literature on the commutativity of discrete-time systems as well [15, 16]. And the research of commutativity is continuing on both analog and digital systems in [17, 18]. In [17], all the second-order commutative pairs of a first-order linear time-varying analogue systems are derived. The decomposition of a second-order linear time-varying systems into its first-order commutative pairs are studied in [18]. This is important for the cascade realization of the second-order linear time-varying systems. In [19], the inverse conditions expressed in terms of the coefficients of the differential equation describing system 𝑩 have been derived for the case of zero initial conditions and shown to be of the same form of the original equations appearing in the literature. Transitivity property of commutativity is first introduced and some general conditions for it are presented in [20], where transitivity of commutativity is fully investigated for first-order systems. It is shown that commutativity of first-order linear timevarying systems with and without initial conditions has always transitivity property. 3 Explicit commutativity conditions for second-order linear time-varying systems have been studied in [21]. On the other hand, no special research as in [20] for the firstorder systems has been done on the transitivity property of commutativity for second-order systems. This paper completes this vacancy in the literature on the base of work in [21]. In this paper, transtivity property of second-order linear time-varying analog systems with and without initial conditions is studied. Section II is devoted to the explicite commutativity conditions for such systems. Section III presents preliminaris, namely inverse commutativity conditions in the case of non-zero inititial states, that are used in the proof of the subseqient section. Section IV deals with transtivity property with and without initial conditions. After giving an example is Section V, the paper ends with conclusions which appears in Section VI. II. Commutativity Conditions for Second-order Systems In this section, the commutativity conditions for two second-order linear time-varying analog systems are reviewed [21]. Let 𝐴 be the systam described by the second-order linear time-varying differential equation 𝑎2 (𝑡)𝑦̈𝐴 (𝑡) + 𝑎1 (𝑡)𝑦̇𝐴 (𝑡) + 𝑎0 (𝑡)𝑦𝐴 (𝑡) = 𝑥𝐴 (𝑡); 𝑡 ≥ 𝑡0 (1a) with the initial conditions at the initial time 𝑡0 ∈ 𝑅 𝑦𝐴 (𝑡0 ), 𝑦̇𝐴 (𝑡0 ). (1b) Where the single (double) dot on the top indicates the first (second) order derivative with respect to time 𝑡 ∈ 𝑅; 𝑥𝐴 (𝑡) and 𝑦𝐴 (𝑡) are the input and output of the system, respectively. Since the system is second-order 𝑎2 (𝑡) ≢ 0. (1c) Further, 𝑎̈ 2 (𝑡), 𝑎̇ 1 (𝑡) and , 𝑎0 (𝑡) are well defined continuous functions, that is 𝑎2 (𝑡) , 𝑎̇ 1 (𝑡), 𝑎0 (𝑡) ∈ 𝐶[𝑡0 , ∞) , hence so 𝑎̇ 2 (𝑡), 𝑎2 (𝑡), 𝑎1 (𝑡) are. 4 It is true that System 𝐴 has a unique continuous solution 𝑦𝐴 (𝑡) with its first and secondorder derivatives for any continuous input function 𝑥𝐴 (𝑡) [22]. Let 𝐵 be another second-order linear time varying system described in a similar way to 𝐴. Hence it is described by 𝑏2 (𝑡)𝑦̈ 𝐵 (𝑡) + 𝑏1 (𝑡)𝑦̇ 𝐵 (𝑡) + 𝑏0 (𝑡)𝑦𝐵 (𝑡) = 𝑥𝐵 (𝑡), 𝑡 ≥ 𝑡0 , (2a) 𝑦𝐵 (𝑡0 ), 𝑦̇ 𝐵 (𝑡𝑜 ), (2b) 𝑏2 (𝑡) ≢ 0. (2c) Where the coefficients 𝑏2 , 𝑏1 , 𝑏0 satisfy the same properties satisfied by 𝑎2 , 𝑎1 , 𝑎0 . For the commutativity of 𝐵 with 𝐴, it is well known that I) i) The coefficients of 𝐵 be expressed in terms of those of 𝐴 through 𝑎2 (𝑡) 0 𝑏2 (𝑡) 0,5 [𝑏1 (𝑡)] = [𝑎1 (𝑡) 𝑎2 (𝑡) 𝑏0 (𝑡) 𝑎0 (𝑡) 𝑓𝐴 (𝑡) 0 𝑘2 0] [𝑘1 ], 1 𝑘0 (3a) . (3b) where 𝑓𝐴 (𝑡) = 𝑎2 −0,5 [2𝑎1 (𝑡)−𝑎̇ 2 (𝑡)] 4 And 𝑘2 , 𝑘1 , 𝑘0 are some constants with 𝑘2 ≠ 0 for 𝐵 is of second-order. Further, it is assumed that 𝐵 can not be obtained from 𝐴 by a constant feed forward and feedback path gains; hence 𝑘1 ≠ 0 [14, 21]. ii) 𝑎0 − 𝑓𝐴 2 𝑎2 −0,5 𝑓𝐴̇ = 𝐴0 , ∀𝑡 ≥ 𝑡0 , (3c) where 𝐴0 is a constant, When the initial conditions in (1b) and (2b) are zero, the conditions i) and ii) are necessary and sufficient conditions fort the commutativity of 𝐴 and 𝐵 under the mentioned conditions (𝑘2 ≠ 0, 𝑘1 ≠ 0). For commutativity with nonzero initial conditions as well, the following additional conditions are required: II) i) 5 𝑦𝐵 (𝑡0 ) = 𝑦𝐴 (𝑡0 ) ≠ 0, (4a) 𝑦̇ 𝐵 (𝑡0 ) = 𝑦̇𝐴 (𝑡0 ); (4b) (𝑘2 + 𝑘0 − 1)2 = 𝑘12 (1 − 𝐴0 ), (4c) ii) iii) 𝑦̇ 𝐵 (𝑡0 ) = −𝑎2 −0,5 (𝑡0 ) [ 𝑘2 +𝑘0 −1 𝑘1 + 𝑓𝐴 (𝑡0 )] 𝑦𝐵 (𝑡0 ); (4d) which are necessary and sufficient together with Eqs. 3a,b,c for commutativity of 𝐴 and 𝐵 under non-zero initial conditions. Note that by the nonzero initial condition it is meant “general values of initial conditions”, so one or two of them may be zero in special cases. In fact, if the output 𝑦(𝑡0 ) is zero, its derivative need to be zero due to Eq. 4d; if not, its derivative may or may not be zero depending on the term in bracket in (4d) is zero or not [21]. III. Inverse Commutativity Conditions for Second Order Unrelaxed Systems For the proof of the transitivity theorems of the following section, we need some formulas which express the inverse commutativity conditions. Although these conditions have been partially treated in [21], initial conditions are all assumed to be zero their; and for the sake of completeness, we express the general results by Lemma 1 and exhibit the complete inverse commutativity conditions for unrelaxed second order linear time-varying systems. In the previous section, the necessary and sufficient conditions for the commutativity of 𝐴 and 𝐵 are expressed dominantly by awnsening “what are the conditions that must be satisfied by 𝐵 to be commutative with 𝐴?” The answer to this question constitutes the inverse commutativity conditions. We express the results by the following Lemma. Lemma 1. The necessary and sufficients conditions given in Eqs. (3) and (4) for the commutativity 𝐴 and 𝐵 can be expressed by the following formulas: I. i) 6 𝑏2 (𝑡) 0 𝑎2 (𝑡) 0,5 [𝑎1 (𝑡)] = [𝑏1 (𝑡) 𝑏2 (𝑡) 𝑎0 (𝑡) 𝑏0 (𝑡) 𝑓𝐵 (𝑡) 𝑓𝐵 (𝑡) = 𝑏2 −0,5 [2𝑏1 (𝑡)−𝑏̇2 (𝑡)] 4 0 𝑙2 0] [𝑙1 ], where 1 𝑙0 (5a) . (5b) ii) II. 𝑏0 − 𝑓𝐵 2 − 𝑏2 0,5 𝑓𝐵̇ = 𝐵0 , ∀𝑡 ≥ 𝑡0 (5c) 𝑦𝐴 (𝑡0 ) = 𝑦𝐵 (𝑡0 ), (6a) 𝑦̇𝐴 (𝑡0 ) = 𝑦̇ 𝐵 (𝑡0 ); (6b) (𝑙2 + 𝑙0 − 1)2 = 𝑙12 (1 − 𝐵0 ), (6c) i) ii) 𝑙 +𝑙0 −1 𝑦̇𝐴 (𝑡0 ) = −𝑏2−0.5 (𝑡0 ) [ 2 𝑙1 + 𝑓𝐵 (𝑡0 )] 𝑦𝐴 (𝑡0 ); (6d) For the proof of the lemma, we solve Eq. 3a for 𝑎2 , 𝑎1 , 𝑎0 in terms of 𝑏2 , 𝑏1 , 𝑏0 as follows: 1 𝑎2 (𝑡) = 𝑘 𝑏2 (𝑡), (7a) 2 𝑎1 (𝑡) = 𝑏1 (𝑡)−𝑘1 𝑎20.5 (𝑡) 𝑘2 1 𝑘 𝑏 0,5 = 𝑘 𝑏1 (𝑡) − 𝑘1 (𝑘2 ) 2 2 2 1 = 𝑘 𝑏1 (𝑡) − 𝑘 2 𝑘1 2 1,5 𝑏20.5 (𝑡). (7b) Before progressing further, let us compute 𝑓𝐴 from (3b) and by using the above formulas for 𝑎2 and 𝑎1 , we obtain 𝑓𝐴 = 1 𝑏2−0.5 𝑏1 𝑘1 𝑏20.5 𝑏̇2 1 𝑏2−0.5 2𝑏1 − 𝑏̇2 2𝑘1 𝑏20.5 [2 ( − 1.5 ) − ] = [ − ] 4 𝑘2 𝑘2 𝑘2 4 𝑘2−0.5 𝑘2 𝑘2 𝑘21.5 = 𝑘2−0.5 𝑏2−0.5 (2𝑏1 − 𝑏̇2 ) 𝑘1 − 4 2𝑘2 Finally, defining 𝑏2−0.5 (2𝑏1 − 𝑏̇2 )/4 as 𝑓𝐵 as in (5b), we have 𝑘 𝑓𝐴 = 𝑘2−0.5 𝑓𝐵 − 2𝑘1 , (8a) 2 or equivalently 7 𝑘 1 𝑓𝐵 = 𝑘20.5 𝑓𝐴 + 2𝑘 0.5 (8b) 2 We now compute 𝑎0 from the last row of (3a) by using (8a) 𝑎0 = 𝑏0 −𝑘1 𝑓𝐴 −𝑘0 𝑘2 1 𝑏 𝑘 𝑘 2 2 2 𝑘2 𝑘 𝑘 1 = 𝑘 𝑏0 − 𝑘 1.5 𝑓𝐵 + 2𝑘12 − 𝑘0 . 2 𝑘 = 𝑘0 − 𝑘1 [𝑘2−0.5 𝑓𝐵 + 2𝑘1 ] − 𝑘0 2 2 2 2 (9) Writing (7a), (7b), and (9) in matrix form we obtain 1 𝑏2 𝑎2 [𝑎1 ] = [𝑏1 𝑎0 𝑏0 0 𝑏20.5 𝑓𝐵 𝑘2 −𝑘1 0 0] 1 , 𝑘21.5 𝑘12 2𝑘22 [ (10) 𝑘 − 𝑘0 ] 2 Comparing with Eq. 5a, we observe that (5a) is valid with 1 𝑙2 [𝑙1 ] = 𝑙0 𝑘2 −𝑘1 . 𝑘21.5 𝑘12 (11a) 𝑘0 [2𝑘22 − 𝑘2 ] Hence (5a) has been proved. For use in the sequel, we solve (11a) for 𝑘𝑖 ’s and obtain 1 𝑘2 [𝑘1 ]= 𝑘0 𝑙2 −𝑙1 , 𝑙21.5 𝑙12 (11b) 𝑙0 [2𝑙22 − 𝑙2 ] which is naturally the dual of Eq. 11a with 𝑘 and 𝑙 iterchanged. By using (11b) in (8a) and (8b), or directly interchanging 𝐴 ↔ 𝐵 and 𝑘𝑖 ↔ 𝑙𝑖 in (8a) and (8b), we obtain the following equations: 𝑙 𝑓𝐵 = 𝑙2−0.5 𝑓𝐴 − 2𝑙1 , 2 𝑙 𝑓𝐴 = 𝑙20.5 (𝑓𝐵 + 2𝑙1 ). 2 8 (11c) (11d) To show (5c), we substitute values of 𝑏2 and 𝑏0 from (3a), value of 𝑓𝐵 from (8b) in the left side of (5c), we obtain 𝑘1 2 𝑏0 − 𝑓𝐵2 − 𝑏20.5 𝑓𝐵̇ = 𝑘2 𝑎0 + 𝑘1 𝑓𝐴 + 𝑘0 − (𝑘20.5 𝑓𝐴 + 2𝑘 0.5 ) − (𝑘2 𝑎2 )0,5 (𝑘20.5 𝑓𝐴̇ ) 2 𝑘2 =𝑘2 𝑎0 + 𝑘1 𝑓𝐴 + 𝑘0 − 𝑘2 𝑓𝐴2 − 𝑘1 𝑓𝐴 − 4𝑘1 − 𝑘2 𝑎20.5 𝑓𝐴̇ 2 2 𝑘 = 𝑘2 (𝑎0 − 𝑓𝐴2 − 𝑎20.5 𝑓𝐴̇ ) + 𝑘0 − 4𝑘1 . 2 Finally using (3c) 𝑏0 − 𝑓𝐵2 − 𝑏20.5 𝑓𝐵̇ 𝑘12 = 𝑘2 𝐴0 + 𝑘0 − 4𝑘2 which is constant for 𝐴0 being constant. Hence (5c) is valid with 𝑘2 𝐵0 = 𝑘2 𝐴0 + 𝑘0 − 4𝑘1 , (12a) 2 or equivalently 1 𝑘2 𝑘 𝐴0 = 𝑘 𝐵0 + 𝑘0 + 4𝑘1 . 2 2 (12b) 2 The dual equations for (12a) and (12b) can be written by using constants 𝑙𝑖 ’s; this is done by using (11b) in (12a) and (12b), or directly interchanging 𝐴 ↔ 𝐵 and 𝑘𝑖 ↔ 𝑙𝑖 in (12a) and (12b). The results are 𝑙2 𝐴0 = 𝑙2 𝐵0 + 𝑙0 − 4𝑙1 , (12c) 2 𝐵0 = 1 𝑙2 𝐴0 − 𝑙0 𝑙2 + 𝑙12 4𝑙22 . (12d) Equations (6a) and (6b) are the same as Eqs. 4a and 4b, respectively, so they do not need to be proved. To prove (6c), we start from (4c); inserting valves of 𝑘𝑖 ’s from (11b) and valve of 𝐴0 from (12b) in, we obtain: 1 𝑙12 𝑙0 𝑙1 2 𝑙12 2 ( + 2 − − 1) = 3 (1 − 𝑙2 𝐵0 − 𝑙0 + ) 𝑙2 2𝑙2 𝑙2 4𝑙2 𝑙2 9 ( 1 − 𝑙0 − 𝑙2 𝑙12 𝑙12 𝑙12 + 2 )2 = 3 (1 − 𝑙2 𝐵0 − 𝑙0 + ) 𝑙2 4𝑙2 2𝑙2 𝑙2 (1 − 𝑙0 − 𝑙2 + 𝑙12 2 𝑙12 𝑙12 ) = (1 − 𝑙 𝐵 − 𝑙 + ) 2 0 0 𝑙2 4𝑙2 2𝑙22 𝑙14 𝑙12 𝑙12 𝑙14 (1 ) ) (1 − 𝑙0 − 𝑙2 ) + 2 + − 𝑙0 − 𝑙2 = ((1 − 𝑙2 𝐵0 − 𝑙0 + 2 ) 𝑙2 4𝑙2 𝑙2 4𝑙2 2 (1 − 𝑙0 − 𝑙2 )2 = 𝑙12 (1 − 𝑙2 𝐵0 − 𝑙0 − 1 + 𝑙0 + 𝑙2 ) 𝑙2 (𝑙2 + 𝑙0 − 1)2 = 𝑙12 (1 − 𝐵0 ) To prove (6d), we start from (4d) and using (4a) and (4b), we write 𝑦̇𝐴 (𝑡0 ) = −𝑎2−0.5 (𝑡0 ) [ 𝑘2 +𝑘0 −1 + 𝑓𝐴 (𝑡0 )] 𝑦𝐴 (𝑡0 ). 𝑘1 Using (5a) for 𝑎2 , (11b) for 𝑘𝑖 ’s and (11d) for 𝑓𝐴 , we proceed 𝑦̇𝐴 (𝑡0 ) = −[𝑙2 𝑏2 (𝑡0 )]−0.5 [ 2 1 𝑙1 𝑙0 + − −1 𝑙2 2𝑙2 𝑙2 2 𝑙1 − 1.5 𝑙2 = −𝑙2−0.5 𝑏2−0.5 (𝑡0 )[− 1 𝑙21.5 𝑙1 𝑙 𝑙 1 + 𝑙20.5 𝑓𝐵 (𝑡0 ) + 2𝑙0.5 ] 𝑦𝐴 (𝑡0 ) 2 𝑙2 1 𝑙 𝑙 1 (𝑙 + 2𝑙12 − 𝑙0 − 1) + 𝑙20.5 𝑓𝐵 (𝑡0 ) + 2𝑙0.5 ] 𝑦𝐴 (𝑡0 ) 2 2 2 𝑙 2 𝑙 𝑙 =−𝑏2−0.5 (𝑡0 )[− 𝑙 − 2𝑙1 + 𝑙0 + 𝑙2 + 𝑓𝐵 (𝑡0 ) + 2𝑙1 ] 1 2 𝑙 +𝑙0 −1 = −𝑏2−0.5 (𝑡0 )[ 2 𝑙1 1 1 2 + 𝑓𝐵 (𝑡0 )] 𝑦𝐴 (𝑡0 ) which is the same equation as (6d). Hence the proof of Lemma 1 is completed. Fact: Comparing Eqs. 4d and 6d, together with the equalities 𝑦𝐴 (𝑡0 ) = 𝑦𝐵 (𝑡0 ), 𝑦̇𝐴 (𝑡0 ) = 𝑦̇ 𝐵 (𝑡0 ), we see that the derivatives 𝑦̇𝐴 (𝑡0 ) and 𝑦̇ 𝐵 (𝑡0 ) are constant multiples of 𝑦𝐴 (𝑡0 ) and 𝑦𝐵 (𝑡0 ). The multipliers are initial time dependent and −𝑎𝑏2−0.5 (𝑡0 ) [ 𝑘2 +𝑘0 −1 𝑘1 𝑙 +𝑙0 −1 + 𝑓𝐵 (𝑡0 )]. 𝑙1 + 𝑓𝐴 (𝑡0 )] = −𝑏2−0.5 (𝑡0 )[ 2 Inserting in value of 𝑏2 from (3a) and value of 𝑓𝐵 (𝑡0 ) from (8b) yields that 𝑙2 +𝑙0 −1 𝑙1 𝑘2 +𝑘0 −1 = 𝑘20.5 [ 10 𝑘1 𝑘 − 2𝑘1 ]. 2 On the other hand, inserting in value of 𝑎2 from (5a) and value of 𝑓𝐴 (𝑡0 ) from (11d) yields 𝑘2 + 𝑘0 − 1 𝑙2 + 𝑙0 − 1 𝑙1 = 𝑙20.5 [ − ]. 𝑘1 𝑙1 2𝑙2 This is the dual of the previous equation. Using the transformations (11a) and (11b) between 𝑘𝑖 ’s and 𝑙𝑖 ’s, it is straightforword to show that the above relations between 𝑘𝑖 ’s and 𝑙𝑖 ’s are valid. IV. Transitivity Property of Commutativity To be able to study the transivity property of commutativity for second-order linear timevarying systems, we should consider a third system 𝐶 of the same type as 𝐴 and 𝐵 considered in Section III. So, let 𝐶 be defined by the following second-order differential equation: 𝑐2 (𝑡)𝑦̈ 𝐶 (𝑡) + 𝑐1 (𝑡)𝑦̇ 𝐶 (𝑡) + 𝑐0 (𝑡)𝑦𝐶 (𝑡) = 𝑥𝐶 (𝑡); 𝑡 ≥ 𝑡0 , (13a) 𝑦𝐶 (𝑡0 ), 𝑦̇ 𝐶 (𝑡0 ), (13b) 𝑐2 (𝑡) ≢ 0, (13c) where 𝑐̈2 (𝑡), 𝑐̇2 (𝑡), 𝑐0 (𝑡) ∈ 𝐶[𝑡0 , ∞). We assume 𝐶 is commutative with 𝐵, to similar relations to Eqs. (3) and (4) can be written as I. i) 𝑏2 𝑐2 [𝑐1 ] = [𝑏1 𝑐0 𝑏0 𝑓𝐵 = 0 𝑏20.5 𝑓𝐵 0 𝑚2 0] [𝑚1 ] , where 1 𝑚0 𝑏2 0,5 (2𝑏1 −𝑏̇2 ) 4 . (14a) (14b) And 𝑚2 , 𝑚1 , 𝑚0 are some constants with 𝑙2 ≠ 0 for 𝐶 is of second-order. Further, we assume that 𝐶 can not be obtained from 𝐵 by constant feed forward and feedback gains; hence 𝑚1 ≠ 0. Moreover, ii) 𝑏0 − 𝑓𝐵 2 − 𝑏2 0,5 𝑓𝐵̇ = 𝐵0 , ∀𝑡 ≥ 𝑡0 11 (14c) where 𝐵0 is a constant. When the initial conditions are non-zero, the following should be satisfied: II. i) 𝑦𝐶 (𝑡0 ) − 𝑦𝐵 (𝑡0 ) ≠ 0, (15a) 𝑦̇ 𝐶 (𝑡0 ) = 𝑦̇ 𝐵 (𝑡0 ), (15b) (𝑚2 + 𝑚0 − 1)2 = 𝑚1 2 (1 − 𝐵0 ), (15c) ii) 𝑦̇ 𝐶 (𝑡0 ) = −𝑏2 −0,5 [ 𝑚2 +𝑚0 −1 𝑚1 + 𝑓𝐵 (𝑡0 )] 𝑦𝐶 (𝑡0 ). (15d) Considering the inverse commutativity conditions derived for 𝐴 and 𝐵, the inverse commutativity conditions for 𝐵 and 𝐶 can be written from Eqs. (5) and (6) by changing 𝐴 → 𝐵 and 𝐵 → 𝐶, ℎ𝑖 → 𝑚𝑖 and 𝑙𝑖 → 𝑛𝑖 in Eqs. (5) and (6). The results are I. i) 𝑐2 𝑏2 [𝑏1 ] = [𝑐1 𝑏0 𝑐0 𝑓𝑐 = 0 𝑐20,5 𝑓𝑐 0 𝑛2 0] [𝑛1 ] , where 1 𝑛0 𝑐2−0,5 (2𝑐1 −𝑐2 ) 4 . (16a) (16b) ii) II. 𝑐0 − 𝑓𝑐2 − 𝑐20,5 𝑓𝑐 = 𝑐0 , ∀𝑡 ≥ 𝑡0 (16c) 𝑦𝐵 (𝑡0 ) = 𝑦𝐶 (𝑡0 ), (17a) 𝑦′𝐵 (𝑡0 ) = 𝑦′𝐶 (𝑡0 ), (17b) (𝑛2 + 𝑛0 − 1)2 = 𝑛12 (1 − 𝑐0 ), (17c) i) ii) 𝑦̇ 𝐵 (𝑡0 ) = −𝑐2−0,5 (𝑡0 ) [ 𝑛2 +𝑛0 −1 𝑛1 Further, Eqs, (8a) and (8b) become 12 + 𝑓𝑐 (𝑡0 )] 𝑦𝐵 (𝑡0 ). (17d) 𝑓𝐵 = 𝑚2−0,5 𝑓𝑐 − 2𝑚1 𝑓𝐶 = 𝑚20,5 𝑓𝐵 + 𝑚1 𝑚2 2𝑚20,5 , (18a) . (18b) The relations between the constants 𝑚𝑖 and 𝑛𝑖 can be written from Eqs. (11a) and (11b) by the replacements 𝑘𝑖 → 𝑚𝑖 , 𝑙𝑖 → 𝑛𝑖 . The results are; 1 𝑛2 [𝑛1 ] = 𝑛0 𝑚2 𝑚1 − , 𝑚21,5 𝑚12 (19a) 𝑚0 [2𝑚22 − 𝑚2 ] 1 𝑛2 𝑛1 𝑚2 [𝑚1 ] = 𝑚0 − 𝑛12 [ 2𝑛22 . 𝑛21,5 (19b) 𝑛 − 𝑛0 ] 2 By using values of 𝑚𝑖 ’s in Eqs. (18a) and (18b) , or directly interchanging 𝐵 ↔ 𝐶, 𝑚𝑖 ↔ 𝑛𝑖 , we obtain 𝑓𝐶 = 𝑛2−0.5 𝑓𝐵 − 2𝑛1 𝑓𝐵 = 𝑛20,5 𝑓𝐶 + 𝑛1 𝑛2 2𝑛20,5 , (19c) ). (19d) Finally, Eqs. (12a, b, c, d) turns out to be 𝑚2 𝐶0 = 𝑚2 𝐵0 + 𝑚0 − 4𝑚1 , 2 𝐵0 = 1 𝑚2 𝐶0 − 𝑚0 𝑚2 + 𝑚12 4𝑚22 𝑛2 𝐵0 = 𝑛2 𝐶0 + 𝑛0 − 4𝑛1 , 2 1 𝑛 𝑛2 𝐶0 = 𝑛 𝐵0 − 𝑛0 + 4𝑛12 . 2 2 , (20a) (20b) (20c) (20d) 2 by the replacements 𝐴 → 𝐵, 𝐵 → 𝐶, 𝑘𝑖 → 𝑚𝑖 , 𝑙𝑖 → 𝑛𝑖 . The preliminaries have been ready now for studying the transitivity property of commutativity. Assuming 𝐵 is commutative with 𝐴 and 𝐶 is commutative with 𝐵, we need 13 to answer that weather 𝐶 is a commutative pair of 𝐴 . The answer is expressed by the following theorems and their proves. Theorem 1: Transitivity property of commutativity for second-order linear time-varying analog systems which cannot be obtained from each other by constant feed forward and feedback gains is always valid under zero initial conditions. Proof: Since it is true by the hypothesis that 𝐵 is commutative with 𝐴, Eqs. (3a) and (3c) are valid; since it is true by hypothesis that 𝐶 is commutative with 𝐵, Eqs. (14a) and (14c) are also valid. To prove Theorem 1, it should be proven that 𝐶 is commutative with 𝐴 under zero initial conditions. Referring to the commutativity conditions for 𝐴 and 𝐵 in Eq. (3a) and replacing 𝐵 by 𝐶, this proof is done by showing the validity of 𝑎2 𝑐2 [𝑐1 ] = [𝑎1 𝑐0 𝑎0 0 𝑎20,5 𝑓𝐴 0 𝑝2 0] [𝑝1 ], 1 𝑝0 (20) where 𝑓𝐴 (𝑡) is given as in Eq. (3b), and the coefficients of 𝐴 already satisfy Eq. (3c) due to the commutativity of 𝐵 with 𝐴; further 𝑝2 , 𝑝1 , 𝑝0 are some constants to be revealed: Using Eq. (14a) first and then Eq. (3a), as well as Eq. (8b) for computing 𝑐0 , we can express 𝑐2 , 𝑐1 , 𝑐0 as follows: 𝑐2 = 𝑚2 𝑏2 = 𝑚2 𝑘2 𝑎2 , 𝑐1 = 𝑚2 𝑏1 + 𝑚1 𝑏20,5 = 𝑚2 (𝑘2 𝑎1 + 𝑘1 𝑎20,5 ) + 𝑚1 (ℎ2 𝑎2 )0,5 = 𝑚2 ℎ2 𝑎1 + (𝑚2 𝑘1 + 𝑚1 𝑘20,5 )𝑎20,5 𝑐0 = 𝑚2 𝑏0 + 𝑚1 𝑓𝐵 + 𝑚0 = 𝑚2 (𝑘2 𝑎0 + 𝑘1 𝑓𝐴 + 𝑘0 ) + 𝑚1 (𝑘20,5 𝑓𝐴 + 𝑘1 2𝑘20,5 = 𝑚2 𝑘2 𝑎0 + (𝑚2 𝑘1 + 𝑚1 𝑘20,5 )𝑓𝐴 + 𝑚2 𝑘0 + 14 ) + 𝑚0 𝑚1 𝑘1 2𝑘20,5 + 𝑚0 . These results can be written as 𝑎2 𝑐2 [𝑐1 ] = [𝑎1 𝑐0 𝑎0 𝑚2 𝑘2 0 0,5 0] [ 𝑚2 𝑘1 + 𝑚1 𝑘2 ], 𝑚1 𝑘1 1 𝑚2 𝑘0 + 0,5 + 𝑚0 0 𝑎20,5 𝑓𝐴 (21a) 2𝑘2 which is exactly in the same form as Eq. (20) with the constants 𝑝2 , 𝑝1 , 𝑝0 ; 𝑚2 𝑘2 𝑝2 0,5 [𝑝1 ] = [ 𝑚2 𝑘1 + 𝑚1 𝑘2 ]. 𝑚 𝑘 𝑝0 𝑚2 𝑘0 + 10,51 + 𝑚0 (21b) 2𝑘2 So, the proof is completed. For the validity of transitivity property for second-order linear time-varying analog systems under non-zero initial conditions, we state the following theorem. Theorem 2: Transitivity property of commutativity of systems considered in Theorem 1 is valid for the non-zero initial conditions of the systems as well. Proof: The proof is done by showing the commutativity of 𝐶 with 𝐴 under non-zero conditions as well. Since 𝐶 and 𝐴 are commutative with non-zero initial conditions Eq. (20) and Eq. (3a) are valid as mentioned in proof of Theorem 1. To complete the proof, we should show that 𝐶 is a commutative pair of 𝐴 under non-zero conditions as well, Eqs. (4ad) are satisfied for systems 𝐶 (instead of 𝐵) and 𝐴. Namely, 𝑦𝐶 (𝑡0 ) = 𝑦𝐴 (𝑡0 ) ≠ 0, (22a) 𝑦̇ 𝐶 (𝑡0 ) = 𝑦̇𝐴 (𝑡0 ), (22b) (𝑝2 + 𝑝0 − 1)2 = 𝑝12 (1 − 𝐴0 ), (22c) 𝑦̇ 𝐶 (𝑡0 ) = −𝑎2−0,5 (𝑡0 ) [ 𝑝2 +𝑝0 −1 𝑝1 + 𝑓𝐴 (𝑡0 )] 𝑦𝐶 (𝑡0 ), (22d) where 𝑘𝑖 ’s in Eq. (3a) for system 𝐵 are replaced by 𝑝𝑖 ‘s in Eq. (20) for system 𝐶. Since, (𝐴, 𝐵) and (𝐵, 𝐶) are commutative under non-zero initial conditions by hypothesis, Eqs. (4a, b) and (17a, b) are satisfied; so it follows that Eqs. (22a) and (22b) are valid. Since, 15 𝐵 and 𝐶 are commutative, in the commutativity conditions (4c) 𝑦′𝐵 (𝑡0 ) and 𝑦𝐵 (𝑡0 ) can be replaced by 𝑦′𝐶 (𝑡0 ) and 𝑦𝐶 (𝑡0 ) due to Eqs. (15a, b); the result is 𝑦̇ 𝐶 (𝑡0 ) = −𝑎2−0,5 (𝑡0 ) [ 𝑘2 +𝑘0 −1 𝑘1 + 𝑓𝐴 (𝑡0 )] 𝑦𝐶 (𝑡0 ). (23) On the other hand, 𝑦̇ 𝐶 (𝑡0 ) and 𝑦𝐶 (𝑡0 ) are related by Eq. (15d). Comparing it with Eq. (23), we write −𝑏2−0,5 (𝑡0 ) [ 𝑚2 +𝑚0 −1 𝑚1 + 𝑓𝐵 (𝑡0 )] 𝑦𝐶 (𝑡0 ) = −𝑎2−0,5 (𝑡0 ) [ 𝑘2 +𝑘0 −1 𝑘1 + 𝑓𝐴 (𝑡0 )] 𝑦𝐶 (𝑡0 ). (24) Since, (𝐴, 𝐵) is a commutative pair, substituting the values of 𝑎2 from Eq. (5a) and 𝑓𝐴 from Eq. (8a) into Eq. (24), we obtain −𝑏2−0,5 (𝑡0 ) [ 𝑚2 + 𝑚0 − 1 + 𝑓𝐵 (𝑡0 )] 𝑦𝐶 (𝑡0 ) 𝑚1 −0,5 𝑏 = − (𝑘2 ) 2 (𝑡0 ) [ 𝑘2 +𝑘0 −1 𝑘1 𝑘 + 𝑘2−0,5 𝑓𝐵 (𝑡0 ) − 2𝑘1 ] 𝑦𝐶 (𝑡0 ). 2 (25) Since, 𝑏2 (𝑡) ≠ 0, 𝑦𝐶 (𝑡0 ) ≠ 0, we can write the above equality as 𝑚2 +𝑚0 −1 𝑚1 + 𝑓𝐵 (𝑡0 ) = 𝑘20,5 [ 𝑘2 +𝑘0 −1 𝑘1 𝑘 − 2𝑘1 ] + 𝑓𝐵 (𝑡0 ). 2 (26) Finally, cancelling 𝑓𝐵 (𝑡0 ), we result with 𝑚2 +𝑚0 −1 𝑚1 = 𝑘20,5 [ 𝑘2 +𝑘0 −1 𝑘1 𝑘 − 2𝑘1 ], 2 (27) which is due to the commutativities of (𝐴, 𝐵) and (𝐵, 𝐶) under non-zero initial conditions. Now, to prove Eq. (22c), we proceed as follows: Using Eq. (21b), we compute 𝑝2 +𝑝0 −1 𝑝1 𝑚 𝑘 𝑚2 𝑘2 +𝑚2 𝑘0 + 10,51 +𝑚0 −1 2𝑘2 = 𝑚2 𝑘1 +𝑚1 𝑘20,5 . (28a) Solving Eq. (27) for 𝑚1 , we have 𝑚1 = 𝑚2 +𝑚0 −1 . 𝑘 +𝑘 −1 𝑘 𝑘20,5 ( 2 0 − 1 ) 𝑘1 2𝑘2 Substituting Eq. (28b) in (28a), we proceed as 16 (28b) 𝑚2 (𝑘2 + 𝑘0 ) + 𝑚0 − 1 + 𝑝2 + 𝑝0 − 1 = 𝑝1 𝑘1 𝑚2 + 𝑚 0 − 1 ] 0,5 [ 𝑘 2𝑘2 𝑘 0,5 ( 2 + 𝑘0 − 1 − 𝑘1 ) 2 𝑘1 2𝑘2 𝑚2 + 𝑚0 − 1 𝑚2 𝑘1 + 𝑘20,5 [ ] 𝑘 + 𝑘0 − 1 𝑘1 ℎ20,5 ( 2 − ) 𝑘1 2𝑘2 [𝑚2 (𝑘2 + 𝑘0 ) + 𝑚0 − 1]𝑘20,5 ( = 𝑘 (𝑚 + 𝑚 − 1) 𝑘2 + 𝑘0 − 1 𝑘1 − ) + 1 2 0,50 𝑘1 2𝑘2 2𝑘 2 𝑘 + 𝑘0 − 1 𝑘1 𝑚2 𝑘1 𝑘20,5 ( 2 − ) + 𝑘20,5 (𝑚2 + 𝑚0 − 1) 𝑘1 2𝑘2 𝑘 (𝑚 + 𝑚0 − 1) 𝑘 + 𝑘0 − 1 𝑘1 [𝑚2 (𝑘2 + 𝑘0 ) + 𝑚0 − 1] ( 2 − )+ 1 2 𝑘1 2𝑘2 2𝑘2 = 𝑘2 + 𝑘0 − 1 𝑘1 𝑚2 𝑘1 ( − ) + (𝑚2 + 𝑚0 − 1) 𝑘1 2𝑘2 𝑘1 𝑘 + 𝑘0 − 1 [𝑚2 + 𝑚0 − 1 − 𝑚2 (𝑘2 + 𝑘0 ) − 𝑚0 + 1] + 2 [𝑚2 (𝑘2 + 𝑘0 ) + 𝑚0 − 1] 2𝑘 ℎ1 = 2 𝑘2 𝑚2 (𝑘2 + 𝑘0 − 1) + 𝑚2 + 𝑚0 − 1 − 𝑚2 1 2𝑘2 𝑘12 (𝑘 ) 𝑚 + 𝑘 + 𝑚 − 1 − 𝑚 2 2 0 0 2 𝑘2 + 𝑘0 − 1 2𝑘2 𝑘2 + 𝑘0 − 1 = = . 𝑘1 𝑘1 𝑘12 𝑚2 (𝑘2 + 𝑘0 ) + 𝑚0 − 1 − 𝑚2 2 2𝑘2 (28𝑐) Using the equality (28c) in Eq. (4c) directly yields Eq. (22c). On the other hand, when Eq. (28c) is used in Eq. (23), this equation results with the proof of Eq. (22d), so does with the completion of the proof of Theorem 2. We now introduce an example to illustrate the results obtained in the paper and to validate the transitivity by computer simulation. V. Example To illustrate the validity of the results obtained in the previous section, consider the system 𝐴 defined by 𝐴: 𝑦′′𝐴 + (3 + sin 𝑡)𝑦′𝐴 + (3.25 + 0.25𝑠𝑖𝑛2 𝑡 + 1.5 sin 𝑡 + 0.5 cos 𝑡)𝑦𝐴 = 𝑥𝐴 , (29a) for which Eq. (3b) yields 17 𝑎2−0,5 [2𝑎1 − 𝑎2 ] 1[2(3 + sin 𝑡) − 0] 6 + 2 sin 𝑡 𝑓𝐴 (𝑡) = = = 4 4 4 = 1.5 + 0.5 sin 𝑡 , (29b) 𝑓 ′𝐴 (𝑡) = 0.5 cos 𝑡 . (29c) To check Eq. (3c), we proceed 𝐴0 = 𝑎0 − 𝑓𝐴2 − 𝑎20,5 𝑓 ′𝐴 = 3.25 + 0.25𝑠𝑖𝑛2 𝑡 + 1.5𝑠𝑖𝑛𝑡 + 0.5𝑐𝑜𝑠𝑡 − (1.5 + 0.5𝑠𝑖𝑛𝑡)2 − 0.5𝑐𝑜𝑠𝑡 = 3.25 + 0.25𝑠𝑖𝑛2 𝑡 + 1.5𝑠𝑖𝑛𝑡 − 2.25 − 1.5𝑠𝑖𝑛𝑡 − 0.25𝑠𝑖𝑛2 𝑡 = 1. (29d) Hence, this expression is constant, that is 𝐴0 = 1. Chosing 𝑘2 = 1, 𝑘1 = −2, 𝑘0 = 0 in Eq. (3a), 𝑎2 𝑏2 [𝑏1 ] = [𝑎1 𝑏0 𝑎0 =[ 0 𝑎20,5 𝑓𝐴 𝑎2 0 1 0,5 1] [−2] = [𝑎1 − 2𝑎2 ] 𝑎0 − 2𝑓𝐴 0 0 1 ] 3 + sin 𝑡 − 2 2 3.25 + 0.25𝑠𝑖𝑛 𝑡 + 0.5 sin 𝑡 + 0.5 cos 𝑡 1 =[ ]. 1 + sin 𝑡 − 2 0.25 + 0.25𝑠𝑖𝑛2 + 0.5 sin 𝑡 + 0.5 cos 𝑡 (30a) So, 𝐴 and 𝐵 are commutative under zero initial conditions. From Eq. (30a), we compute 𝑓𝐵 and 𝐵0 by using Eqs. (5b) and (5c) 𝑓𝐵 = 𝑏2−0,5 (2𝑏1 − 𝑏2 ) (2 + 2 sin 𝑡) = = 0.5 + 0.5 sin 𝑡 , 4 4 𝑓′𝐵 = 0.5 cos 𝑡 , (30b) (30c) 𝐵0 = 𝑏0 − 𝑓𝐵2 − 𝑏20,5 𝑓 ′ 𝐵 = 0.25 + 0.25𝑠𝑖𝑛2 𝑡 + 0.5 sin 𝑡 + 0.5 cos 𝑡 − (0.5 + 0.5 sin 𝑡)2 − 0.5 cos 𝑡 = 0.25 + 0.25𝑠𝑖𝑛2 𝑡 + 0.5 sin 𝑡 − 0.25 − 0.5 sin 𝑡 − 0.25𝑠𝑖𝑛2 𝑡 = 0. (30d) 18 We check the validity of Eq. (12a) by using Eqs. (30d) and (29d): 𝐵0 = 𝑘2 𝐴0 + 𝑘0 − (−2)2 𝑘12 = 1(1) + 0 − = 0. 4𝑘2 4(1) (30e) It can be checked easily by using Eqs. (29b) and (30b) that Eqs. (8a,b) are also correct. Considering the requirements for the non-zero initial conditions at 𝑡0 = 0, Eq. (4) yields 𝑦𝐵 (0) = −(1)2 [ 1+0−1 + 1.5 + 0.5 sin 0] 𝑦𝐴 (0) = −1.5𝑦𝐵 (0). −2 (31a) Hence, for the commutativity of 𝐴 and 𝐵 under non-zero initial conditions as well, due to Eqs. (6a, b) and (31a), 𝑦′𝐴 (0) = 𝑦′𝐵 (0) = −1.5𝑦𝐵 (0) = −1.5𝑦𝐴 (0). (31b) We now consider a third system 𝐶 which is commutative with 𝐵. Therefore, using Eq. (14c) with 𝑚2 = 1, 𝑚1 = 3, 𝑚0 = 3, we have 𝑏2 𝑐2 [𝑐1 ] = [𝑏1 𝑐0 𝑏0 0 𝑏20,5 𝑓𝐵 0 1 0] [3]. 1 3 Inserting values of 𝑏𝑖 ’s from Eq. (30a) and value of 𝑓𝐵 from Eq. (30b) in, we have 𝑐2 1 𝑐 [ 1] = [ 1 + sin 𝑡 𝑐0 0.25 + 0.25𝑠𝑖𝑛2 𝑡 + 0.5 sin 𝑡 + 0.5 cos 𝑡 0 1 0.5 + 0.5 sin 𝑡 1 =[ ]. 4 + sin 𝑡 4.75 + 0.25𝑠𝑖𝑛2 𝑡 + 2 sin 𝑡 + 0.5 cos 𝑡 0 1 0 ] [3 ] 1 3 (32a) Eqs. (16b) and (16c) yield that 𝑓𝐶 = 8 + 2 sin 𝑡 = 2 + 0.5 sin 𝑡 , 4 𝑓′𝐶 = 0.5 cos 𝑡 (32b) (32c) 𝐶0 = 𝑐0 − 𝑓𝐶2 − 𝑐20,5 𝑓 ′ 𝐶 = 4.75 + 0.25𝑠𝑖𝑛2 𝑡 + 2 sin 𝑡 + 0.5 cos 𝑡 − (2 + 0.5 sin 𝑡) 2 − 0.5 cos 𝑡 = 4.75 + 1.25𝑠𝑖𝑛2 𝑡 + 2 sin 𝑡 − 4 − 2𝑠𝑖𝑛𝑡 − 0.25𝑠𝑖𝑛2 𝑡 19 = 0.75. (32d) One can easily check that 𝐶0 and 𝐵0 in Eqs. (32d) and (30d) satisfy relations in Eqs. (20a,b). For the commutativity of 𝐵 and 𝐶 under non-zero initial conditions at time 𝑡0 = 0 as well; Eq. (15) together with Eqs. (30a), (30b) and chosen values of 𝑚𝑖 ’s yield 𝑦̇ 𝐶 (0) = 𝑦̇ 𝐵 (0) = −(1)2 [ = −[ 𝑚2 + 𝑚0 − 1 + 0,5 + 0,5 sin 0] 𝑦𝐶 (𝑡0 ) 𝑚1 1+3−1 + 0,5] 𝑦𝐶 (𝑡0 ) = −1,5𝑦𝐶 (𝑡0 ) = −1,5𝑦𝐵 (𝑡0 ). 3 (33) Considering the transitivity property under non-zero initial conditions, the conditions of Theorem 2 are satisfied. Namely, using the chosen values of 𝑚𝑖 ’s and 𝑘𝑖 ’s , from Eq. (21b), we have 𝑝2 = 1, 𝑝1 = 1, 𝑝0 = 0. And with 𝐴0 = 1 as computed in Eq. (29d), Eq. (22c) is satisfied; (1 + 0 − 1)2 = (1)2 (1 − 1). So does Eq. (22d) 1−0−1 𝑦̇ 𝐶 (0) = −(1)−0,5 ( 1 + 1.5 + 0.5 sin 0) 𝑦𝐶 (0) = −1.5𝑦𝐶 (0). The answer is obviously yes as already shown in Eq. (33). The simulations are done for the inter connection of the above mentioned systems 𝐴, 𝐵, 𝐶. The initial conditions are taken as 𝑦𝐴 (0) = 𝑦𝐵 (0) = 𝑦𝐶 (0) = 1, (34a) 𝑦̇𝐴 (0) = 𝑦̇ 𝐵 (0) = 𝑦̇ 𝐶 (0) = −1.5𝑦𝐴 (0) = −1.5𝑦𝐵 (0) = −1,5𝑦𝐶 (0) = −1.5. (34b) And input is assumed 40sin(10𝜋𝑡). It is observed that 𝐴𝐵, 𝐵𝐴 yield the same response; 𝐵𝐶, 𝐶𝐵 yield the same response; so 𝐶𝐴, 𝐴𝐶 yield to same response. These responses are shown in Fig. 2 by 𝐴𝐵 = 𝐵𝐴, 𝐵𝐶 = 𝐶𝐵, 𝐶𝐴 = 𝐴𝐶, respectively. Hence, transitivity property shows up as if (𝐴, 𝐵) and (𝐵, 𝐶) are commutative pairs so is (𝐴, 𝐶). 20 These simulations and all the subsequent ones are done by MATLAB2010 Simulink Toolbox with fixed time step of 0.02 using ode3b (Bogachi-Shampine) program; the final time is 𝑡 = 10. Fig. 2: Outputs of commutative cascade connections 𝐴𝐵, 𝐵𝐴, 𝐶𝐵, 𝐵𝐶, 𝐶𝐴, 𝐴𝐶 with nonzero initial conditions. The second set of simulations are obtained by zero initial conditions, the conditions of Theorem 1 are satisfied by choosing 𝑚2 = 1, 𝑚1 = −1, 𝑚0 = 3, so that 𝐶 is obtained from 𝐵 through Eq. (14) as 𝑐2 1 1 + sin 𝑡 [𝑐1 ] = [ 𝑐0 0.25 + 0.25𝑠𝑖𝑛2 𝑡 + 0.5 sin 𝑡 + 0,5 cos 𝑡 21 0 1 0.5 + 0.5 sin 𝑡 0 1 0] [−1] 1 3 =[ 1 ]. 2 − sin 𝑡 2 2.75 + 0.25𝑠𝑖𝑛 𝑡 + 0.5 cos 𝑡 Hence, 𝐶 is commutative with 𝐵, and together with 𝐵 being commutative with 𝐴, the conditions of Theorem 1 are satisfied so that 𝐶 is commutative with 𝐴. This is observed in Fig. 3. In this figure, the responses indicated by 𝐴𝐵 = 𝐵𝐴, 𝐵𝐶 = 𝐶𝐵, 𝐶𝐴 = 𝐴𝐶 which are all obtained by zero initial conditions validates the transitivity commutativity, that is Theorem 1 is valid. Fig. 3: The simulations obtained with zero initial conditions 22 property of Finally, the simulations are performed for arbitrary initial conditions 𝑦𝐴 (0) = 0.4, 𝑦̇𝐴 (0) = −0.3, 𝑦𝐵 (0) = 0.2, 𝑦̇ 𝐵 (0) = −0.4, 𝑦𝐶 (0) = −0.5, 𝑦̇ 𝐶 (0) = 0.5. It is observed that (𝐴, 𝐵), (𝐵, 𝐶), (𝐶, 𝐴) are not commutative pairs at all, the plots AB, BA; BC, CA; CA, AC are shown in Fig. 4, respectively. However, since all systems (individually and in pairs as cascade conceded) are asymptotically stable and the effects of non-zero initial conditions die away as time proceeds, and 𝐴, 𝐵, 𝐶 are pairwise commutative with zero initial conditions, the responses of 𝐴𝐵 and 𝐵𝐴, 𝐵𝐶 and 𝐶𝐵, 𝐶𝐴 and 𝐴𝐶 approach each other with increasing time. That is commutativity property and its transitivity gets valid in the steadystate case. 23 Fig. 4: Responses of cascade connection of Systems 𝐴, 𝐵, 𝐶 (which are commutative with zero initial conditions) with arbitrary initial conditions not satisfying commutativity conditions. VI. Conclusions On the base of the commutativity conditions for second order linear time-varying analog systems with nonzero initial conditions, the inverse commutativity conditions are reformulated completely in the form of Lemma 1 by considering the case of non-zero initial conditions. With the obtained results, the transitivity property of commutativity is stated both for relaxed and unrelaxed cases by Theorems 1 and 2, respectively. Througout the study the subsystems considered are assumed not obtainable from each other by any feedforword and feed-back structure, which is a case that needs special treatement due to special commutativity requirements in case of nonzero initial conditions [21]. All the results derived in the paper are well verified by simulations done by MATLAB2010 Simulink Toolbox using ode3b (Bogachi-Shampine) program. Acknowledgments: This study is supported by the Scientific and Technological Research Council of Turkey (TUBITAK) under the project no. 115E952. References [1] M. Borenovic, A. Neskovic, D. Budimir, Space partitioning strategies for indoor wlan positioning with cascade-connected ann structures, International Journal of Neural Systems, 21, 1-15, 2011. [2] D. antic, Z. Jovanovic, V. Nikolic, M. Milojkovic, S. Nikolic, N. Dankovic, Modelling of cascade-connected systems using quise-orthagonal functions, Elektronika Ir Elektrotechnika, 18, 3-8, 2012 [3] K. Nagase, Wave analysis and control of double cascade-connected damped mass24 spring systems, Mechanics Research Communications, 70, 49-57, 2015. [4] B. Samardzic, B. M. Zlatkovic, Analysis of spatial chaos appearance in cascade connected nonlinear electrical circuits, Chaos Solutions and Fractals, 95, 14-20, 2017. [5] T. Kaczorek, Positive time-varying continuous-time linear systems and electrical circuits, The journal of Polish Academy of Sciences, 63, 1-6, 2015. [6] B. Basu, A. Staino, Control of a linear time-varying system with a forward Ricatti formulation in wavelet domain, Journal of Dynamic systems Measurement and Control-Transactions of the ASME, 138, 1-6, 2016. [7] R.K.R. Alla, J.S. Lather, G.L. Pahuja, New delay dependent stability criterion for linear system with time-varying delay using Wirtinger’s inequality, Journal of Engineering Research, 4, 103-116, 2016. [8] J. Wang, C.M. Mak, An active vibration control system with decoupling scheme for linear periodically time-varying systems, Journal of Vibration and Control, 22, 2370-2379, 2016. [9] W. Guan, C. Wang, D.S. Chen, C.Y. Luo, F.F. Su, Recursive principal component analysis with forgetting factor for operational modal analysis of linear time-varying system, International Journal of Applied Electromagnetics and Mechanics, 52, 9991006, 2016. [10] J. Lataire, R. Pintelon, D. Piga, R. Toth, Continuous-time linear time-varying system identification with a frequency-domain kernel-based estimator, IET Control Theory and Applications, 11, 457-465, 2017. [11] R. Boylestad and L. Nashelsky, Electronic Devices and Circuit Theory, Prentice Hall, New Jersey, 2013. [12] E. Marshal, Commutativity of time varying systems, Electronics Letters, 13, 539540, 1977. [13] M. Koksal, An exhaustive study on the commutativity of time-varying systems, International Journal of Control, 47, 1521-1537, 1988. [14] M. Koksal and M. E. Koksal, Commutativity of linear time-varying differential systems with non-zero initial conditions: A review and some new extensions, Mathematical Problems in Engineering, 2011, 1-25, 2011. [15] M. E. Koksal and M. Koksal, Commutativity of cascade connected discrete time linear time-varying systems, 2013 Automatic Control National Meeting TOK’2013, p.1128-1131, 2013. 25 [16] M. Koksal and M. E. Koksal, Commutativity of cascade connected discrete-time linear time-varying systems, Transactions of the Institute of Measurement and Control, 37, 615-622, 2015. [17] M. E. Koksal, The Second order commutative pairs of a first order linear timevarying system, Applied Mathematics and Information Sciences, 9 (1) 1-6, 2015. [18] M. E. Koksal, Decomposition of a second-order linear time-varying differential system as the series connection of two first-order commutative pairs, Open Mathematics, 14, 693-704, 2016. [19] M. E. Koksal, Inverse commutativity conditions for second-order linear timevarying systems, Journal of Mathematics, 2017, 1-14, 2017. [20] M. E. Koksal, Transitivity property of commutativity for linear time varying analog systems, Submitted, arXiv: 1709.04477, 1-22, 2017. [21] M. E. Koksal, Explicit commutativity conditions for second-order linear timevarying systems with non-zero initial conditions, Submitted, arXiv: 1709.04403, 120, 2017. [22] C. A. Desoer, Notes For A Second Course On Linear Systems, Van Nostrand Rheinhold, New York, 1970. 26 

Weakly Supervised Object Detection with Pointwise Mutual Information Rene Grzeszick, Sebastian Sudholt, Gernot A. Fink TU Dortmund University, Germany arXiv:1801.08747v1 [cs.CV] 26 Jan 2018 {rene.grzeszick,sebastian.sudholt,gernot.fink}@tu-dortmund.de Abstract In this work a novel approach for weakly supervised object detection that incorporates pointwise mutual information is presented. A fully convolutional neural network architecture is applied in which the network learns one filter per object class. The resulting feature map indicates the location of objects in an image, yielding an intuitive representation of a class activation map. While traditionally such networks are learned by a softmax or binary logistic regression (sigmoid cross-entropy loss), a learning approach based on a cosine loss is introduced. A pointwise mutual information layer is incorporated in the network in order to project predictions and ground truth presence labels in a non-categorical embedding space. Thus, the cosine loss can be employed in this non-categorical representation. Besides integrating image level annotations, it is shown how to integrate point-wise annotations using a Spatial Pyramid Pooling layer. The approach is evaluated on the VOC2012 dataset for classification, point localization and weakly supervised bounding box localization. It is shown that the combination of pointwise mutual information and a cosine loss eases the learning process and thus improves the accuracy. The integration of coarse point-wise localizations further improves the results at minimal annotation costs. 1. Introduction The classification and localization of objects is one of the main tasks for the understanding of images. Much progress has been made in this field based on the recent developments in Convolutional Neural Networks (CNNs) [12]. The error rates in prominent tasks like ImageNet competitions have been reduced by a large margin over the last five years [16]. While the models become more and more powerful, the required data can still pose a bottleneck. In many localization tasks very detailed annotations are required in order to train a visual detector. Typically, an annotation is required that has the same level of detail as the desired output of the detector, e.g. bounding boxes or even pixel level annotations. As obtaining these annotations is expensive, weakly supervised learning approaches become of broader interest. These methods require a lower level of supervision during training. An object detector can be trained while only labeling images with respect to the presence or absence of certain objects. Similar to supervised tasks, great progress has been made in the field of weakly supervised learning by incorporating CNNs. State-of-the-art approaches in weakly supervised object detection use region proposals in order to localize objects. They evaluate these proposals based on a CNN that is solely trained on image level annotations. In [2] images are evaluated based on a pre-trained classification network, i.e. a VGG16 network that is pre-trained on ImageNet. The convolutional part of this network is evaluated on a given image, computing a feature map. Based on heuristic region proposals, e.g. from selective search or edge boxes, a set of candidate regions is cropped from the feature maps. These candidates are then processed by a Spatial Pyramid Pooling (SPP) [9] layer which is followed by fully connected layers. Only this part is trained in a weakly supervised fashion based on image level annotations and the relation between different candidate regions. In [10] a similar approach is followed. Here, two different loss functions are introduced which incorporate either an additive or subtractive center surround criterion for each candidate region. It has been shown that this allows for improving the results compared to [2]. While these approaches show state-of-the-art performance, the incorporation of region proposals often comes at a high computational cost and is based on heuristic design decisions and expert knowledge (cf. [7]). It has also been shown that with a deeper understanding of CNNs and their activations, the visualization of important filters can be leveraged for localizing objects. These approaches do not include additional region proposals so that they can be learned in an end-to-end fashion. A comparison of recent visualization approaches can be found in [17]. In [14] and [19] CNNs are trained on image level annotations for the task of object detection. The work in [14] applies max pooling for predicting the locations of objects in a weakly supervised manner. A multi-scale training approach is employed in order to learn the locations more accurately. The network training is performed using a binary logistic loss function (also known as cross-entropy loss) which in turn allows to predict a binary vector indicating the presence of multiple objects at once. A similar approach is followed in [19], but in contrast to [14], a global average pooling followed by a softmax is applied. In [19], it is argued that the global average pooling captures the extent of an object rather than just a certain part of an object. Based on the global average of the last filter responses, weights are computed which calculate the importance of each filter for a certain class. This can then be used in order to highlight the presence of certain classes in a so-called class activation map (CAM). These class specific activations show an object’s extent and can therefore be leveraged in order to predict objects in a weakly supervised manner. Besides different approaches for weakly supervised learning, there are a few methods that deal with learning from annotations which require a minimal annotation effort. In [11] CAMs are improved by adding micro annotations. Similar regions are grouped and manually labeled in order to remove false positive detections and obtain a more accurate localization. For example, trains are consistently co-occurring with tracks and thus often falsely recognized in the localization. In [1], point-wise annotations are introduced for the task of semantic segmentation. These provide a coarse localization of objects that also comes with a low annotation effort. It has been shown that the additional effort for point wise annotations is as low as approx. 2.1 sec. per image compared to image level annotations [1]. Such annotations may also provide an interesting cue of information to boost the performance of weakly supervised object detectors. Another interesting aspect of weakly supervised learning with CNNs are the loss functions. For example, in [14] a binary logistic loss function is used. Most prominently this loss is also used in tasks where multiple entities are predicted at once, as, for example, in attribute prediction [3, 8]. In [8], multiple scene attributes are recognized simultaneously in a CNN. It is shown that this approach outperforms traditional per attribute recognizers which are typically SVMs on top of heuristic feature representations or later on SVMs trained on CNN features [15, 20]. The simultaneous prediction is important as training multiple deep networks for each attribute is not suitable. Furthermore, the larger number of samples is an advantage for training. It can be assumed that the network also learns which attributes are typically appearing simultaneously within its feature representations. This idea is followed in [3], where an embedding is computed which encodes the mutual information between two attributes in the data, the pointwise mutual information (PMI) embedding. A CNN is trained based on the feature vectors of the embedding space. The predictions are then also made with respect to the embedding space. Given that this is a continuous, non-categorical, space, traditional softmax or binary logistic loss functions can no longer be used. Thus, a cosine loss function is employed for training the CNN. However, since the predictions are made in the embedding space, the presence of certain attributes can no longer be predicted in a straightforward manner. They are thus predicted using the cosine similarity between the networks output and vectors with a one-hot encoding that indicate the presence of a certain attribute. In this work a fully convolutional network that incorporates PMI for weakly supervised object detection is introduced. The network learns exactly one filter per object class and does not incorporate additional information such as region proposals. The contributions are as follows: The network incorporates a PMI embedding layer and is trained through a cosine loss, yet is still able to predict the presence of objects in a single forward pass. It is furthermore shown how to integrate annotations for either image level annotations or point-wise annotations using a SPP layer. 2. Method A fully convolutional network architecture is proposed which allows for object detection. An overview is given in Fig. 1. The network is designed to learn exactly one filter for each object class (see sec. 2.1). These filters are then followed by a SPP layer which allows for training the network in a weakly supervised fashion (cf. [14, 19]). For example, image level annotations correspond to the first level of the SPP, whereas coarse localizations can be encoded by using multiple levels which indicate the presence of an object in a certain region (as described in sec. 2.2). In order to account for co-occurrences of objects, an integrated learning step is proposed. A PMI layer, more precisely the positive pointwise mutual information (PPMI), is included in the network which projects the object prediction scores into a de-correlated feature space (see sec. 2.3). As the features to be learned in this feature space are in Rn and noncategorical, the cosine loss function is applied for training. The error is backpropagated through the network, including a backprojection of the PMI transformation so that the network still outputs scores for the presence of objects at the last convolutional feature map. 2.1. Fully Convolutional Network The proposed fully convolutional network architecture is similar to many other fully convolutional networks and based on the VGG networks [18]. Here, the fully connected layers of the VGG16 architecture are replaced by two additional convolution layers: one with 512 filters and one with exactly one filter per object class. Thus, instead of learning a global mapping of filters to object classes as in the CAM approach (cf. [14, 19]), the network learns exactly one filter ... of an image. For training the network based on image tags or point-wise annotations, the output is projected into an embedding space using the PPMI layer. The ground truth annotations are projected to the same embedding space and then a cosine loss is computed. For evaluation, a sigmoid layer can be used in order to derive probability scores from the class activations. Weakly supervised object detection can be performed by processing the response of each pixel of the class activation map by a sigmoid function. This results in probability scores which indicate the location of objects. ... 2.2. Integrating coarse point-wise localizations Figure 1: Overview of the proposed fully convolutional network architecture. During training both ground truth and predictions may be processed on image level or may include coarse localizations which are encoded by an Spatial Pyramid Pooling (SPP) layer. Both vectors are projected into an embedding space using the positive pointwise mutual information (PPMI) transformation which is derived from the training data. A cosine loss is computed for training. During testing, the output of the last convolutional layer can be either used for a presence prediction or a weakly supervised classification based on the class activation maps. which is responsible for indicating the presence of an object class at a certain location. This behavior is rather similar to networks for semantic segmentation [13]. The per class activations of the network are, therefore, easily interpretable, i.e., by a human user. For a classification task this map can be processed by a pooling layer in order to indicate the presence of an object in an image. Here, it is proposed to use a SPP layer that employs average pooling. This allows to compute a presence score for a global presence but also for certain regions Incorporating an SPP layer in the network architecture allows for encoding additional coarse localizations for an object’s presence within an image. The presence of an object can be encoded for each tile in the pyramid. Such an encoding can be combined with bounding boxes or with even simpler forms of annotations. Most interestingly, pointwise annotations allow to indicate the presence of an object in a certain region. A human annotator is asked to click on an object, therefore, indicating it’s presence by a single point within the image. These point-wise annotations require a minimal manual effort [1]. As each tile indicates the presence of an object in a certain region, the SPP approach will generate different levels of granularity. Each tile that contains a point of a certain object, is labeled with this object class being present. The feature vector that is used for training is the concatenation of multiple SPP tiles. An illustration is shown in Fig. 2. Therefore, when encoding the presence with a binary vector that shall be learned by the network, multiple co-occurrences are created within this vector. In the given example, the presence of a dog in the upper left, as well as the upper right tile of the image at the first level of the pyramid co-occurs with the presence of a dog at image level. This co-occurrence will always occur for the tiles at the image level and the finer levels of detail. 2.3. Encoding co-occurences with pointwise mutual information Due to the location encoding by the SPP layer, as well as the natural co-occurrences of objects in images, the binary label vectors will exhibit recurring co-occurrences. In order to take these into account, a feature space is computed that captures the likelihood that any two labels may co-occur in a given image. This feature space is then used for training the network. Following the idea of [3], the PMI can be computed in order to measure the mutual information between labels and find correlations within the data. Here, all object occurrences within the ground truth annotations of the training Object Presence Upper Left Upper Right 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 Lower Left 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 Lower Right 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 Image Level 0 0 0 0 0 0 0 0 0 0 0 1 0 0 0 0 0 0 0 0 Figure 2: Illustration of the point-wise labels and the resulting co-occurrences in the resulting pyramid feature vector. Here, the two dogs are annotated with points-wise annotations (in blue) in the upper left and upper right tile (indicated in red). The resulting feature vector indicates the presence of a certain class in the respective tiles using a one. A zero indicates it’s absence. The final representation is the concatenation of all tiles. data can be used for computing   p(Oi , Oj ) PMI = log p(Oi )p(Oj ) (i,j) , (1) where p(Oi , Oj ) represents the probability of object i and j to occur together, p(Oi ) and p(Oj ) are the priors for object i and j respectively. This can either be evaluated on image level or for the complete pyramid in the SPP. In contrast to [3], the PPMI, which is defined by PPMI = max(0, P M I) , (2) is used in the proposed approach. For object detection, the presence of objects occurring together is most important, ignoring negative correlations. This matrix can then be expressed by PPMI = U · Σ · U T (3) where Σ is a diagonal matrix with the eigenvalues in the √ diagonal. Then, E = U Σ is considered as a transformation matrix so that the P P M I = E · E T . In [3] this approach was not only used for computing a transformation, but also in order to reduce dimensionality. In the presented approach, the dimensionality is preserved, which is important in the following as it allows for reconstructing the original feature vector. Note that for unobserved co-occurrences of two object classes P (Oi , Oj ) equals zero. Therefore, the P M I matrix is not necessarily positive semidefinite so that the eigenvalues in Σ could become negative. Without reducing the dimensionality, it is therefore imperative to use the PPMI, which yields √ positive semidefinite matrix, instead of PMI as otherwise Σ could become complex. For projecting a feature vector x into the embedding space, the transformation E · x is applied. In order to integrate this into the CNN, an additional layer is introduced that implements the embedding E (see Fig. 1). The PPMI transformation is learned from the training samples before training the CNN. When training the CNN, the embedding matrix E is encoded in a single fully connected layer for which the weights are not updated during the stochastic gradient descent. This layer is used in order to project the scores as well as the ground truth annotations indicating an objects presence in the image or a certain region of the SPP representation into a new embedding space. In contrast to logistic regression, where a non-continuous space (binary vectors) is used, the embedding space is continuous. Since the features in this space are in Rn and non-categorical, the cosine loss function is computed between the networks output ŷ and a ground truth feature vector y: loss(ŷ, y) = 1 − ŷ T y ||ŷ|| · ||y|| (4) The cosine loss is chosen over the L2 -loss as it can be expected that distances between high-dimensional target vectors are better represented by an angle than the Euclidean distance which typically suffers from the curse of dimensionality. Moreover, the cosine loss computes to zero for an infinite amount of points for a given target while the L2 loss only equals to zero for a single point (the target itself). It is reasonable to assume that this trait benefits the learning process. When training the network with backpropagation, the PPMI layer computes a backprojection from the embedding space and reconstructs the original feature vector x. The error is, therefore, also evaluated with respect to the class scores. In [3] the network is trained solely in the embedding space and thus predicts a vector in the embedding space. The presence of an attribute had therefore to be predicted Annotation Detail Global Global Global Global (*) 2×2 2×2 2×2 Loss Loss Function Layer(s) Binary logistic – PPMI + Cosine – Oquab et. al. [14] (full images) Oquab et. al. [14] (weakly supervised) Binary logistic Finest Level Binary logistic Pyramid PPMI + Cosine Pyramid mAP image Level 2×2 76.9% 26.0% 80.3% 33.1% 76.0% – 81.8% – 75.5% 34.2% 77.1% 34.4% 82.1% 35.2% (*) An additional multi scale analysis is carried out. Table 1: Mean average precision for the classification in the VOC2012 dataset. based on the cosine distance between a binary vector indicating the presence of a single attribute and the PMI output. In the proposed approach, the class scores are directly obtained by a forward pass through the network. 3. Evaluation The proposed approach is evaluated on the VOC2012 dataset [5]. Additional coarse localizations are provided by the point-wise annotations published in [1]. These annotations were created by crowd workers, which were asked to indicate the location of an object by clicking on it. The approach is evaluated for three tasks. First, the classification task indicating the presence of an object in an image. Second, the point-wise localization, following the setup of [14], where the highest activation in a feature map is taken in order to predict a single point indicating the location of an object. Third, the weakly supervised localization is evaluated based on the correct localization (CorLoc) accuracy [4, 10]. While, for the first two tasks, the networks are trained on the train set and evaluated on the validation set of the VOC2012 benchmark (the test set is not available for the point-wise annotations), the CorLoc metric is typically evaluated on a training set and therefore evaluated on the complete trainval split of the dataset (cf. [10]). The training images are rescaled so that the shortest side is 512px in length. All networks are trained with a batch size of 256 for 2, 000 iterations, which equals 512, 000 images or 90/45 epochs on the train/trainval split. The first 600 iterations are trained with a learning rate of 0.0001 which is then increased to 0.001. Random data augmentations are applied, which include translation (up to 5%), rotation (up to 5 deg), Gaussian noise (σ = 0.02) and vertical mirroring. The networks that are trained with global, image level annotations are initialized using the ImageNet weights for the VGG16 networks. The networks which are trained with coarse point-wise localization are initialized using the weights from the image level training. 3.1. Classification Table 1 reports the classification accuracy on the VOC2012 dataset as the mean average precision (mAP). The accuracy is evaluated with respect to the presence of an object in an image or in any of the tiles of a 2 × 2 subdivision of the image. In the latter case, each tile is evaluated independently. The prediction scores for each tile are compared to the point-wise annotations and the average over all tiles is reported. A CNN using binary logistic loss is compared to one using the proposed combination of PPMI and a cosine loss. The results show that the performance on image level and also when using the highly correlated point-wise annotations can be improved by the PPMI embedding. It can also be seen that incorporating the coarse localizations which are derived from the point-wise annotations helps improving the image level results as well as the predictions for the more detailed 2 × 2 regions. When comparing the image level results to the ones published in [14], similar results are achieved. While outperforming the full image setup of [14] the results are slightly below the weakly supervised setup which applied an additional multi-scale analysis during training. Note that our training configuration is more similar to the full image setup as a fixed image size is used. 3.2. Localization For evaluating the localization accuracy the protocol designed in [14] is followed. The class-wise activations after the sigmoid computation are rescaled to the original image size. Using the maximum activation within a feature map, one detection point can be reported for each object class. A point prediction is considered as correct if it is within a ground truth bounding box (± 18px) of the same class, as in [14]. Each point is then associated with its respective probability score and the mAP is computed. The results are reported in Tab. 2. The PPMI embedding improves the results significantly compared to the binary Annotation Detail Global Global Global 2×2 BBoxes Loss Loss Function Layer(s) Binary logistic – PPMI + Cosine – Oquab et. al. [14] (weakly supervised) PPMI + Cosine Pyramid R-CNN [6]; results reported in [14] mAP localization 69.8% 76.5% 74.5% 78.1% 74.8% Table 2: Results of the point-wise localization, following the setup in [14]. logistic loss. Here, it can be seen that the proposed network provides a better localization result than the approach in [14]. Similar to the classification results, the additional coarse localizations allow for further improving the results at the cost of a minimal annotation effort. 3.3. CorLoc Last, the correct localization (CorLoc) accuracy has been evaluated. Here, the results are provided for the VOC2012 trainval set. Given an image and a target class, the CorLoc describes the percentage of images where a bounding box has been predicted that correctly localizes an object of the target class. An intersection over union (IoU) of 50% is required for a prediction to be considered as correct. For predicting a localization, the approach of [19] is followed. Note that the network is able to predict multiple localizations for different object classes at once. Given a target class, all pixels with an activation of more than 10% of the maximum activation for the target class are chosen as foreground pixels. The bounding box around the largest connected region is chosen as the object prediction. The results are shown in Tab. 3. The combination of PPMI and a cosine loss improves the localization by a margin compared to a binary logistic loss. Again, the coarse localizations are able to produce more precise results. Note that recently an approach has been proposed that achieves a CorLoc of 54.8% by incorporating selective search data and explicitly optimizing for bounding box predictions [10]. In contrast, the proposed approach is trained in an end-to-end fashion without additional selective search data. 3.4. Qualitative results Exemplary results are shown in Fig. 3. The examples are taken from the CNN that has been trained for the CorLoc on the VOC2012 trainval set. The network has been trained using the proposed combination of PPMI and a cosine loss. The annotations are coarse localizations derived from point-wise annotations for 2 × 2 tiles. The left column shows the input image with the bounding boxes of the target class shown in green. The middle column shows the predicted object region after thresholding the class activa- tion map. The right side shows the class activation map as derived from the network. The class activation map is the output of a single feature map where each pixel’s intensity has been processed by a sigmoid function. It can be observed that the desired objects are nicely localized in the feature maps. Even in the error case, the activations are reasonable as multiple screens are placed close to each other, making it difficult to distinguish them in a weakly supervised fashion. 4. Conclusion In this work a novel approach for weakly supervised object detection with CNNs has been proposed. The network incorporates the positive pointwise mutual information (PPMI) and a cosine loss function for learning. It is shown that this approach eases the learning process, improving the results compared to a binary logistic loss based on categorical feature vectors. A fully convolutional network architecture has been used for the weakly supervised detection. A single feature map is learned for each class, creating an intuitive representation for the presence of an object class in an image. Furthermore, an SPP layer is incorporated in the network instead of a global pooling operation. This allows for incorporating coarse localizations, i.e., in the form of point-wise annotations. These annotations require a minimal manual effort, but provide additional information that can be leveraged for weakly supervised localization. The evaluation on the VOC2012 dataset shows that the combination of PPMI and a cosine loss improves the results for classification, point localization as well as the CorLoc. Furthermore, the additional point-wise annotations helps in steering the learning process and further improve the results for all three tasks at a minimal annotation cost. 5. Acknowledgment This work has been supported by **** an anonymous institution *** . Annotation Detail Global Global Global (*) 2×2 Loss Loss Function Layer(s) Binary logistic – PPMI + Cosine – Kolesnikov et. al. [11] PPMI + Cosine Pyramid Initialization ImgNet ImgNet SPL1 (*) Requires additional selective search data CorLoc 26.2% 39.2% 54.8% 43.4% Table 3: CorLoc on the VOC2012 trainval set with an IoU of 50%. Figure 3: Qualitative results for the best performing network using 2×2 tiles for coarse localizations, derived from point-wise annotations. (left) Original images with annotated bounding boxes in green, predicted bounding boxes in blue and red for correct and incorrect predictions respectively. (middle) thresholded activations as used for bounding box computation. (right) class activation map as computed by the CNN. References [1] A. Bearman, O. Russakovsky, V. Ferrari, and L. Fei-Fei. Whats the point: Semantic segmentation with point supervision. In European Conference on Computer Vision (ECCV), pages 549–565. Springer, 2016. [2] H. Bilen and A. Vedaldi. Weakly supervised deep detection networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2846–2854, 2016. [3] F. Chollet. Information-theoretical label embeddings for large-scale image classification. arXiv preprint arXiv:1607.05691, 2016. [4] T. Deselaers, B. Alexe, and V. Ferrari. Weakly supervised localization and learning with generic knowledge. International Journal of Computer Vision (IJCV), 100(3):275–293, 2012. [5] M. Everingham, L. Van Gool, C. K. I. Williams, J. Winn, and A. Zisserman. The PASCAL Visual Object Classes Challenge 2011 (VOC2011) Results. http://www.pascalnetwork.org/challenges/VOC/voc2011/workshop/index.html, 2011. [6] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 580–587, 2014. [7] R. Girshick, J. Donahue, T. Darrell, and J. Malik. Regionbased convolutional networks for accurate object detection and segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(1):142–158, 2016. [8] R. Grzeszick, S. Sudholt, and G. A. Fink. Optimistic and pessimistic neural networks for scene and object recognition. 2016. [9] K. He, X. Zhang, S. Ren, and J. Sun. Spatial Pyramid Pooling in Deep Convolutional Networks for Visual Recognition. IEEE transactions on pattern analysis and machine intelligence, 37(9):1904–1916, 2015. [10] V. Kantorov, M. Oquab, C. M., and I. Laptev. Contextlocnet: Context-aware deep network models for weakly supervised localization. In Proc. European Conference on Computer Vision (ECCV), 2016. [11] A. Kolesnikov and C. H. Lampert. Improving weaklysupervised object localization by micro-annotation. Proc. British Machine Vision Conference (BMVC), 2016. [12] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature, 521(7553):436–444, 2015. [13] J. Long, E. Shelhamer, and T. Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3431–3440, 2015. [14] M. Oquab, L. Bottou, I. Laptev, and J. Sivic. Is object localization for free?-weakly-supervised learning with convolutional neural networks. In Proc. IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 685– 694, 2015. [15] G. Patterson, C. Xu, H. Su, and J. Hays. The sun attribute database: Beyond categories for deeper scene understanding. [16] [17] [18] [19] [20] International Journal of Computer Vision, 108(1-2):59–81, 2014. O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, A. C. Berg, and L. Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115(3):211–252, 2015. W. Samek, A. Binder, G. Montavon, S. Lapuschkin, and K.R. Müller. Evaluating the visualization of what a deep neural network has learned. IEEE Transactions on Neural Networks and Learning Systems, 2016. K. Simonyan and A. Zisserman. Very Deep Convolutional Networks for Large-Scale Image Recognition. CoRR, abs/1409.1, 2014. B. Zhou, A. Khosla, L. A., A. Oliva, and A. Torralba. Learning Deep Features for Discriminative Localization. In Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2016. B. Zhou, A. Khosla, A. Lapedriza, A. Torralba, and A. Oliva. Places: An image database for deep scene understanding. CoRR, abs/1610.02055, 2016. 

Prediction risk for the horseshoe regression Anindya Bhadra 1 Jyotishka Datta 2 Yunfan Li 1 Nicholas G. Polson 3 and Brandon Willard 3 Abstract arXiv:1605.04796v2 [math.ST] 13 Jun 2017 Predictive performance in shrinkage regression suffers from two major difficulties: (i) the amount of relative shrinkage is monotone in the singular values of the design matrix and (ii) the amount of shrinkage does not depend on the response variables. Both of these factors can translate to a poor prediction performance, the risk of which can be estimated unbiasedly using Stein’s approach. We show that using a component-specific local shrinkage term that can be learned from the data under a suitable heavy-tailed prior, in combination with a global term providing shrinkage towards zero, can alleviate both these difficulties and consequently, can result in an improved risk for prediction. Demonstrations of improved prediction performance over competing approaches in a simulation study and in a pharmacogenomics data set confirm our theoretical findings. Keywords: global-local priors; principal components; shrinkage regression; Stein’s unbiased risk estimate. 1 Introduction Prediction using shrinkage regression techniques such as ridge regression (Hoerl and Kennard, 1970) and principal components regression or PCR (Jolliffe, 1982) remain popular in high-dimensional problems. Shrinkage methods enjoy a number of advantages over selection-based methods such as the lasso (Tibshirani, 1996) and comfortably outperform them in predictive performance in certain situations. Prominent among these is when the predictors are correlated and the resulting lasso estimate is unstable, but ridge or PCR estimates are not (see, e.g, the discussion in Chapter 3 of Hastie et al., 2009). Polson and Scott (2012a) showed, following a representation originally devised by Frank and Friedman (1993), that many commonly used high-dimensional shrinkage regression estimates, such as the estimates of ridge regression, regression with g-prior (Zellner, 1986) and PCR, can be viewed as posterior means under a unified framework of “global” shrinkage prior on the regression coefficients that are suitably orthogonalized. They went on to demonstrate these global shrinkage regression models suffer from two major difficulties: (i) the amount of relative shrinkage is monotone in the singular values of the design matrix and (ii) the amount of shrinkage does not depend on the response variables. Both of these factors can contribute to poor out of sample prediction performance, which they demonstrated numerically. Polson and Scott (2012a) further provided numerical evidence that both of these difficulties mentioned above can be resolved by allowing a “local,” component-specific shrinkage term that 1 Address: Department of Statistics, Purdue University, 250 N. University St., West Lafayette, IN 47907, USA. Department of Mathematical Sciences, University of Arkansas, Fayetteville, AR 72701, USA. 3 Address: Booth School of Business, The University of Chicago, 5807 S. Woodlawn Ave., Chicago, IL 60637, USA. 2 Address: 1 can be learned from the data, in conjunction with a global shrinkage parameter as used in ridge or PCR, giving rise to the so-called “global-local” shrinkage regression models. Specifically, Polson and Scott (2012a) demonstrated by simulations that using the horseshoe prior of Carvalho et al. (2010) on the regression coefficients performed well over a variety of competitors in terms of predictive performance, including the lasso, ridge, PCR and sparse partial least squares (Chun and Keles, 2010). However, a theoretical investigation of the conditions required for the horseshoe regression model to outperform a global shrinkage regression model such as ridge or PCR in terms of predictive performance has been lacking. The goal of the current work is to bridge this methodological and theoretical gap by developing formal tools for comparing the predictive performances of shrinkage regression methods. Developing a formal measure to compare predictive performance of competing regression methods is important in both frequentist and Bayesian settings. This is because the frequentist tuning parameter or the Bayesian hyper-parameters can then be chosen to minimize the estimated prediction risk, if prediction of future observations is the main modeling goal. A measure of quadratic risk for prediction in regression models can be obtained either through model-based covariance penalties or through nonparametric approaches. Examples of covariance penalties include Mallow’s C p (Mallows, 1973), Akaike’s information criterion (Akaike, 1974), risk inflation criterion (Foster and George, 1994) and Stein’s unbiased risk estimate or SURE (Stein, 1981). Nonparametric penalties include the generalized cross validation of Craven and Wahba (1978), which has the advantage of being model free but usually produces a prediction error estimate with high variance (Efron, 1983). The relationship between the covariance penalties and nonparametric approaches were further explored by Efron (2004), who showed the covariance penalties to be a Rao-Blackwellized version of the nonparametric penalties. Thus, Efron (2004) concluded that model-based penalties such as SURE or Mallow’s C p (the two coincide for models where the fit is linear in the response variable) offer substantially lower variance in estimating the prediction error, assuming of course the model is true. From a computational perspective, calculating SURE, when it is explicitly available, is substantially less burdensome than performing cross validation, which usually requires several Monte Carlo replications. Furthermore, SURE, which is a measure of quadratic risk in prediction, also has connections with the Kullback–Leiber risk for the predictive density (George et al., 2006). Given these advantages enjoyed by SURE, we devise a general, explicit and numerically stable technique for computing SURE for regression models that can be employed to compare the performances of global as well as horseshoe regressions. The key technique to our innovation is an orthogonalized representation first employed by Frank and Friedman (1993), which results in particularly simple and numerically stable formulas for SURE. Using the developed tools for SURE, we demonstrate that the suitable conditions for success of the horseshoe regression model over global regression models in prediction arise when a certain sparse-robust structure is present in the orthogonalized regression coefficients. Specifically, our major finding is that when a certain principal component corresponding to a low singular value of the design matrix is a strong predictor of the outcomes, global shrinkage methods necessarily shrink these components too much, whereas the horseshoe does not. This results in a substantially increased risk for global regression over the horseshoe regression, explaining why global-local shrinkage such as the horseshoe can overcome the two major difficulties encountered by global shrinkage regression methods. 2 The rest of the article is organized as follows. In Section 2, we demonstrate how several standard shrinkage regression estimates can be reinterpreted as posterior means in an orthogonalized representation of the design matrix. Using this representation, we derive explicit expressions for SURE for global shrinkage and horseshoe regressions in Sections 3 and 4 respectively. Section 5 compares the actual prediction risk (as opposed to data-dependent estimates of the risk, such as SURE) of global and horseshoe regressions and explicitly identifies some situations where the horseshoe regression can outperform global shrinkage methods. A simulation study is presented in 6 and prediction performance of several competing approaches are assessed in a pharmacogenomics data set in Section 7. We conclude by pointing out some possible extensions of the current work in Section 8. 2 Shrinkage regression estimates as posterior means Consider the high-dimensional regression model y = Xβ + e, (1) where y ∈ Rn , X ∈ Rn× p , β ∈ R p and e ∼ N (0, σ2 In ) with p > n. Let X = UDW T be the singular value decomposition of the design matrix. Let D = diag(d1 , . . . , dn ) with d1 ≥ . . . ≥ dn > 0 and Rank( D ) = min(n, p) = n. Define Z = UD and α = W T β. Then the regression problem can be reformulated as: y = Zα + e. (2) The ordinary least squared (OLS) estimate of α is α̂ = ( Z T Z )−1 Z T y = D −1 U T y. Following the original results by Frank and Friedman (1993), several authors have used the well-known orthogonalization technique (Clyde et al., 1996; Denison and George, 2012; Polson and Scott, 2012a) to demonstrate that the estimates of many shrinkage regression methods can be expressed in terms of the posterior mean of the “orthogonalized” regression coefficients α under the following hierarchical model: ind (α̂i | αi , σ2 ) ∼ N (αi , σ2 di−2 ), ind (αi | σ2 , τ 2 , λ2i ) ∼ N (0, σ2 τ 2 λ2i ), (3) (4) with σ2 , τ 2 > 0. The global term τ controls the amount of shrinkage and the fixed λ2i terms depend on the method at hand. Given λi and τ, the estimate for β under the global shrinkage prior, denoted by β̃, can be expressed in terms of the posterior mean estimate for α as follows: α̃i = τ 2 λ2i d2i α̂i , 1 + τ 2 λ2i d2i n β̃ = ∑ α̃i wi , (5) i =1 where α̃i = E(αi | τ, λ2i , X, y); wi is a p × 1 vector and is the ith column of the p × n matrix W and the term τ 2 λ2i d2i /(1 + τ 2 λ2i d2i ) ∈ (0, 1) is the shrinkage factor. The expression from Equation (5) makes it clear that it is the orthogonalized OLS estimates α̂i s that are shrunk. We shall show that this orthogonalized representation is also particularly suitable for calculating the prediction risk 3 estimate. The reason is tied to the independence assumption that is now feasible in Equations (3) and (4). To give a few concrete examples, we note below that several popular shrinkage regression models fall under the framework of Equations (3–4): 1. For ridge regression, λ2i = 1, ∀i, and we have α̃i = τ 2 d2i α̂i /(1 + τ 2 d2i ). 2. For K component PCR, λ2i is infinite for the first K components and then 0. Thus, α̃i = α̂i for i = 1, . . . , K and α̃i = 0 for i = K + 1, . . . , n. 3. For regression with g-prior, λ2i = di−2 and we have α̃i = τ 2 α̂i /(1 + τ 2 ) for i = 1, . . . , n. This shows the amount of relative shrinkage α̃i /α̂i is constant in di for PCR and g-prior and is monotone in di for ridge regression. In none of these cases it depends on the OLS estimate α̂i (consequently, on y). In the next section we quantify the effect of this behavior on the prediction risk estimate. 3 Stein’s unbiased risk estimate for global shrinkage regression Define the fit ỹ = X β̃ = Z α̃, where α̃ is the posterior mean of α. As noted by Stein (1981), the fitted risk is an underestimation of the prediction risk, and SURE for prediction is defined as: n ∂ỹi , ∂yi i =1 SURE = ||y − ỹ||2 + 2σ2 ∑ where the ∑in=1 (∂ỹi /∂yi ) term is also known as the “degrees of freedom” (Efron, 2004). By Tweedie’s formula (Masreliez, 1975; Pericchi and Smith, 1992) that relates the posterior mean with the marginals; we have for a Gaussian model of Equations (3–4) that: α̃ = α̂ + σ2 D −2 ∇α̂ log m(α̂), where m(α̂) is the marginal for α̂. Noting y = Z α̂ yields ỹ = y + σ2 UD −1 ∇α̂ log m(α̂). Using the independence of αi s, the formula for SURE becomes n SURE =σ4 ∑ di−2 i =1  ∂ log m(α̂i ) ∂α̂i 2  n  ∂2 + 2σ2 ∑ 1 + σ2 di−2 2 log m(α̂i ) . ∂α̂i i =1 (6) Thus, the prediction risk estimate for shrinkage regression can be quantified in terms of the first two derivatives of the log marginal for α̂. Integrating out αi from Equations (3–4) yields in all these cases, ind (α̂i | σ2 , τ 2 , λ2i ) ∼ N (0, σ2 (di−2 + τ 2 λ2i )). The marginal of α̂ is given by m(α̂) ∝ n ∏ exp i =1 ( − α̂2i /2 σ2 (di−2 + τ 2 λ2i ) ) , which yields ∂2 log m(α̂i ) −1 = 2 −2 . 2 ∂α̂i σ (di + τ 2 λ2i ) −α̂ ∂ log m(α̂i ) = 2 −2 i 2 2 ; ∂α̂i σ ( di + τ λi ) 4 (7) Therefore, Equation (6) reduces to the following expression for SURE for global shrinkage regressions: SURE = ∑in=1 SUREi , where, SUREi = α̂2i d2i τ 2 λ2i d2i 2 + 2σ . (1 + τ 2 λ2i d2i )2 (1 + τ 2 λ2i d2i ) (8) From a computational perspective, the expression in Equation (8) is attractive, as it avoids costly matrix inversions. For a given σ one can choose τ to minimize the prediction risk, which amounts to a one-dimensional optimization. Note that in our notation, d1 ≥ d2 . . . ≥ dn > 0. Clearly, this is the SURE when λi s are fixed and finite (e.g., ridge regression). For K component PCR, only the first K terms appear in the sum. The di terms are features of the design matrix X and one may try to control the prediction risk by varying τ. When τ → ∞, SURE → 2nσ2 , the risk of prediction with ordinary least squares (unbiased). When τ → 0, we get the mean-only model (zero variance): SURE → ∑in=1 α̂2i d2i . Regression models with τ ∈ (0, ∞) represent a bias-variance tradeoff. Following are the two major difficulties of global shrinkage regression. 1. Note from the first term of Equation (8) that SURE is increased by those components for which α̂2i d2i is large. Choosing a large τ alleviates this problem, but at the expense of an SUREi of 2σ2 even for components for which α̂2i d2i is small (due to the second term in Equation (8)). Thus, it might be beneficial to differentially minimize the effect of the components for which α̂2i d2i is large, while ensuring those for which α̂2i d2i is small make a contribution less than 2σ2 to SURE. Yet, regression models with λi fixed, such as ridge, PCR, regression with g-priors, provide no mechanism for achieving this, since the relative shrinkage, defined as the ratio α̃i /α̂i , equals τ 2 λ2i d2i /(1 + τ 2 λ2i d2i ), and is solely driven by a single quantity τ. 2. Equation (5) shows that the relative shrinkage for α̂i is monotone in di ; that is, those α̂i corresponding to a smaller di are necessarily shrunk more (in a relative amount). This is only sensible in the case where one has reasons to believe the low variance eigen-directions (i.e., principal components) of the design matrix are not important predictors of the response variables, an assumption that can be violated in real data (Polson and Scott, 2012a). In the light of these two problems, we proceed to demonstrate that putting a heavy-tailed prior on λi s, in combination with a suitably small value of τ to enable global-local shrinkage can resolve both these issues. The intuition behind this is that a small value of a global parameter τ enables shrinkage towards zero for all the components while the heavy tails of the local or componentspecific λi terms ensure the components with large values of α̂i di are not shrunk too much, and allow the λi terms to be learned from the data. Simultaneously ensuring both of these factors helps in controlling the prediction risk for both the noise as well as the signal terms. 4 Stein’s unbiased risk estimate for the horseshoe regression The global-local horseshoe shrinkage regression of Polson and Scott (2012a) extends the global shrinkage regression models of the previous section by putting a local (component-specific), heavytailed half-Cauchy prior on the λi terms that allow these terms to be learned from the data, in 5 addition to a global τ. The model equations become: ind (α̂i | αi , σ2 ) ∼ N (αi , σ2 di−2 ), (9) ind (αi | σ2 , τ 2 , λ2i ) ∼ N (0, σ2 τ 2 λ2i ), λi (10) ind ∼ C + (0, 1), (11) with σ2 , τ 2 > 0 and C + (0, 1) denotes a standard half-Cauchy random variable with density p(λi ) = (2/π )(1 + λ2i )−1 . The marginal prior on αi s that is obtained as a normal scale mixture by integrating out λi s from Equations (10) and (11) is called the horseshoe prior (Carvalho et al., 2010). Improved mean square error over competing approaches in regression has been empirically observed by Polson and Scott (2012a) with horseshoe prior on αi s. The intuitive explanation for this improved performance is that a heavy tailed prior of λi leaves the large αi terms of Equation (10) un-shrunk in the posterior, whereas the global τ term provides shrinkage towards zero for all components (see, for example, the discussion by Bhadra et al., 2016b; Carvalho et al., 2010; Polson and Scott, 2012b, and the references therein). However, no explicit formulation of the prediction risk under horseshoe shrinkage is available so far and we demonstrate below the heavy-tailed priors on λi terms, in addition to a global τ, can be beneficial in controlling the overall prediction risk. Under the model of Equations (9–11), after integrating out αi from the first two equations, we have, ind (α̂i | σ2 , τ 2 , λ2i ) ∼ N (0, σ2 (di−2 + τ 2 λ2i )). We have, p(λi ) ∝ 1/(1 + λ2i ). Thus, the marginal of α̂, denoted by m(α̂), is given up to a constant of proportionality by n m(α̂) = ∏ Z ∞ i =1 0 N (α̂i | 0, σ2 (di−2 + τ 2 λ2i )) p(λi )dλi 2 −n/2 ∝(2πσ ) n ∏ Z ∞ i =1 0 α̂2 d2 /2 exp − 2 i i 2 2 2 σ (1 + τ d i λ i )   di 2 (1 + τ d2i λ2i )1/2 1 dλi . 1 + λ2i (12) This integral involves the normalizing constant of a compound confluent hypergeometric distribution that can be computed using a result of Gordy (1998). PROPOSITION 4.1. (Gordy, 1998). The compound confluent hypergeometric (CCH) density is given by CCH( x; p, q, r, s, ν, θ ) = x p−1 (1 − νx )q−1 {θ + (1 − θ )νx }−r exp(−sx ) , B( p, q) H ( p, q, r, s, ν, θ ) for 0 < x < 1/ν, where the parameters satisfy p > 0, q > 0, r ∈ R, s ∈ R, 0 ≤ ν ≤ 1 and θ > 0. Here B( p, q) is the beta function and the function H (·) is given by H ( p, q, r, s, ν, θ ) = ν− p exp(−s/ν)Φ1 (q, r, p + q, s/ν, 1 − θ ), 6 where Φ1 is the confluent hypergeometric function of two variables, given by ∞ Φ1 (α, β, γ, x1 , x2 ) = ∞ (α)m ( β)n m n x x , ( γ )m+n m!n! 1 2 m =0 n =0 ∑ ∑ (13) where ( a)k denotes the rising factorial with ( a)0 = 1, ( a)1 = a and ( a)k = ( a + k − 1)( a)k−1 . We present our first result in the following theorem and show that the marginal m(α̂) and all its derivatives lend themselves to a series representation in terms of the first and second moments of a random variable that follows a CCH distribution. Consequently, we quantify SURE for the horseshoe regression. THEOREM 4.1. Denote m0 (α̂i ) = (∂/∂α̂i )m(α̂i ) and m00 (α̂i ) = (∂2 /∂α̂2i )m(α̂i ). Then, the following holds. A. SURE for the horseshoe shrinkage regression model defined by Equations (9–11) is given by SURE = ∑in=1 SUREi , where the component-wise contribution SUREi is given by SUREi = 2σ 2 − σ4 di−2  m0 (α̂i ) m(α̂i ) 2 + 2σ4 di−2 m00 (α̂i ) . m(α̂i ) (14) B. Under independent standard half-Cauchy prior on λi s, for the second and third terms in Equation (14) we have: α̂i d2 m0 (α̂i ) = − 2i E( Zi ), m(α̂i ) σ α̂2 d4 d2 m00 (α̂i ) = − i2 E( Zi ) + i 4 i E( Zi2 ), m(α̂i ) σ σ and where ( Zi | α̂i , σ, τ ) follows a CCH( p = 1, q = 1/2, r = 1, s = α̂2i d2i /2σ2 , v = 1, θ = 1/τ 2 d2i ) distribution. A proof is given in Appendix A.1. Theorem 4.1 provides a computationally tractable mechanism for calculating SURE for the horseshoe shrinkage regression in terms of the moments of CCH random variables. Gordy (1998) provides a simple formula for all integer moments of CCH random variables. Specifically, he shows if X ∼ CCH( x; p, q, r, s, ν, θ ) then E( X k ) = ( p)k H ( p + k, q, r, s, ν, θ ) , ( p + q)k H ( p, q, r, s, ν, θ ) (15) for integers k ≥ 1. Moreover, as demonstrated by Gordy (1998), these moments can be numerically evaluated quite easily over a range of parameter values and calculations remain very stable. A consequence of this explicit formula for SURE is that the global shrinkage parameter τ can now be chosen to minimize SURE by performing a one-dimensional numerical optimization. Another consequence is that an application of Theorem 3 of Carvalho et al. (2010) shows m0 (α̂i ) ∂ log m(α̂i ) = lim = 0, ∂α̂i |α̂i |→∞ m ( α̂i ) |α̂i |→∞ lim with high probability, where m(α̂i ) is the marginal under the horseshoe prior. Recall that the posterior mean α̃i and the OLS estimate α̂i are related by Tweedie’s formula as α̃i = α̂i + σ2 di−2 ∂ log m(α̂i )/∂α̂i . 7 Thus, α̃i ≈ α̂i , with high probability, as |α̂i | → ∞, for any fixed di and σ for the horseshoe regression. Since α̂i is unbiased for αi , the resultant horseshoe posterior mean is also seen to be unbiased when |α̂i | is large. Compare with the resultant α̃i for global shrinkage regression of Equation (5), which is monotone decreasing in di , and therefore can be highly biased if a true large |αi | corresponds to a small di . Perhaps more importantly, we can use the expression from Theorem 4.1 to estimate the prediction risk of the horseshoe regression for the signal and the noise terms. First we treat the case when |α̂i | is large. We have the following result. THEOREM 4.2. Define si = α̂2i d2i /2σ2 and θi = (τ 2 d2i )−1 . For any si ≥ 1, θi ≥ 1, we have for the horseshoe regression of Equations (9–11) that ( !) ) ( 2 C (1 + s i ) SURE C ( 1 + s ) 2 1 i i 1 − θi (C̃1 + C̃2 ) , − θi2 (C̃1 + C̃2 )2 ≤ ≤ 1 + 2θi (1 + si ) + 3/2 2σ2 s2i s2i s3i si almost surely, where C1 = {1 − 5/(2e)}−1/2 ≈ 3.53, C2 = 16/15, C̃1 = (1 − 2/e) ≈ 0.26, C̃2 = 4/3, are constants. A proof is given in Appendix A.2. Our result is non-asymptotic, i.e., it is valid for any si ≥ 1. However, an easy consequence is that SUREi → 2σ2 , almost surely, as si → ∞, provided τ 2 ≤ di−2 . An intuitive explanation of this result is that component-specific shrinkage is feasible in the horseshoe regression model due to the heavy-tailed λi terms, which prevents the signal terms from getting shrunk too much and consequently, making a large contribution to SURE due to a large bias. With just a global parameter τ, this component-specific shrinkage is not possible. A comparison of SUREi resulting from Theorem 4.2 with that from Equation (8) demonstrates using global-local horseshoe shrinkage, we can rectify a major shortcoming of global shrinkage regression, in that the terms with large si do not make a large contribution to the prediction risk. Moreover, the main consequence of Theorem 4.2, that is SUREi → 2σ2 , almost surely, as si → ∞, holds for a larger class of “global-local” priors, of which the horseshoe is a special case. THEOREM 4.3. Consider the hierarchy of Equations (9–10) and suppose the prior on λi in Equation (11) satisfies p(λ2i ) ∼ (λ2i ) a−1 L(λ2i ) as λ2i → ∞, where f ( x ) ∼ g( x ) means limx→∞ f ( x )/g( x ) = 1. Assume a ≤ 0 and L(·) is a slowly-varying function, defined as lim| x|→∞ L(tx )/L( x ) = 1 for all t ∈ (0, ∞). Then we have SUREi → 2σ2 , almost surely, as si → ∞. A proof is given in Appendix A.3. Densities that satisfy p(λ2i ) ∼ (λ2i ) a−1 L(λ2i ) as λ2i → ∞ are sometimes called regularly varying or heavy-tailed. Clearly, the horseshoe prior is a special case, since for the standard half-Cauchy we have p(λi ) ∝ 1/(1 + λ2i ), which yields by a change of variables p(λ2i ) = (λ2i )−3/2 {λ2i /(1 + λ2i )}, which is of the form (λ2i ) a−1 L(λ2i ) with a = −1/2 since L(λ2i ) = λ2i /(1 + λ2i ) is seen to be slowly-varying. Other priors that fall in this framework are the horseshoe+ prior of Bhadra et al. (2016b), for which p(λi ) ∝ log(λi )/(λ2i − 1) = λi−2 L(λ2i ) with L(λ2i ) = log(λi )λ2i /(λ2i − 1). Ghosh et al. (2016) show that the generalized double Pareto prior (Armagan et al., 2013) and the three parameter beta prior (Armagan et al., 2011) also fall in this framework. Thus, Theorem 4.3 generalizes the main consequence of Theorem 4.2 to a broader class of priors in the asymptotic sense as si → ∞. Next, for the case when |α̂i | is small, we have the following result for estimating the prediction risk of the horseshoe regression. 8 THEOREM 4.4. Define si = α̂2i d2i /2σ2 and θi = (τ 2 d2i )−1 . Then the following statements are true (almost surely) for the horseshoe regression. A. SUREi is an increasing function of si in the interval si ∈ [0, 1] for any fixed τ. B. When si = 0, we have that SUREi is a monotone increasing function of τ, and is bounded in the interval (0, 2σ2 /3], almost surely, when τ 2 d2i ∈ (0, 1]. C. When si = 1, we have that SUREi is bounded in the interval (0, 1.93σ2 ], almost surely, when τ 2 d2i ∈ (0, 1]. A proof is given in Appendix A.4. This theorem establishes that: (i) the terms with smaller si in the interval [0, 1] contribute less to SURE, with the minimum achieved at si = 0 (these terms can be thought of as the noise terms) and (ii) if τ is chosen to be sufficiently small, the terms for which si = 0, has an upper bound on SURE at 2σ2 /3. Note that the OLS estimator has risk 2σ2 for these terms. At si = 0, the PCR risk is either 0 or 2σ2 , depending on whether the term is or is not included. A commonly used technique for shrinkage regressions is to choose the global τ to minimize a data-dependent estimate of the risk, such as CL or SURE (Mallows, 1973). The ridge regression SURE at si = 0 is an increasing function of τ and thus, it might make sense to choose a small τ if all si terms were small. However, in the presence of some si terms that are large, ridge regression cannot choose a very small τ, since the large si terms will then be heavily shrunk and contribute too much to SURE. This is not the case with global-local shrinkage regression methods such as the horseshoe, which can still choose a small τ to mitigate the contribution from the noise terms and rely on the heavy-tailed λi terms to ensure large signals are not shrunk too much. Consequently, the ridge regression risk estimate is usually larger than the global-local regression risk estimate even for very small si terms, when some terms with large si are present along with mostly noise terms. At this point, the results concern the risk estimate (i.e., SURE) rather than risk itself, the discussion of which is deferred until Section 5. To summarize the theoretical findings, Theorem 4.2 together with Theorem 4.4 establishes that the horseshoe regression is effective in handling both very large and very small values of α̂2i d2i . Specifically, Theorem 4.4 asserts that a small enough τ shrinks the noise terms towards zero, minimizing their contribution to SURE. Whereas, according to Theorem 4.2, the heavy tails of the Cauchy priors for the λi terms ensure the large signals are not shrunk too much and ensures a SURE of 2σ2 for these terms, which is an improvement over purely global methods of shrinkage. 5 Prediction risk for the global and horseshoe regressions In this section we compare the theoretical prediction risks of global and global-local horseshoe shrinkage regressions. While SURE is a data-dependent estimate of the theoretical risk, these two quantities are equal in expectation. We use a concentration argument to derive conditions under which the horseshoe regression will outperform global shrinkage regression, e.g., ridge regression, in terms of predictive risk. While the analysis seems difficult for an arbitrary design matrix X, we are able to treat the case of ridge regression for orthogonal design, i.e., X T X = I. Clearly, if the SVD of X is written as X = UDV T , then we have D = I and for ridge regression λi = 1 for all i. 9 Thus, for orthogonal design, Equations (3) and (4) become ind (α̂i | αi , σ2 ) ∼ N (αi , σ2 ), ind (αi | σ2 , τ 2 , λ2i ) ∼ N (0, σ2 τ 2 ), where τ is the global shrinkage parameter. Since the fit in this model is linear in α̂i , SURE is equivalent to Mallow’s CL . Equation (14) of Mallows (1973) shows that if τ is chosen to minimize CL , then the optimal ridge estimate is given in closed form by αi? =  1− nσ2 n ∑i=1 α̂2i  α̂i . Alternatively, the solution can be directly obtained from Equation (8) by taking di = λi = 1 for all i and by setting τ ? = argminτ ∑in=1 SUREi . It is perhaps interesting to note that this “optimal” ridge estimate, where the tuning parameter is allowed to depend on the data, is no longer linear in α̂. In fact, the optimal solution α? can be seen to be closely related to the James–Stein estimate of α and its risk can therefore be quantified using the risk bounds on the James–Stein estimate. As expected due to the global nature of ridge regression, the relative shrinkage αi? /α̂i of the optimal solution only depends on |α̂|2 = ∑in=1 α̂2i but not on the individual components of α̂. Theorem 1 of Casella and Hwang (1982) shows that n−2 R(α, α? ) ( n − 2)2 1− ≤ ≤ 1 − n + | α |2 R(α, α̂) n  1 n − 2 + | α |2  . Consequently, if |α|2 /n → c as n → ∞ then the James–Stein estimate satisfies lim n→∞ R(α, α? ) c = . R(α, α̂) c+1 Thus, α? offers large benefits over the least squares estimate α̂ for small c but it is practically equivalent to the least squares estimate for large c. The prediction risk of the least squares estimate for p > n is simply 2nσ2 , or an average component-specific risk of 2σ2 . We show that when true αi = 0, the component-specific risk bound of the horseshoe shrinkage regression is less than 2σ2 . We have the following result. THEOREM 5.1. Let D = I and let the global shrinkage parameter in the horseshoe regression be τ 2 = 1. When true αi = 0, an upper bound of the component-wise risk of the horseshoe regression is 1.75σ2 < 2σ2 . A proof can be found in Appendix A.5. The proof uses the fact that the actual risk can be obtained by computing the expectation of SURE. We split the domains of integration into three distinct regions and use the bounds on SURE from Theorems 4.2 and 4.4, as appropriate. When true αi is large enough, a consequence of Theorem 4.2 is that the component-specific risk for global-local shrinkage regression is 2σ2 . This is because SURE in this case is almost surely equal to 2σ2 and α̂i is concentrated around true αi . Therefore, it is established that if only a few components of true α are large and the rest are zero in such a way that |α|2 /n is large, then the global-local horseshoe regression outperforms ridge regression in terms of predictive risk. The benefit arises from a lower risk for the αi = 0 terms. On the other hand, if all components of true 10 Table 1: The true orthogonalized regression coefficients α0i , their ordinary least squared estimates α̂i , and singular values di of the design matrix, for n = 100 and p = 500. i α0i α̂i di α̂i di 1 2 ... 5 6 ... 29 30 ... 56 57 ... 66 67 ... 95 96 ... 100 0.10 -0.44 ... -0.13 10.07 ... 0.46 10.47 ... 0.35 10.23 ... -0.00 11.14 ... -0.82 9.60 ... 0.61 0.10 -0.32 ... 0.30 10.22 ... 0.60 11.07 ... 0.57 10.66 ... -0.35 11.52 ... -0.56 10.21 ... 0.91 635.10 3.16 ... 3.05 3.02 ... 2.53 2.51 ... 2.07 2.07 ... 1.90 1.88 ... 1.42 1.40 ... 1.27 62.13 -1.00 ... 0.91 30.88 ... 1.53 27.76 ... 1.18 22.05 ... -0.66 21.70 ... -0.79 14.26 ... 1.15 α are zero or all are large, the horseshoe regression need not outperform ridge regression. 6 Numerical examples We simulate data where n = 100, and consider the cases p = 100, 200, 300, 400, 500. Let B be a p × k factor loading matrix, with all entries equal to 1. Let Fi be k × 1 matrix of factor values, with all entries drawn independently from N (0, 1). The ith row of the n × p design matrix X is generated by a factor model, with number of factors k = 8, as follows: Xi = BFi + ξ i , ξ i ∼ N (0, 0.1), for i = 1, . . . , n. Thus, the columns of X are correlated. Let X = UDW T denote the singular value decomposition of X. The observations y are generated from Equation (2) with σ2 = 1, where for the true orthogonalized regression coefficients α0 , the 6, 30, 57, 67, and 96th components are randomly selected as signals, and the remaining 95 components are noise terms. Coefficients of the signals are generated by a N (10, 0.5) distribution, and coefficients of the noise terms are generated by a N (0, 0.5) distribution. For the case n = 100 and p = 500, some of the true orthogonalized regression coefficients α0 , their ordinary least squared estimates α̂, and the corresponding singular values d of the design matrix, are shown in Table 1. Table 2 lists the SURE for prediction and actual out of sample sum of squared prediction error (SSE) for the ridge, lasso, PCR and horseshoe regressions. Out of sample prediction error of 11 Table 2: SURE and average out of sample prediction SSE (standard deviation of SSE) on one training set and 200 testing sets for the competing methods for n = 100, for ridge regression (RR), the lasso regression (LASSO), the adaptive lasso (A LASSO), principal components regression (PCR) and the horseshoe regression (HS). The lowest SURE in each row is in italics and the lowest average prediction SSE is in bold. A formula for SURE is unavailable for the adaptive lasso. RR p SURE SSE 100 159.02 200 187.38 300 192.78 400 195.02 500 196.11 168.24 (23.87) 174.92 (21.13) 191.91 (22.95) 182.55 (22.70) 188.78 (22.33) LASSO SURE SSE 125.37 140.99 147.83 148.56 159.95 128.98 (18.80) 132.46 (18.38) 145.04 (19.89) 165.63 (21.55) 159.56 (19.94) A LASSO SSE 127.22 (18.10) 151.89 (20.47) 153.64 (21.19) 178.98 (20.12) 186.23 (23.50) PCR SURE SSE SURE SSE 162.23 120.59 126.33 (18.77) 126.99 (17.29) 136.67 (18.73) 143.91 (18.41) 160.11 (20.29) 213.90 260.65 346.19 386.50 179.81 (25.51) 191.33 (22.62) 253.00 (26.58) 292.02 (28.98) 366.88 (39.38) HS 139.32 151.24 147.69 144.97 the adaptive lasso is also included in the comparisons, although we are unaware of a formula for computing the SURE for the adaptive lasso. SURE for ridge and PCR can be computed by an application of Equation (8) and SURE for the horseshoe regression is given by Theorem 4.1. SURE for the lasso is calculated using the result given by Tibshirani and Taylor (2012). In each case, the model is trained on 100 samples. We report the SSE on 100 testing samples, averaged over 200 testing data sets, and their standard deviations. For ridge, lasso, PCR and horseshoe regression, the global shrinkage parameters were chosen to minimize SURE for prediction. In adaptive lasso, the shrinkage parameters were chosen by cross validation due to SURE being unavailable. It can be seen that SURE in most cases are within one standard deviation of the actual out of sample prediction SSE, suggesting SURE is an accurate method for evaluating actual out of sample prediction performance. When p = 100, 200, 300, 400, horseshoe regression has the lowest prediction SSE. When p = 500, SSE of the lasso and horseshoe regression are close, and the lasso performs marginally better. The horseshoe regression also has the lowest SURE in all but one cases. Generally, SURE increases with p for all methods. The SURE for ridge regression approaches the OLS risk, which is 2nσ2 = 200 in these situations. SURE for PCR is larger than the OLS risk and PCR happens to be the poorest performer in most settings. Performance of the adaptive lasso also degrades compared to the lasso and the horseshoe, which remain the two best performers. Finally, the horseshoe regression outperforms the lasso in four out of the five settings we considered. Figure 1 shows contribution to SURE by each component for n = 100 and p = 500, for ridge, PCR, lasso and horseshoe regressions. The components are ordered left to right on the x-axis by decreasing magnitude of di , and SURE for prediction on each component are shown on the yaxis. Note from Table 1 that the 6, 30, 57, 67 and 96th components are the signals, meaning these terms correspond to a large α0 . The PCR risk on the 96th component is 203.22, which is out of range for the y-axis in the plot. For this data set, PCR selects 81 components, and therefore SURE for the first 81 components equal to 2σ2 = 2 and the SURE is equal to α̂2i d2i for i = 82, . . . , 100. 12 6 5 4 3 2 0 1 SURE for prediction 0 20 40 60 80 100 Component 4 3 2 0 1 SURE for prediction 5 6 Figure 1: Component-wise SURE for ridge (blue), PCR (gray), lasso (cyan), and horseshoe regression (red), for n = 100 and p = 500. Signal components are shown in solid squares and noise components shown in blank circles. Dashed horizontal line is at 2σ2 = 2. 0 10 20 30 ^d α 40 50 60 Figure 2: SURE for ridge (blue), PCR (gray), lasso (cyan) and horseshoe regression (red), versus α̂d, where α̂ is the OLS estimate of the orthogonalized regression coefficient, and d is the singular value, for n = 100 and p = 500. Dashed horizontal lines are at 2σ2 = 2 and 2σ2 /3 = 0.67. Component-wise SURE for ridge regression are large on the signal components, and is decreasing as the singular values d decrease on the other components. But due to the large global shrinkage parameter τ ridge must select in presence of both large signals and noise terms, the magnitude of improvement over the OLS risk 2σ2 is small for the noise terms. On the other hand, the horseshoe 13 estimator does not shrink the components with large α̂i di heavily and therefore the horseshoe SURE on the signal components are almost equal to 2σ2 (according to Theorem 4.2). SURE for the horseshoe is also much smaller than 2σ2 on many of the noise components. Lasso also appears to be quite effective for the noise terms, but its performance for the signal components is generally not as effective as the horseshoe. Figure 2 takes a fresh look at the same results and shows component-wise SURE plotted against α̂i di . The signal components as well as the first component in Table 1 have α̂i di > 10. Horseshoe SURE converges to 2σ2 for large α̂i di , as expected from Theorem 4.2. For these components, the SURE for both ridge and lasso are larger than 2σ2 , due to the bias introduced in estimating large signals by these methods (see also Theorem 1 of Carvalho et al., 2010). When α̂i 2 d2i ≈ 0, risks for lasso and horseshoe are comparable, with lasso being slightly better. This is because an estimate can be exactly zero for the lasso, but not for the horseshoe, which is a shrinkage method (as opposed to a selection method). Nevertheless, the upper bound on SURE for the horseshoe regression at 2σ2 /3 when α̂i 2 d2i ≈ 0 and provided τ is chosen to be small enough so that τ 2 ≤ di−2 , as established by Theorem 4.4, can be verified from Figure 2. Additional simulation results are presented in Supplementary Section S.1, where we (i) treat a higher dimensional case (p = 1000), (i) perform comparisons with non-convex MCP (Zhang, 2010) and SCAD (Fan and Li, 2001) regressions, (iii) explore different choices of X and (iv) explore the effect of the choice of α. The main finding is that the horseshoe regression is often the best performer when α has a sparse-robust structure as in Table 1, that is most elements are very small while a few are large so that |α|2 is large. This is consistent with the theoretical results of Section 5. 7 Assessing out of sample prediction in a pharmacogenomics data set We compare the out of sample prediction error of the horseshoe regression with ridge regression, PCR, the lasso, the adaptive lasso, MCP and SCAD on a pharmacogenomics data set. The data were originally described by Szakács et al. (2004), in which the authors studied 60 cancer cell lines in the publicly available NCI-60 database (https://dtp.cancer.gov/discovery development/nci60/). The goal here is to predict the expression of the human ABC transporter genes (responses) using some compounds or drugs (predictors) at which 50% inhibition of cellular growth for the cell lines are induced. The NCI-60 database includes the concentration level of 1429 such compounds, out of which we use 853, which did not have any missing values, as predictors. We investigate the expression levels of transporter genes A1 to A12 (except for A11, which we omit due to missing values), and B1. Thus, in our study X is a n × p matrix of predictors with n = 60, p = 853 and Y is a n-dimensional response vector for each of the 12 candidate transporter genes under consideration. To test the performance of the methods, we split each data set into training and testing sets, with 75% (45 out of 60) of the observations in the training sets. We standardize each response by subtracting the mean and dividing by the standard deviation. We fit the model on the training data, and then calculate mean squared prediction error (prediction MSE) on the testing data. This is repeated for 20 random splits of the data into training and testing sets. The tuning parameters in ridge regression, the lasso, the adaptive lasso, SCAD and MCP are chosen by five-fold cross validation on the training data. Similarly, the number of components in PCR and the global shrinkage parameter τ for horseshoe regression are chosen by cross validation as well. It is possible to use 14 SURE to select the tuning parameters or the number of components, but one needs an estimate of the standard deviation of the errors in high-dimensional regressions. This is a problem of recent interest, as the OLS estimate of σ2 is not well-defined in the p > n case. Unfortunately, some of the existing methods we tried, such as the method of moments estimator of Dicker (2014), often resulted in unreasonable estimates for σ2 , such as negative numbers. Thus, we stick to cross validation here, as it is not necessary to estimate the residual standard deviation in that case. The average prediction MSE over 20 random training-testing splits for the competing methods is reported in Table 3. Average prediction MSE for responses A1, A8 and A10 are around or larger than 1 for all of the methods. Since the responses are standardized before analysis, we might conclude that none of the methods performed well for these cases. Among the remaining nine cases, the horseshoe regression substantially outperforms the other methods for A3, A4, A9, A12 and B1. It is comparable to PCR for A5 and A7, and is comparable to the adaptive lasso for A6, which are the best performers in the respective cases. Overall, the horseshoe regression performed the best in 5 among the total 12 cases we considered. 8 Concluding remarks We outlined some situations where the horseshoe regression is expected to perform better compared to some other commonly used “global” shrinkage or selection alternatives for high-dimensional regression. Specifically, we demonstrated that the global term helps in mitigating the prediction risk arising from the noise terms, and an appropriate choice for the tails of the local terms is crucial for controlling the risk due to the signal terms. For this article we have used the horseshoe prior as our choice for the global-local prior. However, in recent years, several other priors have been developed that fall in this class. This includes the horseshoe+ (Bhadra et al., 2016a,b), the threeparameter beta (Armagan et al., 2011), the normal-exponential-gamma (Griffin and Brown, 2010), the generalized double Pareto (Armagan et al., 2013), the generalized shrinkage prior (Denison and George, 2012) and the Dirichlet–Laplace prior (Bhattacharya et al., 2015). Empirical Bayes approaches have also appeared (Martin and Walker, 2014) and the spike and slab priors have made a resurgence due to recently developed efficient computational approaches (Ročková and George, 2016; Ročková and George, 2014). Especially in the light of Theorem 4.3, we expect the results developed in this article for the horseshoe to foreshadow similar results when many of these alternatives are deployed. A particular advantage of using the horseshoe prior seems to be the tractable expression for SURE, as developed in Theorem 4.1. Whether this advantage translates to some of the other global-local priors mentioned above is an open question. Following the approach of Stein (1981), our risk results are developed in a non-asymptotic setting (finite n, finite p > n). However, global-local priors such as the horseshoe and horseshoe+ are known to be minimax in estimation in the Gaussian sequence model (van der Pas et al., 2014, 2016). For linear regression, frequentist minimax risk results are discussed by Raskutti et al. (2011); and Castillo et al. (2015) have shown that spike and slab priors achieve minimax prediction risk in regression. Whether the prediction risk for the horseshoe regression is optimal in an asymptotic sense is an important question to investigate and recent asymptotic prediction risk results for ridge regression (Dobriban and Wager, 2015) should prove helpful for comparing with global shrinkage methods. Another possible direction for future investigation might be to explore the implications of our 15 Table 3: Average out of sample mean squared prediction error computed on 20 random trainingtesting splits (number of splits out of 20 with lowest prediction MSE), for each of the 12 human ABC transporter genes (A1–A10, A12, B1) in the pharmacogenomics example. Methods under consideration are ridge regression (RR), principal components regression (PCR) , the lasso, the adaptive lasso (A LASSO), the minimax concave penalty (MCP), the smoothly clipped absolute deviation (SCAD) penalty, and the horseshoe regression (HS). Lowest prediction MSE and largest number of splits with the lowest prediction MSE for each response in bold. Response RR PCR LASSO A LASSO MCP SCAD HS A1 1.12 (2) 1.00 (3) 0.77 (1) 0.92 (2) 0.82 (1) 0.93 (4) 0.92 (0) 1.08 (6) 0.57 (4) 1.18 (0) 1.01 (0) 0.53 (1) 1.10 (5) 1.04 (1) 0.91 (0) 0.95 (0) 0.77 (6) 0.92 (0) 0.83 (8) 1.05 (4) 0.64 (0) 1.04 (7) 1.12 (0) 0.59 (0) 1.00 (7) 0.95 (7) 1.11 (0) 0.97 (2) 1.06 (4) 0.98 (3) 0.92 (1) 1.14 (6) 0.81 (0) 1.00 (4) 1.09 (2) 0.70 (3) 1.00 (2) 0.93 (5) 0.90 (0) 0.96 (2) 0.81 (1) 0.86 (5) 0.93 (4) 1.01 (4) 0.67 (6) 1.01 (3) 1.01 (2) 0.63 (2) 1.01 (1) 0.92 (1) 0.92 (1) 0.93 (2) 0.83 (2) 0.87 (0) 0.99 (0) 1.01 (0) 0.77 (0) 1.00 (2) 1.02 (1) 0.91 (1) 1.06 (1) 0.99 (0) 1.06 (0) 0.99 (0) 0.94 (0) 0.90 (2) 0.93 (0) 1.15 (0) 0.68 (1) 1.06 (0) 1.05 (0) 0.70 (3) 1.30 (2) 1.15 (3) 0.65 (18) 0.79 (12) 0.79 (6) 0.95 (6) 0.85 (7) 1.34 (0) 0.55 (9) 1.33 (4) 0.80 (15) 0.46 (10) A2 A3 A4 A5 A6 A7 A8 A9 A10 A12 B1 findings on the predictive density in terms of an appropriate metric, say the Kullback-Leibler loss, following the results of George et al. (2006). A A.1 Proofs Proof of Theorem 4.1 Part A follows from Equation (6) with standard algebraic manipulations. To prove part B, define Zi = 1/(1 + τ 2 λ2i d2i ). Then, from Equation (12) 2 −n/2 m(α̂) = (2πσ ) n ∏ Z 1 i =1 0 exp(−zi α̂2i d2i /2σ2 )di z1/2 i 16  zi τ 2 d2i 1 − zi + zi τ 2 d2i  1 (1 − zi )−1/2 zi−3/2 dzi τdi 2 −n/2 = (2πσ ) n ∏ Z 1 i =1 0 exp(−zi α̂2i d2i /2σ2 )(1 − zi )−1/2     −1 1 1 dzi . + 1 − 2 2 zi τ 2 d2i τ di From the definition of the compound confluent hypergeometric (CCH) density in Gordy (1998), the result of the integral is proportional to the normalizing constant of the CCH density and we have from Proposition 4.1 that, 2 −n/2 m(α̂) ∝ (2πσ ) n α̂2i d2i 1 1 H 1, , 1, 2 2 , 1, ∏ 2 2 2σ τ di i =1   . In addition, the random variable ( Zi | α̂i , σ, τ ) follows a CCH(1, 1/2, 1, α̂2i d2i /2σ2 , 1, 1/τ 2 d2i ) distribution. Lemma 3 of Gordy (1998) gives, d p H ( p, q, r, s, ν, θ ) = − H ( p + 1, q, r, s, ν, θ ). ds p+q This yields after some algebra that,   α̂2i d2i 1 1 , 1, H 2, , 1, 0 2 2 2 2 α̂i d2 2σ m (α̂i ) 2 τ di   2i , =− 2 2 m(α̂i ) 3 H 1, 1 , 1, α̂i di , 1, 1 σ 2 2σ2 τ 2 d2i  2  2 4   di α̂i di α̂2i d2i α̂2i d2i 2 1 1 8 1 1 − H 2, , 1, , 1, + H 3, , 1, , 1, 00 2 2 2 2 2 2 2 3 2 15 2 2σ σ 2σ σ4 m (α̂i ) τ di τ di   = . 2 2 α̂i di m(α̂i ) 1 H 1, 12 , 1, 2σ 2 , 1, τ 2 d2 i The correctness of the assertion α̂i d2 m0 (α̂i ) = − 2i E( Zi ), m(α̂i ) σ and α̂2 d4 d2 m00 (α̂i ) = − i2 E( Zi ) + i 4 i E( Zi2 ), m(α̂i ) σ σ can then be verified using Equation (15), completing the proof. A.2 Proof of Theorem 4.2 Define si = α̂2i d2i /2σ2 and θi = (τ 2 d2i )−1 , withθi ≥ 1, si ≥ 1. From Theorem 4.1, the componentwise SURE is SUREi =2σ2 − 2σ2 E( Zi ) − α̂2i d2i {E( Zi )}2 + 2α̂i 2 d2i E( Zi2 ) =2σ2 [1 − E( Zi ) + 2si E( Zi2 ) − si {E( Zi )}2 ], Thus, 2σ2 [1 − E( Zi ) − si {E( Zi )}2 ] ≤ SUREi ≤ 2σ2 [1 + 2si E( Zi2 )]. 17 (A.1) To find bounds on SURE, we need upper bounds on E( Zi2 ) and E( Zi ). Clearly, θi−1 ≤ {θi + (1 − θi )zi }−1 ≤ 1, when θi ≥ 1. Let ai = log(s5/2 i ) /si . Then ai ∈ [0, 5/ (2e )) when si ≥ 1. Now, R1 E( Zi2 ) 0 1 z2i (1 − zi )− 2 {θi + (1 − θi )zi }−1 exp(−si zi )dzi = R1 1 (1 − zi )− 2 {θi + (1 − θi )zi }−1 exp(−si zi )dzi 0 , An upper bound to the numerator of E( Zi2 ) can be found as follows. Z 1 0 ≤ 1 z2i (1 − zi )− 2 {θi + (1 − θi )zi }−1 exp(−si zi )dzi Z 1 0 = Z ai 0 1 z2i (1 − zi )− 2 exp(−si zi )dzi 1 z2i (1 − zi )− 2 exp(−si zi )dzi + − 12 Z 1 ai 1 z2i (1 − zi )− 2 exp(−si zi )dzi Z ai Z 1 1 z2i exp(−si zi )dzi + exp(− ai si ) z2i (1 − zi )− 2 dzi 0 ai     Z 1 a2i s2i 1 − 12 2 1 − 1 + ai si + exp(− ai si ) + exp(− ai si ) z2i (1 − zi )− 2 dzi = (1 − a i ) 3 2 si ai ≤ (1 − a i ) ≤ {1 − 5/(2e)} = − 12 2 1 + 5/2 3 si si Z 1 0 1 z2i (1 − zi )− 2 dzi C2 C1 + 5/2 , 3 si si R1 1 1 where C1 = {1 − 5/(2e)}− 2 ≈ 3.53 and C2 = 0 z2i (1 − zi )− 2 dzi = Γ (1/2) Γ (3)/Γ (3.5) = 16/15. Similarly, a lower bound on the denominator of E( Zi2 ) is Z 1 0 ≥ 1 (1 − zi )− 2 {θi + (1 − θi )zi }−1 exp(−si zi )dzi θi−1 Z 1 exp(−si zi )dzi 0  1 −1 1 − exp(− si ) = θi ≥ , si θ i (1 + s i ) Thus, combining the upper bound on the numerator and the lower bound on the denominator ! C C 2 1 E( Zi2 ) ≤ θi (1 + si ) + 5/2 . s3i si Thus, SUREi ≤ 2σ2 [1 + 2si E( Zi2 )] ( ≤ 2σ 2 1 + 2θi (1 + si ) 18 C1 C2 + 3/2 s2i si !) . (A.2) An upper bound to the numerator of E( Zi ) can be found as follows. Let ãi = log(s2i )/si . Then, ãi ∈ [0, 2/e) for si ≥ 1. Z 1 0 ≤ 1 zi (1 − zi )− 2 {θi + (1 − θi )zi }−1 exp(−si zi )dzi Z 1 0 = Z ãi 0 1 zi (1 − zi )− 2 exp(−si zi )dzi 1 zi (1 − zi )− 2 exp(−si zi )dzi + ≤ (1 − ãi ) − 12 Z ãi 0 1 = (1 − ãi )− 2 ãi 1 zi (1 − zi )− 2 exp(−si zi )dzi zi exp(−si zi )dzi + exp(− ãi si ) Z 1 ai 1 Z 1 1 1 + 2 s2i si 0 1 zi (1 − zi )− 2 dzi 1 {1 − (1 + ãi si ) exp(− ãi si )} + exp(− ãi si ) s2i ≤ (1 − 2/e)− 2 = Z 1 Z 1 ãi 1 zi (1 − zi )− 2 dzi 1 zi (1 − zi )− 2 dzi C̃1 C̃2 + 2, s2i si R1 1 where C̃1 = (1 − 2/e) ≈ 0.26 and C̃2 = 0 zi (1 − zi )− 2 dzi = Γ (1/2) Γ (2)/Γ (2.5) = 4/3. The lower bound on the denominator is the same as before. Thus, E( Zi ) ≤
θ i (1 + s i ) . C̃ + C̃ 2 1 s2i Thus, SUREi ≥ 2σ2 [1 − E( Zi ) − si {E( Zi )}2 ] ( ) 2 (1 + s i ) 2 2 2 (1 + s i ) ≥ 2σ 1 − θi (C̃1 + C̃2 ) − θi (C̃1 + C̃2 ) . s2i s3i Thus, combining Equations (A.2) and (A.3) we get ( ) ( 2 SUREi (1 + s i ) 2 2 (1 + s i ) − θi (C̃1 + C̃2 ) ≤ 1 − θi (C̃1 + C̃2 ) ≤ 1 + 2θi (1 + si ) 2σ2 s2i s3i (A.3) C1 C2 + 3/2 s2i si !) , for si ≥ 1, θi ≥ 1. A.3 Proof of Theorem 4.3 Our proof is similar to the proof of Theorem 1 of Polson and Scott (2011). Note from Equations (9–10) that integrating out αi we have ind α̂i | λ2i , σ2 , τ 2 ∼ N (0, σ2 (di−2 + τ 2 λ2i )). 19 Let p(λ2i ) ∼ (λ2i ) a−1 L(λ2i ), as λ2i → ∞ where a ≤ 0. Define ui = σ2 (di−2 + τ 2 λ2i ). Then, as in Theorem 1 of Polson and Scott (2011), we have p(ui ) ∼ uia−1 L(ui ), as ui → ∞. The marginal of α̂i is then given by m(α̂i ) = Z √ 1 exp{−α̂2i /(2ui )} p(ui )dui . 2πui An application of Theorem 6.1 of Barndorff-Nielsen et al. (1982) shows that m(α̂i ) ∼ |α̂i |2a−1 L(|α̂i |) as |α̂i | → ∞. Thus, for large |α̂i | (2a − 1) ∂ log L(|α̂i |) ∂ log m(α̂i ) = + . ∂α̂i |α̂i | ∂α̂i (A.4) Clearly, the first term in Equation (A.4) goes to zero as |α̂i | → ∞. For the second term, we need to invoke the celebrated representation theorem by Karamata. A proof can be found in Bingham et al. (1989). RESULT A.1. (Karamata’s representation theorem). A function L is slowly varying if and only if there exists B > 0 such that for all x ≥ B the function can be written in the form   Z x ε(t) dt , L( x ) = exp η ( x ) + t B where η ( x ) is a bounded measurable function of a real variable converging to a finite number as x goes to infinity ε( x ) is a bounded measurable function of a real variable converging to zero as x goes to infinity. Thus, using the properties of η ( x ) and ε( x ) from the result above d log( L( x )) ε( x ) = η 0 (x) + →0 dx x as x → ∞. Using this in Equation (A.4) shows ∂ log m(α̂i )/∂α̂i → 0 as |α̂i | → ∞. By similar calculations, ∂2 log m(α̂i )/∂2 α̂i → 0 as |α̂i | → ∞. From Equation (6) SUREi =σ4 di−2  ∂ log m(α̂i ) ∂α̂i 2 + 2σ 2  1 + σ2 di−2  ∂2 log m(α̂i ) . ∂α̂2i Thus, SUREi → 2σ2 , almost surely, as |α̂i | → ∞. A.4 Proof of Theorem 4.4 The proof of Theorem 4.4 makes use of technical lemmas in Appendix A.6. Recall from Appendix A.1 that if we define Zi = 1/(1 + τ 2 λ2i d2i ) then the density of Zi is given 20 by α̂2 d2 1 1 ( Zi | α̂i , di , τ, σ ) ∼ CCH Zi | 1, , 1, i 2i , 1, 2 2 2 2σ τ di  2  . (A.5) Then SURE is given by SURE = ∑in=1 SUREi with SUREi = 2σ2 [1 − E( Zi ) + 2si E( Zi2 ) − si {E( Zi )}2 ] = 2σ2 [1 − E( Zi ) + si E( Zi2 ) + si Var( Zi )], (A.6) where si = α̂2i d2i /2σ2 . Thus, ∂{SUREi } ∂si ∂E( Zi ) ∂ ∂ + 2σ2 {si E( Zi2 )} + 2σ2 {si Var( Zi )} ∂si ∂si ∂si := I + II + III. −2σ2 = (A.7) Now, as a corollary to Lemma A.1, (∂/∂si )E( Zi ) = {E( Zi )}2 − E( Zi2 ) = −Var( Zi ) < 0, giving I > 0. The strict inequality follows from the fact that Zi is not almost surely a constant for any si ∈ R and (∂/∂si )E( Zi ) is continuous at si = 0. Next, consider II. Define θi = (τ 2 d2i )−1 and let 0 ≤ si ≤ 1. Then, ∂ ∂ {si E( Zi2 )} = E( Zi2 ) + si E( Zi2 ) ∂si ∂si 2 = E( Zi ) + si {E( Zi )E( Zi2 ) − E( Zi3 )} = si E( Zi )E( Zi2 ) + (by Lemma A.1) {E( Zi2 ) − si E( Zi3 )}. Now, clearly, the first term, si E( Zi )E( Zi2 ) ≥ 0. We also have Zi2 − si Zi3 = Zi2 (1 − si Zi ) ≥ 0 a.s. when 0 ≤ Zi ≤ 1 a.s. and 0 ≤ si ≤ 1. Thus, the second term E( Zi2 ) − si E( Zi3 ) ≥ 0. Putting the terms together gives II ≥ 0. Finally, consider III. Denote E( Zi ) = µi . Then, ∂ ∂ {si Var( Zi )} = Var( Zi ) + si {Var( Zi )} ∂si ∂si ∂2 E( Zi ) = Var( Zi ) − si ∂s2i = E{( Zi − µi )2 } − si E{( Zi − µi )3 } (by Lemma A.2) = E[( Zi − µi ) {1 − si ( Zi − µi )}]. 2 Now, ( Zi − µi )2 {1 − si ( Zi − µi )} ≥ 0 a.s. when 0 ≤ Zi ≤ 1 a.s. and 0 ≤ si ≤ 1 and thus, III ≥ 0. Using I, II and III in Equation (A.7) yields SUREi is an increasing function of si when 0 ≤ si ≤ 1, completing the proof of Part A. To prove Part B, we need to derive an upper bound on SURE when si = 0. First, consider si = 0 and 0 < θi ≤ 1. we have from Equation (A.6) that SUREi = 2σ2 (1 − EZi ). By Lemma A.3, (∂/∂θi )E( Zi ) > 0 and SUREi is a monotone decreasing function of θi , where θi = (τ 2 d2i )−1 . Next consider the case where si = 0 and θi ∈ (1, ∞). Define Z̃i = 1 − Zi ∈ (0, 1) when Zi ∈ (0, 1). Then, by Equation (A.11) and a formula on Page 9 of Gordy (1998), we have that Z̃i also follows a CCH 21 distribution. Specifically,  α̂2i d2i 1 2 2 ( Z̃i | α̂i , di , τ, σ ) ∼ CCH Z̃i | , 1, 1, − 2 , 1, τ di , 2 2σ  2 and we have SUREi = 2σ2 E( Z̃i ). Define θ̃i = θi−1 = τ 2 d2i . Then by Lemma A.3, (∂/∂θ̃i )E( Z̃i ) = −Cov( Z̃i , W̃i ) > 0 on 0 < θ̃i < 1. Therefore, SUREi is a monotone increasing function of θ˜i on 0 < θ˜i < 1, or equivalently a monotone decreasing function of θi on θi ∈ (1, ∞). Thus, combining the two cases above, we get that SURE at si = 0 is a monotone decreasing function of θi for any θi ∈ (0, ∞), or equivalently, an increasing function of τ 2 d2i . Since 0 ≤ Z̃i ≤ 1 almost surely, a natural upper bound on SUREi is 2σ2 . However, it is possible to do better provided τ is chosen sufficiently small. Assume that τ 2 ≤ di−2 . Then, since SUREi is monotone increasing in θi , the upper bound on SURE is achieved when θi = (τ 2 d2i )−1 = 1. In this case, E( Zi ) has a particularly simple expression, given by R1 1 zi (1 − zi )− 2 {θi + (1 − θi )zi }−1 dzi E( Zi ) = R 1 0 1 (1 − zi )− 2 {θi + (1 − θi )zi }−1 dzi 0 R1 1 z (1 − zi )− 2 dzi 2 0 i = R1 = . 1 − 3 (1 − zi ) 2 dzi 0 (A.8) Thus, sup SUREi = 2σ2 (1 − EZi ) = 2σ2 /3, completing the proof of Part B. To prove Part C, we first note that when si = 1 we have SUREi = 2σ2 [1 − E( Zi )|si =1 + 2E( Zi2 )|si =1 − {E( Zi )|si =1 }2 ] where E( Zi ) and E( Zi2 ) are evaluated at si = 1. Recall that when θi ≥ 1 and zi ∈ (0, 1) we have θi−1 ≤ {θi + (1 − θi )zi }−1 ≤ 1. Thus, R1 E( Zi2 )|si =1 0 1 z2i (1 − zi )− 2 {θi + (1 − θi )zi }−1 exp(−zi )dzi = R1 1 (1 − zi )− 2 {θi + (1 − θi )zi }−1 exp(−zi )dzi R1 2 1 z (1 − zi )− 2 exp(−zi )dzi 0.459 0 i ≈ θi ≤ = 0.43θi , R 1 −1 1 − 1.076 θ (1 − zi ) 2 exp(−zi )dzi 0 i (A.9) 0 and R1 1 zi (1 − zi )− 2 {θi + (1 − θi )zi }−1 exp(−zi )dzi E( Zi )|si =1 = R 1 0 , − 12 { θ + (1 − θ ) z }−1 exp(− z ) dz ( 1 − z ) i i i i i i 0 R1 1 θi−1 0 zi (1 − zi )− 2 exp(−zi )dzi 0.614 ≥ ≈ θi−1 = 0.57θi−1 . R1 1 − 1.076 (1 − zi ) 2 exp(−zi )dzi 0 Thus, " SUREi ≤ 2σ2 0.57 1− + 0.86θi − θi 22  0.57 θi 2 # . (A.10) When θi = 1, it can be seen that SUREi ≤ 1.93σ2 . A.5 Proof of Theorem 5.1 The proof of Theorem 5.1 makes use of technical lemmas in Appendix A.6. Recall from Appendix A.1 that if we define Zi = 1/(1 + τ 2 λ2i d2i ) then the density of Zi is given by ( Zi | α̂i , di , τ, σ2 ) ∼ CCH ( Zi | 1, 1/2, 1, si , 1, θi ) . (A.11) where si = α̂2i d2i /2σ2 and θi = (τ 2 d2i )−1 . Consider the case where di = 1 for all i and τ 2 = 1, i.e., θi = 1 for all i. From Equation (A.6), the risk estimate is SURE = ∑in=1 SUREi with SUREi = 2σ2 [1 − E( Zi ) + si E( Zi2 ) + si Var( Zi )], ≤ 2σ2 [1 − E( Zi ) + si + si Var( Zi )] = Ři . We begin by showing that the upper bound Ři = 2σ2 [1 − E( Zi ) + si + si Var( Zi )] is convex in si when si ∈ (0, 1). It suffices to show −E( Zi ) and si Var( Zi ) are separately convex. First, (∂2 /∂2 si )E( Zi ) = E{( Zi − µi )3 } ≤ 0, by Lemmas A.2 and A.4, proving −E( Zi ) is convex. Next, ∂2 {si Var( Zi )} = ∂s2i   ∂ ∂ Var( Zi ) + si {Var( Zi )} ∂si ∂si ∂ ∂2 {Var( Zi )} + si 2 {Var( Zi )} ∂si ∂si ∂ = −2E( Zi − µi )3 − si E( Zi − µi )3 (by Lemma A.2) ∂si 3 = −2E( Zi − µi ) + si E( Zi − µi )4 , (by Lemma A.5) = 2 ≥ 0, where the last inequality follows by Lemma A.4. Thus, since Ři is convex, it lies entirely below the straight line joining the two end points for si ∈ (0, 1). But Ři |si =0 ≤ 2σ2 /3 = 0.67σ2 (by Equation (A.8)) and   Ři |si =1 ≤ 2σ2 1 − 0.57 + 1 + 0.43 − (0.57)2 = 3.07σ2 , by Equations (A.9) and (A.10).Thus, by convexity SUREi ≤ Ři ≤ 0.67σ2 + si (3.07 − 0.67)σ2 = (0.67 + 2.4si )σ2 for si ∈ (0, 1) (A.12) We remark here that our simulations suggest SUREi itself is convex, not just the upper bound Ři , although a proof seems elusive. Nevertheless, as we shall see below, the convexity of Ři is sufficient for our purposes. Next, consider the interval si ∈ (1, 3). Noting that both E( Zi ) and E( Zi2 ) are monotone decreasing functions of si we have SUREi ≤ 2σ2 [1 − E( Zi )|si =3 + 2si {E( Zi2 )|si =1 } − si {E( Zi )|si =3 }2 ] 23 But, R1 1 zi (1 − zi )− 2 exp(−3zi )dzi E( Zi )|si =3,θi =1 = R 1 0 0 1 (1 − zi )− 2 exp(−3zi )dzi = 0.35. E( Zi2 )|si =1 < 0.43 from Equation (A.9). Thus, SUREi ≤ 2σ2 [1 − 0.35 + 0.86si − si (0.35)2 }2 ] = 2σ2 (0.65 + 0.74si ) for si ∈ (1, 3). (A.13) Using the upper bound from Theorem 4.2, SUREi ≤ 11.55σ2 for si ≥ 3. (A.14) When αi = 0, we have that α̂i ∼ N (0, σ2 di−2 ). Thus, α̂2i d2i /σ2 ∼ χ2 (1). Since si = α̂2i d2i /2σ2 we have that p(si ) = (π )−1/2 si−1/2 exp(−si ) for si ∈ (0, ∞). Combining Equations (A.12), (A.13) and (A.14) we have Riski = E(SUREi ) ≤ + Z 1 0 Z 3 1 + σ2 (0.67 + 2.4)π −1/2 si−1/2 exp(−si )dsi 2σ2 (0.65 + 0.74)π −1/2 si−1/2 exp(−si )dsi Z ∞ 3 11.55σ2 π −1/2 si−1/2 exp(−si )dsi = 1.75σ2 . A.6 Technical lemmas LEMMA A.1. If Z ∼ CCH( p, q, r, s, ν, θ ), then (∂/∂s)E( Z k ) = E( Z )E( Z k ) − E( Z k+1 ). LEMMA A.2. If Z ∼ CCH( p, q, r, s, ν, θ ), then (∂2 /∂2 s)E( Z ) = −(∂/∂s)Var( Z ) = E{( Z − µ)3 }, where µ = E( Z ). LEMMA A.3. If Z ∼ CCH( p, q, r, s, ν, θ ), then (∂/∂θ )E( Z ) = −Cov( Z, W ), for W = (1 − νZ ){θ + (1 − θ )νZ }−1 . If 0 < θ ≤ 1 then (∂/∂θ )E( Z ) > 0. LEMMA A.4. If Z ∼ CCH( p, q, r, s, 1, 1) with q > p, then E( Z − µ)3 ≤ 0, where µ = E( Z ). LEMMA A.5. If Z ∼ CCH( p, q, r, s, ν, θ ), then (∂/∂s)E( Z − µ)3 = −E{( Z − µ)4 }, where µ = E( Z ). A.6.1 Proof of Lemma A.1 Let, Z ∼ CCH( p, q, r, s, ν, θ ). Then for any integer k R 1/ν zk+ p−1 (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz . E( Z ) = R0 1/ν p−1 (1 − νz )q−1 { θ + (1 − θ ) νz }−r exp(− sz ) dz z 0 k Thus, R 1/ν k+ p −z (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz ∂ k E( Z ) = R0 1/ν ∂s z p−1 (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 24 " R 1/ν − zk+ p−1 (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 R 1/ν p−1 z (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 # R 1/ν −z p (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 × R 1/ν z p−1 (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 = − E ( Z k +1 ) + E ( Z )E ( Z k ). For an alternative proof directly using the H (·) functions, see Appendix D of Gordy (1998). A.6.2 Proof of Lemma A.2 Let, Z ∼ CCH( p, q, r, s, ν, θ ). From Lemma A.1, (∂/∂s)E( Z ) = −E( Z2 ) + {E( Z )}2 = −Var( Z ). Let µ = E( Z ). Then, ∂ ∂2 E( Z ) = − Var( Z ) 2 ∂s ∂s " R 1/ν # 2 z p−1 (1 − νz )q−1 { θ + (1 − θ ) νz }−r exp(− sz ) dz ( z − µ ) ∂ 0 =− R 1/ν p−1 ∂s z (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 R 1/ν (z − µ)2 z p (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz = 0 R 1/ν z p−1 (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 " R 1/ν (z − µ)2 z p−1 (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz − 0 R 1/ν z p−1 (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 # R 1/ν p q−1 { θ + (1 − θ ) νz }−r exp(− sz ) dz z ( 1 − νz ) 0 × R 1/ν z p−1 (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 =Cov( Z, ( Z − µ)2 ) =E[( Z − µ){( Z − µ)2 − E( Z − µ)2 }] =E{( Z − µ)3 } − Var( Z )E( Z − µ) = E{( Z − µ)3 }. A.6.3 Proof of Lemma A.3 Let Z ∼ CCH( p, q, r, s, ν, θ ) and W = (1 − νZ ){θ + (1 − θ )νZ }−1 . Then, R 1/ν p z (1 − νz)q {θ + (1 − θ )νz}−(r+1) exp(−sz)dz ∂ E( Z ) = − R 01/ν ∂θ z p−1 (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 " R 1/ν z p (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 + R 1/ν z p−1 (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 # R 1/ν p−1 q { θ + (1 − θ ) νz }−(r +1) exp(− sz ) dz z ( 1 − νz ) × R0 1/ν z p−1 (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 = − E( ZW ) + E( Z )E(W ) = −Cov( Z, W ). When 0 < θ ≤ 1, it is obvious that Z and W are negatively correlated, and thus −Cov( Z, W ) > 0. 25 A.6.4 Proof of Lemma A.4 Let Z ∼ CCH( p, q, r, s, 1, 1). Then, R1 E( Z − µ ) = 3 0 (z − µ)3 z p−1 (1 − z)q−1 exp(−sz)dz , R 1 p −1 z (1 − z)q−1 exp(−sz)dz 0 which can be seen to have the same sign as the third central moment, or skewness of a Beta( p, q) random variable, which is negative when q > p. A.6.5 Proof of Lemma A.5 Let, Z ∼ CCH( p, q, r, s, ν, θ ). Let µ = E( Z ). Then, ∂ E( Z − µ )3 = − ∂s R 1/ν 0 (z − µ)3 z p (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz R 1/ν 0 z p−1 (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz " R 1/ν + (z − µ)3 z p−1 (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz R 1/ν p−1 z (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 # R 1/ν p q−1 { θ + (1 − θ ) νz }−r exp(− sz ) dz z ( 1 − νz ) 0 × R 1/ν z p−1 (1 − νz)q−1 {θ + (1 − θ )νz}−r exp(−sz)dz 0 0 = − Cov( Z, ( Z − µ)3 ) = − E[( Z − µ){( Z − µ)3 − E( Z − µ)3 }] = − E{( Z − µ)4 } + E( Z − µ)3 E( Z − µ) = −E{( Z − µ)4 }. References Akaike, H. (1974). A new look at the statistical model identification. IEEE Transactions on Automatic Control 19, 716–723. Armagan, A., Clyde, M., and Dunson, D. B. (2011). Generalized beta mixtures of Gaussians. In Shawe-Taylor, J., Zemel, R. S., Bartlett, P., Pereira, F. C. N., and Weinberger, K. Q., editors, Advances in Neural Information Processing Systems 24, pages 523–531. Armagan, A., Dunson, D. B., and Lee, J. (2013). Generalized double Pareto shrinkage. Statistica Sinica 23, 119–143. Barndorff-Nielsen, O., Kent, J., and Sørensen, M. (1982). Normal variance-mean mixtures and z distributions. International Statistical Review 50, 145–159. Bhadra, A., Datta, J., Polson, N. G., and Willard, B. (2016a). Default bayesian analysis with globallocal shrinkage priors. Biometrika 103, 955–969. Bhadra, A., Datta, J., Polson, N. G., and Willard, B. (2016b). The horseshoe+ estimator of ultrasparse signals. Bayesian Analysis to appear, 1–27. 26 Bhattacharya, A., Pati, D., Pillai, N. S., and Dunson, D. B. (2015). Dirichlet-Laplace priors for optimal shrinkage. Journal of the American Statistical Association 110, 1479–1490. Bingham, N. H., Goldie, C. M., and Teugels, J. L. (1989). Regular variation, volume 27 of Encyclopedia of mathematics and its applications. Cambridge University Press, Cambridge. Carvalho, C. M., Polson, N. G., and Scott, J. G. (2010). The horseshoe estimator for sparse signals. Biometrika 97, 465–480. Casella, G. and Hwang, J. T. (1982). Limit expressions for the risk of James–Stein estimators. Canadian Journal of Statistics 10, 305–309. Castillo, I., Schmidt-Hieber, J., Van der Vaart, A., et al. (2015). Bayesian linear regression with sparse priors. The Annals of Statistics 43, 1986–2018. Chun, H. and Keles, S. (2010). Sparse partial least squares regression for simultaneous dimension reduction and variable selection. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 72, 3–25. Clyde, M., Desimone, H., and Parmigiani, G. (1996). Prediction via orthogonalized model mixing. Journal of the American Statistical Association 91, 1197–1208. Craven, P. and Wahba, G. (1978). Smoothing noisy data with spline functions. Numerische Mathematik 31, 377–403. Denison, D. G. T. and George, E. I. (2012). Bayesian prediction with adaptive ridge estimators, volume 8 of IMS Collections, pages 215–234. Institute of Mathematical Statistics, Beachwood, Ohio, USA. Dicker, L. H. (2014). Variance estimation in high-dimensional linear models. Biometrika 101, 269– 284. Dobriban, E. and Wager, S. (2015). High-dimensional asymptotics of prediction: Ridge regression and classification. arXiv preprint arXiv:1507.03003 . Efron, B. (1983). Estimating the error rate of a prediction rule: improvement on cross-validation. Journal of the American Statistical Association 78, 316–331. Efron, B. (2004). The estimation of prediction error: covariance penalties and cross-validation (with discussion). Journal of the American Statistical Association 99, 619–642. Fan, J. and Li, R. (2001). Variable selection via nonconcave penalized likelihood and its oracle properties. Journal of the American statistical Association 96, 1348–1360. Foster, D. P. and George, E. I. (1994). The risk inflation criterion for multiple regression. The Annals of Statistics 22, 1947–1975. Frank, I. E. and Friedman, J. H. (1993). A statistical view of some chemometrics regression tools. Technometrics 35, 109–135. 27 George, E. I., Liang, F., and Xu, X. (2006). Improved minimax predictive densities under KullbackLeibler loss. The Annals of Statistics 34, 78–91. Ghosh, P., Tang, X., Ghosh, M., and Chakrabarti, A. (2016). Asymptotic properties of bayes risk of a general class of shrinkage priors in multiple hypothesis testing under sparsity. Bayesian Analysis 11, 753–796. Gordy, M. B. (1998). A generalization of generalized beta distributions. In Finance and Economics Discussion Series. Division of Research and Statistics, Division of Monetary Affairs, Federal Reserve Board. Griffin, J. E. and Brown, P. J. (2010). Inference with normal-gamma prior distributions in regression problems. Bayesian Analysis 5, 171–188. Hastie, T., Tibshirani, R., and Friedman, J. (2009). The elements of statistical learning: data mining, inference and prediction. Springer, New York, 2nd edition. Hoerl, A. E. and Kennard, R. W. (1970). Ridge regression: Biased estimation for nonorthogonal problems. Technometrics 12, 55–67. Jolliffe, I. T. (1982). A note on the use of principal components in regression. Journal of the Royal Statistical Society: Series C (Applied Statistics) 31, 300–303. Mallows, C. L. (1973). Some comments on C p . Technometrics 15, 661–675. Martin, R. and Walker, S. G. (2014). Asymptotically minimax empirical Bayes estimation of a sparse normal mean vector. Electronic Journal of Statistics 8, 2188–2206. Masreliez, C. (1975). Approximate non-Gaussian filtering with linear state and observation relations. IEEE Transactions on Automatic Control 20, 107–110. Pericchi, L. and Smith, A. (1992). Exact and approximate posterior moments for a normal location parameter. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 54, 793–804. Polson, N. G. and Scott, J. G. (2011). Shrink globally, act locally: sparse Bayesian regularization and prediction. In Bernardo, J. M., Bayarri, M. J., Berger, J. O., Dawid, A. P., Heckerman, D., Smith, A. F. M., and West, M., editors, Bayesian Statistics 9, Oxford. Oxford University Press. Polson, N. G. and Scott, J. G. (2012a). Local shrinkage rules, Lévy processes, and regularized regression. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 74, 287–311. Polson, N. G. and Scott, J. G. (2012b). On the half-Cauchy prior for a global scale parameter. Bayesian Analysis 7, 887–902. Raskutti, G., Wainwright, M. J., and Yu, B. (2011). Minimax rates of estimation for highdimensional linear regression over `q -balls. IEEE Transactions on Information Theory 57, 6976– 6994. Ročková, V. and George, E. I. (2016). The spike-and-slab lasso. Journal of the American Statistical Association to appear,. 28 Ročková, V. and George, E. I. (2014). EMVS: The EM approach to Bayesian variable selection. Journal of the American Statistical Association 109, 828–846. Stein, C. M. (1981). Estimation of the mean of a multivariate normal distribution. The Annals of Statistics 9, 1135–1151. Szakács, G., Annereau, J.-P., Lababidi, S., Shankavaram, U., Arciello, A., Bussey, K. J., Reinhold, W., Guo, Y., Kruh, G. D., Reimers, M., Weinstein, J. N., and Gottesman, M. M. (2004). Predicting drug sensitivity and resistance: Profiling ABC transporter genes in cancer cells. Cancer Cell 6, 129 – 137. Tibshirani, R. (1996). Regression shrinkage and selection via the lasso. Journal of the Royal Statistical Society: Series B (Statistical Methodology) 58, 267–288. Tibshirani, R. J. and Taylor, J. (2012). Degrees of freedom in lasso problems. The Annals of Statistics 40, 1198–1232. van der Pas, S., Kleijn, B., and van der Vaart, A. (2014). The horseshoe estimator: Posterior concentration around nearly black vectors. Electronic Journal of Statistics 8, 2585–2618. van der Pas, S., Salomond, J.-B., and Schmidt-Hieber, J. (2016). Conditions for posterior contraction in the sparse normal means problem. Electronic Journal of Statistics 10, 976–1000. Zellner, A. (1986). On assessing prior distributions and Bayesian regression analysis with g-prior distributions. In Goel, P. K. and Zellner, A., editors, Bayesian Inference and Decision Techniques: Essays in Honor of Bruno de Finetti, pages 233–243. Elsevier, North-Holland. Zhang, C.-H. (2010). Nearly unbiased variable selection under minimax concave penalty. The Annals of statistics 38, 894–942. 29 Supplementary Material to Prediction risk for the horseshoe regression Anindya Bhadra Department of Statistics, Purdue University, 250 N. University Street, West Lafayette, IN 47907, USA. bhadra@purdue.edu Jyotishka Datta Department of Mathematical Sciences, University of Arkansas, Fayetteville, AR 72701, USA. jd033@uark.edu Yunfan Li Department of Statistics, Purdue University, 250 N. University Street, West Lafayette, IN 47907, USA. li896@purdue.edu Nicholas G. Polson and Brandon Willard The University of Chicago Booth School of Business, 5807 S. Woodlawn Ave., Chicago, IL 60637, USA. ngp@chicagobooth.edu, brandonwillard@gmail.com S.1 Additional simulations We provide additional simulation results, complementing the results in Table 2. For each simulation setting, we report SURE when a formula is available. We also report the average out of sample prediction SSE (standard deviation of SSE) computed based on one training set and 200 testing sets. For each setting, n = 100. The methods under consideration are ridge regression (RR), principal components regression (PCR), the lasso, the adaptive lasso (A LASSO), the minimax concave penalty (MCP), the smoothly clipped absolute deviation (SCAD) penalty and the proposed horseshoe regression (HS). The method with the lowest SSE is in bold and that with lowest SURE is in italics for each setting. The features of these additional simulations include the following. 1. We explore a higher dimensional case (p = 1000) for each setting. 2. We incorporate two non-convex regression methods for comparisons. These are SCAD (Fan and Li, 2001) and MCP (Zhang, 2010). 3. We explore different choices of the design matrix X. These include three cases: (i) X is generated from a factor model, where it is relatively ill-conditioned (as in Table 2), (ii) X is generated from a standard normal, where it is well-conditioned and (iii) X is exactly orthogonal, with all singular values equal to 1. These are reported in corresponding table captions. 4. We explore different choices of true α. These include three cases: (i) Sparse-robust α, where most elements of α are close to zero and a few are large, (ii) null α, where all elements of α are zero and (iii) dense α, where all elements are non-zero. Exact settings and the value of ||α||2 are reported in the table captions. The major finding is that the horseshoe regression outperforms the other global shrinkage methods (ridge and PCR) when α is sparse-robust, which is consistent with the theoretical observation in Section 5. It also outperforms the other selection-based methods in this case. On the other hand, the dense α case is most often favorable to ridge regression, while the null α case is favorable to selection-based methods such as the lasso, adaptive lasso, MCP or SCAD, due to the ability of these methods to produce exact zero estimates. S.1 Table S.1: Sparse-robust α (five large coefficients equal to 10 and other coefficients equal to 0.5 or −0.5 randomly, giving ∑in=1 α2i = 523.75); X generated by a factor model with 4 factors, each factor follows a standard normal distribution; d1 /dn is the ratio of largest and smallest singular values of X. RR p d1 /dn SURE SSE 100 2360.43 165.45 200 28.47 188.13 300 22.76 192.35 400 21.81 194.73 500 18.18 196.03 1000 15.20 197.91 159.83 (22.02) 206.39 (28.61) 212.05 (28.50) 199.36 (28.75) 180.12 (27.16) 184.86 (26.42) PCR SURE SSE LASSO SURE SSE 163.80 122.78 217.40 266.84 337.32 410.82 669.69 161.62 (21.28) 244.71 (29.80) 280.25 (32.62) 328.48 (34.79) 379.03 (39.41) 736.69 (56.58) 174.48 155.26 179.45 158.07 196.83 A LASSO SSE MCP SSE SCAD SSE 132.25 (17.57) 148.41 (22.48) 175.46 (22.17) 197.25 (25.02) 223.21 (27.98) 345.26 (36.60) 127.07 (16.71) 154.01 (23.17) 172.18 (21.55) 199.08 (25.52) 224.91 (29.65) 344.04 (37.34) 127.85 (17.19) 157.73 (23.23) 176.29 (22.19) 198.40 (25.31) 226.76 (29.26) 344.04 (37.34) 145.07 (19.39) 162.94 (24.44) 190.09 (26.20) 182.89 (27.41) 173.82 (26.50) 205.28 (29.56) HS SURE SSE 116.01 123.07 (16.43) 152.37 (22.75) 164.17 (22.85) 165.15 (24.67) 161.77 (24.22) 182.18 (25.43) 160.89 157.50 172.63 166.10 191.64 Table S.2: Null α (∑in=1 α2i = 0); X is the same as in Table S.1. RR p SURE SSE 100 88.23 200 121.30 300 125.78 400 113.00 500 90.74 1000 88.86 100.86 (13.20) 107.68 (15.70) 101.36 (13.99) 99.50 (13.12) 101.04 (14.17) 100.34 (14.00) PCR SURE SSE LASSO SURE SSE 92.85 87.36 128.83 139.96 113.50 88.26 85.67 113.28 (14.91) 115.65 (16.28) 124.37 (17.35) 99.41 (13.09) 107.31 (15.08) 103.47 (14.29) 117.90 108.85 102.81 90.26 82.51 100.81 (13.29) 105.77 (15.06) 111.85 (15.37) 111.92 (15.51) 99.49 (14.16) 100.43 (13.90) S.2 A LASSO SSE MCP SSE SCAD SSE 100.70 (13.21) 100.32 (14.80) 101.30 (14.02) 114.62 (15.80) 99.06 (14.04) 99.52 (13.70) 100.81 (13.29) 104.39 (14.93) 104.89 (14.27) 99.40 (13.20) 99.49 (14.16) 100.43 (13.90) 100.81 (13.29) 101.78 (14.89) 102.91 (14.00) 110.30 (15.20) 99.49 (14.16) 100.41 (14.00) HS SURE SSE 92.42 102.31 (13.72) 111.39 (16.12) 112.00 (15.76) 107.20 (14.90) 102.93 (14.68) 104.84 (14.96) 122.29 119.67 113.42 101.55 99.73 Table S.3: Dense α (all coefficients equal to 2, giving ∑in=1 α2i = 400); X is the same as in Table S.1. RR p SURE SSE 100 162.49 200 183.75 300 189.38 400 193.02 500 194.85 1000 197.37 159.94 (21.60) 200.92 (27.97) 209.92 (27.88) 195.01 (28.92) 175.46 (26.52) 181.65 (26.50) PCR SURE SSE LASSO SURE SSE 177.47 194.86 196.06 200.39 197.74 208.87 247.40 175.19 (22.36) 233.12 (31.18) 225.92 (30.15) 217.68 (31.02) 201.18 (29.07) 197.59 (27.76) 211.99 216.01 218.16 220.34 224.75 203.89 (28.11) 232.36 (31.06) 524.84 (69.45) 306.15 (42.65) 743.40 (100.54) 210.78 (29.70) A LASSO SSE MCP SSE SCAD SSE 504.55 (46.13) 960.77 (59.48) 1344.27 (71.29) 1768.05 (78.92) 2154.61 (92.70) 4280.80 (145.28) 491.67 (45.31) 895.83 (60.85) 1298.80 (77.97) 1675.73 (75.86) 2082.54 (92.93) 4075.00 (138.72) 491.67 (45.31) 911.94 (60.19) 1298.80 (77.97) 1675.73 (75.86) 2081.42 (93.37) 4075.00 (138.72) HS SURE SSE 185.46 173.89 (23.63) 228.18 (31.21) 227.55 (29.98) 213.91 (31.14) 188.93 (28.10) 186.48 (26.95) 204.10 206.99 207.81 207.93 203.47 Table S.4: Sparse-robust α (five large coefficients equal to 10 and other coefficients equal to 0.5 or −0.5 randomly, giving ∑in=1 α2i = 523.75); X follows a standard normal distribution; d1 /dn is the ratio of largest and smallest singular values of X. RR PCR p d1 /dn SURE SSE SURE SSE 100 351.2 196.72 228.78 200 5.73 199.84 300 3.63 199.91 400 2.89 199.94 500 2.51 199.95 1000 1.88 199.98 188.67 (29.04) 193.41 (28.36) 217.43 (27.91) 197.47 (27.43) 193.86 (27.26) 185.85 (27.17) 231.34 (34.36) 206.25 (28.28) 8082.93 (320.98) 223.43 (31.13) 273.63 (33.97) 560.45 (45.94) 221.35 8538.46 228.38 256.09 605.64 LASSO SURE SSE 207.63 218.26 222.62 224.53 224.96 222.18 S.3 425.67 (59.68) 1618.40 (211.54) 1926.13 (248.97) 2384.04 (299.58) 471.73 (60.52) 5781.13 (759.08) A LASSO SSE MCP SSE SCAD SSE 2537.23 (112.58) 4849.94 (146.61) 13132.01 (281.12) 9593.41 (210.42) 11980.15 (235.77) 23866.06 (326.06) 2573.27 (128.46) 4915.72 (186.74) 7316.38 (218.82) 9695.47 (323.72) 11991.11 (272.80) 24566.13 (941.16) 2573.27 (128.46) 4964.26 (186.55) 7316.39 (218.83) 9695.47 (323.72) 11991.11 (272.80) 24566.13 (941.16) HS SURE SSE 195.22 188.52 (28.90) 194.14 (28.45) 219.97 (28.13) 197.69 (27.46) 194.17 (27.24) 185.79 (27.17) 201.91 200.92 200.31 200.15 199.96 Table S.5: Null α (∑in=1 α2i = 0); X is the same as in Table S.4. RR p SURE SSE 100 118.45 200 136.93 300 152.52 400 158.64 500 166.06 1000 181.23 119.12 (18.19) 135.02 (21.74) 160.29 (21.61) 159.13 (23.15) 158.83 (23.35) 169.22 (25.25) PCR SURE SSE LASSO SURE SSE 96.35 92.06 96.49 119.00 100.88 98.64 89.95 106.88 (15.18) 100.14 (14.77) 131.94 (18.01) 104.06 (15.59) 98.10 (14.50) 100.66 (14.10) 94.54 118.15 96.30 94.30 87.79 A LASSO SSE MCP SSE SCAD SSE 100.52 (14.20) 100.13 (14.70) 100.49 (14.37) 100.46 (14.82) 97.99 (14.50) 99.80 (14.00) 101.21 (14.33) 100.39 (14.88) 100.71 (14.48) 103.03 (15.04) 100.36 (14.79) 100.66 (14.08) 101.21 (14.33) 102.06 (15.26) 100.71 (14.48) 103.03 (15.04) 100.36 (14.79) 100.51 (14.07) 101.21 (14.33) 100.15 (14.69) 100.71 (14.48) 103.07 (15.11) 100.36 (14.79) 100.07 (14.03) HS SURE SSE 119.11 114.69 (17.47) 126.34 (20.13) 140.08 (18.93) 132.14 (19.62) 131.53 (19.59) 138.94 (21.12) 126.19 140.91 138.62 140.14 141.11 Table S.6: Dense α (all coefficients equal to 2, giving ∑in=1 α2i = 400); X is the same as in Table S.4. RR PCR p SURE SSE SURE SSE 100 193.13 206.31 200 199.76 300 199.88 400 199.92 500 199.94 1000 199.97 188.60 (28.91) 193.93 (28.41) 217.60 (27.93) 196.61 (27.35) 193.02 (27.16) 185.98 (27.17) 200.53 (29.51) 349.52 (37.90) 445.11 (43.70) 618.38 (59.10) 823.27 (62.69) 2108.78 (101.50) 392.38 400.63 627.97 794.03 2116.68 LASSO SURE SSE A LASSO SSE MCP SSE SCAD SSE 222.52 40019.73 (690.71) 80016.11 (983.86) 120071.46 (1191.24) 159926.82 (1418.60) 199982.59 (1550.64) 399770.90 (2359.10) 40063.42 (717.27) 80187.49 (1102.52) 123161.75 (3757.26) 161662.45 (3304.65) 200043.69 (1647.47) 400168.82 (2934.25) 40063.42 (717.27) 80187.49 (1102.52) 123161.75 (3757.26) 161662.45 (3304.65) 200043.69 (1647.47) 400168.82 (2934.25) 224.73 222.50 222.51 225.01 224.77 210.23 (31.36) 316.05 (42.77) 16845.87 (2167.53) 43325.17 (5447.99) 6497.32 (824.72) 3145.06 (411.60) S.4 HS SURE SSE 199.25 191.74 (29.37) 194.48 (28.50) 217.75 (27.92) 196.70 (27.35) 193.33 (27.18) 186.02 (27.17) 200.14 200.02 200.00 200.00 200.03 Table S.7: Sparse-robust α (five large coefficients equal to 10 and other coefficients equal to 0.5 or −0.5 randomly, giving ∑in=1 α2i = 523.75); X with all singular values equal to 1. RR p SURE SSE 100 183.50 200 184.47 300 182.50 400 184.03 500 183.81 1000 185.36 179.99 (25.31) 196.14 (28.79) 192.24 (25.37) 178.58 (25.20) 173.44 (24.08) 166.39 (23.16) PCR SURE SSE LASSO SURE SSE 291.45 139.29 261.65 267.96 311.01 278.35 280.59 275.49 (32.44) 277.35 (33.30) 269.04 (30.40) 287.95 (32.98) 268.85 (30.78) 262.54 (30.01) 135.93 126.35 145.28 147.54 124.52 139.39 (20.36) 150.17 (21.76) 146.72 (18.96) 139.68 (19.20) 139.74 (19.98) 130.61 (18.02) A LASSO SSE MCP SSE SCAD SSE 129.30 (19.20) 128.76 (17.75) 132.26 (17.85) 128.94 (17.40) 126.65 (18.36) 128.83 (17.50) 126.70 (18.79) 129.25 (17.61) 132.05 (17.87) 127.57 (17.42) 127.19 (18.31) 129.78 (17.70) 126.19 (18.83) 129.58 (17.90) 132.17 (17.73) 127.41 (17.43) 126.30 (18.17) 129.47 (17.62) HS SURE SSE 131.81 122.60 (18.72) 131.16 (18.63) 128.72 (17.34) 123.83 (17.02) 120.78 (17.29) 125.49 (17.37) 129.15 119.09 130.13 132.70 119.24 Table S.8: Null α (∑in=1 α2i = 0); X with all singular values equal to 1. RR p SURE SSE 100 94.70 200 115.52 300 98.74 400 96.78 500 88.55 1000 88.87 100.13 (14.71) 103.43 (14.80) 100.49 (14.80) 97.99 (14.49) 99.97 (14.81) 100.94 (14.17) PCR SURE SSE LASSO SURE SSE 97.63 94.54 111.09 99.45 103.88 89.06 91.96 102.62 (15.37) 118.11 (16.91) 113.35 (16.49) 102.24 (15.12) 100.91 (14.72) 107.30 (15.40) 109.16 96.40 94.02 87.74 88.45 100.15 (14.69) 122.81 (17.79) 103.03 (15.04) 103.08 (14.96) 100.65 (14.87) 101.14 (14.14) S.5 A LASSO SSE MCP SSE SCAD SSE 100.13 (14.70) 100.49 (14.37) 100.46 (14.82) 97.99 (14.50) 99.98 (14.83) 100.95 (14.17) 100.15 (14.69) 112.22 (16.20) 103.03 (15.04) 101.71 (14.81) 100.65 (14.87) 101.62 (14.30) 100.15 (14.69) 100.80 (14.53) 103.03 (15.04) 103.17 (14.97) 100.65 (14.87) 101.14 (14.14) HS SURE SSE 99.92 101.44 (15.06) 106.72 (15.47) 102.40 (14.87) 101.01 (14.78) 101.53 (14.98) 102.26 (14.48) 116.62 103.10 99.64 93.63 94.34 Table S.9: Dense α (all coefficients equal to 2, giving ∑in=1 α2i = 400); X with all singular values equal to 1. RR p SURE SSE 100 177.89 200 181.88 300 176.64 400 174.62 500 179.13 1000 179.32 183.69 (25.16) 188.39 (27.10) 193.53 (25.76) 195.83 (26.36) 173.13 (23.18) 173.71 (24.14) PCR SURE SSE LASSO SURE SSE 220.87 203.16 207.49 215.03 248.90 202.78 225.50 200.41 (26.35) 239.95 (33.45) 205.00 (26.70) 249.51 (32.46) 192.97 (25.08) 195.46 (25.78) 214.11 196.19 209.80 214.36 209.35 307.44 (43.19) 255.88 (35.61) 250.76 (32.66) 221.41 (29.85) 201.27 (25.63) 248.20 (30.87) S.6 A LASSO SSE MCP SSE SCAD SSE 502.55 (41.53) 499.88 (41.69) 496.84 (45.74) 495.21 (40.50) 501.67 (38.90) 503.44 (41.48) 505.43 (43.20) 498.84 (42.90) 497.60 (47.41) 494.17 (40.89) 503.19 (38.93) 507.29 (41.38) 505.43 (43.20) 498.84 (42.90) 495.93 (45.90) 494.17 (40.89) 503.19 (38.93) 507.29 (41.38) HS SURE SSE 200.80 204.16 (27.46) 217.46 (30.81) 212.19 (27.68) 206.49 (28.41) 193.40 (25.60) 194.73 (26.67) 205.16 199.33 198.84 205.16 204.51 

General Phase Regularized Reconstruction using Phase Cycling Frank Ong1 , Joseph Cheng2 , and Michael Lustig1 1 Department of Electrical Engineering and Computer Sciences, University of California, Berkeley, California. arXiv:1709.05374v1 [cs.CV] 15 Sep 2017 2 Department of Radiology, Stanford University, Stanford, California. September 19, 2017 Running head: General Phase Regularized Reconstruction using Phase Cycling Address correspondence to: Michael Lustig 506 Cory Hall University of California, Berkeley Berkeley, CA 94720 mlustig@eecs.berkeley.edu This work was supported by NIH R01EB019241, NIH R01EB009690, GE Healthcare, and NSF Graduate Fellowship. Approximate word count: 224 (Abstract) 3925 (body) Submitted to Magnetic Resonance in Medicine as a Full Paper. Part of this work has been presented at the ISMRM Annual Conference 2014 and 2017. 1 Abstract Purpose: To develop a general phase regularized image reconstruction method, with applications to partial Fourier imaging, water-fat imaging and flow imaging. Theory and Methods: The problem of enforcing phase constraints in reconstruction was studied under a regularized inverse problem framework. A general phase regularized reconstruction algorithm was proposed to enable various joint reconstruction of partial Fourier imaging, water-fat imaging and flow imaging, along with parallel imaging (PI) and compressed sensing (CS). Since phase regularized reconstruction is inherently non-convex and sensitive to phase wraps in the initial solution, a reconstruction technique, named phase cycling, was proposed to render the overall algorithm invariant to phase wraps. The proposed method was applied to retrospectively under-sampled in vivo datasets and compared with state of the art reconstruction methods. Results: Phase cycling reconstructions showed reduction of artifacts compared to reconstructions without phase cycling and achieved similar performances as state of the art results in partial Fourier, water-fat and divergence-free regularized flow reconstruction. Joint reconstruction of partial Fourier + water-fat imaging + PI + CS, and partial Fourier + divergence-free regularized flow imaging + PI + CS were demonstrated. Conclusion: The proposed phase cycling reconstruction provides an alternative way to perform phase regularized reconstruction, without the need to perform phase unwrapping. It is robust to the choice of initial solutions and encourages the joint reconstruction of phase imaging applications. 2 1 Introduction Phase variations in MRI can be attributed to a number of factors, including field inhomogeneity, chemical shift, fluid flow, and magnetic susceptibility. These phase structures are often the bases of many useful MRI applications: Image phase from B0 inhomogenity is used for calibration and shimming. Others provide important physiological information in clinical imaging methods, such as chemical shift imaging, phase contrast imaging, and susceptibility imaging. Phase structures in MRI provide opportunities in reconstruction to resolve undersampling artifacts or to extract additional information. For example, in partial Fourier imaging, smoothness of the phase images is exploited to reduce acquisition time by factors close to two. In water-fat imaging, chemical shift induced phase shifts are used to separate water and fat images, while smoothness in B0 field inhomogeneity is exploited to prevent water-fat swaps. These methods have either reduced acquisition time as in the case of partial Fourier imaging, or provided more accurate diagnostic information as in the case of water-fat imaging. In this paper, we study the problem of exploiting these phase structures in a regularized inverse problem formulation. An inverse problem formulation allows us to easily incorporate parallel imaging (1,2) and compressed sensing (3) and utilize various phase regularizations for different applications. Our first contribution is to present a unified reconstruction framework for phase regularized reconstruction problems. In particular, using the same framework, we demonstrate the joint reconstruction of partial Fourier and water-fat imaging, along with parallel imaging (PI) and compressed sensing (CS), and the joint reconstruction of partial Fourier and divergence-free regularized flow imaging, along with PI + CS. Since phase regularized image reconstructions are inherently non-convex and are sensitive to phase wraps in the initial solution, we propose a reconstruction technique, named phase cycling, that enables phase regularized reconstruction to be robust to phase wraps. The proposed phase cycling technique is inspired by cycle spinning in wavelet denoising (4) and phase cycling in balanced steady state free precession. Instead of unwrapping phase before or during the reconstruction, the proposed phase cycling makes the reconstruction invariant to phase wraps by cycling different initial phase solutions during the reconstruction. Our key difference with existing works is that the proposed phase cycling does not solve for the absolute phase in reconstruction, but rather averages out the effects of phase wraps in reconstruction. We provide experimental results showing its robustness to phase wraps in initial solutions. 3 Related Works Reconstruction methods utilizing phase structures were proposed for water-fat reconstruction (5–13), B0 field map estimation (14), partial Fourier reconstruction (15–18) and divergence-free regularized 4D flow reconstruction (19–21). In addition, Reeder et al. (22) demonstrated the feasibility of jointly performing homodyne reconstruction and water-fat decomposition. Johnson et al. (23) presented results in jointly reconstructing water-fat and flow images. A closely related line of work is the separate magnitude and phase reconstruction method (24–27). In particular, Fessler and Noll (24) proposed using alternating minimization for separate magnitude and phase reconstruction, but their method remained sensitive to phase wraps in initial solution. Zibetti et al. (25) and Zhao et al. (26) achieved robustness to phase wraps in initial solution by designing a regularization function that resembles the finite difference penalty, and is periodic in the phase. The regularization function proposed by Zhao et al. differs from the one proposed by Zibetti et al. in that it is edge-preserving using the Huber loss function. One limitation of these methods is that they do not support general magnitude and phase operators, or arbitrary phase regularization. In particular, they cannot be applied to water-fat image reconstruction and flow image reconstruction with general velocity encoding, or regularization functions other than finite difference. This restriction to finite difference penalty can lead to well-known staircase artifacts as shown in Figure 6. In contrast, our proposed method can be applied to general phase imaging methods and supports arbitrary phase regularization as long as its proximal operator can be computed. This allows us to enforce application dependent regularization, such as wavelet sparsity penalty and divergence-free constraint for flow imaging. 2 Theory Forward Model and Applications In this section, we describe the forward models for partial Fourier imaging, water-fat imaging and flow imaging. We then show that these phase imaging methods can be described using a general forward model and hence can be combined within the same framework. For example, in the experimental sections, we combine partial Fourier with water-fat imaging, and partial Fourier with divergence-free regularized flow imaging, under the same framework. Figure 1 provides illustrations of the forward models of these applications. 4 Figure 1: Illustration of forward models y = A(M m · eıP p ) for partial Fourier imaging, water fat imaging and flow imaging. . Partial Fourier Imaging In partial Fourier imaging, a contiguous portion of k-space is not observed and the k-space data is observed through the partial Fourier operator F and sensitivity operator S. Let m be the magnitude images and p be the phase images, our forward model is given by: y = F S (m · eıp ) + η where · is the element-wise multiplication operator, ı denotes √ (1) −1, eıp denotes the element-wise exponential operator of the vector ıp, η is a complex, white Gaussian noise vector, and y represents the acquired k-space data. Traditional partial k-space reconstruction methods assume smoothness of the phase. Therefore, in our formulation, we directly impose smoothness constraint on the phase image p. Sparsity of the magnitude 5 image can also be exploited for compressed sensing applications. Water-fat Imaging In water-fat imaging, we are interested in reconstructing water and fat images with different chemical shift induced off-resonance, but the same B0 inhomogeneity induced off-resonance. In order to resolve these images, k-space measurements are acquired over multiple echo times. Concretely, let E be the number of echo times, mwater and mfat be the water and fat magnitude images respectively, pwater and pfat be the water and fat phase images respectively, ∆f be the frequency difference between water and fat under the single-peak model, and pfield be the B0 field inhomogenity. We denote Fe as the Fourier sampling operator for each echo time te . Our forward model is given by: ye = Fe S

mwater eıpwater + mfat eıpfat eıte 2π∆f eıte pfield + ηe for e = 1, . . . , E (2) We note that the water and fat images have independent phase images pwater and pfat , which are often implicitly captured in existing water-fat separation methods by representing the magnitude images as complex images. These phase images can be attributed to RF pulse with spectral variation and B1 field inhomogeneity. We explicitly represent these phase images because they can be regularized as spatially smooth when partial Fourier imaging is incorporated (22). In order to separate the components, the first-order field inhomogeneity pfield is regularized to be spatially smooth. In addition, sparsity constraints can be imposed on magnitude images for compressed sensing applications. Flow Imaging In three-dimensional phase contrast imaging, we are interested in reconstructing phase images with threedimensional velocity information. Concretely, we define p = (pbg , px , py , pz )> to be the background phase image and velocity images along three spatial dimensions (x, y, z) respectively. The background phase includes the B0 field inhomogeneity, and chemical shift induced phase. We also let V be the number of velocity encodes, and Pv be the velocity encoding vector for velocity encode v. For example, the four point 6 balanced velocity encoding (28) vectors, ignoring scaling, have the form: P1 = (+1, −1, −1, −1) P2 = (+1, +1, +1, −1) (3) P3 = (+1, +1, −1, +1) P4 = (+1, −1, +1, +1) Then, our forward model is given by:
yv = Fv S m · eıPv p + ηv for v = 1, . . . , V (4) Since blood flow is incompressible and hence divergence-free, the velocity images px , py , pz can be constrained to be divergence-free to provide more accurate flow rates (19–21). Smoothness constraint can also be imposed on the background phase image to improve flow accuracy. General Forward Model With the right representations shown in Appendix A, the above phase imaging applications can all be described using the following unified forward model:
y = A M m · eıP p + η (5) where y is the acquired k-space measurements, m contains the magnitude images, p contains the phase images, η is a white Gaussian noise vector, A is the forward operator for the complex images, M is the forward operator for the magnitude image, P is the forward operator for the phase image and · is the point-wise multiplication operator. We note that both m and p are real-valued. Objective function To reconstruct the desired magnitude images m and phase images p, we consider the following regularized least squares function:
2 1 y − A M m · eıP p 2 + gm (m) + gp (p) |2 {z } | {z } Data consistency f (m,p) (6) Regularization g(m,p) where gm and gp are regularization functions for magnitude and phase respectively. For notation, we split our objective function into a data consistency term f (m, p) and a regularization term g(m, p). 7 We note that our objective function is non-convex regardless of the application-dependent linear operators A, M , and P , because the exponential term has periodic saddle points in the data consistency function with respect to p. In addition, the forward model is bi-linear in the magnitude and complex exponential variables. Algorithm Since the phase regularized reconstruction problem is non-convex in general, finding the global minimum is difficult. In fact, finding a local minimum is also difficult with conventional gradient-based iterative methods as saddle points have zero gradients. Instead we aim for a monotonically descending iterative algorithm to ensure the reconstructed result reduces the objective function. In particular, we perform alternating minimization with respect to the magnitude and phase images separately and use the proximal gradient method (29, 30) for each sub-problem, which is guaranteed to descend with respect to the objective function in each iteration. Since the objective function is reduced in each iteration as long as the gradient is non-zero, the algorithm eventually converges to a stationary point with zero gradient, which can be either a local minimum or a saddle point, assuming the initialization is not exactly at a local maximum. A high level illustration is shown in Figure 2. Concretely, applying proximal gradient descent on our objective function (6), we perform a gradient descent step with respect to the data consistency function f (m, p) and a proximal operation with respect to the regularization function g(m, p). Then, we obtain the following update step for magnitude images m with fixed phase images p at iteration n: mn+1 = Pαn gm (mn − αn ∇m f (mn , p)) (7) and the following update step for phase images p with fixed magnitude images m at iteration n: pn+1 = Pαn gp (pn − αn ∇p f (m, pn )) (8) where Pg (x) = argminz 21 kz − xk22 + g(z) denotes the proximal operator for the function g, and αn is the step-size for iteration n. Note that implicitly, we require the regularization functions to have simple proximal operators that can be evaluated efficiently, which is true for most commonly used regularization functions. Examples of proximal operators include wavelet soft-thresholding for wavelet `1 norm, weighting for `2 norm and projection for any indicator function for convex sets. We refer the reader to the manuscript of Parikh and Boyd (30) for an overview. Using the CR calculus (31), we can derive exact expressions for the gradient terms. Substituting the 8 gradient terms, the update steps (7, 8) at iteration n can be explicitly written as:
rn = A∗ y − A(M mn · eıP p ) mn+1 = Pαn gm mn + αn Re(M ∗ (e−ıP p · rn )) (9)
and
rn = A∗ y − A(M m · eıP pn ) (10)

pn+1 = Pαn gp pn + αn Im P ∗ (M m · e−ıP pn · rn ) where rn can be interpreted as the residual complex image at iteration n. Figure 2: Conceptual illustration of the proposed algorithm. The overall algorithm alternates between magnitude update and phase update. Each sub-problem is solved using proximal gradient descent. Phase Cycling While the above algorithm converges to a stationary point, this stationary point is very sensitive to the phase wraps in the initial solution in practice. This is because phase wraps are highly discontinuous and are artifacts from the initialization method. Phase regularization in each iteration penalizes these discontinuities, causing errors to accumulate at the same position over iterations, and can result in significant artifacts in the resulting solution. Figure 4 shows an example of applying smoothing regularization on the phase image 9 Figure 3: Conceptual illustration of the proposed phase cycling reconstruction technique. Phase cycling achieves robustness towards phase wraps by spreading the artifacts caused by regularizing phase wraps spatially. by soft-thresholding its Daubechies wavelet coefficients. While the general noise level is reduced, the phase wraps are also falsely smoothened, causing errors around it, as pointed by the red arrows. These errors accumulate over iterations and cause significant artifacts near phase wraps in the reconstructed image, as pointed out by the yellow arrow in Figure 5. The supporting videos S1 and S2 show the development of the reconstruction results over iterations without and with phase cycling for the experiment in Figure 5, demonstrating the convergence behavior described above. To mitigate these artifacts from regularizing phase wraps, we propose a reconstruction technique, called phase cycling, to make our reconstruction invariant to phase wraps in the initial solution. Our key observation is that even though it is difficult to reliably unwrap phase, the position of these phase wraps can easily be shifted to a different spatial location by adding a constant global phase. Hence, artifacts caused by phase regularization can also be shifted spatially to a different location. Our phase cycling method simply proposes to shift the phase wraps randomly over iterations, illustrated in Figure 3, to prevent significant error accumulation at the same spatial location. Then over iterations, the artifacts are effectively averaged spatially. Concretely, let W be the set of phase wraps generated from the initial solution. Then for each iteration n, we randomly draw a phase wrap wn from W with equal probability, and propose the following phase 10 Figure 4: An example of applying smoothing regularization on the phase image by soft-thresholding its Daubechies wavelet coefficients. While the general noise level is reduced, the phase wraps are also falsely smoothened, causing errors around it, as pointed by the red arrow. These errors accumulate over iterations and cause significant artifacts near phase wraps as shown in Figure 5. cycled update for phase images p with fixed magnitude images m at iteration n: pn+1 = Pαn gp (pn + wn − αn ∇p f (m, pn )) − wn (11) A pseudocode of the proposed algorithm with phase cycling is included in Appendix B. Finally, we note that the phase cycled update steps can be viewed as an inexact proximal gradient method applied on the following robust objective function:
1 y − A M m · eıP p 2 2 2 + gm (m) + 1 X gp (p + w) |W| (12) w∈W where the phase regularization function is averaged over phase wraps. We refer the reader to Appendix C for more details. 3 Methods The proposed method was evaluated on partial Fourier imaging, water-fat imaging and flow imaging applications. Parallel imaging was incorporated in each application. Sensitivity maps were estimated using ESPIRiT (32), an auto-calibrating parallel imaging method, using the Berkeley Advanced Reconstruction 11 Toolbox (BART) (33). All reconstruction methods were implemented in MATLAB (MathWorks, Natick, MA), and run on a laptop with a 2.4 GHz Intel Core i7 with 4 multi-cores, and 8GB memory. Unless specified otherwise, the magnitude image was regularized with the `1 norm on the Daubechies 4 wavelet transform for the proposed method. The number of outer iteration was set to be 100, and the number of inner iteration for both magnitude and phase was set to be 10. The regularization parameters were selected by first fixing the magnitude regularization parameter, and optimizing the phase regularization parameter over a grid of parameters with respect to mean squared error. And then fixing the phase regularization parameter, and optimizing the magnitude regularization parameter similarly. The step-size for the magnitude update was chosen to be 1 λmax (A∗ A)λmax (M ∗ M ) , 1 λmax (A∗ A)λmax (P ∗ P ) max(|M m|2 ) , where λmax denotes the maximum eigenvalue. and the step-size for the phase update was chosen to be Partial Fourier Imaging A fully-sampled dataset from a human knee was acquired on a 3T GE Discovery MR 750 scanner (GE Healthcare, Waukesha, WI) and 8-channel knee coil, with a 3D FSE CUBE sequence, TE/TR = 25/1550 ms, 40 echo train length, image size 320×320×256, and spatial resolution of 0.5×0.5×0.6 mm3 as described in Epperson et al. (34) and is available online at http://www.mridata.org/. Another fully-sampled dataset from a human brain was acquired on 1.5T GE Signa scanner (GE Healthcare, Waukesha, WI) with a 8channel head coil, 3D GRE sequence, TE/TR = 5.2 ms / 12.2 ms, image size of 256 × 256 × 230 and spatial resolution of 1 mm. 2D slices were extracted along the readout direction for the experiments. Image masks for displaying the phase images were created by thresholding the bottom 10% of the magnitude images. A partial Fourier factor of 5/8 was retrospectively applied on both datasets. The brain dataset was further retrospectively under-sampled by 4 with variable density Poisson-disk pattern and a 24 × 24 calibration region. The proposed method with and without phase cycling were applied on the under-sampled datasets and compared. `1 regularization was imposed on the Daubechies 6 wavelet domain for the phase image. The homodyne method from Bydder et al. (17) with `1 regularization on the wavelet domain was applied and compared. The method was chosen for comparison because it was shown to be robust to errors in the phase estimate as it penalizes the imaginary component instead of enforcing the image to be strictly real. The original formulation included only `2 regularization on the real and imaginary components separately. To enable a fair comparison using similar image sparsity models, we modified the method to impose `1 wavelet regularization on the real and imaginary components separately to exploit wavelet sparsity, which achieved strictly smaller mean squared error than the original method. The number of iterations was set to be 1000. The regularization parameters were set similarly to how the magnitude and phase regularization parameters 12 were selected. Besides visual comparison, the quality of the reconstructed magnitude images was evaluated using the peak signal-to-noise ratio (PSNR). Given a reference image xref , and a reconstructed image xrec , PSNR is defined as:  PSNR(xref , xrec ) = 20 log10 max(xref ) kxref − xrec k2  (13) For our proposed method, the magnitude image and phase image were initialized from the zero-filled reconstructed image, that is m = |A> y| and p = ∠(A> y). Water-Fat Imaging Fully sampled water-fat datasets were obtained from the ISMRM Water-Fat workshop toolbox (available online at http://www.ismrm.org/workshops/FatWater12/data.htm). In particular, an axial slice of the liver with 8-channel coil array dataset was used, which also appeared in the paper of Sharma et al. (9). The dataset was acquired on a 3T Signa EXCITE HDx system (GE Healthcare, Waukesha, WI), using a GEinvestigational IDEAL 3D spoiled-gradient-echo sequence at three TE points with T E = [2.184, 2.978, 3.772] ms, BW = ±125 kHz, flip angle of 5 degrees and a 256 × 256 sampling matrix. Image masks for displaying the phase images were created by thresholding the bottom 10% of the root-sum-of-squared of the magnitude images. The liver dataset was retrospectively under-sampled by 4 with a variable density Poisson-Disk sampling pattern. Our proposed method was applied and compared with algorithm of Sharma et al. (9) both for the fully-sampled and under-sampled datasets. The implementation in the Water-Fat workshop toolbox was used with modification to impose the same wavelet transforms as the proposed method, and the default parameters were used. An axial slice of the thigh dataset from the ISMRM Water-Fat workshop toolbox was also used, with the same parameters. The dataset was retrospectively under-sampled by 4 with a variable density Poisson Disk sampling pattern and an additional 9/16 partial Fourier factor. Our proposed method was applied and compared with the result of applying the algorithm of Sharma et al. on the fully-sampled dataset. An `1 regularization on the Daubechies-4 wavelet transform of the image phase was applied. For our proposed method, the field map was initialized as zero images, the magnitude images were initialized as (mwater , mfat )> = |M > A> y|, and the phase images (pwater , pfat )> were extracted from ∠(M > A> y). 13 Divergence-free Regularized Flow Imaging Four 4D flow datasets of pediatric patients with tetrahedral flow encoding were acquired with 20 cardiac phases, 140 slices and an average spatial resolution of 1.2 × 1.2 × 1.5mm3 . The 4D flow acquisitions were performed on a 1.5T Signa Scanner (GE Healthcare, Waukesha, WI) with an 8 channel array using a spoiled gradient-echo-based sequence. The acquisitions were prospectively under-sampled by an average factor of 3.82 with variable density Poisson-disk undersampling. The flip angle was 15 degrees and average TR/TE was 4.94/1.91 ms. The performance of the proposed method was compared with `1-ESPIRiT (32), a PI + CS algorithm. Volumetric eddy-current correction was performed on velocity data. Segmentations for flow calculations were done manually on the aorta (AA) and pulmonary trunk (PT). Net flow rate and regurgitant fraction were calculated for each segmentation. Since the datasets did not contain phase wraps in the region of interest, phase unwrapping was not performed. Image masks for displaying the phase images were created by thresholding the bottom 20% of the magnitude images. One of the `1 ESPIRiT reconstruction result was further processed with divergence-free wavelet denoising (35) to compare with the proposed method. Another flow dataset of pediatric patient, with randomized flow encode undersampling using the VDRad sampling pattern (36) and a partial readout factor of 0.7, was acquired with 20 cardiac phase. The proposed method was applied and compared to `1-ESPIRiT. For our proposed method, an `1 regularization on the divergence-free wavelet transform (35) of the flow images was used to impose divergence-free constraint and an `1 regularization on the Daubechies 4 wavelet coefficients of the background phase was used to impose smoothness. The flow images were initialized as zero images and the background phase image pbg was extracted from the first velocity encode of ∠(A> y). 4 Results In the spirit of reproducible research, we provide a software package in MATLAB to reproduce most of the results described in this paper. The software package can be downloaded from: https://github.com/mikgroup/phase_cycling.git Partial Fourier Imaging Supporting Figure S1 shows the partial Fourier reconstruction results combined with PI on the knee dataset. The figure compares the proposed reconstruction with and without phase cycling along with the homodyne 14 reconstruction method described in Bydder et al. (17). For the reconstruction without phase cycling, significant artifacts near the phase wraps can be seen in the magnitude image, as pointed by the yellow arrow. The proposed reconstruction with phase cycling also performs comparably with the state-of-the-art and did not display significant artifacts. In terms of PSNR, the method of Bydder et al. resulted in 34.55 dB, the proposed method without phase cycling resulted in 32.83 dB, and the proposed method with phase cycling resulted in 34.93 dB. One instance of our Matlab implementation of the proposed method took 4 minutes and 6 seconds. Figure 5 shows the partial Fourier reconstruction results combined with PI and CS on the brain dataset with the proposed method. The figure compares the proposed reconstruction with and without phase cycling. Again, without phase cycling, significant artifacts can be seen in the magnitude image near phase wraps in the initial solution, pointed by the yellow arrow. Reconstruction with phase cycling does not display these artifacts. Figure 6 shows the results for the method of Bydder et al. and Zhao et al. As pointed out by the red arrows, the method of Bydder et al. shows higher error in the magnitude image compared to proposed method with phase cycling in these regions. The method of Zhao et al. shows higher magnitude image error in general, and displays staircase artifacts in the phase image, which are common in total variation regularized images, as pointed by the yellow arrow. In terms of PSNR, the method of Bydder et al. resulted in 33.64 dB, the method of Zhao et al. resulted in 30.62 dB, the proposed method without phase cycling resulted in 30.35 dB, and the proposed method with phase cycling resulted in 34.91 dB. One instance of our Matlab implementation of the proposed method took 1 minute and 53 seconds. We note that the severity of the artifact in the reconstruction without phase cycling for the brain dataset is much stronger than for the knee dataset because a higher regularization was needed to obtain lower reconstruction mean squared error. This is because the brain dataset was further under-sampled for CS. Larger regularization led to larger thresholding errors around the phase wraps and hence more significant artifacts in the resulting reconstructed images. Water-Fat Imaging Supporting Figure S2 shows the water-fat reconstruction results on the liver dataset, combined with PI and CS for the under-sampled case. For the fully sampled case, the water and fat images from the proposed method with phase cycling are comparable with the state-of-the-art water-fat reconstruction result using the method of Sharma et al. (9). Reconstruction from under-sampled data with the proposed method also results in similar image quality as the fully-sampled data and is consistent with the result shown in Sharma et al. (9). One instance of our Matlab implementation of the proposed method took 8 minutes and 27 seconds. 15 Figure 5: Partial Fourier + PI reconstruction results on a knee dataset. Without phase cycling, significant artifacts can be seen in the magnitude image near phase wraps in the initial solution, pointed by the yellow arrow. With phase cycling, these artifacts were reduced and the result is comparable to the robust iterative partial Fourier method with `1 wavelet described in Bydder et al. Figure 7 shows the water-fat reconstruction results on the thigh dataset, combined with partial Fourier, PI and CS. Our proposed method produces similar water and fat images on a partial Fourier dataset, as the fully-sampled reconstruction using the method of Sharma et al. (9). This demonstrates the feasibility of performing joint partial Fourier and water fat image reconstruction along with PI and CS using the proposed method. One instance of our Matlab implementation of the proposed method took 8 minutes and 23 seconds. 16 Figure 6: Partial Fourier + PI + CS reconstruction results on a brain dataset for the proposed method. Similar to Supplementary Figure S1, without phase cycling, siginificant artifacts can be seen in the magnitude image near phase wraps in the initial solution, as pointed by the yellow arrow. Divergence-free Regularized Flow Imaging Figure 8 shows the net flow rates calculated for four patient data. Both `1-ESPIRiT reconstruction and proposed reconstruction resulted in similar flow rates. Maximum difference in regurgitant fractions was 2%. 17 Figure 7: Partial Fourier + PI + CS reconstruction comparison results on the same brain dataset in Figure 5. The method of Bydder shows higher error in the magnitude image compared to proposed method with phase cycling in regions pointed out by the red arrows. The method of Zhao et al. shows higher magnitude image error in general, and displays staircase artifacts in the phase image, which are common in total variation regularized images, as pointed by the yellow arrow. 18 Figure 8: Water-fat + PI + CS reconstruction result on a liver dataset with three echoes. Both the method from Sharma et al. and our proposed method produce similar water and fat images on the fully-sampled dataset and the retrospectively under-sampled dataset. The comparison showed that in these four cases, our proposed method provided comparable measurements to those obtained from `1-ESPIRiT, which were shown to correlate with 2D phase contrast flow rates (37). A representative velocity image and speed map are shown in Figure 9. Visually, the proposed method reduces incoherent artifacts and noise compared to the other results, especially in regions pointed by the red arrows where there should be no fluid flow. Figure 10 shows the result of reconstructing a partial readout 4D flow dataset with randomized velocity encoding. The reconstructed magnitude image has significantly reduced blurring from partial readout, compared to the `1-ESPIRiT result. We also note that the velocity images are not masked and that velocities in low magnitude regions are naturally suppressed with the proposed reconstruction. One instance of our Matlab implementation of the proposed method took on the order of three hours. 19 Figure 9: Water-fat + partial Fourier + PI + CS reconstruction result on a thigh dataset. Our proposed method produce similar water and fat images on a under-sampled partial Fourier dataset compared to the fully-sampled reconstruction using the method from Sharma et al. Discussion In this work, we describe a unified framework and algorithm for phase regularized reconstructions. By presenting various phase regularized image reconstructions within the same framework, we are able to combine 20 Figure 10: Net flow rates across 4 studies for the proposed method compared with `1-ESPIRiT, calculated across segmentations of aorta (AA) and pulmonary trunk (PT). Both `1-ESPIRiT reconstruction and proposed reconstruction result in similar flow rates. Figure 11: Divergence-free wavelet regularized flow imaging + PI + CS reconstruction result on a 4D flow dataset, compared with `1-ESPIRiT and `1-ESPIRiT followed by divergence-free wavelet denoising. Visually, the proposed method reduces incoherent artifacts and noise compared to the other results, especially in regions pointed by the red arrows where there should be no fluid flow. 21 Figure 12: Divergence-free wavelet regularized flow imaging + PI + CS reconstruction results on a 4D flow dataset. The reconstructed magnitude image has significantly reduced blurring from partial readout, compared to the `1ESPIRiT result. 22 these phase sensitive imaging methods to improve the reconstructed results: Our result in partial Fourier + water-fat imaging + PI + CS shows the ability to further accelerate water-fat imaging with PI + CS by combining it with partial Fourier, while achieving comparable image quality as the fully sampled images. Our result in divergence-free flow + partial Fourier + PI + CS also shows the ability to achieve improvements in apparent image resolution over standard PI + CS reconstruction. The proposed framework shows promise and encourages the use of joint application of phase sensitive imaging techniques. One advantage of our proposed method is that more appropriate phase regularization can be enforced for each application. In particular, we compared our proposed method with the methods of Bydder et al., and Zhao et al. for partial Fourier imaging. The method of Bydder et al. requires imposing regularization on the image imaginary component, which does not necessarily correlate with regularization of the phase image. The method of Zhao et al., on the other hand, only supports finite difference penalties, which can result in staircase artifacts, shown in Figure 6. Our proposed method can impose a more general class of phase regularization, and in the case of partial Fourier imaging, the proposed method with Daubechies-6 wavelet sparsity constraint on the phase image resulted in the highest PSNR among the compared methods. For water-fat imaging, we compared our proposed method with the method of Sharma et al., and achieved similar image quality for both fully-sampled and under-sampled datasets. With phase cycling, our proposed method can be extended to include partial Fourier imaging in a straightforward way. Since the method of Sharma et al. solves for the absolute phase, it is unclear whether their proposed restricted subspace model can be extended to enforce smoothness of water and fat phase images, which often contain phase wraps, and more effort is needed to investigate such extension. Finally, through comparing the proposed method on divergence-free flow imaging and l1-ESPIRiT with divergence-free wavelet denoising, we have shown the advantage of joint PI + CS reconstruction utilizing phase structure over separate application of PI + CS reconstruction followed by phase image denoising. Further improvement in these applications can be obtained by incorporating additional system information. For example, multi-peak model can be used in water-fat image reconstruction by extending the forward model to become:   ye = Fe S mwater eıpwater + mfat eıpfat  J X   aj eıte 2π∆fj  eıte pfield  + ηe for e = 1, . . . , E (14) j=1 where J, aj and ∆fj are the number of fat peaks, the relative amplitude for the jth peak, and the frequency difference between the jth fat peak and water respectively. Temporal constraints can also be added in 4D flow reconstruction to improve the result. One example would be separately enforcing group sparsity on the Daubechies wavelet coefficients of the magnitude images over time, and divergence-free wavelet coefficients of velocity phase images over time, via the group `1 norm. 23 Similar to other iterative methods, our proposed method requires parameter selections for the regularization functions. One disadvantage of our proposed method over standard PI + CS methods is that we have two parameters to tune for, and thus requires more effort in selecting these parameters. On the other hand, since phase values are bounded between π and −π, we found that the phase regularization parameter often translates fairly well between experiments with similar settings. That is, a fixed phase regularization parameter results in similar smoothness for different scans with similar undersampling factor and noise level. Since we aim for the joint reconstruction of phase images, phase unwrapping becomes a difficult task. The proposed phase cycling provides an alternative and complementary way of regularizing phases during reconstruction, without the need for phase unwrapping. In cases when the absolute phase is needed, for example in flow imaging with low velocity encodes, phase unwrapping can be performed on the phase regularized reconstructed images, which has the potential to improve phase unwrapping performance due to the reduced artifacts. Finally, we note that the proposed method with phase cycling still requires a good phase initialization, as the overall objective is non-convex. Our proposed phase cycling only enables the ability to regularize phase without phase unwrapping. In particular, if the initial phase estimate is randomly initialized, then the resulting reconstruction will be much worse than the one with a good initialization. Finding a good phase initialization for all phase imaging methods remains an open problem. Phase initialization for our proposed method is still manually chosen depending on the application, as described in the Methods section. In our experiments, we found that often either initializing as the phase of the zero-filled reconstructed image, or simply a zero-valued phase image provides good initialization for our proposed method. We note that a better initial solution using existing PI + CS methods, such as `1-ESPIRiT, can potentially result in a more accurate solution for our proposed method. In our experiments, we found that zero-filled images as starting solutions have already provided adequate starting points with much lower computational cost, and initializing with `1 ESPIRiT did not provide noticeable improvements. However, for more aggressively accelerated scans, we expect a better initialization can improve the reconstructed result. Conclusion The proposed phase cycling reconstruction provides an alternative way to perform phase regularized reconstruction, without the need of performing phase unwrapping. The proposed method showed reduction of artifacts compared to reconstructions without phase cycling. The advantage of supporting arbitrary regularization functions, and general magnitude and phase linear models was demonstrated by comparing with 24 other state-of-the-art methods. Joint reconstruction of partial Fourier + water-fat imaging + PI + CS, and partial Fourier + divergence-free regularized flow imaging + PI + CS were demonstrated. The proposed method unifies reconstruction of phase sensitive imaging methods and encourages their joint application. A Phase imaging representation under the general forward model We first consider representing partial Fourier imaging under the general forward model. Let us define I to be the identity operator with input and output size equal to the image spatial size. We also define the magnitude, phase and complex linear operators to be M = I, P = I, and A = F S. Then partial Fourier imaging forward model can be represented as y = A(M m · eıP p ) + η. Next, to represent water-fat imaging under the general forward model, let us define the magnitude images m to be (mwater , mfat )> , and phase images to be (pwater , pfat , pfield )> . We also define the magnitude operator M , phase operator P , and the complex operator A to be:     I 0 t1 I 0          F1 S F1 Sejt1 2π∆f 0 I t1  0 I       ..   .    M =  ..  P =   A= .          I 0 tE  I 0 0 0     0 I 0 I tE ... .. . 0 ... FE S 0 FE SejtE 2π∆f      (15) where the 2 × 2 identity block in M is repeated E times. Then the water-fat imaging forward model (2) can be represented as y = A(M m · eıP p ) + η. Finally we consider representing flow imaging under the general forward model. For concreteness, we consider the four point balanced velocity encoding as an example. Recall that we define p = (pbg , px , py , pz )> . Let us define the magnitude operator M , phase operator P , and the complex operator A to be:       I I −I −I −I F1 S 0 0 0             I  I +I +I −I   0 F2 S 0 0        M =  P =  A=  I  I +I −I +I   0 0 F3 S 0        I I −I +I +I 0 0 0 F4 S Then the flow imaging forward model (4) can be represented as y = A(M m · eıP p ) + η. 25 (16) B Pseudocode for phase regularized reconstruction with phase cycling Algorithm 1 summarizes the proposed reconstruction method with phase cycling. C Phase Cycling as an Inexact Proximal Gradient Method In this section, we show that our proposed phase-cycling is an instance of the inexact proximal splitting method described in Sra’s paper (38). Following its result, the proposed phase cycling converges to an inexact stationary point. Concretely, the inexact proximal splitting method in Sra’s paper considers the following minimization problem: minimize f (p) + g(p) (17) pn+1 = Pαn g (pn − αn ∇f (pn ) + αn en ) (18) p and utilizes the following update steps: where en is the error at iteration n. The results in Sra’s paper (38) show that the algorithm converges to a solution that is close to a nearby stationary point, with distance proportional to the iteration error en . To translate this result to phase cycling, we need to express the error in the proximal operation as additive error. In the context of phase cycling, our objective function consists of f (p) = 21 ky − A(M m · eıP p )k22 and P 1 g(p) = |W| w∈W gp (p + w). Let us define the regularization function error at iteration n to be n (p) = P 1 gp (p + wn ) − |W| w∈W gp (p + w), then the proposed phase cycling update step can be written as: (19) pn+1 = Pαn g+αn n (pn − αn ∇f (pn )) Now, we recall that the proximal operator is defined as Pg (x) = minimize 12 kz − xk22 + g(z). Then using z the first-order optimality condition, we obtain that z ∗ = Pg (x) if and only if z ∗ = x − ∇g(z ∗ ), where ∇g(z ∗ ) is a subgradient of g(z ∗ ). Hence, we can rewrite equation (19) as, pn+1 = pn − αn ∇f (pn ) − αn ∇g(pn+1 ) − αn ∇n (pn+1 ) = Pαn g (pn − αn ∇f (pn ) − αn ∇n (pn+1 )) 26 (20) Algorithm 1 Pseudocode for phase regularized reconstruction with phase cycling Input: y - observed k-space m0 - initial magnitude images p0 - initial phase images A - complex linear operator M - magnitude linear operator P - phase linear operator W - set of phase wraps N - number of outer iterations K - number of inner iterations for magnitude and phase update Output: mN - reconstructed magnitude images pN - reconstructed phase images Function: αm = 1 λmax (A∗ A)λmax (M ∗ M ) for n = 0, . . . , N − 1 do // Update m with fixed pn mn,0 = mn for k = 0, . . . , K − 1 do
rn,k = A∗ y − A(M mn,k · eıP pn ) mn,k+1 = Pαm gm mn,k + αm Re(M ∗ (e−ıP pn · rn,k ))
end for mn+1 = mn,K // Update p with fixed mn+1 pn,0 = pn αn = 1 λmax (A∗ A)λmax (P ∗ P ) max(|M mn+1 |2 ) for k = 0, . . . , K − 1 do Randomly draw wn,k ∈ W
rn,k = A∗ y − A(M mn+1 · eıP pn,k )

pn,k+1 = Pαn gp pn,k + wn,k + αn Im P ∗ (M mn+1 · e−ıP pn,k · rn,k ) − wn,k end for pn+1 = pn,K end for 27 Hence, the proposed phase cycling can be viewed as an inexact proximal gradient method with error in each iteration as en = ∇(pn+1 ) and converges to a point close to a stationary point. We note that this error is often bounded in practice. For example, if we consider the `1 wavelet penalty g(p) = λkΨpk1 , then the gradient of g(p) is bounded by λ, and hence the error en is bounded. References [1] Pruessmann KP, Weiger M, Scheidegger MB, Boesiger P et al. SENSE: sensitivity encoding for fast MRI. Magnetic resonance in medicine 1999; 42:952–962. [2] Griswold MA, Jakob PM, Heidemann RM, Nittka M, Jellus V, Wang J, Kiefer B, Haase A. Generalized Autocalibrating Partially Parallel Acquisitions (GRAPPA). Magnetic Resonance in Medicine 2002; 47:1202–1210. [3] Lustig M, Donoho D, Pauly JM. Sparse MRI: The application of compressed sensing for rapid MR imaging. Magnetic Resonance in Medicine 2007; 58:1182–1195. [4] Coifman RR, Donoho DL. Translation-invariant de-noising. Wavelets and statistics 1995; 103:125–150. [5] Reeder SB, Pineda AR, Wen Z, Shimakawa A, Yu H, Brittain JH, Gold GE, Beaulieu CH, Pelc NJ. Iterative decomposition of water and fat with echo asymmetry and least-squares estimation (IDEAL): Application with fast spin-echo imaging. Magnetic Resonance in Medicine 2005; 54:636–644. [6] Reeder SB, Wen Z, Yu H, Pineda AR, Gold GE, Markl M, Pelc NJ. Multicoil Dixon chemical species separation with an iterative least-squares estimation method. Magnetic Resonance in Medicine 2003; 51:35–45. [7] Yu H, Reeder SB, Shimakawa A, Brittain JH, Pelc NJ. Field map estimation with a region growing scheme for iterative 3-point water-fat decomposition. Magnetic Resonance in Medicine 2005; 54:1032– 1039. [8] Doneva M, Börnert P, Eggers H, Mertins A, Pauly J, Lustig M. Compressed sensing for chemical shift-based water-fat separation. Magnetic Resonance in Medicine 2010; 64:1749–1759. [9] Sharma SD, Hu HH, Nayak KS. Accelerated water-fat imaging using restricted subspace field map estimation and compressed sensing. Magnetic Resonance in Medicine 2011; 67:650–659. [10] Lu W, Hargreaves BA. Multiresolution field map estimation using golden section search for water-fat separation. Magnetic Resonance in Medicine 2008; 60:236–244. 28 [11] Hernando D, Haldar JP, Sutton BP, Ma J, Kellman P, Liang ZP. Joint estimation of water/fat images and field inhomogeneity map. Magnetic Resonance in Medicine 2008; 59:571–580. [12] Reeder SB, McKenzie CA, Pineda AR, Yu H, Shimakawa A, Brau AC, Hargreaves BA, Gold GE, Brittain JH. Water–fat separation with IDEAL gradient-echo imaging. Journal of Magnetic Resonance Imaging 2007; 25:644–652. [13] Eggers H, Börnert P. Chemical shift encoding-based water-fat separation methods. Journal of Magnetic Resonance Imaging 2014; 40:251–268. [14] Funai AK, Fessler JA, Yeo DTB, Olafsson VT, Noll DC. Regularized field map estimation in MRI. IEEE Transactions on Medical Imaging 2008; 27:1484–1494. [15] Noll DC, Nishimura DG, Macovski A. Homodyne Detection in Magnetic Resonance Imaging. IEEE Transactions on Medical Imaging 2004; 10:1–10. [16] Haacke E, Lindskogj E, Lin W. A fast, iterative, partial-fourier technique capable of local phase recovery. Journal of Magnetic Resonance (1969) 1991; 92:126–145. [17] Bydder M, Robson MD. Partial fourier partially parallel imaging. Magnetic Resonance in Medicine 2005; 53:1393–1401. [18] Li G, Hennig J, Raithel E, Büchert M, Paul D, Korvink JG, Zaitsev M. ”An L1-norm phase constraint for half-Fourier compressed sensing in 3D MR imaging”. Magnetic Resonance Materials in Physics, Biology and Medicine 2015; 28:459–472. [19] Loecher M, Santelli C, Wieben O, Kozerke S. Improved L1-SPIRiT Reconstruction with a Phase Divergence Penalty for 3D Phase-Contrast Flow Measurements. In: Proceedings of the 21st Annual Meeting of International Society of Magnetic Resonance in Medicine, Salt Lake City, Utah, USA, November 2012 p. 1355. [20] Ong F, Uecker M, Tariq U, Hsiao A, Alley M, Vasanawala S, Lustig M. Compressed sensing 4d flow reconstruction using divergence-free wavelet transform. In: Proceedings of the 22nd Annual Meeting of International Society of Magnetic Resonance in Medicine, Milan, Italy, 2014. p. 0326. [21] Santelli C, Loecher M, Busch J, Wieben O, Schaeffter T, Kozerke S. Accelerating 4D flow MRI by exploiting vector field divergence regularization. Magnetic Resonance in Medicine 2016; 75:115–125. [22] Reeder SB, Hargreaves BA, Yu H, Brittain JH. Homodyne reconstruction and IDEAL water-fat decomposition. Magnetic Resonance in Medicine 2005; 54:586–593. 29 [23] Johnson KM, Wieben O, Samsonov AA. Phase-contrast velocimetry with simultaneous fat/water separation. Magnetic Resonance in Medicine 2010; 63:1564–1574. [24] Fessler JA, Noll DC. Iterative image reconstruction in MRI with separate magnitude and phase regularization. Biomedical Imaging: Nano to Macro 2004; pp. 489–492. [25] Zibetti MV, DePierro AR. Separate magnitude and phase regularization in mri with incomplete data: Preliminary results. In: Biomedical Imaging: From Nano to Macro, 2010 IEEE International Symposium on, Rotterdam, The Netherlands, 2010. pp. 736–739. [26] Zhao F, Noll DC, Nielsen JF, Fessler JA. Separate magnitude and phase regularization via compressed sensing. IEEE Transactions on Medical Imaging 2012; 31:1713–1723. [27] Zibetti MVW, DePierro AR. Improving compressive sensing in mri with separate magnitude and phase priors. Multidimensional Systems and Signal Processing 2017; 28:1109–1131. [28] Pelc NJ, Bernstein MA, Shimakawa A, Glover GH. Three-Direction Phase-Contrast. Journal of Magnetic Resonance Imaging 2005; 1:1–9. [29] Combettes PL, Pesquet JC. “Proximal Splitting Methods in Signal Processing”, pp. 185–212. Springer New York, New York, NY, 2011. [30] Parikh N, Boyd S et al. Proximal algorithms. Foundations and Trends R in Optimization 2014; 1:127– 239. [31] Kreutz-Delgado, Ken. The complex gradient operator and the CR-calculus. arXiv preprint 2009; arXiv:0906.4835. [32] Uecker M, Lai P, Murphy MJ, Virtue P, Elad M, Pauly JM, Vasanawala SS, Lustig M. ESPIRiTan eigenvalue approach to autocalibrating parallel MRI: Where SENSE meets GRAPPA. Magnetic Resonance in Medicine 2013; 71:990–1001. [33] Uecker M, Ong F, Tamir JI, Bahri D, Virtue P, Cheng JY, Zhang T, Lustig M. Berkeley Advanced Reconstruction Toolbox . In: Proceedings of the 22nd Annual Meeting of International Society of Magnetic Resonance in Medicine, Toronto, 2015. p. 2486. [34] Epperson K, Sawyer AM, Lustig M, Alley M, Uecker M. Creation of fully sampled MR data repository for compressed sensing of the knee. In: Proceedings of the 22nd Annual Meeting for Section for Magnetic Resonance Technologists, Salt Lake City, Utah, USA, 2013. [35] Ong F, Uecker M, Tariq U, Hsiao A, Alley MT, Vasanawala SS, Lustig M. Robust 4D flow denoising using divergence-free wavelet transform. Magnetic Resonance in Medicine 2015; 73:828–842. 30 [36] Cheng JY, Zhang T, Ruangwattanapaisarn N, Alley MT, Uecker M, Pauly JM, Lustig M, Vasanawala SS. Free-breathing pediatric MRI with nonrigid motion correction and acceleration. Journal of Magnetic Resonance Imaging 2015; 42:407–420. [37] Hsiao A, Lustig M, Alley MT, Murphy M, Chan FP, Herfkens RJ, Vasanawala SS. Rapid pediatric cardiac assessment of flow and ventricular volume with compressed sensing parallel imaging volumetric cine phase-contrast MRI. AJR Am J Roentgenol 2012; 198:W250–259. [38] Sra S. Scalable nonconvex inexact proximal splitting. Advances in Neural Information Processing Systems 2012; pp. 530–538. 31 List of Figures 1. Illustration of forward models y = A(M m · ejP p ) for partial Fourier imaging, water fat imaging and flow imaging. 2. Conceptual illustration of the proposed algorithm. The overall algorithm alternates between magnitude update and phase update. Each sub-problem is solved using proximal gradient descent. 3. Conceptual illustration of the proposed phase cycling reconstruction technique. Phase cycling achieves robustness towards phase wraps by spreading the artifacts caused by regularizing phase wraps spatially. 4. An example of applying smoothing regularization on the phase image by soft-thresholding its Daubechies wavelet coefficients. While the general noise level is reduced, the phase wraps are also falsely smoothened, causing errors around it, as pointed by the red arrow. These errors accumulate over iterations and cause significant artifacts near phase wraps as shown in Figure 5. 5. Partial Fourier + PI + CS reconstruction results on a brain dataset for the proposed method. Similar to Supplementary Figure S1, without phase cycling, significant artifacts can be seen in the magnitude image near phase wraps in the initial solution, as pointed by the yellow arrow. 6. Partial Fourier + PI + CS reconstruction comparison results on the same brain dataset in Figure 5. The method of Bydder shows higher error in the magnitude image compared to proposed method with phase cycling in regions pointed out by the red arrows. The method of Zhao et al. shows higher magnitude image error in general, and displays staircase artifacts in the phase image, which are common in total variation regularized images, as pointed by the yellow arrow. 7. Water-fat + partial Fourier + PI + CS reconstruction result on a thigh dataset. Our proposed method produces similar water and fat images on a undersampled partial Fourier dataset compared to the fully-sampled reconstruction using the method from Sharma et al. 8. Net flow rates across 4 studies for the proposed method compared with `1-ESPIRiT, calculated across segmentations of aorta (AA) and pulmonary trunk (PT). Both `1-ESPIRiT reconstruction and proposed reconstruction result in similar flow rates. 9. Divergence-free wavelet regularized flow imaging + PI + CS reconstruction result on a 4D flow dataset, compared with `1-ESPIRiT and `1-ESPIRiT followed by divergence-free wavelet denoising. Visually, the proposed method reduces incoherent artifacts and noise compared to the other results, especially in regions pointed by the red arrows where there should be no fluid flow. 32 10. Divergence-free wavelet regularized flow imaging + PI + CS reconstruction results on a 4D flow dataset. The reconstructed magnitude image has significantly reduced blurring from partial readout, compared to the `1-ESPIRiT result. 33 Supporting Figures S1 Partial Fourier + PI reconstruction results on a knee dataset. Without phase cycling, significant artifacts can be seen in the magnitude image near phase wraps in the initial solution, pointed by the yellow arrow. With phase cycling, these artifacts were reduced and the result is comparable to the robust iterative partial Fourier method with `1 wavelet described in Bydder et al. S2 Water-fat + PI + CS reconstruction result on a liver dataset with three echoes. Both the method from Sharma et al. and our proposed method produce similar water and fat images on the fully-sampled dataset and the retrospectively undersampled dataset. Supporting Videos S1 Reconstructed magnitude and phase images over iterations for the proposed method without phase cycling. Errors accumulate over iterations near phase wraps and cause significant artifacts in the resulting images. S2 Reconstructed magnitude and phase images over iterations for the proposed method with phase cycling. With phase cycling, no significant artifacts like the ones in Supplementary Video S1 are seen during the iterations. 34 

Joint optimization of transmission and propulsion in aerial communication networks arXiv:1710.01529v1 [cs.SY] 4 Oct 2017 Omar J. Faqir, Eric C. Kerrigan, and Deniz Gündüz Abstract— Communication energy in a wireless network of mobile autonomous agents should be considered as the sum of transmission energy and propulsion energy used to facilitate the transfer of information. Accordingly, communication-theoretic and Newtonian dynamic models are developed to model the communication and locomotion expenditures of each node. These are subsequently used to formulate a novel nonlinear optimal control problem (OCP) over a network of autonomous nodes. It is then shown that, under certain conditions, the OCP can be transformed into an equivalent convex form. Numerical results for a single link between a node and access point allow for comparison with known solutions before the framework is applied to a multiple-node UAV network, for which previous results are not readily extended. Simulations show that transmission energy can be of the same order of magnitude as propulsion energy allowing for possible savings, whilst also exemplifying how speed adaptations together with power control may increase the network throughput. U2 U1 a1 p a22 + δ22 AP (0, 0) Fig. 1: Geometric configuration for simulation setups featuring N = 1 (black) and N = 2 (green) nodes. Speeds along these paths may be variable or fixed. The altitudes and lateral displacements of U1 , U2 are a1 = a2 = 1000 m and δ1 = 0, δ2 = 1000 m, respectively. I. I NTRODUCTION We aim to derive a control strategy to minimize communication energy in robotic networks. In particular, uninhabited aerial vehicle (UAV) networks are considered, with results being generalizable to broader classes of autonomous networks. A dynamic transmission model, based on physical layer communication-theoretic bounds, and a mobility model for each node is considered alongside a possible network topology. As a cost function, we employ the underused interpretation of communication energy as the sum of transmission energy and propulsion energy used for transmission, i.e. when a node changes position to achieve a better channel. For simulation purposes we consider the two wireless network setups shown in Figure 1. We first present the most basic scenario consisting of a single agent U1 moving along a predefined linear path while offloading its data to a stationary access point (AP). We compare results for variable and fixed speeds, before studying a two-agent single-hop network. For UAV networks, research efforts largely break down into two streams: the use of UAVs in objective based missions (e.g. search and pursuit [1], information gathering/mobile sensor networks [2], [3]), and use as supplementary network links [4]. Optimal completion of these macro goals has been addressed in the literature, but there The support of the EPSRC Centre for Doctoral Training in High Performance Embedded and Distributed Systems (HiPEDS, Grant Reference EP/L016796/1) is gratefully acknowledged. O. J. Faqir and Deniz Gündüz are with the Department of Electrical & Electronics Engineering, Imperial College London, SW7 2AZ, U.K. ojf12@ic.ac.uk, d.gunduz@ic.ac.uk Eric C. Kerrigan is with the Department of Electrical & Electronic Engineering and Department of Aeronautics, Imperial College London, London SW7 2AZ, U.K. e.kerrigan@imperial.ac.uk is no necessary equivalence between optimal task-based and energy-efficient operations. Efforts concerning mobility focus on mobile (in which node mobility models are random) or vehicular (where mobility is determined by higher level objectives and infrastructure) ad-hoc networks [5]. Since neither are fully autonomous networks, mobility is not available as a decision variable. The work in [6] introduced the concept of proactive networks, where certain nodes are available as mobile relays. However, the focus is on relay trajectory design and a simplistic transmission model is assumed, inherently prohibiting energy efficiency. The related problem of router formation is investigated in [7] using realistic models of communication environments. We assume hard path constraints, possibly due to the existence of higher level macro objectives, but allow changes in trajectory along the path by optimizing their speed (as in [8], we define a trajectory as being a time-parameterized path). Use of fixed paths does not restrict our results as most UAV path planning algorithms operate over longer time horizons and are generally restricted to linear or circular loiter trajectories [8]. A linear program (LP) is used in [9] to determine how close a rolling-robot should move before transmission in order to minimize total energy. However, the linear motion dynamics used restricts applicability of the model. Similarly to our current work, [10] uses a single mobile UAV relay to maximize data throughput between a stationary source-destination pair. An optimal trajectory for an a priori transmission scheme is iteratively found. Similarly, for a given trajectory, the optimal relaying scheme may be obtained through water-filling over the source-torelay and relay-to-receiver channels. Our contribution differs from the above works in terms of the formulation of a more general nonlinear convex OCP for finding joint transmission and mobility strategies to minimize communication energy. We solve this problem, exemplifying possible savings for even just a single node. As a final point, we show analytically and numerically that, even at fixed speeds, the optimal transmission scheme for a twouser multiple-access channel(MAC) is counter-intuitive and not captured by naı̈ve transmission policies. II. P ROBLEM D ESCRIPTION Consider N homogeneous mobile nodes Un , n ∈ N , {1, . . . , N }, traveling along linear non-intersecting trajectories at constant altitudes an and lateral displacements δn over a time interval T , [0, T ]. The trajectory of node Un is denoted by t 7→ (qn (t), δn , an ), relative to a single stationary AP U0 at position (0, 0, 0) in a three dimensional space. At t = 0, Un is initialized with a data load of Dn bits, which must all be offloaded to U0 by time t = T . We consider a cooperative network model, in which all nodes cooperate to offload all the data in the network to the AP by relaying each other’s data. Each node has a data buffer of capacity M bits, which limits the amount of data it can store and relay. CN (q, p) , {r ≥ 0 | fm (q, p, r, S) ≤ 0, ∀S ⊆ N } , We employ scalar additive white Gaussian noise (AWGN) channels. For UAV applications, we assume all links are dominated by line-of-sight (LoS) components, resulting in flat fading channels, meaning all signal components undergo similar amplitude gains [11]. All nodes have perfect information regarding link status, which in practice may be achieved through feedback of channel measurements, while the overhead due to channel state feedback is ignored. Similar to [12], for a given link from source node Un to receiver node Um , the channel gain ηnm (·) is expressed as (1) where qnm , qn − qm , constant G represents transmit and receive antenna gains and α ≥ 1 the path loss exponent. We define anm and δnm in a similar fashion. The channel gain is inversely related to the Euclidean distance between nodes. Each node has a single omnidirectional antenna of maximum transmit power of Pmax Watts. We consider half duplex radios; each node transmits and receives over orthogonal frequency bands. Accordingly, a different frequency band is assigned for each node’s reception, and all messages destined for this node are transmitted over this band, forming a MAC. We do not allow any coding (e.g. network coding) or combining of different data packets at the nodes, and instead consider a decode-and-forward-based routing protocol at the relay nodes [13]. The resulting network is a composition of Gaussian MACs, for each of which the set of achievable rate tuples defines a polymatroid capacity region [14]. If N (2) where q is the tuple of the differences qnm in positions between the N users and the receiver, p ∈ P N is the tuple of transmission powers allocated by the N users on this channel, and P , [0, Pmax ] is the range of possible transmission powers for each user. fm (·) is a nonlinear function bounding CN (q, p), given by X fm (q, p, r, S) , rn − n∈S Bm log2 A. Communication Model G ηnm (qnm ) , p 2α , 2 + q2 a2nm + δnm nm nodes simultaneously transmit independent information to the same receiver, the received signal is a superposition of the transmitted signals scaled by their respective channel gains, plus an AWGN term. We model the achievable data rates using Shannon capacity, which is a commonly used upper bound on the practically achievable data rates subject to average power constraints. Due to the convexity of the capacity region, throughput maximization does not require time-sharing between nodes [14], but may be achieved through successive interference cancellation (SIC). Consider a single MAC consisting of N users Un , n ∈ N , transmitting to a receiver Um , m 6∈ N . The capacity region CN (·, ·), which denotes the set of all achievable rate tuples r, is defined as X ηnm (qnm )pn 1+ σ2 n∈S ! , (3) where rn is the nth component of r, Bm is the bandwidth allocated to Um , and σ 2 is the receiver noise power. Consider the example (Section IV-B) where we do not allow relaying. This gives rise to a MAC with N = 2 transmitters U1 , U2 and the AP U0 . The capacity region C2 (q, p) is the set of non-negative tuples (r1 , r2 ) that satisfy   η10 (q10 )p1 (4a) r1 ≤ B0 log2 1 + σ2   η20 (q20 )p2 r2 ≤ B0 log2 1 + (4b) σ2   η10 (q10 )p1 + η20 (q20 )p2 (4c) r1 + r2 ≤ B0 log2 1 + σ2 for all (p1 , p2 ) ∈ P 2 . The first two bounds restrict individual user rates to the single-user Shannon capacity. Dependence between U1 and U2 leads to the final constraint, that the sum rate may not exceed the point-to-point capacity with full cooperation. For transmit powers (p1 , p2 ) these constraints trace out the pentagon shown in Figure 2. The sum rate is maximized at any point on the segment L3 . Referring to SIC, the rate pair at boundary point R(1) is achieved if the signal from source U2 is decoded entirely before source U1 , resulting in the signal from U2 being decoded at a higher interference rate than the signal from U1 . At R(2) the opposite occurs. B. Propulsion Energy Model The electrical energy used for propulsion in rolling robots has been modeled as a linear or polynomial function of speed achievable data rates is bounded above by a set of 2|N | − 1 nonlinear submodular functions fm (·, ·, ·, ·), where |·| applied to a set denotes the cardinality operator. Exponential growth in the number of nodes is a computational intractability. Hence, results are limited to small or structured networks where only a subset of nodes use each MAC. The trajectory of node Un is denoted by the tuple r1 L1 R(1) L3 R(2) L2 Yn , (pn , rn , sn , qn , vn , v̇n , Fn ), r2 Fig. 2: Capacity region for a given power policy across two parallel channels, with corner rate pairs labeled as R(1) = (1) (1) (2) (2) (r1 , r2 ) and R(2) = (r1 , r2 ) and line segments labeled as L1 , L2 , L3 . (7) where qn (t) is the node’s position at time t and sn (t) the state of its storage buffer subject to maximum memory of M bits. The optimal control problem that we want to solve is N Z T X min pn (t) + vn (t)Fn (t)dt (8a) p,r,s,q,v,F n=1 0 s.t. ∀n ∈ N , m ∈ {N , N + 1}, t ∈ T , S ⊆ N in [9], [15] respectively. We take a more general approach, restricting the fixed wing UAV to moving at strictly positive speeds and using Newtonian laws as a basis, as in [16]. The function Ω(·) models the resistive forces acting on node Un in accordance with the following assumption. Assumption 1: The resistive forces acting on each node Un may be modeled by the function x 7→ Ω(x) such that x 7→ xΩ(x) is convex on x ∈ [0, ∞) and ∞ on x ∈ (−∞, 0). Comparatively, in the fixed wing model proposed in [17], the drag force of a UAV traveling at constant altitude at subsonic speed v is Ω(v) = ρCD0 Sv 2 2L2 + 2 (πe0 AR )ρSv 2 (5) where the first term represents parasitic drag and the second term lift-induced drag. Parasitic drag is proportional to v 2 , where ρ is air density, CD0 is the base drag coefficient, and S is the wing area. Lift induced drag is proportional to v −2 , where e0 is the Oswald efficiency, AR the wing aspect ratio and L the induced lift [17]. For fixed-altitude flight, L must be equal to the weight of the craft W = mg. The power required to combat drag is the product of speed and force. The propulsion force Fn (·) must satisfy the force balance equation Fn (t) − Ω(vn (t)) = mn v̇n (t), (6) where mn is the node mass, vn (t) is the speed and v̇n (t) is the acceleration. The instantaneous power used for propulsion is the product vn (t)Fn (t), with the total propulsion energy taken as the integral of this power over T . We assume vn (t) ≥ 0, ∀t ∈ T , which is valid for fixed wing aircrafts. Thrust is restricted to the range [Fmin , Fmax ]. C. General Continuous-Time Problem Formulation We formulate the problem in continuous-time. At time t, node Un , n ∈ N can transmit to any node Um , m ∈ {N , N + 1}\{n} at a non-negative data rate rnm (t) using transmission power pnm (t). The sum power used in all outgoing transmissions from Un is denoted by pn (t). From this, the set of fm (q(t), p(t), r(t), S \ {m}) ≤ 0 ṡn (t) = N X rmn (t) − N +1 X (8b) rnm (t) (8c) sn (0) = Dn , sn (T ) = 0 qn (0) = Qn,init , qn (T ) = Qn,final (8d) (8e) vn (0) = vn,init Fn (t) = mn v̇n (t) + Ω(vn (t)) (8f) (8g) q̇n (t) = ζn vn (t) Yn,min ≤ Yn (t) ≤ Yn,max (8h) (8i) m6=n m6=n The cost function (8a) is the sum of nodal transmission and propulsion energies. Constraint (8b) bounds the achievable data rate to within the receiving nodes’ capacity region, and (8c) updates the storage buffers with sent/received data. Constraints (8d) act as initial and final constraints on the buffers, while (8e)–(8h) ensure all nodes travel from their initial to final destinations without violating a Newtonian force-acceleration constraint; ζn ∈ {−1, 1} depending on whether the position qn (t) decreases or increases, respectively, if the speed vn (t) ≥ 0. The final constraint (8i) places simple bounds on the decision variables, given by Yn,min , (0, 0, 0, −∞, Vmin , −∞, Fmin ), (9a) Yn,max , (Pmax , ∞, M, ∞, Vmax , ∞, Fmax ), (9b) where 0 ≤ Vmin ≤ Vmax and Fmin ≤ Fmax . The above optimal control problem may then be fully discretized using optimal control solvers, such as ICLOCS [18]. Before simulation results are presented we prove that this problem admits an equivalent convex form under certain conditions. III. C ONVEXITY A NALYSIS Efficient convex programming methods exist, which may be used in real-time applications. We first show that the nonlinear data rate constraints (8b) are convex in both positions and transmission power. We then show that the non-linear equality constraint (8g) may be substituted into the cost function, convexifying the cost function. This, however, turns the previously simple thrust bound Fmin ≤ Fn (t) into a concave constraint, resulting in a convex OCP if thrust bounds are relaxed. The absence of thrust bounds arises when considering a fixed trajectory, or is a reasonable assumption if the speed range is sufficiently small. Lemma 1: The rate constraints (8b) are convex in powers and positions for all path loss exponents α ≥ 1. Proof: By writing the channel gains as an explicit function of node positions, for receiver Um each of the capacity region constraints is of the form X rn (t)− n∈S Bm log2 G X pn (t) 1+ 2 2 +q 2 α σ (a2nm + δnm nm (t) ) n∈S ! ≤ 0. (10) Since the non-negative weighted sum of functions preserves convexity properties, without loss of generality we take S to be a singleton, and drop subscripts. We also drop time dependencies. The above function is the composition of two functions φ1 ◦ φ2 (·), respectively defined as φ1 (r, φ2 (·)) , r − B log2 (1 + φ2 (·)), G p φ2 (p, q) , 2 2 . σ (a + δ 2 + q 2 )α (11) (12) The function (p, q) 7→ φ2 (p, q) is concave on the domain R+ ×R. We show this by dropping constants and considering the simpler function h(x, y) , xy −2α with Hessian # " −2α 0 y −2α−1 2 (13) ∇ h(x, y) = 2α(2α−1)x , −2α y −2α−1 y −2α−2 which is negative semi-definite, because it is symmetric with non-positive sub-determinants. Therefore, φ2 is jointly concave in both power and the difference in positions over the specified domain. φ1 is convex and non-increasing as a function of φ2 . Since the composition of a convex, non-increasing function with a concave function is convex [19], all data rate constraint functions are convex functions of (r, p, q). The posynomial objective function is not convex over the whole of its domain and the logarithmic data rate term prevents the use of geometric programming (GP) methods. Lemma 2: The following problem Z T min Fn (t)vn (t)dt (14a) vn ,Fn 0 s.t. ∀t ∈ T Fn (t) − Ω(vn (t)) = mn v̇n (t) Fmin ≤ fm (t) ≤ Fmax (14b) (14c) vn (t) ≥ 0 (14d) vn (0) = vn,init (14e) of minimizing propulsion energy of a single node Un , subject to initial and final conditions, admits an equivalent convex form for mappings vn (t) 7→ Ω(vn (t)) satisfying Assumption 1 and force bounds (Fmin , Fmax ) = (−∞, ∞). Proof: By noting that Fn (t) = Ω(vn (t)) + mn v̇n (t), we move the equality into the cost function, rewriting the problem as (15) min φ(vn ) s.t. (14c)–(14e), vn where φ(vn ) , Z |0 T Z T vn (t)v̇n (t)dt . vn (t)Ω(vn (t))dt + {z } |0 {z } φ1 (vn ) (16) φ2 (vn ) We now show that both φ1 (·) and φ2 (·) are convex. Starting with the latter, by performing a change of variable, the analytic cost is derived by first noting that φ2 (vn ) is the change in kinetic energy Z vn (T )
mn 2 φ2 (vn ) = mn vdv = vn (T ) − vn2 (0) , (17) 2 vn (0) which is a convex function of vn (T ) subject to fixed initial conditions (14d); in fact, it is possible to drop the vn2 (0) term completely without affecting the argmin. By Assumption 1, vn (t) 7→ vn (t)Ω(vn (t)) is convex and continuous on the admissible domain of speeds. Since integrals preserve convexity, the total cost function φ(·) is also convex. Removal of thrust F as a decision variable results in the set VF , {vn | Fmin ≤ Ω(vn (t)) + mn v̇n (t) ≤ Fmax }. (18) Even if Ω(·) is convex on the admissible range of speeds, the lower bound represents a concave constraint not admissible within a convex optimization framework. Therefore, dropping constraints on thrust results in a final convex formulation of Z T
mn 2 min vn (t)Ω(vn (t))dt + vn (T ) − vn2 (0) (19a) vn 2 0 s.t. ∀t ∈ T Vmin ≤ vn ≤ Vmax (19b) vn (0) = vn,init . (19c) Addition of bounds vn ∈ VF naturally results in a difference of convex (DC) problem [20] that may be solved through exhaustive or heuristic procedures. Theorem 1: In the absence of constraints on thrust, the general problem (8) admits an equivalent convex form. Proof: Non-convexities in this formulation arise from the posynomial function of speed v(t) and thrust Fm (t) in the cost function (8a), the nonlinear force balance equality (8g), and the capacity region data rate constraints (8b). The cost function is a superposition of the energies used by each node for propulsion and transmission. By noting that there is no coupling between nodes or between propulsion and transmission powers in this cost, the transformation used in Lemma 2 may be used to eliminate the nonlinear equality. We eliminate Fn (t) and v̇n (t) and move the nonlinear equality into the objective function, simultaneously convexifying the objective to get "Z # N T X mn 2 v (T ) pn (t) + vn (t)Ω(vn (t))dt + min p,r,s,q,v 2 n 0 n=1 s.t. ∀n ∈ N , m ∈ {N , N + 1}, t ∈ T , v ∈ V N , S ⊆ N 35 100 30 80 25 60 20 15 40 10 20 5 0 (8b)–(8f), (8h), Ỹn,min ≤ Ỹn (t) ≤ Ỹn,max 0 0 where Ỹn (t) , (pn (t), rn (t), sn (t), qn (t), vn (t)), and the bounds Ỹn,min , and Ỹn,max are similarly changed. It follows from Lemma 1 that all data rate constraints in (8b) are also convex, therefore the whole problem is convex. IV. S IMULATION R ESULTS A. Single Node A single mobile node U1 of mass 3 kg traveling at fixed altitude a = 1000 m and lateral displacement δ = 0 m, depicted in Figure 1, is considered first. In this section, simulation results are presented for the problem of minimizing the total communication energy to offload all data to U0 . This is compared to a water-filling solution [21] for minimizing the transmission energy. Subscripts denoting different nodes have been dropped in the remainder of this section. Specifically, we use Ω(·) of the form  ∞, ∀x ∈ (−∞, 0) Ω(x) , (21) CD1 x2 + CD2 x−2 , ∀x ∈ [0, ∞), where CD1 = 9.26 × 10−4 is the parasitic drag coefficient and CD2 = 2250 is the lift induced drag coefficient [17]. Simulation results are shown in Figure 3 for a storage buffer initialized to D = 75 MB and speeds restricted in the range [Vmin , Vmax ] = [30, 100] km/h. This results in a total energy expenditure of 309.50 kJ, where 105.05 kJ is due to transmission and 204.51 kJ is due to propulsion. Of this, only 48.01 kJ of extra propulsion energy is used to vary speed on top of the base energy required to traverse the distance at a constant speed. Furthermore, the problem would have been infeasible if the node was restricted to a constant speed of 65 km/h. We note that, with the given parameterization, it is possible to transmit up to 78 MB of data in the defined time interval. B [Hz] 105 M [GB] 1 Pmax [W] 100 α 1.5 400 600 800 1000 1200 (a) Optimal transmission power and propulsion force used by U1 . 10 5 10 35 8 30 25 The open source primal dual Interior Point solver Ipopt v.3.12.4 has been used through the MATLAB interface. Table I contains parameters common to the following experiments. Force constraints are relaxed in all experiments. From [5], the speed of a typical UAV is in the range 30 to 460 km/h. All nodes are initialized to their average speeds vn,init = (Vmax + Vmin )/2, We assume all nodes move in symmetric trajectories around the AP such that Qn,final = −Qn,init = (T /2)vn,init . σ2 [W] 10−10 200 T [min] 20 TABLE I: Dynamic model parameters that have been used across all simulation results. 6 20 15 4 10 2 5 0 0 0 200 400 600 800 1000 1200 (b) Associated achieved data rate and velocity profile of U1 . Fig. 3: Simulation results for the single-node problem, with trajectories shown as solid, and bounds shown as dashed lines. In comparison, if the speed of U1 is fixed, then the maximum transmittable data is approximately 56 MB, using 120.00 kJ of transmission energy. Although considerably more energy is used, the optimal power policy for a fixed trajectory is characterized by a water-filling solution, an equivalent proof of which may be found in [21]. This problem results in a one dimensional search space, easily solved through such algorithms as binary search. B. Multiple Nodes We now investigate the transmission energy problem for two nodes, traveling in parallel trajectories at fixed speeds such that Vmax = Vmin = 65 km/h, as depicted by the green lines in Figure 1. Relaying is not allowed, as may be the case if no bandwidth is allocated to U1 and U2 to receive each other’s transmissions, equivalently turning them into pure source nodes. Simulation results are presented in Figure 4. U1 is closer to the AP at all times, and therefore is advantaged in that it experiences more favorable channel conditions. The disadvantaged node U2 transmits for a longer duration due to the smaller relative change in its channel gain. The interior point algorithm converged after 42 iterations to a minimum energy of 52.707kJ and 26.77kJ for U1 and U2 , respectively, for a starting data load of D1 = D2 = 25 MB. It is notable that the advantaged node uses considerably more transmission energy than the disadvantaged node. Referring to [22], which derives two-user optimal power allocations that achieve arbitrary rate tuples on the boundary of C we explain this as follows. From Figure 2, the optimal rate pairs for given transmit powers p1 and p2 lie on the R EFERENCES 100 80 60 40 20 0 200 400 600 800 1000 1200 (a) Transmit powers of nodes U1 and U2 . 7 6 5 4 3 2 1 0 200 400 600 800 1000 1200 (b) Associated transmission rates achieved by nodes U1 and U2 . Fig. 4: Simulation results for the two-node transmission power problem. segment L3 . Equivalently, ∃̺ ∈ [0, 1] such that the rate (∗) (∗) pair for an arbitrary point R(∗) = (r1 , r2 ) on L3 is given by the interpolation R(∗) = ̺ · R(1) + (1 − ̺) · R(2) . We may interpret ̺ as being the priority assigned to each transmitting node by the U0 when SIC is being carried out. ̺ = 1 means that data from U1 is being decoded second, subject to a lower noise rate, while ̺ = 0 means the opposite decoding order. We may think of the mapping t 7→ ̺(t) as a time-varying priority. However, by calculating ̺(t) from the optimum powers and rates seen in Figure 4, we find that ̺(t) = 0, ∀t ∈ T such that p1 (t) > 0, p2 (t) > 0. In other words, the disadvantaged node is always given priority, which is why it uses less energy at the optimum, even though it always experiences a worse channel gain. V. C ONCLUSIONS We have presented a general optimization framework for joint control of propulsion and transmission energy for single/multi-hop communication links in robotic networks. The relaxation of transmission constraints to theoretic capacity bounds, with relatively mild assumptions on the mobility model, results in a nonlinear but convex OCP. We showed that optimizing over a fixed path, as opposed to a fixed trajectory, increases the feasible starting data by at least 30% for just a single node. For the fixed-trajectory two-node MAC simulation, the optimal solution has been presented and analyzed. Immediate extensions of this work include higher fidelity models, and analysis of the relay network encompassed in problem (8). Considering the overarching goal of real-time control, further developments will be closedloop analysis of the control strategy, and consideration of the computational burden and energy expenditure [3], [23] in the network. [1] J. Ko, A. Mahajan, and R. Sengupta, “A network-centric UAV organization for search and pursuit operations,” in Aerospace Conference Proceedings, 2002. IEEE, vol. 6, pp. 6–6, IEEE, 2002. [2] S. Wang, A. Gasparri, and B. Krishnamachari, “Robotic message ferrying for wireless networks using coarse-grained backpressure control,” in Globecom Workshops (GC Wkshps), 2013 IEEE, pp. 1386–1390, IEEE, 2013. [3] M. Thammawichai, S. P. Baliyarasimhuni, E. C. Kerrigan, and J. B. Sousa, “Optimizing communication and computation for multi-UAV information gathering applications,” arXiv preprint arXiv:1610.04091, 2016. [4] P. Zhan, K. Yu, and A. L. Swindlehurst, “Wireless relay communications with unmanned aerial vehicles: Performance and optimization,” IEEE Transactions on Aerospace and Electronic Systems, vol. 47, no. 3, pp. 2068–2085, 2011. [5] I. Bekmezci, O. K. Sahingoz, and Ş. Temel, “Flying ad-hoc networks (FANETs): A survey,” Ad Hoc Networks, vol. 11, no. 3, pp. 1254– 1270, 2013. [6] W. Zhao and M. H. Ammar, “Message ferrying: Proactive routing in highly-partitioned wireless ad hoc networks,” in Distributed Computing Systems, 2003. FTDCS 2003. Proceedings. The Ninth IEEE Workshop on Future Trends of, pp. 308–314, IEEE, 2003. [7] Y. Yan and Y. Mostofi, “Robotic router formation in realistic communication environments,” IEEE Transactions on Robotics, vol. 28, no. 4, pp. 810–827, 2012. [8] P. Sujit, S. Saripalli, and J. B. Sousa, “Unmanned aerial vehicle path following: A survey and analysis of algorithms for fixed-wing unmanned aerial vehicles,” IEEE Control Systems, vol. 34, no. 1, pp. 42–59, 2014. [9] Y. Yan and Y. Mostofi, “To go or not to go: On energy-aware and communication-aware robotic operation,” IEEE Transactions on Control of Network Systems, vol. 1, no. 3, pp. 218–231, 2014. [10] Y. Zeng, R. Zhang, and T. J. Lim, “Throughput maximization for mobile relaying systems,” arXiv preprint arXiv:1604.02517, 2016. [11] D. Tse and P. Viswanath, Fundamentals of wireless communication. Cambridge university press, 2005. [12] L. Ren, Z. Yan, M. Song, and J. Song, “An improved water-filling algorithm for mobile mimo communication systems over time-varying fading channels,” in Electrical and Computer Engineering, 2004. Canadian Conference on, vol. 2, pp. 629–632, IEEE, 2004. [13] D. Gunduz and E. Erkip, “Opportunistic cooperation by dynamic resource allocation,” IEEE Transactions on Wireless Communications, vol. 6, no. 4, 2007. [14] D. N. C. Tse and S. V. Hanly, “Multiaccess fading channels. I. polymatroid structure, optimal resource allocation and throughput capacities,” IEEE Transactions on Information Theory, vol. 44, no. 7, pp. 2796–2815, 1998. [15] Y. Mei, Y.-H. Lu, Y. C. Hu, and C. G. Lee, “Energy-efficient motion planning for mobile robots,” in Robotics and Automation, 2004. Proceedings. ICRA’04. 2004 IEEE International Conference on, vol. 5, pp. 4344–4349, IEEE, 2004. [16] U. Ali, H. Cai, Y. Mostofi, and Y. Wardi, “Motion and communication co-optimization with path planning and online channel estimation,” arXiv preprint arXiv:1603.01672, 2016. [17] Y. Zeng and R. Zhang, “Energy-efficient UAV communication with trajectory optimization,” IEEE Transactions on Wireless Communications, vol. 16, no. 6, pp. 3747–3760, 2017. [18] P. Falugi, E. Kerrigan, and E. Van Wyk, “Imperial College London Optimal Control Software User Guide (ICLOCS).” http://www.ee.ic.ac.uk/ICLOCS/, 2010. [19] S. Boyd and L. Vandenberghe, Convex optimization. Cambridge university press, 2004. [20] A. L. Yuille and A. Rangarajan, “The concave-convex procedure,” Neural computation, vol. 15, no. 4, pp. 915–936, 2003. [21] S. Wolf, “An introduction to duality in convex optimization,” Network, vol. 153, 2011. [22] R. S. Cheng and S. Verdú, “Gaussian multiaccess channels with ISI: Capacity region and multiuser water-filling,” IEEE Transactions on Information Theory, vol. 39, no. 3, pp. 773–785, 1993. [23] S. Nazemi, K. K. Leung, and A. Swami, “QoI-aware tradeoff between communication and computation in wireless ad-hoc networks,” in Personal, Indoor, and Mobile Radio Communications (PIMRC), 2016 IEEE 27th Annual International Symposium on, pp. 1–6, IEEE, 2016. 

1 Quality and Diversity Optimization: A Unifying Modular Framework Abstract—The optimization of functions to find the best solution according to one or several objectives has a central role in many engineering and research fields. Recently, a new family of optimization algorithms, named Quality-Diversity optimization, has been introduced, and contrasts with classic algorithms. Instead of searching for a single solution, Quality-Diversity algorithms are searching for a large collection of both diverse and high-performing solutions. The role of this collection is to cover the range of possible solution types as much as possible, and to contain the best solution for each type. The contribution of this paper is threefold. Firstly, we present a unifying framework of Quality-Diversity optimization algorithms that covers the two main algorithms of this family (Multi-dimensional Archive of Phenotypic Elites and the Novelty Search with Local Competition), and that highlights the large variety of variants that can be investigated within this family. Secondly, we propose algorithms with a new selection mechanism for Quality-Diversity algorithms that outperforms all the algorithms tested in this paper. Lastly, we present a new collection management that overcomes the erosion issues observed when using unstructured collections. These three contributions are supported by extensive experimental comparisons of Quality-Diversity algorithms on three different experimental scenarios. QD-algorithm Collection of diverse and high-performing solutions Quality arXiv:1708.09251v1 [cs.NE] 12 May 2017 Antoine Cully and Yiannis Demiris, Senior Member, IEEE Coverage search space descriptor space Previously encountered solution (not stored) Solution contained in the collection Fig. 1: The objective of a QD-algorithm is to generate a collection of both diverse and high-performing solutions. This collection represents a (model free) projection of the highdimensional search space into a lower dimensional space defined by the solution descriptors. The quality of a collection is defined by its coverage of the descriptor space and by the global quality of the solutions that are kept in the collection. evolutionary algorithms are used to generate neural networks, robot behaviors, or objects [9], [10]. However, from a more general perspective and in contrast Index Terms—Optimization Methods, Novelty Search, Quality- with Artificial Evolution, Natural Evolution does not produce Diversity, Behavioral Diversity, Collection of Solutions. one effective solution but rather an impressively large set of different organisms, all well adapted to their respective environment. Surprisingly, this divergent search aspect of I. I NTRODUCTION Natural Evolution is rarely considered in engineering and Searching for high-quality solutions within a typically high- research fields, even though the ability to provide a large dimensional search space is an important part of engineering and diverse set of high-performing solutions appears to be and research. Intensive work has been done in recent decades promising for multiple reasons. to produce automated procedures to generate these solutions, For example, in a set of effective solutions, each provides which are commonly called “Optimization Algorithms”. The an alternative in the case that one solution turns out to be less applications of such algorithms are numerous and range from effective than expected. This can happen when the optimization modeling purposes to product design [1]. More recently, process takes place in simulation, and the obtained result does optimization algorithms have become the core of most machine not transfer well to reality (a phenomenon called the reality learning techniques. For example, they are used to adjust gap [11]). In this case, a large collection of solutions can the weights of neural networks in order to minimize the quickly provide a working solution [4]. Maintaining multiple classification error [2], [3], or to allow robots to learn new solutions and using them concurrently to generate actions or behaviors that maximize their velocity or accuracy [4], [5]. predict actions when done by other agents has also been shown Inspired by the ability of natural evolution to generate to be very successful in bioinspired motor control and cognitive species that are well adapted to their environment, Evolutionary robotics experiments[12]. Computation has a long history in the domain of optimization, Moreover, most artificial agents, like robots, should be able particularly in stochastic optimization [6]. For example, evolu- to exhibit different types of behavior in order to accomplish tionary methods have been used to optimize the morphologies their mission. For example, a walking robot needs to be able and the neural networks of physical robots [7], and to infer to move not only forwards, but in every direction and at the equations behind collected data [8]. These optimization different speeds, in order to properly navigate in its environment. abilities are also the core of Evolutionary Robotics in which Similarly, a robotic arm needs to be able to reach objects at different locations rather than at a single, predefined target. A. Cully and Y. Demiris are with the Personal Robotics Laboratory, Department of Electrical and Electronic Engineering, Imperial College London, Despite this observation, most optimization techniques that U.K. (e-mail: a.cully@imperial.ac.uk; y.demiris@imperial.ac.uk). are employed to learn behaviors output only a single solution: 2 the one which maximizes the optimized function [10], [7], [5]. tend to the global-optimum and the diversity of the produced Learning generic controllers that are able to solve several tasks walking behaviors will not be enough to properly control the is particularly challenging, as it requires testing each solution robot. For instance, it will not contain slow behaviors, which on several scenarios to assess their quality [13]. The automatic are essential for the robot’s manoeuvrability. This example creation of a collection of behaviors is likely to overcome these illustrates that sampling the entire range of possible solutions limitations and will make artificial agents more versatile. is not always related to searching for the local optima, and why The diversity of the solutions could also be beneficial for it may be useful to have the diversity preservation mechanism the optimization process itself. The exploration process may not correlated with the performance function, but rather based find, within the diversity of the solutions, stepping stones that on differences in the solution type. allow the algorithm to find even higher-performing solutions. Similarly, the algorithms may be able to solve a given problem B. Searching for Diverse Solutions faster if they can rely on solutions that have been designed Following this idea of a non performance-based diversity for different but related situations. For example, modifying mechanism, the Novelty Search algorithm [18] introduces the an existing car design to make it lighter might be faster than idea of searching for solutions that are different from the inventing a completely new design. previous ones, without considering their quality. This concept Attracted by these different properties several recent works, is applied by optimizing a “novelty score” that characterizes the such as Novelty Search with Local Competition [14] and the difference of a solution compared to those already encountered, MAP-Elites algorithm [15], started to investigate the question which are stored in a “novelty archive”. The novelty archive is of generating large collections of both diverse and high- independent from the population of the evolutionary algorithm. performing solutions. Pugh et al. [16], [17] nicely named this The novelty score is computed as the average distance of question as the Quality-Diversity (QD) challenge. the k-nearest neighboring solutions that currently are in the After a brief description of the origins of QD-algorithms novelty archive, while the distances are computed according in the next section, we unify these algorithms into a single to a user-defined solution descriptor (also called a behavioral modular framework, which opens new directions to create characterization, or behavioral descriptor [18], [13]). When the QD-algorithms that combine the advantages of existing meth- novelty score of a solution exceeds a pre-defined threshold, ods (see section III). Moreover, we introduce a new QD- this solution is added to the archive and thus used to compute algorithm based on this framework that outperforms the the novelty score of future solutions. existing approaches by using a new selective pressure, named The main hypothesis behind this approach is that, in some the “curiosity score”. We also introduce a new archive cases, the optimal solutions cannot be found by simply maximanagement approach for unstructured archives, like the mizing the objective function. This is because the algorithm novelty archive [18]. The performance of these contributions is first needs to find stepping stones that are ineffective according assessed via an extensive experimental comparison involving to the objective function, but lead to promising solutions numerous variants of QD-algorithms (see section IV). After afterwards. A good illustration of this problem is the “deceptive the conclusion, we introduce the open-source library designed maze” [18] in which following the objective function inevitably for this study, which can be openly used by interested readers leads to a dead-end (a local extremum). The algorithm has to (see section VI). investigate solutions that lead the agent further from the goal before being able to find solutions that actually solve the task. The authors of Novelty Search also introduced the “Novelty II. R ELATED W ORKS AND D EFINITIONS Search with Local Competition” algorithm (NSLC) [14], in While the notion of Quality-Diversity is relatively recent, which the exploration focuses on solutions that are both novel the problem of finding multiple solutions to a problem is a (according to the novelty score) and locally high-performing. long-standing challenge. The main insight consists of comparing the performance of a solution only to those that are close in the descriptor space. A. Searching for Local Optima This is achieved with a “local quality score” that is defined This challenge was first addressed by multimodal function as the number of the k-nearest neighboring solutions in the optimization algorithms, including niching methods in Evolu- novelty archive with a lower performance (e.g., slower walking tionary Computation [19], [20], [21], which aim to find the speed [14]) than the considered solution. The exploration is then local optima of a function. These algorithms mainly involve achieved with a multi-objective optimization algorithm (e.g., niche and genotypic diversity preservation mechanisms [21], NSGA-II [26]) that optimizes both the novelty and local quality like clustering [22] and clearing [23] methods. scores of the solutions. However, the local quality score does However, in many applications, some interesting solutions not influence the threshold used to select whether an individual are not captured by the local-optima of the fitness function. For is added to the novelty archive. The final result of NSLC is example, it is important for walking robots to be able to control the population of the optimization algorithm, which contains the walking speeds, however, there is no guarantee that the solutions that are both novel and high-performing compared performance function (i.e., the walking speed [24], [25]) will to other local solutions. In other words, the population gathers show local-optima that are diverse enough to provide a complete solutions that are both different from those saved in the novelty range of walking speeds. Typically, if the optimized function is archive, and high-performing when compared to similar types mono-modal (i.e., without local-optima), the population would of solutions. 3 The first applications of NSLC consisted of evolving both the morphology and the behavior of virtual creatures in order to generate a population containing diverse species, ranging from slow and massive quadrupeds to fast and lightweight unipedal hoppers by comparing velocity only between similar species [14]. In this experiment, the solution descriptor was defined as the height, the mass and the number of active joints, while the quality of the solutions was governed by their walking speed. At the end of the evolutionary process, the population contained 1,000 different species. These results represent the very first step in the direction of generating a collection of diverse and high-performing solutions covering a significant part of the spectrum of possibilities. C. Gathering and Improving these Solutions into Collections optimizing each solution independently (at least 5 times faster and about 10 times more accurate [13]). By “recycling” and improving solutions that are usually discarded by traditional evolutionary algorithms, the algorithm is able to quickly find necessary stepping stones. This observation correlates with the earlier presented hypothesis that QD-algorithms are likely to benefit from the diversity contained in the collection to improve their optimization and exploration abilities. However, it has been noticed that the archive improvement mechanism may “erode” the borders and alter the coverage of the collection [29]. Indeed, there are cases where the new, and better, solution found by the algorithm is less novel than the one it will replace in the archive. For instance, if high-performance can be more easily achieved for a solution in the middle of the descriptor space, then it is likely that the solutions near the borders will progressively be replaced by slightly better, but less novel, solutions. In addition to eroding the borders of the collection, this phenomenon will also increase the density of solutions in regions with a high performance. For instance, this phenomenon has been observed in the generation of collections containing different walking and turning gaits [29]. The novelty archive of the original NSLC algorithm had a better coverage of the descriptor space (but with lower performance scores) than the one from the BR-Evolution, because it is easier for the algorithms to find solutions that make the robot walk slowly rather than solutions that make it walk fast or execute complex turning trajectories (In section III-A2 of this paper, we introduce a new archive management mechanism that overcomes these erosion issues). Instead of considering the population of NSLC as the result of the algorithms, Cully et al. [13] suggested to consider the novelty archive as the result. Indeed, the aim of the novelty archive is to keep track of the different solution types that are encountered during the process, and thus to cover as much as possible of the entire descriptor space. Therefore, the novelty archive can be considered as a collection of diverse solutions on its own. However, the solutions are stored in the collection without considering their quality: as soon as a new type of solution is found, it is added to archive. While this procedure allows the archive to cover the entire spectrum of the possible solutions, in the original version of NSLC only the first encountered solution of each type is added to the archive. This implies that when finding a better solution for a solution type already present in the archive, this solution is not added to the archive. This mechanism prevents the archive D. Evolving the Collection from improving over time. Following different inspirations from the works presented Based on this observation, a variant of NSLC, named above, the Multi-dimensional Archive of Phenotypic Elites “Behavioral Repertoire Evolution”(BR-Evolution [13]), has (MAP-Elites) algorithm [15] has been recently introduced. been introduced to progressively improve the archive’s quality While this algorithm was first designed to “illuminate” the by replacing the solutions that are kept in the archive with landscape of objective functions [30], it showed itself to be an better ones as soon as they are found. This approach has been effective algorithm to generate a collection of solutions that are applied to generate “Behavioral Repertoires” in robotics, which both diverse and high-performing. The main difference with consists of a large collection of diverse, but effective, behaviors NSLC and BR-Evolution is that, in MAP-Elites, the population for a robotic agent in a single run of an evolutionary algorithm. of the algorithms is the collection itself, and the selection, It has also been used to produce collections of walking gaits, mutations and preservation mechanisms directly consider the allowing a virtual six-legged robot to walk in every direction solutions that are stored in the collection. and at different speeds. The descriptor space is defined as the In MAP-Elites, the descriptor space is discretized and final position of the robot after walking for 3 seconds, while represented as a grid. Initially, this grid is empty and the the quality score corresponds to an orientation error. As we algorithm starts with a randomly generated set of solutions. reproduce this experiment in this paper, we provide additional After evaluating each solution and recording its associated descriptions and technical details in section IV-C. descriptor, these solutions are potentially added to the correThe concepts introduced with BR-Evolution have also later sponding grid cells. If the cell is empty, then the solution is been employed in the Novelty-based Evolutionary Babbling added to the grid, otherwise, only the best solution among the (Nov-EB) [27] that allows a robot to autonomously discover new one and the one already in the grid is kept. After the the possible interactions with objects in its environment. This initialization, a solution is randomly selected via a uniform work draws a first link between the QD-algorithms and the distribution among those in the grid, and is mutated. The domain of developmental robotics, which is also studied in new solution obtained after the mutation is then evaluated and several other works (see [28] for overview). fitted back in the grid following the same procedure as in One of the main results that has been demonstrated with BR- the initialization. This selection/mutation/evaluation loop is Evolution experiments is that this algorithm is able to generate repeated several millions times, which progressively improves an effective collection of behaviors several times faster than by the coverage and the quality of the collection. 4 Definition II.1: Quality-Diversity optimization In one of its first applications, MAP-Elites was used to generate a large collection of different but effective ways A Quality-Diversity optimization algorithm aims to to walk in a straight line by using differently the legs of produce a large collection of solutions that are both as a six-legged robot. This collection of behaviors was then used diverse and high-performing as possible, which covers to allow the robot to quickly adapt to unforeseen damage a particular domain, called the descriptor space. conditions by selecting a new walking gait that still works in spite of the situation [4]. The same algorithm has also been used to generate behavioral repertoires containing turning gaits, While this definition is shared with the existing literature, similarly to the work described previously, and it was shown we also stress the importance of the coverage regularity of the that MAP-Elites generates better behavior collections while produced collections. In the vast majority of the applications being faster than the BR-Evolution algorithm [31]. presented previously, not only is the coverage of importance but The behaviors contained in these collections can be seen as its uniformity is as well. For example, in the locomotion tasks, locomotion primitives and thus can be combined to produce an even coverage of all possible turning abilities of the robot complex behaviors. Following this idea, the Evolutionary is required to allow the execution of arbitrary trajectories [29]. Based on this definition, the overall performance of a QDRepertoire-Based Control (EvoRBC [32]) evolves a neural network, called the “arbitrator”, that selects the appropriate algorithm is defined by the quality of the produced collection behavior in the repertoire, which was previously generated with of solutions according to three criteria: MAP-Elites. This approach has been applied on a four-wheeled 1) the coverage of the descriptor space; steering robot that has to solve a navigation task through a 2) the uniformity of the coverage; and maze composed of several sharp angles, and a foraging task 3) the performance of the solution found for each type. in which the robots needs to collect and consume as many objects as possible. F. Understanding the Underlying Mechanisms These applications take advantage of the non-linear dimenIn addition to direct applications, several other works focus sionality reduction provided by MAP-Elites. Indeed, both on studying the properties of QD-algorithms. For example, applications select behaviors from the descriptor space, which Lehman et al. [37] revealed that extinction events (i.e., erasing is composed of fewer than a dozen of dimensions (respectively, a significant part of the collection) increases the evolvability 36 to 6 dimensions [4] and 8 to 2 dimensions [32]), while the of the solutions [38] and allow the process to find higherparameter space often consists of several dozen dimensions. performing solutions afterwards. For example, with MAP-Elites, MAP-Elites has been employed in several other applications, erasing the entire collection except 10 solutions every 100 000 including the generation of different morphologies of soft generations increases the number of filled cells by 20% and the robots [15], or the production of images that are able to fool average quality of the solutions by 50% in some experimental deep neural networks [33]. It has also been used to create setups [37]. “innovation engines” that are able to autonomously synthesize In other studies, Pugh et al. [16], [17] analyzed the impact of pictures that resemble to actual objects (e.g., television, bagel, the alignment between the solution descriptor and the quality strawberry) [34]. score on both Novelty-based approaches (including NSLC) and However, the obligation to discretize the descriptor space MAP-Elites. For example, if the descriptor space represents may be limiting for some applications, and the uniform the location of the robot in a maze, and the quality score random selection may not be suitable for particularly large represents the distance between this position and the exit, collections, as it dilutes the selection pressure. Indeed, the then the descriptor space and the quality score are strongly uniform random selection of individuals among the collection aligned because the score can be computed according to the makes the selection pressure inversely proportional to the descriptor. The experimental results show that in the case number of solutions actually contained in the collection. A of such alignments with the quality score, then novelty-based simple way to mitigate this limitation is to use a biased approaches are more effective than MAP-Elites, and vice-versa. selection according to the solution performance or according Another study also reveals that the choice of the encoding to its novelty score (like introduced by Pugh et al. [16], (the mapping between the genotype and the phenotype) [17]). Another direction consists in having a number of cells critically impacts the quality of the produced collections [39]. irrespective of the dimensionality descriptor space, for example The experimental results link these differences to the locality by using computational geometry to uniformly partition the of the encoding (i.e., the propensity of the encoding to produce high-dimensional descriptor space into a pre-defined number of similar behaviors after a single mutation). In other words, the regions [35], or by using Hierarchical Spatial Partitioning [36]. behavioral diversity provided by indirect encoding, which is known to empower traditional evolutionary algorithms [40], appears to be counterproductive with MAP-Elites, while the locality of direct encodings allows MAP-Elites to consistently E. Quality-Diversity Optimization fill the collection of behaviors. Based on the seminal works presented previously [14], [15], These different works illustrate the interest of the community [13] and the formulation of Pugh et al. [16], [17], we can in QD-algorithms and that our understanding of the underlying outline a common definition: dynamics is only in its early stages. However, very few works 5 compare MAP-Elites and NSLC on the same applications (the • Finally, several scores, like the novelty, the local compefew exceptions being [16], [17], [31], [36]), or investigate tition, or the curiosity (defined in section III-B3) score, alternative approaches to produce collections of solutions. One are updated. of the goals of this paper is to introduce a new and common These four steps repeat until a stopping criterion is reached framework for these algorithms to exploit their synergies and (typically, a maximum number of iterations) and the algorithm to encourage comparisons and the creation of new algorithms. outputs the collection stored in the container. More details The next section introduces this framework. can be found in the pseudo-code of the algorithm, defined in Algorithm 1. In the following subsections, we will detail III. A UNITED AND MODULAR FRAMEWORK FOR different variants of the container, as well as the selection QD-O PTIMIZATION ALGORITHMS operators. As presented in the previous section, most works using or comparing QD-algorithms consider either MAP-Elites or NSLC-based algorithms, or direct comparisons of these two A. Containers The main purpose of a container is to gather all the solutions algorithms. These comparisons are relevant because of the distinct origins of these two algorithms. However, they only found so far into an ordered collection, in which only the provide high-level knowledge and do not provide much insight best and most diverse solutions are kept. One of the main of properties or particularities which make one algorithm better differences between MAP-Elites and NSLC is the way the collection of solutions is formed. While MAP-Elites relies on an than the other. In this section, we introduce a new and common framework N-dimensional grid, NSLC uses an unstructured archive based for QD-algorithms, which can be instantiated with different on the Euclidean distance between solution descriptors. These operators, such as different selection or aggregation operators, two different approaches constitute two different container similarly to most evolutionary algorithms. This framework types. In the following, we will detail their implementation demonstrates that MAP-Elites and NSLC can be formulated as and particularities. 1) The Grid: MAP-Elites employs an N-dimensional grid the same algorithm using a different combination of operators. Indeed, specific configurations of this framework are equivalent to form the collection of solutions [15], [4]. The descriptor to MAP-Elites or NSLC. However, this framework opens new space is discretized and the different dimensions of the grid perspectives as some other configurations lead to algorithms correspond to the dimensions of the solution descriptor. With that share the advantages of both MAP-Elites and NSLC. For this discretization, each cell of the grid represents one solution example, it can be used to design an algorithm that is as simple type. In the initial works introducing MAP-Elites, only one as MAP-Elites but working on an unstructured archive (rather solution can be contained in each cell. However, one can than a grid), or to investigate different selection pressures like imagine having more individuals per cell (like in [17] in NSLC. Moreover, this decomposition of the algorithms allows which two individuals are kept). Similarly, in the case of us to draw conclusions on the key elements that make an multi-objective optimization, each cell can represent the Pareto algorithm better than the others (e.g., the selective pressure or front for each solution type. Nevertheless, these considerations are outside the scope of this paper. the way to form the collection). This new formulation is composed of two main operators: a) Procedure to add solutions into the container: The 1) a container, which gathers and orders the solutions into procedure to add an individual to the collection is relatively a collection, and 2) the selection operator, which selects straight forward. If the cell corresponding to the descriptor of the solutions that will be altered (via mutations and cross- the individual is empty, then the individual is added to the grid. over) during the next batch (or generation). The selection Otherwise, if the cell is already occupied, only the individual operator is similar to the selection operators used in traditional with the highest performance is kept in the grid. evolutionary algorithms, except that it considers not only b) Computing the novelty/diversity of a solution: The the current population, but all the solutions contained in the inherent structure of the grid provides an efficient way to container as well. Other operators can be considered with this compute the novelty of each solution. Instead of considering new formulation, like the traditional mutation or cross-over the average distance of the k-nearest neighbors as a novelty operators. However, in this paper we only consider the operators score, like suggested in [18], here we can consider the number described above that are specific to QD-algorithms. of filled cells around the considered individual. The density of After a random initialization, the execution of a QD- filled cells of the sub-grid defined around the individual is a algorithm based on this framework follows four steps that good indicator of the novelty of the solution. Similarly to the are repeated: “k” parameter used in the k-nearest neighbors, the sub-grid is defined according to a parameter that governs its size, which • The selection operator produces a new set of individuals (bparents ) that will be altered in order to form the new is defined as ±k cells around the individual (in each direction). In this case, the score needs to be minimized. batch of evaluations (boffspring ). • The individuals of boffspring are evaluated and their 2) The Archive: The novelty archive introduced in the performance and descriptor are recorded. Novelty Search algorithm consists of an unstructured collection • Each of these individuals is then potentially added to of solutions that are only organized according to their descriptor the container, according to the solutions already in the and their Euclidean distance. As introduced in the BR-Evolution collection. algorithm [13], the novelty archive can be used to form the 6 Algorithm 1 QD-Optimization algorithm ( I iterations) (A ← ∅) for iter = 1 → I do if iter == 1 then bparents ← random() boffspring ← random() else bparents ← selection(A, boffspring ) boffspring ← random variation(bparents ) for each indiv ∈ boffspring do {desc, perf} ← eval(indiv) if add to container(indiv) then curiosity(parent(indiv))+ = Reward else curiosity(parent(indiv))− = Penalty update container() return A . Creation of an empty container. . The main loop repeats during I iterations. . Initialization. . The first 2 batches of individuals are generated randomly. . The next controllers are generated using the container and/or the previous batch. . Selection of a batch of individuals from the container and/or the previous batch. . Creation of a randomly modified copy of bparents (mutation and/or crossover). . Evaluation of the individual and recording of its descriptor and performance. . “add to container” returns true if the individual has been added to the container. . The parent gets a reward by increasing its curiosity score (typically Reward = 1). . Otherwise, its curiosity score is decreased (typically Penalty = 0.5). . Update of the attributes of all the individuals in the container (e.g. novelty score). collection of solutions by substituting solutions when better ones are found. In contrast with the grid container presented previously, the descriptor space here is not discretized and the structure of the collection autonomously emerges from the encountered solutions. a) Procedure to add solutions into the container: The management of the solutions is crucial with this container because it affects both the quality, and the final coverage of the collection. A first attempt was proposed in the BR-Evolution algorithm [13] by extending the archive management of the Novelty Search [18]: an individual is added to the archive if its novelty score exceeds a predefined threshold (which can be adjusted over time), or if it outperforms its nearest neighbor in the archive. In the second case, the nearest neighbor is removed from the archive and only the best of the two individuals is kept. While this archive management is relatively simple, further experiments reveal underlying limitations [29]. First, an individual with the same (or very close) descriptor as another individual can be added to the archive. Indeed, the novelty score, which is based on the average distance of the k-nearest neighbors, can still be relatively high even when two individuals are close if the rest of the collection is further. One of the consequences of using the novelty score as a criterion to add the solution in the container is that the collection is likely to show an uneven density of solutions [13], [29]. For example, experiments in these works show collections that contain a high density of solutions in certain regions (the inter-individuals distance being notably lower than the Novelty Score threshold used to add individual into the collection). While this property can be interesting for some applications, it mainly originates from a side effect. Second, the same experiments reveal that the replacement of individuals by better ones can erode the border of the collection, as discussed in the previous section. Indeed, in some cases, the individuals in the center of the collection show better performance than the ones in its border (because of the intrinsic structure of the performance function or because the center has been more intensively explored). This can lead to the replacement of individuals that are on the border of the collection by individuals that are closer to the center. This is an important limitation as it reduces the A l C Quality Zone dominating I1 𝜀 x N1 I2 B Q1 I1 N1 𝜀 x Q1 Novelty Fig. 2: Management of collections of solutions based on an unstructured archive. A) A solution is directly added to the collection if its nearest neighbor from the collection is further than l. B) Conversely, if the distance is smaller than l (i.e., if the circles overlap), the new solution is not automatically added to the collection, but competes against its nearest neighbor. If the new solution dominates the one already in the collection, then the new solution replaces the previous one. C) In the strict -domination, a solution dominates another one if the progress in one objective is larger than the decrease in the other objective (up to a predefined value ). coverage of the collection, as shown in [29]. In order to mitigate these limitations, we propose the following new way to manage solutions in the archive. A solution is added to the archive if the distance to its nearest neighbor exceeds a predefined threshold l (see Fig. 2.A). This parameter defines the maximal density of the collection. The threshold is similar to the novelty score threshold used in the original Novelty Search algorithm, except that in this case we only consider the distance of the nearest neighbor, and not the average distance of the k-nearest ones. If the distance between the new individual and its nearest neighbor is lower than l, then this new individual can potentially replace its nearest neighbor in the collection. This is only the case if its distance from its second nearest neighbor exceeds the l parameter, such that the distance among the solutions is preserved (see Fig. 2.B) and if it improves the overall quality of the collection. A new individual can improve the overall collection in two ways: 1) if it has a better quality, which increases the total quality of the collection or 2) if it has a better novelty score, which means that it extends the coverage 7 of the collection. This can be seen as two objectives that For these reasons, the choice of the most suitable container need to be maximized. From this perspective, we can use depends more on the considered applications, rather than on the definition of Pareto dominance to decide if an individual their performance. Therefore, while we will consider both of should replace another one already in the collection. Therefore, the containers in the experimental section of this paper, we a simple criterion could be to replace an existing individual, will not directly compare their results, as the comparison may only if it is dominated by the new one. However, this criterion not be fair and may be irrelevant with respect to the considered is very difficult to reach, as the new individual should be both applications. better and more diverse than the previous one. This prevents These two containers have been designed to provide uniform most new individuals from being added to the collection, which coverage of the descriptor space. However, experiments reveal limits the quality of the produced collections. that the accumulation of density on specific regions of the In order to soften this criterion, we introduce a variant of the descriptor space is a key factor for the Novelty Search -dominance [41], that we name the exclusive -dominance. In algorithm, as it allows the novelty score to constantly change this variant, we tolerate the dominating individual being worse over time. To avoid this issue, one can use an additional than the other individual according to one of the objectives (up container in which the density accumulates and that drives to a predefined percentage governed by ), only if it is better the exploration, while the other container gathers the collection on the other objective by at least the same amount (see Fig. that will be return to the user. In this paper, we will only focus 2.C). This criterion is more strict than the original -dominance, on variants using only one container, however we will consider which allows an individual to be dominated by another one extending the framework presented in this paper to multiple that is worse on both objectives. From a mathematical point containers in future works. of view, an individual x1 dominates x2 if these three points are verified: B. Selection Operators 1) N (x1 ) >= (1 − ) ∗ N (x2 ) The second main difference between MAP-Elites and NSLC 2) Q(x1 ) >= (1 − ) ∗ Q(x2 ) is the way the next batch, or population2 , of solutions is selected 3) (N (x1 )−N (x2 ))∗Q(x2 ) > −(Q(x1 )−Q(x2 ))∗N (x2 ) before being evaluated. On the one hand, MAP-Elites forms with N corresponding to the Novelty Score and Q to the the next batch by randomly selecting solutions that are already Quality (or performance) of an individual, which both need in the collection. On the other hand, NSLC relies on the current to be maximized1 . This set of conditions makes the addition population of solutions and selects the individuals that are both of new individuals in the container more flexible, but rejects novel and locally high-performing (according to a Pareto front). individuals that do not improve the collection. This difference is of significant importance as MAP-Elites uses The experimental results presented in section IV demonstrate the entire collection of solutions, while NSLC only considers that this new archive management overcomes the limitation of a smaller set of solutions. the previous approaches by producing collections with similar Similarly to the concept of containers, different approaches coverage and quality compared with the grid-based container. for selecting the individuals of the next batch can be considered. b) Computing the novelty of a solution: With the archive- In the following subsections, we will present several selection based container, the computation of the novelty score can be methods that can be employed with both container types. done with the traditional approach proposed by Lehman and 1) No Selection: A naive way to generate the next batch Stanley [18], which consists of the average distance of the of evaluation is to generate random solutions. However, this k-nearest neighbors. approach is likely ineffective because it makes the QD3) Partial Conclusion: These two different types of con- algorithm equivalent to a random sampling of the search tainers both present advantages and disadvantages. On the one space. In general, this approach provides an intuition about hand, the grid-based container provides a simple and effective the difficulty of the task and can be used as a base-line when way to manage the collection. However, it requires discretizing comparing alternative approaches. the descriptor space beforehand, which can be problematic 2) Uniform Random Selection: A second way to select for example if the discretization is not adequate, or needs to solutions that will be used in the next batch is to select solutions be changed over time. On the other hand, the archive-based with a uniform probability from those that are already in the container offers more flexibility, as it only requires the definition collection. This approach is the one used in MAP-Elites and of a distance in the descriptor space. For example, specific follows the idea that promising solutions are close to each other. distances can be used to compare complex descriptors, like In addition to being relatively simple, this approach has the images, without a strong knowledge of the structure of the advantage of being computationally effective. However, one of descriptor space (e.g., number of dimensions or limits) [27]. its main drawbacks is that the selection pressure decreases as However, this advantage is a disadvantage as well, because it the number of solutions in the collection increases (the chance implies that the algorithm needs to find the appropriate structure for a solution to be selected being inversely proportional to of the collection on its own, which represents an additional 2 We use the word batch instead of generation because most of the approaches challenge compared to the grid-based container. 1 This definition could very likely be generalized to more than two objectives, but this question is beyond the scope of this paper. presented in this paper can be used in a “steady state”, selecting and evaluating only one individual at each iteration. However, considering the selection and evaluation in batches allows the algorithm to execute the evaluation in parallel, which increases the computational efficiency of the algorithm. 8 the number of solutions in the collection), which is likely to be ineffective with large collections. 3) Score Proportionate Selection: An intuitive way to mitigate the loss of selection pressure from the random selection is to bias the selection according to a particular score. Similarly to traditional evolutionary algorithms, the selection among the solutions of the collection can be biased according to their quality (fitness), following the roulette wheel or the tournamentbased selection principles [42]. Other scores can also be considered to bias the selection. For example, the novelty score of each solution can substitute for the quality score for fostering the algorithm to focus on solutions that are different from the others. In addition to these two scores, in this paper we introduce a new score, named the Curiosity Score, that can be used to bias the selection and which is defined as follows: Definition III.1: Curiosity Score fitter than any yet existing [46]. One important aspect shared by these two definitions is that the score or the evolvability may dynamically change over time according to the state of the population or the collection, which is rarely considered in evolvability’s definitions. For instance, the definition often used in Evolutionary Computation [38], [45], [30], which considers that the evolvability captures the propensity of random variations to generate phenotypic diversity, depends on the genome of the individual but not on the state of the population. 4) Population-based Selection: All selection approaches described so far select the individuals from the solutions contained in the collection. This represents one of the main differences introduced by MAP-Elites compared to NSLC and traditional evolutionary algorithms, as the collection becomes the “population” of the algorithm and this population progressively grows during the evolutionary process. However, to handle the selection, we can consider employing populations The curiosity score represents the propensity of an in parallel to the collection. This is in line with the Novelty individual to generate offspring that are added to the Search algorithm which computes the novelty score based collection. on the Collection (the Novelty archive), but instead uses a traditional population to handle the selection. A practical implementation (see Algorithm 1) consists of This approach can be included in the framework proposed increasing the curiosity score of an individual (initially set to in this paper by initializing the population with the first batch zero) each time one of its offspring is added to the collection. and then, after each batch evaluation, a new population can be Conversely, when an offspring fails to be added to the archive generated based on the individuals of the current population (because it is not novel or high-performing enough), the (boffspring ) and their parents (bparents ). Classic selection Curiosity Score is decreased. In this paper, we use respectively approaches, like tournament or score proportionate, can be 1 and −0.5 for the reward and the penalty values. With this employed to select the individuals that will be part of the implementation, individuals may gain momentum, but this next population. Like in the collection-based selection, the means that such individual will be selected more often, making selection can be biased according to either the quality, novelty their score more likely to rapidly decrease. or curiosity scores. We named this score “Curiosity” because it encourages 5) Pareto-based Selection: The population-based selection the algorithm to focus on individuals that produce interesting approach can be extended to multi-objective selection, via solutions, until nothing new is produced. In other words, the the Pareto ranking, by taking inspiration from the NSGA-II algorithm focuses on regions of the search space as long as algorithm [26]. In this paper, we will particularly consider they produce interesting results, then, when the algorithm a Pareto-based selection operator that takes both the novelty gets “bored”, it focuses its attention on different regions. This score and the local quality score (number of neighbors that behavior is similar to the one of the “Intelligent Adaptive outperform the solution) of the individuals into account. This Curiosity” [43], while the implementation and the inspiration selection operator is similar to the selection procedure of NSLC. are strictly different. 6) Partial Conclusion: These different selection operators A similar approach has recently been introduced to bias can all be equally used with both of the containers presented the selection by using the same kind of successful offspring in the previous section. While the choice of the container counter [44]. The difference is that, in this paper, the counter is influences the type of the produced results (e.g., unstructured initialized to a fixed value (i.e., 10 in [44]) instead of starting or discretized descriptor space, see section III-A3), the selecat 0 like with the curiosity score, and that when an offspring tion operators will only influence the quality of the results. is added to the collection, the counter is not incremented (like Therefore, it is of importance to know which operators provide with the curiosity score), but rather reset to its maximal value. the best collection of solutions. In the following section, we This difference make the selection process more forgivable, as provide a first answer to this question by comparing the only one successful offspring is enough to make its parent very collections produced by the different selection operators and likely to be selected again. While it would be interesting to by investigating their behaviors. compare the effect of these two different, but related, methods, this comparison is out of the scope of this paper. IV. E XPERIMENTAL C OMPARISONS Although there is no overall agreement on the definition of evolvability [45], we can note that our definition of the curiosity To compare the different combinations of containers and score shares similarities with some of the first definitions of selection operators, we consider three experimental scenarios evolvability, like the one given by Lee Altenberg who defines that take place in simulation: 1) a highly redundant robotic the evolvability as the ability of a population to produce variants arm discovering how to reach points in its vicinity, 2) a virtual 9 TABLE I: The different combinations of containers and selection operators that are evaluated in this paper. The variants in bold are tested on the three experimental scenarios while the others are only tested on the first one. Variant name arch no selection arch random arch pareto arch fitness arch novelty arch curiosity arch pop fitness arch pop novelty arch pop curiosity grid no selection grid random grid pareto grid fitness grid novelty grid curiosity grid pop fitness grid pop novelty grid pop curiosity NSLC Container archive archive archive archive archive archive archive archive archive grid grid grid grid grid grid grid grid grid grid Selection Op. noselection random Pareto Score-based Score-based Score-based Population-based Population-based Population-based noselection random Pareto Score-based Score-based Score-based Population-based Population-based Population-based Population & archive based Considered Value Novelty & Local Quality Fitness Novelty Curiosity Fitness Novelty Curiosity Novelty & Local Quality Fitness Novelty Curiosity Fitness Novelty Curiosity Novelty & Local Quality Related approach Random Search / Motor Babbling MAP-Elites with Novelty [16] Traditional EA Novelty Search [18] Random Search / Motor Babbling MAP-Elites [15] Traditional EA Novelty Search with Local Competition [14] six-legged robot learning to walk in every direction, and 3) the same robot searching for a large number of ways to walk on a straight line. In addition to the tested combinations of containers and selection operators, we include the original Novelty Search with Local Competition algorithm (NSLC, [14]) in our experimental comparisons in order to assess the influence of the lack of density accumulation in the descriptor space, as discussed in section III-A3. Like in [16], all individuals of the population are potentially added to a grid container (the same as the one used with the others variants) after each generation. We then used the produced grid container to compare NSLC with the other variants. For this experiment, we used the implementation of NSLC provided by the Sferesv2 framework [47]. In the experiments presented in this paper, we only consider direct encodings with genomes that are small and fixed in size. It would be interesting to see how the conclusion drawn from these experiments hold with large genomes, genomes of increasing complexity over generations, or indirect encodings. For instance, [39] highlights that indirect encodings may have a negative impact on QD-algorithms. However, these further considerations are out of the scope of this paper and will be considered in future works. be improved either by finding additional individuals or by improving those already in the collection. It corresponds to the metric named “Quality Diversity” used in [16]. 4) Total Novelty: This metric is similar to the previous one, except that the sum considers the novelty score and not the quality value. This metric indicates if the collection is well distributed over the description space or rather if the solutions are highly concentrated. This metric will not be considered for collections produced with the grid-based container because the distribution of the solutions is forced by the grid. Other metrics: In [15], [39], the authors presented other metrics to compare collections produced by MAP-Elites. However, the main idea of these metrics is to normalize the quality of each solution by the maximal quality that can be achieved by each type of solution (i.e., by each grid cell). To infer the highest possible quality for each cell, the authors selected the best solution found by all the algorithms over all the replicates. However, this approach is only feasible with the grid-based container because the continuous descriptor space used in the archive-based container makes it challenging to associate and compare each “solution type”. For this reason, in this paper we decided to only consider the four metrics presented previously. A. Quality Metrics B. The Redundant Arm In order to compare the quality of the collections generated 1) Experimental Setup: In this first experimental comparison, by each variant, we define four quality metrics that characterize we consider a redundant and planar robotic arm with 8 degrees both the coverage and the performance of the solutions: of freedom that needs to discover how to reach every point 1) Collection Size: indicates the number of solutions con- in its vicinity. The quality function captures the idea that all tained in the collection and thus refers to the proportion of the joints of the arm should contribute equally to the movement, descriptor space that is covered by the collection. which allows quick transitions from one configuration to the 2) Maximal Quality: corresponds to the quality of the best next one. This constraint is defined by the variance of the solution contained in the collection and indicates if the global angular position of the joints when the robot reaches its final extremum of the performance function has been found (if it is configuration, and needs to be minimized by the algorithm. known). This experimental setup illustrates the need of quality-diversity 3) Total Quality: is the sum of the qualities over all the algorithms because it needs to simultaneously find a solution solutions contained in the collection. This metric provides for all the reachable positions and to optimize the quality information on the global quality of the collection as it can function for each of them. 10 No_selection Random Pareto No_selection Random Pareto Fitness Novelty Curiosity Fitness Novelty Curiosity NSLC 0 -0.1 Pop_fitness Pop_curiosity Pop_novelty Pop_fitness Pop_novelty Pop_curiosity Quality (rad2) -0.2 -0.3 -0.4 -0.5 -0.6 -0.7 Archive-based container Grid-based container Fig. 3: Typical collections of solutions produced with QD-algorithms. These collections consist of several thousand colored dots or cells that represent the final position of the gripper. The color of each dot or cell indicates the quality of the solution (lighter is better). TABLE II: Parameter values used the experiments. Parameters Batch size No. of Iterations Descriptor size Genotype size Genotype type Crossover Mutation rate for each parameter Mutation type Grid container: Grid size Sub-grid depth Archive container: l  k NSLC variant: ρinit k First exp 200 50.000 2 8 disabled Second exp 200 10.000 2 36 sampled values disabled Third exp 200 20.000 6 36 sampled values disabled 12.5% 5% 5% Polynomial Random new value Random new value 100 ∗ 100 ±3 cells 100 ∗ 100 ±5 cells 5 cells/dim ±1 cells 0.01 0.1 15 0.01 0.1 15 0.25 0.1 15 0.01 15 0.01 15 1 15 real values To simulate the robotic arm, we consider its kinematic structure, which provides the location of its gripper according to the angular position of all joints. The solutions that are optimized by the algorithms consist of a set of angular positions that govern the final configuration of the different joints of the robot. Neither the trajectory of the robot between its initial and final positions, nor internal collisions are simulated in this experiment. The solution descriptor is defined as the final position of the gripper, which is then normalized according to a square bounding box to have values between 0 and 1. The size of the bounding box is 2 ∗ 1.1 ∗ L, where L is the total length of the robot when totally deployed (the factor 1.1 is used to leave some room between the border of the descriptor space and the robot). The center of the box corresponds to the location of the robot’s base. An extensive set of configurations from the QD-algorithm framework (see algorithm 1) has been tested on this experimental setup (see Table I), and the execution of each of those variants has been replicated 20 times. The parameter values used for this experiment can be found in Table II. 2) Results: A typical collection of solutions produced by each of the tested variants is pictured in Figure 3. The collections using the archive-based container appear very similar to those using the other container type. This similarity, which holds in the other experiments as well, demonstrates that the archive management introduced in this paper successfully address the erosion issues described previously. Theoretically, the ideal result homogeneously covers a quasi-circular region and the performance (i.e., the color) should be arranged in concentric shapes resembling cardioids (inverted, heart-shaped curves)3 . This type of collection is found using the random, the fitness or the curiosity-based selection operators (over the collection) regardless of the container type used, as well as with the NSLC algorithm. The novelty based selection with the archive-based container also produces such a collection, while this is not the case with the grid-based container. It is interesting to note that the no-selection approach, which can be considered as a motor babbling or random search, is unable to produce the desired result. While the coverage is decent, the quality of the gathered solutions is not satisfactory. None of the population-based variants managed to produce a collection that both covers all the reachable space and contains high-performing solutions. This result could be explained by a convergence of the population toward specific regions of the collection. Typically, the population considering the fitness is likely to converge toward regions with high quality, whereas the population considering the novelty score converges to the 3 We can demonstrate that the points with the highest performance are located on a curve resembling a cardioid by computing the position of the end-effector for which all angular positions of the joints are set to the same angle (from −π/2 to +π/2). 11 5350 Archive 5000 5340 4000 5330 3000 5320 2000 5310 1000 0 0 Grid Maximal Quality Collection size 6000 1 2 3 5300 4 5 4.9 × 104 5 × 104 0 -0.05 -0.1 -0.15 -0.2 -0.25 -0.3 -0.35 -0.4 -0.45 0 7000 6280 0 6000 6260 -0.05 5000 6240 -0.1 6220 4000 -0.2 6180 -0.25 2000 6160 -0.3 1000 6140 -0.35 0 0 1 2 3 6120 4.9 4 5 × 104 Number of Iterations 2 3 2 -0.15 6200 3000 1 5 × 104 -0.4 0 Total Quality 0 -0.002 -0.004 -0.006 -0.008 -0.01 -0.012 -0.014 -0.016 -0.018 4 5 4.9 × 104 1 2 3 4 5 × 104 5 × 104 × 10-3 4500 4000 3500 3000 2500 2000 1500 1000 500 0 0 4000 -4 3000 -6 2000 -8 1000 -10 4.9 5 × 104 Number of Iterations 0 0 92.1 70 60 4320 1 2 3 4300 4 5 4.9 × 104 5040 5020 5000 3 4 5 × 104 4980 4.9 91.8 30 5 × 104 5060 2 91.9 40 5080 1 92 50 4310 5100 5000 92.2 80 4330 5120 0 92.3 90 4340 6000 -2 Total Novelty 100 4350 5 × 104 20 0 1 2 3 4 5 × 104 91.7 4.9 Number of Iterations pareto pop_curiosity pop_novelty pop_fitness curiosity novelty fitness random no_selection NSLC 5 × 104 Number of Iterations Fig. 4: Progression of the quality metrics in the redundant arm experiment. The first row depicts the results from variants using the archive-based container, while the second row considers variants with the grid-based container. Because of the difficulty to distinguish the different variants, a zoom on the best variants during the last 1000 batches is pictured on the right of each plot. The middle lines represent the median performance over the 20 replications, while the shaded areas extend to the first and third quartiles. In this experiment, the quality score is negative, thus in order to get a monotonic progression in the “Total Quality” metric, +1 is added to the Quality to have a positive score. border of the collection. The results of the variant using a population with the curiosity score could be explained by the difficulty to keep track of all individuals with a relatively small population (200 individuals in the population compared to about 6.000 in the entire collection). The curiosity score is dynamic, and changes during the evolutionary process (an individual can have a high curiosity score at one moment, for example if it reaches a new region of the descriptor space, and can have a very low curiosity score later during the process, for instance when the region becomes filled with good solutions). Therefore, it is likely that the individuals with the highest curiosity score are not contained in the population. score is continuous and the distribution of the solutions in the collection adds some variability in the novelty score, which makes it impossible to have several individuals with the lowest novelty score. While the Pareto-based selection is designed to be similar to the NSLC algorithm, by keeping in the population individuals that both have a high novelty and local-competition scores, we can see that the collection produced by NSLC is significantly better than the Pareto-based selection approach. We can explain this poor performance by the presence of a Pareto-optimal solution in this scenario. Indeed, the solution in which the robot has all his joint positions set to zero has the best fitness Moreover, we can observe different results between the grid- and is located on the border of the collection, which provides based and the archive-based container variants considering a high novelty score. It is worth noting that we can mitigate the novelty score. This difference is likely to originate from this issue by implementing a toroidal distance or container the fact that the novelty score is computed differently in (like in [17]), when such a representation is compatible with these two container types. Indeed, while in the archive- the descriptor space. This is not the case in our experiments. based container the novelty score follows the formulation A behavior that reaches one end of the reachable space of introduced by Lehman and Stanley [18], in the grid-based the robot is not meant to be considered similar to individuals container, the novelty score is computed based on the number that reach the opposite end of the reachable space. For these of individuals in the neighboring cells (see section III-A1b). reasons, the population is then very likely to converge to this Both of these expressions capture the density of solutions Pareto-optimal solution and thus, to neglect certain regions around the considered individuals. However, in the grid based of the collection. The size of the population is probably a container, the novelty score is discretized (because it is related limiting factor as well. A large number of equivalent solutions to the number of neighboring solutions). This discretization in terms of Pareto-dominance exist (all those in the center of is likely to have a strong impact on score-based selection the collection with the highest fitness), which makes it difficult variants using the novelty score because all individuals in the for the population to cover the entire descriptor space. center of the collection will have the same and lowest novelty NSLC is not impacted in the same way because the score (because of all neighboring cells being filled). In the original archive management allows the density to constantly score-based selection, individuals with the lowest score have accumulate around over-explored regions (for instance by nearly no chance of being selected, which makes the selection varying the novelty threshold, as described in [14]). Thanks focus on the border of the collection. This behavior is not to this feature, the novelty score constantly changes over time observed with the archive-based container because the novelty and makes pareto optimal solutions disappear quickly. Indeed, 12 the regions that contain pareto optimal solutions will rapidly This experimental setup has first been introduced in [13]. see their density increased making the novelty score of the Each potential solution consists of a set of 36 parameters (6 corresponding individuals less competitive compared with the per leg) that define the way each of the robot’s joint is moving rest of the population. (the controller is the same as the one used in [4]). During the It is important to note that the NSLC variant uses two evaluation of a solution, the robot executes the behavior defined containers and one population during the evolutionary process. by the parameters for three seconds, and its final position and The population and one of the containers (the novelty archive) orientation are recorded. The descriptor space is defined by the are used to drive the exploration process, while the second final position of the robot (X and Y coordinates), while the container (a grid-based container) gathers the collection that quality of the solution corresponds to the orientation error with respect to a desired orientation, which encourages the robot will be delivered to the user. The variations of the quality metrics (see Fig. 4) demonstrate to follow circular trajectories. These kinds of trajectories are that among all tested variants, the best collections are provided interesting for planning purposes as any arbitrary trajectory by variants which perform the selection based on the entire can be decomposed as a succession of circular arcs. In order to be able to chain circular trajectories, the robot needs to be collection. The coverage, maximal quality, total quality, and total novelty aligned with the tangent of these circles at the beginning and of the collections produced with selection operators considering the end of each movement. We can note that only one circular the entire collections is higher than those using population- trajectory goes through both the initial and final positions of based selection (all p-values < 7e − 8 from the Wilcoxon rank the robot with its tangent aligned with the initial orientation of sum tests4 , except for the “(grid/arch) pop fitness” approaches the robot. The difference between the final orientation of the which are not significantly different in terms of maximal quality robot and the direction of the tangent of this unique circular and for “grid novelty” which performs significantly worse than trajectory defines the orientation error, which is minimized by the other collection-based approaches). The only exception the QD algorithms (more details can be found in [13]). is the novelty-based selection with the grid-based container, The usage of the physical simulator makes the experiments which is unable to correctly fill the center of the collection, as significantly longer (between 4 and 5 hours are required to it focuses on its borders. perform 10,000 batches with one variant). For this reason, We can note that the variant using the Pareto-based selection the number of generations has been decreased to 10,000 and with the archive-based container produces collections that are only 10 variants (those in bold in Table I) are considered for better than those from variants using population-based selection, this experiment. This sub-set of variants includes variants that but worse than those produced by variants that consider the are related to MAP-Elites, NSLC, Motor Babbling, traditional entire collection for the selection. However, the Pareto-based population-based EA and the variant considering the curiosity selection shows the best results according to the maximal score over the entire collection. The execution of each of quality metrics. those variants has been replicated 10 times. The value of the While the difference among variants using the entire collec- parameters used for this experiment can be found in Table II. tion in the selection with the grid-based container is negligible, the curiosity-based selection appears to be significantly better 2) Results: From a high-level point of view, the same (even if the difference is small) than the other selection conclusion as previously can be drawn based on the resulting approaches on all the metrics with the archive-based container collections (see Fig. 5): The variants “no selection” and (all p-values< 2e − 4 for all the metrics except for the “pop fitness” produce worse collections than the other variants, total novelty in which p-values< 0.01). This observation while the variants “random”, “curiosity” and NSLC generate demonstrates that relying on individuals with a high-propensity the best collections. In this experiment, the “Pareto” variant to generate individuals that are added to the collection is a performs better than in the previous one. This result can be promising selection heuristic. explained by the absence of a unique Pareto-optimal solution. We can observe that the NSLC variant performs significantly The quality metrics indicate that the “curiosity” variants, better than the pareto-based approach and that its performance on both the grid and the archive containers, significantly is close to, but lower than, those of the variants that use outperform the other algorithms (see Fig. 6, all p-values < 0.01, selection operators considering the entire collections. except when compared to arch random in terms of total novelty in which p-value = 0.05). These results also demonstrate that this second experimental scenario is more challenging for the C. The Robot Walking in Every Direction 1) The Experimental Setup: In this second experimental algorithms, as the difference in the metrics is clear and the setup, we consider a six-legged robot in a physical simulator. performance of the naive “no selection” is very low. The objective of the QD-algorithms is to produce a collection of behaviors that allows the robot to walk in every direction and at different speeds. 4 The reported p-values should be compared to a threshold α (usually set to0.05) which is corrected to deal with the “Multiple Comparisons problem”. In this paper, all our conclusions about the significance of a difference is given by correcting α according to the Holm-Bonferroni method [48]. In this experiment, the NSLC variant shows similar results to the “random” variant (which corresponds to the MAP-Elites algorithm). In particular, the final size of the collection and the final total quality are not significantly different (p-values< 0.61). However, the performance of the “curiosity” approach remains significantly better on both aspects (p-values< 0.0047) compared to NSLC. 13 Population-based Selection wrt Fitness Pareto-based Selection Random Selection (novelty and local quality) (over the entire collection) Curiosity-based Selection (over the entire collection) Archive-based Collection No Selection Novelty Search with Local Competition -180 Back Grid-based Collection Front 1m -160 -140 -120 -100 -80 Quality (degree) -60 -40 -20 0 Right Left -1m 1m -1m Fig. 5: Typical collections of solutions produced by considered variants in the experiment with the virtual legged-robot learning to walk in every direction. The center of each collection corresponds to the starting position of the robot and the vertical axis represents the front axis of the robot. The position of each colored pixel or dot represent the final position of the robot after walking for 3 seconds and its color depicts the absolute (negative) difference between the robot orientation and the desired orientation. Lighter colors indicate better solutions. The collections are symmetrical because the robot learns how to walk both forward and backward. This possibility, as well as the overall shape of the collection is not predefined but rather autonomously discovered by the algorithms. Total Quality 7 3500 3000 2500 2000 1500 60 50 40 3 30 1 0 0 4000 6000 8000 10000 6000 0 10 2000 × 10 4000 6000 8000 10000 4000 6000 8000 10000 5 0 0.5 1 1.5 2 × 10 3000 4 2000 1000 2 0 2000 4000 6000 8000 Number of Iterations 10000 0 0 2000 4000 6000 8000 10000 pareto pop_fitness curiosity random no_selection NSLC Number of Iterations 800 600 400 500 0.2 200 0 0.5 1 1.5 4 12000 6 1000 0.3 14000 8 1200 1000 0.4 Number of Iterations 4000 Grid 2000 1400 1500 0.5 1000 0 1600 0.6 2000 10 0 Total Novelty 1800 2000 0.7 3000 0 Total Quality 2500 0.8 4000 20 5000 0 1 0.9 5000 70 4 500 Collection size Maximal Quality 6000 80 5 2 2000 Total Novelty 5 6 1000 0 × 10 Archive 8 4000 2 × 10 0 6000 1 5000 0.8 6000 0.6 0.5 1 1.5 2 × 10 1.2 8000 0 4 10000 Grid Archive Collection size 4500 4 4000 3000 2000 4000 0.4 2000 0 0 0.5 1 1.5 2 × 10 4 Number of Iterations 0.2 1000 0 0.5 1 1.5 2 × 10 4 Number of Iterations 0 0 0.5 1 1.5 2 × 10 0 0 0.5 1 1.5 2 × 10 4 Number of Iterationss 4 pareto pop_fitness curiosity random no_selection NSLC Number of Iterations Fig. 6: Progression of three quality metrics in the turning legged- Fig. 7: Progression of the four quality metrics in the experiment robot experiment. The progression of the maximal quality is not with the legged-robot learning different ways to walk in a depicted because all the variants found at least one solution with straight line. The first row depicts the results from variants the highest possible quality (i.e., 0) in fewer than 1.000 batches. using the archive-based container, while the second row The first row depicts the results from variants using the archive- considers variants with the grid-based container. The middle based container, while the second row considers variants with lines represent the median performance over the 10 replications, the grid-based container. The middle lines represent the median while the shaded areas extend to the first and third quartiles. performance over the 10 replications, while the shaded areas extend to the first and third quartiles. In this experiment, the quality score is negative, thus in order to get a monotonic The experiment has been replicated 10 times and the other progression in the “Total Quality” metric, +180 is added to parameters of the algorithm can be found in Table II. the Quality to have positive score. 2) Results: From a general point of view, the same conclusion as in the previous experiments can be drawn from the progression of quality metrics (see Fig.7)5 . Variants selecting D. The Robot Walking with Different Gaits individuals from the whole collection significantly outperform, 1) The Experimental Setup: In this third experimental setup, in terms of coverage, total quality and diversity, those that we use the same virtual robot as in the previous experiment consider populations (all the p-values< 2e − 4). In particular, with the same controller. However, in this case the robot has the curiosity-based selection operator shows the best results to learn a large collection of gaits to walk in a straight line as both with the grid-based and the archive-based containers. For instance, one can note that the total quality achieved by the fast as possible. This scenario is inspired by [4]. In this experiment, the quality score is the traveled distance after walking for 3 seconds, and the solution descriptor is the proportion of time that each leg is in contact with the ground. The descriptor space has thus 6 dimensions in this experiment. 5 Visualizations of the collections are not provided in this experiment because of the high-dimensionality of the descriptor-space. While the grid-based collections could have been depicted with the same approach as in [4], this approach cannot be applied with the archive-based container. 14 random selection (second best approach) after 20,000 batches, is achieved by the curiosity-based selection after only 11,000 batches with the archive-based container and 13,500 batches with the grid-based container. In contrast with the previous experiment, the “no selection” variants manage to achieve good coverage (about half of the coverage produced by the variants using the collection-wise selection). However, they show the worst results according to the total quality and the maximal quality metrics. The variants using the population-based selection with respect to the performance show the opposite results. While the coverage of this variant is the worst among all the evaluated variants with both of the container types, this selection approach, which is similar to a traditional EA, found the solutions with the best quality (the fastest way to walk). In particular, the performance achieved with this variant significantly outperforms the best solutions compared to every other variant, even those using the collection-wise selection (pvalues< 0.0017). This observation shows that the best variants tested so far are not always able to find the global extremum of the quality. The quality difference between the “pop fitness” variants and the others is smaller with the grid-based container than with the archive-based. This quality difference could be explained by the difference in the collection sizes, or the additional difficulty of finding the inherent structure of the collection for the archive-based container. The Pareto-based variants are low-performing in this experiment. They show neither a good coverage (similar to the “no selection” or the “pop fitness” variants) nor a good maximal quality (lower than the variants with a collection-wise selection). It is difficult to understand the reasons for such a low performance in this experiment, as the behavioral space is 6 dimensional, making it hard to visualize. However, it is likely that it happens for the same reasons as in the previous experiments, like a premature convergence to the border of the collection (which show relatively bad performance), or the existence of a Pareto-optimal solution. In contrast with the Pareto-based variants, NSLC achieves good coverage of the behavioral space in this experiment, while smaller than the “random” and “curiosity” ones. However, the maximal quality found on the produced collection is lower than most of the considered variants (p-values< 0.03746 except with the “no selection” variant, p-value= 0.9696), and the global quality of the collections is equivalent to those of the Pareto-based variant. V. C ONCLUSION AND D ISCUSSION In this paper, we presented three new contributions. First, we introduced a new framework that unifies QD-algorithms, showing for example that MAP-Elites and the Novelty Search with Local Competition are two different configurations of the same algorithm. Second, we suggested a new archive management procedure that copes with the erosion issues observed with the previous approaches using unstructured archives (like BR-evolution). This new procedure demonstrates good results 6 These p-values do not reject the null-hypothesis based on the HolmBonferroni method with a α = 0.05, but reject it with α = 0.1. as it allows the algorithms to produce unstructured collections with the same coverage as those with grid containers, which was not the case with the previous management procedure [31]. Finally, we proposed a new selective pressure specific for QDalgorithms, named “curiosity score” that shows very promising results by outperforming all the existing QD-algorithms on all the experiments presented in this paper. In addition to these three contributions, we presented the results of an experimental comparison between a large number of QD-algorithms, including MAP-Elites and NSLC. One of the main results that can be outlined from these experiments is that selection operators considering the collection instead of a population showed better performance on all scenarios. We can hypothesize that this results from the inherent diversity of solutions contained in the collection. Indeed, several works suggest that maintaining the behavioral diversity in populations of evolutionary algorithms (via additional objective for example) is a key factor to avoid local extremum and to find promising stepping stones [40], [18]. Another fundamental lesson learned from the experiments presented in this paper is about the importance of allowing the density of solutions to increase in diverse regions of the archive to obtain the full effectiveness the NSLC. This can be achieved by varying the novelty-score threshold or via probabilistic addition to the archive[37]. While such mechanisms are often used in the literature, their importance is rarely highlighted by experimental comparisons like in this paper. In particular, we demonstrated that algorithms using the novelty score, but with archives in which the density does not increase, are unable to show similar results to NSLC, because they are severely impacted by certain aspects of the fitness landscape (e.g., presence of Pareto-optimal solutions). This unified and modular framework for QD-algorithms is intended to encourage new research directions via novel container types, selection operators, or selective pressures that are specific to this domain. We expect that the emergence of new QD-algorithms will provide insights about the key factors for producing the best collection of solutions. VI. Q UALITY D IVERSITY L IBRARY The source code of the QD-algorithm framework is available at https://github.com/sferes2/modular QD. It is based on the Sferesv2 framework [47] and implements both the grid-based and archive-based containers and several selection operators, including all those that have been evaluated in this paper. The source code of the experimental setups is available at the same location and can be used by interested readers to investigate and evaluate new QD-algorithms. The implementation allows researchers to easily implement and evaluate new combinations of operators, while maintaining high execution speed. For this reason, we followed the policybased design in C++ [49], which allows developers to replace the behavior of the program simply by changing the template declarations of the algorithm. For example, changing from the grid-based container to the archive-based one only requires changing “container::Grid” to “container::Archive” in the template definition of the QD-algorithm object. Moreover, the 15 modularity provided by this design pattern does not add any overhead, contrary to classic Object-Oriented Programming design. Interested readers are welcome to use and to contribute to the source code. ACKNOWLEDGMENT This work was supported by the EU Horizon2020 project PAL (643783-RIA). The authors gratefully acknowledge the support from the members of the Personal Robotics Lab. R EFERENCES [1] A. Antoniou and W.-S. Lu, Practical optimization: algorithms and engineering applications. Springer Science & Business Media, 2007. [2] D. E. Rumelhart, G. E. Hinton, and R. J. Williams, “Learning representations by back-propagating errors,” Cognitive modeling, vol. 5, no. 3, p. 1, 1988. [3] S. J. Russell and P. Norvig, Artificial intelligence: a modern approach, 2003. [4] A. Cully, J. Clune, D. Tarapore, and J.-B. Mouret, “Robots that can adapt like animals,” Nature, vol. 521, no. 7553, pp. 503–507, 2015. [5] J. Kober, J. A. Bagnell, and J. Peters, “Reinforcement learning in robotics: A survey,” International Journal of Robotics Research, vol. 32, no. 11, p. 1238, 2013. [6] J. C. Spall, Introduction to stochastic search and optimization: estimation, simulation, and control. John Wiley & Sons, 2005, vol. 65. [7] H. Lipson and J. B. Pollack, “Automatic design and manufacture of robotic lifeforms,” Nature, vol. 406, no. 6799, pp. 974–978, 2000. [8] M. Schmidt and H. Lipson, “Distilling free-form natural laws from experimental data,” science, vol. 324, no. 5923, pp. 81–85, 2009. [9] A. E. Eiben and J. Smith, “From evolutionary computation to the evolution of things,” Nature, vol. 521, no. 7553, pp. 476–482, 2015. [10] J. Bongard, V. Zykov, and H. Lipson, “Resilient machines through continuous self-modeling,” Science, vol. 314, no. 5802, 2006. [11] S. Koos, J.-B. Mouret, and S. Doncieux, “The transferability approach: Crossing the reality gap in evolutionary robotics,” Evolutionary Computation, IEEE Transactions on, vol. 17, no. 1, pp. 122–145, 2013. [12] Y. Demiris, L. Aziz-Zadeh, and J. Bonaiuto, “Information processing in the mirror neuron system in primates and machines,” Neuroinformatics, vol. 12, no. 1, pp. 63–91, 2014. [13] A. Cully and J.-B. Mouret, “Behavioral repertoire learning in robotics,” in Proceedings of the 15th annual conference on Genetic and Evolutionary Computation. ACM, 2013, pp. 175–182. [14] J. Lehman and K. O. Stanley, “Evolving a diversity of virtual creatures through novelty search and local competition,” in Proceedings of the 13th annual conference on Genetic and Evolutionary Computation. ACM, 2011, pp. 211–218. [15] J.-B. Mouret and J. Clune, “Illuminating search spaces by mapping elites,” arXiv preprint arXiv:1504.04909, 2015. [16] J. K. Pugh, L. Soros, P. A. Szerlip, and K. O. Stanley, “Confronting the challenge of quality diversity,” in Proceedings of the 2015 on Genetic and Evolutionary Computation Conference. ACM, 2015, pp. 967–974. [17] J. K. Pugh, L. B. Soros, and K. O. Stanley, “Quality diversity: A new frontier for evolutionary computation,” Frontiers in Robotics and AI, vol. 3, p. 40, 2016. [18] J. Lehman and K. O. Stanley, “Abandoning objectives: Evolution through the search for novelty alone,” Evolutionary Computation, vol. 19, no. 2, pp. 189–223, 2011. [19] D. E. Goldberg and J. Richardson, “Genetic algorithms with sharing for multimodal function optimization,” in Genetic algorithms and their applications: Proceedings of the Second International Conference on Genetic Algorithms. Hillsdale, NJ: Lawrence Erlbaum, 1987, pp. 41–49. [20] S. W. Mahfoud, “Niching methods for genetic algorithms,” Urbana, vol. 51, no. 95001, pp. 62–94, 1995. [21] G. Singh and K. Deb Dr, “Comparison of multi-modal optimization algorithms based on evolutionary algorithms,” in Proceedings of the 8th annual conference on Genetic and Evolutionary Computation. ACM, 2006, pp. 1305–1312. [22] X. Yin and N. Germay, “A fast genetic algorithm with sharing scheme using cluster analysis methods in multimodal function optimization,” in Artificial neural nets and genetic algorithms. Springer, 1993. [23] A. Pétrowski, “A clearing procedure as a niching method for genetic algorithms,” in Evolutionary Computation, 1996., Proceedings of IEEE International Conference on. IEEE, 1996, pp. 798–803. [24] D. J. Lizotte, Practical bayesian optimization. University of Alberta, 2008. [25] N. Kohl and P. Stone, “Policy gradient reinforcement learning for fast quadrupedal locomotion,” in Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), vol. 3. IEEE, 2004, pp. 2619–2624. [26] K. Deb, A. Pratap, S. Agarwal, and T. Meyarivan, “A fast and elitist multiobjective genetic algorithm: Nsga-ii,” Evolutionary Computation, IEEE Transactions on, vol. 6, no. 2, pp. 182–197, 2002. [27] C. Maestre, A. Cully, C. Gonzales, and S. Doncieux, “Bootstrapping interactions with objects from raw sensorimotor data: a novelty search based approach,” in Development and Learning and Epigenetic Robotics (ICDL-EpiRob), 2015 Joint IEEE International Conference on. IEEE, 2015, pp. 7–12. [28] F. Benureau and P.-Y. Oudeyer, “Behavioral diversity generation in autonomous exploration through reuse of past experience,” Frontiers in Robotics and AI, vol. 3, p. 8, 2016. [29] A. Cully and J.-B. Mouret, “Evolving a behavioral repertoire for a walking robot,” Evolutionary Computation, 2015. [30] J. Clune, J.-B. Mouret, and H. Lipson, “The evolutionary origins of modularity,” Proceedings of the Royal Society of London B: Biological Sciences, vol. 280, no. 1755, 2013. [31] A. Cully, “Creative adaptation through learning,” Ph.D. dissertation, Université Pierre et Marie Curie, 2015. [32] M. Duarte, J. Gomes, S. M. Oliveira, and A. L. Christensen, “Evorbc: Evolutionary repertoire-based control for robots with arbitrary locomotion complexity,” in Proceedings of the 25th annual conference on Genetic and Evolutionary Computation. ACM, 2016. [33] A. Nguyen, J. Yosinski, and J. Clune, “Deep neural networks are easily fooled: High confidence predictions for unrecognizable images,” in Conference on Computer Vision and Pattern Recognition. IEEE, 2015. [34] A. M. Nguyen, J. Yosinski, and J. Clune, “Innovation engines: Automated creativity and improved stochastic optimization via deep learning,” in Proceedings of the 2015 Annual Conference on Genetic and Evolutionary Computation. ACM, 2015, pp. 959–966. [35] V. Vassiliades, K. Chatzilygeroudis, and J.-B. Mouret, “Scaling up map-elites using centroidal voronoi tessellations,” arXiv preprint arXiv:1610.05729, 2016. [36] D. Smith, L. Tokarchuk, and G. Wiggins, “Rapid phenotypic landscape exploration through hierarchical spatial partitioning,” in International Conference on Parallel Problem Solving from Nature. Springer, 2016, pp. 911–920. [37] J. Lehman and R. Miikkulainen, “Enhancing divergent search through extinction events,” in Proceedings of the 2015 on Genetic and Evolutionary Computation Conference. ACM, 2015, pp. 951–958. [38] D. Tarapore and J.-B. Mouret, “Evolvability signatures of generative encodings: beyond standard performance benchmarks,” Information Sciences, vol. 313, pp. 43–61, 2015. [39] D. Tarapore, J. Clune, A. Cully, and J.-B. Mouret, “How do different encodings influence the performance of the map-elites algorithm?” in Genetic and Evolutionary Computation Conference, 2016. [40] J.-B. Mouret and S. Doncieux, “Encouraging behavioral diversity in evolutionary robotics: An empirical study,” Evolutionary Computation, vol. 20, no. 1, pp. 91–133, 2012. [41] M. Laumanns, L. Thiele, K. Deb, and E. Zitzler, “Combining convergence and diversity in evolutionary multiobjective optimization,” Evolutionary Computation, vol. 10, no. 3, pp. 263–282, 2002. [42] D. E. Goldberg and K. Deb, “A comparative analysis of selection schemes used in genetic algorithms,” Foundations of genetic algorithms, 1991. [43] P.-Y. Oudeyer, F. Kaplan, and V. V. Hafner, “Intrinsic motivation systems for autonomous mental development,” Evolutionary Computation, IEEE Transactions on, vol. 11, no. 2, pp. 265–286, 2007. [44] J. Lehman, S. Risi, and J. Clune, “Creative generation of 3d objects with deep learning and innovation engines,” in Proceedings of the 7th International Conference on Computational Creativity, 2016. [45] M. Pigliucci, “Is evolvability evolvable?” Nature Reviews Genetics, vol. 9, no. 1, pp. 75–82, 2008. [46] L. Altenberg et al., “The evolution of evolvability in genetic programming,” Advances in genetic programming, vol. 3, pp. 47–74, 1994. [47] J.-B. Mouret and S. Doncieux, “Sferes v2: Evolvin’in the multi-core world,” in Evolutionary Computation (CEC), 2010 IEEE Congress on. IEEE, 2010, pp. 1–8. [48] J. P. Shaffer, “Multiple hypothesis testing,” Annual review of psychology, vol. 46, no. 1, pp. 561–584, 1995. [49] A. Alexandrescu, Modern C++ design: generic programming and design patterns applied. Addison-Wesley, 2001. 

THE RATIONAL HOMOLOGY OF THE OUTER AUTOMORPHISM GROUP OF F7 arXiv:1512.03075v2 [math.GR] 19 Jan 2016 LAURENT BARTHOLDI Abstract. We compute the homology groups H∗ (Out(F7 ); Q) of the outer automorphism group of the free group of rank 7. We produce in this manner the first rational homology classes of Out(Fn ) that are neither constant (∗ = 0) nor Morita classes (∗ = 2n − 4). 1. Introduction The homology groups Hk (Out(Fn ); Q) are intriguing objects. On the one hand, they are known to “stably vanish”, i.e. for all n ∈ N we have Hk (Out(Fn ); Q) = 0 as soon as k is large enough [3]. Hatcher and Vogtmann prove that the natural maps Hk Out(Fn ) → Hk Aut(Fn ) and Hk Aut(Fn ) → Hk Aut(Fn+1 ) are isomorphisms for n ≥ 2k + 2 respectively n ≥ 2k + 4, see [4, 5]. On the other hand, Hk (Out(Fn ); Q) = 0 for k > 2n − 3, since Out(Fn ) acts geometrically on a contractible space (the “spine of outer space”, see [2]) of dimension 2n − 3. Combining these results, the only k ≥ 1 for which Hk (Out(Fn ); Q) could possibly be non-zero are in the range n2 − 2 < k ≤ 2n − 3. Morita conjectures in [9, page 390] that H2n−3 (Out(Fn ); Q) always vanishes; this would improve the upper bound to k = 2n − 4, and H2n−4 (Out(Fn ); Q) is also conjectured to be non-trivial. We shall see that the first conjecture does not hold. Indeed, the first few values of Hk (Out(Fn ); Q) may be computed by a combination of human and computer work, and yield n\k 0 1 2 3 4 5 6 7 8 9 10 11 2 1 0 3 1 0 0 0 4 1 0 0 0 1 0 1 0 0 0 0 0 0 0 5 6 1 0 0 0 0 0 0 0 1 0 7 1 0 0 0 0 0 0 0 1 0 0 1 The values for n ≤ 6 were computed by Ohashi in [12]. They reveal that, for n ≤ 6, only the constant class (k = 0) and the Morita classes k = 2n − 4 yield non-trivial homology. The values for n = 7 are the object of this Note, and reveal that the picture changes radically: Theorem. The non-trivial homology groups Hk (Out(F7 ); Q) occur for k ∈ {0, 8, 11} and are all 1-dimensional. P Previously, only the rational Euler characteristic χQ (Out(F7 )) = (−1)k dim Hk (Out(F7 ); Q) was known [10], and shown to be 1. These authors computed in fact the rational Euler characteristics up to n = 11 in that paper and the sequel [11]. 2. Methods We make fundamental use of a construction of Kontsevich [6], explained in [1]. We follow the simplified description from [12]. Let Fn denote the free group of rank n. This parameter n is fixed once and for all, and will in fact be omitted from the notation as often as possible. An admissible graph of rank n is a Date: 18 January 2016. Partially supported by ANR grant ANR-14-ACHN-0018-01. 1 2 LAURENT BARTHOLDI graph G that is 2-connected (G remains connected even after an arbitrary edge is removed), without loops, with fundamental group isomorphic to Fn , and without vertex of valency ≤ 2. P Its degree is deg(G) := v∈V (G) (deg(v) − 3). In particular, G has 2n − 2 − deg(G) vertices and 3n − 3 − deg(G) edges, and is trivalent if and only if deg(G) = 0. If Φ is a collection of edges in a graph G, we denote by G/Φ the graph quotient, obtained by contracting all edges in Φ to points. A forested graph is a pair (G, Φ) with Φ an oriented forest in G, namely an ordered collection of edges that do not form any cycle. We note that the symmetric group Sym(k) acts on the set of forested graphs whose forest contains k edges, by permuting the forest’s edges. For k ∈ N, let Ck denote the Q-vector space spanned by isomorphism classes of forested graphs of rank n with a forest of size k, subject to the relation (G, πΦ) = (−1)π (G, Φ) for all π ∈ Sym(k). Note, in particular, that if (G, Φ) ∼ (G, πΦ) for an odd permutation π then (G, Φ) = 0 in Ck . These spaces (C∗ ) form a chain complex for the differential ∂ = ∂C − ∂R , defined respectively on (G, Φ) = (G, {e1 , . . . , ep }) by ∂C (G, Φ) = p X (−1)i (G/ei , Φ \ {ei }), ∂R (G, Φ) = i=1 p X (−1)i (G, Φ \ {ei }), i=1 and the homology of (C∗ , ∂) is H∗ (Out(Fn ); Q). The spaces Ck may be filtered by degree: let Fp Ck denote the subspace spanned by forested graphs (G, Φ) with deg(G/Φ) ≤ p. The differentials satisfy respectively ∂C (Fp Ck ) ⊆ Fp Ck−1 , ∂R (Fp Ck ) ⊆ Fp−1 Ck−1 . A spectral sequence argument gives (1) 2 Hp (Out(Fn ); Q) = Ep,0 = ker(∂C |Fp Cp ) ∩ ker(∂R |Fp Cp ) . ∂R (ker(∂C |Fp+1 Cp+1 )) Note that if (G, Φ) ∈ Fp Cp then G is trivalent. We compute explicitly bases for the vector spaces Fp Cp , and matrices for the differentials ∂C , ∂R , to prove the theorem. 3. Implementation We follow for n = 7 the procedure sketched in [12]. Using the software program nauty [8], we enumerate all trivalent graphs of rank n and vertex valencies ≥ 3. The libraries in nauty produce a canonical ordering of a graph, and compute generators for its automorphism group. We then weed out the non-2-connected ones. For given p ∈ N, we then enumerate all p-element oriented forests in these graphs, and weed out those that admit an odd symmetry. These are stored as a basis for Fp Cp . Let ap denote the dimension of Fp Cp . For (G, Φ) a basis vector in Fp Cp , the forested graphs that appear as summands in ∂C (G, Φ) and ∂R (G, Φ) are numbered and stored in a hash table as they occur, and the matrices ∂C and ∂R are computed as sparse matrices with ap columns. The nullspace ker(∂C |Fp Cp ) is then computed: let bp denote its dimension; then the nullspace is stored as a sparse (ap × bp )-matrix Np . The computation is greatly aided by the fact that ∂C is a block matrix, whose row and column blocks are spanned by {(G, Φ) : G/Φ = G0 } for all choices of the fully contracted graph G0 . The matrices Np are computed using the linear algebra library linbox [7], which provides exact linear algebra over Q and finite fields. Finally, the rank cp of ∂R ◦ Np is computed, again using linbox. By (1), we have dim Hp (Out(Fn ); Q) = bp − cp − cp+1 . THE RATIONAL HOMOLOGY OF THE OUTER AUTOMORPHISM GROUP OF F7 3 For memory reasons (the computational requirements reached 200GB of RAM at its peak), some of these ranks were computed modulo a large prime (65521 and 65519 were used in two independent runs). Computing modulo a prime can only reduce the rank; so that the values cp we obtained are underestimates of the actual ranks of ∂R ◦ Np . However, we also know a priori that bp − cp − cp+1 ≥ 0 since it is the dimension of a vector space; and none of the cp we computed can be increased without at the same time causing a homology dimension to become negative, so our reduction modulo a prime is legal. For information, the parameters ap , bp , cp for n = 7 are as follows: p ap bp cp 0 365 365 0 1 3712 1784 364 2 23227 5642 1420 3 ≈105k 14766 4222 4 ≈348k 28739 10544 5 ≈854k 39033 18195 6 ≈1.6m 38113 20838 7 ≈2.3m 28588 17275 8 ≈2.6m 16741 11313 9 ≈2.1m 6931 5427 10 ≈1.2m 1682 1504 The largest single matrix operations that had to be performed were computing the nullspace of a 2038511 × 536647 matrix (16 CPU hours) and the rank modulo 65519 of a (less sparse) 1355531 × 16741 matrix (10 CPU hours). The source files used for the computations are available as supplemental material. Compilation requires g++ version 4.7 or later, a functional linbox library, available from the site http://www.linalg.org, as well as the nauty program suite, available from the site http://pallini.di.uniroma1.it. It may also be directly downloaded and installed by typing ‘make nauty25r9’ in the directory in which the sources were downloaded. Beware that the calculations required for n = 7 are prohibitive for most desktop computers. Conclusion Computing the dimensions of the homology groups is only the first step in understanding them; much more interesting would be to know visually, or graph-theoretically, where these non-trivial classes come from. It seems almost hopeless to describe, via computer experiments, the non-trivial class in degree 8. It may be possible, however, to arrive at a reasonable understanding of the non-trivial class in degree 11. This class may be interpreted as a linear combination w of trivalent graphs on 12 vertices, each marked with an oriented spanning forest. There are 376365 such forested graphs that do not admit an odd symmetry. The class w ∈ Q376365 is an Z-linear combination of 70398 different forested graphs, with coefficients in {±1, . . . , ±16}. For example, eleven graphs occur with coefficient ±13; four of them have indices 25273, 53069, 53239, 53610 respectively, and are, with the spanning tree in bold, 8 11 3 4 0 2 5 0 11 5 9 8 10 1 6 1 2 3 7 6 10 9 4 7 11 ≈376k 179 178 4 LAURENT BARTHOLDI 5 8 7 4 7 8 11 1 5 2 9 0 3 4 10 6 0 11 3 1 6 9 2 10 The coefficients of w, and corresponding graphs, are distributed as ancillary material in the file w_cycle, in format ‘coefficient [edge1 edge2 ...]’, where each edge is ‘x-y’ or ‘x+y’ to indicate whether the edge is absent or present in the forest. Edges always satisfy x ≤ y, and the forest is oriented so that its edges are lexicographically ordered. Edges are numbered from 0 while graphs are numbered from 1. There are no multiple edges. Acknowledgments I am grateful to Alexander Berglund and Nathalie Wahl for having organized a wonderful and stimulating workshop on automorphisms of free groups in Copenhagen in October 2015, when this work began; to Masaaki Suzuki, Andy Putman and Karen Vogtmann for very helpful conversations that took place during this workshop; and to Jim Conant for having checked the cycle w (after finding a mistake in its original signs) with an independent program. References [1] James Conant and Karen Vogtmann, On a theorem of Kontsevich, Algebr. Geom. Topol. 3 (2003), 1167–1224, DOI 10.2140/agt.2003.3.1167. MR2026331 (2004m:18006) [2] Marc Culler and Karen Vogtmann, Moduli of graphs and automorphisms of free groups, Invent. Math. 84 (1986), no. 1, 91–119, DOI 10.1007/BF01388734. MR830040 (87f:20048) [3] Søren Galatius, Stable homology of automorphism groups of free groups, Ann. of Math. (2) 173 (2011), no. 2, 705–768, DOI 10.4007/annals.2011.173.2.3. MR2784914 (2012c:20149) [4] Allen Hatcher and Karen Vogtmann, Homology stability for outer automorphism groups of free groups, Algebr. Geom. Topol. 4 (2004), 1253–1272, DOI 10.2140/agt.2004.4.1253. MR2113904 (2005j:20038) [5] Allen Hatcher, Karen Vogtmann, and Nathalie Wahl, Erratum to: “Homology stability for outer automorphism groups of free groups [Algebr. Geom. Topol. 4 (2004), 1253–1272 (electronic); MR 2113904] by Hatcher and Vogtmann, Algebr. Geom. Topol. 6 (2006), 573–579 (electronic), DOI 10.2140/agt.2006.6.573. MR2220689 (2006k:20069) [6] Maxim Kontsevich, Formal (non)commutative symplectic geometry, The Gel′fand Mathematical Seminars, 1990–1992, Birkhäuser Boston, Boston, MA, 1993, pp. 173–187. MR1247289 (94i:58212) [7] LinBox — Exact Linear Algebra over the Integers and Finite Rings, Version 1.1.6, The LinBox Group, 2008. [8] Brendan D. McKay and Adolfo Piperno, Practical graph isomorphism, II, J. Symbolic Comput. 60 (2014), 94–112, DOI 10.1016/j.jsc.2013.09.003, available at arXiv:1301.1493. MR3131381 [9] Shigeyuki Morita, Structure of the mapping class groups of surfaces: a survey and a prospect, Proceedings of the Kirbyfest (Berkeley, CA, 1998), Geom. Topol. Monogr., vol. 2, Geom. Topol. Publ., Coventry, 1999, pp. 349–406 (electronic), DOI 10.2140/gtm.1999.2.349, (to appear in print). MR1734418 (2000j:57039) [10] Shigeyuki Morita, Takuya Sakasai, and Masaaki Suzuki, Computations in formal symplectic geometry and characteristic classes of moduli spaces, Quantum Topol. 6 (2015), no. 1, 139–182, DOI 10.4171/QT/61. MR3335007 , Integral Euler characteristic of Out F11 , Exp. Math. 24 (2015), no. 1, 93–97, DOI [11] 10.1080/10586458.2014.956373. MR3305042 [12] Ryo Ohashi, The rational homology group of Out(Fn ) for n ≤ 6, Experiment. Math. 17 (2008), no. 2, 167–179. MR2433883 (2009k:20118) École Normale Supérieure, Paris and Georg-August-Universität zu Göttingen E-mail address: laurent.bartholdi@gmail.com 

ABOUT VON NEUMANN’S PROBLEM FOR LOCALLY COMPACT GROUPS arXiv:1702.07955v1 [math.GR] 25 Feb 2017 FRIEDRICH MARTIN SCHNEIDER Abstract. We note a generalization of Whyte’s geometric solution to the von Neumann problem for locally compact groups in terms of Borel and clopen piecewise translations. This strengthens a result of Paterson on the existence of Borel paradoxical decompositions for non-amenable locally compact groups. Along the way, we study the connection between some geometric properties of coarse spaces and certain algebraic characteristics of their wobbling groups. 1. Introduction In his seminal article [18] von Neumann introduced the concept of amenability for groups in order to explain why the Banach-Tarski paradox occurs only for dimension greater than two. He proved that a group containing an isomorphic copy of the free group F2 on two generators is not amenable. The converse, i.e., the question whether every non-amenable group would have a subgroup being isomorphic to F2 , was first posed in print by Day [5], but became known as the von Neumann problem (or sometimes von Neumann-Day problem). The original question has been answered in the negative by Ol’šanskiı̆ [19]. However, there are very interesting positive solutions to variants of the von Neumann problem in different settings: a geometric solution by Whyte [27], a measure-theoretic solution by Gaboriau and Lyons [9] and its generalization to locally compact groups by Gheysens and Monod [14], as well as a Baire category solution by Marks and Unger [13]. Whyte’s geometric version reads as follows. Theorem 1.1 (Theorem 6.2 in [27]). A uniformly discrete metric space of uniformly bounded geometry is non-amenable if and only if it admits a partition whose pieces are uniformly Lipschitz embedded copies of the 4-regular tree. In particular, the above applies to Cayley graphs of finitely generated groups and in turn yields a geometric solution to the von Neumann problem. The aim of the present note is to extend Whyte’s relaxed version of the von Neumann conjecture to the realm of locally compact groups. For this purpose, we need to view the result from a slightly different perspective. Given a uniformly discrete metric space X, its wobbling group (or group of bounded displacement ) is defined as W (X) := {α ∈ Sym(X) | ∃r ≥ 0 ∀x ∈ X : d(x, α(x)) ≤ r}. Wobbling groups have attracted growing attention in recent years [11, 12, 3]. Since the 4-regular tree is isomorphic to the standard Cayley graph of F2 , one can easily Date: 28th February 2017. 2010 Mathematics Subject Classification. Primary 22D05, 43A07, 20E05, 20F65. This research has been supported by funding of the German Research Foundation (reference no. SCHN 1431/3-1) as well as by funding of the Excellence Initiative by the German Federal and State Governments. 1 2 FRIEDRICH MARTIN SCHNEIDER reformulate Whyte’s in terms of semi-regular subgroups. Let us recall that a subgroup G ≤ Sym(X) is said to be semi-regular if no non-identity element of G has a fixed point in X. Corollary 1.2 (Theorem 6.1 in [27]). A uniformly discrete metric space X of uniformly bounded geometry is non-amenable if and only if F2 is isomorphic to a semi-regular subgroup of W (X). For a finitely generated group G, the metrics generated by any two finite symmetric generating sets containing the neutral element are equivalent and hence give rise to the very same wobbling group W (G). It is easy to see that W (G) is just the group of piecewise translations of G, i.e., a bijection α : G → G belongs to W (G) if and only if exists a finite partition P of G such that ∀P ∈ P ∃g ∈ G : α|P = λg |P . Furthermore, we note that the semi-regularity requirement in the statement above cannot be dropped: in fact, van Douwen [6] showed that W (Z) contains an isomorphic copy of F2 , despite Z being amenable. As it turns out, F2 embeds into the wobbling group of any coarse space of positive asymptotic dimension (see Proposition 4.3 and Remark 4.4). We are going to present a natural counterpart of Corollary 1.2 for general locally compact groups. Let G be a locally compact group. We call a bijection α : G → G a clopen piecewise translation of G if there exists a finite partition P of G into clopen subsets such that ∀P ∈ P ∃g ∈ G : α|P = λg |P , i.e., on every member of P the map α agrees with a left translation of G. It is easy to see that the set C (G) of all clopen piecewise translations of G constitutes a subgroup of the homeomorphism group of the topological space G and that the mapping Λ : G → C (G), g 7→ λg embeds G into C (G) as a regular, i.e., semi-regular and transitive, subgroup. Similarly, a bijection α : G → G is called a Borel piecewise translation of G if there exists a finite partition P of G into Borel subsets with ∀P ∈ P ∃g ∈ G : α|P = λg |P . Likewise, the set B(G) of all Borel piecewise translations of G is a subgroup of the automorphism group of the Borel space of G and contains C (G) as a subgroup. For a locally compact group G, both B(G) and C (G) are reasonable analogues of the wobbling group. Yet, the mere existence of an embedding of F2 as a semiregular subgroup of B(G), or even C (G), does not prevent G from being amenable. In fact, there are many examples of compact (thus amenable) groups that admit F2 as a (non-discrete) subgroup and hence as a semi-regular subgroup of C (G). For example, since F2 is residually finite, it embeds into the compact group formed by the product of its finite quotients. Therefore, we have to seek for a topological analogue of semi-regularity, which amounts to a short discussion. Remark 1.3. Let X be a set. A subgroup G ≤ Sym(X) is semi-regular if and only if there exists a (necessarily surjective) map ψ : X → G such that ψ(gx) = gψ(x) for all g ∈ G and x ∈ X. Obviously, the latter implies the former. To see the converse, let σ : X → X be any orbit cross-section for the action of G on X, i.e., σ(X) ∩ Gx = {σ(x)} for every x ∈ X. Since G is semi-regular, for each x ∈ X there is a unique ψ(x) ∈ G such that ψ(x)σ(x) = x. For all g ∈ G and x ∈ X, we have gψ(x)σ(gx) = gψ(x)σ(x) = gx = ψ(gx)σ(gx), which readily implies that ψ(gx) = gψ(x). So, ψ : X → G is as desired. The purpose of this note is to show the following. ABOUT VON NEUMANN’S PROBLEM FOR LOCALLY COMPACT GROUPS 3 Theorem 1.4. Let G be a locally compact group. The following are equivalent. (1) G is not amenable. (2) There exist a homomorphism ϕ : F2 → C (G) and a Borel measurable map ψ : G → F2 such that ψ ◦ ϕ(g) = λg ◦ ψ for all g ∈ F2 . (3) There exist a homomorphism ϕ : F2 → B(G) and a Borel measurable map ψ : G → F2 such that ψ ◦ ϕ(g) = λg ◦ ψ for all g ∈ F2 . We remark that any map ϕ as in (2) or (3) of Theorem 1.4 has to be injective. In view of the discussion above, we also note that for finitely generated discrete groups the statement of Theorem 1.4 reduces to Whyte’s geometric solution to the von Neumann problem. More specifically, the existence of a map ψ as in (2) or (3) above may be thought of as a Borel variant of the semi-regular embedding condition in Corollary 1.2. In general, we cannot arrange for ψ to be continuous, as there exist non-amenable connected locally compact groups Both (2) and (3) of Theorem 1.4 may be considered relaxed versions of containing F2 as a discrete subgroup: according to a result of Feldman and Greenleaf [7], if H is a σ-compact metrizable closed (e.g., countable discrete) subgroup of a locally compact group G, then the right coset projection G → H \ G, x 7→ Hx admits a Borel measurable cross-section τ : H \G → G, and hence the H-equivariant map ψ : G → H, x 7→ xτ (Hx)−1 is Borel measurable, too. This particularly applies if H ∼ = F2 is discrete. The proof of Theorem 1.4 combines a result of Rickert resolving the original von Neumann problem for almost connected locally compact groups (Theorem 3.3) with a slight generalization of Whyte’s result for coarse spaces (Theorem 2.2) and in turn refines an argument of Paterson proving the existence of Borel paradoxical decompositions for non-amenable locally compact groups [21]. In fact, Theorem 1.4 implies Paterson’s result [21]. Corollary 1.5 (Paterson [21]). A locally compact group G is non-amenable if and only if it admits a Borel paradoxical decomposition, i.e., there exist finite partitions P and Q of G into Borel subsets and gP , hQ ∈ G (P ∈ P, Q ∈ Q) such that [ [ G = · gP P ∪· · hQ Q. P ∈P Q∈Q This note is organized as follows. Building on some preparatory work concerning coarse spaces done in Section 2, we prove Theorem 1.4 in Section 3. Since our approach to proving Theorem 1.4 involves wobbling groups, and there has been recent interest in such groups, we furthermore include some complementary remarks about finitely generated subgroups of wobbling groups in Section 4. 2. Revisiting Whyte’s result Our proof of Theorem 1.4 will make use of Whyte’s argument [27] – in the form of Corollary 2.3. More precisely, we will have to slightly generalize his result from metric spaces to arbitrary coarse spaces. However, this will just require very minor adjustments, and we only include a proof for the sake of completeness. For convenience, let us recall some terminology from coarse geometry as it may be found in [24]. For a relation E ⊆ X × X on a set X and x ∈ X, A ⊆ X, let [ E[x] := {y ∈ X | (x, y) ∈ E}, E[A] := {E[z] | z ∈ A}. A coarse space is a pair (X, E ) consisting of a set X and a collection E of subsets of X × X (called entourages) such that • the diagonal ∆X = {(x, x) | x ∈ X} belongs to E , 4 FRIEDRICH MARTIN SCHNEIDER • if F ⊆ E ∈ E , then also F ∈ E , • if E, F ∈ E , then E ∪ F, E −1 , E ◦ F ∈ E . A coarse space (X, E ) is said to have bounded geometry if ∀E ∈ E ∀x ∈ X : E[x] is finite, and (X, E ) has uniformly bounded geometry if ∀E ∈ E ∃m ≥ 0 ∀x ∈ X : |E[x]| ≤ m. Among the most important examples of coarse spaces are metric spaces: if X is a metric space, then we obtain a coarse space (X, EX ) by setting EX := {E ⊆ X × X | sup{d(x, y) | (x, y) ∈ E} < ∞}. Another crucial source of examples of coarse spaces is given by group actions. Indeed, if G is a group acting on a set X, then we obtain a coarse space (X, EG ) of uniformly bounded geometry by EG := {R ⊆ X × X | ∃E ⊆ G finite : R ⊆ {(x, gx) | x ∈ X, g ∈ E}}. Note that the coarse structure induced by a finitely generated group G acting on itself by left translations coincides with the coarse structure on G generated by the metric associated with any finite symmetric generated subset of G containing the neutral element. Now we come to amenability. Adopting the notion from metric coarse geometry, we call a coarse space (X, E ) of bounded geometry amenable if ∀θ > 1 ∀E ∈ E ∃F ⊆ X finite, F 6= ∅ : |E[F ]| ≤ θ|F |, which is (easily seen to be) equivalent to saying that ∃θ > 1 ∀E ∈ E ∃F ⊆ X finite, F 6= ∅ : |E[F ]| ≤ θ|F |. This definition is compatible with the existing notion of amenability for group actions (Proposition 2.1). Recall that an action of a group G on a set X is amenable if the space ℓ∞ (X) of all bounded real-valued functions on X admits a G-invariant mean, i.e., there exists a positive linear functional µ : ℓ∞ (X) → R with µ(1) = 1 and µ(f ◦ g) = µ(f ) for all f ∈ ℓ∞ (X) and g ∈ G. Proposition 2.1 (cf. Rosenblatt [25]). An action of a group G on a set X is amenable if and only if the coarse space (X, EG ) is amenable. Proof. Generalizing Følner’s work [8] on amenable groups, Rosenblatt [25] showed that an action of a group G on a set X is amenable if and only if ∀θ > 1 ∀E ⊆ G finite ∃F ⊆ X finite, F 6= ∅ : |EF | ≤ θ|F |, which is easily seen to be equivalent to the amenability of (X, EG ).  Let us turn our attention towards Theorem 1.1. A straightforward adaptation of Whyte’s original argument readily provides us with the following only very slight generalization (Theorem 2.2). For a binary relation E ⊆ X × X, we will denote the associated undirected graph by Γ(E) := (X, {{x, y} | (x, y) ∈ E}). Furthermore, let gr(f ) := {(x, f (x)) | x ∈ X} for any map f : X → Y . Our proof of Theorem 2.2 will utilize the simple observation that, for a map f : X → X, the graph Γ(gr(f )) is a forest, i.e., it contains no cycles, if and only if f has no periodic points, which means that P (f ) := {x ∈ X | ∃n ≥ 1 : f n (x) = x} is empty. Theorem 2.2. Let d ≥ 3. A coarse space (X, E ) of bounded geometry is nonamenable if and only if there is E ∈ E such that Γ(E) is a d-regular forest. ABOUT VON NEUMANN’S PROBLEM FOR LOCALLY COMPACT GROUPS 5 Proof. (⇐=) Due to a very standard fact about isoperimetric constants for regular trees [2, Example 47], if E ⊆ X × X is symmetric and Γ(E) is a d-regular tree, then |E[F ]| ≥ (d − 1)|F | for every finite subset F ⊆ X. Of course, this property passes to d-regular forests, which readily settles the desired implication. (=⇒) Suppose that (X, E ) is not amenable. Then there is a symmetric entourage E ∈ E such that |E[F ]| ≥ d|F | for every finite F ⊆ X. Consider the symmetric relation R := E \ ∆X ⊆ X × X. Since |R[x]| < ∞ for every x ∈ X and |R[F ]| ≥ |E[F ] \ F | ≥ |E[F ]| − |F | ≥ (d − 1)|F | for every finite subset F ⊆ X, the Hall harem theorem [1, Theorem H.4.2] asserts that there exists a function f : X → X with gr(f ) ⊆ R and |f −1 (x)| = d − 1 for all x ∈ X. Notice that f does not have any fixed points as R ∩ ∆X = ∅. Since the set of f -orbits of its elements S partitions the set P (f ), we may choose a subset P0 ⊆ P (f ) such that P (f ) = · x∈P0 {f n (x) | n ∈ N}. Furthermore, choose functions g, h : P0 × N → X such that, for all x ∈ P0 and n ≥ 1, • g(x, 0) = x and h(x, 0) = f (x), • {g(x, n), h(x, n)} ∩ P (f ) = ∅, • f (g(x, n)) = g(x, n − 1) and f (h(x, n)) = h(x, n − 1). It follows that g and h are injective functions with disjoint ranges. Now we define f∗ : X → X by setting   g(z, n + 2) if x = g(z, n) for z ∈ P0 and even n ≥ 0,  g(z, n − 2) if x = g(z, n) for z ∈ P and odd n ≥ 3, 0 f∗ (x) := f 2 (x) if x = h(z, n) for z ∈ P 0 and n ≥ 2,    f (x) otherwise for x ∈ X. We observe that gr(f∗ ) ⊆ gr(f 2 )−1 ∪ gr(f 2 ) ∪ gr(f ). In particular, gr(f∗ ) ⊆ E ◦ E and therefore gr(f∗ ) ∈ E . Moreover, it follows that P (f∗ ) ⊆ P (f ). However, for every x ∈ P (f ), there exists a smallest m ∈ N such that f m (x) ∈ P0 , and we conclude that f∗m+1 (x) = f∗ (f m (x)) = g(f m (x), 2) ∈ / P (f ) and hence f∗m+1 (x) ∈ / P (f∗ ), which readily implies that x ∈ / P (f∗ ). Thus, P (f∗ ) = ∅. In particular, f∗ has no fixed points. Furthermore,   (f −1 (x) ∪ {g(z, n − 2)}) \ {g(z, n + 1)} if x = g(z, n) for z ∈ P0     and even n ≥ 2,     (f −1 (x) ∪ {g(z, n + 2)}) \ {g(z, n + 1)} if x = g(z, n) for z ∈ P0     and odd n ≥ 1, −1 f∗ (x) = −1  (f (x) ∪ {h(z, n + 2)}) \ {h(z, n + 1)} if x = h(z, n) for z ∈ P0      and n ≥ 1,    −1  (f (x) ∪ {h(z, 2)}) \ {z} if x = f (z) for z ∈ P0 ,     −1 f (x) otherwise and thus |f∗−1 (x)| = d−1 for each x ∈ X. Hence, Γ(gr(f∗ )) is a d-regular forest.  Just as Theorem 1.1 corresponds to Corollary 1.2, we can translate Theorem 2.2 into an equivalent statement about wobbling groups. Given a coarse space (X, E ), we define its wobbling group (or group of bounded displacement ) as W (X, E ) := {α ∈ Sym(X) | gr(α) ∈ E }. Since the 4-regular tree is isomorphic to the standard Cayley graph of the free group on two generators, we now obtain the following consequence of Theorem 2.2. 6 FRIEDRICH MARTIN SCHNEIDER Corollary 2.3. A coarse space X of bounded geometry is non-amenable if and only if F2 is isomorphic to a semi-regular subgroup of W (X). We note that Corollary 2.3 for group actions has been applied already (though without proof) in the recent work of the author and Thom [26, Corollary 5.12], where a topological version of Whyte’s result for general (i.e., not necessarily locally compact) topological groups in terms of perturbed translations is established. In the present note, Corollary 2.3 will be used to prove Theorem 1.4, which generalizes Whyte’s result to locally compact groups by means of clopen and Borel piecewise translations and is in turn quite different to [26, Corollary 5.12]. 3. Proving the main result In this section we prove Theorem 1.4. For the sake of clarity, recall that a locally compact group G is said to be amenable if there is a G-invariant1 mean on the space Cb (G) of bounded continuous real-valued functions on G, i.e., a positive linear map µ : Cb (G) → R with µ(1) = 1 and µ(f ◦ λg ) = µ(f ) for all f ∈ Cb (G) and g ∈ G. In preparation of the proof of Theorem 1.4, we note the following standard fact, whose straightforward proof we omit. Lemma 3.1. Let H be a subgroup of a locally compact group G and consider the usual action of G on the set G/H of left cosets of H in G. If µ : ℓ∞ (G/H) → R is a G-invariant mean and ν : Cb (H) → R is an H-invariant mean, then a G-invariant mean ξ : Cb (G) → R is given by ξ(f ) := µ(xH 7→ ν((f ◦ λx )|H )) (f ∈ Cb (G)). It is a well-known fact (see Section 2 in [10]) that a locally compact group G (considered together with a left Haar measure) is amenable if and only if there exists a G-invariant mean on L∞ (G), i.e., a positive linear map µ : L∞ (G) → R such that µ(1) = 1 and µ(f ◦ λg ) = µ(f ) for all f ∈ Cb (G) and g ∈ G. An easy calculation now provides us with the following. Lemma 3.2. Let G be a locally compact group. (1) A mean µ : L∞ (G) → R is G-invariant if and only if µ is B(G)-invariant. (2) Let H be a locally compact group, let ϕ : H → B(G) be a homomorphism and ψ : G → H be Borel measurable with ψ ◦ ϕ(g) = λg ◦ ψ for all g ∈ H. If G is amenable, then so is H. Proof. (1) Clearly, B(G)-invariance implies G-invariance. To prove the converse, suppose that µ is G-invariant. Let α ∈ B(G) and let P be a finite partition of G into Borel subsets and gP ∈ G (P ∈ P) with α|P = λgP |P for each P ∈ P. Now, X X µ(f ◦ α) = µ ((f ◦ α) · 1P ) = µ ((f ◦ λgP ) · 1P ) P ∈P = X P ∈P = X P ∈P P ∈P    X = µ (f · (1gP P )) µ f · 1P ◦ λg−1 P P ∈P
µ f · 1α(P ) = µ(f ) for every f ∈ L∞ (G), as desired. (2) Let ν : L∞ (G) → R be a G-invariant mean. Define µ : Cb (H) → R by µ(f ) := ν(f ◦ ψ) (f ∈ Cb (H)). It is easy to see that µ is a mean. Furthermore, (1) asserts that µ(f ◦ λg ) = ν(f ◦ λg ◦ ψ) = ν(f ◦ ψ ◦ ϕ(g)) = ν(f ◦ ψ) = µ(f ) 1In case of ambiguity, invariance shall always mean left invariance. ABOUT VON NEUMANN’S PROBLEM FOR LOCALLY COMPACT GROUPS for all f ∈ Cb (H) and g ∈ H. Hence, µ is H-invariant. 7  We note that Lemma 3.2 readily settles the implication (3)=⇒(1) of Theorem 1.4. The remaining part of the proof of Theorem 1.4 will rely on some structure theory for locally compact groups – most importantly the following remarkable result of Rickert [22] building on [23]. We recall that a locally compact group G is said to be almost connected if the quotient of G by the connected component of its neutral element is compact. Theorem 3.3 (Theorem 5.5 in [22]). Any almost connected, non-amenable, locally compact group has a discrete subgroup being isomorphic to F2 . Now everything is prepared to prove our main result. Proof of Theorem 1.4. Evidently, (2) implies (3) as C (G) is a subgroup of B(G). Furthermore, (3) implies (1) due to Lemma 3.2 and the non-amenability of F2 . (1)=⇒(2). Let G be a non-amenable locally compact group. It follows by classical work of van Dantzig [4] that any locally compact group contains an almost connected, open subgroup (see, e.g., [20, Proposition 12.2.2 (c)]). Choose any almost connected, open (and hence closed) subgroup H of G. We will distinguish two cases depending upon whether H is amenable. H is not amenable. According to Theorem 3.3, H contains a discrete subgroup F being isomorphic to F2 , and so does G. By a result of Feldman and Greenleaf [7], the right coset projection π : G → F \ G, x 7→ F x admits a Borel measurable cross-section, i.e., there exists a Borel measurable map τ : F \ G → G such that π ◦ τ = idF \G . Clearly, the F -equivariant map ψ : G → F, x 7→ xτ (F x)−1 is Borel measurable. This readily settles the first case: the maps ϕ : F2 ∼ = F → C (G), g 7→ λg and ψ are as desired. H is amenable. Since G is not amenable, Lemma 3.1 implies that the action of G on the set G/H is not amenable. By Proposition 2.1, this means that the coarse space X := (G/H, EG ) is not amenable. Due to Corollary 2.3, there exists an embedding ϕ : F2 = F (a, b) → W (X) such that ϕ(F2 ) is semi-regular. Thus, by definition of W (X), there exists some finite subset E ⊆ G such that ∀x ∈ {a, b} ∀z ∈ X ∃g ∈ E : ϕ(x)(z) = gz. Hence, we find a finite partition P of X along with gP , hP ∈ E (P ∈ P) such that ϕ(a)|P = λgP |P and ϕ(b)|P = λhP |P for every P ∈ P. Consider the projection π : G → G/H, x 7→ xH. Since H is an open subgroup of G, the quotient topology on G/H, i.e., the topology induced by π, is discrete. So, π −1 (P) = {π −1 (P ) | P ∈ P} is a finite partition of G into clopen subsets. What is more, [ [ [ G = · π −1 (ϕ(a)(P )) = · π −1 (gP P ) = · gP π −1 (P ), P ∈P P ∈P P ∈P [ [ [ G = · π −1 (ϕ(b)(P )) = · π −1 (hP P ) = · hP π −1 (P ). P ∈P P ∈P P ∈P Therefore, we may define ϕ : {a, b} → C (G) by setting ϕ(a)|π−1 (P ) = λgP |π−1 (P ) , ϕ(b)|π−1 (P ) = λhP |π−1 (P ) ∗ (P ∈ P). ∗ Consider the unique homomorphism ϕ : F2 → C (G) satisfying ϕ |{a,b} = ϕ. Since π ◦ ϕ(x) = ϕ(x) ◦ π for each x ∈ {a, b}, it follows that π ◦ ϕ∗ (w) = ϕ(w) ◦ π for every w ∈ F2 . Appealing to Remark 1.3, we find a mapping ψ : G/H → F2 such that ψ(ϕ(w)(z)) = wψ(z) for all w ∈ F2 and z ∈ G/H. Since the quotient space 8 FRIEDRICH MARTIN SCHNEIDER G/H is discrete, the map ψ ∗ := ψ ◦ π : G → F2 is continuous and therefore Borel measurable. Finally, we note that ψ ∗ (ϕ∗ (w)(x)) = ψ(π(ϕ∗ (w)(x))) = ψ(ϕ(w)(π(x))) = wψ(π(x)) = wψ ∗ (x) for all w ∈ F2 and x ∈ G, as desired. This completes the proof.  Let us deduce Paterson’s result [21] from Theorem 1.4. Proof of Corollary 1.5. (⇐=) This is clear. (=⇒) Let G be a non-amenable locally compact group. By Theorem 1.4, there exist a homomorphism ϕ : F2 → B(G) and a Borel measurable map ψ : G → F2 with ψ ◦ ϕ(g) = λg ◦ ψ for all g ∈ F2 . Consider any paradoxical decomposition of F2 given by P, Q, (gP )P ∈P , (hQ )Q∈Q . Taking a common refinement of suitable finite Borel partitions of G corresponding to the elements ϕ(gP ), ϕ(hQ ) ∈ B(G) (P ∈ P, Q ∈ Q), we obtain a finite Borel partition R of G along with mappings σ : P × R → G and τ : Q × R → G such that ϕ(gP )|R = λσ(P,R) |R ϕ(hQ )|R = λτ (Q,R) |R for all P ∈ P, Q ∈ Q, and R ∈ R. By ψ being Borel measurable, the refinements ψ −1 (P) ∨ R and ψ −1 (Q) ∨ R are finite Borel partitions of G. What is more, [ [ G = · ψ −1 (gP P ) ∪· · ψ −1 (hQ Q) P ∈P Q∈Q [ [ = · ϕ(gP )(ψ −1 (P )) ∪· · ϕ(hQ )(ψ −1 (Q)) P ∈P = Q∈Q [ · ϕ(gP )(ψ −1 (P ) ∩ R) ∪· (P,R)∈P×R = [ · [ · ϕ(hQ )(ψ −1 (Q) ∩ R) (Q,R)∈Q×R σ(P, R)(ψ −1 (P,R)∈P×R (P ) ∩ R) ∪· [ · τ (Q, R)(ψ −1 (Q) ∩ R). (Q,R)∈Q×R Thus, the data ψ −1 (P) ∨ R, ψ −1 (Q) ∨ R, (σ(P, R))(P,R)∈P×R , (τ (Q, R))(Q,R)∈Q×R constitute a Borel paradoxical decomposition of G.  4. Further remarks on wobbling groups We are going to conclude with some additional remarks about wobbling groups, which we consider noteworthy complements of Corollary 2.3. As van Douwen’s result [6] shows, the presence of F2 as a subgroup of the wobbling group does not imply the non-amenability of a coarse space. As it turns out, containment of F2 is just a witness for positive asymptotic dimension (Proposition 4.3). Let us once again recall some terminology from [24]. The asymptotic dimension asdim(X, E ) of a coarse space (X, E ) is defined as the infimum of all those n ∈ N such that, for every E ∈ E , there exist C0 , . . . , Cn ⊆ P(X) with S S • X = C0 ∪ . . . ∪ Cn , • (C S × D) ∩ E = ∅ for all i ∈ {0, . . . , n} and C, D ∈ Ci with C 6= D, • {C × C | C ∈ Ci , i ∈ {0, . . . , n}} ∈ E . The concept of asymptotic dimension was first introduced for metric spaces by Gromov [15] and later extended to coarse spaces by Roe [24]. We refer to [24] for a thorough discussion of asymptotic dimension, related results and examples. As we aim to describe positive asymptotic dimension in algebraic terms, we will unravel the zero-dimensional case in the following lemma. Let us denote by [R] the equivalence relation on a set X generated by a given binary relation R ⊆ X × X. ABOUT VON NEUMANN’S PROBLEM FOR LOCALLY COMPACT GROUPS 9 Lemma 4.1. Let (X, E ) be a coarse space. Then asdim(X, E ) = 0 if and only if [E] ∈ E for every E ∈ E . Proof. (=⇒) Let E ∈ E . Without loss of generality, assume that E contains ∆X . As asdim(X, E ) = 0, there exists C0 ⊆ P(X) such that S (1) X = C0 , (2) S (C × D) ∩ E = ∅ for all C, D ∈ C0 with C 6= D, (3) {C × C | C ∈ C0 } ∈ E . As ∆X ⊆ E, assertion (2) implies that any two distinct members of C0 are disjoint. Hence, S(1) gives that C0 is a partition of X. By (2), the induced equivalence relation R := {C × C | C ∈ C0 } contains E, thus [E]. By (3), it follows that [E] ∈ E . (⇐=) Let E ∈ E . It is straightforward to check that C0 := {[E][x] | x ∈ X} has the desired properties. Hence, asdim(X, E ) = 0.  Our proof of Proposition 4.3 below will rely upon the following slight modification of the standard argument for residual finiteness of free groups. For an element w ∈ F2 = F (a, b), let us denote by |w| the length of w with respect to the generators a and b, i.e., the smallest integer n ≥ 0 such that w can be represented as a word of length n in the letters a, a−1 , b, b−1 . Lemma 4.2. Let w ∈ F2 with w 6= e and let M := {0, . . . , 2|w|}. Then there exists a homomorphism ϕ : F2 → Sym(M ) such that ϕ(w) 6= e and |ϕ(v)(i) − i| ≤ 2|v| for all i ∈ M and v ∈ F2 . Proof. Let (k0 , . . . , kn ) ∈ (Z \ {0})n × ZP and (ℓ0 , . . .P , ℓn ) ∈ Z × (Z \ {0})n such that n k0 ℓ0 kn ℓn w = a b · · · a b . Of course, |w| = i=0 |ki | + ni=0 |ℓi |. Let Xi−1 Xi−1 Xi Xi−1 |ℓj | |kj | + |ℓj |, βi := |kj | + αi := j=0 j=0 j=0 j=0 for i ∈ {0, . . . , n} and let βn+1 := |w|. We will define a map ϕ : {a, b} → Sym(M ). First, let us define ϕ(a) ∈ Sym(M ) by case analysis as follows: if i ∈ [2αj , 2βj+1 ] for some j ∈ {0, . . . , n} with kj > 0, then  i + 2 if i is even and i ∈ [2αj , 2βj+1 − 2],    i − 1 if i = 2β , j+1 ϕ(a)(i) :=  i − 2 if i is odd and i ∈ [2αj + 3, 2βj+1 − 1],    i − 1 if i = 2αj + 1, if i ∈ [2αj , 2βj+1 ] for some j ∈ {0, . . . , n} with kj < 0, then  i − 2 if i is even and i ∈ [2αj + 2, 2βj+1 ],    i + 1 if i = 2α , j ϕ(a)(i) :=  i + 2 if i is odd and i ∈ [2αj + 1, 2βj+1 − 3],    i + 1 if i = 2βj+1 − 1, S and if i ∈ / {[2αj , 2βj+1 ] | j ∈ {0, . . . , n}, kj 6= 0}, then ϕ(a)(i) := i. Analogously, let us define ϕ(b) ∈ Sym(M ) by case analysis as follows: if i ∈ [2βj , 2αj ] for some j ∈ {0, . . . , n} with ℓj > 0, then  i + 2 if i is even and i ∈ [2βj , 2αj − 2],    i − 1 if i = 2α , j ϕ(b)(i) :=  i − 2 if i is odd and i ∈ [2βj + 3, 2αj − 1],    i − 1 if i = 2βj + 1, 10 FRIEDRICH MARTIN SCHNEIDER if i ∈ [2βj , 2αj ] for some j ∈ {0, . . . , n} with ℓj < 0, then  i − 2 if i is even and i ∈ [2βj + 2, 2αj ],    i + 1 if i = 2β , j ϕ(b)(i) :=  i + 2 if i is odd and i ∈ [2βj + 1, 2αj − 3],    i + 1 if i = 2αj − 1, S and if i ∈ / {[2βj , 2αj ] | j ∈ {0, . . . , n}, ℓj 6= 0}, then ϕ(b)(i) := i. It is easy to check that ϕ(a) and ϕ(b) are well-defined permutations of M , and that moreover |ϕ(x)(i) − i| ≤ 2 for each x ∈ {a, b} and all i ∈ M . Considering the unique homomorphism ϕ∗ : F2 → Sym(M ) with ϕ∗ |{a,b} = ϕ, we observe that
ϕ∗ (w)(0) = ϕ(a)kn ϕ(b)ℓn · · · ϕ(a)k0 ϕ(b)ℓ0 (0) = 2|w| and thus ϕ∗ (w) 6= e. Also, |ϕ∗ (v)(i) − i| ≤ 2|v| for all i ∈ M and v ∈ F2 .  For the sake of clarity, we recall that a group is locally finite if each of its finitely generated subgroups is finite. For a subset S of a group G, we will denote by hSi the subgroup of G generated by S. Proposition 4.3. Let X be a coarse space of uniformly bounded geometry. The following are equivalent. (1) asdim(X) > 0. (2) W (X) is not locally finite. (3) F2 embeds into W (X). Proof. We will denote by E the coarse structure of X. (2)=⇒(1). Let us recall a general fact: for a finite group G and any set M , the group GM is locally finite. Indeed, considering a finite subset S ⊆ GM and the induced equivalence relation R := {(x, y) ∈ M × M | ∀α ∈ S : α(x) = α(y)} on M , we observe that N := {R[x] | x ∈ M } is finite, due to G and S being finite. The map π : M → N, x 7→ R[x] induces a homomorphism ϕ : GN → GM , α 7→ α ◦ π. Evidently, S is contained in the finite group ϕ(GN ), and so is hSi. Suppose now that asdim(X) = 0. Consider a finite subset S ⊆ W (X). We aim to show that H := hSi is finite. To this end, we first observe that Y ϕ: H → Sym(Hx), α 7→ (α|Hx )x∈X x∈X S constitutes a well-defined embedding. Since D := {gr(α) | α ∈ S} belongs to E , Lemma 4.1 asserts that E := [D] ∈ E , too. Note that gr(α) ∈ E for all α ∈ H. Hence, Hx ⊆ E[x] for every x ∈ X. Due to X having uniformly bounded geometry, there exists m ≥ 0 such that |E[x]| ≤ m and thus |Hx| Q ≤ m for every x ∈ X. Now, let M := {0, . . . , m − 1}. It follows that the group x∈X Sym(Hx) is isomorphic to a subgroup of Sym(M )X , and so is H by virtue of ϕ. Since H is finitely generated and Sym(M )X is locally finite by the remark above, this implies that H is finite. (3)=⇒(2). This is trivial. (1)=⇒(3). Suppose that asdim(X) > 0. By Lemma 4.1, there exists E ∈ E such that [E] ∈ / E .SWithout loss of generality, we may assume that ∆X ⊆ E = E −1 . Hence, [E] = {E n | n ∈ N}. For each n ∈ N, let us define  Tn := x ∈ X n+1 |{x0 , . . . , xn }| = n + 1, ∀i ∈ {0, . . . , n − 1} : (xi , xi+1 ) ∈ E . Claim. For every n ∈ N and every finite subset F ⊆ X, there exists x ∈ Tn such that {x0 , . . . , xn } ∩ F = ∅. Proof of claim. Let n ∈ N and let F ⊆ X be finite. Put ℓ := (n + 1)(|F | + 1). Since E ∈ E and [E] ∈ / E , we conclude that E ℓ * E ℓ−1 . Let x0 , . . . , xℓ ∈ X such ℓ−1 that (x0 , xℓ ) ∈ /E and (xi , xi+1 ) ∈ E for every i ∈ {0, . . . , ℓ − 1}. As ∆X ⊆ E, it ABOUT VON NEUMANN’S PROBLEM FOR LOCALLY COMPACT GROUPS 11 follows that |{x0 , . . . , xℓ }| = ℓ + 1. Applying the pigeonhole principle, we find some j ∈ {0, . . . , ℓ−n} such that {xj , . . . , xj+n }∩F = ∅. Hence, y0 := xj , . . . , yn := xj+n are as desired.  Since N := F2 \ {e} is countable, we may recursively apply the claim above and choose a family (xw )w∈N such that (i) xw ∈ T2|w| for every w ∈ N , (ii) {xw,0 , . . . , xw,2|w| } ∩ {xv,0 , . . . , xv,2|v| } = ∅ for any two distinct v, w ∈ N . Let w ∈ N and define Dw := {xw,0 , . . . , xw,2|w| }. Due to Lemma 4.2, there exists a homomorphism ϕw : F2 → Sym(Dw ) such that ϕw (w) 6= e and ϕw (v)(xw,i ) ∈ {xw,j | j ∈ {0, . . . , 2|w|}, |i − j| ≤ 2|v|} for all v ∈ F2 , i ∈ {0, . . . , 2|w|}. Since (xw,i , xw,i+1 ) ∈ E for i ∈ {0, . . . , 2|w| − 1}, it follows that gr(ϕw (v)) ⊆ E 2|v| for all v ∈ F2 . As Dw and Dv are disjoint for any distinct v, w ∈ N , we may define a homomorphism ϕ : F2 → Sym(X) by setting ( ϕw (v)(x) if x ∈ Dw for some w ∈ N, ϕ(v)(x) := x otherwise for v ∈ F2 and x ∈ X. By construction, ϕ is an embedding, and furthermore [ gr(ϕ(v)) ⊆ ∆X ∪ {gr(ϕw (v)) | w ∈ N } ⊆ E 2|v| ∈ E for every v ∈ F2 . Hence, the image of ϕ is contained in W (X), as desired.  Remark 4.4. The assumption of uniformly bounded geometry in Theorem 1.4 is needed only to prove that (2) implies (1). In fact, a similar argument as in the proof of (1)=⇒(3) (not involving Lemma 4.2 though) shows that the wobbling group of any coarse Q space not having uniformly bounded geometry contains an isomorphic copy of n∈N Sym(n), hence F2 . One might wonder whether Proposition 4.3 could have been deduced readily from van Douwen’s result [6] on F2 embedding into W (Z). However, there exist uniformly discrete metric spaces of uniformly bounded geometry and positive asymptotic dimension whose wobbling group does not contain an isomorphic copy of W (Z) (see Example 4.7). We clarify the situation in Proposition 4.5. As usual, a group is called residually finite if it embeds into a product of finite groups, and a group is called locally residually finite if each of its finitely generated subgroups is residually finite. Let us recall from [24] that a map f : X → Y between two coarse spaces X and Y is bornologous if, for every entourage E of X, the set {(f (x), f (y)) | (x, y) ∈ E} is an entourage of Y . Proposition 4.5. Let X be a coarse space. The following are equivalent. (1) There is a bornologous injection from Z into X. (2) W (X) is not locally residually finite. (3) W (X) contains a subgroup being isomorphic to W (Z). Remark 4.6. (i) For groups there is no difference between positive asymptotic dimension and the existence of a bornologous injection of Z: a group has asymptotic dimension 0 if and only if it is locally finite, and any group which is not locally finite admits a bornologous injection of Z by a standard compactness argument (see, e.g., [17, IV.A.12]). However, for arbitrary coarse spaces, even of uniformly bounded geometry, the situation is slightly different (see Example 4.7). (ii) One may equivalently replace Z by N in item (1) of Proposition 4.5: on the one hand, the inclusion map constitutes a bornologous injection from N into Z; on 12 FRIEDRICH MARTIN SCHNEIDER the other hand, there is a bornologous bijection f : Z → N given by ( 2n if n ≥ 0, f (n) := (n ∈ Z). 2|n| − 1 if n < 0 Unless explicitly stated otherwise, we always understand N as being equipped with the coarse structure generated by the usual (i.e., Euclidean) metric. (iii) Any bornologous injection f : X → Y between two coarse spaces X and Y induces an embedding ϕ : W (X) → W (Y ) via ( f (α(f −1 (y))) if y ∈ f (X), ϕ(α)(y) := (α ∈ W (X), y ∈ Y ). y otherwise Hence, by (ii), the groups W (N) and W (Z) mutually embed into each other, and thus Z may equivalently be replaced by N in item (3) of Proposition 4.5. Proof of Proposition 4.5. (1)=⇒(3). This is due to Remark 4.6(iii). (3)=⇒(2). It suffices to show that W (Z) is not locally residually finite. A result of Gruenberg [16] states that, for a finite group F , the restricted wreath product F ≀ Z = F (Z) ⋊ Z (i.e., the lamplighter group over F ) is residually finite if and only Sn−1 if F is abelian. For n ≥ 1, the action of Sym(n) ≀ Z on Z = · r=0 nZ + r given by   (α, m).(nk + r) := n(m + k) + αm+k (r) α ∈ Sym(n)(Z) , m, k ∈ Z, 0 ≤ r < n defines an embedding of Sym(n) ≀ Z into Sym(Z), the image of which is contained in W (Z) as supz∈Z |z − (α, m).z| ≤ n(|m| + 1) for every (α, m) ∈ Sym(n) ≀ Z. Since the embedded lamplighter groups are finitely generated and not residually finite for n ≥ 3, it follows that W (Z) is not locally residually finite. (2)=⇒(1). Let E denote the coarse structure of X. If X does not have bounded geometry, then there exist E ∈ E and x ∈ X such that E[x] is infinite, and any thus existing injection f : Z → X with f (Z) ⊆ E[x] is bornologous. Hence, we may without loss of generality assume that X has bounded geometry. On the other hand, there must exist E ∈ E and x ∈ X such that [E][x] is infinite. Otherwise, W (X) would S have to be locally residually finite: for any finite subset F ⊆ W (X), since E := {gr(α) | α ∈ F } ∈ E , the homomorphism Y
hF i → Sym([E][x]), α 7→ α|[E][x] x∈X x∈X would embed hF i into a product of finite groups. So, let E ∈ E and x ∈ X such that [E][x] is infinite. S Without loss of generality, we may assume that ∆X ⊆ E = E −1 . Therefore, [E] = {E n | n ∈ N}. We conclude that E n [x] 6= E n+1 [x] and thus  Rn := f ∈ X N f0 = x, |{f0 , . . . , fn }| = n + 1, ∀i ∈ N : (fi , fi+1 ) ∈ E is non-empty for all Q n ∈ N. As (Rn )n∈N is a chain T of closed subsets of the compact topological space m∈N E m [x], we have R := n∈N Rn 6= ∅. Since any member of R is a bornologous injection from N into X, this implies (1) by Remark 4.6(ii).  Example 4.7. Let I be a partition of N into finite intervals with supI∈I |I| = ∞. Consider the metric space X := (N, d) given by ( |x − y| if x, y ∈ I for some I ∈ I , d(x, y) := (x, y ∈ N). max(x, y) otherwise It is easy to see that X has uniformly bounded geometry. Moreover, by Lemma 4.1 and the unboundedness assumption for the interval lengths, it follows that X has positive asymptotic dimension. On the other hand, essentially by finiteness of the considered intervals, there is no bornologous injection from N into X. Due to Proposition 4.5, this readily implies that W (Z) does not embed into W (X). ABOUT VON NEUMANN’S PROBLEM FOR LOCALLY COMPACT GROUPS 13 The interplay between certain geometric properties of coarse spaces on the one hand and algebraic peculiarities of their wobbling groups on the other is a subject of recent attention [12, 3]. It would be interesting to have further results in that direction, e.g., to understand if (and how) specific positive values for the asymptotic dimension may be characterized in terms of wobbling groups. Acknowledgments The author would like to thank Andreas Thom for interesting discussions about Whyte’s variant of the von Neumann conjecture, as well as Warren Moors and Jens Zumbrägel for their helpful comments on earlier versions of this note. References [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] Tullio Ceccherini-Silberstein and Michel Coornaert, The Hall Harem Theorem, In: Cellular Automata and Groups, Springer Monographs in Mathematics, Springer Berlin Heidelberg (2010), pp. 391–401. Tullio Ceccherini-Silberstein, Rostislav I. Grigorchuk, and Pierre de la Harpe, Amenability and paradoxical decompositions for pseudogroups and discrete metric spaces, Proc. Steklov Inst. Math. 224 (1999), pp. 57–97. Yves Cornulier, Irreducible lattices, invariant means, and commensurating actions, Math. Z. 279 (2015), no. 1–2, pp. 1–26. David van Dantzig, Zur topologischen Algebra. III. Brouwersche und Cantorsche Gruppen, Compositio Math. 3 (1936), pp. 408–426. Mahlon M. Day, Amenable semigroups, Illinois J. Math. 1 (1957), pp. 509–544. Eric K. van Douwen, Measures invariant under actions of F2 , Topology and its Applications 34 (1990), pp. 53–68. Jacob Feldman and Frederick P. Greenleaf, Existence of Borel transversals in groups, Pacific J. Math. 25 (1968), pp. 455–461. Erling Følner, On groups with full Banach mean value, Math. Scand. 3 (1955), pp. 243–254. Damien Gaboriau and Russell Lyons, A measurable-group-theoretic solution to von Neumann’s problem, Invent. Math. 177 (2009), no. 3, pp. 533–540. Frederick P. Greenleaf, Invariant means on topological groups and their applications, Van Nostrand Mathematical Studies, No. 16. Van Nostrand Reinhold Co., New York-Toronto, Ont.-London, 1969, pp. ix+113. Kate Juschenko and Nicolas Monod, Cantor systems, piecewise translations and simple amenable groups, Ann. of Math. (2) 178 (2013), no. 2, pp. 775–787. Kate Juschenko and Mikael de la Salle, Invariant means for the wobbling group, Bull. Belg. Math. Soc. Simon Stevin 22 (2015), no. 2, pp. 281–290. Andrew Marks and Spencer Unger, Baire measurable paradoxical decompositions via matchings, Adv. Math. 289 (2016), pp. 397–410. Maxime Gheysens and Nicolas Monod, Fixed points for bounded orbits in Hilbert spaces, August 2015, arXiv: 1508.00423[math.GR], to appear in Annales scientifiques de l’École normale supérieure. Michail Gromov, Asymptotic invariants of infinite groups. In: Geometric group theory, Vol. 2 (Sussex, 1991). London Math. Soc. Lecture Note Ser. 182, Cambridge Univ. Press, Cambridge, 1993, pp. 1–295. Karl W. Gruenberg, Residual properties of infinite soluble groups, Proc. London Math. Soc. (3) 7 (1957), pp. 29–62. Pierre de la Harpe, Topics in geometric group theory, Chicago Lectures in Mathematics. University of Chicago Press, Chicago, IL, 2000, pp. vi+310. John von Neumann, Über die analytischen Eigenschaften von Gruppen linearer Transformationen und ihrer Darstellungen, Math. Z. 30 (1929), no. 1, pp. 3–42. Alexander Ju. Ol’šanskiı̆, On the question of the existence of an invariant mean on a group, Uspekhi Mat. Nauk 35 (1980), no. 4(214), pp. 199–200. Theodore W. Palmer, Banach algebras and the general theory of ∗-algebras. Vol. 2, Encyclopedia of Mathematics and its Applications 79. Cambridge University Press, Cambridge, 2001, pp. i–xii and 795–1617. Alan L. T. Paterson, Nonamenability and Borel paradoxical decompositions for locally compact groups, Proc. Amer. Math. Soc. 96 (1986), pp. 89–90. Neil W. Rickert, Amenable groups and groups with the fixed point property, Trans. Amer. Math. Soc. 127 (1967), pp. 221–232. 14 FRIEDRICH MARTIN SCHNEIDER [23] Neil W. Rickert, Some properties of locally compact groups, J. Austral. Math. Soc. 7 (1967), pp. 433–454. [24] John Roe, Lectures on coarse geometry, University Lecture Series 31. American Mathematical Society, Providence, RI, 2003, pp. viii+175. [25] Joseph M. Rosenblatt, A generalization of Følner’s condition, Math. Scand. 33 (1973), no. 3, pp. 153–170. [26] Friedrich M. Schneider and Andreas B. Thom, On Følner sets in topological groups, August 2016, arXiv: 1608.08185[math.GR]. [27] Kevin Whyte, Amenability, bi-Lipschitz equivalence, and the von Neumann conjecture, Duke Math. J. 99 (1999), no. 1, pp. 93–112. Institute of Algebra, TU Dresden, 01062 Dresden, Germany Current address: Department of Mathematics, The University of Auckland, Private Bag 92019, Auckland 1142, New Zealand E-mail address: martin.schneider@tu-dresden.de 

1 Adaptive Coded Caching for Fair Delivery over Fading Channels arXiv:1802.02895v1 [cs.IT] 7 Feb 2018 Apostolos Destounis, Member, IEEE, Asma Ghorbel, Student Member, IEEE Georgios S. Paschos, Senior Member, IEEE, and Mari Kobayashi, Senior Member, IEEE, Abstract—The performance of existing coded caching schemes is sensitive to the worst channel quality, a problem which is exacerbated when communicating over fading channels. In this paper, we address this limitation in the following manner: in short-term, we allow transmissions to subsets of users with good channel quality, avoiding users with fades, while in long-term we ensure fairness among users. Our online scheme combines the classical decentralized coded caching scheme [1] with (i) joint scheduling and power control for the fading broadcast channel, as well as (ii) congestion control for ensuring the optimal long-term average performance. We prove that our online delivery scheme maximizes the alpha-fair utility among all schemes restricted to decentralized placement. By tuning the value of alpha, the proposed scheme enables to balance between different operating points on the average delivery rate region. We demonstrate via simulations that our scheme outperforms two baseline schemes: (a) standard coded caching with multicast transmission, limited by the worst channel user yet exploiting the global caching gain; (b) opportunistic scheduling with unicast transmissions exploiting only the local caching gain. Index Terms—Broadcast channel, coded caching, fairness, Lyapunov optimization. I. I NTRODUCTION A key challenge for the future wireless networks is the increasing video traffic demand, which reached 70% of total mobile IP traffic in 2015 [2]. Classical downlink systems cannot meet this demand since they have limited resource blocks, and therefore as the number K of simultaneous video transfers increases, the per-video throughput vanishes as 1/K. Recently it was shown that scalable per-video throughput can be achieved if the communications are synergistically designed with caching at the receivers. Indeed, the recent breakthrough of coded caching [3] has inspired a rethinking of wireless downlink. Different video sub-files are cached at the receivers, and video requests are served by coded multicasts. By careful selection of sub-file caching and exploitation of the wireless broadcast channel, the transmitted signal is simultaneously useful for decoding at users who requested different video files. This scheme has been theoretically proven to scale well, and therefore has the potential to resolve the challenge of A. Destounis and G. S. Paschos are with the Mathematical and Algorithmic Sciences Lab, France Research Center - Huawei Technologies Co. Ltd., 20 quai de Point du Jour, 92100 Boulogne-Bilancourt, France. Emails: firstname.lastname@hauwei.com M. Kobayashi and A. Ghorbel are with the Laboratoire des Signaux et Systèmes (L2S), CentraleSupélec, Université Paris-Saclay, 3, Rue Joliot-Curie, 91192 Gif sur Yvette, France. Emails: firstname.lastname@centralesupelec.fr Part of the work in this paper has been presented at the 15th International Symposium on Modeling and Optimization in Mobile, Ad Hoc, and Wireless Networks (WiOpt), Telecom ParisTech, Paris, France, 15th - 19th May, 2017. downlink bottleneck for future networks. Nevertheless, several limitations hinder its applicability in practical systems [4]. In this work, we take a closer look to the limitations that arise from the fact that coded caching was originally designed for a symmetric error-free shared link. If instead we consider a realistic wireless channel, we observe that coded caching faces a short-term limitation. Namely, its performance is limited by the user in the worst channel condition because the wireless multicast capacity is determined by the worst user [5, Chapter 7.2]. This is in stark contrast with standard downlink techniques such as opportunistic scheduling [6]–[8], which serve the user with the best instantaneous channel quality. Thus, a first challenge is to modify coded caching for exploitation of fading peaks, similar to the opportunistic scheduling. In addition to the fast fading consideration, there is also a long-term limitation due to a network topology. Namely, the ill-positioned users, e.g. users at the cell edge, may experience consistently poor channel quality during a whole video delivery. The classical coded caching scheme is designed to provide video files at equal data rates to all users, which leads to illpositioned users consuming most of the air time and hence driving the overall system performance to low efficiency. In the literature of wireless scheduling without caches at receivers, this problem has been resolved by the use of fairness among user throughputs [7]. By allowing poorly located users to receive less throughput than others, precious air time is saved and the overall system performance is greatly increased. Since the sum throughput rate and equalitarian fairness are typically the two extreme objectives, past works have proposed the use of alpha-fairness [9] which allows to select the coefficient α and drive the system to any desirable tradeoff point in between of the two extremes. Previously, the alpha-fair objectives have been studied in the context of (i) multiple user activations [6], (ii) multiple antennas [10] and (iii) broadcast channels [11]. However, in the presence of caches at user terminals, the fairness problem is further complicated by the interplay between user scheduling and designing codewords for multiple users. In particular, we wish to shed light into the following questions: Which user requests shall we combine together to perform coded caching? How shall we schedule a set of users to achieve our fairness objective while adapting to timevarying channel quality? To address these questions, we study the content delivery over a realistic block-fading broadcast channel, where the channel quality varies across users and time. Although the decisions of user scheduling and codeword design are inher- 2 ently coupled, we design a scheme which decouples these two problems, while maintaining optimality through a specifically designed queueing structure. On the transmission side, we select the multicast user set dynamically depending on the instantaneous channel quality and user urgency captured by queue lengths. On the coding side, we adapt the codeword construction of [1] to the set of users chosen by the appropriate routing which depends also on the past transmission side decisions. Combining with an appropriate congestion controller, we show that this approach yields our alpha-fair objective. More specifically, our approaches and contributions are summarized below: 1) We design a novel queueing structure which decouples the channel scheduling from the codeword construction. Although it is clear that the codeword construction needs to be adaptive to channel variation, our scheme ensures this through our backpressure that connects the user queues and the codeword queues. Hence, we are able to show that this decomposition is without loss of optimality (see Theorem 6). 2) We then provide an online policy consisting of (i) admission control of new files into the system; (ii) combination of files to perform coded caching; (iii) scheduling and power control of codeword transmissions to subset of users on the wireless channel. We prove that the longterm video delivery rate vector achieved by our scheme is a near optimal solution to the alpha-fair optimization problem under the restriction to policies that are based on the decentralized coded caching scheme [1]. 3) Through numerical examples, we demonstrate the superiority of our approach versus (a) standard coded caching with multicast transmission limited by the worst channel condition yet exploiting the global caching gain, (b) opportunistic scheduling with unicast transmissions exploiting only the local caching gain. This shows that our scheme not only is the best among online decentralized coded caching schemes, but moreover manages to exploit opportunistically the time-varying fading channels. A. Related work Since coded caching was first introduced in [3] and its potential was recognized by the community, substantial efforts have been devoted to quantify the gain in realistic scenarios, including decentralized placement [1], non-uniform popularities [12], [13], and more general network topologies (e.g. [14]–[16]). A number of recent works have studied coded caching by replacing the original perfect shared link with wireless channels [17]–[22]. In particular, the harmful effect of coded caching over wireless multicast channels has been highlighted recently [17], [18], [21], [23], while similar conclusions and some directions are given in [17], [18], [20], [23]. Although [23] consider the same channel model and address a similar question as in the current work, they differ in their objectives and approaches. [23] highlights the scheduling part and provides rigorous analysis on the long-term average per-user rate in the regime of large number of users. In the current work, a new queueing structure is proposed to deal Fig. 1. Decentralized coded caching for N = K = 3 over block-fading broadcast channel jointly with admission control, routing, as well as scheduling for a finite number of users. Furthermore, most of existing works have focused on offline caching where both cache placement and delivery phases are performed once without capturing the random and asynchronous nature of video traffic. The works [24], [25] addressed partly the online aspect by studying cache eviction strategies, the delivery delay, respectively. In this work, we will explore a different online aspect. Namely, we assume that the file requests from users arrive dynamically and the file delivery is performed continuously over time-varying fading broadcast channels. Finally, online transmission scheduling over wireless channels has been extensively studied in the context of opportunistic scheduling [6] and network utility maximization [26]. Prior works emphasize two fundamental aspects: (a) the balancing of user rates according to fairness and efficiency considerations, and (b) the opportunistic exploitation of the time-varying fading channels. There have been some works that study scheduling policies over a queued-fading downlink channel; [27] gives a maxweight-type of policy and [28] provides a throughput optimal policy based on a fluid limit analysis. Our work is the first to our knowledge that studies coded caching in this setting. The new element in our study is the joint consideration of user scheduling with codeword construction for the coded caching delivery phase. II. C ODED C ACHING OVER W IRELESS C HANNELS A. System Model We consider a content delivery network where a server (or a base station) wishes to convey requested files to K user terminals over a wireless channel; in Fig. 1 we give an example with K=3. The wireless channel is modeled by a standard block-fading broadcast channel, such that the channel state remains constant over a slot and changes from one slot to another in an i.i.d. manner. Each slot is assumed to allow for Tslot channel uses. The channel output of user k in any channel use of slot t is given by p x(t) + ν k (t), (1) y k (t) = hk (t)x 3 T where the channel input x ∈ C slot is subject to the xk2 ] ≤ P Tslot ; ν k (t) ∼ NC (0, ITslot ) power constraint E[kx are additive white Gaussian noises with covariance matrix identity of size Tslot , assumed independent of each other; {hk (t) ∈ C} are channel fading coefficients independently distributed across time. At each slot t, the channel state h(t) = (h1 (t), . . . , hK (t)) is perfectly known to the base station while each user knows its own channel realization. We follow the network model considered in [3] as well as its follow-up works. The server has an access to N equally popular files W1 , . . . , WN , each F bits long, while each user k is equipped with cache memory Zk of M F bits, where M ∈ {0, 1, . . . , N }. We restrict ourselves to decentralized cache placement [1]. More precisely, each user k independently caches a subset of MNF bits of file i, chosen uniformly at random for i = 1, . . . , N , under its memory constraint of M F bits. For later use, we let m = M N denote the normalized memory size. By letting Wi|J denote the sub-file of Wi stored exclusively in the cache memories of the user set J, the cache memory Zk of user k after decentralized placement is given by Zk = {Wi | J : J ⊆ {1, . . . , K}, k ∈ J, ∀i = 1, . . . , N }. (2) Under the assumption of large file size (F → ∞), we use the law of large numbers to calculate the size of each sub-file (measured in bits) as the following K−|J| |Wi | J | = m|J| (1 − m) F. (3) Once the requests of all users are revealed, decentralized coded caching proceeds to the delivery of the requested files (delivery phase). Assuming that user k demands file k, and writing dk = k, the server generates and conveys the following codeword simultaneously useful to the subset of users J: VJ = ⊕k∈J Wk|J\{k} , (4) where ⊕ denotes the bit-wise XOR operation. The central idea of coded caching is to create a codeword simultaneously useful to a subset of users by exploiting the receiver side information established during the placement phase. This multicasting operation leads to a gain: let us consider the uncoded delivery such that sub-files are sent sequentially. The total number of transmissions intended to |J| users is equal to |J|×|Wk|J\{k} |. The coded delivery requires the transmission of |Wk|J\{k} |, yielding a reduction of a factor |J|. It can be shown that the transmitted signal as per (4) can be decoded correctly with probability 1 by all intended receivers. In order to further illustrate the placement and delivery of decentralized coded caching, we provide a three-user example in Fig. 1. Example 1. Let us assume that user 1, 2, 3, requests file A, B, C, respectively. After the placement phase, a given file A will be partitioned into 8 sub-files, one per user subset. Codewords to be sent are the following: • A∅ , B∅ and C∅ to user 1, 2 and 3, respectively. • A2 ⊕ B1 is intended to users {1, 2}. Once received, user 1 decodes A2 by combining the received codeword with B1 given in its cache. Similarly user 2 decodes B1 . The • same holds for codeword B3 ⊕ C2 to users {2, 3} and codeword A3 ⊕ C1 to users {1, 3}, respectively. A23 ⊕ B13 ⊕ C12 is intended users {1, 2, 3}. User 1 can decode A23 by combining the received codeword with {B13 , C12 } given in its cache. The same approach is used for user 2, 3 to decode B13 , C12 respectively. In order to determine the user throughput under this scheme we must inspect the achievable transmission rate per codeword, then determine the total time to transmit all codewords, and finally extract the user throughput. To this aim, the next subsection will specify the transmission rates of each codeword by designing a joint scheduling and power allocation to subsets of users. B. Degraded Broadcast Channel with Private and Common Messages The placement phase creates 2K − 1 independent sub-files {VJ }J⊆{1,...,K} , each intended to a subset of users. We address the question on how the transmitter shall convey these subfiles while opportunistically exploiting the underlying wireless channel. We start by remarking that the channel in (1) for a given channel realization h is stochastically degraded BC which achieves the same capacity region as the physically degraded BC [5, Sec. 5]. The capacity region of the degraded broadcast channel for K private messages and a common message is well-known [5]. Here, we consider the extended setup where the transmitter wishes to convey 2K − 1 mutually independent messages, denoted by {MJ }, where MJ denotes the message intended to the users in subset J ⊆ {1, . . . , K}. We require that each user k must decode all messages {MJ } for J 3 k. By letting RJ denote the multicast rate of the 2K −1 is message MJ , we say that the rate-tuple R ∈ R+ achievable if there exists encoding and decoding functions which ensure the reliability and the rate condition as the slot duration Tslot is taken arbitrarily large. The capacity region is defined as the supremum of the achievable rate-tuple as shown in [23], where the rate is measured in bit/channel use. h) of a K-user degraded Theorem 1. The capacity region Γ (h Gaussian broadcast channel with fading gains h1 ≥ · · · ≥ hK and 2K − 1 independent messages {MJ } is given by R1 ≤ log(1 + h1 p1 ) Pk X 1 + hk j=1 pj RJ ≤ log Pk−1 1 + hk j=1 pj J⊆{1,...,k}:k∈J for non-negative variables {pk } such that (5) k = 2, . . . , K (6) PK k=1 pk ≤ P . Proof. The proof is quite straightforward and is based on ratesplitting and the private-message region of degraded broadcast channel. For completeness, see details in Appendix IX-A. The achievability builds on superposition coding at the transmitter and successive interference cancellation at receivers. For K = 3, the transmit signal is simply given by x = x1 + x2 + x3 + x12 + x13 + x23 + x123 , 4 where xJ denotes the signal corresponding to the message MJ intended to the subset J ⊆ {1, 2, 3}. We suppose that all {xJ : J ⊆ {1, . . . , K}} are mutually independent Gaussian distributed random variables satisfying the power constraint. User 3 (the weakest user) decodes M̃3 = {M3 , M13 , M23 , M123 } by treating all the other messages as noise. User 2 decodes first the messages M̃3 and then jointly decodes M̃2 = {M2 , M12 }. Finally, user 1 (the strongest user) successively decodes M̃3 , M̃2 and, finally, M1 . Later in our online coded caching scheme, we will need to characterize specific boundary points of the capacity region h) that maximize a weighted sum rate. To this end, it Γ (h suffices to consider the weighted sum rate maximization: X max θJ rJ . (7) h) r ∈Γ (h J:J⊆{1,...,K} We first simplify the problem using the following theorem. Theorem 2. The weighted sum rate maximization with 2K −1 variables in (7) reduces to a simpler problem with K variables, given by Pk K X 1 + hk j=1 pj (8) max θ̃k log Pk−1 . p 1 + hk j=1 pj k=1 K where p = (p1 , . . . , pK ) ∈ R+ is a positive real vector satisfying the total power constraint, and θ̃k denotes the largest weight for user k θ̃k = max K:k∈K⊆{1,...,k} θK . Proof. The proof builds on the simple structure of the capacity region. We remark that for a given power allocation of users 1 to k − 1, user k sees 2k−1 messages {MJ } for all J such that k ∈ J ⊆ {1, . . . , k} with the equal channel gain. For a given set of {pj }k−1 j=1 , the capacity region of these messages is a simple hyperplane characterized by 2k−1 vertices R̃ke i for i = 1, . . . , 2k−1 , where R̃k is the sum rate of user k in the RHS of (6) and e i is a vector with one for the i-th entry and zero for the others. Therefore, the weighted sum rate is maximized for user k by selecting the vertex corresponding to the largest weight, denoted by θ̃. This holds for any k. We provide an efficient algorithm to solve this power allocation problem as a special case of the parallel Gaussian broadcast channel studied in [29, Theorem 3.2]. Following [29], we define the rate utility function for user k given by uk (z) = θ̃k − λ, 1/hk + z (9) where λ is a Lagrangian multiplier. The optimal solution corresponds to selecting the user with the maximum rate utility at each z and the resulting power allocation for user k is   ∗ pk = z : [max uj (z)]+ = uk (z) (10) j with λ satisfying " θ̃k 1 P = max − k λ hk # . + (11) Throughout the paper, we assume that each slot is arbitrarily large to achieve transmission rates of the whole capacity region of the broadcast channel (as given above) without errors, for each possible channel realization. This is necessary to ensure the successful decoding of each sub-file at the receivers. C. Application to Online Delivery In this subsection, we wish to apply the superposition encoding over different subsets of users, proposed in the previous subsection to the online delivery phase of decentralized coded caching. Compared to the original decentralized coded caching in [1], we introduce here the new ingredients: i) at each slot, the superposition based delivery scheme is able to serve multiple subsets of users, such that each user shall decode multiple sub-files; ii) users’ requests arrive randomly and each user decodes a sequence of its requested files. In the original framework [1], [3], the vector of user requests, denoted by d = (d1 , . . . , dK ), is assumed to be known by all users. This information is necessary for each user to recover its desired sub-files by operating XOR between the received signal and the appropriate sub-files available in its cache content. Let us get back to the three-user example in Fig. 1. Upon the reception of A2 ⊕ B1 , user 1 must identify both its desired sub-file identity (A2 ) and the combined sub-file available in its cache (B1 ). Similarly upon the reception of A23 ⊕ B13 ⊕ C12 , user 1 must identify its desired sub-file A23 and the combined sub-files B13 , C12 . In the case of a single request per user, the base station simply needs to disseminate the vector of user requests. However, if user requests arrive dynamically and the delivery phase is run continuously, we associate a header to identify each sub-file (combined files index and intended receivers) as we discuss in details in Section V-C. At the end of the whole transmission as t → ∞, each receiver decodes its sequence of requested files by applying a decoding function ξk to the sequence of the received signals y tk = (yy k (1), . . . , y k (t)), that of its channel state hk (1), . . . , h k (t)), its cache Zk . Namely, the output h tk = (h of the k-th user’s decoding function at slot t is given by F D̂k (t) ξk (t) = ξk (Zk , y tk , h tk ) ∈ F2 (12) where D̂k (t) is defined to be the number of decoded files by user k up to slot t. Under the assumption that Tslot is arbitrarily large, each receiver can successfully decode the sequence of the encoded symbols and reconstruct its requested files. III. P ROBLEM F ORMULATION After specifying the codeword generation and the transmission scheme over the broadcast channel, this section will formulate the problem of alpha-fair file delivery. Now we are ready to define the feasible rate region as the set of the average number of successfully delivered files for K users. We let rk denote time average delivery rate of user k, measured in files par slot. We let Λ denote the set of all feasible delivery rate vectors. Definition 1 (Feasible rate). A rate vector r = (r1 , . . . , rK ), measured in file/slot, is said to be feasible r ∈ Λ if there exist a file combining and transmission scheme such that 5 rk = lim inf t→∞ Dk (t) . t (13) where Dk (t) denotes the number of successfully delivered files to user k up to t. It is worth noticing that as t → ∞ the number of decoded files D̂k (t) shall coincide with the number of successfully delivered files Dk (t) under the assumptions discussed previously. In contrast to the original framework [1], [3], our rate metric measures the ability of the system to continuously and reliably deliver requested files to the users. Since finding the optimal policy is very complex in general, we restrict our study to a specific class of policies given by the following mild assumptions: Definition 2 (Admissible class policies Π CC ). The admissible policies have the following characteristics: 1) The caching placement and delivery follow the decentralized scheme [1]. 2) The users request distinct files, i.e. the IDs of the requested files of any two users are different. Since we restrict our action space, the feasibility rate region, denoted by ΛCC , under the class of policies Π CC is smaller than the one for the original problem Λ. However, the joint design of caching and online delivery appears to be a very hard problem; note that the design of an optimal code for coded caching alone is an open problem and the proposed solutions are constant factor approximations. Restricting the caching strategy to the decentralized scheme proposed in [1] makes the problem amenable to analysis and extraction of conclusions for general cases such as the general setup where users may not have the symmetrical rates. Additionally, if two users request the same file simultaneously, it is efficient to handle exceptionally the transmissions as naive broadcasting instead of using the decentralized coded caching scheme, yielding a small efficiency benefit but complicating further the problem. Note, however, the probability that two users simultaneously request the same parts of video is very low in practice, hence to simplify our model we exclude this consideration altogether. Our objective is to solve the fair file delivery problem: ∗ r =arg max r∈ΛCC K X g(rk ), (14) k=1 where the utility function corresponds to the alpha fair family of concave functions obtained by choosing: ( (d+x)1−α , α 6= 1 1−α g(x) = (15) log(1 + x/d), α = 1 for some arbitrarily small d > 0 (used to extend the domain of the functions to x = 0). Tuning the value of α changes the shape of the utility function and consequently drives the system performance r ∗ to different operating points: (i) α = 0 yields max sum delivery rate, (ii) α → ∞ yields max-min delivery rate [9], (iii) α = 1 yields proportionally fair delivery rate [30]. Choosing α ∈ (0, 1) leads to a tradeoff between max sum and proportionally fair delivery rates. Fig. 2. Illustration of the feasibility region and different performance operating points for K = 2 users. Point A corresponds to a naive adaptation of [3] on our channel model, while the rest points are solutions to our fair delivery problem. The optimization (14) is designed to allow us tweak the performance of the system; we highlight its importance by an example. Suppose that for a 2-user system Λ is given by the convex set shown on Fig. 2. Different boundary points are obtained as solutions to (14). If we choose α = 0, the system is operated at the point that maximizes the sum r1 + r2 . The choice α → ∞ leads to the maximum r such that r1 = r2 = r, while α = 1 maximizes the sum of logarithms. The operation point A is obtained when we always broadcast to all users at the weakest user rate and use [3] for coded caching transmissions. Note that this results in a significant loss of efficiency due to the variations of the fading channel, and consequently A lies in the interior of Λ. To reach the boundary point that corresponds to α → ∞ we need to carefully group users together with good instantaneous channel quality but also serve users with poor average channel quality. This shows the necessity of our approach when using coded caching in realistic wireless channel conditions. IV. Q UEUED D ELIVERY N ETWORK This section presents the queued delivery network and then the feasible delivery rate region, based on stability analysis of the queueing model. A. Queueing Model At each time slot t, the controller admits ak (t) files to be delivered to user k, and hence ak (t) is a control variable. We equip the base station with the following types of queues: 1) User queues to store admitted files, one for each user. The buffer size of queue k is denoted by Sk (t) and expressed in number of files. 2) Codeword queues to store codewords to be multicast. There is one codeword queue for each subset of users I ⊆ {1, . . . , K}. The size of codeword queue I is denoted by QI (t) and expressed in bits. A queueing policy π performs the following operations: (i) it decides how many files to admit into the user queues Sk (t) in the form of (ak (t)) variables, (ii) it combines files destined to different users to create multiple codewords. When a new codeword is form in this way, we denote this with 6 the codeword routing control variable σJ , that denotes the number of combinations among files from the subset J f users according to the coded caching scheme in [3], (iii) it decides the encoding function for the wireless transmission. Below we explain in detail the queue operations and the queue evolution: 1) Admission control: At the beginning of each slot, the controller decides how many requests for each user, ak (t) should be pulled into the system from the infinite reservoir. 2) Codeword Routing: The admitted files for user k are stored in queues Sk (t) for k = 1, . . . , K. At each slot, files from subsets of these queues are combined into codewords by means of the decentralized coded caching encoding scheme. Specifically, the decision at slot t for a subset of users J ⊆ {1, .., K}, denoted by σJ (t) ∈ {0, 1, . . . , σmax }, refers to the number of combined requests for this subset of users. 1 The size of the user queue Sk evolves as: X +  (16) Sk (t + 1) = Sk (t) − σJ (t) + ak (t) | {z } J:k∈J number of | {z } admitted files number of files combined into codewords If σJ (t) > 0, the server creates codewords by applying (4) for this subset of users as a function of the cache contents {Zj : j ∈ J}. 3) Scheduling: The codewords intended to the subset I of users are stored in codeword queue whose size is given by QI (t) for I ⊆ {1, . . . , K}. Given the instantaneous channel realization h (t) and the queue state {QI (t)}, the server performs multicast scheduling and rate allocation. Namely, at slot t, it determines the number µI (t) of bits per channel use to be transmitted for the users in subset I. By letting bJ,I denote the number of bits generated for codeword queue I ⊆ J when coded caching is performed to the users in J, codeword queue I evolves as X  + + bJ,I σJ (t) QI (t + 1) = QI (t) − Tslot µI (t) | {z } J:I⊆J number of bits | {z } multicast to I number of bits created by combining files (17) where bJ,I = m|I| (1 − m)|J|−|I|−1 . A control policy is fully specified by giving the rules with a(t), σ(t), µ(t)} are taken at every slot which the decisions {a t. The first step towards this is to characterize the set of feasible delivery rates, ΛCC , which is the subject of the next subsection. B. Feasibility Region The main idea here is to characterize the set of feasible file delivery rates via characterizing the stability performance of the queueing system. To this end, let ak = 1 It is worth noticing that standard coded caching lets σ = 1 for J = J {1, . . . , K} and zero for all the other subsets. On the other hand, uncoded caching can be represented by sigmaJ = 1 for J = k, k ∈ 1, ...., K. Our scheme can, therefore be seen as a combination of both, which explains its better performance. lim sup 1t t→∞ Pt−1 t=0 E [ak (t)] , denote the time average number of admitted files for user k. We use the following definition of stability: Definition 3 (Stability). A queue S(t) is said to be (strongly) stable if T −1 1 X lim sup E [S(t)] < ∞. T →∞ T t=0 A queueing system is said to be stable if all its queues are stable. Moreover, the stability region of a system is the set of all vectors of admitted file rates such that the system is stable. If the queueing system we have introduced is stable the rate of admitted files (input rate) is equal to the rate of successfully decoded files (output rate), hence we can characterize the system performance by means of the stability region of our h) denote the capacity region for queueing system. We let Γ (h a fixed channel state h, as defined in Theorem 1. Then we have the following: Theorem 3 (Stability region). Let Γ CC be a set to which a rate vectorP of admitted files a belongs to, if and only if there exist µ ∈ h∈H φh Γ (h), σ I ∈ [0, σmax ], ∀I ⊆ {1, . . . , K} such that: X σ J ≥ ak , ∀k = 1, . . . , K (18) J:k∈J Tslot µI ≥ X bJ,I σ J , ∀I ⊆ {1, 2, ..., K}. (19) J:I⊆J Then, the stability region of the system is the interior of Γ CC , where the above inequalities are strict. Constraint (18) says that the aggregate service rate is greater than the arrival rate, while (19) implies that the long-term average rate for the subset J is greater than the arrival rate of the codewords intended to this subset. In terms of the queueing system defined, these constraints impose that the service rates of each queue should be greater than their arrival rates, thus rendering them stable 2 . The proof of this theorem relies on existence of static policies, i.e. randomized policies whose decision distribution depends only on the realization of the channel state. See the Appendix, Section IX-B for a definition and results on these policies. Since the channel process h (t) is a sequence of i.i.d. realizations of the channel states (the same results hold if, more generally, h (t) is an ergodic Markov chain), we can obtain any admitted file rate vector a in the stability region by a a(t), σ(t), µ(t)} Markovian policy, i.e. a policy that chooses {a based only the state of the system at the beginning of time h(t), S (t), Q (t)}, and not the time index itself. This slot t, {h S (t), Q (t)) evolves as a Markov chain, therefore implies that (S our stability definition is equivalent to that Markov chain being ergodic with every queue having finite mean under the stationary distribution. Therefore, if we develop a policy that keeps user queues S (t) stable, then all admitted files will, at some point, be combined into codewords. Additionally, if 2 We restrict vectors a to the interior of Γ CC , since arrival rates at the boundary are exceptional cases of no practical interest, and require special treatment. 7 codeword queues Q (t) are stable, then all generated codewords will be successfully conveyed to their destinations. This in turn means that all receivers will be able to decode the admitted files that they requested: Lemma 4. The region of all feasible delivery rates ΛCC is the same as the stability region of the system, i.e. ΛCC = Int(Γ CC ). Proof. Please refer to Appendix IX-C. Lemma 4 implies the following Corollary. ak (t) = γk,max 1{Uk (t) ≥ Sk (t)} k=1 the system is stable. This implies that the solution to the original problem (14) in terms of the long-term average rates is equivalent to the new problem in terms of the admission rates stabilizing the system. Next Section provides a set of the explicit solutions to this new problem. V. P ROPOSED O NLINE D ELIVERY S CHEME (23) For every subset J ⊆ {1, . . . , K}, routing combines σJ (t) demands of users in J given by    X X bJ,I Q (t) . (24) Sk (t) > σJ (t) = σmax 1 I   F2 I:I⊆J k∈J Corollary 5. Solving (14) is equivalent to finding a policy π K such that X aπ =arg max (20) gk (ak ) s.t. We present our on-off policy for admission control and routing. For every user k, admission control chooses ak (t) demands given by B. Scheduling and Transmission In order to stabilize all codeword queues, the scheduling and resource allocation explicitly solve the following weighted sum rate maximization at each slot t where the weight of the subset J corresponds to the queue length of QJ X µ(t) = arg max QJ (t)rJ . (25) h(t)) r ∈Γ (h J⊆{1,...,K} We propose to apply the power allocation algorithm in subsection II-B to solve the above problem by sorting users in a decreasing order of channel gains and treating QJ (t) as θJ . Algorithm 1 summarizes our online delivery scheme. A. Admission Control and Codeword Routing Our goal is to find a control policy that optimizes (20). To this aim, we need to introduce one more set of queues. These queues are virtual, in the sense that they do not hold actual file demands or bits, but are merely counters to drive the control policy. Each user k is associated with a queue Uk (t) which evolves as follows: + Uk (t + 1) = [Uk (t) − ak (t)] + γk (t) (21) where γk (t) represents the arrival process to the virtual queue and is an additional control parameter. We require these queues to be stable: The actual mean file admission rates are greater than the virtual arrival rates and the control algorithm actually seeks to optimize the time average of the virtual arrivals γk (t). However, since Uk (t) is stable, its service rate, which is the actual admission rate, will be greater than the rate of the virtual arrivals, therefore giving the same optimizer. Stability of all other queues will guarantee that admitted files will be actually delivered to the users. With thee considerations, Uk (t) will be a control indicator such that when Uk (t) is above Sk (t) then we admit files into the system else we set ak (t) = 0. In particular, we will control the way Uk (t) grows over time using the actual utility objective gk (.) such that a user with rate x and rapidly increasing utility gk (x) (steep derivative at x) will also enjoy a rapidly increasing Uk (t) and hence admit more files into the system. In our proposed policy, the arrival process to the virtual queues are given by γk (t) = arg max 0≤x≤γk,max [V gk (x) − Uk (t)x] (22) In the above, V > 0 is a parameter that controls the utilitydelay tradeoff achieved by the algorithm (see Theorem 6). C. Practical Implementation When user requests arrive dynamically and the delivery phase is run continuously, it is not clear when and how the base station shall disseminate the useful side information to each individual users. This motivates us to consider a practical solution which associates a header to each sub-file Wi|J for i = 1, . . . , N and J ⊆ {1, . . . , K} . Namely, any sub-file shall indicate the following information prior to message symbols: a) the indices of files; b) the identities of users who cache (know) the sub-files 3 . At each slot t, the base station knows the cache contents of all users Z K , the sequence of the channel state h t , as well as that of the demand vectors d t . Given this information, the base station constructs and transmits either a message symbol or a header at channel use i in slot t as follows. ( h dt ft,i (d , Z K ) if header (26) xi (t) = m ft,i ({Wdk (τ ) : ∀k, τ ≤ t}, h t ) if message h m where ft,i , ft,i denotes the header function, the message encoding function, respectively, at channel use i in slot t. Example 2. We conclude this section by providing an example of our proposed online delivery scheme for K = 3 users as illustrated in Fig. 3. We focus on the evolution of codeword queues between two slots, t and t + 1. The exact backlog of codeword queues is shown in Table I. Given the routing and scheduling decisions (σJ (t) and µJ (t)), we provide the new states of the queues at the next slot in the same Table. 3 We assume here for the sake of simplicity that the overhead due to a header is negligible. This implies in practice that each of sub-files is arbitrarily large. 8 TABLE I C ODEWORD QUEUES EVOLUTION FOR µ{1,2} (t) > 0, µ{1,2,3} (t) > 0 AND σ{1,2} (t) = σ{1} (t) = 1. QJ (t) Output µ{1,2} (t) > 0, µ{1,2,3} (t) > 0 Input σ{1,2} (t) = σ{1} (t) = 1 QJ (t + 1) Q{1} A∅ Q{2} B∅ Q{3} C∅ Q{1,2} A2 ⊕ B 1 Q{1,3} A 3 ⊕ C1 Q{2,3} B3 ⊕ C2 Q{1,2,3} A23 ⊕ B13 ⊕ C12 - - - A2 ⊕ B 1 - - A23 ⊕ B13 ⊕ C12 E∅ ; E3 - - - - B∅ E∅ ; E3 C∅ A 3 ⊕ C1 B3 ⊕ C2 - D∅ ; D3 {FJ }1∈J / A∅ ; D∅ ; D3 {FJ }1∈J / E1 ⊕ D2 E13 ⊕ D23 E1 ⊕ D2 E13 ⊕ D23 Algorithm 1 Proposed delivery scheme 1: PLACEMENT (same as [1]): 2: Fill the cache of each user k Zk = {Wi | J : J ⊆ {1, . . . , K}, k ∈ J, ∀i = 1, . . . , N }. 3: DELIVERY: 4: for t = 1, . . . , T 5: Decide the arrival process to the virtual queues γk (t) = arg max [V gk (x) − Uk (t)x] 0≤x≤γk,max Decide the number of admitted files ak (t) = γk,max 1{Uk (t) ≥ Sk (t)} . 7: Update the virtual queues + Uk (t + 1) = [Uk (t) − ak (t)] + γk (t) 8: Decide the number of files to be combined    X X bJ,I Q (t) . σJ (t) = σmax 1 Sk (t) > I   F2 6: k∈J 9: I:I⊆J Scheduling decides the instantaneous rate P µ(t) = arg max J⊆{1,...,K} QJ (t)rJ . h(t)) r ∈Γ (h Fig. 3. An example of the queueing model for a system with 3 users. Dashed lines represent wireless transmissions, solid circles files to be combined and solid arrows codewords generated. 10: Update user queues queues:  and codeword + P Sk (t + 1) = Sk (t) − J:k∈J σJ (t) + ak (t), X + QI (t + 1) = [QI (t) − Tslot µI (t)] + bJ,I σJ (t). J:I⊆J We suppose that h1 (t) > h2 (t) > h3 (t). The scheduler uses (25) to allocate positive rates to user set {1, 2} and {1, 2, 3} given by µ{1,2} , µ{1,2,3} and multicasts the superposed signal x(t) = B∅ + B3 ⊕ C2 . User 3 decodes only B3 ⊕ C2 . User 2 decodes first B3 ⊕ C2 , then subtracts it and decodes B∅ . Note that the sub-file B∅ is simply a fraction of the file B whereas the sub-file B3 ⊕ C2 is a linear combination of two fractions of different files. In order to differentiate between each subfile, each user uses the data information header existing in the received signal. In the next slot, the received sub-files are evacuated from the codeword queues. For the routing decision, the server decides at slot t to combine D requested by user 1 with E requested by user 2 and to process F requested by user 1 uncoded. Therefore, we have σ{1,2} (t) = σ{1} (t) = 1 and σJ (t) = 0 otherwise. Given this codeword construction, codeword queues have inputs that change its state in the next slot as described in Table I. D. Performance Analyis Here we present the main result of the paper, by proving that our proposed online algorithm achieves near-optimal performance for all policies within the class Π CC : Theorem 6. Let rπk the mean time-average delivery rate for user k achieved by the proposed policy. Then K X k=1 lim sup T →∞ 1 T T −1 X t=0 gk (rπk ) ≥ max r ∈ΛCC K X k=1 gk (rk ) − B V n o B + V PK g (γ k=1 k max,k ) E Q̂(t) ≤ , 0 where Q̂(t) is the sum of all queue lengths at the beginning of time slot t, thus a measure of the mean delay of file delivery. The quantities B an 0 are constants that depend on the statistics of the system and are given in the Appendix. The above theorem states that, by tuning the constant V , the utility resulting from our online policy can be arbitrarily close to the optimal one, where there is a tradeoff between the guaranteed optimality gap O(1/V ) and the upper bound on the total buffer length O(V ). We note that these tradeoffs 9 TABLE II PARAMETERS QI (t) Sk (t) Uk (t) σJ (t) µI (t) ak (t) γk (t) Zk Dk (t) Ak (t) codeword queue storing XOR-packets intended users in I. user queue storing admitted files for user k. virtual queue for the admission control. decision variable of number of combined requests for users J in [0, σmax ]. decision variable for multicast transmission rate to users I. decision variable of the number of admitted files for user k in [0, γmax ]. the arrival process to the virtual queue in [0, γmax ], given by eq. (22). cache content for user k number of successfully decoded files by user k up to slot t. number of (accumulated) requested files by user k up to slot t. rk λk bJ,I Tslot h) Γ (h H φh time average delivery rate equal to lim inf t→∞ kt in files/slot. mean of the arrival process. length of codeword intended to users I from applying coded caching for user in J. number of channel use per slot. the capacity region for a fixed channel state h . the set of all possible channel states. the probability that the channel state at slot t is h ∈ H. D (t) are in direct analogue to the converge error vs step size of the subgradient method in convex optimization. Sketch of proof. For proving the Theorem, we use the Lyapunov function   K X X 1 1 2  L(t) =  Uk2 (t) + Sk2 (t) + Q (t) 2 I 2 F K k=1 I∈2 and specifically the related drift-plus-penalty quantity, defined E {L(t + 1) − L(t)|S(t), Q(t), U(t)} − nP as: o K VE The proposed k=1 g(γk (t))|S(t), Q(t), U(t) . algorithm is such that it minimizes (a bound on) this quantity. The main idea is to use this fact in order to compare the evolution of the drift-plus-penalty under our policy and two ”static” policies, that is policies that take random actions (admissions, demand combinations and wireless transmissions), drawn from a specific distribution, based only on the channel realizations (and knowledge of the channel statistics). We can prove from Theorem 4 that these policies can attain every feasible delivery rate. The first static policy is one such that it achieves the stability of the system for an arrival rate vector a 0 such that a 0 + δ ∈ ∂ΛCC . Comparing with our policy, we deduce strong stability of all queues and the bounds on the queue lengths by using a Foster-Lyapunov type of criterion. In order to prove near-optimality, we consider a static P policy that admits file requests at rates a ∗ = arg maxa k gk (ak ) and keeps the queues stable in a weaker sense (since the arrival rate is now in the boundary ΛCC ). By comparing the drift-plus-penalty quantities and using telescopic sums and Jensen’s inequality on the time average utilities, we obtain the near-optimality of our proposed policy. The full proof, as well as the expressions for the constants B and 0 , are in Section IX-D of the Appendix (equations (35) and (41) - (42), respectively). VI. DYNAMIC F ILE R EQUESTS In this Section, we extend our algorithm to the case where there is no infinite amount of demands for each user, rather each user requests a finite number of files at slot t. Let Ak (t) be the number of files requested by user k at the beginning of slot t. We assume it is an i.i.d. random process with mean λk and such that Ak (t) ≤ Amax almost surely. 4 In this case, the alpha fair delivery problem is to find a delivery rate r that solves Maximize K X gk (rk ) k=1 s.t. r ∈ ΛCC rk ≤ λk , ∀k ∈ {1, ..., K}, where the additional constraints rk ≤ λk denote that a user cannot receive more files than the ones actually requested. The fact that file demands are not infinite and come as a stochastic process is dealt with by introducing one ”reservoir queue” per user, Lk (t), which stores the file demands that have not been admitted, and an additional control decision on how many demands to reject permanently from the system, dk (t). At slot t, no more demands then the ones that arrived at the beginning of this slot and the ones waiting in the reservoir queues can be admitted, therefore the admission control must have the additional constraint ak (t) ≤ Ak (t) + Lk (t), ∀k, t, and a similar restriction holds for the number of rejected files from the system, dk (t). The reservoir queues then evolve as Lk (t + 1) = Lk (t) + Ak (t) − ak (t) − dk (t). The above modification with the reservoir queues has only an impact that further constrains the admission control of files to the system. The queuing system remains the same as described in Section V, with the user queues S (t), the codeword queues Q (t) and the virtual queues U (t). Similar to the case with infinite demands we can restrict ourselves to policies that are functions only of the system state at S (t), Q (t), L (t), A (t), h (t), U (t)} without loss time slot t, {S 4 The assumptions can be relaxed to arrivals being ergodic Markov chains with finite second moment under the stationary distribution 10 of optimality. Furthermore, we can show that the alpha fair optimization problem equivalent to the problem of controlling the admission rate. That is, we want to find a policy π such that aπ = arg max K X gk (ak ) • Standard coded caching: We use decentralized coded caching among all K users. For the delivery, nonopportunistic TDMA transmission is used. The server sends a sequence of codewords {VJ } at the worst transmission rate. The number of packets to be multicast in order to satisfy one demand for each user is given by [1] k=1 S (t), Q (t), U (t)) are strongly stable s.t. the queues (S ak (t) ≤ min[amax,k , Lk (t) + Ak (t)], ∀t ≥ 0, ∀k The rules for scheduling, codeword generation, virtual queue arrivals and queue updating remain the same as in the case of infinite demands in subsections C and D of Sec. V. The only difference is that there are multiple possibilities for the admission control; see [7] and Chapter 5 of [31] for more details. Here we propose that at each slot t, any demand that is not admitted get rejected (i.e. the reservoir queues hold no demands), the admission rule is n o 1 K (1 − m) 1 − (1 − m) . (30) m Thus the average delivery rate (in file per slot) is symmetric, and given as the following   Tslot rk = E log(1 + P min hi ) . (31) Ttot (K, m)F i∈{1,...,K} Ttot (K, m) = We consider a system with normalized memory of m = 0.6, power constraint P = 10dB, file size F = 103 bits and number of channel uses per slot Tslot = 102 . The channel coefficient hk (t) follows an exponential distribution with mean βk . aπk (t) = Ak (t)1{Uk (t)≥Sk (t)} , (27) We compare the three algorithms for the cases where the objective of the system is sum rate maximization (α = 0) and and the constants are set as γk,max , σmax ≥ Amax . Using the proportional fairness (α = 1) in two different scenarios. The same ideas employed in the performance analysis of the case results depicted in Fig. 4 consider a deterministic channel with with infinite demands and the ones employed in [7], we can two classes of users of K/2 each: strong users with βk = 1 and prove that the O(1/V ) − O(V ) utility-queue length tradeoff weak users with β = 0.2. For Fig. 5, we consider a symmetric k of Theorem 6 holds for the case of dynamic arrivals as well. block fading channel with βk = 1 for all users. Finally, for Fig. 6, we consider a system with block fading channel and VII. N UMERICAL E XAMPLES two classes of users: K/2 strong users with βk = 1 and K/2 In this section, we compare our proposed delivery scheme weak users with βk = 0.2. with two other schemes described below, all building on the It is notable that our proposed scheme outperforms the decentralized cache placement described in (2) and (3). unicast opportunistic scheme, which maximizes the sum rate if only private information packets are to be conveyed, and • Our proposed scheme: We apply Algorithm 1 for t = 105 slots. Using the scheduler (25), we calculate µJ (τ ) standard coded caching which transmit multicast packets with denoting the rate allocated to a user set J at slot τ ≤ t. the worst user channel quality. In Fig. 4 for the deterministic channel scenario with α = 0, As defined in (13), the long-term average rate of user k the unicast opportunistic scheme serves only the K/2 strong measured in file/slot is given by log(1+P ) Pt P users in TDMA at a constant rate equal to Tslot = F 1−m Tslot limt→∞ 1t τ =1 J:k∈J µJ (τ ) 0.865 in file/slot. For the standard coded caching, the sum rate rk = . (28) (1 − m)F increases linearly with the number of users. This is because the Notice that the numerator corresponds to the average multicast rate is constant over the deterministic channel and number of useful bits received over a slot by user k and the behavior of Ttot (K, m) is almost constant with m = 0.6, the denominator (1 − m)F corresponds to the number of which makes the per-user rate almost constant. For the symmetric fading channel in Fig 5, the performance bits necessary to recover one file. of unicast opportunistic and that of standard coded caching • Unicast opportunistic scheduling: For any request, the schemes are limited due to the lack of global caching gain server sends the remaining (1 − m)F bits to the correand vanishing multicast rate, respectively. sponding user without combining any files. Here we only Finally, for the case of both fading and differences in the exploit the local caching gain. In each slot the transmitter mean SNRs, we can see from Fig. 6 that, again, our proposed sends with full power to the following user scheme outperforms the unicast opportunistic scheduling and log (1 + hk (t)P ) standard coded caching both in terms of sum rate and in terms , k ∗ (t) = arg max k Tk (t)α of proportional fair utility. P µ (τ ) In all scenarios, the relative merit of our scheme increases k ≤t−1 where Tk (t) = 1≤τ(t−1) is the empirical average as the number of users grows. This can be attributed to the fact rate for user k up to slot t. The resulting long-term that our scheme can exploit any available multicast opportuniaverage rate of user k measured in file/slot is given by ties. Our result here implies that, in realistic wireless systems, Pt Tslot limt→∞ 1t τ =1 log(1 + P hk (τ ))1{k = k ∗ (τ )} coded caching can indeed provide a significant throughput rk = . increase when an appropriate joint design of routing and (1 − m)F opportunistic transmission is used. Regarding the proportional (29) 11 (a) Rate (α = 0) vs K. (b) Proportional fair utility (α = 1) vs K. Fig. 4. Deterministic channel with different SNR. (a) Rate (α = 0) vs K. (b) Proportional fair utility (α = 1) vs K. Fig. 5. Symmetric fading channel. (a) Rate (α = 0) vs K. Fig. 6. Fading channels with two groups of users, each with different average SNR. (b) Proportional fair utility (α = 1) vs K. 12 fair objective, we can see that the average sum utility increases with a system dimension for three schemes although our proposed scheme provides a gain compared to the two others. VIII. C ONCLUSIONS In this paper, we studied coded caching over wireless fading channels in order to address its limitation governed by the user with the worst fading state. By formulating an alpha-fair optimization problem with respect to the long-term average delivery rates, we proposed a novel queueing structure that allowed us to obtain an optimal algorithm for joint file admission control, codeword construction and wireless transmissions. The main conclusion is that, by appropriately combining the multicast opportunities and the opportunism due to channel fading, coded caching can lead to significant gains in wireless systems with fading. Low-complexity algorithms which retain the benefits of our approach as well as a delay-constrained delivery scheme, are left as interesting topics of future investigation. IX. A PPENDIX : P ROOFS A. Proof of Theorem 1 Let MJ be the message for all the users in J ⊆ [K] and of size 2nRJ . We first show the converse. It follows that the set of 2K − 1 independent messages {MJ : J ⊆ [K], J 6= ∅} can be partitioned as K [ {MJ : k ∈ J ⊆ [k]}. (32) k=1 We can now define K independent mega-messages M̃k := P {MJ : k ∈ J ⊆ [k]} with rate R̃k := J: k∈J⊆[k] RJ . Note that each mega-message k must be decoded at least by user k reliably. Thus, the K-tuple (R̃1 , . . . , R̃K ) must lie inside the private-message capacity region of the K-user BC. Since it is a degraded BC, the capacity region is known [5], and we have Pk 1 + hk j=1 pj (33) R̃k ≤ log Pk−1 , k = 2, . . . , K, 1 + hk j=1 pj PK for some pj ≥ 0 such that j=1 pj ≤ P . This establishes the converse. To show the achievability, it is enough to use rate-splitting. Specifically, the transmitter first assembles the original messages into K mega-messages, and then applied the standard K-level superposition coding [5] putting the (k − 1)-th signal on top of the k-th signal. The k-th signal has average power pk , k ∈ [K]. At the receivers’ side, if the rate of the megamessages are inside the private-message capacity region of the K-user BC, i.e., the K-tuple (R̃1 , . . . , R̃K ) satisfies (33), then each user k can decode the mega-message k. Since the channel is degraded, the users 1 to k − 1 can also decode the megamessage k and extract its own message. Specifically, each user j can obtain MJ (if J 3 j), from the mega-message M̃k when k ∈ J ⊆ [k]. This completes the achievability proof. B. Static policies An important concept for characterizing the feasibility region and proving optimality of our proposed policy is the one we will refer to here as ”static policies”. The concept is that decisions taken according to these policies depend only on the channel state realization (i.e. the uncontrollable part of the system) as per the following definition: Definition 4 (Static Policy). Any policy that selects the a(t), σ(t), µ(t)} according to a probability control variables {a distribution that depends only on the channel state h (t) will be called a static policy. It is clear from the definition that all static policies belong to the set of admissible policies for our setting. An important case is where actually admission control a (t) and codeword routing σ(t) are decided at random and independently of everything and transmissions µ(t) are decided at by a distribution that depends only on the channel state realization of the slot: It can be shown using standard arguments in stochastic network optimization (see for example [6], [7], [26], [31]) that the optimal long term file delivery vector and any file delivery vector in the stability region of the queueing system can be achieved by such static policies, as formalized by the following Lemmas: Lemma 7 (Static Optimal Policy). Define a policy π ∗ ∈ Π CC that in each slot where the channel states are h works as follows: (i) it pulls random user demands with mean ā∗k , and it gives the virtual queues arrivals with mean γ k = a∗k as well (ii) the number of combinations for subset J is a random variable with mean σ ∗J and uniformly bounded by σmax , (iii) selects one out of K + 1 suitably defined rate vectors µl ∈ h), l = 1, .., K + 1 with probability ψl,hh . The parameters Γ (h above are selected such that they solve the following problem: max a s.t. K X gk (a∗k ) k=1 X σ ∗J ≥ a∗k , ∀k ∈ {1, .., K} J:k∈J X J:I⊆J bJ,I σ ∗J ≤ Tslot X h φh K+1 X h), ∀I ⊆ {1, 2, ..., K} ψl,hh µlI (h l=1 Then, π ∗ results in the optimal delivery rate vector (when all possible policies are restricted to set Π CC ). Lemma 8 (Static Policy for the δ− interior of Γ CC ). Define a policy π δ ∈ Π CC that in each slot where the channel states are h works as follows: (i) it pulls random user demands with a +δ) ∈ Γ CC , and gives the virtual queues mean aδk such that (a random arrivals with mean γ k ≤ ak + 0 for some 0 > 0 (ii) the number of combinations for subset J is a random variable with mean σ δJ and uniformly bounded by σmax , (iii) selects h), l = one out of K + 1 suitably defined rate vectors µl ∈ Γ (h δ 1, .., K + 1 with probability ψl,h . The parameters above are h 13 P h) is the capacity assumed to be irrevocable and h ∈H φh Γ (h region of the wireless channel, the above implies that there is not enough wireless capacity to satisfy a long term file delivery / ΛCC , finishing the proof. 5 rate vector of a . Therefore, a ∈ selected such that: X σ δ J ≥  + aδk , ∀k ∈ {1, .., K} J:k∈J X bJ,I σ δJ ≤  + Tslot J:I⊆J X h φh K+1 X δ l h), ∀I ∈ 2K ψl,h h µI (h l=1 for some appropriate  < δ. Then, the system under π δ has mean incoming rates of a δ and is strongly stable. C. Proof of Lemma 4 We prove the Lemma in two parts:
(i) first we
prove that c c Γ CC ⊆ ΛCC and (ii) then that Γ CC ⊆ ΛCC . For the first part, we show that if a ∈ Int(Γ CC ) then also λ ∈ ΛCC , that is the long term file delivery rate vector observed by the users as per (13) is r = a . Denote Ak (t) the number of files that have been admitted to the system for user k up to slot t. Also, note that due to our restriction on the class of policies Π CC and our assumption about long enough blocklengths, there are no errors in decoding the files, therefore the number of files correctly decoded for user k till slot t is Dk (t) . From Lemma 8 it follows that there exists a static policy π RAN D , the probabilities of which depending only on the channel state realization at each slot, for which the system is strongly stable. Since the channels are i.i.d. random with a finite state space and queues are measured in files and bits, the system now evolves as a discrete time Markov chain (S(t), Q(t), H(t)), which can be checked that is aperiodic, irreducible and with a single communicating class. In that case, strong stability means that the Markov chain is ergodic with finite mean. Further, this means that the system reaches to the set of states where all queues are zero infinitely often. Let T [n] be the number of timeslots between the n−th and (n + 1)−th visit to this set (we make the convention that T [0] is the time slot that this state is reached for the first time). In addition, let Ãk [n], D̃k [n] be the number of demands that arrived and were delivered in this frame, respectively. Then, since within this frame the queues start and end empty, we have Ãk [n] = D̃k [n], ∀n, ∀k. In addition since the Markov chain is ergodic, PN Ãk [n] A(t) ak = lim = lim Pn=0 N t→∞ N →∞ t n=0 T [n] and PN D̃k [n] D(t) rk = lim = lim Pn=0 N t→∞ N →∞ t n=0 T [n] Combining the three expressions, r = a thus the result follows. We now proceed to show the second part, that is given any arrival rate vector a that is not in the stability region of the queueuing system we cannot have a long term file delivery rate vector r = a . Indeed, sinceP a ∈ / Γ CC , for any possible h) there will be σ satisfying (18), for every µ ∈ h ∈H φh Γ (h some subset(s) of users for which the corresponding inequality (19) is violated. Since codeword generation decisions are D. Proof ot Theorem 6 We first look at static policies, which take random decisions based only on the channel realizations. We focus on two such policies: (i) one that achieves the optimal utility, as described in Lemma 7 and (ii) one that achieves (i.e. admits and stabilizes the system for that) a rate vector in the δ− interior of ΛCC (for any δ > 0), as described in Lemma 8. Then, we show that our proposed policy minimizes a bound on the drift of the quadratic Lyapunov function and compare with the two aforementioned policies: Comparison with the second policy proves strong stability of the system under our proposed policy, while comparison with the first one proves almost optimality. From Lemma 4 and Corollary 5, it suffices to prove that under the online policy the queues are strongly stable and the resulting time average admission rates maximize the desired utility function subject to minimum rate constraints. The proof of the performance of our proposed policy is based on applying Lyapunov optimization theory [26] with the following as Lyapunov function (where we have defined Z(t) = (S(t), Q(t), U(t)) to shorten the notation)   K X Q2 (t) 1 X 2 I . Z ) = L(S, Q, U) = Uk (t) + Sk2 (t) + L(Z 2 F2 K I∈2 k=1 We then define the drift of the aforementioned Lyapunov function as ∆L(Z) = E {L(Z(t + 1)) − L(Z(t))|Z(t) = Z} , where the expectation is over the channel distribution and possible randomizations of the control policy. Using the queue evolution equations (16), (17), (21) and the fact that ([x]+ )2 ≤ x2 , we have ∆L(Z(t)) ≤B    X QI (t)  X + E bI,J σJ (t) − Tslot µI (t) Z(t) 2   F J:I⊆J I∈2K ( ) K X X + Sk (t)E ak (t) − σI (t) Z(t) k=1 + X I:k∈I Uk (t)E {γk (t) − ak (t)|Z(t)} , (34) k=1K 5 We would also need to check the boundary of Γ CC . Note, however, that by similar arguments we can show that for each vector on ∂Γ CC we need to achieve a rate vector on the boundary of the capacity region of the wireless channel. Since, as mentioned in the main text, we do not consider boundaries in this work, we can discard these points. 14 some δ > 0, we get that there exist , 0 > 0 such that (the superscript π denotes the quantities under our proposed policy) where B= K X k=1  !2  1 2 γk,max + 2 X σmax ∆Lπ (Z(t)) ≤ B + V  I:k∈I 1 X X 2 + (σmax bI,J ) 2F 2 I∈2K J:I⊆J o 2 X X n Tslot 2 + E (log (1 + P h (t))) . k 2 2F 2 K I∈2 K X E {gk (aπk (t))} − V K X gk (aδk ) k=1  k=1 K X X QJ (t)  −  Sk (t) + F2 K J∈2 k=1 (35) − 0 k∈I Note that B is a finite constant that depends only on P the parameters of the system. Adding the quantity K −V k=1 E {gk (γk (t))|Z(t)} to both hands of (34) and rearranging the right hand side, we have the drift-plus-penalty expression K X Uk (t) (38) k=1 Since ak (t) ≤ γmax,k ∀t, it follows that gk (aπk ) < gk (γmax,k ). In addition, gk (x) ≥ 0, ∀x ≥ 0 therefore ∆Lπ (Z(t)) ≤ B + V K X E {gk (γmax,k )} k=1 ∆L(Z(t)) − V K X  E {gk (γk (t))|Z(t)} ≤ −  k=1 B+ K X k=1 E {−V gk (γk (t)) + γk (t)Uk (t)|Z(t)} − 0 k=1 + X   E {σJ (t)|Z(t)}  X X QI (t) bI,J − Sk (t) 2 F I:I⊆J J∈2K + K X k:k∈J (Sk (t) − Uk (t)) E {ak (t)|Z(t)} k=1 − K X X QJ (t) Tslot E {µJ (t)|Z(t)} F2 K (36) J∈2 Now observe that the proposed scheme π minimizes the right hand side of (36) given any channel state h (t) (and hence in expectation over the channel state distributions). Therefore, for every vectors a ∈ [1,P γmax ]K , γ ∈ [1, γmax ]K , σ ∈ M h) that denote time Conv({0, .., σmax } ), µ ∈ h ∈H φh Γ (h averages of the control variables achievable by any static (i.e. depending only on the channel state realizations) randomized policies it holds that K X  X QJ (t)  Sk (t) + F2 K Uk (t) J∈2 (39) k=1 Using the the Foster-Lyapunov criterion, the above inequality implies that the system Z(t)(S(t), Q(t), U(t)) under our proposed policy π has a unique stationary probability distribution, under which the mean queue lengths are finite 6 . Moreover,   T −1  X K  X X 1 QJ (t) lim sup E + (S (t) + U (t)) k k  K F2  T →∞ T t=0 k=1 J∈2 PK B + V k=1 gk (γmax,k ) ≤ .  (40) Therefore the queues are strongly stable under our proposed policy. In order to prove the part of Theorem 6 regarding the guaranteed bound on the average queue lengths, we first note that the above inequality holds for every  > 0 and define 0 as 0 = argmax  (41) >0 K X ∆Lπ (Z(t)) − V B−V + K X k=1 K X gk (γ k ) + k=1 K X ! X Sk (t) ak − σJ J:k∈J  X QJ (t) J Uk (t) (γ k − ak ) k=1 k=1 + s.t. 1 ∈ ΛCC . E {gk (γkπ (t))} ≤ F2  X  bJ,I σ I − Tslot µJ  (37) I:J⊆I We will use (37) to compare our policy with the specific static policies defined in Lemmas 7, 8. Proof of strong stability: Replacing the time averages we get from the static stabilizing policy π δ of Lemma 8 for (42) Following the same arguments as in Section IV of [7], we can show that the Right Hand Side of (40) is bounded from below by PK B + V k=1 gk (γmax,k ) , 0 therefore proving the requested bound on the long-term average queue lengths. We now proceed to proving the near-optimality of our proposed policy. Proof of near optimal utility: Here we compare π with the static optimal policy π ∗ from Lemma 7. Since π ∗ takes decisions irrespectively of the queue lengths, we can replace 6 For the utility-related virtual queues, note that if g 0 (0) < ∞, then k Uk (t) < V gk0 (0) + γk,max , i.e. their length is deterministically bounded 15 quantities a , σ, µ on (37) with the time averages corresponding to π ∗ , i.e. a ∗ , σ ∗ , µ∗ . From the inequalities in Lemma 7 we have V K X E {gk (γkπ (t))} ≥ V k=1 K X gk (a∗k ) − B + ∆Lπ (Z(t)) k=1 Taking expectations over Z(t) for both sides and summing the inequalities for t = 0, 1, .., T − 1 and dividing by V T we get T −1 K K X B 1 XX E {Lπ (Z(0))} E {gk (γkπ (t))} ≥ gk (a∗k ) − − T t=1 V VT k=1 k=1 + E {Lπ (Z(T ))} VT Assuming E {Lπ (Z(0))} < ∞ (this assumption is standard in this line of work, for example it holds if the system starts empty), since E{Lπ (Z(T ))} > 0, ∀T > 0, taking the limit as T goes to infinity gives T −1 K K X B 1 XX E {gk (γkπ (t))} ≥ gk (a∗k ) − T →∞ T V t=1 lim k=1 k=1 In addition, since gk (x) are concave, Jensen’s inequality implies ! K K T X X 1X π π gk (γ k ) = gk lim E{γk (t)} T →∞ T t=0 k=1 k=1 T −1 K 1 XX E {gk (γkπ (t))} T →∞ T t=1 ≥ lim k=1 ≥ K X gk (a∗k ) − k=1 B . V Finally, since the virtual queues Uk (t) are strongly stable, it holds aπk > γ πk . We then have K X k=1 gk (aπk ) > K X k=1 gk (γ πk ) ≥ K X k=1 gk (a∗k ) − B , V which proves the near optimality of our proposed policy π. R EFERENCES [1] M. A. Maddah-Ali and U. Niesen, “Decentralized coded caching attains order-optimal memory-rate tradeoff,” IEEE/ACM Transactions on Networking, vol. 23, no. 4, pp. 1029–1040, Aug. 2015. [2] CISCO, “Visual networking index: Forecast and methodology 20152020,” white paper, Jun. 2016. [3] M. A. Maddah-Ali and U. Niesen, “Fundamental limits of caching,” IEEE Transactions on Information Theory, vol. 60, no. 5, pp. 2856– 2867, Mar. 2014. [4] G. Paschos, E. Baştuğ, I. Land, G. Caire, and M. Debbah, “Wireless caching: Technical misconceptions and business barriers,” IEEE Communications Magazine, vol. 54, no. 8, pp. 16–22, Aug. 2016. [5] A. El Gamal and Y.-H. Kim, Network information theory. Cambridge University Press, 2011. [6] A. L. Stolyar, “On the asymptotic optimality of the gradient scheduling algorithm for multiuser throughput allocation,” Operations research, vol. 53, no. 1, pp. 12–25, Jan. 2005. [7] M. J. Neely, E. Modiano, and C.-P. Li, “Fairness and optimal stochastic control for heterogeneous networks,” IEEE/ACM Transactions on Networking, vol. 16, no. 2, pp. 396–409, Apr. 2008. [8] R. Knopp and P. A. Humblet, “Information capacity and power control in single-cell multiuser communications,” in IEEE International Conference on Communications (ICC ’96), Seattle, WA, USA, 1995, pp. 331–335. [9] J. Mo and J. Walrand, “Fair end-to-end window-based congestion control,” IEEE/ACM Transactions on Networking, vol. 8, no. 5, pp. 556– 567, Oct. 2000. [10] H. Shirani-Mehr, G. Caire, and M. J. Neely, “Mimo downlink scheduling with non-perfect channel state knowledge,” IEEE Transactions on Communications, vol. 58, no. 7, pp. 2055–2066, Jul. 2010. [11] G. Caire, R. R. Muller, and R. Knopp, “Hard fairness versus proportional fairness in wireless communications: The single-cell case,” IEEE Transactions on Information Theory, vol. 53, no. 4, pp. 1366–1385, Apr. 2007. [12] M. Ji, A. M. Tulino, J. Llorca, and G. Caire, “Order-optimal rate of caching and coded multicasting with random demands,” IEEE Transactions on Information Theory, vol. 63, no. 6, pp. 3923–3949, Jun. 2017. [13] U. Niesen and M. A. Maddah-Ali, “Coded caching with nonuniform demands,” IEEE Transactions on Information Theory, vol. 63, no. 2, pp. 1146–1158, Feb. 2017. [14] M. Ji, G. Caire, and A. F. Molisch, “Fundamental limits of distributed caching in d2d wireless networks,” in 2013 IEEE Information Theory Workshop (ITW), Sevilla, Spain, 2013, pp. 1–5. [15] N. Karamchandani, U. Niesen, M. A. Maddah-Ali, and S. N. Diggavi, “Hierarchical coded caching,” IEEE Transactions on Information Theory, vol. 62, no. 6, pp. 3212–3229, Jun. 2016. [16] S. P. Shariatpanahi, S. A. Motahari, and B. H. Khalaj, “Multi-server coded caching,” IEEE Transactions on Information Theory, vol. 62, no. 12, pp. 7253–7271, Dec. 2016. [17] W. Huang, S. Wang, L. Ding, F. Yang, and W. Zhang, “The performance analysis of coded cache in wireless fading channel,” arXiv preprint arXiv:1504.01452, 2015. [18] J. Zhang and P. Elia, “Wireless coded caching: A topological perspective,” in 2018 IEEE International Symposium on Information Theory (ISIT), Aachen, Germany, 2017, pp. 401–405. [19] ——, “Fundamental limits of cache-aided wireless BC: Interplay of coded-caching and CSIT feedback,” IEEE Transactions on Information Theory, vol. 63, no. 5, pp. 3142–3160, May 2017. [20] S. S. Bidokhti, M. Wigger, and R. Timo, “Noisy broadcast networks with receiver caching,” arXiv preprint arXiv:1605.02317, 2016. [21] K.-H. Ngo, S. Yang, and M. Kobayashi, “Cache-aided content delivery in mimo channels,” in 54th Annual Allerton Conference on Communication, Control, and Computing (Allerton), Monticello, IL, USA, 2016. [22] S. P. Shariatpanahi, G. Caire, and B. H. Khalaj, “Multi-antenna coded caching,” arXiv preprint arXiv:1701.02979, 2017. [23] R. Combes, A. Ghorbel, M. Kobayashi, and S. Yang, “Utility optimal scheduling for coded caching in general topologies,” arXiv preprint arXiv:1801.02594, 2018. [24] R. Pedarsani, M. A. Maddah-Ali, and U. Niesen, “Online coded caching,” IEEE/ACM Transactions on Networking, vol. 24, no. 2, pp. 836–845, Apr. 2016. [25] U. Niesen and M. A. Maddah-Ali, “Coded caching for delay-sensitive content,” in 2015 IEEE International Conference on Communications (ICC), London, UK, 2015, pp. 5559–5564. [26] M. J. Neely, “Stochastic network optimization with application to communication and queueing systems,” Synthesis Lectures on Communication Networks, vol. 3, no. 1, pp. 1–211, 2010. [27] K. Seong, R. Narasimhan, and J. M. Cioffi, “Queue proportional scheduling via geometric programming in fading broadcast channels,” IEEE Journal on Selected Areas in Communications, vol. 24, no. 8, pp. 1593–1602, Aug. 2006. [28] A. Eryilmaz, R. Srikant, and J. R. Perkins, “Throughput-optimal scheduling for broadcast channels,” in Proc. of SPIE, vol. 4531, 2001, pp. 70–78. [29] D. N. Tse, “Optimal power allocation over parallel gaussian broadcast channels,” 2001. [30] F. Kelly, “Charging and rate control for elastic traffic,” European transactions on Telecommunications, vol. 8, no. 1, pp. 33–37, Jan. 1997. [31] L. Georgiadis, M. J. Neely, L. Tassiulas et al., “Resource allocation and cross-layer control in wireless networks,” Foundations and Trends R in Networking, vol. 1, no. 1, pp. 1–144, 2006. 

arXiv:1801.04369v1 [math.ST] 13 Jan 2018 Is profile likelihood a true likelihood? An argument in favor O.J. Maclaren Department of Engineering Science, University of Auckland January 16, 2018 Abstract Profile likelihood is the key tool for dealing with nuisance parameters in likelihood theory. It is often asserted, however, that profile likelihood is not a true likelihood. One implication is that likelihood theory lacks the generality of e.g. Bayesian inference, wherein marginalization is the universal tool for dealing with nuisance parameters. Here we argue that profile likelihood has as much claim to being a true likelihood as a marginal probability has to being a true probability distribution. The crucial point we argue is that a likelihood function is naturally interpreted as a maxitive possibility measure: given this, the associated theory of integration with respect to maxitive measures delivers profile likelihood as the direct analogue of marginal probability in additive measure theory. Thus, given a background likelihood function, we argue that profiling over the likelihood function is as natural (or as unnatural, as the case may be) as marginalizing over a background probability measure. Keywords: Estimation; Inference; Profile Likelihood; Marginalization; Nuisance Parameters; Idempotent Integration; Maxitive Measure Theory 1 1 Introduction Consider the opening sentence from the entry on profile likelihood in the Encyclopedia of Biostatistics (Aitkin 2005): The profile likelihood is not a likelihood, but a likelihood maximized over nuisance parameters given the values of the parameters of interest. Numerous similar assertions that profile likelihood is not a ‘true’ likelihood may be found throughout the literature and various textbooks, and is apparently the accepted viewpoint of the statistical community. Importantly, this includes the ‘pure’ likelihood literature, which generally accepts a lack of systematic methods for dealing with nuisance parameters, while still recommending profile likelihood as the most general, albeit ‘ad-hoc’, solution (see e.g. Royall 1997, Rohde 2014, Edwards 1992, Pawitan 2001). Similarly, recent monographs on characterizing statistical evidence presents favorable opinions of the likelihood approach but criticize the lack of general methods for dealing with nuisance parameters (Aitkin 2010, Evans 2015). The various justifications given, however, appear to the present author to rather vague and unconvincing. For example, suppose we modified the above quotation to refer to marginal probability instead of profile likelihood: A marginal probability is not a probability, but a probability distribution integrated over nuisance variables given the values of the variables of interest. The above would be a perfectly fine characterization of a marginal probability if the “not a probability, but” part was dropped, i.e. A marginal probability is a probability distribution integrated over nuisance variables given the values of the variables of interest. Simply put: the fact that a marginal probability is obtained by integrating over a ‘background’ probability distribution does not prevent the marginal probability from being a true probability. The crucial observation in the case of marginal probability is that integration over variables takes probability distributions to probability distributions. 2 The purpose of the present article is to point out that there is an appropriate notion of integration over variables that takes likelihood functions to likelihood functions via maximization. This notion of integration is based on the idea of idempotent analysis, wherein one replaces a standard algebraic operation such as addition in a given mathematical theory with another basic algebraic operation, defining a form of ‘idempotent addition’, to obtain a new analogous, self-consistent theory (Maslov 1992, Kolokoltsov & Maslov 1997). In this case one simply replaces the usual ‘addition’ operations, including the usual (Lebesgue) integration, with ‘maximization’ operations, including taking supremums, to obtain a new, ‘idempotent probability theory’. Maximization in this context is understood algebraically as an idempotent addition operation, hence the terminology. While perhaps somewhat exotic at first sight, this idea finds direct applications in e.g. large deviation theory (Puhalskii 2001) and, most relevantly, possibility theory, fuzzy set theory and pure-likelihood-based decision theory (Dubois et al. 1997, Cattaneo 2013, 2017). 2 Likelihood as a possibility measure Though apparently not well known in the statistical literature, likelihood theory is known in the wider literature on uncertainty quantification to have a natural correspondence to possibility theory rather than to probability theory (Dubois et al. 1997, Cattaneo 2013, 2017). This has perhaps been obscured by the usefulness of likelihood methods as tools in probabilistic statistical inference. It is not our intention to review this wider literature in detail here (see e.g. Dubois et al. 1997, Cattaneo 2013, 2017, Augustin et al. 2014, Halpern 2017, for more), but to simply point out the implications of this correspondence. In particular, likelihood theory interpreted as a possibilistic, rather than probabilistic theory can be summarized as: Probability theory with addition replaced by maximization. As indicated above, this is sometimes known as, for example, ‘idempotent measure theory’, ‘maxitive measure theory or ‘possibility’ theory, among other names (see e.g. Dubois et al. 1997, Cattaneo 2013, 2017, Augustin et al. 2014, Halpern 2017, Maslov 1992, 3 Kolokoltsov & Maslov 1997, Puhalskii 2001, for more). This correspondence perhaps explains the preponderance of maximization methods in likelihood theory, including the methods of maximum likelihood and profile likelihood. The most important consequence of this perspective is that the usual Lebesgue integration with respect to an additive measure, as in probability theory, becomes, in likelihood/possibility theory, a different type of integration, defined with respect to a maxitive measure. Again, the key point is simply that addition operations (including summation and integration) are replaced by maximization operations (or taking supremums in general). For completeness, we contrast the key axioms of possibility theory with those of probability theory. Given a set of possibilities of Ω, assumed to be discrete for the moment for simplicity, and for two discrete sets of possibilities A, B ⊆ Ω the key axioms of elementary possibility theory are (Halpern 2017): poss(∅) = 0 (1) poss(Ω) = 1 poss(A ∪ B) = max{poss(A), poss(B)} which can be contrasted with those of elementary probability theory: prob(∅) = 0 (2) prob(Ω) = 1 prob(A ∪ B) = sum{prob(A), prob(B)} where A and B are required to be disjoint in the probabilistic case, but this is not strictly required in the possibilistic case. Given a ‘background’ or ‘starting’ likelihood measure, likelihood theory can be developed as a self-contained theory of possibility, where derived distributions are manipulated according to the first set of axioms above. This is entirely analogous to developing probability theory from a background measure, with derived distributions manipulated according to the second set of axioms. As our intention is to consider methods for obtaining derived distributions by ‘eliminating’ nuisance parameters, we need not consider here where the starting measure comes from (but see the Discussion). 4 To make the correspondences of interest clear in what follows, we first present probabilistic marginalization as a special case of a pushforward measure or, equivalently, as a special case of a general (not necessarily 1-1) change of variables. We then consider the possibilistic analogues. 3 Pushforward probability measures and the delta function method for general changes of variable Given a probability measure µ over a random variable x ∈ Rn with associated density ρ, define the new random variable t = T (x) where T : Rn → Rm . This variable is distributed according to the pushforward measure T ⋆ µ, i.e. t ∼ T ⋆ µ. The density of t, here denoted by q = T ⋆ ρ, is conveniently calculated via the delta function method which is valid for arbitrary changes of variables (not necessarily 1-1): q(t) = [T ⋆ ρ](t) = Z δ(t − T (x))ρ(x)dx (3) As a side point, we note that this method of carrying out arbitrary transformations of variables is standard in statistical physics (see e.g. Van Kampen 1992), but is apparently less common in statistics (see the articles Au & Tam 1999, Khuri 2004, aimed at highlighting this method to the statistical community). 3.1 Marginalization via the delta function method The above means that we can interpret marginalization to a component x1 , say, as a special case of a (non-1-1) deterministic change of variables via: ρ(x1 ) = Z δ(x1 − projX1 (x))ρ(x)dx (4) where projX1 (x) is simply the projection of x to its first coordinate. Thus marginalization can be thought of as the pushforward under the projection operator and as a special case of a general (not necessarily 1-1) change of variables t = T (x). 5 4 Profile likelihood as marginal possibility and an extension to general changes of variable As we have repeatedly stressed above, likelihood theory interpreted as a possibilistic, and hence maxitive, measure theory simply means that addition operations such as the usual Lebesgue integration are replaced by maximization operations such as taking the supremum. Consider first then the analogue of a marginal probability density, which we will call a marginal possibility distribution and denote by Lp . Starting from a ‘background’ likelihood measure L(x) we ‘marginalize’ in the analogous manner to before: Lp (x1 ) = sup{δ(x1 − projX1 (x))L(x)} = sup{x|projX 1 (x)=x1 } {L(x)} (5) This is again simply the pushforward under the projection operator, but here under a different type of ‘integration’ - i.e. the operation of taking a supremum. Of course, this is just the usual profile likelihood for x1 . As above, we need not be restricted to marginal possibility distributions: we can consider arbitrary functions of the parameter t = T (x). This leads to an analogous pushforward operation of L(x) to Lp (t) that we denote by ⋆p : Lp (t) = [T ⋆p L](t) = sup{δ(t − T (x))L(x)} = sup{x|T (x)=t} {L(x)} (6) which again corresponds to the usual definition of profile likelihood. 5 5.1 Discussion Objections to profile likelihood As discussed, it is frequently asserted that profile likelihood is not a true likelihood (Aitkin 2005, Royall 1997, Pawitan 2001, Rohde 2014, Evans 2015). Common reasons include: that it is obtained from a likelihood via maximization (Aitkin 2005), that it is not based directly on observable quantities (Royall 1997, Pawitan 2001, Rohde 2014) and that it lacks particular repeated sampling properties (Royall 1997, Cox & Barndorff-Nielsen 1994). 6 None of the above objections appear to the present author to apply to the following: given a starting or ‘background’ likelihood function, profile likelihood satisfies the axioms of possibility theory, in which the basic additivity axiom of probability theory is replaced by a maxitivity axiom. Profile likelihood is simply the natural possibilistic counterpart to marginal probability, where additive integration is replaced by a maxitive analogue. We thus argue that, if marginal probability is a ‘true’ probability, then profile likelihood should likewise be considered a ‘true’ likelihood, at least when likelihood theory is interpreted in a possibilistic manner. 5.2 Fixed data Regarding the second two objections mentioned above: observable quantities and repeated sampling properties, it is important to note that the given data must be held fixed to give a consistent background likelihood over which to profile. Given fixed data one has a fixed possibility measure and thus can consider ‘marginal’ - i.e. profile - likelihoods. In contrast, repeated sampling will produce a distribution of such possibility measures, and these may or may not have good frequentist properties. None of this is in contrast to marginal probability: changing the distribution over which we marginalize changes the resulting marginal probability. Of course, despite this caveat, profile likelihood often does have good repeated sampling properties (Royall 1997, Cox & Barndorff-Nielsen 1994) and also plays a key role in frequentist theory, though we do not discuss this further here. 5.3 Why? A natural question, perhaps, is why worry about whether profile likelihood is a true likelihood? One answer is that profile likelihood is a widely used tool but is often dismissed as ‘ad-hoc’ or lacking proper justification. This gives the impression that, for example, likelihood theory is lacking in comparison with e.g. Bayesian theory in terms of systematic methods for dealing with nuisance parameters. By understanding that profile likelihood does in fact have a systematic basis in terms of possibility theory practitioners and students can better understand and reason about a widely popular and useful tool. Understanding the connection to possibilistic as opposed to probabilistic reasoning may also help explain 7 why profile likelihood has emerged as a particularly promising method of identifiability analysis (Raue et al. 2009), where identifiability is traditionally a prerequisite for probabilistic analysis. 5.4 Ignorance The possibilistic interpretation of likelihood also helps understand the representation of ignorance. While probabilistic ignorance is not preserved under arbitrary changes of variables (e.g. non-1-1 transformations), even in the discrete case, possibilistic ignorance is in the following sense: if we take the maximum likelihood over a set of possibilities, such as {x | T (x) = t} for each t, rather than summing them, a flat ‘prior likelihood’ (Edwards 1969, 1992) over x becomes a flat prior likelihood over t. On the other hand, a flat prior probability over x in general becomes non-flat over t under non-1-1 changes of variable. Thus a profile prior likelihood has what, in many cases, may be desirable properties as a representation of prior ignorance (see the discussion in Edwards 1969, 1992, for more on likelihood and the representation of ignorance). 6 Conclusions We have argued that profile likelihood has as much claim to being a true likelihood as a marginal probability has to being a true probability distribution. In the case of marginal probability, integration over variables takes probability distributions to probability distributions, while in the case of likelihood, maximization takes likelihood functions to likelihood functions. Maximization can be considered in this context as an alternative (idempotent) notion of integration, and a likelihood function as a maxitive possibility measure. This gives a self-consistent theory of possibilistic statistical analysis with a well-defined method of treating nuisance parameters. 8 References Aitkin, M. (2005), Profile likelihood, in ‘Encyclopedia of Biostatistics’, John Wiley & Sons, Ltd. Aitkin, M. (2010), Statistical Inference: An Integrated Bayesian/Likelihood Approach, Chapman & Hall/CRC Monographs on Statistics & Applied Probability, CRC Press. Au, C. & Tam, J. (1999), ‘Transforming variables using the dirac generalized function’, Am. Stat. 53(3), 270–272. Augustin, T., Coolen, F. P. A., de Cooman, G. & Troffaes, M. C. M. (2014), Introduction to imprecise probabilities, John Wiley & Sons. Cattaneo, M. E. G. (2013), ‘Likelihood decision functions’, Electron. J. Stat. 7, 2924–2946. Cattaneo, M. E. G. V. (2017), ‘The likelihood interpretation as the foundation of fuzzy set theory’, Int. J. Approx. Reason. . Cox, D. R. & Barndorff-Nielsen, O. E. (1994), Inference and asymptotics, Chapman and Hall, London. Dubois, D., Moral, S. & Prade, H. (1997), ‘A semantics for possibility theory based on likelihoods’, J. Math. Anal. Appl. 205(2), 359–380. Edwards, A. W. F. (1969), ‘Statistical methods in scientific inference’, Nature 222(5200), 1233–1237. Edwards, A. W. F. (1992), ‘Likelihood, expanded ed’, Johns Hopkins University Press, Baltimore . Evans, M. (2015), Measuring Statistical Evidence Using Relative Belief, Chapman & Hall/CRC Monographs on Statistics & Applied Probability, CRC Press. Halpern, J. Y. (2017), Reasoning about Uncertainty, MIT Press. Khuri, A. I. (2004), ‘Applications of dirac’s delta function in statistics’, Internat. J. Math. Ed. Sci. Tech. 35(2), 185–195. 9 Kolokoltsov, V. & Maslov, V. P. (1997), Idempotent Analysis and Its Applications, Springer Science & Business Media. Maslov, V. P. (1992), Idempotent Analysis, American Mathematical Soc. Pawitan, Y. (2001), In All Likelihood: Statistical Modelling and Inference Using Likelihood, Oxford science publications, OUP Oxford. Puhalskii, A. (2001), Large Deviations and Idempotent Probability, CRC Press. Raue, A., Kreutz, C., Maiwald, T., Bachmann, J., Schilling, M., Klingmüller, U. & Timmer, J. (2009), ‘Structural and practical identifiability analysis of partially observed dynamical models by exploiting the profile likelihood’, Bioinformatics 25(15), 1923–1929. Rohde, C. A. (2014), Introductory Statistical Inference with the Likelihood Function:, Springer International Publishing. Royall, R. (1997), Statistical Evidence: A Likelihood Paradigm, CRC Press. Van Kampen, N. G. (1992), Stochastic processes in physics and chemistry, Vol. 1, Elsevier. 10 

Functional Map of the World arXiv:1711.07846v2 [cs.CV] 23 Nov 2017 Gordon Christie1 Neil Fendley1 James Wilson2 Ryan Mukherjee1 1 The Johns Hopkins University Applied Physics Laboratory 2 DigitalGlobe {gordon.christie,neil.fendley,ryan.mukherjee}@jhuapl.edu james.wilson@digitalglobe.com Abstract 1 We present a new dataset, Functional Map of the World (fMoW), which aims to inspire the development of machine learning models capable of predicting the functional purpose of buildings and land use from temporal sequences of satellite images and a rich set of metadata features. The metadata provided with each image enables reasoning about location, time, sun angles, physical sizes, and other features when making predictions about objects in the image. Our dataset consists of over 1 million images from over 200 countries1 . For each image, we provide at least one bounding box annotation containing one of 63 categories, including a “false detection” category. We present an analysis of the dataset along with baseline approaches that reason about metadata and temporal views. Our data, code, and pretrained models have been made publicly available. FD flooded road … gsd: 0.5349 utm: 21J timestamp: 2014-07-08T14:10:29Z … off_nadir_angle_dbl: 21.865 … FD flooded road 2 2 gsd: 0.5087 utm: 21J timestamp: 2017-03-26T14:04:08Z … off_nadir_angle_dbl: 17.407 ... Figure 1: In fMoW, temporal sequences of images, metadata and bounding boxes for 63 categories (including “false detections”) are provided. In this example, if we only look inside the yellow box for the right image, we will only see road and vegetation. In the left image, we will only see water, and potentially predict this to be a lake. However, by observing both views of this area we can now reason that this sequence contains a flooded road. 1. Introduction Satellite imagery presents interesting opportunities for the development of object classification methods. Most computer vision (CV) datasets for this task focus on images or videos that capture brief moments [22, 18]. With satellite imagery, temporal views of objects are available over long periods of time. In addition, metadata is also available to enable reasoning beyond visual information. For example, by combining temporal image sequences with timestamps, models may learn to differentiate office buildings from multi-unit residential buildings by observing whether or not their parking lots are full during business hours. Models may also be able to combine certain metadata parameters with observations of shadows to estimate object heights. In addition to these possibilities, robust models must be able to generalize to unseen areas around the world that may include different building materials and unique architectural styles. Enabling the aforementioned types of reasoning requires 1 207 1 a large dataset of annotated and geographically diverse satellite images. In this work, we present our efforts to collect such a dataset, entitled Functional Map of the World (fMoW). fMoW has several notable features, including a variable number of temporal images per scene and an associated metadata file for each image. The task posed for our dataset falls in between object detection and classification. That is, for each temporal sequence of images, at least one bounding box is provided that maps to one of 63 categories, including a “false detection” (FD) category that represents content not characterized by the other 62 categories. These boxes are intended to be used as input to a classification algorithm. Figure 1 shows an example. Collecting a dataset such as fMoW presents some interesting challenges. For example, one consideration would be to directly use crowdsourced annotations provided by OpenStreetMap2 (OSM). However, issues doing so include 2 https://www.openstreetmap.org of the total 247 ISO Alpha-3 country codes are present in fMoW. 1 inconsistent, incorrect, and missing annotations for a large percentage of buildings and land use across the world. Moreover, OSM may only provide a single label for the current contents of an area, making it difficult to correctly annotate temporal views. Another possibility is to use the crowd to create annotations from scratch. However, annotating instances of a category with no prior information is extremely difficult in a large globally-diverse satellite dataset. This is due in part to the unique perspective that satellite imagery offers when compared with ground-based datasets, such as ImageNet [22]. Humans are seldom exposed to aerial viewpoints in their daily lives and, as such, objects found in satellite images tend to be visually unfamiliar and difficult to identify. Buildings can also be repurposed throughout their lifetime, making visual identification even more difficult. For these reasons, we use a multiphase process that combines map data and crowdsourcing. Another problem for fMoW is that full annotation is made very difficult by the increased object density for certain categories. For example, single-unit residential buildings often occur in dense clusters alongside other categories, where accurately discriminating and labeling every building would be very time-consuming. To address this shortcoming, we propose providing bounding boxes as part of algorithm input as opposed to requiring bounding box output, which would be more akin to a typical detection dataset. This avoids full image annotation issues that stem from incomplete map data and visual unfamiliarity. Imagery does not have to be fully annotated, as algorithms are only asked to classify regions with known contents. This allows us to focus collection on areas with more accurate map data and limit annotations to a small number of category instances per image. Our contributions are summarized as follows: (1) To the best of our knowledge, we provide the largest publicly available satellite dataset containing bounding box annotations, metadata and revisits. This enables joint reasoning about images and metadata, as well as long-term temporal reasoning for areas of interest. (2) We present methods based on CNNs that exploit the novel aspects of our dataset, with performance evaluation and comparisons, which can be applied to similar problems in other application domains. Our code, data, and pretrained models have all been publicly released3 . In the following sections, we provide an analysis of fMoW and baseline methods for the task. lected and annotated. Recently, there have been several, mostly successful, attempts to leverage techniques that were founded on first-person imagery and apply them to remote sensing data [13, 19, 28]. However, these efforts highlight the research gap that has developed due to the lack of a large dataset to appropriately characterize the problems found in remote sensing. fMoW offers an opportunity to close this gap by providing, to the best of our knowledge, the largest quantity of labeled satellite images that has been publicly released to date, while also offering several features that could help unify otherwise disparate areas of research around the multifaceted problem of processing satellite imagery. We now highlight several of these areas where we believe fMoW can make an impact. Reasoning Beyond Visual Information Many works have extended CV research to simultaneously reason about other modules of perception, such as joint reasoning about language and vision [2, 14, 21], audio and vision [10], 2D and 3D information [3], and many others. In this work, we are interested in supporting joint reasoning about temporal sequences of images and associated metadata features. One of these features is UTM zone, which provides location context. In a similar manner, [24] shows improved image classification results by jointly reasoning about GPS coordinates and images, where several features are extracted from the coordinates, including high-level statistics about the population. Although we use coarser location features (UTM zones) than GPS in this work, we do note that using similar features would be an interesting study. Multi-view Classification Satellite imagery offers a unique and somewhat alien perspective on the world. Most structures are designed for recognition from ground level. For example, buildings often have identifying signs above entrances that are not visible from overhead. As such, it can be difficult, if not impossible, to identify the functional purpose of a building from a single overhead image. One of the ways in which fMoW attempts to address this issue is by providing multiple temporal views of each object, when available. Along these lines, several works in the area of video processing have been able to build upon advancements in single image classification [15, 6, 30] to create networks capable of extracting spatio-temporal features. These works may be a good starting point, but it is important to keep in mind the vastly different temporal resolution on which these datasets operate. For example, the YouTube-8M dataset [1], on which many of these video processing algorithms were developed, contains videos with 30 frames per second temporal resolution that each span on the order of minutes. Satellites, on the other hand, typically cannot capture imagery with such dense temporal resolution. Revisit times vary, but it is not uncommon for satellites to require multiple days before they can image the same location; it is possible for months to go by before they can get 2. Related Work While large datasets are nothing new to the vision community, they have typically focused on first-person or ground-level imagery [22, 18, 1, 8, 9, 7, 17]. This is likely due in part to the ease with which this imagery can be col3 https://github.com/fMoW 2 an unobstructed view. As such, temporal views in fMoW span multiple years as opposed to minutes. Techniques that attempt to capture features across disjoint periods of time, such as [20], are likely better candidates for the task. Perhaps the most similar work to ours in terms of temporal classification is PlaNet [26]. They pose the image localization task as a classification problem, where photos are classified as belonging to a particular bucket that bounds a specific area on the globe. They extend their approach to classify the buckets of images in photo albums taken in the same area. A similar approach is used in one of our baseline methods for fMoW. Another recent work similar to fMoW is TorontoCity [25]. They provide a large dataset that includes imagery and LiDAR data collected by airplanes, low-altitude unmanned aerial vehicles, and cars in the greater Toronto area. While they present several tasks, the two that are related to land-use classification are zoning classification and segmentation (e.g., residential, commercial). Aerial images included in TorontoCity were captured during four different years and include several seasons. While this is an impressive dataset, we believe fMoW is more focused on satellite imagery and offers advantages in geographic diversity. Satellite Datasets One of the earliest annotated satellite datasets similar to fMoW is the UC Merced Land Use Dataset, which offers 21 categories and 100 images per category with roughly 30cm resolution and image sizes of 256x256 [29]. While some categories from this dataset overlap with fMoW, we believe fMoW offers several advantages in that we have three times the number of categories, localized objects within the images, and multiple orders of magnitude more images per category. We also provide metadata, temporal views, and multispectral images. SpaceNet [5], a recent dataset that has received substantial attention, contains both 30cm and 50cm data of 5 cities. For the most part, the data in SpaceNet currently includes building footprints. However, earlier this year, point of interest (POI) data was also released into SpaceNet. This POI data includes the locations of several categories within Rio de Janeiro. Unrelated to SpaceNet, efforts have also been made to label data from Google Earth, with the largest released thus far being the AID [27] and NWPU-RESISC45 [4] datasets. The AID dataset includes 10,000 images of 30 categories, while the NWPU-RESISC45 dataset includes 31,500 images of 45 categories. In comparison, fMoW offers over 1,000,000 images of 63 categories. Datasets derived from Google Earth imagery lack associated metadata, temporal views, and multispectral data, which would typically be available to real-world systems. million images, collection resources, plan to collect temporal views, and discussions with researchers in the CV community, we set a goal of including between 50 and 100 categories. We searched sources such as the OSM Map Features4 list and NATO Geospatial Feature Concept Dictionary5 for categories that highlight some of the challenges discussed in Section 2. For example, “construction site” and “impoverished settlement” are categories from our dataset that may require temporal reasoning to identify, which presents a unique challenge due to temporal satellite image sequences typically being scattered across large time periods. We also focused on grouping categories according to their functional purpose, which should encourage the development of approaches that reason about contextual information, both visually and in the associated metadata. Beyond research-based rationales for picking certain categories, we had some practical ones as well. Before categories could be annotated within images, we needed to find locations where we have high confidence of their existence. This is where maps play a crucial role. “Flooded road”, “debris or rubble”, and “construction site” were the most difficult categories to collect because open source data does not generally contain temporal information. However, with more careful search procedures, reuse of data from humanitarian response campaigns, and calculated extension of keywords to identify categories even when not directly labeled, we were able to collect temporal stacks of imagery that contained valid examples. All imagery used in fMoW was collected from the DigitalGlobe constellation6 . Images were gathered in pairs, consisting of 4-band or 8-band multispectral imagery in the visible to near-infrared region, as well as a pan-sharpened RGB image that represents a fusion of the high-resolution panchromatic image and the RGB bands from the lowerresolution multispectral image. 4-band imagery was obtained from either the QuickBird-2 or GeoEye-1 satellite systems, whereas 8-band imagery was obtained from WorldView-2 or WorldView-3. More broadly, fMoW was created using a three-phase workflow consisting of location selection, image selection, and bounding box creation. The location selection phase was used to identify potential locations that map to our categories while also ensuring geographic diversity. Potential locations were drawn from several Volunteered Geographic Information (VGI) datasets, which were conflated and curated to remove duplicates. To ensure diversity, we removed neighboring locations within a specified distance (typically 500m) and set location frequency caps for categories that have severely skewed geographic distributions. These two factors helped reduce spatial density while also encouraging 3. Dataset Collection 4 https://wiki.openstreetmap.org/wiki/Map_Features 5 https://portal.dgiwg.org/files/?artifact_id=8629 Prior to the dataset collection process for fMoW, a set of categories had to be identified. Based on our target of 1 6 https://www.digitalglobe.com/resources/ satellite-information 3 the selection of locations from disparate geographic areas. The remaining locations were then processed using DigitalGlobe’s GeoHIVE7 crowdsourcing platform. Members of the GeoHIVE crowd were asked to validate the presence of categories in satellite images, as shown in Figure 2. able locations from the United States. Many of these “wind farm” instances were invalidated by the crowd, likely due to the difficulty of identifying tall, thin structures in satellite imagery, particularly when the satellite image is looking straight down on the tower. The “barn”, “construction site”, “flooded road”, and “debris or rubble” categories are also examples that contain some geographic bias. In the case of the “barn” category, the bias comes from the distribution of “barn” tags in OSM, which are predominately located in Europe, whereas the other three categories contain geographic bias as a result of the more complex feature selection process, mentioned earlier, that was required for these categories. 4. Dataset Analysis Here we provide some statistics and analysis of fMoW. Two versions of the dataset are publicly available: • fMoW-full The full version of the dataset includes pan-sharpened RGB images and 4/8-band multispectral images (MSI), which are both stored in TIFF format. Pan-sharpened images are created by “sharpening” lower-resolution MSI using higher-resolution panchromatic imagery. All pan-sharpened images in fMoW-full have corresponding MSI, where the metadata files for these images are nearly identical. • fMoW-rgb An alternative JPEG compressed version of the dataset, which is provided since fMoW-full is very large. For each pan-sharpened RGB image we simply perform a conversion to JPEG. For MSI images, we extract the RGB channels and save them as JPEGs. For all experiments presented in this paper, we use fMoW-rgb. We also exclude RGB-extracted versions of the MSI in fMoW-rgb as they are effectively downsampled versions of the pan-sharpened RGB images. Figure 2: Sample image of what a GeoHIVE user might see while validating potential fMoW dataset features. Instructions can be seen in the top-left corner that inform users to press the ‘1’, ‘2’, or ‘3’ keys to validate existence, nonexistence, or cloud obscuration of a particular object. The image selection phase comprised of a three-step process, which included searching the DigitalGlobe satellite imagery archive, creating image chips, and filtering out cloudy images. Approximately 30% of the candidate images were removed for being too cloudy. DigitalGlobe’s IPE Data Architecture Highly-available Objectstore (IDAHO) service was used to process imagery into pan-sharpened RGB and multispectral image chips in a scalable fashion. These chips were then passed through a CNN architecture to classify and remove any undesirable cloud-covered images. Finally, images that passed through the previous two phases were sent to a curated and trusted crowd for bounding box annotation. This process involved a separate interface from the first phase, one that asked crowd users to draw bounding boxes around the category of interest in each image and provided some category-specific guidance for doing so. The resulting bounding boxes were then graded by second trusted crowd to assess quality. In total, 642 unique GeoHIVE users required a combined total of approximately 2,800 hours to annotate category instances for fMoW. Even after multiple crowd validation procedures and implementing programmatic methods for ensuring geographic diversity, there were several categories that contained some bias. For example, the “wind farm” category does not contain very many examples from the United States, even though the initial location selection phase returned 1,938 vi- 4.1. fMoW Splits We have made the following splits to the dataset: • seq This is the sequestered portion of the dataset that is not currently publicly available. It will be released after it is used for final testing in the public challenge centered around the dataset8 . • train Contains 65.2% and 72.13% of the total bounding boxes with and without seq included, respectively. • val Contains 11.4% and 12.6% of the total bounding boxes with and without seq included, respectively. This set was made representative of test, so that validation can be performed. • test Contains 13.8% and 15.3% of the total bounding boxes with and without seq included, respectively. 7 https://geohive.digitalglobe.com 8 https://www.iarpa.gov/challenges/fmow.html 4 Instances per Category 90000 80000 8 band 70000 4 band 60000 3 band 50000 40000 30000 20000 10000 false detection airport airport hangar airport terminal amusement park aquaculture archaeological site barn border checkpoint burial site car dealership construction site crop field dam debris or rubble educational institution electric substation factory or powerplant fire station flooded road fountain gas station golf course ground transportation station helipad hospital impoverished settlement interchange lake or pond lighthouse military facility multi-unit residential nuclear powerplant office building oil or gas facility park parking lot or garage place of worship police station port prison race track railway bridge recreational facility road bridge runway shipyard shopping mall single-unit residential smokestack solar farm space facility stadium storage tank surface mine swimming pool toll booth tower tunnel opening waste disposal water treatment facility wind farm zoo 0 Figure 3: This shows the total number of instances for each category (including FD) in fMoW across different number of bands. These numbers include the temporal views of the same areas. fMoW-full consists of 3 band imagery (pan-sharpened RGB), as well as 4 and 8 band imagery. In fMoW-rgb, the RGB channels of the 4 and 8 band imagery are extracted and saved as JPEG images. The total number of bounding box instances for each category can be seen in Figure 3. UTM zones in the metadata. Figure 5 illustrates the frequency of sequences within the UTM zones on earth, where the filled rectangles each represent a different UTM zone. Green colors represent areas with higher numbers of sequences, while blue regions have lower counts. As seen, fMoW covers much of the globe. The images captured for fMoW also have a wide range of dates, which allows algorithms to analyze areas on earth over long periods of time in some cases. Figure 6 shows a distributions for years and local times (converted from UTC) in which the images were captured. The average time difference between the earliest and most recent images in each sequence is approximately 3.8 years. 4.2. fMoW Statistics Variable length sequences of images are provided for each scene in the dataset. Figure 4 shows what percentage of the sequences in the dataset belong to each sequence length. 21.2% of the sequences contain only 1 view. Most (95%) of the sequences contain 10 or less images. Percentage of Dataset Number of Temporal Views Distribution 25% 20% 15% 5. Baselines and Methods 10% Here we present 5 different approaches to our task, which vary by their use of metadata and temporal reasoning. All experiments were performed using fMoW-rgb. Two of the methods presented involve fusing metadata into a CNN architecture. The following provides a summary of the metadata features that are used, as well as any preprocessing operations that are applied: • UTM Zone One of 60 UTM zones and one of 20 latitude bands are combined for this feature. We convert these values to 2 coordinate values, each between 0 and 1. This is done by taking the indices of the values within the list of possible values and then normalizing. • Timestamp The year, month, day, hour, minute, second, and day of the week are extracted from the timestamp and added as separate features. The timestamp provided in the metadata files is in Coordinated Universal Time (UTC). 5% 0% 1 3 5 7 9 11 13 15 17 Number of Temporal Views 19 20+ Figure 4: This shows the distribution of the number of temporal views in our dataset. The number of temporal views is not incremented by both the pan-sharpened and multispectral images. These images have almost identical metadata files and are therefore not counted twice. The maximum number of temporal views for any area in the dataset is 41. A major focus of the collection effort was global diversity. In the metadata, we provide UTM zones, which typically refer to 6◦ longitude bands (1-60). We also concatenate letters that represent latitude bands (total of 20) to the 5 major axis. – Sun Azimuth Angle in degrees (0-360◦ ) of clockwise rotation off north to the sun. – Sun Elevation Angle in degrees (0-90◦ ) of elevation, measured from the horizontal, to the sun. • Image+box sizes The pixel dimensions of the bounding boxes and image size, as well as the fraction of the image width and height that the boxes occupy are added as features. After preprocessing the metadata features, we perform mean subtraction and normalization using values calculated for train + val. A full list of metadata features and their descriptions can be found in the appendix. It is worth noting here that the imagery in fMoW is not registered, and while many sequences have strong spatial correspondence, individual pixel coordinates in different images do not necessarily represent the same positions on the ground. As such, we are prevented from easily using methods that exploit registered sequences. The CNN used as the base model in our various baseline methods is DenseNet-161 [12], with 48 feature maps (k=48). During initial testing, we found this model to outperform other models such as VGG-16 [23] and ResNet50 [11]. We initialize our base CNN models using the pretrained ImageNet weights, which we found to improve performance during initial tests. Training is performed using a crop size of 224x224, the Adam optimizer [16], and an initial learning rate of 1e-4. Due to class imbalance in our dataset, we attempted to weight the loss using class frequencies, but did not observe any improvement. To merge metadata features into the model, the softmax layer of DenseNet is removed and replaced with a concatenation layer to merge DenseNet features with preprocessed metadata features, followed by two 4096-d fully-connected layers with 50% dropout layers, and a softmax layer with 63 outputs (62 main categories + FD). An illustration of this base model is shown in Figure 7. Figure 5: This shows the geographic diversity of fMoW. Data was collected from over 400 unique UTM zones (including latitude bands). This helps illustrate the number of images captured in each UTM zone, where more green colors show UTM zones with a higher number of instances, and more blue colors show UTM zones with lower counts. 11.6% Time Distribution 2010 25% 2011 5.4% 4.6% 5.3% 2012 6.0% 2016 7.1% 29.8% 2013 2015 (a) 15% 10% 5% 0% 11.7% 18.5% 20% 2014 00:00-09:30 09:30-10:00 10:00-10:30 10:30-11:00 11:00-11:30 11:30-12:00 12:00-12:30 12:30-13:00 13:00-13:30 13:30-14:00 14:00-24:00 2002-2009 2017 (b) Figure 6: Distribution over (a) years the images were captured, and (b) time of day the images were captured (UTC converted to local time for this figure). • GSD Ground sample distance, measured in meters, is provided for both the panchromatic and multispectral bands in the image strip. The panchromatic images used to generate the pan-sharpened RGB images have higher resolution than the MSI, and therefore have smaller GSD values. These GSD values, which describe the physical sizes of pixels in the image, are used directly without any preprocessing. • Angles These identify the angle at which the sensor is imaging the ground, as well as the angular location of the sun with respect to the ground and image. These features can be added without preprocessing. The following angles are provided: – Off-nadir Angle Angle in degrees (0-90◦ ) between the point on the ground directly below the sensor and the center of the image swath. – Target Azimuth Angle in degrees (0-360◦ ) of clockwise rotation off north to the image swath’s DenseNet Softmax 4096 4096 Extract Features Concat gsd: 0.5219 utm: 30T timestamp: 2016-02-04T12:29:21Z … off_nadir_angle_dbl: 10.154 ... Figure 7: An illustration of our base model used to fuse metadata features into the CNN. This model is used as a baseline and also as a feature extractor (without softmax) for providing features to an LSTM. Dropout layers are added after the 4096-d FC layers. 6 We test the following approaches with fMoW: • LSTM-M An LSTM architecture trained using temporal sequences of metadata features. We believe training solely on metadata helps understand how important images are in making predictions, while also providing some measure of bias present in fMoW. • CNN-I A standard CNN approach using only images, where DenseNet is fine-tuned after ImageNet. Softmax outputs are summed over each temporal view, after which an argmax is used to make the final prediction. The CNN is trained on all images across all temporal sequences of train + val. • CNN-IM A similar approach to CNN-I, but with metadata features concatenated to the features of DenseNet before the fully connected layers. • LSTM-I An LSTM architecture trained using features extracted from CNN-I. • LSTM-IM An LSTM architecture trained using features extracted from CNN-IM. All of these methods are trained on train + val. Since tight bounding boxes are typically provided for category instances in the dataset, we add a context buffer around each box before extracting the region of interest from the image. We found that it was useful to provide more context for categories with smaller sizes (e.g., single-unit residential) and less context for categories that generally cover larger areas (e.g., airports). Per-category F1 scores for test are shown in Table 1. From the results, it can be observed that, in general, the LSTM architectures show similar performance to our approaches that sum the probabilities over each view. Some possible contributors to this are the large quantity of singleview images provided in the dataset, and that temporal changes may not be particularly important for several of the categories. CNN-I and CNN-IM are also, to some extent, already reasoning about temporal information while making predictions by summing the softmax outputs over each temporal view. Qualitative results that show success and failure cases for LSTM-I are shown in Figure 8. Qualitative results are not shown for the approaches that use metadata, as it is much harder to visually show why the methods succeed in most cases. It could be argued that the results for approaches using metadata are only making improvements because of bias exploitation. To show that metadata helps beyond inherent bias, we removed all instances from the test set where the metadata-only baseline (LSTM-M) is able to correctly predict the category. The results of this removal, which can be found in Table 2, show that metadata can still be useful for improving performance. To further confirm the importance of temporal reasoning, we compare the methods presented above with two additional methods, CNN-I-1 and CNN-IM-1, which make LSTM-M CNN-I LSTM-I airport airport hangar airport terminal amusement park aquaculture archaeological site barn border checkpoint burial site car dealership construction site crop field dam debris or rubble educational institution electric substation factory or powerplant fire station flooded road fountain gas station golf course ground transportation station helipad hospital impoverished settlement interchange lake or pond lighthouse military facility multi-unit residential nuclear powerplant office building oil or gas facility park parking lot or garage place of worship police station port prison race track railway bridge recreational facility road bridge runway shipyard shopping mall single-unit residential smokestack solar farm space facility stadium storage tank surface mine swimming pool toll booth tower tunnel opening waste disposal water treatment facility wind farm zoo 0.599 0.447 0.017 0.023 0.622 0.514 0.016 0.292 0.000 0.019 0.101 0.053 0.514 0.158 0.381 0.157 0.000 0.000 0.028 0.625 0.085 0.022 0.220 0.114 0.067 0.012 0.538 0.142 0.000 0.037 0.426 0.227 0.000 0.011 0.522 0.025 0.076 0.362 0.068 0.444 0.087 0.234 0.030 0.295 0.000 0.488 0.000 0.117 0.429 0.204 0.424 0.000 0.174 0.140 0.200 0.362 0.030 0.141 0.526 0.071 0.044 0.540 0.039 0.728 0.859 0.721 0.697 0.746 0.754 0.524 0.695 0.333 0.852 0.741 0.372 0.888 0.806 0.403 0.495 0.849 0.443 0.409 0.296 0.727 0.785 0.860 0.658 0.812 0.387 0.410 0.833 0.721 0.715 0.509 0.385 0.720 0.198 0.789 0.626 0.775 0.638 0.246 0.692 0.611 0.898 0.703 0.907 0.722 0.821 0.371 0.615 0.688 0.735 0.912 0.824 0.825 0.921 0.824 0.920 0.891 0.723 0.867 0.595 0.854 0.939 0.566 0.729 0.800 0.665 0.715 0.727 0.762 0.491 0.684 0.404 0.859 0.797 0.373 0.872 0.798 0.607 0.475 0.869 0.459 0.494 0.285 0.705 0.779 0.916 0.694 0.856 0.404 0.506 0.678 0.650 0.755 0.564 0.414 0.762 0.218 0.773 0.638 0.787 0.658 0.237 0.698 0.650 0.886 0.755 0.919 0.738 0.814 0.351 0.629 0.703 0.755 0.921 0.737 0.850 0.921 0.802 0.913 0.918 0.737 0.897 0.570 0.816 0.948 0.582 0.853 0.884 0.677 0.746 0.898 0.811 0.574 0.717 0.523 0.827 0.747 0.318 0.930 0.864 0.474 0.548 0.858 0.536 0.483 0.638 0.814 0.761 0.899 0.713 0.831 0.426 0.750 0.905 0.687 0.779 0.597 0.445 0.600 0.228 0.844 0.662 0.700 0.712 0.201 0.736 0.695 0.919 0.761 0.903 0.747 0.889 0.368 0.662 0.717 0.772 0.927 0.875 0.818 0.928 0.870 0.906 0.960 0.754 0.949 0.604 0.853 0.959 0.598 0.837 0.837 0.699 0.759 0.868 0.805 0.607 0.707 0.515 0.846 0.770 0.358 0.926 0.886 0.488 0.557 0.872 0.544 0.523 0.795 0.840 0.772 0.875 0.719 0.820 0.458 0.704 0.909 0.694 0.828 0.655 0.451 0.552 0.225 0.865 0.698 0.732 0.735 0.329 0.667 0.726 0.892 0.813 0.906 0.756 0.885 0.351 0.662 0.684 0.768 0.931 0.889 0.819 0.924 0.880 0.907 0.954 0.777 0.942 0.670 0.879 0.968 0.611 Average 0.193 0.679 0.688 0.722 0.734 false_detection CNN-IM LSTM-IM Table 1: F1 scores for different approaches on test. Color formatting was applied to each column independently. The average values shown at the bottom of the table are calculated without FD scores. predictions for each individual view. We then have all other methods repeat their prediction over the full sequence. This is done to show that, on average, seeing an area multiple times outperforms single-view predictions. We note that these tests are clearly not fair for some categories, such as 7 LSTM-I: Construction Site CNN-I: Educational Institution LSTM-I: Debris or Rubble LSTM-I: Flooded Road CNN-I: Hospital GT: Debris or Rubble CNN-I: False Detection LSTM-I: Construction Site GT: Construction Site CNN-I: False Detection GT: False Detection GT: False Detection Figure 8: Qualitative examples from test of the image-only approaches. The images presented here show the extracted and resized images that are passed to the CNN approaches. The top two rows show success cases for LSTM-I, where CNN-I was not able to correctly predict the category. The bottom two rows show failure cases for LSTM-I, where CNN-I was able to correctly predict the category. We also note that sequences with ≥9 views were chosen. The second row was trimmed to keep the figure consistent. However, we note that variable temporal views are provided for throughout the dataset. 6. Conclusion and Discussion “construction site”, where some views may not even contain the category. However, we perform these tests for completeness to confirm our expectations. Results are shown in Table 3. Per-category results can be found in the appendix. LSTM-M CNN-I LSTM-I CNN-IM LSTM-IM 0 0.685 0.693 0.695 0.702 We present fMoW, a dataset that consists of over 1 million image and metadata pairs, of which many are temporal views of the same scene. This enables reasoning beyond visual information, as models are able to leverage temporal information and reason about the rich set of metadata features (e.g., timestamp, UTM zone) provided for each image. By posing a task in between detection and classification, we avoid the inherent challenges associated with collecting a large, geographically diverse, detection dataset, while still allowing for models to be trained that are transferable to real-world detection systems. Different methods were presented for this task that demonstrate the importance of reasoning about metadata and temporal information. All code, data, and pretrained models have been made publicly available. We hope that by releasing the dataset and code, other researchers in the CV community will find new and interesting ways to further utilize the metadata and temporal changes to a scene. We also hope to see fMoW being used to train models that are able to assist in humanitarian efforts, such as applications involving disaster relief. Table 2: Results on test instances where the metadataonly baseline (LSTM-M) is not able to correctly predict the category. These are the average F1 scores not including FD. These results show that metadata is important beyond exploiting bias in the dataset. CNN-I-1 CNN-I LSTM-I CNN-IM-1 CNN-IM LSTM-IM 0.618 0.678 0.684 0.666 0.722 0.735 Table 3: Average F1 scores, not including FD, for individual images from test. CNN-I-1 and CNN-IM-1 make predictions for each individual view. All other methods repeat their prediction over the full sequence. 8 Acknowledgements This work would not have been possible without the help of everyone on the fMoW Challenge team, who we thank for their contributions. A special thanks to: Kyle Ellis, Todd Bacastow, Alex Dunmire, and Derick Greyling from DigitalGlobe; Rebecca Allegar, Jillian Brennan, Dan Reitz, and Ian Snyder from Booz Allen Hamilton; Kyle Bowerman and Gődény Balázs from Topcoder; and, finally, Myron Brown, Philippe Burlina, Alfred Mayalu, and Nicolas Norena Acosta from JHU/APL. We also thank the professors, graduate students, and researchers in industry and from the CV community for their suggestions and participation in discussions that helped shape the direction of this work. The material in this paper is based upon work supported by the Office of the Director of National Intelligence (ODNI), Intelligence Advanced Research Projects Activity (IARPA), via Contract 2017-17032700004. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of the ODNI, IARPA, or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Governmental purposes notwithstanding any copyright annotation therein. path (String). 6. Pan Resolution Ground sample distance of panchromatic band (pan-GSD) in the image strip, measured in meters (Double). start, end, min, and max values are also included. start and end represent the pan-GSD for the first and last scan lines, respectively. min and max represent the minimum and maximum pan-GSD for all scan lines, respectively. 7. Multi Resolution Ground sample distance of multispectral bands (multi-GSD) in the image strip, measured in meters (Double). start, end, min, and max values are also included. start and end represent the multi-GSD for the first and last scan lines, respectively. min and max represent the minimum and maximum multi-GSD for all scan lines, respectively. 8. Target Azimuth Azimuth angle of the sensor with respect to the center of the image strip, measured in degrees (Double). start, end, min, and max values are also included. start and end represent the target azimuth for the first and last scan lines, respectively. min and max represent the minimum and maximum target azimuth for all scan lines, respectively. 9. Sun Azimuth Azimuth angle of the sun measured from north, clockwise in degrees, to the center of the image strip, measured in degrees (Double). min and max values are also included. min and max represent the minimum and maximum sun azimuth for all scan lines, respectively. 10. Sun Elevation Elevation angle of the sun measured from the horizontal, measured in degrees (Double). min and max values are also included. min and max represent the minimum and maximum sun elevation for all scan lines, respectively. 11. Off-Nadir Angle The off nadir angle of the satellite with respect to the center of the image strip, measured in degrees (Double). start, end, min, and max values are also included. start and end represent the off-nadir angle for the first and last scan lines, respectively. min and max represent the minimum and maximum off-nadir angle for all scan lines, respectively. Appendix Overview In this document, we provide: Appendix I: Descriptions of the metadata features and distributions for country codes and UTM zones. Appendix II: Additional collection details. Appendix III: Additional results. Appendix IV: Examples from our dataset. Appendix I. Metadata Features and Statistics 1. ISO Country Code ISO Alpha-3 country code (String). There are a total of 247 possible country codes, 207 of which are present in fMoW. 2. UTM Zone Universal Transverse Mercator. There are 60 UTM zones, which are 6◦ in width. We provide a number for the UTM zone (1-60), along with a letter representing the latitude band. There are a total of 20 latitude bands, which range from “C” to “X” (“I” and “O” are not included). 3. Timestamp UTC timestamp. Datetime format (Python): “%Y-%m-%dT%H:%M:%SZ” (String). 4. Cloud Cover Fraction of the image strip, not image chip, that is completely obscured by clouds on a scale of 0-100 (Integer). 5. Scan Direction The direction the sensor is pointed when collecting an image strip. Either “Forward”, when the image is collected ahead of the orbital path or “Reverse” when the image is taken behind the orbital Country Codes Here we show the counts for each unique country code in fMoW. Counts are incremented once for each sequence instead of once per metadata file. [(“USA”, 18750), (“FRA”, 7470), (“ITA”, 6985), (“RUS”, 6913), (“CHN”, 6597), (“DEU”, 4686), (“GBR”, 4496), (“BRA”, 3820), (“CAN”, 3128), (“TUR”, 2837), (“JPN”, 2542), (“IDN”, 2448), (“ESP”, 2402), (“AUS”, 2105), (“DZA”, 1849), (“IND”, 1804), (“UKR”, 1735), (“CZE”, 1713), (“POL”, 1386), (“MEX”, 1274), (“ARG”, 1248), (“NLD”, 1236), (“SYR”, 1224), (“BEL”, 1190), (“PHL”, 1179), (“IRQ”, 1129), (“EGY”, 1041), (“ZAF”, 924), (“CHL”, 888), (“LTU”, 871), (“LBY”, 863), (“KOR”, 809), (“CHE”, 788), (“LVA”, 772), (“PRT”, 722), (“YEM”, 9 701), (“BLR”, 601), (“GRC”, 592), (“AUT”, 572), (“SVN”, 570), (“ARE”, 566), (“IRN”, 540), (“COL”, 509), (“TWN”, 509), (“TZA”, 475), (“NZL”, 465), (“PER”, 459), (“HTI”, 417), (“KEN”, 405), (“NGA”, 383), (“VEN”, 378), (“PRK”, 371), (“ECU”, 351), (“IRL”, 335), (“MYS”, 328), (“BOL”, 313), (“FIN”, 288), (“KAZ”, 268), (“MAR”, 266), (“TUN”, 257), (“CUB”, 256), (“EST”, 247), (“SAU”, 246), (“HUN”, 222), (“THA”, 219), (“NPL”, 196), (“HRV”, 187), (“NOR”, 183), (“SVK”, 175), (“SEN”, 172), (“BGD”, 171), (“HND”, 167), (“SWE”, 166), (“BGR”, 165), (“HKG”, 154), (“DNK”, 153), (“MDA”, 147), (“ROU”, 142), (“ZWE”, 141), (“SRB”, 140), (“GTM”, 140), (“DOM”, 134), (“LUX”, 133), (“SDN”, 132), (“VNM”, 126), (“URY”, 120), (“CRI”, 119), (“SOM”, 112), (“ISL”, 110), (“LKA”, 110), (“QAT”, 108), (“PRY”, 107), (“SGP”, 106), (“OMN”, 105), (“PRI”, 95), (“NIC”, 87), (“NER”, 85), (“SSD”, 82), (“UGA”, 79), (“SLV”, 79), (“JOR”, 78), (“CMR”, 77), (“PAN”, 74), (“PAK”, 72), (“UZB”, 70), (“CYP”, 67), (“KWT”, 67), (“ALB”, 66), (“CIV”, 65), (“BHR”, 65), (“GIN”, 64), (“MLT”, 63), (“JAM”, 62), (“AZE”, 62), (“GEO”, 60), (“SLE”, 59), (“ETH”, 58), (“LBN”, 57), (“ZMB”, 55), (“TTO”, 54), (“LBR”, 52), (“BWA”, 51), (“ANT”, 50), (“BHS”, 50), (“MNG”, 46), (“MKD”, 45), (“GLP”, 45), (“COD”, 45), (“KO-”, 42), (“BEN”, 42), (“GHA”, 41), (“MDG”, 36), (“MLI”, 35), (“AFG”, 35), (“ARM”, 33), (“MRT”, 33), (“KHM”, 32), (“CPV”, 31), (“TKM”, 31), (“MMR”, 31), (“BFA”, 29), (“BLZ”, 29), (“NCL”, 28), (“AGO”, 27), (“FJI”, 26), (“TCD”, 25), (“MTQ”, 25), (“GMB”, 23), (“SWZ”, 23), (“BIH”, 21), (“CAF”, 19), (“GUF”, 19), (“PSE”, 19), (“MOZ”, 18), (“NAM”, 18), (“SUR”, 17), (“GAB”, 17), (“LSO”, 16), (“ERI”, 15), (“BRN”, 14), (“REU”, 14), (“GUY”, 14), (“MAC”, 13), (“TON”, 13), (“ABW”, 12), (“PYF”, 12), (“TGO”, 12), (“BRB”, 12), (“VIR”, 11), (“CA-”, 11), (“DJI”, 11), (“FLK”, 11), (“MNE”, 11), (“KGZ”, 11), (“ESH”, 10), (“LCA”, 10), (“BMU”, 10), (“COG”, 9), (“ATG”, 9), (“BDI”, 9), (“GIB”, 8), (“LAO”, 8), (“GNB”, 8), (“DMA”, 8), (“KNA”, 8), (“GNQ”, 7), (“RWA”, 7), (“BTN”, 7), (“TJK”, 6), (“TCA”, 5), (“VCT”, 4), (“WSM”, 3), (“IOT”, 3), (“AND”, 3), (“ISR”, 3), (“AIA”, 3), (“MDV”, 2), (“SHN”, 2), (“VGB”, 2), (“MSR”, 2), (“PNG”, 1), (“MHL”, 1), (“VUT”, 1), (“GRD”, 1), (“VAT”, 1), (“MCO”, 1)] 1086), (“51R”, 1069), (“36S”, 1046), (“35T”, 1038), (“36R”, 1037), (“49M”, 1026), (“48M”, 1021), (“10T”, 1010), (“53S”, 1001), (“10S”, 955), (“14R”, 935), (“19T”, 928), (“30S”, 912), (“17S”, 875), (“17R”, 874), (“43P”, 854), (“50S”, 796), (“36U”, 767), (“50R”, 751), (“33S”, 751), (“32S”, 746), (“14S”, 730), (“34T”, 728), (“12S”, 716), (“37M”, 705), (“13S”, 676), (“37T”, 667), (“36T”, 653), (“15S”, 629), (“55H”, 618), (“34S”, 604), (“29S”, 600), (“38P”, 598), (“15T”, 586), (“22J”, 585), (“18Q”, 549), (“15R”, 539), (“35S”, 511), (“10U”, 497), (“21H”, 492), (“36V”, 491), (“19H”, 482), (“48R”, 476), (“49S”, 459), (“48S”, 446), (“49Q”, 444), (“29T”, 438), (“16P”, 429), (“56H”, 425), (“14Q”, 422), (“40R”, 420), (“39R”, 413), (“39U”, 406), (“18N”, 385), (“35J”, 383), (“37V”, 380), (“50T”, 379), (“56J”, 355), (“34V”, 351), (“43V”, 347), (“29U”, 346), (“38U”, 345), (“17M”, 328), (“38T”, 323), (“19P”, 323), (“51S”, 317), (“54H”, 311), (“49R”, 295), (“34H”, 293), (“22K”, 293), (“48N”, 276), (“20H”, 273), (“50Q”, 268), (“28P”, 262), (“18L”, 260), (“24M”, 258), (“24L”, 256), (“21J”, 255), (“41V”, 254), (“13T”, 254), (“47N”, 253), (“40U”, 253), (“45R”, 251), (“43Q”, 245), (“51Q”, 243), (“51T”, 240), (“39S”, 239), (“19K”, 238), (“19Q”, 237), (“59G”, 236), (“43R”, 234), (“12T”, 230), (“49T”, 227), (“41U”, 223), (“32V”, 219), (“30V”, 212), (“13Q”, 212), (“40V”, 210), (“16R”, 210), (“20T”, 210), (“38R”, 204), (“36J”, 203), (“46T”, 200), (“45T”, 197), (“44U”, 196), (“15Q”, 190), (“50L”, 190), (“32P”, 184), (“60H”, 182), (“47P”, 182), (“20P”, 181), (“24K”, 178), (“17Q”, 178), (“35K”, 169), (“20J”, 168), (“11U”, 165), (“18H”, 164), (“52T”, 163), (“11T”, 161), (“36N”, 158), (“39V”, 157), (“20K”, 157), (“39Q”, 155), (“12U”, 149), (“38V”, 147), (“18P”, 147), (“23L”, 147), (“18G”, 146), (“31N”, 146), (“19J”, 142), (“33P”, 141), (“40Q”, 136), (“13R”, 136), (“47T”, 132), (“47R”, 126), (“48U”, 124), (“32R”, 123), (“15P”, 121), (“39P”, 117), (“48P”, 117), (“33R”, 116), (“45U”, 113), (“43S”, 111), (“44N”, 109), (“54T”, 109), (“32N”, 109), (“36W”, 108), (“17P”, 108), (“36P”, 105), (“31R”, 104), (“56K”, 101), (“20Q”, 101), (“39T”, 97), (“16Q”, 96), (“29R”, 95), (“25L”, 92), (“45Q”, 91), (“46Q”, 91), (“48T”, 90), (“44Q”, 89), (“42V”, 87), (“29N”, 87), (“43U”, 86), (“4Q”, 86), (“47Q”, 85), (“48Q”, 84), (“30N”, 83), (“19G”, 82), (“25M”, 81), (“42Q”, 80), (“44P”, 80), (“20L”, 77), (“50J”, 77), (“53U”, 76), (“38N”, 75), (“27W”, 75), (“44R”, 75), (“33V”, 74), (“34R”, 72), (“49L”, 70), (“36M”, 69), (“40S”, 69), (“12R”, 68), (“37P”, 68), (“52R”, 65), (“14T”, 64), (“50U”, 62), (“35H”, 62), (“50H”, 61), (“28R”, 60), (“54U”, 59), (“46V”, 58), (“44T”, 56), (“21K”, 56), (“55G”, 56), (“22L”, 56), (“35P”, 55), (“31P”, 54), (“29P”, 54), (“35R”, 52), (“30R”, 51), (“19U”, 50), (“53T”, 49), (“46U”, 49), (“50N”, 48), (“47S”, 48), (“42R”, 48), (“37Q”, 47), (“19L”, 47), (“14U”, 47), (“28Q”, 46), (“37N”, 45), (“19F”, 45), (“42U”, 44), (“36K”, 42), (“37R”, 40), (“37W”, 40), UTM Zones Here we show the counts for each unique UTM zone in fMoW. Counts are incremented once for each sequence instead of once per metadata file. [(“31U”, 5802), (“32T”, 4524), (“33T”, 4403), (“30U”, 4186), (“32U”, 3864), (“33U”, 3315), (“31T”, 3150), (“18T”, 2672), (“17T”, 2339), (“34U”, 2049), (“37S”, 1718), (“30T”, 1686), (“37U”, 1672), (“23K”, 1627), (“18S”, 1481), (“11S”, 1388), (“16T”, 1283), (“54S”, 1244), (“38S”, 1229), (“31S”, 1227), (“35U”, 1137), (“35V”, 1116), (“52S”, 1115), (“16S”, 1110), (“51P”, 10 (“41S”, 38), (“42S”, 38), (“38Q”, 37), (“30P”, 37), (“42T”, 36), (“35L”, 36), (“46R”, 36), (“52U”, 35), (“60G”, 35), (“27V”, 34), (“45V”, 34), (“35W”, 34), (“13U”, 34), (“35M”, 34), (“18M”, 32), (“17L”, 32), (“41W”, 32), (“17N”, 31), (“21N”, 31), (“23M”, 30), (“21L”, 29), (“28S”, 28), (“58K”, 28), (“22M”, 28), (“41R”, 27), (“18R”, 27), (“10V”, 26), (“57U”, 26), (“34K”, 26), (“49U”, 25), (“6V”, 25), (“38L”, 25), (“20G”, 25), (“33L”, 24), (“60K”, 24), (“55K”, 23), (“51N”, 23), (“22H”, 22), (“22N”, 22), (“47V”, 22), (“41T”, 21), (“44V”, 21), (“36Q”, 21), (“46S”, 20), (“22T”, 20), (“34N”, 19), (“20U”, 19), (“12Q”, 19), (“12V”, 19), (“19N”, 18), (“31Q”, 18), (“21M”, 18), (“52L”, 18), (“56V”, 18), (“52V”, 18), (“23J”, 16), (“45W”, 16), (“9U”, 16), (“34J”, 16), (“27P”, 16), (“43W”, 15), (“1K”, 14), (“33M”, 14), (“40W”, 14), (“40K”, 14), (“43T”, 14), (“55T”, 14), (“51U”, 13), (“53K”, 13), (“34M”, 13), (“32M”, 13), (“37L”, 13), (“21P”, 12), (“50P”, 12), (“35N”, 12), (“6K”, 11), (“59H”, 11), (“33K”, 11), (“20M”, 11), (“49N”, 11), (“5Q”, 10), (“6W”, 10), (“26Q”, 10), (“39L”, 10), (“47U”, 10), (“34W”, 10), (“50K”, 10), (“8V”, 10), (“20S”, 10), (“40T”, 9), (“51V”, 9), (“42W”, 8), (“60W”, 8), (“53H”, 8), (“50V”, 8), (“20F”, 8), (“53L”, 7), (“18F”, 7), (“35Q”, 7), (“30Q”, 7), (“44S”, 7), (“15M”, 7), (“5V”, 7), (“54J”, 7), (“39W”, 6), (“49P”, 6), (“50M”, 6), (“19V”, 6), (“21F”, 6), (“20N”, 5), (“14P”, 5), (“34P”, 5), (“53J”, 5), (“38M”, 5), (“51K”, 5), (“29Q”, 4), (“11R”, 4), (“49V”, 4), (“48V”, 4), (“51M”, 4), (“38W”, 4), (“33N”, 4), (“45S”, 4), (“27Q”, 4), (“55J”, 3), (“19M”, 3), (“53V”, 3), (“2W”, 3), (“32Q”, 3), (“2L”, 3), (“16M”, 3), (“57W”, 3), (“43M”, 3), (“53W”, 2), (“43N”, 2), (“52J”, 2), (“28M”, 2), (“56T”, 2), (“33H”, 2), (“21T”, 2), (“44W”, 2), (“15V”, 1), (“33W”, 1), (“60V”, 1), (“18K”, 1), (“31M”, 1), (“54M”, 1), (“58P”, 1), (“58W”, 1), (“40X”, 1), (“58G”, 1), (“57V”, 1), (“16U”, 1), (“59K”, 1), (“52N”, 1), (“2K”, 1), (“33Q”, 1), (“34Q”, 1), (“11V”, 1), (“56W”, 1), (“26P”, 1), (“28W”, 1), (“59W”, 1), (“38K”, 1), (“26S”, 1), (“7L”, 1), (“56U”, 1), (“55V”, 1)] existence, non-existence, and cloud cover. Appendix II. Dataset Collection Introduced in the main paper, CNN-I-1 and CNN-IM-1 make predictions for each individual view. All other methods repeat their prediction over the full sequence. Again, we note that these tests are clearly not fair to some categories, such as “construction site”, where some views may not even contain the category. However, we show results for these tests for completeness. Only the average values, which do not include “false detection” results, are shown in the main paper. We show per-category results in Table 4. Figure 9: Sample image (“wind farm”) of what a GeoHIVE user might see while validating potential fMoW features. Instructions can be seen in the top-left corner that inform users to press the ‘1’, ‘2’, or ‘3’ keys to validate existence, non-existence, or cloud obscuration of a particular object. For validation of object localization, a different interface is used that asks users to draw a bounding box around the object of interest after being given an initial seed point. The visualization for this is shown in Figure 10, and the seed point can be seen as a green dot located on the object of interest. Users are additionally provided some instructions regarding how large of a box to draw, which may vary by object class. This interface is more complex than the location selection interface, which is why it is performed after object existence can be confirmed and non-cloudy high-quality imagery is obtained. A smaller and more experienced group of users is also used for this task to help ensure the quality of the annotations. Appendix III. Additional Results The location selection phase was used to identify potential locations that map to our categories while also ensuring geographic diversity. Potential locations were drawn from several Volunteered Geographic Information (VGI) datasets, which were conflated and curated to remove duplicates and ensure geographic diversity. The remaining locations were then processed using DigitalGlobe’s GeoHIVE crowdsourcing platform. Members of the GeoHIVE crowd were asked to validate the presence of categories in satellite images, as shown in Figure 9. The interface uses centerpoint location information to draw a circle around a possible object of interest. The interface then asks users to rapidly verify the existence of a particular label, as extracted from the VGI datasets, using the ‘1’, ‘2’, and ‘3’ keys to represent Appendix IV. Dataset Examples Figure 11 shows one example for each category in our dataset. For viewing purposes, regions within the full image chip were extracted using the scaled bounding box coordinates for the categories. For the baseline approaches 11 CNN-I-1 CNN-I LSTM-I airport airport hangar airport terminal amusement park aquaculture archaeological site barn border checkpoint burial site car dealership construction site crop field dam debris or rubble educational institution electric substation factory or powerplant fire station flooded road fountain gas station golf course ground transportation station helipad hospital impoverished settlement interchange lake or pond lighthouse military facility multi-unit residential nuclear powerplant office building oil or gas facility park parking lot or garage place of worship police station port prison race track railway bridge recreational facility road bridge runway shipyard shopping mall single-unit residential smokestack solar farm space facility stadium storage tank surface mine swimming pool toll booth tower tunnel opening waste disposal water treatment facility wind farm zoo 0.662 0.815 0.664 0.653 0.698 0.642 0.458 0.598 0.232 0.736 0.664 0.286 0.853 0.755 0.297 0.461 0.771 0.406 0.337 0.253 0.659 0.691 0.852 0.627 0.703 0.331 0.429 0.804 0.615 0.634 0.520 0.388 0.548 0.180 0.692 0.563 0.710 0.560 0.187 0.630 0.540 0.847 0.645 0.864 0.667 0.781 0.388 0.549 0.643 0.665 0.784 0.748 0.844 0.895 0.746 0.859 0.841 0.642 0.789 0.475 0.815 0.899 0.471 0.737 0.864 0.746 0.726 0.751 0.743 0.532 0.678 0.268 0.788 0.712 0.436 0.879 0.805 0.330 0.517 0.852 0.461 0.382 0.254 0.744 0.779 0.906 0.691 0.814 0.385 0.484 0.852 0.700 0.727 0.564 0.433 0.575 0.229 0.757 0.624 0.778 0.637 0.216 0.646 0.614 0.893 0.704 0.908 0.712 0.847 0.416 0.617 0.700 0.756 0.862 0.878 0.866 0.933 0.789 0.916 0.874 0.741 0.852 0.562 0.842 0.932 0.531 0.732 0.819 0.685 0.757 0.736 0.767 0.507 0.675 0.311 0.802 0.771 0.423 0.871 0.778 0.536 0.482 0.865 0.461 0.450 0.240 0.720 0.806 0.926 0.733 0.866 0.395 0.546 0.691 0.625 0.751 0.627 0.472 0.759 0.245 0.767 0.653 0.791 0.642 0.225 0.621 0.657 0.880 0.759 0.925 0.728 0.806 0.326 0.622 0.705 0.762 0.884 0.788 0.903 0.920 0.754 0.903 0.878 0.765 0.880 0.516 0.786 0.934 0.563 0.825 0.904 0.647 0.659 0.852 0.752 0.531 0.635 0.367 0.750 0.662 0.252 0.902 0.785 0.296 0.537 0.786 0.488 0.368 0.553 0.749 0.704 0.895 0.663 0.730 0.356 0.720 0.898 0.604 0.734 0.575 0.429 0.588 0.201 0.761 0.611 0.645 0.631 0.165 0.668 0.596 0.899 0.697 0.868 0.710 0.861 0.397 0.589 0.645 0.673 0.828 0.859 0.853 0.895 0.779 0.857 0.918 0.674 0.914 0.478 0.783 0.928 0.552 0.840 0.905 0.696 0.758 0.901 0.798 0.624 0.697 0.465 0.821 0.716 0.347 0.933 0.839 0.365 0.585 0.847 0.534 0.471 0.634 0.811 0.767 0.932 0.734 0.834 0.447 0.764 0.912 0.661 0.805 0.630 0.517 0.650 0.225 0.824 0.658 0.694 0.703 0.199 0.710 0.656 0.936 0.762 0.911 0.742 0.899 0.390 0.676 0.711 0.792 0.852 0.885 0.871 0.930 0.837 0.894 0.949 0.749 0.943 0.583 0.841 0.950 0.606 0.821 0.835 0.726 0.782 0.846 0.790 0.622 0.682 0.497 0.830 0.748 0.407 0.929 0.861 0.439 0.601 0.859 0.542 0.516 0.809 0.857 0.785 0.898 0.764 0.804 0.468 0.691 0.927 0.676 0.854 0.685 0.523 0.494 0.213 0.859 0.685 0.704 0.729 0.317 0.642 0.729 0.924 0.794 0.909 0.758 0.900 0.411 0.675 0.658 0.782 0.882 0.971 0.879 0.921 0.848 0.881 0.947 0.777 0.932 0.632 0.864 0.972 0.637 Average 0.618 0.678 0.684 0.666 0.722 0.735 false_detection (a) ground transportation station (b) helipad Figure 10: Sample images of the interface used to more precisely localize objects within an image. In each example, a green dot is placed near the center of the pertinent object. Users are able to draw a bounding box by clicking and dragging. Instructions at the top of each example inform the user how to use the interface and also provide any category-specific instructions that may be relevant. Comments regarding issues such as clouds or object misclassification can be entered near the bottom of the page before submitting an annotation. CNN-IM-1 CNN-IM LSTM-IM Table 4: F1 scores for different approaches on an individual image basis. Color formatting was applied to each column independently. The average values shown at the bottom of the table are calculated without the false detection scores. CNN-I-1 and CNN-IM-1 make predictions for each individual view. All other methods repeat their prediction over the full sequence. to keep in mind that the images for each category in the full dataset vary in quality, have different weather conditions (e.g., snow cover), contain drastically different context (e.g., desert vs. urban), have different levels of difficulty for recognition, and other variations. presented in the main paper, smaller boxes were given more context than larger boxes, and therefore it may appear for some categories with smaller sizes (e.g., smoke stacks) that there is a lot more context than expected. It is important 12 airport airport hangar burial site factory or powerplant airport terminal car dealership construction site false detection helipad fire station hospital amusement park aquaculture crop field dam flooded road impoverished settlement fountain interchange lake or pond archaeological site debris or rubble barn border checkpoint educational institution gas station golf course lighthouse military facility electric substation ground transportation station multi-unit residential nuclear powerplant office building oil or gas facility park parking lot or garage place of worship police station port prison race track railway bridge recreational facility road bridge runway shipyard shopping mall single-unit residential toll booth smokestack tower solar farm tunnel opening space facility stadium waste disposal storage tank water treatment facility Figure 11: One example per category in fMoW. 13 surface mine wind farm swimming pool zoo References [17] I. Krasin, T. Duerig, N. Alldrin, V. Ferrari, S. Abu-El-Haija, A. Kuznetsova, H. Rom, J. Uijlings, S. Popov, A. Veit, S. Belongie, V. Gomes, A. Gupta, C. Sun, G. Chechik, D. Cai, Z. Feng, D. Narayanan, and K. Murphy. Openimages: A public dataset for large-scale multi-label and multiclass image classification. Dataset available from https://github.com/openimages, 2017. [1] S. Abu-El-Haija, N. Kothari, J. Lee, P. Natsev, G. Toderici, B. Varadarajan, and S. Vijayanarasimhan. YouTube-8M: A Large-Scale Video Classification Benchmark. arXiv preprint arXiv:1609.08675, 2016. [2] S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. Lawrence Zitnick, and D. Parikh. VQA: Visual Question Answering. In ICCV, 2015. [3] A. Chang, A. Dai, T. Funkhouser, M. Halber, M. Nießner, M. Savva, S. Song, A. Zeng, and Y. Zhang. Matterport3d: Learning from rgbd data in indoor environments. arXiv preprint arXiv:1709.06158, 2017. [4] G. Cheng, J. Han, and X. Lu. Remote Sensing Image Scene Classification: Benchmark and State of the Art. Proc. IEEE, 2017. [5] N. DigitalGlobe, CosmiQ Works. Spacenet. Dataset available from https://aws.amazon.com/publicdatasets/spacenet/, 2016. [6] 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. [7] M. Everingham, S. A. Eslami, L. Van Gool, C. K. Williams, J. Winn, and A. Zisserman. The Pascal Visual Object Classes Challenge: A Retrospective. IJCV, 2015. [8] L. Fei-Fei, R. Fergus, and P. Perona. One-Shot Learning of Object Categories. PAMI, 2006. [9] G. Griffin, A. Holub, and P. Perona. Caltech-256 Object Category Dataset. 2007. [10] D. Harwath and J. R. Glass. Learning Word-Like Units from Joint Audio-Visual Analysis. ACL, 2017. [11] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR, 2016. [12] G. Huang, Z. Liu, L. van der Maaten, and K. Q. Weinberger. Densely Connected Convolutional Networks. CVPR, 2017. [13] N. Jean, M. Burke, M. Xie, W. M. Davis, D. B. Lobell, and S. Ermon. Combining Satellite Imagery and Machine Learning to Predict Poverty. Science, 2016. [14] A. Karpathy and L. Fei-Fei. Deep Visual-Semantic Alignments for Generating Image Descriptions. In CVPR, 2015. [15] A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, and L. Fei-Fei. Large-scale Video Classification with Convolutional Neural Networks. In CVPR, 2014. [16] D. Kingma and J. Ba. Adam: A Method for Stochastic Optimization. ICLR, 2014. [18] T.-Y. Lin, M. Maire, S. Belongie, J. Hays, P. Perona, D. Ramanan, P. Dollár, and C. L. Zitnick. Microsoft COCO: Common Objects in Context. In ECCV, 2014. [19] D. Marmanis, M. Datcu, T. Esch, and U. Stilla. Deep Learning Earth Observation Classification Using ImageNet Pretrained Networks. GRSL, 2016. [20] A. Miech, I. Laptev, and J. Sivic. Learnable pooling with Context Gating for video classification. arXiv preprint arXiv:1706.06905, 2017. [21] Y. Pan, T. Mei, T. Yao, H. Li, and Y. Rui. Jointly Modeling Embedding and Translation to Bridge Video and Language. In CVPR, 2016. [22] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 2015. [23] K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. [24] K. Tang, M. Paluri, L. Fei-Fei, R. Fergus, and L. Bourdev. Improving Image Classification with Location Context. In ICCV, 2015. [25] S. Wang, M. Bai, G. Mattyus, H. Chu, W. Luo, B. Yang, J. Liang, J. Cheverie, S. Fidler, and R. Urtasun. Torontocity: Seeing the world with a million eyes. ICCV, 2017. [26] T. Weyand, I. Kostrikov, and J. Philbin. Planet-photo geolocation with convolutional neural networks. In ECCV, 2016. [27] G.-S. Xia, J. Hu, F. Hu, B. Shi, X. Bai, Y. Zhong, L. Zhang, and X. Lu. AID: A Benchmark Data Set for Performance Evaluation of Aerial Scene Classification. TGRS, 2017. [28] G.-S. Xia, X.-Y. Tong, F. Hu, Y. Zhong, M. Datcu, and L. Zhang. Exploiting deep features for remote sensing image retrieval: A systematic investigation. arXiv preprint arXiv:1707.07321, 2017. [29] Y. Yang and S. Newsam. Bag-Of-Visual-Words and Spatial Extensions for Land-Use Classification. In ACM GIS, 2010. [30] J. Yue-Hei Ng, M. Hausknecht, S. Vijayanarasimhan, O. Vinyals, R. Monga, and G. Toderici. Beyond Short 14 Snippets: Deep Networks for Video Classification. In CVPR, 2015. 15 

MPISE: Symbolic Execution of MPI Programs Xianjin Fu1,2 , Zhenbang Chen1,2 , Yufeng Zhang1,2 , Chun Huang1,2 ,Wei Dong1 and Ji Wang1,2 1 arXiv:1403.4813v3 [cs.DC] 15 Sep 2014 2 School of Computer, National University of Defense Technology, P. R. China Science and Technology on Parallel and Distributed Processing Laboratory, National University of Defense Technology, P. R. China {xianjinfu,zbchen,yfzhang,chunhuang,wdong,wj}@nudt.edu.cn Abstract. Message Passing Interfaces (MPI) plays an important role in parallel computing. Many parallel applications are implemented as MPI programs. The existing methods of bug detection for MPI programs have the shortage of providing both input and non-determinism coverage, leading to missed bugs. In this paper, we employ symbolic execution to ensure the input coverage, and propose an on-the-fly schedule algorithm to reduce the interleaving explorations for non-determinism coverage, while ensuring the soundness and completeness. We have implemented our approach as a tool, called MPISE, which can automatically detect the deadlock and runtime bugs in MPI programs. The results of the experiments on benchmark programs and real world MPI programs indicate that MPISE finds bugs effectively and efficiently. In addition, our tool also provides diagnostic information and replay mechanism to help understanding bugs. 1 Introduction In the past decades, Message Passing Interface (MPI) [19] has become the de facto standard programming model for parallel programs, especially in the filed of high performance computing. A significant part of parallel programs were written using MPI, and many of them are developed in dozens of person-years [14]. Currently, the developers of MPI programs usually use traditional methods to improve the confidence of the programs, such as traditional testing and debugging [2][1]. In practice, developers may waste a lot of time in testing, but only a small part of behavior of the program is explored. MPI programs have the common features of concurrent systems, including non-determinism, possibility of deadlock, etc. These features make the shortage of testing in coverage guarantee more severe. Usually, an MPI program will be run as several individual processes. The nature of non-determinism makes the result of an MPI program depend on the execution order of the statements in different processes. That is to say, an MPI program may behave differently with a same input on different executions. Hence, sometimes it is harder to find the bugs in an MPI program by a specific program execution. To improve the reliability of MPI programs, many techniques have been proposed. Basically, we can divide the existing work into two categories: static analysis methods [21][22][23] and dynamic analysis methods [27] [24]. A static method analyzes an MPI program without actually running it. The analysis can be carried out on code level [22] or model level [21]. Usually, a static method needs to make an abstraction of the MPI program under analysis [21][23]. Therefore, many static methods suffer the false alarm problem. Dynamic methods, such as testing and runtime verification, need to run the analyzed MPI programs and utilize the runtime information to do correctness checking [17][20][26], online verification [25][27], debugging [1][4], etc. Traditional testing methods work efficiently in practice by checking the correctness of a run under a given test harness. However, testing methods cannot guarantee the coverage on non-determinism even after many runs of a same program input. Other dynamic analysis methods, such as ISP [25], provide the coverage guarantee over the space of non-determinism and scale well, but they are still limited to program inputs. While TASS [22] employs symbolic execution and model checking to verify MPI programs, it only works on small programs due to the limited support of runtime library models. In this paper, we use symbolic execution to reason about all the inputs and try to guarantee the coverage on both input and non-determinism. We symbolically execute the statements in each process of an MPI program to find input-related bugs, especially runtime errors and deadlocks. For the non-determinism brought by the concurrent features, we use an on-the-fly scheduler to reduce the state space to be explored in the analysis, while ensuring the soundness and completeness. Specially, to handle the non-determinism resulted from the wildcard receives in MPI programs, we dynamically match the source of a wildcard receive into all the possible specific sources in a lazy style, which avoids the problem of missing bugs. Furthermore, unlike the symbolic execution plus model checking method in [22], which uses an MPI model to simulate the runtime behaviors of MPI library, we use a true MPI library as the model, which enables us to analyze real-world MPI programs. To summarize, our paper has the following main contributions: firstly, we propose an on-the-fly scheduling algorithm, which can reduce unnecessary interleaving explorations while ensuring the soundness and completeness; secondly, when attacking the non-determinism caused by wildcard receives, we propose a technique, called lazy matching, to avoid blindly matching each process as the source of a wildcard receive, which may lead to false positives; finally, we have implemented our approach in a tool called MPISE, and conducted extensive experiments to justify its effectiveness and efficiency in finding bugs in MPI programs. The rest of this paper is organized as follows. Section 2 introduces the background and shows the basic idea of MPISE by motivating examples. Section 3 describes the details of the algorithms implemented in MPISE. Section 4 explains our implementation based on Cloud9 and shows the experimental results. Finally, Sections 5 discusses the related work and the conclusion is drawn in Section 6. 2 Background and Motivating Example In this section, we briefly describe symbolic execution and the scope of the MPI APIs we are concerned with, then show how our algorithm works by motivating examples. 2.1 Symbolic execution Symbolic execution [16] is a SAT/SMT based program analysis technique originally introduced in the 1970s. With the significant improvement in SAT/SMT techniques and computing power, symbolic execution draws renewed interests recently. The main idea is, rather than using concrete values, symbolic execution uses symbolic values as input values, and keeps tracking the results of numerical operations on symbolic values. Hence, the result of a program under symbolic execution will be symbolic expressions. Most importantly, symbolic execution uses a constraint of symbolic values, called path condition (PC), to represent a path of a program. At the beginning, the path condition is true. When encountering a branch statement, symbolic execution explores both directions of the branch. For exploring one direction, symbolic execution records (i.e., conjunction) the condition cond corresponding to the direction in PC and queries an underlying solver with P C ∧ cond to decide whether this direction is feasible. If the answer is yes, symbolic execution will continue to execute the statements following the direction, and PC is update to be P C ∧ cond; otherwise, it means the direction is infeasible, thus symbolic execution backtracks to the branch statement, and starts to explore the other direction. The selection of which direction of a branch to explore first can be random or according to some heuristics. Once symbolic execution reaches the end of a program, the accumulated PC represents the constraints that the inputs need to satisfy to drive the program to the explored path. Therefore, we can consider symbolic execution as a function that computes a set of PCs for a program. Naturally, we can use the PCs of the program to do automatic test generation [8][9], bug finding [8][13], verification [12], etc. According to the before explanation, symbolic execution is a precise program analysis technique, because each PC represents a real feasible path of the program under analysis. Therefore, when used for bug finding, symbolic execution does not suffer from the false alarm problem, and the bugs found are real bugs. Whereas, one of the major challenge symbolic execution faces is path space exploration, which is theoretically exponential with the number the branches in the program. 2.2 MPI Programs An MPI program is a sequential program in which some MPI APIs are used. The running of an MPI program usually consists of a number of parallel processes, say P0 , P1 , ..., Pn−1 , that communicate via message passings based on MPI APIs and the supporting platform. The message passing operators we consider in this paper include: – Send(dest) -send a message to Pdest (dest = 0, . . . , n − 1), which is the destination process of the Send operation. Note that only synchronous communications are considered in this paper, so this operation blocks until a matching receive has been posted. – Recv(src) -receive a message from Psrc (src = 0, . . . , n − 1, AN Y ), which is the source process of the Recv operation. Note that the src can take the wildcard value “ANY”, which means this Recv operation expects messages from any process. Because Send and Recv are synchronous, a Send/Recv that fails to match with a corresponding Recv/Send would result in a deadlock. – Barrier() -synchronization of all processes, which means the statements of any process should not be issued past this barrier until all the processes are synchronized. Therefor, an MPI program is expected to eventually reach such a state that all the processes reach their barrier calls. If this does not hold, there would be a deadlock. The preceding three MPI operations are the most important operations we consider in this paper. Actually, they cover the most frequently used synchronous communications in MPI programs. 2.3 Motivating Examples Usually, an MPI program is fed with inputs to perform a computational task, and the bugs of the program may be input-dependent. On the other side, due to the non-determinism feature, even with same inputs, one may find that bugs occur “sometimes”. Consider the MPI program in Fig 1, if the program runs with an input that is not equal to ‘a’, the three processes will finish normally with two matched Send and Recv, as indicated by Fig 2(a). However, if the program is fed with the input ‘a’, a deadlock may happen, in case that P roc1 receives a message from P roc2 first by a wildcard receive, and then it waits a message from P roc2 and P roc0 also expects P roce1 to receive a message, as shown in Fig 2(c). Therefore, tools that do not provide input space coverage would surely fail to detect this bug if the program is not fed with ‘a’. Even if one is lucky enough to run the program with ‘a’, we may still fail to detect the bug if the wildcard receive is matched with P roc0 , e.g., the case in Fig 2(b). Thus, for detecting deadlock bugs, we need both input coverage and nondeterminism coverage guarantee. The basic idea of our method is: we employ symbolic execution to cover all possible inputs, and explore all the possible matches of a wildcard receive by matching it to any possible source. To be more detailed, since we only symbolically execute one process at a time, we need to decide the exploration order of the processes. Usually, each process of an MPI program has a rank, we always start from the smallest ranked process and switch to another process until the current process needs synchronization, such as sending or receiving a message. Thus, the switches during symbolic execution happen on-the-fly. Specifically, things become more complex when encountering a Recv(ANY) statement, where we need to delay the selection of 1 2 3 i n t main ( i n t argc , char ∗∗ argv ) { i n t x , y , myrank ; MPI Comm comm = MPI COMM WORLD; 4 M P I I n i t (& argc , &argv ) ; MPI Comm rank (comm, &myrank ) ; i f ( myrank==0) { x = 0; MPI Ssend(&x , 1 , MPI INT , 1 , 9 9 , comm) ; } e l s e i f ( myrank==1) { i f ( argv [ 1 ] [ 0 ] ! = ’ a ’ ) // a r g c i s e x a c t l y 2 MPI Recv(&x , 1 , MPI INT , 0 , 9 9 , comm, NULL) ; else MPI Recv(&x , 1 , MPI INT , MPI ANY SOURCE, 9 9 , comm, NULL) ; 5 6 7 8 9 10 11 12 13 14 15 16 MPI Recv(&y , 1 , MPI INT , 2 , 9 9 , comm, NULL) ; } e l s e i f ( myrank==2){ x = 20; MPI Ssend(&x , 1 , MPI INT , 1 , 9 9 , comm) ; } MPI Finalize () ; return 0 ; 17 18 19 20 21 22 23 24 } Fig. 1. Example showing the need for both input and non-determinism coverage the corresponding sending process until all the possible sending statements are encountered. For the MPI program in Fig 1 run in three processes, we start from P roc0 , i.e., the process with rank 0. When executing to line 9, a Send is encountered, which means a synchronization is needed. From the send statement, we know it needs to send a message to P roc1 . Thus, we switch to P roc1 and do symbolic execution from the beginning. When the branch statement at line 12 is encountered, and argv[1][0] is symbolic (we suppose it has a symbolic value X), the condition X 6= ‘a’ is added to the path condition of the true side and its negation to the false side. We mark here as a backtrack point and has two paths to follow, which are explained as follows: X 6= ‘a’: If we explore the true side first, the path condition, i.e., X 6= ‘a’, is fed to the solver to check the feasibility of the path. Apparently, the solver will answer yes, thus we can continue the symbolic execution of P roc1 . Then, Recv(0) is meet and it is exactly matched with the send in P roc0 . Therefore, both processes advance, and P roc0 ends while P roc1 goes to Recv(2). In a same manner, P roc1 gets asleep, we switch to P roc2 . Again the two operations matches, the whole execution will end normally, as shown in Fig 2(a). X == ‘a’: This side is also feasible. The symbolic execution of P roc1 will encounter Recv(ANY), and switches to P roc2 . After executing the Send at Proc0 Send(1) Recv(2) Proc1 Recv(0) Proc2 Send(1) Send(1) Send(1) Recv(ANY) Recv(2) Send(1) (a) X6=a Recv(ANY) Recv(2) Send(1) (b) X==a and wildcard (c) X==a and wildcard receive matches with receive matches with P roc0 P roc2 Fig. 2. Three cases of the program in Fig 1 Line 20, there is no process that can be switched to. All the possible sending processes of the Recv(ANY) in P roc1 are determined. Thus, now we begin to handle the Recv(ANY) by matching it with each possible sending. Suppose we match the Recv(ANY) with the Send of P roc0 , we continue to execute P roc1 . We encounter another Recv at Line 17 that expects to receive a message from P roc2 , then P roc1 and P roc2 advance, and finally the whole execution ends normally, as indicated by Fig 2(b). On the other hand, if the Recv(ANY) is matched with the Send of P roc2 , when encountering the Recv in P roc1 , symbolic execution will switch to P roc2 , but P roc2 has finished. Then, P roc0 and P roc1 can not terminated. Hence, a deadlock is detected, as shown in Fig 2(c). In summary, the deadlock, which may happen in the program in Fig 1 when run in three processes, can only be encountered when the input starts with ‘a’ and the Recv(ANY) in the second process is matched with the Send in the third process. By using our approach, MPISE can detect it automatically. The details of our symbolic execution algorithms will be introduced in the next section. 3 Symbolic execution algorithms In this section, we will introduce a general framework for symbolic execution of MPI programs first, and then present a scheduling algorithm during the symbolic execution. Furthermore, to attack the non-determinism brought by wildcard receives, we will present a refined scheduling method, which can ensure the exploration of all the possible matches of a wildcard receive. To start with, we introduce some notions first. When symbolically executing a sequential program, the symbolic executor keeps tracking of states, each of which consists of a map that records the symbolic/concrete value of each variable, a program counter and a path condition. For an MPI program, a state during symbolic execution is composed by the states of the parallel processes. A state s0 is said to be the successor of a state s, if s0 can be obtained by symbolically executing a statement in one process. With the notion of state, we define a deadlock to be the state that has no successor, and at which there is at least one process that does not terminate. Recall that symbolic execution will do state forking when encountering a branch statement. For MPI programs, in addition to branch statements, the concurrency nature can also result in state forking. Theoretically, for the current state, if there are more than one process, say n, that can be executed, there are n possible successor states. Hence, besides the number of branch statements, the number of parallel processes also makes the path space increase exponentially. Algorithm 1 presents a general framework for symbolic execution of MPI programs. Algorithm 1: Symbolic Execution Framework 1 2 3 4 5 6 7 8 9 10 Search(M P , n, slist){ Active = {P0 , . . . , Pn } ; Inactive = ∅; NextProcCandidate= -1; worklist = {initial state}; while (worklist is not empty) do s = pick next state; p = Scheduler (s); if p 6= null then stmt = the next statement of p; SE(s, p, stmt); } Basically, the symbolic execution procedure is a worklist-based algorithm. The input consists of an MPI program, the number of the parallel running processes and the symbolic variables. At the beginning, only the initial state, i.e., composed by the initial states of all the processes, is contained in the worklist. Then, new states can be derived from the current state and put into the worklist. State exploration is done if there is no state in the worklist. Because of state forking, we usually have a way for space exploration, such as depth first search (DFS) and breadth first search (BFS). Clearly, it is hard or even impossible to explore the whole path space. In fact, for the state forking introduced by the concurrent feature, sometimes there is no need to add all the possible successor states to the worklist, which can still capture the behavior of the program precisely in our context. Hence, different from the usual symbolic execution algorithm, in our algorithm, we first select a state from worklist (Line 5, where a search algorithm can be used), then we make a decision (Line 6, the details of which will be given in Section 3.1) of which process is scheduled for symbolic execution. Finally, we symbolically execute the next statement of the scheduled process, in which some new states may be generated. Basically, for the non-communication statements in an MPI program, the symbolic execution semantics is same as usual. In the following of this section, we will concentrate on explaining the scheduling of the processes and the handling of the communication operations. 3.1 On-the-fly scheduling With the general framework in Algorithm 1, we introduce our scheduler here, aiming to avoid naively exploring the interleavings of all the processes. For each process of an MPI program during symbolic execution, the process is active if it is not asleep. Usually, we make a process asleep when the process needs to communicate but the corresponding process is not ready, whose details will be given in Algorithm 3. We maintain the current status of each process via two sets: Active and Inactive. At beginning, all the processes are contained in Active. If a process is made to be asleep, it will be removed from Active and added to Inactive. Because we schedule the processes on-the-fly, we use a global variable N extP rocCandidate to denote the index of the next process to symbolically execute. The following Algorithm 2 gives how to do scheduling. Algorithm 2: Scheduling the Next Process for Symbolic Execution 1 2 3 4 5 6 7 8 9 10 Scheduler(s){ if N extP rocCandidate! = −1 and P rocN extP rocCandidate is active then N ext = N extP rocCandidate; N extP rocCandidate = −1; return P rocN ext ; else if Active 6= ∅ then return the process p0 with the smallest rank in Active; if Inactive 6= ∅ then Report Deadlock; } First, we check whether there is a next process that needs to be executed and is also active. If there exists one, the process identified by N extP rocCandidate will be selected, and the next process global variable is reset (Line 1∼5); otherwise, we return the active process with the smallest rank if exists (Line 6∼7). Finally, if there is no active process that can be scheduled, and the Inactive set is non-empty, i.e., there exists at least one process that does not terminate, we report that a deadlock is found (Line 8∼9). Now, we explain how to symbolically execute each statement. In Algorithm 3, we mainly give the handling for MPI APIs considered in this paper. The local statements in each process do not influence the other processes, and the symbolic execution of basic statements, such as assignment and branch, is the same with the traditional approach [8]. Hence, the symbolic execution of local statements is omitted for the sake of space. In Algorithm 3, Advance(S) denotes the procedure in which the program counter of each process in S will be advanced, and Match(p, q) denotes the procedure in which the synchronization between p and q happens, i.e., the receiver receives the data sent by the sender, and the program counters of p and q will be both advanced. If a Send(dest) is encountered and there is a process in Inactive that matches the statement, we move that process from Inactive to Active (Line 5) and advance the two processes (Line 6). If there is no process that can receive the message, we add this process into Inactive set (Line 8), and switch to the destination process of the send operation (Line 9). The execution of a receive operation is similar, except that when the receive operation is a wildcard receive, we make the current process asleep (the reason will be explained in Section 3.2). Algorithm 3: Symbolic Execution of a Statement 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 SE(s, p, stmt){ switch kindof (stmt) do case Send (dest) if stmt has a matched process q ∈ Inactive then Inactive = Inactive \ {q}; Active = Active ∪ {q}; Match(p, q); else Inactive = Inactive ∪ {p}; N extP rocCandidate = dest; return; case Recv (src) if src != MPI ANY SOURCE then if stmt has a matched process q ∈ Inactive then Inactive = Inactive \ {q}; Active = Active ∪ {q}; Match(p, q); else Inactive = Inactive ∪ {p}; N extP rocCandidate = src; 20 else Inactive = Inactive ∪ {p}; 21 return; 19 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 case Barrier if mcb == ∅ then mcb = {P0 , . . . , Pn } \ {p}; Inactive = Inactive ∪ {p}; else mcb = mcb \ {p}; if mcb == ∅ then Advance({P0 , . . . , Pn }); Inactive = ∅; Active = {P0 , . . . , Pn }; else Inactive = Inactive ∪ {p}; return; case Exit Active = Active \ {p}; return; Advance({p}); } For handling barriers, we use a global variable mcb to denote the rest processes that need to reach a barrier for a synchronization. When a barrier state- ment is encountered, if mcb is empty, we initialize mcb to be the set containing the rest processes (Line 24) and add the current process into Inactive (Line 25). If mcb is not empty, we remove the current process from mcb . Then, if mcb is empty, i.e., all the processes have reached a barrier, we can advance all the processes (Line 29) and make all the processes active (Line 30); otherwise, we add the current process into Inactive set (Line 32). When encountering an Exit statement, which means the current process terminates, we remove the current process from Active (Line 35). In summary, according to the two algorithms, the symbolic execution process will continue to execute the active process with the smallest rank until a preemption happens caused by an unmatched MPI operation. From a state in symbolic execution, we do not put all the possible states into the worklist, but only the states generated by the current process. This is the reason why we call it on-the-fly scheduling. Actually, we only explore a sub space of the whole program path space, but without sacrificing the ability of finding deadlock bugs. The correctness of our on-the-fly scheduling algorithms is guaranteed by the following theorem, whose proof is given in appendix. Theorem 1. Given a path of an MPI program from the initial state to a deadlocked state, there exists a path from the initial state to the same deadlocked state obtained by the on-the-fly scheduling. And vice versa. 3.2 Lazy matching algorithm Note that so far, we do not treat wildcard receives. Actually, wildcard receives are one of the major reasons of non-determinism. Clearly, we cannot blindly rewrite a wildcard receive. For example, in Fig 3(a), if we force the wildcard receive in P roc1 to receive from P roc2 , a deadlock will be reported, which actually will not happen. In addition, if we rewrite a wildcard receive immediately when we find a possible match, we still may miss bugs. As shown in Fig 3(b), if we match the wildcard receive in P roc0 with the send in P roc1 , the whole symbolic execution will terminate successfully, thus a deadlock, which will appear when the wildcard receive is matched with the send in P roc2 , is missed. P roc0 P roc1 P roc2 Send(1) Recv(ANY) local statements (a) Blind rewriting of a wildcard receive P roc0 P roc1 P roc2 Recv(ANY) ; Recv(2) Send(0) Send(0) (b) Eager rewriting of a wildcard receive Fig. 3. Rewriting of a wildcard statement To solve this problem, we employ a lazy style approach instead of an eager one. That is, we delay the selection of the send candidate of a wildcard receive until the whole symbolic execution procedure blocks. To be detailed, when the symbolic execution encounters a wildcard receive, we would make the current process asleep (Line 20 in Algorithm 3), waiting for all possible senders. When a matched send is found, the current process will also be made asleep, and we switch to the next active process. When there is no process that can be scheduled, i.e., all the processes are in Inactive, we match the wildcard receive to each possible matched send by forking a successor state for each one. Thus, Algorithm 2 needs to be refined to handle wildcard receives. The refined parts are given as follows. Algorithm 4: Refined Scheduling for Handling Wildcard Receives 1 2 3 4 5 6 7 8 9 Scheduler(s){ ... if Inactive 6= ∅ then if Exists a Recv(ANY) process in Inactive then P S = Inactive; for each Recv(ANY) process p ∈ Inactive do for each matched process q ∈ Inactive of p do Inactive = P S \ {p, q}; Active = {p, q}; AddState(s, p, q); return null; 10 else Report Deadlock; 11 12 13 } For each process encountering a wildcard receive in Inactive, we add a new state for each of its matched sender processes (Line 9). The AddState(s, p, q) denotes a procedure that does the synchronization between p and q, advances both p and q, and adds the new state to the worklist. Thus, we are exploring all the possible cases of a wildcard receive. If there are multiple Recv(ANY) processes, we are interleaving the matches of all the processes. The example in Fig 4 demonstrates this situation. When all the processes are asleep, if we match the Recv(ANY) in P roc1 with the send in P roc0 first, no deadlock will be detected; otherwise, if we match the Recv(ANY) in P roc2 with the send in P roc3 first, a deadlock will be detected. P roc0 P roc1 P roc2 P roc3 Send(to:1) Recv(from:ANY) Recv(from:ANY) Send(to:2) Recv(from:3) Send(to:1) Fig. 4. Multiple wildcard receives Therefore, after considering wildcard receives, the matches of different wildcard receives are not independent. We deals with this problem by naively interleaving the match orders of wildcard receives. This leads to redundant interleavings, but dose not miss interleaving-specific deadlocks. The optimization is left to our future work. The proof of correctness of our handling for wildcard receives is provided in appendix. 4 4.1 Implementation and Experiments Implementation We have implemented our approach as a tool, called MPISE, based on Cloud9 [7], which is a distributed symbolic executor for C programs. Cloud9 enhances KLEE [8] by enabling the support of most POSIX interfaces and parallelism. The architecture of MPISE is shown in Fig 5. Process number and other arguments C-MPI programs LLVM-GCC Compiler LLVM bytecode MPISE (executor,scheduler, test generator) Hooked TOMPI lib （ LLVM bytecode） MPISE (replayer) Test cases Deadlock, assert failure Fig. 5. The architecture of MPISE. The target MPI programs written in C is fed into LLVM-GCC compiler to obtain the LLVM bytecode, which will be linked with a pre-compiled library, i.e., TOMPI [11], as well as the POSIX runtime library. Then, the linked executable program will be symbolically executed. Basically, TOMPI is a platform that uses multi-threads to simulate the running of an MPI program. TOMPI provides a subset of MPI interfaces, which contains all the MPI APIs we consider in this paper. An MPI program can be compiled and linked with TOMPI libraries to generate a multi-thread executable, which is supposed to generate the same output as that of the parallel running of the MPI program. Hence, we use TOMPI as the underlying MPI library. By using TOMPI, we can use the support for concurrency in Cloud9 to explore the path space of an MPI program run with a specific number of processes. When a path ends or a deadlock is detected, MPISE records all the information of the path, including the input, the orders of message passings, etc. For each path, we generate a corresponding test case, based on which one can use replayer to reproduce a concrete path. Compared with Cloud9, our implementation of MPISE consists of the following new features: – New scheduler. Cloud9 employs a none-preemptive scheduler, i.e., a process would keep being executed until it gives up, such as encountering an explicit preemption call or process exit. Clearly, we need a new scheduler for MPISE. We have implemented our on-the-fly scheduler that can schedule the MPI processes according to the algorithms in Sections 3.1 & 3.2. – Environment support for MPI APIs. Cloud9 does not “recognize” MPI operations, while MPISE makes the symbolic engine know MPI operations based on TOMPI, including MPI Send, MPI Ssend, MPI Recv, MPI Barrier, etc. The message passing APIs are dealt specially for scheduling, while other MPI APIs are treated as normal function callings. – Enhanced Replay. MPISE can replay each generated test case of an MPI program, which can help user to diagnosis bugs such as deadlock and assertion failure. The replayer of MPISE extends the replayer component of Cloud9 by using the on-the-fly schedule when replaying a test case. During replaying, the replayer uses the recorded input to feed the program, and follows the recorded schedules to schedule the processes. – Enhanced POSIX model. MPISE heavily depends on the library models it uses. However, the POSIX model provided by Cloud9 is not sufficient for us to symbolically execute MPI programs. The reason is we need to maintain a process specific data area for each process when symbolically executing each process. Because we use multi-thread programs to simulate the behaviour of MPI programs, we have improved the mechanism for multi-thread programs in Cloud9 to support maintaining thread specific data. 4.2 Experimental evaluation We have conducted extensive experiments to validate the effectiveness and scalability of MPISE. All the experiments were conducted on a Linux server with 32 cores and 250 GB memory. Using MPISE to analyze the programs in the Umpire test suite [25], we have successfully analyzed 33 programs, i.e., either no deadlock detected or detecting a deadlock as expected. The Umpire test case are input-independent, i.e. the inputs have nothing to do with whether a deadlock happens. Hence, we conduct the experiments on the programs with input-dependent deadlocks. The conducted test cases mainly cover two typical situations of deadlock [18]: point-to-point ones and collective ones. Point-to-point deadlocks are usually caused by (1). a send/receive routine has no corresponding receive/send routine; (2). a send-receive cycle may exist due to the improper usage of send and receive. Collective deadlocks are typically caused by (1). missed collective routines (such as Barrier); (2). improper ordering of some point-to-point or/and collective routines. In our experiments, we also use ISP and TASS to analyze the programs. Fig 6 displays the experimental results, including those of ISP and TASS. In Fig 6, we divide the experimental results into two categories: input independent programs and input dependent ones. For each category, we select programs that can deadlock caused by different reasons, including head to head receive, wait all, receive any, etc. For each input dependent program, we generate the input randomly when analyzing the program with ISP, and analyze the program for 10 times, expecting to detect a deadlock. The execution time of analyzing each input dependent program with ISP is the average time of the 10 times of runnings. According to the experimental results, we can conclude as follows: MPISE can detect the deadlock in all the programs. ISP misses the deadlock for all the input dependent programs. TASS fails to analyze most of programs. Program Input Independent Input Dependent anysrc-deadlock.c basic-deadlock.c collect-misorder.c waitall-deadlock.c bcast-deadlock.c complex-deadlock.c waitall-deadlock2.c barrier-deadlock.c head-to-head.c rr-deadlock.c recv-any-deadlock.c cond-bcast.c collect-misorder.c waitall-deadlock3.c ISP TASS Result Time(s) Result Time(s) Deadlock 0.126 Fail 1.299 Deadlock 0.022 Fail 1.227 Deadlock 0.022 Fail 0.424 Deadlock 0.024 Fail 1.349 Deadlock 0.021 Fail 0.493 Deadlock 0.023 Fail 1.323 Deadlock 0.024 Fail 1.349 No 0.061 Fail 0.863 No 0.022 Fail 1.542 No 0.022 Fail 1.244 No 0.022 Deadlock 1.705 No 0.021 No 1.410 No 0.023 Deadlock 1.682 No 0.104 Fail 1.314 MPISE Result Time(s) Deadlock 1.59 Deadlock 1.46 Deadlock 1.48 Deadlock 1.49 Deadlock 1.40 Deadlock 1.46 Deadlock 1.48 Deadlock 1.71 Deadlock 1.67 Deadlock 1.67 Deadlock 1.70 Deadlock 1.63 Deadlock 1.85 Deadlock 1.78 Fig. 6. Experimental results Thus, MPISE outperforms ISP and TASS for all the programs in Fig 6. The reason is, MPISE uses symbolic execution to have an input coverage guarantee, and the scheduling algorithms ensures that any deadlock caused by the MPI operations considered in this paper will not be missed. In addition, we utilize TOMPI and Cloud9 to provide a better environment support for analyzing MPI programs. The reason of the common failure of TASS is that TASS does not support many APIs, such as fflush(stdout) of POSIX and MPI Get Processor Name of MPI, and needs manually modifying the analyzed programs. For each program, the analysis time using MPISE is longer than that of using ISP or TASS. The reason is two fold: firstly, we need to symbolically execute the bytecodes including those of the underlying MPI library, i.e., TOMPI. For example, for the input dependent program cond-barrier-deadlock.c, the number of the executed instructions is 302625. Secondly, the time used by MPISE includes the linking time of the target program byte code and TOMPI library. In addition, we need to record states and do solving during symbolic execution, which also needs more time than dynamic analysis. For the rest programs in Umpire test suite, MPISE either reports a deadlock that actually does not exist or aborts during symbolic execution. The reason is we only consider the synchronous communications in MPI programs, or some advanced MPI operations, such as MPI Type vector, are not supported by MPISE. The improvement with respect to these aspects is our future work. To validate the scalability of MPISE, we use MPISE to analyze three real world MPI programs, including an MPI program (CPI) for calculating π and two C MPI programs (DT and IS) from NSA Parallel Benchmarks (NPB) 3.3 [3] with class S. The LOC of DT is 1.2K, and the program needs an input that is either BH, WH or SH for showing the communication graph name. IS is an MPI program for integer sorting, and the LOC of IS is 1.4K. MPISE can analyze these three programs successfully, and no deadlock is found. We make Symbolic exectuion time of CPI Instruction count of CPI 1.7 symbolic execution time(s) 4.00E+05 Inctruction count 3.50E+05 3.00E+05 2.50E+05 2.00E+05 1.50E+05 1.00E+05 5.00E+04 1.65 1.6 1.55 1.5 1.45 0.00E+00 2 4 6 8 10 12 14 CPI CPI 2 4 Number of Processes (a) Instruction count of CPI 10 12 14 Symbolic exectuion time of DT Instruction count of DT 25 symbolic execution time(s) 2.00E+07 Inctruction count 8 (b) Symbolic execution time of CPI 2.50E+07 1.50E+07 1.00E+07 5.00E+06 DT 0.00E+00 6 8 10 12 20 15 10 5 DT 0 6 14 8 10 12 14 Number of Processes Number of Processes (c) Instruction count of DT (d) Symbolic execution time of DT Symbolic exectuion time of IS Instruction count of IS 26 3.20E+07 symbolic execution time(s) 3.15E+07 Inctruction count 6 Number of Processes 3.10E+07 3.05E+07 3.00E+07 2.95E+07 2.90E+07 2.85E+07 2.80E+07 2 4 6 8 Number of Processes (e) Instruction count of IS IS 25.5 25 24.5 24 23.5 23 22.5 22 21.5 21 IS 2 4 6 8 Number of Processes (f) Symbolic execution time of IS Fig. 7. The experimental results under different numbers of processes the input symbolic and symbolically execute all the three MPI programs under different numbers of parallel processes. The experimental results are displayed Fig 7. Because IS can only be run with 2n (n ≥ 1) processes, we do not have results for the case of 6 processes. From Fig 7, we can observe that, for all the three programs, the number of the executed instructions and the symbolic execution time do not increase exponentially with respect to the number of processes. It justifies that MPISE avoids the exponential increasing of instructions or symbolic execution time caused by the parallelism by the on-the-fly scheduling algorithms. Note that we make the input of DT symbolic ones, and this program aborts early when fed with input BH and the process number that is less than 12, this explains the sudden rise of both analyze time and instructions when the number of processes goes from 10 to 12 in Fig.7(c) and Fig.7(d). 5 Related Work There are already some existing work for improving the reliability of MPI programs [14]. Generally, they often fall into one of the following two categories: debugging and testing methods, and verification methods. Debugging and testing tools often scale well, but depend on concrete inputs to run MPI programs, expecting to find or locate bugs. Debugging tools such as TotalView [4] and DDT [1] are often effective when the bugs can be replayed consistently. Whereas, for MPI programs, reproducing a concurrency bug caused by non-determinism is itself a challenging problem. Another kind of tools, such as Marmot [17], the Intel Trace Analyzer and Collector [20] and MUST [15], intercept MPI calls at runtime and record the running information of an MPI program, and check runtime errors, deadlock or analyze performance bottlenecks based on the recorded runtime information. These tools often need to recompile or relink MPI programs, and also depend on the inputs and the scheduling of each running. Another line of tools are verification tools. Dynamic verification tools, such as ISP [25] and DAMPI [27], provide a coverage guarantee over the space of MPI non-determinism. When two or more matches of a non-deterministic operation, such as wildcard receive is detected, the program will be re-executed, and each running using a specific match. Hence, these tools can find the bug relying on a particular choice when non-deterministic operations are encountered, but also depend on the inputs that are fed to run the program. TASS [22] tackles the limitation by using symbolic execution to reason about all the inputs of an MPI program, but its feasibility is limited by the simple MPI model used, which is justified in Section 4.2. There are few static analysis work for MPI program. Greg proposes in [6] a novel data flow analysis notion, called parallel control flow graph (pCFG), which can capture the interaction behavior of an MPI program with arbitrary number of processes. Based on pCFG, some static analysis activities can be carried out. However, the static analysis based on pCFG is hard to be automated. Compared with the existing work, MPISE symbolically executes MPI programs and uses an on-the-fly scheduling algorithm to handle non-determinism, and provides the coverage guarantee on both input and non-determinism. In addition, MPISE uses a realistic MPI library, i.e., TOMPI [11], to be the MPI model used. Therefore, more realistic MPI programs can be analyzed automatically by MPISE, without modifying the programs manually. Furthermore, since the MPI library and the symbolic executor are loosely coupled by hooking the library, it is not hard to switch to another implementation of MPI library to improve the precision and the feasibility of symbolic execution. 6 Conclusion MPI plays a significant role in parallel programming. To improve the reliability of the softwares implemented using MPI, we propose MPISE in this paper to use symbolic execution to analyze MPI programs, targeting to find the bugs of an MPI program automatically. Existing work on analyzing MPI programs suffers problems in different aspects, such as scalability, feasibility and input or non-determinism coverage. We employ symbolic execution to tackle the input coverage problem, and propose an on-the-fly algorithm to reduce the interleaving explorations for non-determinism coverage, while ensuring the soundness and completeness. We have implemented a prototype of MPISE as an adoption of Cloud9, and conducted extensive experiments. The experimental results show that MPISE can find bugs effectively and efficiently. MPISE also provides diagnostic information and utilities to help people understand a bug. For future work, there are several aspects. In one aspect, we plan to support non-blocking MPI operations, which are widely used in nowadays MPI programs. In another aspect, we want to refine our MPI model further, e.g., using a more realistic library, to improve the precision of symbolic execution. Finally, we are also concerned with improving the scalability of MPISE and the analysis of production-level MPI programs. References 1. Allinea DDT. http://www.allinea.com/products/ddt/. 2. The GNU debugger. http://www.gnu.org/software/gdb/. 3. NSA parallel benchmarks. http://www.nas.nasa.gov/Resources/Software/npb. html. 4. Totalview Software. http://www.roguewave.com/products/totalview. 5. Christel Baier and Joost-Pieter Katoen. Principles of Model Checking (Representation and Mind Series). The MIT Press, 2008. 6. Greg Bronevetsky. Communication-sensitive static dataflow for parallel message passing applications. In Proceedings of the 7th Annual IEEE/ACM International Symposium on Code Generation and Optimization, CGO ’09, pages 1–12. IEEE Computer Society, 2009. 7. Stefan Bucur, Vlad Ureche, Cristian Zamfir, and George Candea. Parallel symbolic execution for automated real-world software testing. In Proceedings of the Sixth Conference on Computer Systems, EuroSys ’11, pages 183–198, New York, NY, USA, 2011. ACM. 8. Cristian Cadar, Daniel Dunbar, and Dawson Engler. KLEE: Unassisted and automatic generation of high-coverage tests for complex systems programs. In Proceedings of the 8th USENIX Conference on Operating Systems Design and Implementation, OSDI ’08, pages 209–224, Berkeley, CA, USA, 2008. USENIX Association. 9. Cristian Cadar, Patrice Godefroid, Sarfraz Khurshid, Corina S. Păsăreanu, Koushik Sen, Nikolai Tillmann, and Willem Visser. Symbolic execution for software testing in practice: Preliminary assessment. In Proceedings of the 33rd International Conference on Software Engineering, ICSE ’11, pages 1066–1071, New York, NY, USA, 2011. ACM. 10. Edmund M. Clarke, Jr., Orna Grumberg, and Doron A. Peled. Model checking. MIT Press, Cambridge, MA, USA, 1999. 11. Erik Demaine. A threads-only MPI implementation for the development of parallel programs. In Proceedings of the 11th International Symposium on High Performance Computing Systems, pages 153–163, 1997. 12. Xianghua Deng, Jooyong Lee, and Robby. Bogor/Kiasan. A k-bounded symbolic execution for checking strong heap properties of open systems. In Proceedings of the 21st IEEE/ACM International Conference on Automated Software Engineering, ASE ’06, pages 157–166, 2006. 13. P. Godefroid, M.Y. Levin, D. Molnar, et al. Automated whitebox fuzz testing. In Proceedings of the Network and Distributed System Security Symposium, NDSS ’08, 2008. 14. Ganesh Gopalakrishnan, Robert M. Kirby, Stephen Siegel, Rajeev Thakur, William Gropp, Ewing Lusk, Bronis R. De Supinski, Martin Schulz, and Greg Bronevetsky. Formal analysis of MPI-based parallel programs. Commun. ACM, 54(12):82–91, December 2011. 15. Tobias Hilbrich, Joachim Protze, Martin Schulz, Bronis R. de Supinski, and Matthias S. Müller. MPI runtime error detection with must: Advances in deadlock detection. In Proceedings of the International Conference on High Performance Computing, Networking, Storage and Analysis, SC ’12, pages 30:1–30:11, Los Alamitos, CA, USA, 2012. IEEE Computer Society Press. 16. J.King. Symbolic execution and program testing. Communications of the ACM, 19(7):385–394, 1976. 17. Bettina Krammer, Katrin Bidmon, Matthias S. Müller, and Michael M. Resch. Marmot: An MPI analysis and checking tool. In Gerhard R. Joubert, Wolfgang E. Nagel, Frans J. Peters, and Wolfgang V. Walter, editors, PARCO, volume 13 of Advances in Parallel Computing, pages 493–500. Elsevier, 2003. 18. Glenn R. Luecke, Yan Zou, James Coyle, Jim Hoekstra, and Marina Kraeva. Deadlock detection in MPI programs. Concurrency and Computation: Practice and Experience, 14(11):911–932, 2002. 19. Message Passing Interface Forum. MPI: A message-passing interface standard,version 2.2. http://www.mpi-forum.org/docs/, 2009. 20. Patrick Ohly and Werner Krotz-Vogel. Automated MPI correctness checking what if there was a magic option? In 8th LCI International Conference on HighPerformance Clustered Computing, South Lake Tahoe, California, USA, May 2007. 21. Stephen F. Siegel. Verifying parallel programs with MPI-spin. In Proceedings of the 14th European Conference on Recent Advances in Parallel Virtual Machine and Message Passing Interface, PVM/MPI ’07, pages 13–14, Berlin, Heidelberg, 2007. Springer-Verlag. 22. StephenF. Siegel and TimothyK. Zirkel. TASS: The toolkit for accurate scientific software. Mathematics in Computer Science, 5(4):395–426, 2011. 23. Michelle Mills Strout, Barbara Kreaseck, and Paul D. Hovland. Data-flow analysis for MPI programs. In Proceedings of the 2006 International Conference on Parallel Processing, ICPP ’06, pages 175–184, Washington, DC, USA, 2006. IEEE Computer Society. 24. Sarvani Vakkalanka, Ganesh Gopalakrishnan, and RobertM. Kirby. Dynamic verification of mpi programs with reductions in presence of split operations and relaxed orderings. In Aarti Gupta and Sharad Malik, editors, Computer Aided Verification, volume 5123 of Lecture Notes in Computer Science, pages 66–79. Springer Berlin Heidelberg, 2008. 25. Sarvani S. Vakkalanka, Subodh Sharma, Ganesh Gopalakrishnan, and Robert M. Kirby. Isp: A tool for model checking mpi programs. In Proceedings of the 13th ACM SIGPLAN Symposium on Principles and Practice of Parallel Programming, PPoPP ’08, pages 285–286, New York, NY, USA, 2008. ACM. 26. Jeffrey S. Vetter and Bronis R. de Supinski. Dynamic software testing of mpi applications with umpire. In Proceedings of the 2000 ACM/IEEE Conference on Supercomputing, Supercomputing ’00, Washington, DC, USA, 2000. IEEE Computer Society. 27. Anh Vo, Sriram Aananthakrishnan, Ganesh Gopalakrishnan, Bronis R. de Supinski, Martin Schulz, and Greg Bronevetsky. A scalable and distributed dynamic formal verifier for MPI programs. In Proceedings of the 2010 ACM/IEEE International Conference for High Performance Computing, Networking, Storage and Analysis, SC ’10, pages 1–10, Washington, DC, USA, 2010. IEEE Computer Society. A Theorems and Proofs In order to model MPI programs, we introduce the notion of transition system in [5]: Definition 1 Transition System A transition system TS is a tuple (S, Act, →, I, AP, L), where – – – – – – S is a set of states, Act is a set of actions, →⊆ S × Act × S is a transition relation, I ∈ S is a set of initial states, AP is a set of atomic propositions, and L : S → 2AP is a labeling function. For an action act and a state s, if there is a transition τ = hs, act, s0 i ∈→, we act say that act is enabled in state s, and τ is denoted as s → s0 . If there are no such τ , act is disabled in s. We denote the set of actions enabled in s as enabled(s), act i.e., {act | s1 → s2 ∧ s1 = s}. Note that all transitions systems in this paper is assumed to be action deterministic, i.e. for a sate s ∈ State and an action α ∈ Act, s has at most one transition labeled with α to another state. Hence if α ∈ Act is enabled in state s, we also use α(s) to denote the unique α-successor α of s, i.e. s → α(s). And we use execution to describe the behavior of the transition system. An execution in a transition system is an alternative sequence of states and actions π = s0 α1 s1 α2 . . . , αn sn starts from a initial sate s0 and ends at a terminal state, αi+1 where si → si+1 holds for 0 ≤ i < n. We use |π| to denote the length of π, and |π| = n. To model an MPI process, we define the model more precisely: S = Loc × Eval(V ar) is a set of states, where Loc is the locations of a process, and Eval(V ar) denote the set of variable evaluations that assign values to variables. Act = {s, r, b}, which means we only care about the blocking synchronous send and receive MPI operations as well as the collective operation barrier. We use dest(op), where op ∈ {s, r}, to denote the destination of a send or a receive. The above definition refers to only one process. However, the running of an MPI program typically consists of many processes. Therefore, we need mechanisms to provide the operational model for parallel runnings in terms of transition systems. Definition 2 Parallel composition Let T Si = (Si , Acti , →i , Ii , APi , Li ) i = 1, 2, . . . , n be n transition systems. The transition system T S = T S1 9H T S2 9H · · · 9H T Sn is defined to be: T S = (S1 × S2 × . . . Sn , Actg , →, I1 × I2 × . . . In , AP1 ∪ AP2 ∪ . . . APn , Lg ) where the transition relation → is defined by the following rule: – for matched actions α, β ∈ H = {s, r, b} in distinct processes: α si →i s0i ∧ β sj →j s0j ∧ match(α, β) SR hs1 , . . . , si , . . . , sj , . . . sn i −−→ hs1 , . . . , s0i , . . . , s0j , . . . sn i here match(α, β) if and only if (α = s∧β = r)∨(α = r∧β = s), dest(α) = j, and dest(β) = i, SR is the compositional global action of s and r. – for matched actions α = b in distinct processes: α s1 →1 s01 ∧ α . . . si →i s0i ∧ α . . . sn →n s0n B hs1 , . . . , si , . . . sn i − → hs01 , . . . , s0i , . . . s0n i Here B is the compositional global action of the local action b of each process. S The labeling function is defined by Lg (hs1 , . . . , si , . . . , sn i) = 1≤i≤n L(si ). And note that actions in Actg are also introduced in the above two rules. The composition of transition systems gives a global view of a directed graph G = (S, T ), where the nodes in S are global states and an edge in T is a global transition with action SR or B. Note that the idea behind our on-the-fly schedule is that, for a global state, we only choose some of the transitions to move on and discard the others. Hence, we only explore a subgraph. To describe this subgraph, we first introduce some notions here. Given a global state σ in a composed transition system, we fix a total order on actions enabled in σ according to weight(act), where act ∈ enabled(σ), δ = act σ → σ 0 , and  1, if act = B; weight(act) = i, if act = SR , σ = hs1 , . . . , sn i, σ 0 = hs1 , . . . , s0i , . . . , s0j , . . . sn i. i.e., act1 < act2 iff weight(act1 ) < weight(act2 ). When an action act ∈ enabled(s) has the minimal weight, we say that act ranks first in enabled(s). Definition 3 Let G̃ = (S, T̃ ) be a subgraph of G, where T̃ is defined as follows: [ T̃ = {τ | τ = hs, act, s0 i ∧ act ranks first in enabled(s)}. s∈S We can see that this G̃ is formed according to the on-the-fly schedule, in which we always schedule the active process with the smallest rank. Now, we can present the main theorem, which guarantees the completeness and soundness of the on-the-fly schedule. Theorem 1. Given an execution π in G from a global initial state σ0 to a deadlocked global state σ, there exists an execution T from σ0 to σ in G̃ such that |T | = |π|. And vice versa. To prove the above theorem, we introduce some notions first. An independent relation I ∈ Act × Act is a relation, satisfying the following two conditions: for any state s ∈ S, with α, β ∈ enabled(s) and α 6= β, – Enabledness α ∈ enabled(β(s)) , β ∈ enabled(α(s)). – Commutativity α(β(s)) = β(α(s)). α Recall that α(s) denotes the unique α successor of s, i.e. if s → s0 holds, then s0 = α(s). The dependency relation D is the complement of I, i.e., D = Act×Act−I. The method we constructing a subgraph G̃ from a graph G actually falls into the ample set method proposed in [10], which expands only a part of the transitions at each state, for a path not considered by the method, there is an equivalent path with respect to the specification considered. Among the four conditions C0-C3 for selecting ample(s) ⊆ enabled(s), the first two are as follows: C0 ample(s) = ∅ if and only if enabled(s) = ∅. αn+1 α α C1 Let s0 →0 s1 , . . . , →n sn → t be a finite execution fragment, if αn+1 depends on β ∈ ample(s0 ), then there exists an αi ∈ ample(s0 ), where 0 ≤ i < n + 1, i.e., along every execution in the full state graph G that starts at s, an action appears in the execution fragment starts at s which is dependent on an action in ample(s) cannot be executed without an action in ample(s) occurring first. C2 makes sure that: when ample(s) 6= enabled(s), the actions in ample(s) do not change the label of states with respect to the verifying properties. Here the property we concerned with is deadlock-reachability. Since we commute independent actions by applying commutativity of independent relation, we do not need C2. C3 ensures that a cycle in the reduced graph needs to satisfy some requirements. Owing to the acyclicity of the state space in out context, we do not need C3. In our circumstance, we define ample(s) as following:  {B} if B ∈ enabled(s); ample(s) = {SR} else if SR ∈ enabled(s) ∧ SR ranks first in enabled(s). Note that for a state s ∈ S, if enabled(s) 6= ∅, |ample(s)| = 1, i.e., ample(s) has only one element. Hence, C0 surely holds. To check whether C1 holds on the full state graph generated by our schedule algorithm, we first introduce some properties of the full state graph. Clearly, according to the definition of parallel composition, only SR actions can be enabled simultaneously at a global state, and the SR actions enabled at a same state are αj−1 αj α independent. So, given a execution fragment π = s0 →0 s1 . . . , → sj → sj+1 starting from s0 ∈ S, αj depends on an action β ∈ ample(s0 ). We want to prove that there exists a αk , where 0 ≤ k < j, and αk ∈ ample(s0 ). Recall the definition of dependent relation D, we know that αj and β should meet one of the following cases: 1. @s ∈ S such that αj , β ∈ enabled(s); 2. for any state s ∈ S, if αj , β ∈ enabled(s), then αj 6∈ enabled(β(s)) or β 6∈ enabled(αj (s)), i.e., either αj disables β or β disables αj ; 3. for any state s ∈ S, if αj , β ∈ enabled(s) and αj ∈ enabled(β(s)) ∧ β ∈ enabled(αj (s)), then αj (β(s)) 6= β(αj (s)), i.e., αj and β are not commutative. Because only SR actions can be enabled simultaneously at a global state, both case 2 and case 3 cannot hold under our context. Therefore, only case 1 holds for the state graph generated by our method, i.e., αj and β should not be both enabled at any state. Based on this result, we can get C1 holds in our context by contradiction. Assume that {α0 . . . αj−1 } ∩ ample(s0 ) = ∅ holds for the execution π. Because β is enabled at s0 , α0 and β are independent, hence β is also enabled at s1 . In addition, α1 is also enabled at s1 and α1 6= β, so α1 and β are also independent. In the same way, we can get that each αi (0 ≤ i < j) and β are independent. Thus, by using commutativity, we can get β and αj are both enabled at sj , which violates case 1. Hence, the condition C1 holds. Proof. One direction, because G̃ is a subgraph of G, the execution T from δ0 to δ in G̃ is also an execution from δ0 to δ in G, hence we got an execution π = T , and |T | = |π|. The other direction is a little more complex. The basic idea is to construct a corresponding execution in the subgraph gradually based on the ample set of each state passed in π. Let π be an execution in G from δ0 to δ. We construct a finite sequence of executions π0 , π1 , . . . , πn , where π0 = π and n = |π|. Each execution πi is constructed based on the before execution πi−1 . For example, π1 is constructed from π0 , i.e., π, according to the first action execution in π. Thus, we want to prove that the last execution πn is an execution in the subgraph, and shares the same first and last states with π. We can prove it by presenting the construction method of each step. We decompose each πi into two execution fragments, i.e., πi = ηi ◦ θi , where ηi is of length i and ηi ◦ θi is the concatenation of the two execution fragments. Assuming that we have constructed π0 , . . . , πi , we now turn to construct πi+1 = ηi+1 ◦ θi+1 . Let s0 be the last state of the execution fragment ηi and α be the first action of θi . Note that s0 is also the first state of the execution fragment θi , i.e., α0 =α α1 α2 θi = s0 −→ s1 −→ s2 −→ . . . s|θi | There are two cases: α1 α2 α A. α ∈ ample(s0 ). Then ηi+1 = ηi ◦ (s0 → α(s0 )) and θi+1 = s1 −→ s2 −→ . . . s|θi | . B. α 6∈ ample(s0 ). Note that s|θi | = σ is a deadlock state, hence no action can be enabled at σ. Therefore, for any action β ∈ ample(s0 ), some actions that appear in θi must be dependent on β. The reason is: if all the actions that appears in θi is independent of β, then all the actions in θi cannot disable β, hence β would be enabled at s|θi | = σ, which violates the premiss that σ is a deadlock state. Therefore, for any action β in ample(s0 ), we can find an action αj that appears in θi , and αj depends on β. According to C1, there must exist an action β 0 ∈ ample(s0 ), such that β 0 occurs before αj . Because there may exist multiple actions that are in ample(s0 ) and occur before αj , we take the first one, say αk and αk ∈ ample(s0 ). So, αk is the first one among the elements of ample(s0 ) that occur in θi . Clearly, the actions before αk , i.e., α0 , ..., αk−1 , are independent with αk . Hence, we can construct the following execution by using the commutativity condition k times: α|θi |−1 αk−1 αk+1 αk+2 α α1 α=α s|θi | . . . . −→ αk (sk ) −→ sk+2 −→ . . . −→ ξ = s0 →k αk (s0 ) −→0 αk (s1 ) −→ α In this case ηi+1 = ηi ◦ (s0 →k αk (s0 )) and θi+1 is the execution fragment α that is obtained from ξ by removing the first transition s0 →k αk (s0 ). Clearly, πi and πi+1 share the same last state. So, π0 = δ and πn share the same last state of the execution, namely πn is also an execution from δ0 to δ in G. In addition, according to the construction procedure, |π| = |πn | holds. αj Most importantly, in execution πn , for any 0 ≤ j < n, such that sj −→ sj+1 , αj ∈ ample(sj ) holds. Therefore, πn is also an execution from δ0 to δ in G̃, and we take this execution as T .  To prove the correctness and soundness of our lazy matching algorithm, we need to deal with wildcard receives. Hence the rules of parallel composition of transition systems need to be refined. Instead of redefine match to make it work with wildcard receives, we make a new rule for matched send and wildcard receives, to distinct it with source specific receives. – for matched actions α, β ∈ H = {s, r∗ } in distinct processes, where r∗ is the wildcard receive: α si →i s0i ∧ β sj →j s0j ∧ match(α, β) SR∗ hs1 , . . . , si , . . . , sj , . . . sn i −−−→ hs1 , . . . , s0i , . . . , s0j , . . . sn i here match(α, β) if and only if α = s ∧ β = r∗ , dest(α) = j, and dest(β) = AN Y , SR∗ is the compositional global action of s and r∗ . We also need to redefine the subgraph G̃ because we have a new kind of global transitions. Definition 4 Let T̃ = S subtran(s) , where subtrans(s) is defined as: s∈S  B   {s → B(s)} if B ∈ enabled(s); subtran(s) = {s SR → SR(s)} else if SR ∈ enabled(s)∧SR ranks first in enabled(s);   act {s → act(s)} else if act ∈ enabled(s). Let G̃∗ = (S, T̃ ), which is a subgraph of the full state graph. Clearly, we can see that G̃∗ is the subgraph we formed according to the on-the-fly schedule plus lazy matching. Accordingly, we define ample(s) as:  if B ∈ enabled(s);  {B} else if SR ∈ enabled(s) ∧ SR ranks first in enabled(s); ample(s) = {SR}  enabled(s) other. And the following theorem addresses the correctness and soundness of lazy matching algorithm: Theorem 2. Given any execution π in G from a global state σ0 to a deadlocked global state σ, there exists an execution T from σ0 to σ in G̃∗ such that |T | = |π|. And vice versa. To prove theorem 2, we first check the conditions C0 and C1. Clearly, C0 holds. To check C1, we should point out that in the new global state graph, an action SR ∈ enabled(s) is independent with the rest actions in enabled(s). In addition, only SR∗ can disable other SR∗ actions, which is ensured by the semantics of wildcard receives. Same as before, given a execution fragment π = αj−1 αj α s0 →0 s1 . . . , → sj → sj+1 starting from s0 ∈ S, αj depends on an action β ∈ ample(s0 ). We want to prove that there exists a αk , where 0 ≤ k < j, and αk ∈ ample(s0 ). We discuss the two cases of ample(s0 ): 1. ample(s0 ) = enabled(s0 ). Clearly, C1 holds. 2. ample(s0 ) 6= enabled(s0 ). Thus, ample(s0 ) = {SR} and β = SR. We then discuss two cases: 1) αj ∈ enabled(s0 ) \ ample(s0 ), according to the observation, SR is independent with each of the rest actions in enabled(s0 ), so αj and SR are independent. Therefore, it is a contradiction, thus this case never happens. 2) αj 6∈ enabled(s0 ). For an SR action, the dependent relation of it can only be one case, i.e., @s ∈ S such that αj , β ∈ enabled(s). Because SR will never be disabled by any other action, same as the idea for proving C1 for the case without wildcard receives, we can prove that β occurs before αj . In total, we can get that C1 holds on the new state graph. Proof. We have concluded that the condition C0 and C1 still holds in the new full state graph. Hence, the procedure of the proof for theorem 2 is basically the same with that of theorem 1.  

ENUMERATIVE ASPECTS OF NULLSTELLENSATZ CERTIFICATES arXiv:1607.05031v1 [math.CO] 18 Jul 2016 BART SEVENSTER: , JACOB TURNER: Abstract. Using polynomial equations to model combinatorial problems has been a popular tool both in computational combinatorics as well as an approach to proving new theorems. In this paper, we look at several combinatorics problems modeled by systems of polynomial equations satisfying special properties. If the equations are infeasible, Hilbert’s Nullstellensatz gives a certificate of this fact. These certificates have been studied and exhibit combinatorial meaning. In this paper, we generalize some known results and show that the Nullstellensatz certificate can be viewed as enumerating combinatorial structures. As such, Gröbner basis algorithms for solving these decision problems may implicitly be solving the enumeration problem as well. Keywords: Hilbert’s Nullstellensatz, Polynomial Method, Enumerative Combinatorics, Algorithmic Combinatorics 1. Introduction Polynomials and combinatorics have a long common history. Early in both the theories of graphs and of matroids, important polynomial invariants were discovered including the chromatic polynomial, the Ising model partition function, and the flow polynomial, many of which are generalized by the Tutte polynomial [29, 30, 5, 26]. The notion of using generating functions is ubiquitous in many areas of combinatorics and many these polynomials can be viewed as such functions. Another general class of polynomials associated to graphs is the partition function of an edge coloring model [9]. On the other hand, polynomials show up not just as graph parameters. Noga Alon famously used polynomial equations to actually prove theorems about graphs using his celebrated “Combinatorial Nullstellensatz” [1]. His approach often involved finding a set of polynomial equations whose solutions corresponded to some combinatorial property of interest and then studying this system. Modeling a decision problem of finding a combinatorial structure in a graph by asking if a certain system of polynomial equations has a common zero is appealing from a purely computational point of view. This allows these problems to be approached with well-known algebraic algorithms from simple Gaussian elimination to Gröbner basis algorithms [10, 8]. Using polynomial systems to model these problems also seems to work nicely with semi-definite programming methods [17, 19, 18, 27]. This approach was used in [21], where the question of infeasibility was considered. If a set of polynomial equations is infeasible, Hilbert’s Nullstellensatz implies that there is a set of polynomials acting as a certificate for this infeasibility. For any : Korteweg-de Vries Institute for Mathematics, University of Amsterdam, 1098 XG Amsterdam, Netherlands. 1 2 BART SEVENSTER: , JACOB TURNER: given finite simple graph, a polynomial system whose solutions corresponded to independent sets of size k was originally formulated by László Lovász [22], but other articles have studied the problem algebraically [20, 28]. One of the interesting results in [21] was to show that for these particular systems of polynomials, the Nullstellensatz certificate contained a multivariate polynomial with a natural bijection between monomials and independent sets in the graph. As such, the Nullstellensatz certificate can be viewed as an enumeration of the independent sets of a graph. Furthermore, the independence polynomial can quickly be recovered from this certificate. Later, when modeling the set partition problem, a similar enumeration occurred in the Nullstellensatz certificate [24]. This paper is directly inspired by these two results and we look at the different systems of polynomials given in [21] and show that this phenomenon of enumerative Nullstellensatz certificates shows up in all of their examples. We explain this ubiquity in terms of inversion in Artinian rings. One important example considered in [21] was k-colorable subgraphs. This problem has also been studied by analyzing polynomial systems in [2, 14, 25, 13]. We generalize the polynomial systems used in [21] to arbitrary graph homomorphisms. We also consider existence of planar subgraphs, cycles of given length, regular subgraphs, vertex covers, edge covers, and perfect matchings. On the one hand, these results may be viewed negatively as they imply that a certificate for infeasibility contains much more information than necessary to settle a decision problem. There have also been papers on attempting efficient computations of Nullstellensatz certificates ([11, 12]) and we should expect that often this will be very hard. On the other hand, if one wishes to enumerate combinatorial structures, our results imply algorithms built from known algebraic techniques to solve this problem. One polynomial system that does not fall into the general setting of the other examples is that of perfect matchings. While there is a natural system of equations modeling this problem that does have enumerative Nullstellensatz certificates, there is another system of polynomials generating the same ideal that does not. We spend the latter part of this paper investigating the second set and try to achieve some partial results in explaining what combinatorial information is contained in the Nullstellensatz certificates. This paper is organized as follows. In Section 2, we review the necessary background on the Nullstellensatz and make our definitions precise. We then present our motivating example, independent sets, and explain how this particular problem serves as a prototype for other interesting problems. In Section 3, we prove a sufficient condition for a Nullstellensatz certificate to enumerate combinatorial structures and give several examples, some of them new and some reformulations of old ones, that satisfy this property. Lastly, in Section 4, we look at a system of polynomials whose solutions are perfect matchings that do not satisfy this sufficient condition and prove some results about the certificates. 2. Background Given a system of polynomials f1 , . . . , fs P Crx1 , . . . , xn s, consider the set Vpf1 , . . . , fs q “ tpα1 , . . . , αn q P Cn | f1 pα1 , . . . , αn q “ ¨ ¨ ¨ “ fs pα1 , . . . , αn q “ 0u. We call such a set an variety (by an abuse of language, we use the term even for reducible and non-reduced sets in this paper). In particular, the empty set is a ENUMERATIVE ASPECTS OF NULLSTELLENSATZ CERTIFICATES 3 variety and if Vpf1 , . . . , fs q “ H, we say that the system of polynomials f1 , . . . , fs is infeasible. One version David Hilbert’s famous Nullstellensatz states that a system f1 , . . . , fs P Crx1 , . . . , xn s is infeasible ř if and only if there exists polynomials β1 , . . . , βs P Crx1 , . . . , xn s such that βi fi “ 1 (cf. [8]). The set of polynomials β1 , . . . , βs are called a Nullstellensatz certificate for the infeasibility of the system. The degree of the Nullstellensatz certificate is defined to be maxtdegpβ1 q, . . . , degpβs qu. We note that a Nullstellensatz certificate is dependent on the choice of polynomials defining the system. We will revisit this point later. The second observation is that Nullstellensatz certificates aren’t unique. Often the Nullstellensatz certificates of greatest interest are those of minimum degree. Research into an “effective Nullstellensatz” has yielded general bounds for the degree of a Nullstellensatz certificate of a system of polynomials [4, 15]. As such, the following general algorithm for finding such certificates has been proposed (cf. [11, 12]). Suppose we have a system řs of s equations in n variables, f1 , . . . , fs . We want to find β1 , . . . , βs such that i“1 βi fi “ 1. Let Mn,k denote the set of monomials of degree ď k in n variables. Algorithm 1 Basic Outline of the NulLA Algorithm. Suppose we know from some theorem that a Nullstellensatz certificate must have degree at most d. i “ 0. while i ď d do Ź test every degree for a certificate ř For i P rss, βi :“ M PMn,d αM M . Let L be the empty set of linear equations. for M P Mn,d do ř Determine the coefficient of M in βi fi , LM . if M “ 1 then Append LM “ 1 to L. else Append LM “ 0 to L. Solve the system L if possible. if L has a solution then Output ”Yes” and Exit While Loop. if i “ d then Output ”No”. else i ÞÑ i ` 1. We summarize the above pseudocode. First guess at the degree of the Nullstellensatz certificate and then consider ř generic polynomials βi in variables x1 , . . . , xn of said degree. Then the condition βi fi “ 1 can be reformulated as system of linear equations whose solutions give the coefficients each βi should have. If the linear system has no solution, the guessed degree is increased. The general degree bounds guarantee that this process will terminate eventually, implicitly finding a valid certificate of minimal degree, provided the initial polynomial system was infeasible. This algorithm is similar to the XL style Gröbner basis algorithms studied in algebraic cryptography [7, 3]. One way to understand the complexity of this algorithm is to look at the Nullstellensatz certificates that get produced for different systems of polynomials. This algorithm is one of the main motivations behind the inquiry into Nullstellensatz certificates. 4 BART SEVENSTER: , JACOB TURNER: In this paper, we will consider combinatorial problems modeled by systems of polynomials in Crx1 , . . . , xn s. The problems we consider will all come from the theory of finite graphs and all varieties will be zero dimensional. 2.1. Motivating Example. Let us give the first example of a polynomial system modeling a graph problem: independent sets. Lovász gave the following set of polynomials for determining if a graph has an independent set of size m. Proposition 2.1 ([22]). Given a graph G, every solution of the system of equations x2i ´ xi “ 0, i P V pGq, řn xi xj “ 0, ti, ju P EpGq, i“1 xi “ m corresponds to an independent set in G of size m. It is not hard to see that the equations in Proposition 2.1 define a zero-dimensional variety whose points are in bijection with independent sets of size m. The first equation says that xi “ 0, 1 for all i P V pGq. So every vertex is either in a set or not. The second equation says that two adjacent vertices cannot both be in the same set. The last equation says that precisely m vertices are in the set. If this system is infeasible, then the Nullstellensatz certificate has been explicitly worked out [21]. Many of properties of the certificate encodes combinatorial data. For example, one does not need to appeal to general effective Nullstellensatz bounds as the degree of the Nullstellensatz certificate can be taken to be the independent set number of the graph in question. Combining several theorems from that paper, the following is known. Theorem 2.2 ([21]). Suppose the system of equations in Proposition 2.1 are infeasible. Then the minimum degree Nullstellensatz certificate is unique and has degree equal to αpGq, the independence number of G. If 1 “ Ap´m ` n ÿ i“1 xi q ` ÿ ti,juPEpGq Qij xi xj ` n ÿ Pi px2i ´ xi q, i“1 with certificate polynomials A, Qij , and Pi , then the degree of this certificate is realized by A; degpPi q ď αpGq ´ 1 and degpQij q ď αpGq ´ 2. Furthermore, if the certificate is of minimum degree, the monomials in A with non-zero coefficients can ś be taken to be precisely those of the form iPI xi where I is an independent set of G. Lastly, the polynomials A, Qi , and Pi all have positive real coefficients. So Theorem 2.2 tells us for a minimum degree certificate, the polynomial A enumerates all independent sets of G. The coefficients of the monomials, however, will not necessarily be one, so it is not precisely the generating function for independents sets. The precise coefficients were worked out in [21]. In the next section, we show that a similar theorem will hold for many other examples of zero-dimensional varieties coming from combinatorics. 3. Rephrasing Nullstellensatz certificates as inverses in an Artinian ring Recall that a ring S is called Artinian is it satisfies the descending chain condition on ideals, i.e. for every infinite chain of ideals ¨ ¨ ¨ Ĺ I2 Ĺ I1 Ă S, there is some i ENUMERATIVE ASPECTS OF NULLSTELLENSATZ CERTIFICATES 5 such that Ik “ Ik`1 for all k ě i. Equivalently, viewed as a left module over itself, S is finite dimensional, meaning it contains finitely many monomials. Let V be a variety in Cn defined by the equations f1 , . . . , fs . Although not standard, we do not assume that V is reduced or irreducible, which is to say that often the ideal xf1 , . . . , fs y will not be radical or prime. An ideal IpVq :“ xf1 , . . . , fs y is radical if g ` P IpVq ùñ g P IpVq. The quotient ring CrVs :“ Crx1 , . . . , xn s{IpVq is called the coordinate ring of V. It is an elementary fact from algebraic geometry that V is zero dimensional if and only if CrVs is Artinian. This is true even for non-reduced varieties. The polynomial systems coming from combinatorics are designed to have a solution if some combinatorial structure exists. In Subsection 2.1, the combinatorial structure of interest was independent sets in a graph. Many of the polynomials systems that show up in examples have a particular set of equations that always play the same role. Given a polynomial system in Crx1 , . . . , xn s, we call a subset of the variables xi for i P I Ď rns indicator variablesřif the polynomial system includes the equations x2i ´ xi “ 0 for i P I and ´m ` iPI xi “ 0 for some m P N. This was the case for the example in Subsection 2.1. The indicator variables often directly correspond combinatorial objects, e.g. edges or vertices in a graph. The equations x2i ´ xi “ 0 means that object i is ř either in some structure or not. Then the equation ´m ` iPI xi “ 0 says that there must be m objects in the structure. The other equations in the polynomial system impose conditions the structure must satisfy. Now suppose we are given an infeasible polynomial system f1 “ 0, . . . , fs “ 0 in R :“ Cry1 , . . . , yn , x1 , . . . , xp s, where řn y1 , . . . , yn are indicator variables. Without loss of generality, let f1 “ ´m ` i“1 yi for some m P N. Then one way to find a Nullstellensatz certificate for this polynomial system is to find the ř inverse of f1 in s R{xf2 , . . . , fs y, which we denote β1 , and then express f1 β1 ´ 1 as i“2 βi fi . The polynomials β1 , β2 , . . . , βs will be a Nullstellensatz certificate. Throughout the rest of this section, we consider an infeasible polynomial system f1 “ 0, . . . , fs “ 0 in A :“ řnCry1 , . . . , yn , x1 , . . . , xp s where y1 , . . . , yn are indicator variables and f1 “ ´m ` i“1 yi for some m P N not equal to zero. We let R :“ A{xf2 , . . . , fs y and V “ SpecpRq be the variety defined by f2 , . . . , fs . Lemma 3.1.řThe ring R is Artinian if and only if for every a P N, the polynomial n system ´a ` i“1 yi “ 0, f2 “ 0, . . . , fs “ 0 has finitely many solutions. Proof. We look at the variety defined by the equations f2 , . . . , fs . Since y1 , . . . , yn are indicator variables, the equations yi2 ´ yi “ 0 are among the equations f2 “ 0, . . . , fs “ 0. Thus any solution to these equations must have every yi is equal to řneither zero or one. Thus there are only finitely many a P N such that ´a ` i“1 yi “ 0, f2 “ 0, . . . , fs “ 0 has a solution as a must be less than or equal to n. Furthermore, each such polynomial system has finitely many solutions so the system f2 “ 0, . . . , fs “ 0 only has finitely many solutions. Thus V is zero dimensional and the ring R is Artinian. Conversely, if R is Artinian, there can be only finitely many solutions to the system f2 “ 0, . . . , fs “ 0 and thus ř only a finite subset of n  those solutions can also be a solution to the equation ´a ` i“1 yi “ 0. From here on out, we assume that R is Artinian as this will be the case in every example we consider. This is the consequence of the fact that the polynomial 6 BART SEVENSTER: , JACOB TURNER: systems are designed to have solutions corresponding to some finite combinatorial structure inside of some larger, yet still finite, combinatorial object. In our examples, we are always looking for some graph structure inside a finite graph. The examples we consider also satisfy another property that we shall assume throughout the rest of this section unless otherwise stated. Definition 3.2. The system of polynomials f2 “ 0, . . . , fs “ 0 in indicator variables y1 , . . . , yn is called subset closed if (a) There is a solution of the system where y1 “ ¨ ¨ ¨ “ yn “ 0 and (b) Let I Ď rns and χI piq “ 1 if i P I and 0 else. If there is a solution of the system where yi “ χI piq, then for all J Ď I, there is a solution of the system where yj “ χJ pjq. It is very easy to describe the monomials with non-zero coefficients appearing in the inverse of f1 in the ring R. We look at the variety defined by the polynomials f2 , . . . , fs which consists of finitely many points. Suppose we are given a solution to the system f2 “ 0, . . . , fs “ 0: y1 “ d1 , . . . , yn “ dn , and x1 “ a1 , . . . , xp “ ap . We can then map this solution to the point pd1 , . . . , dn q P t0, 1un . We see that each solution to the system f2 , . . . , fs can be associated to a point on the n-dimensional hypercube t0, 1un . Let B be the subset of t0, 1un which is the image of this mapping. n ś Given d “ pd1 , . . . , dn q P t0, 1u , we can associate to it the monomial yd :“ i,di “1 yi . For any b “ pb1 . . . , bn q P Bzp0, . . . , 0q, the monomial yb regarded as a function restricted to V is not identically zero as there is a point in V where yi “ 1 for all bi “ 1 in pb1 , . . . , bn q. Lemma 3.3. If f1 “ 0, . . . , fs “ 0 is subset closed, then for d “ pd1 , . . . , dn q R B, yd is in the ideal generated by f2 , . . . , fs . Proof. If this were not the case, there would be a point v “ pc1 , . . . , cn , γ1 , . . . , γp q P V such that yd pvq “ 1. Let I be the set ti P rns|ci “ 1u. If J “ ti P rns|di “ 1u, we see that J Ď I and that there is a solution where yi “ χI piq. So there must be a solution where yi “ χJ piq by the property of being subset closed. This implies that pd1 , . . . , dn q P B since di “ χJ piq, and so we have a contradiction. This means that for d R B, there is some k P N such that ydk “ 0 in the ring R. The exponent k may a priori be greater than one as we have not assumed that the ideal xf2 , . . . , fs y is radical. However, we note that for any d P t0, 1un , ydk “ yd for all k P N because the equations yi2 ´ yi for all i P rns are among the polynomials f2 , . . . , fs . Thus for d R B, yd “ 0 in the ring R.  We can thus conclude that the monomials in the indicator variables in R are in bijection with combinatorial structures satisfying the constraints of the polynomial system. In Proposition 2.1, the ring Crx1 , . . . , x|V pGq| s modulo ś the first two sets of polynomials gives a ring whose monomials are of the form iPI xi , where I indexes vertices in an independent set of G. Lemma 3.4. Given b1 , . . . , bk P B, there are no polynomial relations of the form řk i“1 ai ybi “ 0 for ai P Rě0 in R except for all ai “ 0. Similarly if all ai P Rď0 . Proof. We note that because of the polynomials yi2 ´ yi “ 0, any monomial yβi can only take the value of zero or one when restricted to V. The only set of non-negative real numbers whose sum is zero is the trivial case where all are zero. The proof for the second assertion is the same as the first.  ENUMERATIVE ASPECTS OF NULLSTELLENSATZ CERTIFICATES 7 Theorem 3.5. There is a Nullstellensatz certificate β1 , . . . , βs for the system f1 “ 0, . . . , fs “ 0 such that the non-zero monomials of β1 are precisely the monomials yb for b P B. Proof. We look at the Nullstellensatz certificate β1 , . . . , βn where β1 f1 “ 1 in R so řs we can express β1 f1 ´ 1 “ i“2 βi fi in A. We now analyze what β1 must look like. First of all we note that over C, there is a power series expansion ˆ ˙i 1 1 ÿ t “´ ´m ` t m iě0 m řn viewed as a function in t. Replacing t with i“1 yi , we get a power series in the indicator variables that includes every monomial in these variables with a negative coefficient. We consider the partial sums ˙i k ˆ n ÿ 1 ÿ t , t“ yi ´ m i“0 m i“1 first taken modulo the ideal generated by the polynomials of the form yi2 ´ yi . This gives a sum where the monomials are yd for d P t0, 1un whose coefficients are all negative real numbers. We know from Lemma 3.3, that the monomials in β1 can be taken of the form yb for b P B; the other monomials can be expressed in terms of f2 , . . . , fs . So then those monomials that are equal to zero modulo f2 , . . . , fs can ř ř be removed, giving us another sum. If there is a relation of the form 1 i ai ydi “ j cj ydj , we ignore it. Such relations simply mean that there is a nonunique way to represent this sum in R. Lastly, these partial sums converge to the inverse β1 of f1 in R which is supported on the monomials of the form yb for b P B, using the assumption that R is Artinian.  Theorem 3.5 guarantees the existence of a Nullstellensatz certificate such that every possible combinatorial structure satisfying the constraints of f2 , . . . , fs is encoded in a monomial in the indicator variables appearing with non-zero coefficient in β1 . This is precisely the Nullstellensatz certificate found in Theorem 2.2 since any subset of an independent set is again an independent set. As it so happens, in Theorem 2.2 this Nullstellensatz certificate is also of minimal degree. However, it is not necessarily the case that the certificate given in Theorem 3.5 is minimal. The most obvious way for minimality to fail is by reducing by the linear relations among the monomials yb for b P B with both negative and positive coefficients. However, if all such linear relations are homogeneous polynomials, we show these relations cannot reduce the degree of the Nullstellensatz certificate. Definition 3.6. We say that the ideal f2 , . . . , fs has only homogeneous linear ř` relations among the indicator variables if every equation is of the form i“1 ai ydi “ 0 for ai P R and di P t0, 1un and has the property that not all ai are positive or all negative and that degpydi q “ degpydj q for all i, j P r`s. Lemma 3.7. Let β1 , . . . , βs be a minimal Nullstellensatz certificate for the system f1 “ 0, . . . , fs “ 0, and suppose that β1 only has positive real coefficients. Then after adding homogeneous relations in the indicator variables to the system, there is a minimal degree Nullstellensatz certificate of the form β1 , β21 , . . . , βs1 . 8 BART SEVENSTER: , JACOB TURNER: Proof. By adding homogeneous relations in the indicator variables, we claim it is impossible to reduce the degree of β1 if it only has positive real coefficients. Let us try to remove the monomials of highest degree in β1 by adding homogeneous linear relations in the indicator variables. We apply the first linear homogeneous relation to seeřwhich monomials we can remove. The relations are of the form řk ` a y “ j“k`1 aj ybj where the ybk are all monomials of a given degree d i b i i“1 and all ak P Rě0 . Thus we can potentially remove some of the monomials of degree d in β1 using such relations, but not all of them. This is true not matter how many linear homogeneous relations we apply; there will always be some monomials of highest degree remaining. However, we might be able to reduce the degree of the other βi , i ě 2 to get a Nullstellensatz certificate β1 , β21 , . . . , βs1 such that the degree of this new certificate is less than the degree of the original.  Given the Nullstellensatz certificate β1 , . . . , βs guaranteed by Theorem 3.5, we note that f1 β1 ´ 1 must consist entirely of monomials of the form yd for d R B. If D “ maxtdegpyb q| b P Bu, then f1 β1 ´ 1 must be of degree D ` 1 as f1 is linear. Let M denote the set of monomials (in A) of f1 β1 ´ 1. We consider the following hypothetical situation where the Nullstellensatz certificate guaranteed by Theorem 3.5 is not of minimal degree. Given a monomial řs µ P M , we have that µ “ i“2 µi fi for some polynomials µi , although this is not unique. We define s ÿ degR pµq :“ maxtdegpµi q, over all equalities µ “ µi fi u. i“2 Then the degree of β1 , . . . , βs is maxtD, degR pµqfor µ P M u. Suppose that the degree of the certificate is ě D`2 and that for every maxtdegR pyi ¨ µq|µ P M u ă maxtdegR pµq|µ P M u for every i P rns. Then define β11 :“ β1 ` m´1 pf1 β1 ´ 1q, which has degree D ` 1. Now we note that f1 β11 ´ 1 “ f1 β1 ` m´1 f1 pf1 β1 ´ 1q ´ 1 “ f1 β1 ´ f1 β1 ` 1 ` m´1 n ÿ yi pf1 β1 ´ 1q ´ 1 i“1 “ m´1 n ÿ yi pf1 β1 ´ 1q. i“1 řn However, since every monomial in m´1 i“2 yiř pf1 β1 ´ 1q is of the form yi ¨ µ for s some µ P M , we can express this polynomial as i“1 βs1 fs where degpβs1 q ă degpβs q since maxtdegR pyi ¨ µq|µ P M u ă maxtdegR pµq|µ P M u for every i P rns. So we have found a Nullstellensatz certificate with smaller degree. In the hypothetical situation above, we were able to drop the degree of the Nullstellensatz certificate by increasing the degree of β1 by one. However, this construction can be iterated and it may be that the degree of β1 must be increased several times before the minimal degree certificate is found. This depends on how high the degrees of β2 , . . . , βs are. It may also be the case that adding lower degree monomials also lowers the degree of the certificate. ENUMERATIVE ASPECTS OF NULLSTELLENSATZ CERTIFICATES 9 Lemma 3.8. If β1 , . . . , βs be the Nullstellensatz certificate guaranteed by Theorem 3.5 and f2 , . . . , fs have only linear homogeneous relations in the indicator variables. If maxtdegpβi q, i P rssu “ degpβi q, then this is a Nullstellensatz certificate of minimal degree. Proof. For any Nullstellensatz certificate β11 , . . . , βs1 , we may assume without loss of generality that the monomials of β1 form a subset of those in β11 . Indeed we may use Lemma 3.3 to say that all monomials yb for b P B must appear in β11 unless there are linear homogeneous relations in the indicator variables. By Lemma 3.7, reducing β11 alone by these relations will not reduce the degree of β11 , . . . , βs1 . However, this implies that degpβ11 q ě degpβ1 and thus that the degree of the Nullstellensatz certificate β11 , . . . , βs1 has degree ě degpβ1 q, which is the degree of the certificate β1 , . . . , βs by assumption. So β1 , . . . , βs is a Nullstellensatz certificate is of minimal degree.  Lemma 3.8 tells us that the Nullstellensatz certificate given in Theorem 3.5 is naı̈vely more likely to be of minimal degree when the degrees of f2 , . . . , fs are high with respect to the number of variables, implying that the degrees of β2 , . . . , βs are low. Proposition 3.9. Let f1 , . . . , fs have only homogeneous linear relations in the indicator variables and β1 , . . . , βs be a Nullstellensatz certificate. Let β1 be supported on the monomials yb1 , . . . , yb` for b1 , . . . , b` P B. Then if for all j P rns and k P r`s, řs yj ybk “ i“2 αi fi satisfies degpαi q ď degpybk q for all i “ 2, . . . , s, β1 , . . . , βs is a minimum degree Nullstellensatz certificate. In addition, if f2 “ 0, . . . , fs “ 0 is a polynomial system entirely in the indicator variables and there are only homogeneous linear relations, then there is a Nullstellensatz certificate of minimal degree of the form β1 , β21 , . . . , βs1 , where β1 is the coefficient polynomial guaranteed by Theorem 3.5. Proof. We know that f1 β1 ´1 contains řsonly monomials of the form yj ybk for j P rns and k P r`s. By assumption yj ybk “ i“2 αi fi and satisfies degpαi q ď degpybk q for all i “ 2, . . . , s. This implies degpβ1 q ě degpβi q for all i “ 2, . . . , s. Then apply Lemma 3.8. If we restrict our attention to a system f1 “ 0, . . . , fs “ 0 only in the indicator variables, all homogeneous linear řs relations are in the indicator variables. Furthermore, any equation yj ybk “ i“2 αi fi is a homogeneous linear relation in the indicator variables since these are the only variables in the equation. So these too can be ignored. We can then apply Lemma 3.7.  Proposition 3.9 is a generalization of Theorem 2.2 as we can see from Proposition 2.1 that all of the variables are indicator variables. 3.1. Some examples with indicator variables. We now reproduce the first theorem from [21]. This theorem establishes several polynomial systems for finding combinatorial properties of graphs and we shall see that all of them (except for one, which we have omitted from the theorem) satisfy the conditions of Theorem 3.5. Afterwards, we shall present a few new examples that also use indicator variables. Theorem 3.10 ([21]). BART SEVENSTER: , JACOB TURNER: 10 1. A simple graph G “ pV, Eq with vertices numbered 1, . . . , n and edges numbered 1, . . . , e has a planar subgraph with m edges if and only if the following system of equations has a solution: ´m ` ř ti,juPE zij “ 0. 2 ´ zij “ 0 zij śn`e s“1 pxtiuk ´ sq “ 0 śn`e s“1 pytijuk ´ sq “ 0 ztiju pytijuk ´ xtiuk ´ ∆tij,iuk q “ 0 ś3 ztuvu k“1 pytuvuk ´ xtiuk ´ ∆tuv,iuk q “ 0 ś3 ztiju ztuvu k“1 pytijuk ´ ytuvuk ´ ∆tij,uvuk q “ 0 ś3 ztiju ztuvu k“1 pytuvuk ´ ytijuk ´ ∆tuv,ijuk q “ 0 ś3 pxtiuk ´ xtjuk ´ ∆ti,juk q “ 0 śk“1 3 ´ xtiuk ´ ∆tj,iuk q “ 0 k“1 px śtjuk n`e´1 p∆tij,uvuk ´ dq “ 0 d“1 ś n`e´1 p∆tij,iuk ´ dq “ 0 d“1 For k “ 1, 2, 3: ˆ ź sk pxtiuk ´ xtjuk q for for for for for for for for for for for all ti, ju P E. k “ 1, 2, 3 and every i P rns. k “ 1, 2, 3 and ti, ju P E. i P rns, ti, ju P E, k P r3s. i P rns, tu, vu P E, u, v ‰ i. every ti, ju, tu, vu P E. every ti, ju, tu, vu P E. i, j P rns. i, j P rns. ti, ju, tu, vu P E, k P r3s. ti, ju P E, k P r3s. ˙ ź pxtiuk ´ ytuvuk q ź i,jPrns iPrns ti,ju, iăj tu,vuinE tu,vuPE pytijuk ´ ytuvuk q “ 1. 2. A graph G “ pV, Eq with vertices labeled 1, . . . , n has a k-colorable subgraph with m edges if and only if the following systems of equations has a solution: ř ´m ` ti,juPE yij “ 0 2 yij ´ yij “ 0 for ti, ju P E. xki ´ 1 “ 0 for i P rns. yij pxk´1 ` xk´2 xj ` ¨ ¨ ¨ ` xk´1 q “ 0 for ti, ju P E. i i j 3. Let G “ pV, Eq be a simple graph with maximum vertex degree ∆ and vertices labeled 1, . . . , n. Then g has a subgraph with m edges and edge-chromatic number ∆ if and only if the following system of equations has a solution: ř ´m ` ti,juPE yij “ 0 2 yij ´ yij “ 0 for ti, ju P E. yij px∆ ij ´ 1q “ 0 for ti, ju P E. ź si pxij ´ xik q “ 1 for i P rns. j,kPN piq jăk We can also look at the a system of polynomials that ask if there is a subgraph of graph homomorphic to another given graph. This is a generalization of Part 3 of Theorem 3.10 as k-colorable subgraphs can be viewed as subgraphs homomorphic to the complete graph on k vertices. Proposition 3.11. Given two simple graphs G (with vertices labeled 1, . . . , n) and H, there is a subgraph of G “ pV, Eq with m edges homomorphic to H if and only if the following system of equations has a solution: ENUMERATIVE ASPECTS OF NULLSTELLENSATZ CERTIFICATES ´m ` ˆ ˙ ř jPN piq yij ř yij ś tv,wuPEpHq ś vPV pHq ti,juPE 2 yij ´ yij “ 0 yij “ 0 pzi ´ xv q “ 0 pzi ` zj ´ xv ´ xw q “ 0 11 for all ti, ju P E. for all i P rns. for all ti, ju P E. Proof. The variables yij are the indicator variables which designate whether or not an edge of G is included in the subgraph. The third set of equations says that if at least one of the edges incident to vertex i is included in the subgraph, then vertex i must map to a vertex v P V pHq. The last set of equations says that if the edge ti, ju is included in the subgraph, its endpoints must be mapped to the endpoints of an edge in H.  We see that all of these systems of equations have indicator variables and that each system has only finitely many solutions. So Lemma 3.1 says that the rings formedř by taking a quotient by the ideal generated by all equations not of the form n ´m ` i“1 yi gives an Artinian ring. We also see that a subset of k-colorable subgraph is k-colorable, a subset of a planar subgraph is planar, and a subgraph of a k-edge colorable subgraph is k-edge colorable. Lastly, if a subgraph F of G is homomorphic to H, so is any subgraph of F by restricting the homomorphism. So all of these systems are subset closed. Corollary 3.12. If the first system of equations in Theorem 3.10 is infeasible, there is a Nullstellensatz certificate β1 , . . . , βs such that the monomials in β1 are monomials in the variables yi in bijections with the planar subgraphs. If the second system in Theorem 3.10 is infeasible, the same holds example that the monomials are in bijection with the k-colorable subgraphs. If the third system in Theorem 3.10 is infeasible, the same holds except that the monomials are in bijection with the k-edge colorable subgraphs. Lastly, if the system of equations in Proposition 3.11 is infeasible, the same holds except that the monomials are in bijection with the subgraphs homomorphic to H. Proof. This follows directly from Theorem 3.5 and Theorem 3.10.  None of the examples in Theorem 3.10 satisfy the conditions of Proposition 3.9 as the indicator variables are a proper subset of the variables in the system. The following theorem gives a few examples that only involve indicator variables. While only three of the four following examples satisfies the conditions of Proposition 3.9, we will see that Theorem 3.5 can be useful in understanding minimum degree certificates if we can analyze the equations directly. Definition 3.13. Given a graph G, we say that a subgraph H cages a vertex v P V pGq if every edge incident to v in G is an edge in H. Theorem 3.14. 1. A graph G “ pV, Eq with vertices labeled 1, . . . , n has a regular spanning subgraph with m edges if and only if the following system of equations has a solution: ř ´m ` ti,juPE yij “ 0 2 yij ř ř ´ yij “ 0 for all ti, ju P E. for every i, ` P rns. jPN piq yij “ kPN p`q yk` 12 BART SEVENSTER: , JACOB TURNER: Furthermore, if the system is infeasible, there is a minimal degree Nullstellensatz certificate of the form β1 , β21 , . . . , βs1 , where β1 is the coefficient polynomial guaranteed in Theorem 3.5. 2. A graph G “ pV, Eq with vertices labeled 1, . . . , n has a k-regular subgraph with m edges if and only if the following system of equations has a solution: ř ´m ` ti,juPE yij “ 0 2 yij ´ yij “ 0 for all ti, ju P E. ř ř p jPN piq yij qp jPN piq yij ´ kq “ 0 for every i P rns. Furthermore, if the system is infeasible, if there exists an edge in a maximum kregular subgraph such that for both of its endpoints, there is an edge incident to it that is in no maximum k-regular subgraph, then there is a minimal degree Nullstellensatz certificate of the form β1 , β21 , . . . , βs1 , where β1 is the coefficient polynomial guaranteed in Theorem 3.5. 3. A graph G “ pV, Eq with vertices labeled 1, . . . , n has a vertex cover of size m if and only if the following system of equations has a solution: ř ´pn ´ mq ` iPrns yi “ 0 yi2 ´ yi “ 0 for all i P rns. yi yj “ 0 for all ti, ju P E. Furthermore, if the system is infeasible, there is a Nullstellensatz certificate β1 , . . . , βs of minimal degree such the monomials in β1 are in bijection with the independent sets of G. 4. A graph G with vertices labeled 1, . . . , n and e edges has an edge cover of size m if and only if the following system of equations has a solution: ř ´pe ´ mq ` ti,juPE yij “ 0 2 ś yij ´ yij “ 0 for all ti, ju P E. jPN piq yij “ 0 for all i P rns. Furthermore, if the system is infeasible, there is a minimal degree Nullstellensatz certificate β1 , . . . , βs such that the monomials of β1 correspond to the subgraphs of G that cage no vertex of G. Proof. We first prove Part 1. First we show that a solution to the system imply the existence of a regular spanning subgraph of size m. The indicator variables correspond to edges that will either be in a subgraph satisfying the last set of equations or not. The last set of equations say that every pair of vertices must be incident to the same number of edges in the subgraph. The last equations are homogeneous linear equations and so we use Proposition 3.9 to prove that Nullstellensatz certificate guaranteed in Theorem 3.5 is a minimal degree certificate. Now we move to Part 2. Once again, the indicator variables correspond to edges that will either be in the subgraph or not. The last set of equations say that the number of that every vertex must be incident to k edges in the subgraph or 0 edges. The last equations are not homogeneous linear relations. Now suppose that there is an edge ti, ju that is in a maximum k-regular subgraph and ti, `1 u and t`2 , ju are edges in none. ENUMERATIVE ASPECTS OF NULLSTELLENSATZ CERTIFICATES 13 The polynomial β1 in the certificate β1 , . . . , βs given by Theorem 3.5 contains a monomial for every k-regular subgraph. At least one of these monomials contains in which yij appears: ř the variable ř yij . There are only two ř linear relations ř p sPN piq yis qp sPN piq pyis ´ kqq “ 0 and p sPN piq ysi qp sPN pjq pysj ´ kqq “ 0. The former equation involves the variable yi`1 and the latter the variable y`2 j . But neither of these variables appear in monomials of maximal degree by assumption. Therefore monomials of maximal degree involving yij cannot be gotten rid of by the polynomials f2 , . . . , fs . So the total degree of β1 cannot be reduced. Now we prove Part 3. We first consider a different system modeling vertex cover: ř ´m ` iPrns xi “ 0 x2i ´ xi “ 0 for all i P rns. pxi ´ 1qpxj ´ 1q “ 0 for all ti, ju P E. In this system, the indicator variables correspond to vertices that will be either in a vertex cover or not. The last set of equations say that for every edge, at least one of its endpoints must be included the in the vertex cover. However, this system is not subset closed, in fact it is the opposite. If a set is a vertex cover, so is any superset. So for Theorem 3.5 to be applicable, we make the variable change ´yi “ xi ´ 1. Plugging this variable change in gives us the equations in the statement in the theorem and is now subset closed. However, it defines an isomorphic ideal. We then note that these equations model independent set on the same graph and use Theorem 2.2. Lastly, we prove Part 4. Like in Part 3, we first consider the following system: ř ´m ` ti,juPE xij “ 0 x2ij ´ xij “ 0 for all ti, ju P E. ś jPN piq pxij ´ 1q “ 0 for all i P rns. In this system, the indicator variables correspond to edges that are in the edge cover or not. The last equations say that for every vertex, at least one of its incident edges must be in the edge cover. Once again, this system is the opposite of being subset closed: any superset of an edge cover is an edge cover. So once again we make a variable substitution, this time ´yij “ xij ´ 1. Plugging in gives us the system in the statement of the theorem. We then use Proposition 3.9, noting there are no linear relations among the indicator variables, and note that those square free ś monomials that get sent to zero are those divisible by a monomial of the form jPN pjq xij . If a monomial is not divisible by a monomial of such a form, it corresponds to a subgraph that cages no vertex.  We see from Part 2 of Theorem 3.14 that whether or not a minimal degree Nullstellensatz certificate exists that enumerates all combinatorial structures satisfying the polynomial constraints is sensitive to the input data. We also from the proofs of Parts 3 and 4 how Theorem 3.5 might not be applicable. However, in the case of a superset closed system, it is generally possible to change it to a subset closed system using the change of variables exhibited in the proof of Theorem 3.14. While the systems of equations for Parts 3 and 4 of Theorem 3.14 are not the most obvious ones, because they can be obtained from a more straightforward system by a linear change of basis, we have the following Corollary. BART SEVENSTER: , JACOB TURNER: 14 Corollary 3.15. Part 1. A graph G “ pV, Eq with vertices labeled 1, . . . , n has a vertex cover of size m if and only if the following system of equations has a solution: ř ´m ` iPrns xi “ 0 x2i ´ xi “ 0 for all i P rns. pxi ´ 1qpxj ´ 1q “ 0 for all ti, ju P E. Furthermore, if the system is infeasible, the degree of a minimum degree Nullstellensatz certificate is the independence number of G. Part 2. A graph G “ pV, Eq with vertices labeled 1, . . . , n and e edges has an edge cover of size m if and only if the following system of equations has a solution: ř ´m ` ti,juPE xij “ 0 x2ij ´ xij “ 0 for all ti, ju P E. ś jPN piq pxij ´ 1q “ 0 for all i P rns. Furthermore, if the system is infeasible, the degree of a minimum degree Nullstellensatz certificate is equal to the maximum number of edges a subgraph of G can have such that no vertex of G is caged. Proof. For both parts, the correctness of the system of equations was proven in Theorem 3.14. Furthermore, these systems are equivalent to the systems in Parts 3 and 4, respectively, in Theorem 3.14 after an invertible linear change of basis. This means that if β1 , . . . , βs is a minimum degree Nullstellensatz certificate for the systems in Theorem 3.14, then applying an an invertible linear change of basis to the variables in the βi preserves their degrees. Thus the degrees any minimum degree Nullstellensatz certificate for the systems in the statement of the corollary must be the same.  4. Perfect Matchings We now turn our attention to the problem of determining if a graph has a perfect matching via Nullstellensatz certificate methods. Unlike many of the problems considered in the previous section, this problem is not NP-complete. Edmond’s blossom algorithm determines if a graph G “ pV, Eq has a perfect matching in time Op|E||V |1{2 q. The following set of equations has been proposed for determining if a graph G has a perfect matching. ř (1) xij “ 1 xij xjk “ 0 jPN piq i P V pGq @i P V pGq, j, k P N piq where N piq denotes the neighborhood of vertex i. The first equation says that a vertex must be incident to at least one edge in a perfect matching and the second equation says it can be incident to at most one edge. So indeed, these equations are infeasible if and only if G does not have a perfect matching. However, there is not yet a complete understanding of the Nullstellensatz certificates if this system is infeasible [23]. ENUMERATIVE ASPECTS OF NULLSTELLENSATZ CERTIFICATES 15 We note that the equations x2ij ´xij can easily be seen to be in the ideal generated by the Equations 1. Thus the variables xij are indicator variables. However, there ř is no equation of the form ´m ` xij , so we are not in the situation required to apply Theorem 3.5. That said, there still exists a Nullstellensatz certificate such that the non-zero monomials are precisely those corresponding to matchings in the graph G. We observe that a matching on a graph G corresponds precisely to an independent set of its line graph LpGq. In fact, there is a bijection between independent sets of LpGq and matchings of G. This suggests a different set of equations for determining perfect matchings of G that mimic those in Proposition 2.1. x2ij ´ xij “ 0, ti, ju P EpGq, ř xij xik “ 0, ti, ju, ti, ku P EpGq, “ |V pGq|{2. ti,juPEpGq xij (2) It quickly follows from Proposition 2.1 that the solutions to this system forms a zero-dimensional variety whose solutions correspond to perfect matchings. However, we also know from Theorem 2.2 that if the system is infeasible then there is a unique minimum degree Nullstellensatz certificate whose degree is the size of a maximum matching of G. Furthermore, the coefficient polynomial for the equař tion xij “ |V pGq|{2 in this certificate has monomials precisely corresponding to matchings in G. Equations 1 and Equations 2 define the same variety as a set. We now want to find a way of turning a Nullstellensatz certificate for Equations 2 into a Nullstellensatz certificates for Equations 2. This should be possible if Equations 1 and Equations 2 both define the same ideal. It is sufficient to show that both generate a radical ideal. We have the following lemma (cf. [16]). Proposition 4.1 (Seidenberg’s Lemma). Let I Ă krx1 , . . . , xn s be a zero dimensional ideal. Suppose that for every i P rns, there is a non-zero polynomial gi P I X krxi s such that gi has no repeated roots. Then I is radical. We see that the polynomials of the form x2ij ´xij satisfy the conditions of Proposition 4.1 and so both ideals are indeed radical. By Theorem 2.2, if Equations 2 are infeasible, we have a Nullstellensatz certificate of the form ˙ ˆ ÿ ÿ ÿ |V pGq| ` xij ` Qijk xij xik ` Pij px2ij ´ xij q, 1“A ´ 2 ti,ju‰ti,ku ti,juPEpGq ti,juPEpGq PEpGq where A is a polynomial whose monomials are in bijection with matchings of G and all coefficients are positive real numbers. If Equations 1 are infeasible, we denote by the polynomials ∆i and Θijk a Nullstellensatz certificate such that „ ÿ  ÿ ÿ 1“ ∆i p xij q ´ 1 ` Θijk xij xik . iPV pGq jPN piq ti,ju‰ti,kuPEpGq Proposition 4.2. If Equations 1 are infeasible, then there is a Nullstellensatz certificate ∆i , i P V pGq, and Θijk for ti, ju ‰ tj, ku P EpGq such that (a) The degree of each ∆i is the size of a maximal ś matching of G. (b) For every matching M of G, the monomial ti,juPM xij appears with non-zero coefficient in ∆i for all i P V pGq. BART SEVENSTER: , JACOB TURNER: 16 (c) The degree of Θijk is less than or equal to the degree of ∆` for all i, j, k, ` P V pGq. Proof. First we note that „ ÿ  1 ÿ |V pGq| ` p xij q ´ 1 “ ´ 2 2 iPV pGq jPN piq Then for ti, ju P EpGq we have that ˆ Pij px2ij ´ xij q “ Pij xij r´1 ` ÿ So if A, Pij and that if we set xij . ti,juPEpGq ÿ xik s ´ ti,kuPEpGq Qijk ÿ ˙ xij xik . ti,kuPEpGq j‰k are a Nullstellensatz certificate for Equations 2, then we see ÿ 1 Pij xij and A` 2 jPN piq ÿ :“ Qijk ´ Pij xij xik , ∆i :“ Θijk ti,kuPEpGq j‰k that we get a Nullstellensatz certificate for Equations 1. Since degpPij q ă degpAq and both have only positive real coefficients, degp∆i q “ degpAq, which is the size of a maximal matching of G, using Theorem 2.2. This also implies Part (b) of the statement. Lastly, we note that since degpQijk q ď degpAq ´ 2 that Θijk has degree at most degpAq “ degp∆i q, again using Theorem 2.2.  While Proposition 4.2 implies the existence of an enumerative Nullstellensatz certificate similar to that in Theorem 2.2, it is not necessarily of minimal degree. In fact many times it will not be. Consider the following result. Theorem 4.3. A loopless graph G has a degree zero Nullstellensatz certificate β1 , . . . , βs for Equations 1 if and only if G is bipartite and the two color classes are of unequal size. Furthermore, we can choose such a Nullstellensatz certificate such that for each non-zero βi , |βi |´1 can be take to be equal to the difference in size of the independent sets. ř Proof. Let G “ pV, Eq and fi “ jPN piq xij ´ 1 for i P V . Suppose the graph G is 1 bipartite and has two color classes A and B, such that |A| ą |B|. Let c “ |A|´|B| , then we have that ÿ ÿ cfi ` ´cfj “ 1, iPA jPB so this gives a Nullstellensatz certificate of degree 0 for G. Conversely, suppose that G has a degree zero Nullstellensatz certificate β1 , ..., βs . Clearly, the coefficients of the equations of the form xij xjk have to be zero. Now for some vertex vi , let the equation fi have βi “ c, for c P C. Then for all j P N piq we have that βj “ ´c. Repeating this argument, we see that G can not have any odd cycles. Furthermore, for the sum to be unequal to 0, we need the sizes of the color classes to be unequal.  We see that as the size of the graphs being considered grows, the difference in the degree of Nullstellensatz certificate given in Proposition 4.2 and a the degree of a minimal degree certificate can grow arbitrarily large since Theorem 4.3 gives an infinite family of graphs with degree zero Nullstellensatz certificates. ENUMERATIVE ASPECTS OF NULLSTELLENSATZ CERTIFICATES 17 We analyze the time complexity of the NulLA algorithm if it is promised a connected bipartite graph with independent sets of unequal size for returning the result that the equations are infeasible. The algorithm first assumes that the polynomial equations has a Nullstellensatz certificate of degree zero, ř which we know from Theorem 4.3 to be true in this case. Letting fi “ jPN piq xij ´ 1 and i g i “ xij xik , then the algorithm will try to find constants αi and βjk such that řjk ř i i αi fi ` βjk xij xik “ 1. However, we immediately see that βjk “ 0 for all ti, ju, ti, ku P EpGq. So we consider an augmented matrix M |v with columns labeled by the constants i αi , βjk and rows for each linear relation that will be implied among the constants, which we now determine. Each variable xij , ti, ju P EpGq, appears as a linear term in exactly two polynomials: fi and fj . We see that this imposes the relation αi ` αj “ 0 for ti, ju P EpGq. Because of the ´1 appearing in each fi , we also have ř 1 i that iPrns αi “ ´1. Lastly, since each gjk has no monomial in common with gji 1 k1 , i there are no relations among the βjk . So we see that the number of rows of M is |EpGq| ` 1. The matrix M can then be described as follows: If we restrict to the columns labeled by the αi , we get a copy of the incidence matrix of G along with an extra i row of all one’s added to the bottom. The columns labeled by βjk are all zero columns. The augmented column v has a zero in every entry except the last, which is contains a negative one. The NulLA algorithm seeks to determine if this linear system has a solution. Since we have a matrix with |V pGq| nontrivial columns and |EpGq| ` 1 ě |V pGq| rows. This takes time Ωp|V pGq|ω q to run (where ω is some constant ą 2, although conjectured to asymptotically approach 2, depending on the complexity of matrix multiplication [6]). However, two-coloring a graph and counting the size of the two independent sets can be done in time Op|V pGq| ` |EpGq|q “ Op|V pGq|2 q. So even in the best case scenario, the NulLA algorithm is not an optimal algorithm. 4.1. Nullstellensatz certificates for Odd Cliques. We now turn our attention to another question inspired by Proposition 4.2. When is the Nullstellensatz certificate given in that theorem of minimal degree? Surprisingly, this turns out to be the case for odd cliques. This is especially unappealing from an algorithmic standpoint as any graph with an odd number of vertices clearly cannot have a perfect matching. Throughout the rest of the is section, we take G “ Kn , for n odd. To prove our result, we will work over the ring R “ Crxij | ti, ju P rns, i ‰ js{I where I is the ideal generated by the polynomials x2ij ´ xij and xij xik for ti, ju, ti, ku P EpKn q. We will be doing linear algebra over this ring as it is the coefficient polynomials ∆i that we are most interested in. Furthermore, we note that adding the equations x2ij ´ xij does not increase the degree of the polynomials ∆i in a certificate and it is convenient to ignore square terms. ř Workingř over R, we now want to find polynomials ∆i , i P rns, such that iPrns ∆i p jPN piq xij ´ 1q “ 1. Our goal is to prove that each ∆i has degree tn{2u, which is the size of a maximum matching in Kn , for n odd. We already knew from Theorem 4.3 that any Nullstellensatz certificate for Kn must be of degree at least one. For the proof of the statement, it will be convenient BART SEVENSTER: , JACOB TURNER: 18 to alter our notation. We now denote the variable xij for e “ ti, ju P EpKn q as xe . We will also write ∆v for v P V pKn q. Theorem 4.4. The Nullstellensatz certificate given in Proposition 4.2 is a minimal degree certificate for Kn , n odd. Proof. By Proposition 4.2 we know that there exists a Nullstellensatz certificate of degree tn{2u. We work in the ring R :“ Crxij : i ‰ j P rnss{I, where I is generated by the second set of equations inśEquations 1. Let M be the set of matchings of Kn , and, for M P M let xM “ ePM xe . Since we are working in R, by Lemma 3.3, we can write ÿ αv,M xM . ∆v “ M PM A Nullstellensatz certificate gives us that in R ˛ ¨ ÿ ÿ xe ´ 1‚ “ 1. ∆v ˝ ePN pvq vPV pKn q The coefficient of xM is given by ÿ ÿ ÿ ´αv,M ` αv,M ze , ePM vPe vPV pKn q which has to equal zero in a Nullstellensatz certificate if |M | ą 0. Now, if there is a Nullstellensatz certificate of degree l ă tn{2u, then, if |M | “ l ` 1, we see that ÿ RM :“ pαu,M ze ` αv,M ze q “ 0, e“tu,vuPM and by edge transitivity of G, summing over these relations implies that ÿ ÿ αv,M “ 0. vPV pKn q M PM |M |“l Furthermore, we have 0“ ÿ ÿ ÿ ÿ ´αv,M ` vPV pKn q αv,M ze ùñ ePM vPe M PM |M |“l ÿ ÿ 0“ αv,M ze ePM vPe Then summing over the linear relations in the second line above gives ÿ ÿ 0 “ pl ´ 1q αv,M vPV pKn q M PM |M |“l´1 Repeating this, we obtain that ÿ αv,H “ 0, vPV which contradicts the assumption that the ∆v give a Nullstellensatz certificate as we must have ÿ αv,H “ ´1. vPV ENUMERATIVE ASPECTS OF NULLSTELLENSATZ CERTIFICATES 19 Thus we can conclude that there is no Nullstellensatz certificate where all ∆v have degree at most tn{2u ´ 1 in R. But we know from Proposition 4.2 that there exists a Nullstellensatz certificate where each ∆v has degree tn{2u and all other coefficient polynomials have degree at most tn{2u. So this Nullstellensatz certificate is of minimal degree.  So we see that using NulLA to determine if a graph has a perfect matching using Equations 1 can be quite problematic. Since any graph with and odd number of vertices cannot have a perfect matching, the NulLA algorithm does a `lot of ˘ work: a`i for every i P rtn{2us, it determines if a system of linear equations in where i ` ˘ a “ n2 variables, which is the number of monomials in the variables xij of degree i. However, the NulLA algorithm could be made smarter by having it reject any graph on odd vertices before doing any linear algebra. This leads us to an open question: Question 1. Is there a family of graphs, each with even size, none of which have a perfect matching, such that the Nullstellensatz certificate given in Proposition 4.2 is of minimal degree? We actually implemented the NulLA algorithm to try and find examples of graphs with high degree Nullstellensatz certificates for Equations 1. The only ones were graphs containing odd cliques. This leads us to wonder if there are natural ”bad graphs” for the degree of the Nullstellensatz certificate and if their presence as a subgraph determines the minimal degree. Formally: Question 2. Are there finitely many families of graphs G1 , . . . Gk such that the degree of a minimal degree Nullstellensatz certificate for Equations 1 of a graph G is determined by the largest subgraph of G contained in one of the families Gi ? 5. Conclusion In tackling decision problems, it is often a natural idea to rephrase them in some other area of mathematics and use algorithms from said area to see if performance can be improved. The NulLA algorithm is inspired by the idea of rewriting combinatorial decision problems as systems of polynomials and then using Gröbner basis algorithms from computational algebraic geometry to decide this problems quickly. Amazingly, from a theoretical point of view, the rewriting of these problems as polynomial systems is not just a change of language. Lots of combinatorial data seems to come packaged with it. Throughout this paper, we have seen time and again that simply trying to solve the decision problem in graph theory might actually involve enumerating over many subgraphs. The theory of Nullstellensatz certificates is fascinating for the amount of extra information one gets for free by simply writing these problems as polynomial systems. From an algorithmic viewpoint, our results suggest that one should be cautious about using the NulLA algorithm as a practical tool. The NulLA algorithm always finds a minimal degree certificate, and our theorems show that such certificates may entail solving a harder problem that the one intended. However, we do not know which minimal degree Nullstellensatz certificate will get chosen: maybe there are others that are less problematic algorithmically. Certainly, however, work should be done to understand which minimal degree Nullstellensatz certificates will actually be found by the algorithm if there is to be 20 BART SEVENSTER: , JACOB TURNER: any hope in actual computational gains. We have analyzed the worst case scenario, but it is unclear how often it will arise in practice. Acknowledgments. We would like to thank Jeroen Zuiddam for coding a working copy of the NulLA algorithm for our use. The research leading to these results has received funding from the European Research Council under the European Union’s Seventh Framework Programme (FP7/2007-2013) / ERC grant agreement No 339109. References [1] Noga Alon and M Tarsi. Combinatorial nullstellensatz. Combinatorics Probability and Computing, 8(1):7–30, 1999. [2] Noga Alon and Michael Tarsi. Colorings and orientations of graphs. Combinatorica, 12(2):125–134, 1992. [3] Gwénolé Ars, Jean-Charles Faugere, Hideki Imai, Mitsuru Kawazoe, and Makoto Sugita. Comparison between xl and gröbner basis algorithms. In Advances in CryptologyASIACRYPT 2004, pages 338–353. Springer, 2004. [4] W Dale Brownawell. Bounds for the degrees in the nullstellensatz. Annals of Mathematics, 126(3):577–591, 1987. [5] Thomas Brylawski and James Oxley. The Tutte polynomial and its applications. Matroid applications, 40:123–225, 1992. [6] Don Coppersmith and Shmuel Winograd. On the asymptotic complexity of matrix multiplication. SIAM Journal on Computing, 11(3):472–492, 1982. [7] Nicolas Courtois, Alexander Klimov, Jacques Patarin, and Adi Shamir. Efficient algorithms for solving overdefined systems of multivariate polynomial equations. In Advances in Cryptology-EUROCRYPT 2000, pages 392–407. Springer, 2000. [8] David Cox, John Little, and Donal O’shea. Ideals, varieties, and algorithms, volume 3. Springer, 1992. [9] Pierre de la Harpe and Vaughan Frederick Randal Jones. Graph invariants related to statistical mechanical models: examples and problems. Journal of Combinatorial Theory, Series B, 57(2):207–227, 1993. [10] Jesús A de Loera. Gröbner bases and graph colorings. Beiträge zur algebra und geometrie, 36(1):89–96, 1995. [11] Jesús A De Loera, Jon Lee, Peter N Malkin, and Susan Margulies. Hilbert’s nullstellensatz and an algorithm for proving combinatorial infeasibility. In Proceedings of the twenty-first international symposium on Symbolic and algebraic computation, pages 197–206. ACM, 2008. [12] Jesús A De Loera, Jon Lee, Peter N Malkin, and Susan Margulies. Computing infeasibility certificates for combinatorial problems through hilberts nullstellensatz. Journal of Symbolic Computation, 46(11):1260–1283, 2011. [13] Jesús A De Loera, Susan Margulies, Michael Pernpeintner, Eric Riedl, David Rolnick, Gwen Spencer, Despina Stasi, and Jon Swenson. Graph-coloring ideals: Nullstellensatz certificates, gröbner bases for chordal graphs, and hardness of gröbner bases. In Proceedings of the 2015 ACM on International Symposium on Symbolic and Algebraic Computation, pages 133–140. ACM, 2015. [14] Shalom Eliahou. An algebraic criterion for a graph to be four-colourable. 1992. [15] János Kollár. Sharp effective nullstellensatz. Journal of the American Mathematical Society, 1(4):963–975, 1988. [16] Martin Kreuzer and Lorenzo Robbiano. Computation Commutative Algebra I. Springer Verlag, Heidelberg, 2000. [17] Jean Lasserre. Polynomials nonnegative on a grid and discrete optimization. Transactions of the American Mathematical Society, 354(2):631–649, 2002. [18] Monique Laurent. Semidefinite representations for finite varieties. Mathematical Programming, 109(1):1–26, 2007. [19] Monique Laurent and Franz Rendl. Semidefinite programming and integer programming. Handbooks in Operations Research and Management Science, 12:393–514, 2005. ENUMERATIVE ASPECTS OF NULLSTELLENSATZ CERTIFICATES 21 [20] Shuo-Yen Robert Li et al. Independence numbers of graphs and generators of ideals. Combinatorica, 1(1):55–61, 1981. [21] J. a. Loera, J. Lee, S. Margulies, and S. Onn. Expressing combinatorial problems by systems of polynomial equations and hilbert’s nullstellensatz. Comb. Probab. Comput., 18(4):551–582, July 2009. [22] László Lovász. Stable sets and polynomials. Discrete mathematics, 124(1):137–153, 1994. [23] Susan Margulies. private communication, 2016. [24] Susan Margulies, Shmuel Onn, and Dmitrii V Pasechnik. On the complexity of hilbert refutations for partition. Journal of Symbolic Computation, 66:70–83, 2015. [25] Yuri Vladimirovich Matiyasevich. Some algebraic methods for calculation of the number of colorings of a graph. Zapiski Nauchnykh Seminarov POMI, 283:193–205, 2001. [26] James G Oxley. Matroid theory, volume 1997. Oxford university press Oxford, 1992. [27] Pablo A Parrilo. Semidefinite programming relaxations for semialgebraic problems. Mathematical programming, 96(2):293–320, 2003. [28] Aron Simis, Wolmer V Vasconcelos, and Rafael H Villarreal. On the ideal theory of graphs. Journal of Algebra, 167(2):389–416, 1994. [29] William T Tutte. A contribution to the theory of chromatic polynomials. Canad. J. Math, 6(80-91):3–4, 1954. [30] Neil White. Combinatorial geometries, volume 29. Cambridge University Press, 1987. 

Routing and Staffing when Servers are Strategic Ragavendran Gopalakrishnan Xerox Research Centre India, Bangalore, Karnataka 560103, Ragavendran.Gopalakrishnan@xerox.com arXiv:1402.3606v4 [cs.GT] 23 Mar 2016 Sherwin Doroudi Tepper School of Business, Carnegie Mellon University, Pittsburgh, PA 15213, sdoroudi@andrew.cmu.edu Amy R. Ward Marshall School of Business, University of Southern California, Los Angeles, CA 90089, amyward@marshall.usc.edu Adam Wierman Department of Computing and Mathematical Sciences, California Institute of Technology, Pasadena, CA 91125, adamw@caltech.edu Traditionally, research focusing on the design of routing and staffing policies for service systems has modeled servers as having fixed (possibly heterogeneous) service rates. However, service systems are generally staffed by people. Furthermore, people respond to workload incentives; that is, how hard a person works can depend both on how much work there is, and how the work is divided between the people responsible for it. In a service system, the routing and staffing policies control such workload incentives; and so the rate servers work will be impacted by the system’s routing and staffing policies. This observation has consequences when modeling service system performance, and our objective in this paper is to investigate those consequences. We do this in the context of the M /M /N queue, which is the canonical model for large service systems. First, we present a model for “strategic” servers that choose their service rate in order to maximize a tradeoff between an “effort cost”, which captures the idea that servers exert more effort when working at a faster rate, and a “value of idleness”, which assumes that servers value having idle time. Next, we characterize the symmetric Nash equilibrium service rate under any routing policy that routes based on the server idle time (such as the longest idle server first policy). We find that the system must operate in a quality-driven regime, in which servers have idle time, in order for an equilibrium to exist. The implication is that to have an equilibrium solution the staffing must have a first-order term that strictly exceeds that of the common square-root staffing policy. Then, within the class of policies that admit an equilibrium, we (asymptotically) solve the problem of minimizing the total cost, when there are linear staffing costs and linear waiting costs. Finally, we end by exploring the question of whether routing policies that are based on the service rate, instead of the server idle time, can improve system performance. Key words : service systems; staffing; routing; scheduling; routing; strategic servers Subject classifications : Primary: Queues: applications, limit theorems; secondary: Games/group decisions: noncooperative 1. Introduction. There is a broad and deep literature studying the scheduling and staffing of service systems that bridges operations research, applied probability, and computer science. This literature has had, and is continuing to have, a significant practical impact on the design of call centers (see, for example, the survey papers [18] and [1]), health care systems (see, for example, the recent book [29]), and large-scale computing systems (see, for example, the recent book [26]), among other areas. Traditionally, this literature on scheduling and staffing has modeled the servers of the system as having fixed (possibly heterogeneous) service rates and then, given these rates, scheduling and staffing policies are proposed and analyzed. However, in reality, when the servers are people, the rate a server chooses to work can be, and often is, impacted by the scheduling and staffing policies used by the system. For example, if requests are always scheduled to the “fastest” server whenever that server is available, then this server may have the incentive to slow her rate to avoid being overloaded with 1 2 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic work. Similarly, if extra staff is always assigned to the division of a service system that is the busiest, then servers may have the incentive to reduce their service rates in order to ensure their division is assigned the extra staff. The previous two examples are simplistic; however, strategic behavior has been observed in practice in service systems. For example, empirical data from call centers shows many calls that last near 0 seconds [18]. This strategic behavior of the servers allowed them to obtain “rest breaks” by hanging up on customers – a rather dramatic means of avoiding being overloaded with work. For another example, academics are often guilty of strategic behavior when reviewing for journals. It is rare for reviews to be submitted before an assigned deadline since, if someone is known for reviewing papers very quickly, then they are likely to be assigned more reviews by the editor. Clearly, the strategic behavior illustrated by the preceding examples can have a significant impact on the performance provided by a service system. One could implement a staffing or scheduling policy that is provably optimal under classical scheduling models, where servers are nonstrategic, and end up with far from optimal system performance as a result of undesirable strategic incentives created by the policy. Consequently, it is crucial for service systems to be designed in a manner that provides the proper incentives for such “strategic servers”. In practice, there are two approaches used for creating the proper incentives for strategic servers: one can either provide structured bonuses for employees depending on their job performance (performance-based payments) or one can provide incentives in how scheduling and staffing is performed that reward good job performance (incentive-aware scheduling). While there has been considerable research on how to design performance-based payments in the operations management and economics communities; the incentives created by scheduling and staffing policies are much less understood. In particular, the goal of this paper is to initiate the study of incentive-aware scheduling and staffing policies for strategic servers. The design of incentive-aware scheduling and staffing policies is important for a wide variety of service systems. In particular, in many systems performance-based payments such as bonuses are simply not possible, e.g., in service systems staffed by volunteers such as academic reviewing. Furthermore, many service systems do not use performance-based compensation schemes; for example, the 2005 benchmark survey on call center agent compensation in the U.S. shows that a large fraction of call centers pay a fixed hourly wage (and have no performance-based compensation) [3]. Even when performance-based payments are possible, the incentives created by scheduling and staffing policies impact the performance of the service system, and thus impact the success of performance-based payments. Further, since incentive-aware scheduling and staffing does not involve monetary payments (beyond a fixed employee salary), it may be less expensive to provide incentives through scheduling and staffing than through monetary bonuses. Additionally, providing incentives through scheduling and staffing eliminates many concerns about “unfairness” that stem from differential payments to employees. Of course, the discussion above assumes that the incentives created by scheduling and staffing can be significant enough to impact the behavior. A priori it is not clear if they are, since simply changing the scheduling and staffing policies may not provide strong enough incentives to strategic servers to significantly change service rates, and thus system performance. It is exactly this uncertainty that motivates the current paper, which seeks to understand the impact of the incentives created by scheduling and staffing, and then to design incentive-aware staffing and scheduling policies that provide near-optimal system performance without the use of monetary incentives. 1.1. Contributions of this paper. This paper makes three main contributions. We introduce a new model for the strategic behavior of servers in large service systems and, additionally, we initiate the study of staffing and routing in the context of strategic servers. Each of these contributions is described in the following. Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic 3 Modeling Strategic Servers (Sections 2 and 3): The essential first step for an analysis of strategic servers is a model for server behavior that is simple enough to be analytically tractable and yet rich enough to capture the salient influences on how each server may choose her service rate. Our model is motivated by work in labor economics that identifies two main factors that impact the utility of agents: effort cost and idleness. More specifically, it is common in labor economics to model agents as having some “effort cost” function that models the decrease in utility which comes from an increase in effort [12]. Additionally, it is a frequent empirical observation that agents in service systems engage in strategic behavior to increase the amount of idle time they have [18]. The key feature of the form of the utility we propose in Section 2 is that it captures the inherent trade-off between idleness and effort. In particular, a faster service rate would mean quicker completion of jobs and might result in a higher idle time, but it would also result in a higher effort cost. In Section 3 of this paper, we apply our model in the context of a M /M /N system, analyzing the first order condition, and provide a necessary and sufficient condition for a solution to the first order condition to be a symmetric equilibrium service rate (Theorem 4). In addition, we discuss the existence of solutions to the first order condition, and provide a sufficient condition for a unique solution (Theorem 5). These results are necessary in order to study staffing and routing decisions, as we do in Sections 4 and 5; however, it is important to note that the model is applicable more generally as well. Staffing Strategic Servers (Section 4): The second piece of the paper studies the impact strategic servers have on staffing policies in multi-server service systems. The decision of a staffing level for a service system has a crucial impact on the performance of the system. As such, there is a large literature focusing on this question in the classical, nonstrategic, setting, and the optimal policy is well understood. In particular, the number of servers that must be staffed to ensure stability in a conventional M /M /N queue with arrival rate λ and fixed service rate µ should be strictly larger than the offered load, λ/µ. However, when there are linear staffing and waiting costs, the economically optimal number of servers to staff is more. Specifically, the optimal policy employs the square root of the offered load more servers [8]. This results in efficient operation, because the system loading factor λ/(N µ) is close to one; √ and maintains quality of service, because the customer wait times are small (on the order of 1/ λ). Thus, this is often referred to as the Quality and Efficiency Driven (QED) regime or as square-root staffing. Our contribution in this paper is to initiate the study of staffing strategic servers. In the presence of strategic servers, the offered load depends on the arrival rate, the staffing, and the routing, through the servers’ choice of their service rate. We show that an equilibrium service rate exists only if the number of servers staffed is order λ more than the aforementioned square-root staffing (Theorem 7). In particular, the system must operate in a quality-driven regime, in which the servers have idle time, instead of the quality-and-efficiency driven regime that arises under square-root staffing, in which servers do not have idle time. Then, within the set of policies that admit an equilibrium service rate, we (asymptotically) solve the problem of minimizing the total cost, when there are linear staffing costs and linear waiting costs (Theorem 8). Routing to Strategic Servers (Section 5): The final piece of this paper studies the impact of strategic servers on the design of scheduling policies in multi-server service systems. When servers are not strategic, how to schedule (dispatch) jobs to servers in multi-server systems is well understood. In particular, the most commonly proposed policies for this setting include Fastest Server First (FSF), which dispatches arriving jobs to the idle server with the fastest service rate; Longest Idle Server First (LISF), which dispatches jobs to the server that has been idle for the longest period of time; and Random, which dispatches the job to each idle server with equal probability. When strategic servers are not considered, FSF is the natural choice for reducing the mean response time (though it is not optimal in general [16, 35]). However, in the context of strategic servers the story changes. In particular, we prove that FSF has no symmetric equilibria when strategic servers 4 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic are considered, even when there are just two servers. Further, we prove that LISF, a commonly suggested policy for call centers due to its fairness properties, has the same, unique, symmetric equilibrium as random dispatching. In fact, we prove that there is a large policy-space collapse – all routing policies that are idle-time-order-based are equivalent in a very strong sense (Theorem 9). With this in mind, one might suggest that Slowest Server First (SSF) would be a good dispatch policy, since it could incentivize servers to work fast; however, we prove that, like FSF, SSF has no symmetric equilibria (Theorem 10). However, by “softening” SSF’s bias toward slow servers, we are able to identify policies that are guaranteed to have a unique symmetric equilibrium and provide mean response times that are smaller than that under LISF and Random (Theorem 11). A key message provided by the results described above is that scheduling policies must carefully balance two conflicting goals in the presence of strategic servers: making efficient use of the service capacity (e.g., by sending work to fast servers) while still incentivizing servers to work fast (e.g., by sending work to slow servers). While these two goals are inherently in conflict, our results show that it is possible to balance them in a way that provides improved performance over Random. 1.2. Related work. As we have already described, the question of how to route and staff in many-server systems when servers have fixed, nonstrategic, service rates is well-studied. In general, this is a very difficult question, because the routing depends on the staffing and vice versa. However, when all the servers serve at the same rate, the routing question is moot. Then, [8] shows that square-root staffing, first introduced in [17] and later formalized in [25], is economically optimal when both staffing and waiting costs are linear. Furthermore, square root staffing is remarkably robust: there is theoretical support for why it works so well for systems of moderate size [30], and it continues to be economically optimal both when abandonment is added to the M /M /N model [19] and when there is uncertainty in the arrival rate [33]. Hence, to study the joint routing and staffing question for more complex systems, that include heterogeneous servers that serve at different rates and heterogeneous customers, many authors have assumed square root staffing and show how to optimize the routing for various objective functions (see, for example, [4, 23, 6, 40, 41]). In relation to this body of work, this paper shows that scheduling and routing results for classical many-server systems that assume fixed service rates must be revisited when servers exhibit strategic behavior. This is because they may not admit a symmetric equilibrium service rate in the case of square-root staffing (see Section 4) or be feasible in the case of Fastest Server First routing (see Section 5). Importantly, the Fastest Server First routing policy mentioned earlier has already been recognized to be potentially problematic because it may be perceived as “unfair”. The issue from an operational standpoint is that there is strong indication in the human resource management literature that the perception of fairness affects employee performance [15, 14]. This has motivated the analysis of “fair” routing policies that, for example, equalize the cumulative server idleness [7, 38], and the desire to find an optimal “fair” routing policy [5, 42]. Another approach is to formulate a model in which the servers choose their service rate in order to balance their desire for idle time (which is obtained by working faster) and the exertion required to serve faster. This leads to a non-cooperative game for a M /M /N queue in which the servers act as strategic players that selfishly maximize their utility. Finally, the literature that is, perhaps, most closely related to the current paper is the literature on queueing games, which is surveyed in [28]. The bulk of this literature focuses on the impact of customers acting strategically (e.g., deciding whether to join and which queue to join) on queueing performance. Still, there is a body of work within this literature that considers settings where servers can choose their service rate, e.g., [31, 21, 10, 11]. However, in all of the aforementioned papers, there are two servers that derive utility from some monetary compensation per job or per unit of service that they provide, and there are no staffing decisions. In contrast, our work considers systems with more than two servers, and considers servers that derive utility from idle time (and Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic 5 have a cost of effort). The idea that servers value idle time is most similar to the setting in [20], but that paper restricts its analysis to a two server model. Perhaps the closest previous work to the current paper in analysis spirit is [2], which characterizes approximate equilibria in a market with many servers that compete on price and service level. However, this is similar in theme to [31, 10] in the sense that they consider servers as competing firms in a market. This contrasts with the current paper, where our focus is on competition between servers within the same firm. 2. A model for strategic servers. The objective of this paper is to initiate an investigation into the effects of strategic servers on classical management decisions in service systems, e.g., staffing and routing. We start by, in this section, describing formally our model for the behavior of a strategic server. The term “strategic server” could be interpreted in many ways depending on the server’s goal. Thus, the key feature of the model is the utility function for a strategic server. Our motivation comes from a service system staffed by people who are paid a fixed wage, independent of performance. In such settings, one may expect two key factors to have a first-order impact on the experience of the servers: the amount of effort they put forth and the amount of idle time they have. Thus, a first-order model for the utility of a strategic server is to linearly combine the cost of effort with the idle time of the server. This gives the following form for the utility of server i in a service system with N servers: Ui (µ) = Ii (µ) − c(µi ), i ∈ {1, . . . , N }, (1) where µ is a vector of the rate of work chosen by each server (i.e., the service rate vector), Ii (µ) is the time-average idle time experienced by server i given the service rate vector µ, and c(µi ) is the effort cost of server i. We take c to be an increasing, convex function which is the same for all servers. We assume that the strategic behavior of servers (choosing a utility-maximizing service rate) is independent of the state of the system and that the server has complete information about the steady state properties of the system when choosing a rate, i.e., they know the arrival rate, scheduling policy, staffing policy, etc., and thus can optimize Ui (µ). The key feature of the form of the utility in (1) is that it captures the inherent trade-off between idleness and effort. The idleness, and hence the utility, is a steady state quantity. In particular, a faster service rate would mean quicker completion of jobs and might result in higher idle time in steady state, but it would also result in a higher effort cost. This trade-off then creates a difficult challenge for staffing and routing in a service system. To increase throughput and decrease response times, one would like to route requests to the fastest servers, but by doing so the utility of servers decreases, making it less desirable to maintain a fast service rate. Our model should be interpreted as providing insight into the systemic incentives created by scheduling and staffing policies rather than the transitive incentives created by the stochastic behavior of the system. Our focus in this paper will be to explore the consequences of strategic servers for staffing and routing in large service systems, specifically, in the M /M /N setting. However, the model is generic and can be studied in non-queueing contexts as well. To quickly illustrate the issues created by strategic servers, a useful example to consider is that of a M /M /1 queue with a strategic server. Example 1 (The M/M/1 queue with a strategic server). In a classic M /M /1 system, jobs arrive at rate λ into a queue with an infinite buffer, where they wait to obtain service from a single server having fixed service rate µ. When the server is strategic, instead of serving at a fixed rate µ, the server chooses her service rate µ > λ in order to maximize the utility in (1). To understand what service rate will emerge, recall that in a M /M /1 queue with µ > λ the steady 6 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic state fraction of time that the server is idle is given by I(µ) = 1 − µλ . Substituting this expression into (1) means that the utility of the server is given by the following concave function: U (µ) = 1 − λ − c(µ). µ We now have two possible scenarios. First, suppose that c′ (λ) < 1/λ, so that the cost function does not increase too fast. Then, U (µ) attains a maximum in (λ, ∞) at a unique point µ⋆ , which is the optimal (utility maximizing) operating point for the strategic server. Thus, a stable operating point emerges, and the performance of this operating point can be derived explicitly when a specific form of a cost function is considered. On the other hand, if c′ (λ) ≥ 1/λ, then U (µ) is strictly decreasing in (λ, ∞) and hence does not attain a maximum in this interval. We interpret this case to mean that the server’s inherent skill level (as indicated by the cost function) is such that the server must work extremely hard just to stabilize the system, and therefore should not have been hired in the first place. For example, consider the class of cost functions c(µ) = cE µp . If c(λ) < p1 , then µ⋆ solves µ⋆ c(µ⋆ ) = 1   p+1 λ λ ⋆ > λ. On the other hand, if c(λ) ≥ p1 , then U (µ) is strictly decreasing , which gives µ = cE p p in (λ, ∞) and hence does not attain a maximum in this interval. Before moving on to the analysis of the M /M /N model with strategic servers, it is important to point out that the model we study focuses on a linear trade-off between idleness and effort. There are certainly many generalizations that are interesting to study in future work. One particularly interesting generalization would be to consider a concave (and increasing) function of idle time in the utility function, since it is natural that the gain from improving idle time from 10% to 20% would be larger than the gain from improving idle time from 80% to 90%. A preliminary analysis highlights that the results in this paper would not qualitatively change in this context.1 3. The M /M /N queue with strategic servers. Our focus in this paper is on the staffing and routing decisions in large service systems, and so we adopt a classical model of this setting, the M /M /N , and adjust it by considering strategic servers, as described in Section 2. The analysis of staffing and routing policies is addressed in Sections 4 and 5, but before moving to such questions, we start by formally introducing the M /M /N model, and performing some preliminary analysis that is useful both in the context of staffing and routing. 3.1. Model and notation. In a M /M /N queue, customers arrive to a service system having N servers according to a Poisson process with rate λ. Delayed customers (those that arrive to find all servers busy) are served according to the First In First Out (FIFO) discipline. Each server is fully capable of handling any customer’s service requirements. The time required to serve each customer is independent and exponential, and has a mean of one time unit when the server works at rate one. However, each server strategically chooses her service rate to maximize her own (steady state) utility, and so it is not a priori clear what the system service rates will be. 1 Specifically, if g(Ii (µ)) replaces Ii (µ) in (1), all the results in Section 3 characterizing equilibria service rates are maintained so long as g ′′′ < 0, except for Theorem 5, whose sufficient condition would have to be adjusted to accommodate g. In addition, our results could be made stronger depending on the specific form of g. For example, if g is such that limµi →µi + Ui (µ) = −∞, then, a preliminary analysis reveals that it would not be necessary to impose the stability constraint µi > λ/N exogenously. Moreover, every solution to the symmetric first order condition (9) would be a symmetric equilibrium (i.e., the sufficient condition of Theorem 4 as generalized for this case by Footnote 2 would automatically be satisfied). Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic 7 In this setting, the utility functions that the servers seek to maximize are given by Ui (µ; λ, N, R) = Ii (µ; λ, N, R) − c(µi), i ∈ {1, . . . , N }, (2) where µ is the vector of service rates, λ is the arrival rate, N is the number of servers (staffing level), and R is the routing policy. Ii (µ; λ, N, R) is the steady state fraction of time that server i is idle. c(µ) is an increasing, convex function with c′′′ (µ) ≥ 0, that represents the server effort cost. Note that, as compared with (1), we have emphasized the dependence on the arrival rate λ, staffing level N , and routing policy of the system, R. In the remainder of this article, we expose or suppress the dependence on these additional parameters as relevant to the discussion. In particular, note that the idle time fraction Ii (and hence, the utility function Ui ) in (2) depends on how arriving customers are routed to the individual servers. There are a variety of routing policies that are feasible for the system manager. In general, the system manager may use information about the order in which the servers became idle, the rates at which servers have been working, etc. This leads to the possibility of using simple policies such as Random, which chooses an idle server to route to uniformly at random, as well as more complex policies such as Longest/Shortest Idle Server First (LISF/SISF) and Fastest/Slowest Server First (FSF/SSF). We study the impact of this decision in detail in Section 5. Given the routing policy chosen by the system manager and the form of the server utilities in (2), the situation that emerges is a competition among the servers for the system idle time. In particular, the routing policy yields a division of idle time among the servers, and both the division and the amount of idle time will depend on the service rates chosen by the servers. As a result, the servers can be modeled as strategic players in a noncooperative game, and thus the operating point of the system is naturally modeled as an equilibrium of this game. In particular, a Nash equilibrium of this game is a set of service rates µ⋆ , such that, Ui (µ⋆i , µ⋆−i ; R) = max Ui (µi , µ⋆−i ; R), λ µi > N (3) where µ⋆−i = (µ⋆1 , . . . , µ⋆i−1 , µ⋆i+1 , . . . , µ⋆N ) denotes the vector of service rates of all the servers except server i. Note that we exogenously impose the (symmetric) constraint that each server must work at a rate strictly greater than Nλ in order to define a product action space that ensures the stability of the system.2 Such a constraint is necessary to allow steady state analysis, and does not eliminate any feasible symmetric equilibria. We treat this bound as exogenously fixed, however in some situations a system manager may wish to impose quality standards on servers, which would correspond to imposing a larger lower bound (likely with correspondingly larger payments for servers). Investigating the impact of such quality standards is an interesting topic for future work. Our focus in this paper is on symmetric Nash equilibria. With a slight abuse of notation, we say that µ⋆ is a symmetric Nash equilibrium if µ⋆ = (µ⋆ , . . . , µ⋆ ) is a Nash equilibrium (solves (3)). Throughout, the term “equilibrium service rate” means a symmetric Nash equilibrium service rate. We focus on symmetric Nash equilibria for two reasons. First, because the agents we model intrinsically have the same skill level (as quantified by the effort cost functions), a symmetric 2 One can imagine that servers, despite being strategic, would endogenously stabilize the system. To test this, one could study a related game where the action sets of the servers are (0, ∞). Then, the definition of the idle time Ii (µ) must be extended into the range of µ for which the system is overloaded; a natural way to do so is to define it to be zero in this range, which would ensure continuity at µ for which the system is critically loaded. However, it is not differentiable there, which necessitates a careful piecewise analysis. A preliminary analysis indicates that in  λ this scenario, no µ ∈ 0, N can ever be a symmetric equilibrium, and then, the necessary and sufficient condition of Theorem 4 would become U (µ⋆ , µ⋆ ) ≥ limµ1 →0+ U (µ1 , µ⋆ ), which is more demanding than (10) (e.g., it imposes a finite upper bound on µ⋆ ), but not so much so that it disrupts the staffing results that rely on this theorem (e.g., Lemma 1 still holds). 8 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic equilibrium corresponds to a fair outcome. As we have already discussed, this sort of fairness is often crucial in service organizations [15, 14, 5]. A second reason for focusing on symmetric equilibria is that analyzing symmetric equilibria is already technically challenging, and it is not clear how to approach asymmetric equilibria in the contexts that we consider. Note that we do not rule out the existence of asymmetric equilibria; in fact, they likely exist, and it would be interesting to study whether they lead to better or worse system performance than their symmetric counterparts. 3.2. The M /M /N queue with strategic servers and Random routing. Before analyzing staffing and routing in detail, we first study the M /M /N queue with strategic servers and Random routing. We focus on Random routing first because it is, perhaps, the most commonly studied policy in the classical literature on nonstrategic servers. Further, this importance is magnified by a new “policy-space collapse” result included in Section 5.1.1, which shows that all idle-time-order-based routing policies (e.g., LISF and SISF) have equivalent steady state behavior, and thus have the same steady state behavior as Random routing. We stress that this result stands on its own in the classical, nonstrategic setting of a M /M /N queue with heterogeneous service rates, but is also crucial to analyze routing to strategic servers (Section 5). The key goal in analyzing a queueing system with strategic servers is to understand the equilibria service rates, i.e., show conditions that guarantee their existence and characterize the equilibria when they exist. Theorems 4 and 5 of Section 3.2.2 summarize these results for the M /M /N queue with Random routing. However, in order to obtain such results we must first characterize the idle time in a M /M /N system in order to be able to understand the “best responses” for servers, and thus analyze their equilibrium behavior. Such an analysis is the focus of Section 3.2.1. 3.2.1. The idle time of a tagged server. In order to characterize the equilibria service rates, a key first step is to understand the idle time of a M /M /N queue. This is, of course, a well-studied model, and so one might expect to be able to use off-the-shelf results. While this is true when the servers are homogeneous (i.e., all the server rates are the same), for heterogeneous systems, closed form expressions are challenging to obtain in general, and the resulting forms are quite complicated [22]. To characterize equilibria, we do need to understand the idle time of heterogeneous M /M /N queues. However, due to our focus on symmetric equilibria, we only need to understand a particular, mild, form of heterogeneity. In particular, we need only understand the best response function for a “deviating server” when all other servers have the same service rate. Given this limited form of heterogeneity, the form of the idle time function simplifies, but still remains quite complicated, as the following theorem shows. Theorem 1. Consider a heterogenous M /M /N system with Random routing and arrival rate λ > 0, where N − 1 servers operate at rate µ > Nλ , and a tagged server operates at rate µ1 > µ1 = + (λ − (N − 1)µ) . The steady state probability that the tagged server is idle is given by:   −1   ρ ErlC(N, ρ)  µ  ρ    1− 1+ 1− I(µ1 , µ; λ, N ) = 1 − , N N µ1 N − ρ + 1 − µ1  µ where ρ = λµ , and ErlC(N, ρ) denotes the Erlang C formula, given by: ErlC(N, ρ) = ρN N N ! N −ρ PN −1 ρj ρN N j=0 j! + N ! N −ρ . (4) 9 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic In order to understand this idle time function more, we derive expressions for the first two derivatives of I with respect to µ1 in the following theorem. These results are crucial to the analysis of equilibrium behavior. Theorem 2. The first two partial derivatives of I with respect to µ1 are given by     µ1 µ1 ∂I ErlC(N, ρ) I2 λ  ErlC(N, ρ)   + 1−  (5) = 2     1 + ∂µ1 µ1 N − ρ µ µ N − ρ + 1 − µ1 2 N − ρ + 1 − µµ1 µ       2  µ ErlC(N, ρ) ∂2I ρ ErlC(N, ρ) µ 2I 3 λ  ErlC(N, ρ) 1 1   + N − 1−  =− 3  2  1 +   3  1 −  µ1 µ1 µ1 ∂µ21 µ1 N − ρ µ µ N − ρ + 1 − N − ρ+1− µ N − ρ+1− µ µ (6) Importantly, it can be shown that the right hand side of (5) is always positive, and therefore, the idle time is increasing in the service rate µ1 , as expected. However, it is not clear through inspection of (6) whether the second derivative is positive or negative. Our next theorem characterizes the second derivative, showing that the idle time could be convex at µ1 = µ1 to begin with, but if so, then as µ1 increases, it steadily becomes less convex, and is eventually concave. This behavior adds considerable complication to the equilibrium analysis. Theorem 3. The second derivative of the idle time satisfies the following properties: † ∂2I ∂2I (a) There exists a threshold µ†1 ∈ [µ1 , ∞) such that ∂µ 2 > 0 for µ < µ1 < µ1 , and ∂µ2 < 0 for 1 1 µ†1 < µ1 < ∞. ∂2I ∂3I (b) ∂µ 2 > 0 ⇒ ∂µ3 < 0. 1 1 1 We remark that it is possible that the threshold µ† could be greater than Nλ , so, restricting the service rate of server 1 to be greater than Nλ does not necessarily simplify the analysis. 3.2.2. Symmetric equilibrium analysis for a finite system. The properties of the idle time function derived in the previous section provide the key tools we need to characterize the symmetric equilibria service rates under Random routing for a M /M /N system. To characterize the symmetric equilibria, we consider the utility of a tagged server, without loss of generality, server 1, under the mildly heterogeneous setup of Theorem 1. We denote it by U (µ1 , µ; λ, N ) = I(µ1 , µ; λ, N ) − c(µ1 ) (7) For a symmetric equilibrium in ( Nλ , ∞), we explore the first order and second order conditions for U as a function of µ1 to have a maximum in (µ1 , ∞). The first order condition for an interior local maximum at µ1 is given by: ∂U =0 ∂µ1 =⇒ ∂I = c′ (µ1 ) ∂µ1 (8) Since we are interested in a symmetric equilibrium, we analyze the symmetric first order condition, obtained by plugging in µ1 = µ in (8):    ∂U λ λ λ = 0 =⇒ N − + ErlC N, = c′ (µ) (9) ∂µ1 µ1 =µ N 2 µ2 µ µ Now, suppose that µ⋆ > Nλ satisfies the symmetric first order condition (9). Then, µ1 = µ⋆ is a stationary point of U (µ1 , µ⋆ ). It follows then, that µ⋆ will be a symmetric equilibrium for the servers (satisfying (3)) if and only if U (µ1 , µ⋆ ) attains a global maximum at µ1 = µ⋆ in the interval ( Nλ , ∞). While an obvious necessary condition for this is that U (µ⋆ , µ⋆ ) ≥ U ( Nλ , µ⋆ ), we show, perhaps surprisingly, that it is also sufficient, in the following theorem. 10 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic Theorem 4. µ⋆ > Nλ is a symmetric equilibrium if and only if it satisfies the symmetric first order condition (9), and the inequality U (µ⋆ , µ⋆ ) ≥ U ( Nλ , µ⋆ ), i.e.,  λ c(µ) ≤ c N   −1 !−1 ρ ErlC(N, ρ) ρ 1+ 1− + . + 1− N N N −1  (10) Finally, we need to understand when the symmetric first order condition (9) admits a feasible solution µ⋆ > Nλ . Towards that, we present sufficient conditions for at least one feasible solution, as well as for a unique feasible solution.
Theorem 5. If c′ Nλ < λ1 , then the symmetric first order condition (9) has at least one solu


2
tion for µ in Nλ , ∞ . In addition, if 2 Nλ c′ Nλ +
Nλ c′′ Nλ ≥ 1, then the symmetric first order condition (9) has a unique solution for µ in Nλ , ∞ . In the numerical results that follow, we see instances of zero, one, and two equilibria.3 Interestingly, when more than one equilibrium exists, the equilibrium with the largest service rate, which leads to best system performance, also leads to highest server utility, and hence is also most preferred by the servers, as the following theorem shows. µ⋆1 Theorem 6. If the symmetric first order condition (9) has two solutions, say µ⋆1 and µ⋆2 , with > µ⋆2 > Nλ , then U (µ⋆1 , µ⋆1 ) > U (µ⋆2 , µ⋆2 ). 3.3. Numerical examples. Because of the complexity of the expression for the equilibrium service rate(s) given by the first order condition (9) and the possibility of multiple equilibria, we discuss a few numerical examples here in order to provide intuition. In addition, we point out some interesting characteristics that emerge as a consequence of strategic server behavior. We present two sets of graphs below: one that varies the arrival rate λ while holding the staffing level fixed at N = 20 (Figure 1), and one that varies the staffing level N while holding the arrival rate fixed at λ = 2 (Figure 2). In each set, we plot the following two equilibrium quantities: (a) service rates, and (b) mean steady state waiting times. Note that the graphs in Figure 2 only show data points corresponding to integer values of N ; the thin line through these points is only meant as a visual tool that helps bring out the pattern. Each of the four graphs shows data for three different effort cost functions: c(µ) = µ, c(µ) = µ2 , and c(µ) = µ3 , which are depicted in red, blue, and green respectively. The data points in Figure 2 marked × and ⋄ correspond to special staffing levels N ao,2 and N opt,2 respectively, which are introduced later, in Section 4. The first observation we make is that there are at most two equilibria. Further, for large enough values of the minimum service rate Nλ , there is no equilibrium. (In Figure 1(a) where N is fixed, this happens for large λ, and in Figure 2(a) where λ is fixed, this happens for small N .) On the other hand, when the minimum service rate Nλ is small enough, there is a unique equilibrium; for this range, even if the symmetric first order condition (9) has another solution greater than Nλ , it fails to satisfy (10). If an intermediate value of Nλ is small enough for (9) to have two feasible solutions, but not too small so that both solutions satisfy (10), then there are two equilibria. The second observation we make is that the two equilibria have very different behaviors. As illustrated in Figure 1(a), the larger equilibrium service rate first increases and then decreases while 3 In general, the symmetric first order condition (9) can be rewritten as µ2 c′ (µ) + λ λ (ρ − ErlC (N, ρ)) − = 0. N2 N Note that, when the term ρ − ErlC(N, ρ) is convex in µ, it follows that the left hand side of the above equation is also convex in µ, which implies that there are at most two symmetric equilibria. 11 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic µ⋆ ⋆ log10 (W ) c(µ) = µ3 3 0.4 c(µ) = µ3 2 ◦ c(µ) = µ 0.2 c(µ) = µ • 0 c(µ) = µ2 c(µ) = µ2 • 4 λ 6 -3 • • -6 0 2 4 6 λ (a) Service Rates (b) Mean Steady State Waiting Times Figure 1. Equilibrium behavior as a function of the arrival rate when the staffing level is fixed at N = 20, for three different effort cost functions: linear, quadratic, and cubic. The dotted line in (a) is µ = λ/N = λ/20. ⋆ log10 (W ) µ⋆ 0.4 c(µ) = µ3 • • • • × • 3 ⋄ × • • • c(µ) = µ2 • • • • • • • • c(µ) = µ3 • • • • • • • • ⋄ × • ⋄ 0 • • • -3 × • 6 c(µ) = µ2 • × × • • • • • × • 12 • • • • • • • • • • • 0 6 12 (a) Service Rates 18 N • • • -6 N 18 ⋄ • • ⋄ ⋄ c(µ) = µ × • ⋄ ⋄ c(µ) = µ 0.2 • • • (b) Mean Steady State Waiting Times Figure 2. Equilibrium behavior as a function of the staffing level when the arrival rate is fixed at λ = 2, for three different effort cost functions: linear, quadratic, and cubic. The dotted curve in (a) is µ = λ/N = 2/N . The data points marked × and ⋄ correspond to N ao,2 and N opt,2 respectively. the corresponding mean steady state waiting time in Figure 1(b) steadily increases. In contrast, as the smaller equilibrium service rate increases, the corresponding mean steady state waiting time decreases. The relationship between the equilibrium service rates and waiting times is similarly inconsistent in Figure 2. This behavior is not consistent with results from classical, nonstrategic models, and could serve as a starting point to explaining empiric observations that are also not consistent with classical, nonstrategic models. For example, the non-monotonicity of service rate in workload is consistent with behavior observed in a hospital setting in [32]. 4. Staffing strategic servers. One of the most studied questions for the design of service systems is staffing. Specifically, how many servers should be used for a given arrival rate. In the classical, nonstrategic setting, this question is well understood. In particular, as mentioned in the introduction, square-root staffing is known to be optimal when there are linear staffing and waiting costs [8]. In contrast, there is no previous work studying staffing in the context of strategic servers. The goal of this section is to initiate the study of the impact that strategic servers have on staffing. To get a feeling for the issues involved, consider a system with arrival rate λ and two possible staffing 12 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic policies: N1 = λ and N2 = 2λ, where Ni is the number of servers staffed under policy i given arrival rate λ. Under N1 , if the servers work at any rate slightly larger than 1, then they will have almost no idle time, and so they will have incentive to work harder. However, if servers are added, so that the provisioning is as in N2 , then servers will have plentiful idle time when working at rate 1, and thus not have incentive to work harder. Thus, the staffing level has a fundamental impact on the incentives of the servers. The above highlights that one should expect significant differences in staffing when strategic servers are considered. In particular, the key issue is that the staffing level itself creates incentives for the servers to speed up or slow down, because it influences the balance between effort and idle time. Thus, the policies that are optimal in the nonstrategic setting are likely suboptimal in the strategic setting, and vice versa. The goal of the analysis in this section is to find the staffing level that minimizes costs when the system manager incurs linear staffing and waiting costs, within the class of policies that admit a symmetric equilibrium service rate. However, the analysis in the previous section highlights that determining the exact optimal policy is difficult, since we only have an implicit characterization of a symmetric equilibrium service rate in (9). As a result, we focus our attention on the setting where λ is large, and look for an asymptotically optimal policy. As expected, the asymptotically optimal staffing policy we design for the case of strategic servers differs considerably from the optimal policies in the nonstrategic setting. In particular, in order for a symmetric equilibrium service rate to exist, the staffing level must be order λ larger than the optimal staffing in the classical, nonstrategic setting. Then, the system operates in a quality-driven (QD) regime instead of the quality-and-efficiency-driven (QED) regime that results from squareroot staffing. This is intuitive given that the servers value their idle time, and in the QD regime they have idle time but in the QED regime their idle time is negligible. The remainder of this section is organized as follows. We first introduce the cost structure and define asymptotic optimality in Section 4.1. Then, in Section 4.2, we provide a simple approximation of a symmetric equilibrium service rate and an asymptotically optimal staffing policy. Finally, in Section 4.3, we compare our asymptotically optimal staffing policy for strategic servers with the square-root staffing policy that is asymptotically optimal in the nonstrategic setting. 4.1. Preliminaries. Our focus in this section is on a M /M /N queue with strategic servers, as introduced in Section 3. We assume Random routing throughout this section. It follows that our results hold for any “idle-time-order-based” routing policy (as explained in the beginning of Section 3.2 and validated by Theorem 9). The cost structure we assume is consistent with the one in [8], under which square-root staffing is asymptotically optimal when servers are not strategic. In their cost structure, there are linear staffing and waiting costs. One difference in our setting is that there may be multiple equilibrium service rates. In light of Theorem 6, we focus on the largest ⋆ symmetric equilibrium service rate, and assume W denotes the mean steady state waiting time in a M /M /N queue with arrival rate λ, and strategic servers that serve at the largest symmetric equilibrium service rate (when there is more than one equilibrium).4 Then, the total system cost is ⋆ C ⋆ (N, λ) = cS N + wλW , where cS is the per-unit staffing cost and w is the per-unit waiting cost. The ⋆ superscript indicates that the mean steady state waiting time, and hence, the cost function, depends on the (largest) symmetric equilibrium service rate µ⋆ , which in turn depends on N and λ. 4 Note that the staffing policy we derive in this section (Theorem 8) will be asymptotically optimal regardless of which equilibrium service rate the servers choose. Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic 13 The function C ⋆ (N, λ) is well-defined only if a symmetric equilibrium service rate, under which the system is stable, exists. Furthermore, we would like to rule out having an unboundedly large symmetric equilibrium service rate because then the server utility (1) will be large and negative – and it is hard to imagine servers wanting to participate in such a game. Definition 1. A staffing policy N λ is admissible if the following two properties hold: (i) There exists a symmetric equilibrium µ⋆,λ under which the system is stable (λ < µ⋆,λ N λ ) for all large enough λ. (ii) There exists a sequence of symmetric equilibria {µ⋆,λ , λ > 0} for which lim supλ→∞ µ⋆,λ < ∞. If the requirement (ii) in the above definition is not satisfied, then the server utility will approach −∞ as the service rates become unboundedly large. The servers will not want to participate in such a game. As long as the requirement (ii) is satisfied, we can assume the server payment is sufficient to ensure that the servers have positive utility. We let Π denote the set of admissible staffing policies. We would like to solve for N opt,λ = arg min C ⋆ (N, λ). (11) N ∈Π However, given the difficulty of deriving N opt,λ directly, we instead characterize the first order growth term of N opt,λ in terms of λ. To do this, we consider a sequence of systems, indexed by the arrival rate λ, and let λ become large. Our convention when we wish to refer to any process or quantity associated with the system having arrival rate λ is to superscript the appropriate symbol by λ. In particular, N λ denotes the staffing level in the system having arrival rate λ, and µ⋆,λ denotes an equilibrium service rate ⋆,λ (assuming existence) in the system with arrival rate λ and staffing level N λ . We assume W equals the mean steady state waiting time in a M /M /N λ queue with arrival rate λ when the servers work at the largest equilibrium service rate. The associated cost is C ⋆,λ (N λ ) = cS N λ + wλW ⋆,λ . (12) Given this setup, we would like to find an admissible staffing policy N λ that has close to the minimum cost C ⋆,λ (N opt,λ ). Definition 2. A staffing policy N λ is asymptotically optimal if it is admissible (N λ ∈ Π) and C ⋆,λ (N λ ) = 1. λ→∞ C ⋆,λ (N opt,λ ) lim In what follows, we use the o and ω notations to denote the limiting behavior of functions. Formally, for any two real-valued functions f (x), g(x) that take nonzero values for sufficiently large (x) x, we say that f (x) = o(g(x)) (equivalently, g(x) = ω(f (x))) if limx→∞ fg(x) = 0. In other words, f is dominated by g asymptotically (equivalently, g dominates f asymptotically). 4.2. An asymptotically optimal staffing policy. The class of policies we study are those that staff independently of the equilibrium service rates, which are endogenously determined according to the analysis in Section 3.2. More specifically, these are policies that choose N λ purely as a function of λ. Initially, it is unclear what functional form an asymptotically optimal staffing policy can take in the strategic server setting. Thus, to begin, it is important to rule out policies that cannot be asymptotically optimal. The following proposition does this, and highlights that asymptotically optimal policies must be asymptotically linear in λ. Proposition 1. Suppose N λ = f (λ) + o(f (λ)) for some function f. If either f (λ) = o(λ) or f (λ) = ω(λ), then the staffing policy N λ cannot be asymptotically optimal. 14 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic Intuitively, if f (λ) = o(λ), understaffing forces the servers to work too hard, their service rates growing unboundedly (and hence their utilities approaching −∞) as λ becomes large. On the other hand, the servers may prefer to have f (λ) = ω(λ) because the overstaffing allows them to be lazier; however, the overstaffing is too expensive for the system manager. Proposition 1 implies that to find a staffing policy that is asymptotically optimal, we need only search within the class of policies that have the following form: 1 N λ = λ + o(λ), for a ∈ (0, ∞). a (13) However, before we can search for the cost-minimizing a, we must ensure that the staffing (13) guarantees the existence of a symmetric equilibria µ⋆,λ for all large enough λ. It turns out that this is only true when a satisfies certain conditions. After providing these conditions (see Theorem 7 in the following), we evaluate the cost function as λ becomes large to find the a⋆ (defined in (17)) under which (13) is an asymptotically optimal staffing policy (see Theorem 8). Equilibrium characterization. The challenge in characterizing equilibria comes from the complexity of the first order condition derived in Section 3. This complexity drives our focus on the large λ regime. The first order condition for a symmetric equilibrium (9) is equivalently written as     ErlC(N λ , λ/µ) λ λ µ 1+ − λ = µ3 c′ (µ). (14) Nλ Nλ N Under the staffing policy (13), when the limit λ → ∞ is taken, this becomes a(µ − a) = µ3 c′ (µ). (15) Since µ3 c′ (µ) > 0, it follows that any solution µ has µ > a. Therefore, under the optimistic assumption that a symmetric equilibrium solution µ⋆,λ converging to the aforementioned solution µ exists, it follows that λ/µ⋆,λ < λ/a for all large enough λ. In words, the presence of strategic servers that value their idle time forces the system manager to staff order λ more servers than the offered load λ/µ⋆,λ . In particular, since the growth rate of N λ is λ/a, the system will operate in the quality-driven regime. The properties of the equation (15) are easier to see when it is rewritten as 1 µ2 ′ 1 = 2 c (µ) + . a a µ (16) Note that the left-hand side of (16) is a constant function and the right-hand side is a convex function. These functions either cross at exactly two points, at exactly one point, or never intersect, depending on a. That information then can be used to show whether or not there exists a solution to the first order condition (14), depending on the value of a in the staffing policy (13). Theorem 7. The following holds for all large enough λ. (i) Suppose a > 0 is such that there exists µ2 > µ1 > 0 that solve (16). Then, there exist two solutions that solve (14). (ii) Suppose a > 0 is such that there exists exactly one µ1 > 0 that solves (16). (a) Suppose N λ − λa ≥ 0. Then, there exist two solutions that solve (14). (b) Otherwise, if N λ − λa < −3, then there does not exist a solution µλ to (14). Furthermore, for any ǫ > 0, if µλ solves (14), then |µλ − µ| < ǫ for some µ that solves (16). Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic 15 We are not sure if there exists a solution in the case of N λ − a1 λ ∈ [−3, 0); however, given that we are focusing on a large λ asymptotic regime, the range [-3,0) is vanishingly small. Moving forward, once the existence of a solution to the first order condition (14) is established, to conclude that solution is a symmetric equilibrium service rate also requires verifying the condition (10) in Theorem 4. This can be done for any staffing policy (13) under which the system operates in the quality driven-regime. Lemma 1. tion (14), if For any staffing policy N λ and associated µλ that satisfies the first order condilim inf λ→∞ N λ µλ = d > 1 and lim sup µλ < ∞, λ λ→∞ then µ⋆,λ = µλ is a symmetric equilibrium for all large enough λ. Under the conditions for the existence of a solution to the first order condition (14) in Theorem 7, it is also true that the conditions of Lemma 1 are satisfied. In particular, there exists a bounded sequence {µλ } having N λ µλ µλ o(λ) lim inf = lim inf + µλ > 1. λ→∞ λ→∞ λ a λ This then guarantees that, for all large enough λ, there exists a solution µ⋆,λ to (14) that is a symmetric equilibrium, under the conditions of Theorem 7. There are either multiple symmetric equilibria for each λ or 0, because from Theorem 7 there are either multiple or zero solutions to the first order condition (14). These symmetric equilibria will be close when there exists exactly one µ that solves (16); however, they may not be close when there exist two µ that solve (16). We show in the following that this does not affect what staffing policy should be asymptotically optimal. Optimal staffing. Given the characterization of symmetric equilibria under a staffing policy (13), we can now move to the task of optimizing the staffing level, i.e., optimizing a. The first step is to characterize the associated cost, which is done in the following proposition. Proposition 2. Suppose a > 0 is such that there exists µ > 0 that solves (16). Then, under the staffing policy (13), C ⋆,λ (N λ ) 1 → cS , as λ → ∞. λ a Proposition 2 implies that to minimize costs within the class of staffing policies that satisfy (13), the maximum a under which there exists at least one solution to (16) should be used. That is, we should choose a to be a⋆ := sup A, where A := {a > 0 : there exists at least one solution µ > 0 to (16)} . Lemma 2. (17) a⋆ ∈ A is finite. Importantly, this a⋆ is not only optimal among the class of staffing policies that satisfy (13), it is asymptotically optimal among all admissible staffing policies. In particular, the following theorem shows that as λ becomes unboundedly large, no other admissible staffing policy can asymptotically achieve strictly lower cost than the one in (13) with a = a⋆ . Theorem 8. If N ao,λ satisfies (13) with a = a⋆ , then N ao,λ is admissible and asymptotically optimal. Furthermore, 1 C ⋆,λ (N opt,λ ) C ⋆,λ (N ao,λ ) = lim = cS ⋆ . λ→∞ λ→∞ λ λ a lim 16 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic Note that an inspection of the proof of Theorem 8 shows that it holds regardless of which ⋆,λ ⋆,λ equilibrium service rate is used to define W . Hence, even though we have defined W to be the mean steady state waiting time when the servers serve at the largest equilibrium service rate, this is not necessary. The staffing policy N ao,λ in Theorem 8 will be asymptotically optimal regardless of which equilibrium service rate the servers choose. Though the above theorem characterizes an asymptotically optimal staffing level, because the definition of a⋆ is implicit, it is difficult to develop intuition. To highlight the structure more clearly, the following lemma characterizes a⋆ for a specific class of effort cost functions. Lemma 3. Suppose c(µ) = cE µp for some cE ∈ [1, ∞) and p ≥ 1. Then, " 1 #(p+1)/p   p1   p+1 p+1 (p + 1) 1 ⋆ ⋆ < 1, <µ = a = (p + 2) cE p(p + 2) cE p(p + 2)2 and a⋆ and µ⋆ are both increasing in p. Furthermore,    <  there are 2 non-negative solutions to (16) there is no non-negative solution to (16) . if a > a⋆ , then    = there is exactly one solution to (16) There are several interesting relationships between the effort cost function and the staffing level that follow from Lemma 3. First, for fixed p, a⋆ (p) ↓ 0 as cE → ∞. In words, the system manager must staff more and more servers as effort becomes more costly. Second, for fixed cE , since a⋆ (p) is increasing in p, the system manager can staff less servers when the cost function becomes “more convex”. The lower staffing level forces the servers to work at a higher service rate since since µ⋆ (p) is also increasing in p. We will revisit this idea that convexity is helpful to the system manager in the next section. 4.3. Contrasting staffing policies for strategic and nonstrategic servers. One of the most crucial observations that the previous section makes about the impact of strategic servers on staffing is that the strategic behavior leads the system to a quality-driven regime. In this section, we explore this issue in more detail, by comparing to the optimal staffing rule that arises when servers are not strategic, and then attempting to implement that staffing rule. Nonstrategic servers. Recall that, for the conventional M /M /N queue (without strategic servers), square-root staffing minimizes costs as λ becomes large (see equation (1), Proposition 6.2, and Example 6.3 in [8]). So, we can define λ Cµλ (N ) = cS N + wλW µ to be the cost associated with staffing N nonstrategic servers that work at the fixed service rate µ. Further, Nµopt,λ = arg min Cµλ (N ) N> λ µ is the staffing level that minimizes expected cost when the system arrival rate is λ and the service rate is fixed to be µ. So, the staffing rule s λ λ (18) NµBMR,λ = + y ⋆ µ µ 17 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic is asymptotically optimal in the sense that lim λ→∞ Cµλ (NµBMR,λ ) Cµλ (Nµopt,λ ) = 1. o  −1 n y , where α(y) = 1 + with h(·) being the hazard rate Here, y ⋆ := arg miny>0 cS y + wα(y) y h(−y) φ(x) function of the standard normal distribution, namely, h(x) := 1−Φ(x) with φ(x) = √12π e− Rx Φ(x) = −∞ φ(t)dt. The staffing rule (18) is the famous square-root safety staffing rule. x2 2 and Contrasting strategic and nonstrategic servers. In order to compare the case of strategic servers to the case of nonstrategic servers, it is natural to fix µ in (18) to the limiting service rate that results from using the optimal staffing rule N ao,λ defined in Theorem 8. We see that N ao,λ , where µ⋆ solves (16) for a = a⋆ , because any solution staffs order λ more servers than NµBMR,λ ⋆ to (16) has a > µ. When the effort cost function is c(µ) = cE µp for p ≥ 1, we know from Lemma 3 and Theorem 7 (since the a⋆ is unique) that µ⋆,λ → µ⋆ as λ → ∞, where µ⋆ is as given in Lemma 3. Then, the difference in the staffing levels is     1 1 1 1 − λ + o(λ) = λ + o(λ). N ao,λ − NµBMR,λ = ⋆ a⋆ µ⋆ a⋆ p + 2 Since a⋆ = a⋆ (p) is increasing in p from Lemma 3, we see that the difference N ao,λ − NµBMR,λ ⋆ decreases to 0 as the cost function becomes “more convex”. This is consistent with our observation at the end of the previous subsection that convexity is helpful to the system manager. It is natural to wonder if a system manager can force the servers to work harder by adopting the staffing policy suggested by the analysis of nonstrategic servers, i.e., s λ λ ⋆,BMR,λ ⋆ N = ⋆,λ + y . (19) µ µ⋆,λ The interpretation of this staffing rule requires care, because the offered load λ/µ⋆,λ is itself a function of the staffing level (and the arrival rate) through an equilibrium service rate µ⋆,λ . The superscript ⋆ emphasizes this dependence. The first question concerns whether or not the staffing policy (19) is even possible in practice, because the staffing level depends on an equilibrium service rate and vice versa. More specifically, for a given staffing level, the servers relatively quickly arrive at an equilibrium service rate. Then, when system demand grows, the system manager increases the staffing, and the servers again arrive at an equilibrium service rate. In other words, there are two games, one played on a faster time scale (that is the servers settling to an equilibrium service rate), and one played on a slower time scale (that is the servers responding to added capacity). To analyze the staffing policy (19), note that the first order condition for a symmetric equilibrium (9) is equivalently written as    λ ⋆,BMR,λ λ ⋆,BMR,λ + ErlC N , = µc′ (µ). N − 2 ⋆,BMR,λ µ µ (N ) λ/µ 18 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic Then, if µλ is a solution to the first order condition under the staffing N ⋆,BMR,λ from (19), from substituting N ⋆,BMR,λ into the above expression, µλ must satisfy    p p λ/µλ λ ⋆ λ ⋆ λ λ = µλ c′ (µλ ).  2 y λ/µ + ErlC λ/µ + y λ/µ , λ p µ λ/µλ + y ⋆ λ/µλ   p As λ becomes large, since ErlC λ/µ + y λ/µ, µλ is bounded above by 1, the left-hand side of the above expression has limit 0. Furthermore, the right-hand side of the above equation is nonnegative and increasing as a function of µ. Hence any sequence of solutions µλ to the first order condition has the limiting behavior µλ → 0, as λ → ∞, which cannot be a symmetric equilibrium service rate because we require the servers to work fast enough to stabilize the system. One possibility is to expand the definition of an equilibrium service rate in (1) to allow the servers to work exactly at the lower bound λ/N . In fact, the system manager may now be tempted to push the servers to work even faster. However, faster service cannot be mandated for free – there must be a trade-off; for example, the service quality may suffer or the salaries should be higher. 4.4. Numerical examples. In order to understand how well our asymptotically optimal staffing policy N ao,λ performs in comparison with the optimal policy N opt,λ for finite λ, and how fast the corresponding system cost converges to the optimal cost, we present some results from numerical analysis in this section. We consider two staffing policies: (i) N opt,λ (defined in (11)), and (ii) N ao,λ (defined in Theorem 8 and (17)) where we ignore the o(λ) term of (13). For each, we first round up the staffing level if necessary, and then plot the following two equilibrium quantities as a function of the arrival rate λ: (a) service rates µ⋆,λ (if there is more than one, we pick the largest), and (b) normalized costs C ⋆,λ /λ. We calculate N opt,λ numerically, by iterating over the staffing levels that admit equilibria (and we choose the lowest cost when there are multiple equilibria). These plots are shown in Figure 3 for three different effort cost functions: c(µ) = µ, c(µ) = µ2 , and c(µ) = µ3 , which are depicted in red, blue, and green respectively. For each color, the curve with the darker shade corresponds to N opt,λ and the curve with the lighter shade corresponds to N ao,λ . The horizontal dashed lines correspond to the limiting values as λ → ∞. An immediate first observation is the jaggedness of the curves, which is a direct result of the discreteness of the staffing levels N opt,λ and N ao,λ . In particular, as the arrival rate λ increases, the equilibrium service rate µ⋆,λ decreases (respectively, the equilibrium normalized cost C ⋆,λ /λ increases) smoothly until the staffing policy adds an additional server, which causes a sharp increase (respectively, decrease). The jaggedness is especially pronounced for smaller λ, resulting in a complex pre-limit behavior that necessitates asymptotic analysis in order to obtain analytic results. However, despite the jaggedness, the plots illustrate clearly that both the equilibrium service rates and normalized costs of the optimal policy N ao,λ converge quickly to those of the optimal policy N opt,λ , highlighting that our asymptotic results are predictive at realistically sized systems. 5. Routing to strategic servers. Thus far we have focused our discussion on staffing, assuming that jobs are routed randomly to servers when there is a choice. Of course, the decision of how to route jobs to servers is another crucial aspect of the design of service systems. As such, the analysis of routing policies has received considerable attention in the queueing literature, when 19 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic C ⋆,λ /λ µ⋆,λ 0.4 c(µ) = µ3 8 0.2 c(µ) = µ2 c(µ) = µ c(µ) = µ c(µ) = µ2 4 c(µ) = µ3 0 9 18 (a) Service Rates 27 λ 0 9 18 27 λ (b) Normalized Costs Figure 3. Equilibrium behavior as a function of the arrival rate for the optimal and asymptotically optimal staffing policies, for three different effort cost functions: linear, quadratic, and cubic. servers are not strategic. In this section, we begin to investigate the impact of strategic servers on the design of routing policies. In the classical literature studying routing when servers are nonstrategic, a wide variety of policies have been considered. These include “rate-based policies” such as Fastest Server First (FSF) and Slowest Server First (SSF); as well as “idle-time-order-based policies” such as Longest Idle Server First (LISF) and Shortest Idle Server First (SISF). Among these routing policies, FSF is a natural choice to minimize the mean response time (although, as noted in the Introduction, it is not optimal in general). This leads to the question: how does FSF perform when servers are strategic? In particular, does it perform better than the Random routing that we have so far studied? Before studying optimal routing to improve performance, we must first answer the following even more fundamental question: what routing policies admit symmetric equilibria? This is a very challenging goal, as can be seen by the complexity of the analysis for the M /M /N under Random routing. This section provides a first step towards that goal. The results in this section focus on two broad classes of routing policies idle-time-order-based policies and rate-based policies, which are introduced in turn in the following. 5.1. Idle-time-order-based policies. Informally, idle-time-order-based policies are those routing policies that use only the rank ordering of when servers last became idle in order to determine how to route incoming jobs. To describe the class of idle-time-order-based policies precisely, let I(t) be the set of servers idle at time t > 0, and, when I(t) 6= ∅, let s(t) = (s1 , . . . , s|I(t)| ) denote the ordered vector of idle servers at time t, where server sj became idle before server sk whenever j < k. For n ≥ 1, let Pn = ∆({1, . . . , n}) denote the set of all probability distributions over the set {1, . . . , n}. An idle-time-order-based routing policy is defined by a collection of probability distributions p = {pS }S∈2{1,2,...,N} \∅ , such that pS ∈ P|S| , for all S ∈ 2{1,2,...,N } \∅. Under this policy, at time t, the next job in queue is assigned to idle server sj with probability pI(t) (j). Examples of idle-time-order-based routing policies are as follows. 1. Random. An arriving customer that finds more than one server idle is equally likely to be routed to any of those servers. Then, pS = (1/|S|, . . . , 1/|S|) for all S ∈ 2{1,2,...,N } \∅. 2. Weighted Random. Each such arriving customer is routed to one of the idle servers with probabilities that may depend on the order in which the servers became idle. For example, if |S| + 1 − j pS (j) = P|S| , j ∈ S, for sj ∈ S, for all S ∈ 2{1,2,...,N } \∅, n=1 n 20 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic then the probabilities are decreasing according to the order in which the servers became idle. 1 |S|(|S|+1) P |S|(|S|+1)− 2 Note that j pS (j) = = 1. 1 |S|(|S|+1) 2 3. Longest Idle Server First (Shortest Idle Server First). Each such arriving customer is routed to the server that has idled the longest (idled the shortest). Then, pS = (1, 0, . . . , 0) (pS = (0, . . . , 0, 1)) for all S ⊆ {1, 2, . . . , N }. 5.1.1. Policy-space collapse. Surprisingly, it turns out that all idle-time-order-based policies are “equivalent” in a very strong sense — they all lead to the same steady state probabilities, resulting in a remarkable policy-space collapse result, which we discuss in the following. Fix R to be some idle-time-order-based routing policy, defined through the collection of probability distributions p = {pS }∅6=S⊆{1,2,...,N } . The states of the associated continuous time Markov chain are defined as follows: • State B is the state where all servers are busy, but there are no jobs waiting in the queue. • State s = (s1 , s2 , . . . , s|I| ) is the ordered vector of idle servers I. When I = ∅, we identify the empty vector s with state B. • State m (m ≥ 0) is the state where all servers are busy and there are m jobs waiting in the queue (i.e., there are N + m jobs in the system). We identify state 0 with state B. When all servers are busy, there is no routing, and so the system behaves exactly as a M /M /1 queue with arrival rate λ and service rate µ1 + · · · + µN . Then, from the local balance equations, the associated steady state probabilities πB and πm for m = 0, 1, 2, . . ., must satisfy m πm = (λ/µ) πB where µ = N X µj . (20) j=1 One can anticipate that the remaining steady state probabilities satisfy Y µs for all s = (s1 , s2 , . . . , s|I| ) with |I| > 0, πs = πB λ s∈I (21) and the following theorem verifies this by establishing that the detailed balance equations are satisfied. Theorem 9. All idle-time-order-based policies have the steady state probabilities that are uniquely determined by (20)-(21), together with the normalization constraint that their sum is one. Theorem 9 is remarkable because there is no dependence on the collection of probability distributions p that define R. Therefore, it follows that all idle-time-order-based routing policies result in the same steady state probabilities. Note that, concurrently, a similar result has been discovered independently in the context of loss systems [24]. In relation to our server game, it follows from Theorem 9 that all idle-time-order-based policies have the same equilibrium behavior as Random. This is because an equilibrium service rate depends on the routing policy through the server idle time vector (I1 (µ; R), . . . , IN (µ; R)), which can be found from the steady state probabilities in (20)-(21). As a consequence, if there exists (does not exist) an equilibrium service rate under Random, then there exists (does not exist) an equilibrium service rate under any idle-time-order-based policy. In summary, it is not possible to achieve better performance than under Random by employing any idle-time-order-based policy. 5.2. Rate-based policies. Informally, a rate-based policy is one that makes routing decisions using only information about the rates of the servers. As before, let I(t) denote the set of idle servers at time t. In a rate-based routing policy, jobs are assigned to idle servers only based on their service rates. We consider a parameterized class of rate-based routing policies that we term Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic 21 r-routing policies (r ∈ R). Under an r-routing policy, at time t, the next job in queue is assigned to idle server i ∈ I(t) with probability µr pi (µ, t; r) = Xi µrj j∈I(t) Notice that for special values of the parameter r, we recover well-known policies. For example, setting r = 0 results in Random; as r → ∞, it approaches FSF; and as r → −∞, it approaches SSF. In order to understand the performance of rate-based policies, the first step is to perform an equilibrium analysis, i.e., we need to understand what the steady state idle times look like under any r-routing policy. The following proposition provides us with the required expressions. Proposition 3. Consider a heterogeneous M /M /2 system under an r-routing policy, with arrival rate λ > 0 and servers 1 and 2 operating at rates µ1 and µ2 respectively. The steady state probability that server 1 is idle is given by: h i r 2 µ1 (µ1 + µ2 − λ) (λ + µ2 )2 + µ1 µ2 + µrµ+µ r (λµ1 + λµ2 ) 1 2 h i I1r (µ1 , µ2 ) = , r µ 2 2 1 µ1 µ2 (µ1 + µ2 )2 + (λµ1 + λµ2 ) µ1 + 2µ1 µ2 − µr +µr (µ1 − µ22 ) + (λµ1 )2 + (λµ2 )2 1 2 and the steady state probability that server 2 is idle is given by I2r (µ1 , µ2 ) = I1r (µ2 , µ1 ). Note that we restrict ourselves to a 2-server system for this analysis. This is due to the fact that there are no closed form expressions known for the resulting Markov chains for systems with more than 3 servers. It may be possible to extend these results to 3 servers using results from [37]; but, the expressions are intimidating, to say the least. However, the analysis for two servers is already enough to highlight important structure about the impact of strategic servers on policy design. In particular, our first result concerns the FSF and SSF routing policies, which can be obtained in the limit when r → ∞ and r → −∞ respectively. Recall that FSF is asymptotically optimal in the nonstrategic setting. Intuitively, however, it penalizes the servers that work the fastest by sending them more and more jobs. In a strategic setting, this might incentivize servers to decrease their service rate, which is not good for the performance of the system. One may wonder if by doing the opposite, that is, using the SSF policy, servers can be incentivized to increase their service rate. However, our next theorem (Theorem 10) shows that neither of these policies is useful if we are interested in symmetric equilibria. Recall that our model for strategic servers already assumes an increasing, convex effort cost function with c′′′ (µ) ≥ 0. For the rest of this section, in addition, we assume that c′ ( λ2 ) < λ1 . (Recall that this is identical to the sufficient condition c′ ( Nλ ) < λ1 which we introduced in Section 3.2, on substituting N = 2.)5 Theorem 10. Consider a M /M /2 queue with strategic servers. Then, FSF and SSF do not admit a symmetric equilibrium. Moving beyond FSF and SSF, we continue our equilibrium analysis (for a finite r) by using the first order conditions to show that whenever an r-routing policy admits a symmetric equilibrium, it is unique. Furthermore, we provide an expression for the corresponding symmetric equilibrium service rate in terms of r, which brings out a useful monotonicity property. 5 The sufficient condition c′ ( λ2 ) < λ1 might seem rather strong, but it can be shown that it is necessary for the symmetric first order condition to have a unique solution. This is because, if c′ ( λ2 ) > λ1 , then the function ϕ(µ), defined in (22), ceases to be monotonic, and as a result, for any given r, the first order condition ϕ(µ) = r could have more than one solution. 22 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic Theorem 11. Consider a M /M /2 queue with strategic servers. Then, any r-routing policy that admits a symmetric equilibrium, admits a unique symmetric equilibrium, given by µ⋆ = ϕ−1 (r), where ϕ : ( λ2 , ∞) → R is the function defined by ϕ(µ) = 4(λ + µ) (µ(λ + 2µ)c′ (µ) − λ) . λ(λ − 2µ) (22) Furthermore, among all such policies, µ⋆ is decreasing in r, and therefore, E[T ], the mean response time (a.k.a. sojourn time) at symmetric equilibrium is increasing in r. In light of the inverse relationship between r and µ⋆ that is established by this theorem, the system manager would ideally choose the smallest r such that the corresponding r-routing policy admits a symmetric equilibrium, which is in line with the intuition that a bias towards SSF (the limiting r-routing policy as r → −∞) incentivizes servers to work harder. However, there is a hard limit on how small an r can be chosen (concurrently, how large an equilibrium service rate µ⋆ can be achieved) so that there exists a symmetric equilibrium, as evidenced by our next theorem. Theorem 12. Consider a M /M /2 queue with strategic servers. Then, there exists µ, r ∈ R, with r = ϕ(µ), such that no service rate µ > µ can be a symmetric equilibrium under any r-routing policy, and no r-routing policy with r < r admits a symmetric equilibrium. The proof of this theorem is constructive and we do exhibit an r, however, it is not clear whether this is tight, that is, whether there exists a symmetric equilibrium for all r-routing policies with r ≥ r. We provide a partial answer to this question of what r-routing policies do admit symmetric equilibria in the following theorem. Theorem 13. Consider a M /M /2 queue with strategic servers. Then, there exists a unique symmetric equilibrium under any r-routing policy with r ∈ {−2, −1, 0, 1}. Notice that we show equilibrium existence for four integral values of r. It is challenging to show that all r-routing policies in the interval [−2, 1] admit a symmetric equilibrium. This theorem provides an upper bound on the r of the previous theorem, that is, r ≤ −2. Therefore, if the specific cost function c is unknown, then the system manager can guarantee better performance than Random (r = 0), by setting r = −2. If the specific cost function is known, the system manager may be able to employ a lower r to obtain even better performance. For example, consider a 2server system with λ = 1/4 and one of three different effort cost functions: c(µ) = µ, c(µ) = µ2 , and c(µ) = µ3 . Figure 4 shows the corresponding equilibrium mean response times (in red, blue, and green, respectively). It is worth noting that the more convex the effort cost function, larger the range of r (and smaller the minimum value of r) for which a symmetric equilibrium exists. 6. Concluding remarks. The rate at which each server works in a service system has important consequences for service system design. However, traditional models of large service systems do not capture the fact that human servers respond to incentives created by scheduling and staffing policies, because traditional models assume each server works at a given fixed service rate. In this paper, we initiate the study of a class of strategic servers that seek to optimize a utility function which values idle time and includes an effort cost. Our focus is on the analysis of staffing and routing policies for a M /M /N queue with strategic servers, and our results highlight that strategic servers have a dramatic impact on the optimal policies in both cases. In particular, policies that are optimal in the classical, nonstrategic setting can perform quite poorly when servers act strategically. For example, a consequence of the strategic server behavior is that the cost-minimizing staffing level is order λ larger than square-root staffing, the cost minimizing staffing level for systems with fixed service rate. In particular, any system with strategic servers operates in the quality-driven 23 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic log10 (E[T ]) 1 • c(µ) = µ • c(µ) = µ2 2 3 • • c(µ) = µ3 1 3 • • -18 -12 -6 0 6 12 18 r Figure 4. Equilibrium mean response time (a.k.a. sojourn time) as a function of the policy parameter, r, when the arrival rate is λ = 14 , for three different effort cost functions: linear, quadratic, and cubic. regime at equilibrium (as opposed to the quality-and-efficiency-driven regime that arises under square-root staffing), in which the servers all enjoy non-negligible idle time. The intuitive reason square-root staffing is not feasible in the context of strategic servers is that the servers do not value their idleness enough in comparison to their effort cost. This causes the servers to work too slowly, making idle time scarce. In the economics literature [9, 36], it is common to assume that scarce goods are more highly valued. If we assume that the servers valued their idle time more heavily as the idle time becomes scarcer, then the servers would work faster in order to make sure they achieved some. This suggests the following interesting direction for future research: what is the relationship between the assumed value of idle time in (2) and the resulting cost minimizing staffing policy? Another situation in which servers may not care about idle time becoming scarce is when their compensation depends on their service volume (which is increasing in their service rate). Then, it is reasonable to expect the servers prefer to have negligible idle time. It would be interesting to be able to identify a class of compensation schemes under which that is the case. The aforementioned two future research directions become even more interesting when the class of routing policies is expanded to include rate-based policies. This paper solves the joint routing and staffing problem within the class of idle-time-order-based policies. Section 5 suggests that by expanding the class of routing policies to also include rate-based policies we should be able to achieve better system performance (although it is clear that the analysis becomes much more difficult). The richer question also aspires to understand the relationship between the server idle time value, the compensation scheme, the (potentially) rate-based routing policy, and the number of strategic servers to staff. Finally, it is important to note that we have focused on symmetric equilibrium service rates. We have not proven that asymmetric equilibria do not exist. Thus, it is natural to wonder if there are routing and staffing policies that result in an asymmetric equilibrium. Potentially, there could be one group of servers that have low effort costs but negligible idle time and another group of servers that enjoy plentiful idle time but have high effort costs. The question of asymmetric equilibria becomes even more interesting when the servers have different utility functions. For example, more experienced servers likely have lower effort costs than new hires. Also, different servers can value their idle time differently. How do we design routing and staffing policies that are respectful of such considerations? 24 Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic Acknowledgments. This work was supported by NSF grant #CCF-1101470, AFOSR grant #FA9550-12-1-0359, and ONR grant #N00014-09-1-0751. An important source of inspiration for this work came from discussions with Mor Armony regarding how to fairly assign workload to employees. We would also like to thank Bert Zwart for an illuminating discussion regarding asymptotically optimal staffing policies. References [1] Aksin, Z., M. Armony, V. Mehrotra. 2007. The modern call-center: A multi-disciplinary perspective on operations management research. Prod. Oper. Manag. 16(6) 665–688. [2] Allon, G., I. Gurvich. 2010. Pricing and dimensioning competing large-scale service providers. M&SOM 12(3) 449–469. [3] Anton, J. 2005. One-minute survey report #488: Agent compensation & advancement. Document Tracking Number SRV488-080305. [4] Armony, M. 2005. Dynamic routing in large-scale service systems with heterogeneous servers. Queueing Syst. Theory Appl. 51(3-4) 287–329. [5] Armony, M., A. R. Ward. 2010. Fair dynamic routing in large-scale heterogeneous-server systems. Oper. Res. 58 624–637. [6] Atar, R. 2005. Scheduling control for queueing systems with many servers: Asymptotic optimality in heavy traffic. Ann. Appl. Probab. 15(4) 2606–2650. [7] Atar, R., Y. Y. Shaki, A. Shwartz. 2011. A blind policy for equalizing cumulative idleness. Queueing Syst. 67(4) 275–293. [8] Borst, S., A. Mandelbaum, M. I. Reiman. 2004. Dimensioning large call centers. Oper. Res. 52(1) 17–34. [9] Brock, T. C. 1968. Implications of commodity theory for value change. Psychological Foundations of Attitudes 243–275. [10] Cachon, G. P., P. T. Harker. 2002. Competition and outsourcing with scale economies. Manage. Sci. 48(10) 1314–1333. [11] Cachon, G. P., F. Zhang. 2007. Obtaining fast service in a queueing system via performance-based allocation of demand. Manage. Sci. 53(3) 408–420. [12] Cahuc, P., A. Zylberberg. 2004. Labor Economics. MIT Press. [13] Cheng, S.-F., D. M. Reeves, Y. Vorobeychik, M. P. Wellman. 2004. Notes on equilibria in symmetric games. International Workshop On Game Theoretic And Decision Theoretic Agents (GTDT). 71–78. [14] Cohen-Charash, Y., P. E. Spector. 2001. The role of justice in organizations: A meta-analysis. Organ. Behav. and Hum. Dec. 86(2) 278–321. [15] Colquitt, J. A., D. E. Conlon, M. J. Wesson, C. O. L. H. Porter, K. Y. Ng. 2001. Justice at the millennium: A meta-analytic review of 25 years of organizational justice research. J. Appl. Psychol. 86(3) 425–445. [16] de Véricourt, F., Y.-P. Zhou. 2005. Managing response time in a call-routing problem with service failure. Oper. Res. 53(6) 968–981. [17] Erlang, A. K. 1948. On the rational determination of the number of circuits. E. Brockmeyer, H. L. Halstrom, A. Jensen, eds., The Life and Works of A. K. Erlang. The Copenhagen Telephone Company, 216–221. [18] Gans, N., G. Koole, A. Mandelbaum. 2003. Telephone call centers: Tutorial, review, and research prospects. M&SOM 5(2) 79–141. [19] Garnett, O., A. Mandelbaum, M. Reiman. 2002. Designing a call center with impatient customers. M&SOM 4(3) 208–227. [20] Geng, X., W. T. Huh, M. Nagarajan. 2013. Strategic and fair routing policies in a decentralized service system. Working paper. [21] Gilbert, S. M., Z. K. Weng. 1998. Incentive effects favor nonconsolidating queues in a service system: The principal-agent perspective. Manage. Sci. 44(12) 1662–1669. Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic 25 [22] Gumbel, H. 1960. Waiting lines with heterogeneous servers. Oper. Res. 8(4) 504–511. [23] Gurvich, I., W. Whitt. 2007. Scheduling flexible servers with convex delay costs in many-server service systems. M&SOM 11(2) 237–253. [24] Haji, B., S. M. Ross. 2013. A queueing loss model with heterogenous skill based servers under idle time ordering policies. Working paper. [25] Halfin, S., W. Whitt. 1981. Heavy-traffic limits for queues with many exponential servers. Oper. Res. 29(3) 567–588. [26] Harchol-Balter, M. 2013. Performance Modeling and Design of Computer Systems: Queueing Theory in Action. Cambridge University Press. [27] Harel, A. 1988. Sharp bounds and simple approximations for the Erlang delay and loss formulas. Manage. Sci. 34(8) 959–972. [28] Hassin, R., M. Haviv. 2003. To Queue or Not to Queue: Equilibrium Behavior in Queueing Systems. Kluwer. [29] Hopp, W., W. Lovejoy. 2013. Hospital Operations: Principles of High Efficiency Health Care. Financial Times Press. [30] Janssen, A. J. E. M., J. S.H. van Leeuwaarden, B. Zwart. 2011. Refining square-root safety staffing by expanding Erlang C. Oper. Res. 59(6) 1512–1522. [31] Kalai, E., M. I. Kamien, M. Rubinovitch. 1992. Optimal service speeds in a competitive environment. Manage. Sci. 38(8) 1154–1163. [32] Kc, D. S., C. Terwiesch. 2009. Impact of workload on service time and patient safety: An econometric analysis of hospital operations. Manage. Sci. 55(9) 1486–1498. [33] Kocaga, L., M. Armony, A. R. Ward. 2013. Staffing call centers with uncertain arrival rates and cosourcing. Working paper. [34] Krishnamoorthi, B. 1963. On Poisson queue with two heterogeneous servers. Oper. Res. 11(3) 321–330. [35] Lin, W., P. Kumar. 1984. Optimal control of a queueing system with two heterogeneous servers. IEEE Trans. Autom. Contr. 29(8) 696–703. [36] Lynn, M. 1991. Scarcity effects on value: A quantitative review of the commodity theory literature. Psychol. Market. 8(1) 43–57. [37] Mokaddis, G. S., C. H. Matta, M. M. El Genaidy. 1998. On Poisson queue with three heterogeneous servers. International Journal of Information and Management Sciences 9 53–60. [38] Reed, J., Y. Shaki. 2013. A fair policy for the G/GI/N queue with multiple server pools. Preprint. [39] Saaty, T. L. 1960. Time-dependent solution of the many-merver Poisson queue. Oper. Res. 8(6) 755–772. [40] Tezcan, T. 2008. Optimal control of distributed parallel server systems under the Halfin and Whitt regime. Math. Oper. Res. 33 51–90. [41] Tezcan, T., J. Dai. 2010. Dynamic control of N-systems with many servers: Asymptotic optimality of a static priority policy in heavy traffic. Oper. Res. 58(1) 94–110. [42] Ward, A. R., M. Armony. 2013. Blind fair routing in large-scale service systems with heterogeneous customers and servers. Oper. Res. 61 228–243. [43] Whitt, W. 2002. IEOR 6707: Advanced Topics in Queueing Theory: Focus on Customer Contact Centers. Homework 1e Solutions, see http://www.columbia.edu/~ww2040/ErlangBandCFormulas.pdf. e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic ec1 Routing and Staffing when Servers are Strategic: Technical Appendix Ragavendran Gopalakrishnan, Sherwin Doroudi, Amy R. Ward, and Adam Wierman In this technical appendix, we provide proofs for the results stated in the main body of the manuscript titled: “Routing and Staffing when Servers are Strategic”. The proofs of these results are in the order in which they appear in the main body. PROOFS FROM SECTION 3 Proof of Theorem 1. The starting point of this proof is the expression for the steady state probabilities of a general heterogeneous M /M /N system with Random routing, which was derived in [22]. Before stating this more general result, we first set up the required notation. Let µ1 , µ2 , . . . , µN denote the service rates of the N servers, and let ρj = µλj , 1 ≤ j ≤ N . We assume that PN −1 j=1 ρj > 1 for stability. Let (a1 , a2 , . . . , ak ) denote the state of the system when there are k jobs in the system (0 < k < N ) and the busy servers are {a1 , a2 , . . . , ak }, where 1 ≤ a1 < a2 < · · · < ak ≤ N . Let P (a1 , a2 , . . . , ak ) denote the steady state probability of the system being in state (a1 , a2 , . . . , ak ). Also, let Pk denote the steady state probability of k jobs in the system. Then, P (a1 , a2 , . . . , ak ) = (N − k)! P0 ρa1 ρa2 · · · ρak , N! (EC.1) where P0 , the steady state probability that the system is empty, is given by: P0 = N ! CNN , DN (EC.2) where, for 1 ≤ j ≤ N , values values from N ρ−1 CjN = sum of combinations of j ρ−1 i i = N −j+1 N −j+2 X X a1 =1 a2 =a1 +1 ··· N −j+j−1 X aj−1 =aj−2 +1 aj =aj−1 +1 and DN = N X j! CjN + j=1 Note that, CNN = N Y i=1 C0N N X ρ−1 i and −1 −1 ρ−1 a1 ρa2 · · · ρaj , C1N . C1N − 1 C1N = N X (EC.3) (EC.4) ρ−1 i . i=1 Also, by convention, we write = 1. The steady state probability that a tagged server, say server 1, is idle is obtained by summing up the steady state probabilities of every state in which server 1 is idle: N −1 X X P (a1 , a2 , . . . , ak ) (EC.5) I(µ1 , µ2 , . . . , µN ; λ, N ) = P0 + k=1 2≤a1 ≤···≤ak ≤N We now simplify the expressions above for our special system where the tagged server works at a rate µ1 and all other servers work at rate µ. Without loss of generality, we pick server 1 to be ec2 e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic the tagged server, and we set µ2 = µ3 = · · · = µN = µ, and therefore, ρ2 = ρ3 = · · · = ρN = ρ = µλ . Then, (EC.1) simplifies to: P (a1 , a2 , . . . , ak ) = (N − k)! P0 ρk , 2 ≤ a1 ≤ · · · ≤ ak ≤ N N! (EC.6) In order to simplify (EC.3), we observe that N −1 N −1 CjN = ρ−1 1 Cj−1 + Cj N −1 where the terms Cj−1 and CjN −1 are obtained by applying (EC.3) to a homogeneous M /M /(N −1) system with arrival rate λ and all servers operating at rate µ. This results in:         N −j N −j 1 ρ N j (N − j) ρ ρ + (EC.7) Cj = N ρ1 N j j N The corresponding special cases are given by: C0N = 1, C1N = ρ−1 1 + (N − 1)ρ, and CN = then simplify (EC.4) by substituting for CjN from (EC.7), to obtain: !    N   −1 j X N! ρ ρ ρ 1 ρ ρ DN = + −1 + 1+ N + 1− ρN ρ1 N ρ1 j! ρ1 C1 − 1 j=0      N −1 j X ρ1 ρ ρ ρ ρ N! ρ1   +1+ 1− = 1− ρ1 ρN N ρ j! ρ N − ρ+1− ρ j=0 ρ −N ρ . ρ1 We (EC.8) ρ1 Next, we simplify (EC.2) by substituting for DN from (EC.8), to obtain:      N −1 j N X ρ  ρ ρ ρ ρ1  1 + 1+ P0 =  1 − 1− N ρ j! N ! N − ρ+1− j=0 ρ ρ1 −1   To express P0 in terms of ErlC(N, ρ), the Erlang C formula, we add and subtract the term within, to obtain:     ρ ρ1  P0 = 1− 1− N ρ N−1 X j=0 ρj N ρN + j! N − ρ N! !  ρN  ρ  1 + 1+ N! N − ρ+1− ρ ρ1 N − N −ρ  ρ 1− N which reduces to:  !   N    −1 j N ρN N X ρ N ρ ρ ρ ρ1 ρ 1 N −ρ N !   − + 1− 1− P0 = 1− N ρ j! N − ρ N ! N ρ N − ρ+1− j=0 = N −1 X j=0 j N ρ N ρ + j! N − ρ N ! !−1  Finally, (EC.5) simplifies to: 1 − ρ N  1− ρ1 ρ I(µ1 , µ, µ, . . . , µ; λ, N ) = P0 +   1 + N −1  X k=1 −1 ErlC(N, ρ)    N − ρ + 1 − ρρ1  N −1 P (2, 3, . . . , k + 1) k  ρ ρ1 ρ1 1− ρ N ρN N −ρ N !  −1  −1  (EC.9) ec3 e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic Substituting for P0 from (EC.9) and P (2, 3, . . . , k + 1) from (EC.6), we get: N −1  X  (N − k)! P0 ρk I(µ1 , µ; λ, N ) = P0 + N! k=1 ! N −1  N ρN ρ  X ρk P0 + = 1− N k! N − ρ N ! k=0   −1    ρ ErlC(N, ρ)  ρ1  ρ    1− 1+ 1− = 1− N N ρ N − ρ+1− ρ N −1 k  ρ1 −1    ρ ErlC(N, ρ)  µ  ρ    1− 1+ 1− = 1− , N N µ1 N − ρ + 1 − µ1   as desired.  µ Proof of Theorem 2. We start with the expression for I from (4), and take its first partial derivative with respect to µ1 :  −2          ∂I ρ ρ  µ  µ ∂ ErlC(N, ρ) ρ ErlC(N, ρ) 1 − 1 +       1− =− 1− 1− 1− 1+ ∂µ1 N N µ1 ∂µ1 N µ1 N − ρ + 1 − µµ1 N − ρ + 1 − µµ1           N µ  ρ ErlC(N, ρ) ErlC(N, ρ) µ  ρ 2 ∂  2 ∂         =− 1− I I 1− 1+ 1+ 1− = N − ρ ∂µ1 N µ1 N − ρ ∂µ1 µ1 N − ρ + 1 − µ1 N − ρ + 1 − µ1 µ µ Applying the product rule, and simplifying the expression, we get (5). Next, for convenience, we rewrite (5) as:     ErlC(N, ρ) I2  µ1 µ1 N − ρ ∂I ErlC(N, ρ)    + 1− = 2 1 +     λ ∂µ1 µ1 µ µ N − ρ + 1 − µ1 2 N − ρ + 1 − µµ1 µ (EC.10) Differentiating this equation once more with respect to µ1 by applying the product rule, we get:     ErlC(N, ρ) ErlC(N, ρ) µ1 µ1   + 1−      µ µ N − ρ + 1 − µ1 2 N − ρ + 1 − µµ1 µ     I2 ∂  ErlC(N, ρ) ErlC(N, ρ) µ1 µ1    + 1− + 2   2  1 + µ1 µ1 µ1 ∂µ1 µ µ N − ρ+1− µ N − ρ+1− µ       2I ∂I ErlC(N, ρ) 2I 2 µ21 N − ρ ∂I I2 ∂  µ1 µ1 ErlC(N, ρ)    + 1− = − 3 + 2     1 + µ21 ∂µ1 µ1 I 2 λ ∂µ1 µ1 ∂µ1 µ µ N − ρ + 1 − µ1 2 N − ρ + 1 − µµ1 µ N − ρ ∂2I = λ ∂µ21  2I ∂I 2I 2 − 3 2 µ1 ∂µ1 µ1   1 + Applying the product rule for the second term, and simplifying the expression, we get: 2 ∂2I = ∂µ21 I  ∂I ∂µ1 2 2 − µ1  ∂I ∂µ1     1 λ ErlC(N, ρ) 2I 2 1 + 1 − µ1  −    2 µ1 µ2 N − ρ N − ρ + 1 − µ1 µ N − ρ+1− µ The expression in (6) is then obtained by substituting for some incredibly messy (but straightforward) algebra. ∂I ∂µ1 µ1 µ   from (5), and carefully going through ec4 e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic Proof of Theorem 3. In order to prove this theorem, we make the transformation t=ρ+1− µ1 µ (EC.11) + For example, when µ1 = µ1 = (λ − (N − 1)µ) , t = t = min (ρ + 1, N ). Using this transformation, the µI1 term that appears in the beginning of the expression for the second derivative of the idle time (6) can be written in terms of t as follows. I (N − ρ) (N − t) = µ1 µg(t) where g(t) = N (N − t) (ρ + 1 − t) − ρ (ρ − t) (N − t + ErlC(N, ρ)) Note that g(t) > 0, since I > 0, N > ρ, and from stability, N > t. Substituting this in (6), and using (EC.11) to complete the transformation, we get the following expression for the second derivative of the idle time in terms of t. 2 ∂2I 2λ (N − ρ) f (t) = H(t) = − 2 ∂µ1 µ3 g 3 (t) 3 where we use the notation g 3 (t) to denote (g(t)) , and     2 2 f (t) = (N − t) − ρErlC(N, ρ) (N − t + ErlC(N, ρ)) + N − (ρ − t) (ρ + 1 − t) ErlC(N, ρ) In order to prove the theorem, we now need  to show that (a) There exists a threshold t† ∈ −∞, t such that H(t) < 0 for −∞ < t < t† , and H(t) > 0 for t† < t < t. (b) H(t) > 0 ⇒ H ′ (t) > 0. To show these statements, we prove the following three properties of f and g. • f (t) is a decreasing function of t. • g(t) is a decreasing function of t. • f (0) > 0. In what follows, for convenience, we denote ErlC(N, ρ) simply by C. Differentiating f (t), we get

f ′ (t) = − (N − t)2 − ρC − 2(N − t)(N − t + C) − N − (ρ − t)2 C + 2(ρ − t)(ρ + 1 − t)C

= −3 (N − t)2 + −(ρ − t)2 + (N − ρ) C


= −3 (N − t)2 (1 − C) + (N − t)2 − (ρ − t)2 + (N − ρ) C
= −3 (N − t)2 (1 − C) + ((N − t + ρ − t)(N − ρ) + (N − ρ)) C
= −3 (N − t)2 (1 − C) + (N − t + ρ + 1 − t)(N − ρ)C <0 The last step follows by noting that N − t > 0, ρ + 1 − t ≥ 0, N − ρ > 0, and 0 < ErlC(N, ρ) < 1 when 0 < ρ < N . This shows that f (t) is a decreasing function of t. Next, differentiating g(t), we get g ′ (t) = −N (N − t) − N (ρ + 1 − t) + ρ(ρ − t) + ρ(N − t + C) = −N (N − t + ρ + 1 − t) + ρ(ρ + 1 − t) + ρ(N − t) − ρ(1 − C) = −(N − ρ)(N − t + ρ + 1 − t) − ρ(1 − C) <0 e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic ec5 The last step follows by noting that N − t > 0, ρ + 1 − t ≥ 0, N − ρ > 0, and 0 < ErlC(N, ρ) < 1 when 0 < ρ < N . This shows that g(t) is a decreasing function of t. Finally, evaluating f (0), we get f (0) = (N 2 − ρC)(N + C) + (N − ρ2 )(ρ + 1)C = N 3 − ρ3 C + N 2 C − ρ2 C + N C − ρC 2 = (N 3 − ρ3 ) + ρ3 (1 − C) + (N 2 − ρ2 )C + (N − ρ)C + ρC(1 − C) >0 The last step follows by noting that N − ρ > 0, and 0 < ErlC(N, ρ) < 1 when 0 < ρ < N . We are now ready to prove the statements (a-b).  (a) First, note that because f (t) is decreasing and f (0) > 0, there exists a threshold t† ∈ 0, t such that f (t) > 0 for −∞ < t < t† , and f (t) < 0 for t† < t < t. (Note that if f (t) > 0, then  we let t† = t so that f (t) < 0 in an empty interval.) Next, since g(t) > 0 for all t ∈ −∞, t , the sign of H(t) is simply the opposite of the sign of f (t). Statement (a) now follows directly. (b) Statement (b) is equivalent to showing that f (t) < 0 ⇒ H ′ (t) > 0. Differentiating H(t), we get   2λ(N − ρ)2 g 3 (t)f ′(t) − 3f (t)g 2(t)g ′ (t) ′ H (t) = − µ3 g 6 (t)   2λ(N − ρ)2 g(t)f ′(t) − 3f (t)g ′(t) =− µ3 g 4 (t) Since g(t) > 0, f ′ (t) < 0, and g ′ (t) < 0, it follows that H ′ (t) > 0 whenever f (t) < 0. This concludes the proof. Proof of Theorem 4. The “only if” direction is straightforward. Briefly, it follows from the fact that, by definition, any symmetric equilibrium µ∗ > Nλ must be an interior global maximizer of U (µ1 , µ⋆ ) in the interval µ1 ∈ ( Nλ , ∞). The “if” direction requires more care. We first show that the utility function U (µ1 , µ⋆ ) inherits the properties of the idle time function I(µ1 , µ⋆ ) as laid out in Theorem 3, and then consider the λ two cases when it is either increasing or decreasing ath µ1 = N . † Recall that U (µ1 , µ⋆ ) = I(µ1 , µ⋆ ) − c(µ1 ). Let µ1 ∈ µ1 , ∞ be the threshold of Theorem 3. We subdivide the interval (µ1 , ∞) as follows, in order to analyze U (µ1 , µ⋆ ). • Consider the interval (µ1 , µ†1 ), where, from Theorem 3, we know that I ′′′ (µ1 , µ⋆ ) < 0. Therefore, U ′′′ (µ1 , µ⋆ ) = I ′′′ (µ1 , µ⋆ ) − c′′′ (µ1 ) < 0. This means that U ′′ (µ1 , µ⋆ ) is decreasing in this interval. (Note that this interval could be empty, i.e., it is possible that µ†1 = µ1 .) • Consider the interval (µ†1 , ∞), where, from Theorem 3, we know that I ′′ (µ1 , µ⋆ ) < 0. Therefore, U ′′ (µ1 , µ⋆ ) = I ′′ (µ1 , µ⋆ ) − c′′ (µ1 ) < 0. This means that U (µ1 , µ⋆ ) is concave in this interval. Thus, the utility function U (µ1 , µ⋆ ), like the idle time function I(µ1 , µ⋆ ), may start out as a convex function at µ1 = µ1 , but it eventually becomes concave, and stays concave thereafter. Moreover, because the cost function c is increasing and convex, limµ1 →∞ U (µ1 , µ⋆ ) = −∞, which implies that U (µ1 , µ⋆ ) must eventually be decreasing concave. We now consider two possibilities for the behavior of U (µ1 , µ⋆ ) in the interval ( Nλ , ∞): λ . If µ†1 > Nλ (see Figure 1(a)), U (µ1 , µ⋆ ) would Case (I): U(µ1 , µ⋆ ) is increasing at µ1 = N start out being increasing convex, reach a rising point of inflection at µ1 = µ†1 , and then become increasing concave. (Otherwise, if µ†1 ≤ Nλ , U (µ1 , µ⋆ ) would just be increasing concave to begin with.) It would then go on to attain a (global) maximum, and finally become decreasing concave. This means that the unique stationary point of U (µ1 , µ⋆ ) in this interval must be at this (interior) global maximum. Since U ′ (µ⋆ , µ⋆ ) = 0 (from the symmetric first order condition (9)), µ1 = µ⋆ must be the global maximizer of the utility function U (µ1 , µ⋆ ), and hence a symmetric equilibrium. ec6 e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic λ Case (II): U(µ1 , µ⋆ ) is decreasing at µ1 = N . Because U (µ⋆ , µ⋆ ) ≥ U ( Nλ , µ⋆ ), U (µ1 , µ⋆ ) must eventually increase to a value at or above U ( Nλ , µ⋆ ), which means it must start out being decreasing convex (see Figure 1(b)), attain a minimum, then become increasing convex. It would then follow the same pattern as in the previous case, i.e., reach a rising point of inflection at µ1 = µ†1 , and then become increasing concave, go on to attain a (global) maximum, and finally become decreasing concave. This means that it admits two stationary points – a minimum and a maximum. Since U ′ (µ⋆ , µ⋆ ) = 0 (from the symmetric first order condition (9)) and U (µ⋆ , µ⋆ ) ≥ U ( Nλ , µ⋆ ), µ1 = µ⋆ must be the (global) maximizer, and hence a symmetric equilibrium. U (µ1 , µ⋆ ) U (µ1 , µ⋆ ) • • • • µ† µ1 µ⋆ λ/N µ† µ⋆ µ1 λ/N (a) Case (I) (b) Case (II) Figure EC.1. The graphic depiction of the proof of Theorem 4. Note that if U (µ1 , µ⋆ ) is stationary at µ1 = Nλ , it could either start out increasing or decreasing in the interval ( Nλ , ∞), and one of the two cases discussed above would apply accordingly. Finally, to conclude the proof, note that (10) is equivalent to the inequality U (µ⋆ , µ⋆ ) ≥ U ( Nλ , µ⋆ ), obtained by plugging in and evaluating the utilities using (7) and (4). This completes the proof. Proof of Theorem 5. The symmetric first order condition (9) can be rewritten as  λ ErlC N, µ  = µ2 c′ (µ) N2 λ + −N λ µ
It suffices to show that if λc′ Nλ < 1, then the left hand side and the right hand side intersect at
least once in Nλ , ∞ . We first observe that the left hand side, the Erlang-C function, is shown to be convex and increasing in ρ = µλ (pages 8 and 11 of [43]). This means that it is decreasing and convex in µ. Moreover, ErlC(N, N ) = 1 and ErlC(N, 0) = 0, which means that the left hand side decreases from 1 to 0 in a convex
fashion as µ runs from Nλ to ∞. The right hand side is clearly ∞. Therefore, convex in
µ, and is equal to λc′ Nλ when µ = Nλ , and approaches ∞ as µ approaches
if λc′ Nλ < 1, then the two curves must intersect at least once in Nλ , ∞ . e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic ec7

2
Next, it is sufficient to show that if 2 Nλ c′ Nλ + Nλ c′′ Nλ ≥ 1, then the right hand side is non-decreasing in µ. In order to do so, it suffices to show that   ∂ N2 λ 2 ′ + −N ≥0 µ c (µ) ∂µ λ µ
N2 λ − 2 ≥0 ⇔ µ2 c′′ (µ) + 2µc′ (µ) λ µ  2
Nµ ⇔ µ2 c′′ (µ) + 2µc′ (µ) ≥1 λ
The left hand side is a non-decreasing function of µ, therefore, in the interval Nλ , ∞ , we have   2     
Nµ 2 λ λ λ ′ λ ′′ µ c (µ) + 2µc (µ) ≥ c +2 c ≥ 1. λ N N N N 2 ′′ ′ This completes the proof. Proof of Theorem 6. The utility at any symmetric point can be evaluated as U (µ, µ) = 1 − − c(µ), using (7) and (4). Therefore, it follows that showing U (µ⋆1 , µ⋆1 ) > U (µ⋆2 , µ⋆2 ) is equivalent to showing that c(µ⋆1 ) − c(µ⋆2 ) λ < . ⋆ ⋆ µ1 − µ2 N µ⋆1 µ⋆2 λ Nµ The function c is convex by assumption. It follows that c(µ⋆1 ) − c(µ⋆2 ) ≤ (µ⋆1 − µ⋆2 )c′ (µ⋆1 ). (EC.12) Therefore, rearranging and substituting for c′ (µ⋆1 ) from the symmetric first order condition (9),    λ λ λ c(µ⋆1 ) − c(µ⋆2 ) ≤ 2 ⋆ 2 N − ⋆ + ErlC N, ⋆ . µ⋆1 − µ⋆2 N (µ1 ) µ1 µ1   λ It has been shown (page 14 of [43], and [27]) that ErlC N, µ < Nλµ . Using this, c(µ⋆1 ) − c(µ⋆2 ) λ < 2 ⋆ 2 ⋆ ⋆ µ1 − µ2 N (µ1 )    λ λ λ λ 1 . < N − ⋆ 1− < 2 ⋆ 2 (N ) = ⋆ 2 µ1 N N (µ1 ) N (µ1 ) N µ⋆1 µ⋆2 This completes the proof. PROOFS FROM SECTION 4 Proof of Proposition 1. We first observe that if f (λ) = ω(λ), then C ⋆,λ (N λ ) Nλ ≥ cS → ∞ as λ → ∞. λ λ Since Proposition 2 evidences a staffing policy under which C ⋆,λ (N λ )/λ has a finite limit, having f (λ) = ω(λ) cannot result in an asymptotically optimal staffing policy. Next, we consider the case f (λ) = o(λ). In this case, λ/N λ → ∞. Since any symmetric equilibrium must have µ⋆,λ > λ/N λ from (3), it follows that if there exists a sequence of symmetric equilibria {µ⋆,λ }, then µ⋆,λ → ∞ as λ → ∞. We conclude that such a staffing policy cannot be admissible. ec8 e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic Proof of Theorem 7. We can rewrite (16) as f (µ) = g(µ) where f (µ) = µ2 1 1 and g(µ) = 2 c′ (µ) + . a a µ The two cases of interest (i) and (ii) are as shown in Figure EC.2. Our strategy for the proof is to rewrite (14) in terms of functions f λ and g λ that are in some sense close to f and g. Then, in case (i), the fact that g(µ) lies below f (µ) for µ ∈ [µ1 , µ2 ] implies that f λ and g λ intersect (at least) twice. The case (ii) is more delicate, because the sign of o(λ) determines if the functions f λ and g λ will cross (at least) twice or not at all. (We remark that it will become clear in that part of the proof where the condition o(λ) < −3 is needed.) 1 a + c′ (a) 1 a g(µ) • 1 a • + c′ (a) g(µ) f (µ) • 1 a 0 µ1 µ2 µ 0 a µ1 f (µ) µ a (a) Case (i) (b) Case (ii) Figure EC.2. The limiting first order condition (16). The first step is to rewrite (14) as f λ (µ) = g λ (µ) where   1 Nλ λ λ f (µ) = ErlC N , + λ µ λ  λ 2 N 1 g λ (µ) = µ2 c′ (µ) + . λ µ λ The function g λ converges uniformly on compact sets to g since for any µ > 0, substituting for N λ in (13) shows that  2 ! 2 o(λ) o(λ) sup g λ (µ) − g(µ) ≤ µ2 c′ (µ) → 0, (EC.13) + a λ λ µ∈[0,µ] as λ → ∞. Next, recall ErlC(N, ρ) ≤ 1 whenever ρ/N < 1. Since   1 o(λ) λ 1 λ λ + o(λ), + f (µ) − f (µ) ≤ ErlC λ a µ λ (EC.14) e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic ec9 and ErlC(λ/a + o(λ), λ/µ) ≤ 1 for all µ > a for all large enough λ, the function f λ converges uniformly to f on any compact set [a + ǫ, µ] with µ > a + ǫ and ǫ arbitrarily small. The reason we need only consider compact sets having lower bound a + ǫ is that it is straightforward to see any solution to (16) has µ > a. It is also helpful to note that g λ is convex in µ because
1 d2 λ g (µ) = 2c′ (µ) + 4µc′′ (µ) + µ2 c′′′ (µ) + 2 3 > 0 for µ ∈ (0, ∞), 2 dµ µ and f λ is convex decreasing in µ because ErlC(N, ρ) is convex increasing in ρ (pages 8 and 11 of [43]). We prove (i) and then (ii). Proof of (i): There exists µm ∈ (µ1 , µ2 ) for which f (µm ) > g(µm ). Then, it follows from (EC.13) and (EC.14) that f λ (µm ) > g λ (µm ) for all large enough λ. Also, lim f λ (µ) = µ→∞ and 1 < lim g λ (µ) = ∞. a µ→∞ 1 Nλ < gλ lim f (µ) = + λ λ µ↓λ/N λ λ  λ Nλ  =c ′  λ Nλ  + Nλ λ for all large enough λ, where the inequality follows because c is strictly increasing. Since f λ is convex decreasing and g λ is convex, we conclude that there exist two solutions to (14). Proof of (ii): We prove part (a) and then part (b). Recall that µ1 is the only µ > 0 for which f (µ1 ) = g(µ1 ). Proof of (ii)(a): For part (a), it is enough to show that for all large enough λ, f λ (µ1 ) − g λ (µ1 ) > 0. (EC.15) The remainder of the argument follows as in the proof of part (i). From the definition of f λ and g λ in the second paragraph of this proof, and substituting for N λ , f λ (µ1 ) − g λ (µ1 )  2      µ 2 1 2 1 o(λ) 1 λ 1 o(λ) 1 = − 1 − (µ1 )2 c′ (µ1 ) − (µ1 )2 c′ (µ1 ) c′ (µ1 ) + ErlC + − λ + o(λ), . a µ1 a λ a µ1 λ a λ It follows from f (µ1 ) = g(µ1 ) that 1/a − 1/µ1 − (µ1 /a)2 c′ (µ1 ) = 0, and so, also noting that ErlC(λ/a + o(λ), λ/µ1 ) > 0,    2 o(λ) o(λ) 2 2 ′ 2 ′ f (µ1 ) − g (µ1 ) > . 1 − (µ1 ) c (µ1 ) − (µ1 ) c (µ1 ) λ a λ λ λ (EC.16) Again using the fact that f (µ1 ) = g(µ1 ),   1 1 a 2 2 ′ − = −1 + 2 . 1 − (µ1 ) c (µ1 ) = 1 − 2a a a µ1 µ1 Then, the term multiplying o(λ)/λ in (EC.16) is positive if −1+2 a > 0, µ1 which implies (EC.15) holds for all large enough λ. (EC.17) ec10 e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic To see (EC.17), and so complete the proof of part (ii)(a), note that since µ1 solves (16), and the left-hand side of (16) is convex increasing while the right-hand side is concave increasing, µ1 also solves the result of differentiating (16), which is Algebra shows that
1 1 = 2 µ21 c′′ (µ1 ) + 2µ1 c′ (µ1 ) . 2 µ1 a  2  1 µ1 ′ µ31 ′′ −2 c (µ ) = c (µ1 ). 1 µ1 a2 a2 We next use (16) to substitute for µ2 1 ′ c (µ1 ) a2 to find 3 2 µ3 − = 21 c′′ (µ1 ). µ1 a a Since c is convex, 2 3 − ≥ 0, µ1 a and so 1.5a ≥ µ1 , from which (EC.17) follows. Proof of (ii)(b): Let µλ ∈ (0, ∞) be the minimizer of the function g λ . The minimizer exists because g λ is convex and  
Nλ 2 1 d λ ′ 2 ′′ g (µ) = µc (µ) + µ c (µ) − 2, dµ λ µ which is negative for all small enough µ, and positive for all large enough µ. It is sufficient to show that for all large enough λ g λ (µ) − f λ(µ) > 0 for all µ ∈ [a, µλ ]. (EC.18) This is because for all µ > µλ , g λ is increasing and f λ is decreasing. Suppose we can establish that for all large enough λ 2 ǫλ 1 ≥ − , for all µ ∈ [a, µλ ], µ 3a 2a (EC.19) where ǫλ satisfies ǫλ → 0 as λ → ∞. Since g(µ) ≥ f (µ) for all µ, it follows that  λ 2    λ 2 N a2 N 1 1 g λ (µ) = µ2 c′ (µ) + ≥ a− + . λ µ µ λ µ Substituting for N λ and algebra shows that    λ 2      2  o(λ) N o(λ) 1 1 a a a2 . + = + 2 1− +a 1− a− µ λ µ a λ µ µ λ Then, from the definition of f λ and the above lower bound on g λ , also using the fact that the assumption N λ − λ/a < 0 implies the term o(λ) is negative,       2 o(λ) 2a 1 λ o(λ) a g λ (µ) − f λ(µ) ≥ . − 1 − ErlC N λ , +a 1− λ µ λ µ µ λ Since −ErlC(N λ , λ/µ) > −1 and 1/a − 1/µ > 0 from (16) implies 1 − a/µ > 0,     1 2a λ λ g (µ) − f (µ) ≥ −1 −1 . |o(λ)| λ µ e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic Next, from (EC.19), and so ec11 2a 4 ≥ − ǫλ , µ 3     1 1 λ g (µ) − f (µ) ≥ −ǫ −1 . |o(λ)| λ 3 λ λ The fact that |o(λ)| > 3 and ǫλ → 0 then implies that for all large enough λ, (EC.18) is satisfied. Finally, to complete the proof, we show that (EC.19) holds. First note that µλ as the minimizer of g λ satisfies  
Nλ 2 1 ′ 2 ′′ 2µc (µ) + µ c (µ) − 2 = 0, λ µ and that solution is unique and continuous in λ. Hence µλ → µ1 as λ → ∞. Then, g λ (µλ ) → g(µ1 ) = 1 as λ → ∞. a Furthermore, g λ (µλ ) approaches g(µ1 ) from above; i.e., g λ (µλ ) ↓ 1 as λ → ∞, a because, recalling that the term o(λ) is negative, g λ (µ) = µ2 c′ (µ)  1 o(λ) − a λ 2 + 1 > g(µ) for all µ > 0. µ Therefore, there exists ǫλ → 0 such that g λ (µλ ) = 1 3 ǫλ − , a 4a where the 3/(4a) multiplier of ǫλ is chosen for convenience when obtaining the bound in the previous paragraph. Finally, 1 1 ≥ λ µ µ means that (EC.19) follows if 1 2 1 1 ǫλ 2 λ λ g (µ ) = − . ≥ µλ 3 3a 2 a To see the above display is valid, note that µλ solves
′ g λ (µ) = 0, which from algebra is equivalent to 2g λ (µλ ) − Hence  λ 2
N 3 λ 3 ′′ λ + µ c (µ ) = 0. µλ λ 2g λ (µλ ) − as required. 3 ≤ 0, µλ ec12 e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic Proof of Lemma 1. It is enough to show the inequality (10) of Theorem 4 holds. The function c is convex by assumption. It follows that     λ λ λ c(µλ) − c ≤ µ − c′ (µλ ). (EC.20) Nλ Nλ Plugging in for c′ (µλ ) from the symmetric first order condition (14) yields (after algebra)      ErlC (N λ , λ/µλ ) λ λ λ λ ′ λ λ 1− λ λ + . µ − λ c (µ ) = λ λ 1 − λ λ N µ N µ N µ N Nλ Hence, in order to show the inequality (10) is true, also substituting for ρλ = λ/µλ , it is enough to verify that λ µλ N λ  λ 1− λ λ µ N ⇐⇒  ErlC N λ , λ/µλ λ 1− λ λ + µ N Nλ 1+
! 1 1− λ µλ N λ + ErlC(N λ ,λ/µλ ) N λ −1 Since N λ − 1 < N λ , it is enough to show that  λ ≤ 1− λ λ µ N ≤ λ µλ N λ  λ ErlC(N λ , λ/µλ ) 1+ 1− λ λ + µ N N −1  1  . λ ,λ/µλ ) λ 1 − µλ N λ + ErlC(N λ N −1 !−1 1   ErlC(N λ ,λ/µλ ) λ λ 1 − µλλN λ + 1 − + λ N µλ N λ µλ N λ Nλ   1 − µλλN λ   ⇐⇒ 1≤ ErlC(N λ ,λ/µλ ) λ λ 1 − + λ λ λ λ λ µ N µ N N     λ λ λ ErlC(N λ , λ/µλ ) λ ≤ 1− λ λ 1− λ λ + λ λ ⇐⇒ µλ N λ µ N µ N Nλ µ N 2  λ λ λ λ ErlC(N , λ/µ ) N µ ⇐⇒ −1 . ≤ λ/µλ λ 1 1+ ErlC(N λ ,λ/µλ ) ≤ Since N λ µλ /λ → d > 1 by assumption, the limit of the right-hand side of the above expression is positive, and, since and ErlC(N λ , λ/µλ ) ≤ 1, the limit of the left-hand side of the above expression is 0. We conclude that for all large enough λ, the above inequality is valid. Proof of Proposition 2. Let µ⋆ = arg min{µ > 0 : (16) holds }. Next, recalling that µ⋆ > a, also let µ = µ⋆ −
1 ⋆ µ − a > a, 2 so that the system is stable if all servers were to work at rate µ (λ < µN λ for all large enough λ). It follows from Theorem 7 that, for all large enough λ, any µλ that satisfies the first order condition (14) also satisfies µλ > µ. Hence any symmetric equilibrium µ⋆,λ must also satisfy µ⋆,λ > µ for all large enough λ, and so ⋆,λ λ W < W µ. Therefore, also using the fact that W cS ⋆,λ > 0, it follows that Nλ N λ C ⋆,λ (N λ ) Nλ ⋆,λ λ < = cS + wW < cS + wW µ . λ λ λ λ e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic ec13 Then, since N λ /λ → 1/a as λ → ∞ from (13), it is sufficient to show λ W µ → 0 as λ → ∞. This follows from substituting the staffing N λ = λ/a + o(λ) in (13) into the well-known formula for the steady state mean waiting time in a M /M /N λ queue with arrival rate λ and service rate µ as follows   λ/µ 1 λ λ λ Wµ = ErlC N , λ N λ − λ/µ µ   1/µ λ
= ErlC N λ , µ 1/a − 1/µ λ + o(λ) → 0, as λ → ∞, since ErlC(N λ , λ/µ) ∈ [0, 1] for all λ. Proof of Lemma 2. It follows from the equation a(µ − a) = µ3 c′ (µ) that   + p µ ′ a= 1 − 1 − 4µc (µ) . 2 The condition 4µc′ (µ) ≤ 1 is required to ensure that there is a real-valued solution for a. Hence     + p µ ′ ′ 1 − 1 − 4µc (µ) : 0 ≤ 4µc (µ) ≤ 1 . A= 2 Since c′ (µ) is well-behaved, this implies that A is compact, and, in particular, closed. We conclude that a⋆ = sup A ∈ A, which implies that a⋆ is finite. Proof of Theorem 8. It follows from Proposition 1 that 0 ≤ lim inf λ→∞ N opt,λ N opt,λ ≤ lim sup < ∞, λ λ λ→∞ because any staffing policy that is not asymptotically optimal also is not optimal for each λ. Consider any subsequence λ′ on which either lim inf λ→∞ N opt,λ /λ or lim supλ→∞ N opt,λ /λ is attained, and suppose that ′ 1 N opt,λ (EC.21) → as λ′ → ∞, where a ∈ [0, ∞). λ′ a The definition of asymptotic optimality requires that for each λ′ , there exists a symmetric equi′ librium service rate µ⋆,λ . As in the proof of Lemma 1, it is enough to consider sequences {µλ } that satisfy the first order condition (14). Then, by the last sentence of Theorem 7, any sequence ′ ′ of solutions {µλ } to (14) must be such that |µλ − µ| is arbitrarily small, for λ′ large enough, for some µ that solves (16), given a in (EC.21). In summary, the choice of a in (EC.21) is constrained by the requirement that a symmetric equilibrium service rate must exist. Given that there exists at least one symmetric equilibrium service rate for all large enough λ′ , it follows in a manner very similar to the proof of Proposition 2 that W ⋆,λ′ → 0 as λ′ → ∞, ec14 e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic even though when there are multiple equilibria we may not be able to guarantee which symmetric ′ equilibrium µ⋆,λ the servers choose for each λ′ . We conclude that ′ ′ ′ C ⋆,λ (N opt,λ ) N opt,λ 1 ⋆,λ′ = c + wW → cS , as λ′ → ∞. S ′ ′ λ λ a (EC.22) We argue by contradiction that a in (EC.22) must equal a⋆ . Suppose not. Then, since C ⋆,λ (N ao,λ ) 1 → cS ⋆ as λ → ∞ λ a by Proposition 2 (and so the above limit is true on any subsequence), and a⋆ > a by its definition, it follows that ′ ′ ′ ′ C ⋆,λ (N ao,λ ) < C ⋆,λ (N opt,λ ) for all large enough λ′ . ′ The above inequality contradicts the definition of N opt,λ . The previous argument did not depend on if λ′ was the subsequence on which lim inf λ→∞ N opt,λ /λ or lim supλ→∞ N opt,λ /λ was attained. Hence N opt,λ 1 = ⋆, λ→∞ λ a lim and, furthermore, 1 C ⋆,λ (N opt,λ ) = cS ⋆ . λ→∞ λ a lim Since also , 1 C ⋆ λ (N ao,λ ) = cS ⋆ , lim λ→∞ λ a the proof is complete. Proof of Lemma 3. We first observe that (16) is equivalently written as: 0 = cE pµp+2 − aµ + a2 . The function f (µ) = cE pµp+2 − aµ + a2 attains its minimum value in (0, ∞) at µ=  a cE p(p + 2) 1/(p+1) . The function f is convex in (0, ∞) because f ′′ (µ) > 0 for all µ ∈ (0, ∞) and so µ is the unique minimum. It follows that    <  there are 2 non-negative solutions to (16) there is no non-negative solution to (16) . if f (µ) > 0, then    there is exactly one solution to (16) = Since   p+2 p+2 f (µ) = a p+1 a2− p+1 − △ for △ :=  1 cE p(p + 1) 1  p+1 1−  1 cE p p+1  1 p+2 p+2 ! > 0, e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic ec15 it follows that p if a p+1    <  there are 2 non-negative solutions to (16) there is no non-negative solution to (16) . − △ > 0, then    there is exactly one solution to (16) = The expression for △ can be simplified so that (p + 1) △= (p + 2)  1 cE p(p + 2) 1  p+1 . Then, a⋆ follows by noting that a⋆ = △(p+1)/p and µ⋆ follows by noting that µ⋆ = µ and then substituting for a⋆ . To complete the proof, we must show that a⋆ and µ⋆ are both increasing in p. This is because we have already observed that any solution to (16) has a < µ, and the fact that µ < 1 follows directly from the expression for µ. We first show a⋆ is increasing in p, and then argue that this implies µ⋆ is increasing in p. To see that a⋆ is increasing in p, we take the derivative of log a⋆ (p) and show that this is positive. Since log a⋆ (p) = log(p + 1) − log(p + 2) + 1 log(p + 1) p 1 2 1 − log cE − log p − log(2 + p), p p p it follows that   1 p/(p + 1) − log(p + 1) 1 1 (log a (p)) = − + + 2 log cE 2 p+1 p+2 p p ! p p − log(p) − log(p + 2) p p+2 −2 . − 2 p p2 ⋆ ′ After much simplification, we have 1 1 (log a (p)) = 2 log cE + 2 p p ⋆ ′     p(p + 2)2 p2 + p + 4 log − . p+1 (p + 1)(p + 2) Hence it is enough to show that △(p) = log  p(p + 2)2 p+1  − p2 + p + 4 ≥ 0, for p ≥ 1. (p + 1)(p + 2) This follows because the first term is increasing in p, and has a value that exceeds 1 when p = 1; on the other hand, the second term has a value that is strictly below 1 for all p ≥ 1. Finally, it remains to argue that µ⋆ is increasing in p. At the value µ = µ⋆ g(µ) = µ3 c′ (µ) − aµ + a2 = 0. At the unique point where the minimum is attained, it is also true that g ′ (µ) = µ3 c′′ (µ) + 3µ2 c′ (µ) − a = 0. Since µ3 c′′ (µ) + 3µ2 c′ (µ) is an increasing function of µ, it follows that if a increases, then µ must increase. ec16 e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic PROOFS FROM SECTION 5 Proof of Theorem 9. It is sufficient to verify the detailed balance equations. For reference, it is helpful to refer to Figure EC.3, which depicts the relevant portion of the Markov chain. We require the following additional notation. For all I ⊆ {1, 2, . . . , N }, all states s = (s1 , s2 , . . . , s|I| ), all servers s′ ∈ {1, 2, . . . , N }\I, and integers j ∈ {1, 2, . . . , |I| + 1}, we define the state s[s′ , j] by s[s′ , j] ≡ (s1 , s2 , . . . , sj−1 , s′ , sj , . . . , s|I|). We first observe that: Rate into state s due to an arrival = λ X X |I|+1 ′ πs[s′ ,j] pI∪{s } (j) s′ 6∈I j=1 |I| XX µs′ πB Y  µs  I∪{s′ } =λ p (j) λ s∈I λ s′ 6∈I j=0 Y µs X X = µs′ πs = µs′ πB λ ′ ′ s∈I s 6∈I s 6∈I = Rate out of state s due to a departure. Then, to complete the proof, we next observe that for each s′ 6∈ I: Rate into state s due to a departure = µs|I| π(s1 ,s2 ,...,s|I|−1 ) Y µs = µs|I| πB λ s∈I\{s|I| } s[s′ , 1] s − s1 λp I ∪{ s′ } I λp (1) ′ s[s′ , 2] ... s + s′ λpI∪{s } (2) (1) λpI (2) s ′ I λp s ∪{ } | (|I µs ′ +1 ) λp I (|I | µs|I| ) s − s2 ... s − s|I| For each s′ 6∈ I Figure EC.3. Snippet of the Markov chain showing the rates into and out of state s = (s1 , . . . , s|I| ). For convenience, we use s − sj to denote the state (s1 , s2 , . . . , sj−1 , sj+1 , . . . , s|I| ) and s + s′ to denote the state s[s′ , |I| + 1] = (s1 , s2 , . . . , s|I| , s′ ). e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic ec17 1 (1) λp λ λ µ2 µ1 0 λ 2 µ1 µ2 ··· 3 µ1 + µ2 µ1 + µ2 λ λ(1 − p) 1 (2) Figure EC.4. The M /M /2 Markov chain with probabilistic routing = λπB Y µs s∈I λ = λπs = Rate out of state s due to an arrival. Proof of Proposition 3. In order to derive the steady state probability that a server is idle, we first solve for the steady state probabilities of the M /M /2 system (with arrival rate λ and service rates µ1 and µ2 respectively) under an arbitrary probabilistic routing policy where a job that arrives to find an empty system is routed to server 1 with probabilityr p and server 2 with 1 probability 1 − p. Then, for an r-routing policy, we simply substitute p = µrµ+µ r. 1 2 It should be noted that this analysis (and more) for 2 servers has been carried out by [34]. Prior to that, [39] carried out a partial analysis (by analyzing an r-routing policy with r = 1). However, we rederive the expressions using our notation for clarity. The dynamics of this system can be represented by a continuous time Markov chain shown in Figure EC.4 whose state space is simply given by the number of jobs in the system, except when there is just a single job in the system, in which case the state variable also includes information about which of the two servers is serving that job. This system is stable when µ1 + µ2 > λ and we denote the steady state probabilities as follows: • π0 is the steady state probability that the system is empty. (j) • π1 is the steady state probability that there is one job in the system, served by server j. • For all k ≥ 2, πk is the steady state probability that there are k jobs in the system. We can write down the balance equations of the Markov chain as follows: (1) (2) λπ0 = µ1 π1 + µ2 π1 (1) (λ + µ1 )π1 = λpπ0 + µ2 π2 (2) (λ + µ2 )π1 = λ(1 − p)π0 + µ1 π2 (1) (2) (λ + µ1 + µ2 )π2 = λπ1 + λπ1 + (µ1 + µ2 )π3 ∀k ≥ 3 : (λ + µ1 + µ2 )πk = λπk−1 + (µ1 + µ2 )πk+1 , yielding the following solution to the steady state probabilities: π0 = µ1 µ2 (µ1 + µ2 )2 µ1 µ2 (µ1 + µ2 − λ)(µ1 + µ2 + 2λ) + λ(µ1 + µ2 )(µ22 + 2µ1 µ2 + (1 − p)(µ21 − µ22 )) + λ2 (µ21 + µ22 ) (EC.23) ec18 e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic λ(λ + p(µ1 + µ2 ))π0 µ1 (µ1 + µ2 + 2λ) λ(λ + (1 − p)(µ1 + µ2 ))π0 (2) . π1 = µ2 (µ1 + µ2 + 2λ) (1) π1 = Consequently, the steady state probability that server 1 is idle is given by   λ(λ + (1 − p)(µ1 + µ2 )) (2) π0 . I1 (µ1 , µ2 ; p) = π0 + π1 = 1 + µ2 (µ1 + µ2 + 2λ) Substituting for π0 , we obtain I1 (µ1 , µ2 ; p) = µ1 (µ1 + µ2 − λ) [(λ + µ2 )2 + µ1 µ2 + (1 − p)λ(µ1 + µ2 )] . (EC.24) µ1 µ2 (µ1 + µ2 )2 + λ(µ1 + µ2 ) [µ22 + 2µ1 µ2 + (1 − p)(µ21 − µ22 )] + λ2 (µ21 + µ22 ) Finally, for an r-routing policy, we let p = I1r (µ1 , µ2 ) = I1 (µ1 , µ2 ; p = µr1 µr1 +µr2 to obtain: µr1 ) µr1 + µr2 h i r 2 µ1 (µ1 + µ2 − λ) (λ + µ2 )2 + µ1 µ2 + µrµ+µ r λ(µ1 + µ2 ) 1 2 h i = . µr2 2 2 µ1 µ2 (µ1 + µ2 ) + λ(µ1 + µ2 ) µ2 + 2µ1 µ2 + µr +µr (µ21 − µ22 ) + λ2 (µ21 + µ22 ) 1 2 By symmetry of the r-routing policy, it can be verified that I2r (µ1 , µ2 ) = I1r (µ2 , µ1 ), completing the proof. Proof of
Theorem 10. We first highlight that when all servers operate at the same rate µ ∈ Nλ , ∞ , both FSF and SSF are equivalent to Random routing. Henceforth, we refer to such a configuration as a symmetric operating point µ. In order to prove that there does not exist a symmetric equilibrium under either FSF or SSF, we show that at any symmetric operating point µ, any one server can attain a strictly higher utility by unilaterally setting her service rate to be slightly lower (in the case of FSF) or slightly higher (in the case of SSF) than µ. We borrow some notation from the proof of Proposition 3 where we derived the expressions for the steady state probability that a server is idle when there are only 2 servers under any probabilistic policy, parameterized by a number p ∈ [0, 1] which denotes the probability that a job arriving to an empty system is routed to server 1. Recall that I1 (µ1 , µ2 ; p) denotes the steady state probability that server 1 is idle under such a probabilistic policy, and the corresponding utility function for server 1 is U1 (µ1 , µ2 ; p) = I1 (µ1 , µ2 ; p) − c(µ1). Then, by definition, the utility function for server 1 under FSF is given by:   U1 (µ1 , µ2 ; p = 0)
, µ1 < µ2 F SF U1 (µ1 , µ2 ) = U1 µ1 , µ2 ; p = 21 , µ1 = µ2   U1 (µ1 , µ2 ; p = 1) , µ1 > µ2 . Similarly, under SSF, we have:   U1 (µ1 , µ2 ; p = 1)
SSF U1 (µ1 , µ2 ) = U1 µ1 , µ2 ; p = 21   U1 (µ1 , µ2 ; p = 0) , µ1 < µ2 , µ1 = µ2 , µ1 > µ2 . Note that while the utility function under any probabilistic routing policy is continuous everywhere, the utility function under FSF or SSF is discontinuous at symmetric operating points. This e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic ec19 discontinuity turns out to be the crucial tool in the proof. Let the two servers be operating at a symmetric operating point µ. Then, it is sufficient to show that there exists 0 < δ < µ − λ2 such that U1F SF (µ − δ, µ) − U1F SF (µ, µ) > 0, (EC.25) U1SSF (µ + δ, µ) − U1F SF (µ, µ) > 0. (EC.26) and We show (EC.25), and (EC.26) follows from a similar argument. Note that   1 F SF F SF U1 (µ − δ, µ) − U1 (µ, µ) = U1 (µ − δ,µ; p = 0) − U1 µ, µ; p = 2
= U1 (µ − δ,µ; p = 0) − U1 (µ, µ; p = 0)    1 + U1 (µ, µ; p = 0) − U1 µ, µ; p = 2 Since the first difference, U1 (µ − δ, µ; p = 0) − U1 (µ, µ; p = 0), is zero when δ = 0, and is continuous in δ, it is sufficient to show that the second difference, U1 (µ, µ; p = 0) − U1 (µ, µ; p = 12 ), is strictly positive:     1 1 = I1 (µ, µ; p = 0) − I1 µ, µ; p = U1 (µ, µ; p = 0) − U1 µ, µ; p = 2 2
λ(2µ − λ) >0 using (EC.24) . = (µ + λ)(2µ + λ) This completes the proof. Proof of Theorem 11. The proof of this theorem consists of two parts. First, we show that under any r-routing policy, any symmetric equilibrium µ⋆ ∈ ( λ2 , ∞) must satisfy the equation ϕ(µ⋆ ) = r. This is a direct consequence of the necessary first order condition for the utility function of server 1 to attain an interior maximum at µ⋆ . The second part of the proof involves using the condition c′ ( λ2 ) < λ1 to show that ϕ is a strictly decreasing bijection onto R, which would lead to the following implications: • ϕ is invertible; therefore, if an r-routing policy admits a symmetric equilibrium, it is unique, and is given by µ⋆ = ϕ−1 (r). • ϕ−1 (r) is strictly decreasing in r; therefore, so is the unique symmetric equilibrium (if it exists). Since the mean response time E[T ] is inversely related to the service rate, this establishes that E[T ] at symmetric equilibrium (across r-routing policies that admit one) is increasing in r. We begin with the first order condition for an interior maximum. The utility function of server 1 under an r-routing policy, from (2), is given by U1r (µ1 , µ2 ) = I1r (µ1 , µ2 ) − c(µ1 ) For µ⋆ ∈ (λ/2, ∞) to be a symmetric equilibrium, the function U1r (µ1 , µ⋆ ) must attain a global maximum at µ1 = µ⋆ . The corresponding first order condition is then given by: ∂I1r (µ1 , µ⋆ ) ∂µ1 = c′ (µ⋆ ), (EC.27) µ1 =µ⋆ where I1r is given by Proposition 3. The partial derivative of the idle time can be computed and the left hand side of the above equation evaluates to ∂I1r (µ1 , µ⋆ ) ∂µ1 = µ1 =µ⋆ λ(4λ + 4µ⋆ + λr − 2µ⋆ r) . 4µ⋆ (λ + µ⋆ )(λ + 2µ⋆ ) (EC.28) ec20 e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic Substituting in (EC.27) and rearranging the terms, we obtain: 4(λ + µ⋆ ) (µ⋆ (λ + 2µ⋆ )c′ (µ⋆ ) − λ) = r. λ(λ − 2µ⋆ ) The left hand side is equal to ϕ(µ⋆ ), thus yielding the necessary condition ϕ(µ⋆ ) = r. Next, we proceed to show that if c′ ( λ2 ) < λ1 , then ϕ is a strictly decreasing bijection onto R. Note that the function 4(λ + µ) ϕ(µ) = (µ(λ + 2µ)c′ (µ) − λ) λ(λ − 2µ) is clearly a continuous function in ( λ2 , ∞). In addition, it is a surjection onto R, as evidenced by the facts that ϕ(µ) → −∞ as µ → ∞ and ϕ(µ) → ∞ as µ → λ2 + (using c′ ( λ2 ) < λ1 ). To complete the proof, it is sufficient to show that ϕ′ (µ) < 0 for all µ ∈ ( λ2 , ∞). First, observe that 4ψ(µ) , ϕ′ (µ) = λ(λ − 2µ)2 where ψ(µ) = µ(λ + µ)(λ2 − 4µ2 )c′′ (µ) + (λ3 + 6λ2 µ − 8µ3 )c′ (µ) − 3λ2 . Since c′ ( λ2 ) < λ1 , as µ → λ2 +, ψ(µ) < 0. Moreover, since c′′′ (µ) > 0, for all µ > λ2 , we have   2 !   2 ! λ λ λ c′′′ (µ) − 4 µ − (λ2 + 6λµ + 6µ2)c′′ (µ) − 24 µ2 − c′ (µ) < 0. ψ ′ (µ) = −4µ(λ + µ) µ2 − 2 2 2 It follows that ψ(µ) < 0 for all µ > λ2 . Since ϕ′ (µ) has the same sign as ψ(µ), we conclude that ϕ′ (µ) < 0, as desired. Proof of Theorem 12. From Theorem 11, we know that if a symmetric equilibrium exists, then it is unique, and is given by µ⋆ = ϕ−1 (r), where ϕ establishes a one-to-one correspondence between r and µ⋆ (µ⋆ is strictly decreasing in r and vice versa). Therefore, it is enough to show that there exists a finite upper bound µ > λ2 such that no service rate µ > µ can be a symmetric equilibrium under any r-routing policy. It would then automatically follow that for r = ϕ(µ), no r-routing policy with r ≤ r admits a symmetric equilibrium. We prove this by exhibiting a µ and showing that if µ ≥ µ, then the utility function of server 1, U1r (µ1 , µ), cannot attain a global maximum at µ1 = µ for any r ∈ R. We begin by establishing a lower bound for the maximum utility U1r (µ1 , µ) that server 1 can obtain under any r-routing policy:           λ λ λ λ λ λ r r r r , µ = I1 ,µ −c , ≥ −c = U1 . (EC.29) max U1 (µ1 , µ) ≥ U1 2 2 2 2 2 2 µ1 > λ 2 By definition, if µ⋆ is a symmetric equilibrium under any r-routing policy, then the utility function of server 1, U1r (µ1 , µ⋆ ), is maximized at µ1 = µ⋆ , and hence, using (EC.29), we have λ λ U1r (µ⋆ , µ⋆ ) ≥ U1r ( , ). 2 2 (EC.30) Next, we establish some properties on U1r (µ, µ) that help us translate this necessary condition for a symmetric equilibrium into an upper bound on any symmetric equilibrium service rate. We have, U1r (µ, µ) = 1 − which has the following properties: λ − c(µ), 2µ e-companion to Gopalakrishnan, Doroudi, Ward, and Wierman: Routing and Staffing when Servers are Strategic ec21 • Since c′ ( λ2 ) < λ1 , U1r (µ, µ), as a function of µ, is strictly increasing at µ = λ2 . • U1r (µ, µ) is a concave function of µ. This means that U1r (µ, µ) is strictly increasing at µ = λ2 , attains a maximum at the unique µ† > λ2 that solves the first order condition µ2† c′ (µ† ) = λ2 , and then decreases forever. This shape of the curve U1r (µ, µ) implies that there must exist a unique µ > µ† , such that U1r (µ, µ) = U1r ( λ2 , λ2 ). Since U1r (µ, µ) is a strictly decreasing function for µ > µ† , it follows that if µ⋆ > µ, then, r U1 (µ⋆ , µ⋆ ) < U1r (µ, µ) = U1r ( λ2 , λ2 ), contradicting the necessary condition (EC.30). This establishes the required upper bound µ on any symmetric equilibrium service rate, completing the proof. Proof of Theorem 13. A useful tool for proving this theorem is Theorem 3 from [13], whose statement we have adapted to our model: Theorem EC.1. A symmetric game with a nonempty, convex, and compact strategy space, and utility functions that are continuous and quasiconcave has a symmetric (pure-strategy) equilibrium. We begin by verifying that our 2-server game meets the qualifying conditions of Theorem EC.1: • Symmetry: First, all servers have the same strategy space of service rates, namely, ( λ2 , ∞). Moreover, since an r-routing policy is symmetric and all servers have the same cost function, their utility functions are symmetric as well. Hence, our 2-server game is indeed symmetric. • Strategy space: The strategy space ( λ2 , ∞) is nonempty and convex, but not compact, as required by Theorem EC.1. Hence, for the time being, we modify the strategy space to be [ λ2 , µ + 1] so that it is compact, where µ is the upper bound on any symmetric equilibrium, established in Theorem 12, and deal with the implications of this modification later. • Utility function: U1r (µ1 , µ2 ) is clearly continuous. From Mathematica, it can be verified that the idle time function I1r (µ1 , µ2 ) is concave in µ1 for r ∈ {−2, −1, 0, 1}, and since the cost function is convex, this means the utility functions are also concave. (Unfortunately, we could not get Mathematica to verify concavity for non-integral values of r, though we strongly suspect that it is so for the entire interval [−2, 1].) Therefore, we can apply Theorem EC.1 to infer that an r-routing policy with r ∈ {−2, −1, 0, 1} admits a symmetric equilibrium in [ λ2 , µ+1]. We now show that the boundaries cannot be symmetric equilibria. We already know from Theorem 12 that µ + 1 cannot be a symmetric equilibrium. (We could have chosen to close the interval at any µ > µ. The choice µ + 1 was arbitrary.) To see that λ cannot be a symmetric equilibrium, observe that c′ ( λ2 ) < λ1 implies that U1r (µ1 , λ2 ) is increasing 2 at µ1 = λ2 (using the derivative of the idle time computed in (EC.28)), and hence server 1 would have an incentive to deviate. Therefore, any symmetric equilibrium must be an interior point, and from Theorem 11, such an equilibrium must be unique. This completes the proof. 

Analyzing Adaptive Cache Replacement Strategies Mario E. Consuegra2 , Wendy A. Martinez1 , Giri Narasimhan1 , Raju Rangaswami1 , Leo Shao1 , and Giuseppe Vietri1 arXiv:1503.07624v2 [cs.DS] 24 Apr 2017 1 School of Computing and Information Sciences, Florida International University, Miami, FL 33199, USA. {walem001,giri,raju,gviet001}@fiu.edu 2 Google Inc., Kirkland, WA, USA. April 25, 2017 Abstract Adaptive Replacement Cache (Arc) and CLOCK with Adaptive Replacement (Car) are state-of-theart “adaptive” cache replacement algorithms invented to improve on the shortcomings of classical cache replacement policies such as Lru, Lfu and Clock. By separating out items that have been accessed only once and items that have been accessed more frequently, both Arc and Car are able to control the harmful effect of single-access items flooding the cache and pushing out more frequently accessed items. Both Arc and Car have been shown to outperform their classical and popular counterparts in practice. Both algorithms are complex, yet popular. Even though they can be treated as online algorithms with an “adaptive” twist, a theoretical proof of the competitiveness of Arc and Car remained unsolved for over a decade. We show that the competitiveness ratio of Car (and Arc) has a lower bound of N + 1 (where N is the size of the cache) and an upper bound of 18N (4N for Arc). If the size of cache offered to Arc or Car is larger than the one provided to Opt, then we show improved competitiveness ratios. The important implication of the above results are that no “pathological” worst-case request sequences exist that could deteriorate the performance of Arc and Car by more than a constant factor as compared to Lru. 1 Introduction Megiddo and Modha [MM03,MM04] engineered an amazing cache replacement algorithm that was self-tuning and called it Adaptive Replacement Cache or Arc. Later, Bansal and Modha [BM04] designed another algorithm called Clock with Adaptive Replacement (Car). Extensive experimentation suggested that Arc and Car showed substantial improvements over previously known cache replacement algorithms, including the well-known Least Recently Used or Lru and Clock. On the theoretical side, the seminal work of Sleator and Tarjan [ST85] showed that Lru can be analyzed using the theory of online algorithms. They showed that Lru has a competitiveness ratio of N (where N is the size of the cache). More surprisingly, they also showed that with no prefetching, no online algorithm for cache replacement could achieve a competitiveness ratio less than N , suggesting that under this measure, Lru is optimal. In other words, there exist worst-case request sequences that would prevent any algorithm from being better than N -competitive. While these results are significant, they highlight the difference between theory and practice. Sleator and Tarjan’s techniques analyze online algorithms in terms of their worst-case behavior (i.e., over all possible inputs), which means that other algorithms with poorer competitiveness ratios could perform better in practice. Another way to state this is that the results assume an oblivious adversary who designs inputs for the online algorithms in a way that make them perform as poorly as possible. The upper bound on performance ratio merely guarantees that no surprises are in store, i.e., there is no input designed by an adversary that can make the algorithm perform poorly. 1 Given a fixed size cache, the cache replacement problem is that of deciding which data item to evict from the cache in order to make room for a newly requested data item with the objective of maximizing cache hits in the future. The cache replacement problem has been referred to as a fundamental and practically important online problem in computer science (see Irani and Karlin [Hoc97], Chapter 13) and a “fundamental metaphor in modern computing” [MM04]. The Lru algorithm was considered the most optimal page replacement policy for a long time, but it had the drawback of not being “scan-resistant”, i.e., items used only once could pollute the cache and diminish its performance. Furthermore, Lru is difficult to implement efficiently, since moving an accessed item to the front of the queue is an expensive operation, first requiring locating the item, and then requiring data moves that could lead to unacceptable cache contention if it is to be implemented consistently and correctly. The Clock algorithm was invented by Frank Corbató in 1968 as an efficient one-bit approximation to Lru with minimum overhead [Cor68] and continues to be used in MVS, Unix, Linux, and Windows operating systems [Fri99]. Like Lru, Clock is also not scan-resistant because it puts too much emphasis on “recency” of access and pays no attention to “frequency” of access. So there are sequences in which many other algorithms can have significantly less cost than the theoretically optimal Lru. Since then, many other cache replacement strategies have been developed and have been showed to be better than Lru in practice. These are discussed below in Section 2. An important development in this area was the invention of adaptive algorithms. While regular “online” algorithms are usually designed to respond to input requests in an optimal manner, these selftuning algorithms are capable of adapting to changes in the request sequence caused by changes in the workloads. Megiddo and Modha’s Arc [MM03] is a self-tuning algorithm that is a hybrid of Lfu and Lru. Bansal and Modha’s Car is an adaptivehybrid of Lfu and Clock [BM04]. Experiments show that Arc and Car outperform Lru and Clock for many benchmark data sets [BM04]. Versions of Arc have been deployed in commercial systems such as the IBM DS6000/DS8000, Sun Microsystems’s ZFS, and in PostgreSQL. Unfortunately, no theoretical analysis of the adaptive algorithms, Arc and Car, exist in the literature. The main open question that remained unanswered was whether or not there exist “pathological” request sequences that could force Arc or Car to perform poorly. In this document we show that these two algorithms are O(N )-competitive, suggesting that they are not much worse than the optimal Lru. We also prove a surprising lower bound on the competitiveness that is larger than N . The main contributions of this paper are as follows: 1. For completeness, we provide proofs that Lru and Clock are N -competitive. 2. We prove a lower bound on the competitiveness of Arc and Car of N + 1, proving that there are request sequences where they cannot outperform Lru and Clock. 3. We show that Arc is 4N -competitive. 4. We show that Car is 18N -competitive. 5. We obtain precise bounds for the competitiveness of all four algorithms when the sizes of the caches maintained by them are different from that maintained by Opt. 6. We show that if the size of the cache is twice that of the one allowed for the optimal offline algorithm, then the competitiveness ratio drops to a small constant. We use the method of potential functions to analyze the algorithms. However, the main challenges in solving these problems is that of carefully designing the potential function for the analysis. We discuss the role of the adaptive parameter in the potential function. The contributions of this paper are summarized in Table 1. In this table, N is the size of the cache maintained by the algorithm, while NO is the size of the cache maintained by Opt. The table provides lower bounds (LB) and upper bounds (UB) on the competitiveness ratio when the cache sizes are equal, i.e., N = NO ; it also provides upper bounds when they are not equal. 2 Algorithm Lru Arc Clock Car Compet. Ratio LB N N +1 N N +1 Compet. Ratio UB N 4N N 18N Compet. Ratio UB w/ Unequal Sizes N/(N − NO + 1) 12N/(N − NO + 1) N/(N − NO + 1) 18N/(N − NO + 1) [Ref] [ST85] This paper This paper This paper Table 1: Summary of Results After providing relevant background on cache replacement algorithms in Section 2, we discuss the lower bounds on the competitiveness ratios of Arc and Car in Section 3. Next we prove upper bounds on the competitiveness ratios of Lru, Clock, Arc, and Car in Section 4. Concluding remarks can be found in Section 5. 2 Previous Work on Cache Replacement Strategies Below we give brief descriptions of the four algorithms being discussed in this paper, after which we mention a large collection of other closely related cache replacement algorithms. The Lru Algorithm: Lru evicts the least recently used entry. It tends to perform well when there are many items that are requested more than once in a relatively short period of time, and performs poorly on “scans”. Lru is expensive to implement because it requires a queue with move-to-front operations whenever a page is requested. The Clock Algorithm: On the other hand, Clock was designed as an efficient approximation of Lru, which it achieves by avoiding the move-to-front operation. Clock’s cache is organized as a single “circular” list, instead of a queue. The algorithm maintains a pointer to the “head” of the list. The item immediately counterclockwise to it is the “tail” of the list. Each item is associated with a “mark” bit. Some of the pages in the cache are marked, and the rest are unmarked. When a page hit occurs that page is marked, but the contents of the cache remain unchanged. When a page fault occurs, in order to make room for the requested page, the head page is evicted if the page is unmarked. If the head page is marked, the page is unmarked and the head is moved forward clockwise, making the previous head as the tail of the list. After a page is evicted, the requested page is unmarked and placed at the tail of the list. Clock is inexpensive to implement, but is not scan-resistant like Lru. The Arc Algorithm To facilitate our discussion, we briefly describe the Arc algorithm. As mentioned before, it combines ideas of recency and frequency. Arc’s cache is organized into a “main” part (of size N ) and a “history” part (of size N ). The main part is further divided into two lists, T1 and T2 , both maintained as LRU lists (i.e., sorted by “recency”). T1 focuses on “recency” because it contains pages with short-term utility. Consequently, when an item is accessed for the first time from the disk, it is brought into T1 . Items “graduate” to T2 when they are accessed more than once. Thus, T2 deals with “frequency” and stores items with potential for long-term utility. Additionally, Arc maintains a history of N more items, consisting of B1 , i.e., items that have been recently deleted from T1 , and B2 , i.e., items that have been recently deleted from T2 . History lists are also organized in the order of recency of access. The unique feature of Arc is its self-tuning capability, which makes it scan-resistant. Based on a self-tuning parameter, p, the size of T1 may grow or shrink relative to the size of T2 . The details of the algorithm are fairly complex and non-intuitive. Detailed pseudocode for Arc (Figure 4 from [MM03]) is provided in the Appendix for convenience. It is worth noting that Arc is considered a “universal” algorithm in the sense that it does not use any a priori knowledge of its input, nor does it do any offline tuning. Furthermore, Arc is continuously adapting, since adaptation can potentially happen at every step. 3 It must be noted that our results on Arc assume the “learning rate”, δ, to be equal to 1, while the Arc algorithm as presented by Megiddo and Modha recommended a “faster” learning rate based on experiments on real data. The learning rate is the rate at which the adaptive parameter p is changed as and when needed. The Car Algorithm Inspired by Arc, Car’s cache is organized into two main lists, T1 and T2 , and two history lists, B1 and B2 . Inspired by Clock, both T1 and T2 are organized as “circular” lists, with each item associated with a mark bit. The history lists, B1 and B2 are maintained as simple FIFO lists. We let t1 , t2 , b1 , b2 denote the sizes of T1 , T2 , B1 , B2 , respectively. Also, let t := t1 + t2 . Let lists L1 (and L2 , resp.) be the list of size ℓ1 (ℓ2 , resp.) obtained by concatenating list B1 to the end of “linearized” T1 (concatenating B2 to the tail of T2 , resp.). Note that circular lists are linearized from head to tail. We let T10 and T20 (T11 and T21 , resp.) denote the sequence of unmarked (marked, resp.) pages in T1 and T2 , respectively. The following invariants are maintained by Car for the lists: 1. 0 ≤ t1 + t2 ≤ N 2. 0 ≤ ℓ1 = t1 + b1 ≤ N 3. 0 ≤ ℓ1 + ℓ2 = t1 + t2 + b1 + b2 ≤ 2N 4. t1 + t2 < N =⇒ b1 + b2 = 0 5. t1 + t2 + b1 + b2 ≥ N =⇒ t1 + t2 = N 6. Once t1 + t2 = N and/or ℓ1 + ℓ2 = 2N , they remain true from that point onwards. Car maintains an adaptive parameter p, which it uses as a target for t1 , the size of list T1 . Consequently, N − p is the target for t2 . Using this guiding principle, it decides whether to evict an item from T1 or T2 in the event that a miss requires one of the pages to be replaced. The replacement policy can be summarized into two main points: 1. If the number of items in T1 (barring the marked items at the head of the list) exceeds the target size, p, then evict an unmarked page from T1 , else evict an unmarked page from T2 . 2. If ℓ1 = t1 + b1 = N , then evict a history page from B1 , else evict a history page from B2 . Since the details of the algorithm are complex, the actual pseudocode is provided (Figure 2 from [BM04]) in the Appendix. Other Cache Replacement Algorithms The DuelingClock algorithm [JIPP10] is like Clock but keeps the clock hand at the newest page rather than the oldest one, which allows it to be scan-resistant. More recent algorithms try to improve over Lru by implementing multiple cache levels and leveraging history. In [OOW93] the Lru-K algorithm was introduced. Briefly, the Lru-K algorithm estimates interarrival times from observed requests, and favors retaining pages with shorter interarrival times. Experiments have shown Lru-2 performs better than Lru, and that Lru-K does not show increase in performance over Lru2 [OOW93], but has a higher implementation overhead. It was also argued that Lru-K is optimal under the independence reference model (IRM) among all algorithms A that have limited knowledge of the K most recent references to a page and no knowledge of the future [OOW93]. In essence, the Lru-K algorithm tries to efficiently approximate Least Frequently Used (Lfu) cache replacement algorithm. As K becomes larger, it gets closer and closer to Lfu. It has been argued that Lfu cannot adapt well to changing workloads because it may replace currently “hot” blocks instead of “cold” blocks that had been “hot” in the past. Lfu is implemented as a heap and takes O(log N ) time per request. Another cache replacement algorithm is Lirs [JZ02]. The Lirs algorithm evicts the page with the largest IRR (inter-reference recency). It attempts to keep a small (≈ 1%) portion of the cache for HIR (high inter-reference) pages, and a large (≈ 99%) portion of the cache for LIR (low inter-reference) pages. The Clock-Pro algorithm approximates Lirs efficiently using Clock [JCZ05]. The 2q [JS94] algorithm 4 is scan-resistant. It keeps a FIFO buffer A1 of pages that have been accessed once and a main Lru buffer Am of pages accessed more than once. 2q admits only hot pages to the main buffer. The buffer A1 is divided into a main component that keeps the pages in A1 that still reside in cache, and a history component that remembers pages that have been evicted after one access. The relative sizes of the main and history components are tunable parameters. 2q has time complexity of O(1). Another algorithm that tries to bridge the gap between recency and frequency is Lrfu [LCK+ 01]. This is a hybrid of Lru and Lfu and is adaptive to changes in workload. The time complexity ranges from O(1) for Lru to O(log n) for Lfu. 3 Lower Bounds on Competitiveness Ratio for Arc and Car This section presents our results on the lower bounds for Arc and Car. We also show that the adaptive parameter is critical to both Arc and Car by showing that their non-adaptive versions have an unbounded competitiveness ratio. 3.1 Lower Bound for Arc First, we show a lower bound on the competitiveness ratio for Arc. Theorem 1. The competitiveness ratio of Algorithm Arc has a lower bound of N + 1. Proof. We show that we can generate an unbounded request sequence that causes N + 1 page faults on Arc for every page fault on Opt. The sequence only involves 2N + 1 pages denoted by 1, . . . , 2N + 1. Our example, will take the contents of the cache managed by Arc from configurations 1 through configuration 5, which are shown in Table 2. Note that configuration 1 and configuration 5 are essentially the same to the extent that the value of p is 0 in both, and the number of pages in each of the four parts of the cache are identical. Configuration 1 2 3 4 5 p 0 0 0 1 0 T1 ∅ 2N + 1 ∅ ∅ ∅ T2 1, . . . , N 2, . . . , N 2, . . . , N, 1 3, . . . , N, 1, 2N + 1 1, 2N + 1, 2, . . . , N − 1 B1 ∅ ∅ 2N + 1 ∅ ∅ B2 N + 1, . . . , 2N N + 2, . . . , 2N, 1 N + 2, . . . , 2N N + 2, . . . , 2N, 2 N + 2, . . . , 2N, N Table 2: Example for Lower Bound on Arc’s competitiveness We note that we can obtain configuration 1 from an empty cache with the following request sequence: 2N, 2N, 2N − 1, 2N − 1, . . . , 2, 2, 1, 1. Consider the first half of the above request sequence, which contains a total of 4N requests to 2N new pages, each page requested twice in succession. The first time a page is requested from the first N new pages, it will be put in T1 . The second time the page is requested, it will get moved to T2 . In the second half, if a page not in Arc is requested, Replace will be called, which will move a page from T2 to B2 , and the new page will be placed in T1 . When the same page is requested again, it simply gets moved to T2 . The value of p remains unchanged in this process. It is clear that we get Configuration 1 as a result of the request sequence. We design our request sequence by following the steps below. 1. Make one request to a page 2N + 1 not in Arc. We will assume that this is a brand new page and therefore also causes a fault for Opt and for Arc. The page 2N + 1 will go into T1 and a page in T2 will be demoted to B2 . The contents of Arc is given by Configuration 2 in Table 2. 2. Request any page in B2 . This decreases the value of p but since p is zero it will remain unchanged. Since the size of T1 is more than p Arc will call Replace, which will act on T1 , hence 2N + 1 will be demoted to B1 . Upon requesting page 1 in B2 , we get Configuration 3 in Table 2. 5 3. The next step is to request 2N + 1 again, which will move to T2 , p is increased and a page in T2 is demoted to B2 . Configuration 4 reflects the contents of the cache at this stage. 4. Finally we make N − 2 requests to any pages from B2 . By requesting the pages 2, 3, . . . , N , we end up in Configuration 5 from Table 2. The steps outlined above cause N + 1 page faults for Arc and at most one page fault for Opt. Since we are back to the initial configuration we can repeat this process over again. This concludes the proof that the competitiveness ratio of Arc is at least N + 1. 3.2 Lower Bound for Car Now we prove a similar lower bound for Car. Theorem 2. The competitiveness ratio of Algorithm Car has a lower bound of N + 1. Proof. We show that we can generate an infinite request sequence that causes N + 1 page faults in Car for every page fault on Opt. The sequence only involves 2N + 1 pages denoted by 1, . . . , 2N + 1. Our example, will take the contents of the cache managed by Car from configurations 1 through N + 2 as shown in Table 3. Note that a superscript of 1 on any page in T1 ∪ T2 indicates that it is marked. All others are unmarked. Also note that configuration 1 and configuration N + 2 are essentially the same upon relabeling. First, we show that configuration 1 is attainable, by showing that it can be reached from an empty cache. This is formalized in the following lemma. Lemma 1. We can obtain configuration 1 starting from an empty cache with the following request sequence: 2N, 2N, 2N − 1, 2N − 1, . . . , 2, 2, 1, 1. Proof. The first half of the above request sequence calls each of the N pages 2N, 2N − 1, . . . , N + 1 twice in succession. The first time they are called, they are moved into T1 unmarked. The second time the same page is called it gets marked, but remains in T1 . At the end of the first half, all the N pages requested end up in T1 and are all marked. The next call to new page N , will trigger a call to Replace, which will move all the marked pages in T1 to T2 leaving them unmarked. It will also move one page from T2 to B2 . Finally, the requested page N will be moved to T1 and left unmarked. When requested again, it simply gets marked. When the next page, i.e., N − 1 is requested, it moves marked page N to T2 , moves one more page from T2 to B2 . As the rest of the pages from the request sequences are requested, the previous requested page gets moved to T2 , which in turn demotes one of its pages to B2 . At the end of the process, we get a marked page 1 in T1 . Pages 2, . . . , N are in T2 , unmarked, and pages N + 1, . . . , 2N end up in B2 . This is exactly what we need for configuration 1. Continuing on the proof of Theorem 2, we show what happens when, starting from configuration 1, Car processes the following request sequence. Page 2N + 1: A page in T2 is demoted to B2 , which loses a page; the marked page from T1 is moved to T2 and the new page is moved into T1 . MRU page in B2 : This should have decremented p but remains unchanged since it is already zero. Since the size of T1 is more than p Car will call Replace and 2N + 1 will be demoted to B1 , resulting in configuration 3 in Table 3. Page 2N + 1: It will move to T2 , p is increased and a page in T2 is demoted to B2 . See configuration 4 in Table 3. MRU page from B2 , repeat N − 2 times: It results in configuration N + 2 in Table 3. 6 Config. 1 2 3 4 5 ... N +2 p 0 0 0 1 0 0 0 B1 ∅ ∅ 2N + 1 ∅ ∅ ... ∅ T1 11 2N + 1 ∅ ∅ ∅ ... ∅ T2 2, . . . , N 1, . . . , N − 1 N, 1, . . . , N − 1 2N + 1, N, 1, . . . , N − 2 N − 1, 2N + 1, N, 1, . . . , N − 3 ... 2, . . . , N − 1, 2N + 1, N B2 N + 1, . . . , 2N N, . . . , 2N − 1 N + 1, . . . , 2N − 1 N − 1, N + 1, . . . , 2N − 1 N − 2, N + 1, . . . , 2N − 1 ... 1, N + 1, . . . , 2N − 1 Table 3: Example for Lower Bound on Car’s competitiveness The request sequence detailed above generates N + 1 faults for Car while only N different pages are requested. Thus, Opt could limit itself to at most one fault in this stretch. Opt will fault once during each stretch if the next page is picked to be one that is farthest used in the future. Repeating the above steps an unbounded number of times with appropriate relabeling proves that the competitiveness ratio of Car is lower bounded by N + 1. 3.3 Non-Adaptive Arc and Car are not Competitive It is particularly interesting to note that the non-adaptive version of Car and Arc (called Fixed Replacement cache) [MM03] are not competitive. The following two theorems prove that the competitiveness ratios can be unbounded in this case. Theorem 3. Algorithm Car with fixed p is not competitive. Proof. Suppose that algorithm Car has p fixed instead of being adaptive and 0 < p < N − 1. Recall that p is the target size of T1 and N − p is the target size of T2 . We design a request sequence such that with less than N pages we can generate an infinite number of page faults for Car. The sequence is described as follows: Step 1: Fill up T2 with N − p unmarked pages as described above in the proof of Theorem 2. Step 2: Request the MRU page in B2 . The requested page goes to the tail of T2 as an unmarked page. Since the size of T2 is greater than p we discard the head of T2 . Step 3: Request the MRU page in B2 which is actually the page discarded in Step 2 from T2 . This step is similar to Step 2 and we can continue to repeat this infinitely often, since the page that moves from B2 to T2 get’s unmarked and the page that moves from T2 to B2 goes to MRU. Therefore, we can cycle infinitely many times through N − p + 1 pages triggering an infinite number of faults, while Opt can avoid faults altogether during the cycle. Theorem 4. Algorithm Arc with fixed p is not competitive. Proof. Suppose that algorithm Arc has p fixed instead of being adaptive and 0 < p < N . Recall that p is the target size of T1 and N − p is the target size of T2 . We design a request sequence such that with less than N pages we can generate an infinite number of page faults for Arc. The first step is to fill up T2 (size of T2 = N − p). Next we request the MRU page in B2 . Every time we request a page from B2 , it goes into the top of T2 and thus it increases the size of T2 beyond its target size. It follows that Arc will call Replace and move a page from T2 to the MRU position in B2 . If the MRU page from B2 is repeatedly requested, we will cycle through N − p pages, every time incurring a page fault for Arc, while Opt can avoid faults altogether during the cycle. 7 4 4.1 Analyzing Lru using potential functions The generic approach The standard approach used here is as follows. First, we define a carefully crafted potential function, Φ. As per the strategy of analyzing competitiveness ratios suggested by Sleator and Tarjan [ST85], we then try to show the following inequality: CA + ∆Φ ≤ f (N ) · CO + g(N ), (1) where CA and CO are the costs incurred by the algorithm and by Opt, respectively, ∆Φ is the change in potential, f (N ) is some function of N , the size of the cache. In all of our proofs, we assume that the work involves simultaneously maintaining Opt’s cache as well as the algorithm’s cache. So we can break down the work into two steps, one where only Opt serves and one where only the algorithm serves. When only Opt serves, there are 2 cases: first when Opt has a hit and next when Opt has a miss. Next, we consider the cases when the algorithm serves, once when it has a hit and once when it has a miss. In each case, our goal is to prove the inequality (1) mentioned above, which establishes that f (N ) is the competitiveness ratio of algorithm A. There may be an additive term of g(N ) which is a function of the misses needed to get to some initial configuration for the cache. 4.2 Analyzing Lru using potential functions Assuming that the size of cache given to the competing Opt algorithm is NO ≤ N , the following result was proved by Sleator and Tarjan [ST85] (Theorem 6) for Lru.
N -competitive. Theorem 5. [ST85] Algorithm Lru is N −N O +1 Here we present a complete proof of this well-known result because we believe it is instructive for the other proofs in this paper. Proof. While this was not used in the proof in Sleator and Tarjan [ST85], a potential function that will facilitate the proof of the above theorem is: Φ= P r(x) , NL − NO + 1 x∈D (2) where D is the list of items in Lru’s cache but not in Opt’s cache, and r(x) is the rank of item x in Lru’s list with the understanding that the LRU item has rank 1, while the MRU item has rank equal to the size of the cache [Alb96]. We now show the following inequality: CA + ∆Φ ≤
N · CO + O(N ), N − NO + 1 (3) where CA and CO are the costs incurred by the algorithm and by Opt, respectively, ∆Φ is the change in potential, f (N ) is some function of N , the size of the cache. We assume that the work involves simultaneously maintaining Opt’s cache as well as Lru’s cache. So we can break down the work of Lru into two steps, one where only Opt serves and one where only Lru serves. When only Opt serves, there are 2 cases: first when Opt has a hit and next when Opt has a miss. In either case, the cost for Lru is 0, since only Opt is serving. When Opt has a hit, the cost for Opt is also 0. Furthermore, since Lru’s cache remains untouched, and no changes take place in the contents of Opt’s cache, the ranks of the items in Lru remain unchanged. Thus, ∆Φ = 0. Therefore, the inequality in (3) is trivially satisfied in this case. 8 When Opt has a miss, CO = 1, as before. The item evicted by Opt can contribute the rank of that item NL to increase at most by NL , making the increase in potential function to be bounded by NL −N . Thus, O +1 the inequality in (3) is satisfied. Next, we consider the step where Lru serves the request. As with Opt, when Lru is serving, the cost for Opt is 0. We again consider two cases: first when Lru has a hit and next when Lru has a miss. When Lru has a hit, the cost for Lru is 0. The contents of Lru’s cache may change. The item that was accessed is moved to the MRU position. However, this item is already in Opt’s cache and therefore cannot contribute to a change in potential. Several other items may move down in the cache, thus contributing to a decrease in potential of at most (N − 1). In the worst, case the increase in potential is at most 0. Therefore, the inequality in (3) is again satisfied. Finally, we consider the case when Lru has a miss. As before, CL = 1. Following the previous arguments, an item would be brought into MRU (which is already present in Opt’s cache), a bunch of items may be demoted in rank, and the Lru item will be evicted. The only action that can contribute to an increase is caused by the item that is brought into the MRU location. However, this item is already present in Opt’s cache, and hence cannot contribute to an increase. All the demotions and eviction can only decrease the potential function. Note that before the missed item is brought into Lru’s cache, the contents of Lru’s and Opt’s cache agree in at most NO − 1 items, since Opt just finished serving the request and the item that caused the miss is already in Opt’s cache. Thus there are at least NL − NO + 1 items that contribute their ranks to the potential function. These items either get demoted in rank or get evicted. Either way, the potential function will reduce by a minimum value of NL − NO + 1, although it could more if there are more items that are in Lru and that are not in Opt’s cache. Thus the total change in potential has to be at most NL − NO + 1, and we have CL + ∆Φ ≤ 1 − (NL − NO + 1) NL ≤0= · CO . (NL − NO + 1) NL − NO + 1 Summarizing the costs, we have the following: Step CL ∆Φ Opt Serves Request Opt has a hit 0 0 Opt has a miss 0 ≤ NL Lru Serves Request Lru has a hit 0 ≤0 Lru has a miss 1 ≤ NL − NO + 1 CO 0 1 0 0 The analysis of Lru states that if the sizes of Lru’s and Opt’s caches are NL and NO respectively, and NL if NL ≥ NO , then the competitiveness ratio of Lru is NL −N . Thus Lru is 2-competitive if the size of O +1 Lru’s cache is roughly twice that of Opt’s cache. 4.3 Analyzing the competitiveness of Clock Our result on the competitiveness of Clock is formalized in the following theorem. While this result appears to be known, we have not been able to locate a full proof and we believe this is of value. We therefore present it for the sake of completeness.
N Theorem 6. Algorithm Clock is N −N -competitive. O +1 Proof. Let M0 denote the subsequence of unmarked pages in Clock, ordered counterclockwise from head to tail. Let M1 denote the subsequence of marked pages in Clock, ordered counterclockwise from head to tail. Let q be any page in Clock’s cache. Let P 0 [q] denote the position of an unmarked page q in the 9 ordered sequence M0 , and let P 1 [q] denote the position of a marked page q in M1 . Finally, let R[q] denote the rank of page q defined as follows: R[q] = ( P 0 [q] if q is unmarked, P 1 [q] + |M0 | otherwise. (4) Thus, if q is an unmarked page at the head, then R[q] = 1. By the above definition, the following lemmas are obvious. Lemma 2. If q is any page in Clock’s cache, then R[q] ≤ N . Lemma 3. If a marked page q at the head of Clock’s cache is unmarked and moved to the tail, then R[q] does not change in the process. Let D be the set of pages that are in the cache maintained by Clock, but not in the cache maintained by Opt. We define the potential function as follows: Φ= X R[q] (5) q∈D We prove one more useful lemma about the ranks as defined above. Lemma 4. If an unmarked page at the head of Clock’s cache is evicted from Clock’s cache, and if there is at least one page in D, then Φ decreases by at least 1 in the process. . Proof. All pages, marked or unmarked, will move down by at least one position (reducing the rank of each by at least 1). The decrease in potential for at least one page that is in D will contribute to Φ, guaranteeing that ∆Φ ≤ −1. Let CClock and COpt be the costs incurred by the algorithms Clock and Opt, and let S = σ1 , σ2 , . . . , σm be an arbitrary request sequence. Let S ′ denote the initial subsequence of requests that take place prior to the cache becoming full. Note that exactly N faults are incurred in S ′ , after which the cache remains full. Let S ′′ be the subsequence of S that comes after S ′ . Let CClock and COpt be the cost incurred by the algorithms Clock and Opt respectively. We will prove that for every individual request, σ ∈ S ′′ : CClock (σ) + ∆Φ ≤ N ∗ COpt (σ) (6) As before, we assume that request σ is processed in two distinct steps: first when Opt services the page request and, next when Clock services the request. We will show that inequality (6) is satisfied for both the steps. When only Opt acts in this step, Cclock = 0. If Opt does not fault on this request, then COP T = 0. No change occurs to the contents of the cache maintained by Opt as well as Clock, and the clock hand does not move. Thus, ∆Φ = 0, satisfying inequality 6. If Opt faults on request σ, then COP T = 1 and CClock = 0. The contents of the cache maintained by Opt does change, which could affect the potential function. The potential could increase due to the eviction of a page in Opt. Since by Lemma 2 the rank of the evicted page cannot exceed N , the potential will change by at most N . Thus, inequality 6 is satisfied. Next we consider what happens when Clock services the request. For this case COP T = 0. If Clock does not fault, then Cclock = 0 and the requested page may change from an unmarked status to a marked one. However, since the page is already in the cache maintained by Opt it is not in D and is therefore not considered for the potential function calculations in 5. Thus, inequality 6 is satisfied. 10 Finally, we consider the case when Clock faults, in which case CClock = 1 and COpt = 0. To satisfy inequality 6, ∆Φ needs to be less or equal to -1. When Clock has a miss, if the head page happens to be marked, then Clock will repeatedly unmark the marked head page, moving it to the tail position, until an unmarked head page is encountered. The unmarked head page is then evicted. Each time a marked head page becomes an unmarked tail page, by Lemma 3 its rank does not change. When finally an unmarked head page is evicted, we know that there is at least one page in Opt’s cache that is not in Clock’s cache (i.e., the page that caused the fault). Since there are N pages in the cache maintained by Clock, at least one of those pages is guaranteed not to be part of the cache maintained by Opt. Since there is at least one page in D, by Lemma 4 it is clear that evicting an unmarked head page will decrease the potential function by at least one, which will pay for the Clock’s page fault. We have therefore showed that for every request σ, inequality 6 is satisfied. Since there can be at most N faults for the requests in S ′ , summing up the above inequality for all requests, σ ∈ S, we get CClock (S) ≤ N ∗ COpt (S) + N. This completes the proof of the theorem and the competitiveness analysis of the Clock algorithm. 4.4 Analyzing the Competitiveness of ARC In this paper, we prove two different upper bounds for the competitiveness of Arc. These two proofs use very different potential function. The first one allows for the sizes of the caches maintained by Arc and Opt to be different, while the second one does not allow for it, but provides a tighter bound. We provide both results below. Our first result on the competitiveness of Arc is formalized in the following theorem:
12N Theorem 7. Algorithm Arc is N −N -competitive. O +1 Proof. Let PX [q] be the position of page q in an arbitrary ordered sequence of pages X. When the set is obvious, we will drop the subscript and denote PX [q] simply by P [q]. The set of history pages T1 , T2 , B1 , and B2 will be treated as an ordered sequence of pages ordered from its LRU position to its MRU position. Let Opt and Car be the set of main pages stored in the caches for algorithms Opt and Arc respectively. Let D = Arc \ Opt. As before, we associate each page with a rank value R[q], which is defined as follows:  2PB1 [q]    2P [q] B2 R[q] =  4P T1 [q] + 2b1    4PT2 [q] + 2b2 if if if if q q q q ∈ B1 ∈ B2 ∈ T1 ∈ T2 (7) Finally, we define the potential function as follows: Φ = p + 2t1 + 2 P R[q]
− 3|Arc| N − NO + 1 q∈D (8) The initial value of Φ is 0. If the following inequality (9) is true for any request σ, where ∆Φ is the change in potential caused by serving the request, then when summed over all requests, it proves Theorem 7. CArc (σ) + ∆Φ ≤ 12N COpt (σ) . N − NO + 1 (9) As before, we assume that request σ is processed in two distinct steps: first when Opt serves and, next when Arc serves. We will show that inequality (9) is satisfied for each of the two steps. 11 Step 1: Opt serves request σ Since only Opt acts in this step, CArc = 0, and T1 ∪ T2 does not change. There are two possible cases: either Opt faults on σ or it does not. If Opt does not fault on this request, then it is easy to see that COpt = 0 and ∆Φ = 0, thus satisfying inequality (9). If Opt faults on request σ, then COpt = 1 and some page q, is evicted from the cache maintained by Opt will belong to D after this step and thus its rank will contribute to the potential function, which will increase by two times the rank of q. The maximal positive change in potential will occur when q is the MRU page of either T1 or T2 . In this case the rank of q is given by: R[q] = 4P [q] + b1 (R[q] = 4P [q] + b2 ). The maximum possible values for each of the terms P [q] and b1 will be N , hence the maximum possible rank of 12N q will be 4N + 2N = 6N . Therefore resulting potential change is at most N −N . O +1 Step 2: Arc serves request σ We break down the analysis into four cases. Case 2.1 deals with the case when Arc finds the page in its cache. The other three cases assume that Arc faults on this request because the item is not in T1 ∪T2 . Cases 2.2 and 2.3 assume that the missing page is found recorded in the history in lists B1 and B2 , respectively. Case 2.4 assumes that the missing page is not recorded in history. Case 2.1: Arc has a page hit Clearly, the page was found in T1 ∪ T2 , and CArc = 0. We consider the change of each of terms in the potential function individually. 1. As per the algorithm, p can only change when the page is found in history. (See lines 3 through 10 of Arc(x).) Since the page is not found in Arc’s history, ∆p = 0. 2. If the hit happens in T1 , the page will move to the top of T2 (See line 2 of Arc(x).), which will result in a decrease in t1 . If the hit happens in T2 , the size of t1 will remain the same. The overall change in t1 will be 0. 3. Since Opt has already served the page, the page is in Opt’s cache. Therefore, even if the page’s rank could change when moved from T1 to MRU position of T2 , this rank will not affect the potential since the page is not in D. We, therefore, conclude that ∆Φ = 0, satisfying inequality (9). Next we will analyze the 3 cases when the requested page is not in Arc’s cache. Since CArc = 1, the change in potential must be ≤ −1 in each case in order for inequality (9) to be satisfied. Case 2.2: Arc has a page miss and the missing page is in B1 We consider the two cases, first when Replace moves an item from T1 to B1 and second when it moves an item from T2 to B2 . 1. Case 1: We consider the change in potential function by analyzing each of the 3 terms. • Value of p will either increase by 1 or stay the same in case p = N , we will account for the worst case which is when ∆p = 1. • A new page is being added to MRU of T2 , and Replace is taking the LRU page of T1 to B1 , then 2∆t1 = −2. • The page that moved from B1 to T2 is not in D, therefore the change in its rank will P not affect the potential, the other pages will could only decrease their rank, meaning that 2∆ q∈D R[q] ≤ 0. Since p increases by at most 1 and t1 decreases by at least 2 the total change in potential is at most -1. 2. Case 2: Once again. we consider the change in potential function by analyzing each of the three terms. • Value of p will either increase by 1 or stay the same in case p = N , we will account for the worst case which is when ∆p = 1. 12 • A new page is added to MRU of T2 , and Replace moves the LRU page of T2 to B2 . Thus, there is no change in T1 . • The page that moved from B1 to T2 is not in D, therefore the change in its rank will not affect the potential. Since t1 + t2 = N , it is guaranteed that at least N − NO + 1 pages are not in Opt. For the pages that are in T1 , their ranks will decrease by at least 2 since b1 decreases by 1, and for the pages in T2 their ranks will decrease by at least 2 as well since b2 increasesPby 1 but the LRU R[q] page in T2 will move to B2 , reducing P [q] for all the pages in T2 . The term 2 N q∈D −NO +1 decreases by at least -4. P R[q] Since p increases by at most 1 and 2 N q∈D −NO +1 decreases by at least -4 the total change in potential is at most -3. Case 2.3: Arc has a page miss and the missing page is in B2 When the missing page is in B2 , Arc makes a call to Replace (Line 5) and then executes Lines 18-19. Thus, p is decremented except if it is already equal to 0. We consider two sub cases: ∆p ≤ −1 and ∆p = 0. ∆p ≤ −1: As in Case 2.2, the call to Replace has no effect on t1 . Replace will not increment the rank using a similar analysis as in 2.2 and change in p will at least be -1. The change in the potential function is at most -1. ∆p = 0: Unlike the sub case above when p decreases by 1, the change in p cannot guarantee the required reduction in the potential. We therefore need a tighter argument. We know that there is a call to Replace. Two cases arise and are discussed below. • Replace moves an item in T1 to B1 : Since the LRU page of T1 is moved to the MRU position of B1 , 2∆t1 = −2 and there is no movement of a page in D that could increase the rank. Therefore the total change in the potential function is at most -2. • Replace moves an item in T2 to B2 : p = 0 indicates that T2 has N pages, therefore is guarantee that at least N − NO + 1 pages will not be part of Opt, contributing to the change in potential. The page being moved from T2 to B2 will decrease it’s rank by at least 2, and the rest of the pages in T2 will move down one position (P [q] will decrease by 1) while B2 will remain the same, resulting in a change in the potential function of at most -4. Thus, in each case the potential function decreased by at most -2. Case 2.4: Arc has a page miss and the missing page is not in B1 ∪ B2 1. t1 + b1 = N ; t1 < N ; The LRU page in B1 is evicted. Assume Replace moves a page from T1 to B1 and a new page is brought into T1 (∆t1 = 0, ∆b1 = 0, ∆t2 = 0, ∆b2 = 0). • The term p is not affected. • The term t1 is not affected. • Since t1 + b1 = N , at least N − No + 1 pages in T1 ∪ B1 are not in Opt. If the page is in B1 \ Opt then its rank decreases by 2; if the page is in T1 \ Opt its rank decreases by 4. 2. t1 + b1 = N ; t1 < N ; The LRU page in B1 is evicted. Assume Replace moves a page from T2 to B2 and a new page is brought into T1 (∆t1 = 1, ∆b1 = −1, ∆t2 = 1, ∆b2 = 1). • The term p is not affected. • The term t1 is increased by 1. 13 • Since t1 + t2 = N , at least N − No + 1 pages in T1 ∪ T2 that are not in Opt. If a page, q, is in T1 \ Opt then its rank decreases by 2 (∆R[q] = ∆4 ∗ P [q] + ∆2 ∗ b2 = −2); if the page, q, is in T2 \ Opt its rank decreases by 2 (∆R[q] = ∆4 ∗ P [q] + ∆2 ∗ b2 = −2). 3. t1 + b1 < N ; t1 + t2 + b1 + b2 = 2N ; Assume that the LRU page in B2 is evicted and Replace moves a page from T1 to B1 and a new page is brought into T1 (∆t1 = 0, ∆b1 = 1, ∆t2 = 0, ∆b2 = −1). • The term p is not affected. • The term t1 is not affected. • Here we used the fact that t2 + b2 > N , then at least N − No + 1 pages in T2 ∪B2 are not in Opt. If a page, q, is in T2 \Opt then its rank decreases by 2 (∆R[q] = ∆4∗P [q]+∆2∗b2 = 4∗(0)+2(−1) = −2); if the page, q, is in B2 \ Opt its rank decreases by 2 (∆R[q] = ∆2 ∗ P [q] = 2 ∗ (−1) = −2). 4. t1 + b1 < N ; t1 + t2 + b1 + b2 = 2N ; Assume that the LRU page in B2 is evicted and Replace moves a page from T2 to B2 and a new page is brought into T1 (∆t1 = 1, ∆b1 = 0, ∆t2 = 1, ∆b2 = 0). • The term p is not affected. • The term t1 is increased by 1. • Here we used the fact that t2 + b2 > N , then at least N − No + 1 pages in T2 ∪B2 are not in Opt. If a page, q, is in T2 \Opt then its rank decreases by 2 (∆R[q] = ∆4∗P [q]+∆2∗b2 = 4∗(0)+2(−1) = −2); if the page, q, is in B2 \ Opt its rank decreases by 2 (∆R[q] = ∆2 ∗ P [q] = 2 ∗ (−1) = −2). 5. t1 + b1 < N ; t1 + t2 + b1 + b2 < 2N ; In this case, no pages are evicted from history. Assume that Replace moves a page from T1 to B1 and a new page is brought into T1 (∆t1 = 0, ∆b1 = 1, ∆t2 = 0, ∆b2 = 0) • The term p is not affected. • The term t1 is increased by 1. • Here we cannot say that the rank decreases. Hence the rank term is at most 0. • The term |Arc| increases by 1. 6. t1 + b1 < N ; t1 + t2 + b1 + b2 < 2N ; In this case, no pages are evicted from history. Assume Replace moves a page from T2 to B2 and a new page is brought into T1 (∆t1 = 1, ∆b1 = 0, ∆t2 = −1, ∆b2 = 1) • The term p is not affected. • The term t1 is not affected. • Here we cannot say that the rank decreases. Hence the rank term is at most 0. • The term |Arc| increases by 1. Wrapping up the proof of Theorem 7: Combining the four cases (2.1 through 2.4) proves that inequality (9) is satisfied when Arc serves request σ. This completes the proof of Theorem 7, establishing that the upper bound on the competitiveness of Arc is 12N for the cases where the sizes of Opt and Arc are the same. By analyzing cases where the size of Arc is greater than Opt we can observe that since Arc 12N the greater the size of Arc’s cache relative to the size of Opt’s cache, smaller will be the will be N −N O +1 competitiveness of Arc. 14 4.5 Alternative Analysis of Competitiveness of Arc Below, we prove an improved upper bound on the competitiveness ratio of Arc. As seen below, the potential function is considerably different. Let CA and CO be the costs incurred by the algorithms Arc and Opt. We start with some notation and definitions. If X is the set of pages in a cache, then let M RU (X) and LRU (X) be the most recently and least recently used pages from X. Let M RUk (X) and LRUk (X) be the k most recently and k least recently used pages from X. Let lists L1 (and L2 ) be the lists obtained by concatenating lists T1 and B1 (T2 and B2 , resp.). Let list L be obtained by concatenating lists L1 and L2 . We let ℓ1 , ℓ2 , t1 , t2 , b1 , b2 denote the sizes of L1 , L2 , T1 , T2 , B1 , B2 , respectively. Finally, let t := t1 + t2 and ℓ := ℓ1 + ℓ2 . At any instant of time during the parallel simulation of Opt and Arc, and for any list X, we let M RUk (X) be denoted by T OP (X), where k is the largest integer such that all pages of M RUk (X) are also in the cache maintained by OPT. We let L′1 , L′2 , T1′ , T2′ denote the T OP s of L1 , L2 , T1 , T2 , respectively, with sizes ℓ′1 , ℓ′2 , t′1 , t′2 , respectively. We let b′1 and b′2 denote the sizes of the B1′ = L′1 ∩ B1 and B2′ = L′2 ∩ B2 , respectively. Note that if b′1 > 0 (b′2 > 0, resp.), then all of T1 (T2 , resp.) is in Opt. Finally, we let ℓ′ := ℓ′1 + ℓ′2 . The Arc algorithm ensures that 0 ≤ t ≤ N , 0 ≤ ℓ ≤ 2N and 0 ≤ ℓ1 ≤ N , thus making 0 ≤ ℓ2 ≤ 2N . We assume that algorithm X being analyzed is provided an arbitrary request sequence σ = σ1 , σ2 , . . . , σm . We define the potential function as follows: Φ = p − (b′1 + 2 · t′1 + 3 · b′2 + 4 · t′2 ). (10) The main result of this section is the following theorem: Theorem 8. Algorithm ARC is 4N -competitive. We say that the cache is full if t = N and either t1 + b1 = N or t2 + b2 ≥ N . We will prove the above theorem by proving the following inequality for any request σ that is requested after the cache is full: CA (σ) + ∆Φ ≤ 4N · CO (σ) + 2N, (11) where ∆X represents the change in any quantity X. Summing up the above inequality for all requests would prove the theorem as long as the number of faults prior to the cache becoming full is bounded by the additive term 2N . We make the following useful observation about a full cache. Lemma 5. When the request sequence requests the N -th distinct page, we have t = N , and this remains an invariant from that point onward. No items are discarded from the cache (main or history) until either t1 + b1 = N or ℓ1 + ℓ2 = 2N . By the time the request sequence requests the 2N -th distinct page, we have either t1 + b1 = N or ℓ1 + ℓ2 = 2N . Proof. Once the request sequence requests the N -th distinct page, it is obvious that we will have t = N , since until then, no item is evicted from T1 ∪ T2 ∪ B1 ∪ B2 . (Note that Replace only moves items from the main part to the history, i.e., from T1 ∪ T2 to B1 ∪ B2 .) Also, until then, p does not change. From that point forward, the algorithm never evicts any item from T1 ∪ T2 without replacing it with some other item. Thus, t = N is an invariant once it is satisfied. The history remains empty until the main cache is filled, i.e., t = N. From the pseudocode it is clear that items are discarded from the cache in statements 14, 17, and 21; no discards happen from the cache until either t1 + b1 = N (statement 12) or ℓ1 + ℓ2 = 2N (statement 20). If ℓ1 + ℓ2 = 2N is reached, since t1 + b1 ≤ N , we are guaranteed that t2 + b2 ≥ N and b1 + b2 = N , both of which will remain true from that point onward. Thus, by the time the 2N -th distinct page is requested, we have reached either t1 + b1 = N or ℓ1 + ℓ2 = 2N . We assume that request σ is processed in two distinct steps: first when Opt services the page request and, next when Arc services the request. We will show that inequality (11) is satisfied for each of the two steps. 15 Step 1: Opt services request σ Since only Opt acts in this step, CA = 0, and the contents of Arc’s cache does not change. There are two possible cases: either Opt faults on σ or it does not. Assume that page x is requested on request σ. If Opt does not fault on this request, then CO = 0. Since the contents of the cache maintained by Opt does not change, and neither do the lists L1 and L2 , we have ∆Φ = 0, and CA (σ) + ∆Φ ≤ 4N · CO (σ) ≤ 0. If Opt faults on request σ, then CO = 1. The contents of the cache maintained by Opt does change, which will affect the potential function. Opt will bring in page x into its cache. Assume that it evicts page y from its cache. The entry of page x into Opt’s cache can only decrease the potential function. The exit of page y from Opt’s cache can increase the potential function by at most 4N . The reason is as follows. Since the sum of b′1 , b′2 , t′1 , t′2 cannot exceed the size of Opt’s cache, we have 0 ≤ b′1 + t′1 + b′2 + t′2 ≤ N . Since b′1 + 2t′1 + 3b′2 + 4t′2 ≤ 4(b′1 + t′1 + b′2 + t′2 ), the left hand side cannot decrease by more than 4N . Thus, CA (σ) + ∆Φ1 ≤ 4N , proving inequality (11). Step 2: Arc services request σ There are four possible cases, which correspond to the four cases in Arc’s replacement algorithm. Case 1 deals with the case when Arc finds the page in its cache. The other three cases assume that Arc faults on this request because the item is not in T1 ∪ T2 . Cases 2 and 3 assume that the missing page is found recorded in the history in lists B1 and B2 , respectively. Case 4 assumes that the missing page is not recorded in history. Case I: Arc has a page hit. Clearly, CA = 0. We consider several subcases. In each case, the requested page will be moved to M RU (T2 ) while shifting other pages in T2 down. Case I.1 If the requested page is in T1′ , the move of this page from T1′ to T2′ implies ∆t′1 = −1; ∆t′2 = +1 and ∆Φ = −(2 · ∆t′1 + 4 · ∆t′2 ) = −2. Case I.2 If the requested page is in T2′ , the move of this page to M RU (T2 ) does not change the set of items in T2′ . Thus, ∆t′1 = ∆t′2 = 0 and ∆Φ = 0. Case I.3 If the requested page is in T1 − T1′ , then ∆t′1 = 0; ∆t′2 = +1 and ∆Φ = −4. One subtle point to note is that moving x from T1 − T1′ could potentially increase t′1 if the following conditions are met: x is located just below T1′ in T1 , it is not in Opt’s cache, and the items in T1 immediately below it are in Opt. However, x is already in Opt’s cache and there must be some item above it in T1 that is not in Opt. Case I.4 If the requested page is in T2 − T2′ , then ∆t′2 = +1 and ∆Φ = −4. The subtle point mentioned in Case I.3 also applies here. Next we will analyze the three cases when the requested page is not in Arc’s cache. Since CA = 1, the change in potential must be at most -1 in order for inequality (11) to be satisfied. We make the following useful observations in the form of lemmas. Lemma 6. If Arc has a miss and if the page is not in Arc’s history, we have ℓ′ = t′1 + t′2 + b′1 + b′2 < N . Consequently, we also have ℓ′1 < N and ℓ′2 < N . Proof. Since Opt has just finished serving the request, the page is present in the cache maintained by Opt just before Arc starts to service the request. If Arc has a miss, there is at least one page in the cache maintained by Opt that is not present in the cache maintained by Arc, implying that l′ < N . By definition, ℓ′ = ℓ′1 + ℓ′2 = t′1 + t′2 + b′1 + b′2 . Thus, the lemma holds. Lemma 7. A call to procedure Replace either causes an element to be moved from T1 to B1 or from T2 to B2 . In either case, the change in potential due to Replace, denoted by ∆ΦR , has an upper bound of 1. 16 Proof. Procedure Replace is only called when Arc has a page miss. Clearly, it causes an item to be moved from T1 to B1 or from T2 to B2 . If that item is in T1′ (or T2′ ), then T1 = T1′ (T2 = T2′ , resp.) and the moved item becomes part of B1′ (B2′ , resp.). Because the coefficients of b′1 and t′1 (b′2 and t′2 , resp.) differ by 1, we have ∆ΦR = +1. On the other hand, if that element is in T1 − T1′ (T2 − T2′ , resp.), then B1′ (B2′ , resp.) was empty before the move and remains empty after the move, and thus, ∆ΦR = 0. Lemma 8. On an Arc miss after phase P (0), if T1 = T1′ then the Replace step will not move a page from T2′ to B2 . On the other hand, if T2 = T2′ then Replace will not move a page from T1′ to B1 . Proof. In an attempt to prove by contradiction, let us assume that T1 = T1′ and T2 = T2′ are simultaneously true and Arc has a miss. By Lemma 5, we know that after phase, we have t = t1 + t2 = N , which by our assumption means that t′1 + t′2 = N ; this is impossible by Lemma 6. Thus, if T1 = T1′ , then T2 6= T2′ . Consequently, if LRU (T2 ) is moved to B2 , this item cannot be from T2′ . By a symmetric argument, if T2′ = T2 , then T1 6= T1′ , and LRU (T1 ) is not in T1′ . Case II: Arc has a miss and the missing page is in B1 Note that in this case the value of p will change by +1, unless its value equals N , in which case it has no change. Thus ∆p ≤ 1. If the missing item is in B1′ , then ∆b′1 = −1 and ∆t′2 = +1. Adding the change due to Replace, we get ∆Φ ≤ 1 − (∆b′1 + 4 · ∆t′2 ) + ∆ΦR ≤ −1 If the missing item is in B1 − B1′ , then we have ∆t′2 = 1 and ∆b′1 = 0. Thus, we have ∆Φ ≤ 1 − (∆b′1 + 4 · ∆t′2 ) + ∆ΦR ≤ −2 Case III: Arc has a miss and the missing page is in B2 . Note that in this case the value of p will change by -1, if its value was positive, otherwise it has no change. Thus ∆p ≤ 0. If the requested item is in B2′ , then ∆t′2 = 1, and ∆b′2 = −1. Thus, we have ∆Φ = ∆p − (3 · ∆b′2 + 4 · ∆t′2 ) + ∆ΦR ≤ 0 But this is not good enough since we need the potential change to be at most -1. When ∆p = −1, then we get the required inequality ∆Φ ≤ −1. Clearly, the difficulty is when ∆p = 0, which happens when p = 0. Since the missing item is from b′2 , it implies that B2′ is non-empty and T2′ = T2 . By Lemma 8 above, there must be at least one item in T1 − T1′ , which means that means that t1 > 0. As per the algorithm, since T1 is non-empty and p = 0, we are guaranteed to replace LRU (T1 ), and not an element from T1′ . Therefore, Replace will leave t′1 and b′1 unchanged, implying that ∆ΦR = 0. Thus, we have ∆Φ = ∆p − (3 · ∆b′2 + 4 · ∆t′2 ) + ∆ΦR ≤ −1 If the requested item is from B2 − B2′ , then ∆t′2 = 1, and ∆b′2 = 0. Thus, we have ∆Φ ≤ ∆p − (4 · ∆t′2 ) + ∆ΦR ≤ −3 Case IV: Arc has a miss and the missing page is not in B1 ∪ B2 17 We consider two cases. First, when ℓ1 = N , Arc will evict the LRU (L1 ). Since by Lemma 6, ℓ′1 < N , we know that for this case, b′1 remains unchanged at 0 and ∆t′1 = +1. Thus, ∆Φ ≤ ≤ −(2 · ∆t′1 ) + ∆ΦR −1 On the other hand, if ℓ1 < N , then Arc will evict the LRU (L2 ). Again, if the cache is full (i.e., t1 + t2 = N and ℓ1 + ℓ2 = 2N ), then we know that ℓ2 > N , which means that L′2 6= L2 and LRU (L2 ) is not in L′2 . Thus, deletion of LRU (L2 ) = LRU (B2 ) will not affect b′2 or any of the other quantities in the potential function. Then comes the Replace step, for which a bound has been proved earlier. Finally, a new item is brought in and placed in M RU (T1 ). Thus ∆t′1 ≤ 1. Putting it all together, we have ∆Φ ≤ −(2 · ∆t′1 ) + ∆ΦR ≤ −1 Wrapping up the proof of Theorem 8 Tying it all up, we have shown that inequality (11) holds for every request made after the cache is full, i.e., CA (σ) + ∆Φ ≤ 4N · CO (σ). If we assume that the caches started empty, then the initial potential is 0, while the final potential can be at most 4N . Thus, we have CA (σ) ≤ 4N · CO (σ) + 4N, thus proving Theorem 8. 4.6 Analyzing the Competitiveness of CAR Next, we analyze the competitiveness of Car. The main result of this section is the following: Theorem 9. Algorithm Car is 18N -competitive. Proof. Let PX [q] be the position of page q in an arbitrary ordered sequence of pages X. When the set is obvious, we will drop the subscript and denote PX [q] simply by P [q]. The set of history pages B1 and B2 will be treated as an ordered sequence of pages ordered from its LRU position to its MRU position. The set of main pages T10 (resp., T20 , T11 , and T21 ) will be treated as an ordered sequence of unmarked (resp., unmarked, marked, and marked) pages in T1 (resp, T2 , T1 , and T2 ) ordered from head to tail. Let Opt and Car be the set of (main and history) pages stored in the caches for algorithms Opt and Car respectively. Let D = (T1 ∪ T2 ∪ B1 ∪ B2 ) \ Opt. Thus D consists of pages in Car but not in Opt. We associate each page with a rank value R[q], which is defined as follows:  PB1 [q]      PB2 [q]    2P 0 [q] + b 1 T1 R[q] =  0 2P [q] + b 2  T2    3N + 2PT11 [q] + b1    3N + 2PT21 [q] + b2 if if if if if if q q q q q q ∈ B1 ∈ B2 ∈ T10 ∈ T20 ∈ T11 ∈ T21 (12) Finally, we define the potential function as follows: Φ= X
1 (p + 2(b1 + t1 ) + 3 R[q]) N − NO + 1 q∈D 18 (13) The initial value of Φ is 0. If the following inequality (14) is true for any request σ, where ∆Φ is the change in potential caused by serving the request, then when summed over all requests, it proves Theorem 9. CCar (σ) + ∆Φ ≤
18N COpt (σ). N − NO + 1 (14) As before, we assume that request σ is processed in two distinct steps: first when Opt serves and, next when Car serves. We will show that inequality (14) is satisfied for each of the two steps. Step 1: Opt serves request σ Since only Opt acts in this step, CCar = 0, and T1 ∪ T2 does not change. There are two possible cases: either Opt faults on σ or it does not. If Opt does not fault on this request, then it is easy to see that COpt = 0 and ∆Φ = 0, thus satisfying inequality (14). If Opt faults on request σ, then COpt = 1 and some page, q, is evicted from the cache maintained by Opt. If q is maintained by Car then it follows that q will belong to D after this step and thus its rank will contribute to the potential function, which will increase by three times the rank of q. The maximal positive change in potential will occur when q is the marked head page in T2 . In this case the rank of q is given by: R[q] = 3N + 2P [q] + b2 . The maximal possible values for each of the terms P [q] and b2 will be N , hence the maximum possible rank of q will be 3N + 2N + N = 6N . Therefore resulting potential change is at most 3(6N ) = 18N . Step 2: Car serves request σ We break down the analysis into four cases. Case 2.1 deals with the case when Car finds the page in its cache. The other three cases assume that Car faults on this request because the item is not in T1 ∪T2 . Cases 2.2 and 2.3 assume that the missing page is found recorded in the history in lists B1 and B2 , respectively. Case 2.4 assumes that the missing page is not recorded in history. Case 2.1: Car has a page hit Clearly, the page was found in T1 ∪ T2 , and CCar = 0. We consider the change of each of terms in the potential function individually. 1. As per the algorithm, p can only change when the page is found in history. (See lines 14 through 20 of Car(x).) Since the page is not found in Car’s history, ∆p = 0. 2. Neither the cache nor the history lists maintained by Car will change. Thus, the contribution to the second term in Φ, i.e., 2(b1 + t1 ) does not change. 3. Since Opt has already served the page, the page is in Opt’s cache. Therefore, even if the page gets marked during this hit, its rank value does not change. Thus, the contribution to the last term in Φ, also remains unchanged. We, therefore, conclude that ∆Φ = 0, satisfying inequality (14). Next we will analyze the three cases when the requested page is not in Car’s cache. Since CCar = 1, the change in potential must be at most −1 in each case in order for inequality (14) to be satisfied. Before tackling the three cases, the following lemmas (9 and 10) Pare useful for understanding the potential change caused by the last term in the potential function, i.e., q∈D R[q]. It is worth pointing out that a call to Replace moves either an item from T1 to B1 or from T2 to B2 , which is exactly the premise of Lemma 9 below. Lemma 9. When a page is moved from T1 to B1 (or from T2 to B2 ) its rank decreases by at least 1. Proof. Let q be any page in T1 . In order for q to be moved from T1 to B1 it must have been unmarked and located at the head of T1 . Since PT1 [q] = 1, the rank of q prior to the move must have been R[q] = 2PT1 [q] + b1 = b1 + 2, where b1 is the size of B1 prior to moving q. 19 After q is moved to the MRU position of B1 , R[q] = PB1 [q] = b1 + 1. Thus its rank decreased by 1. The arguments for the move from T2 to B2 are identical with the appropriate changes in subscripts. P Lemma 10. When Car has a page miss, the term q∈D R[q] in the potential function Φ cannot increase. Proof. We examine the rank change based on the original location of the page(s) whose ranks changed and in each case show that the rank change is never positive. Wherever appropriate we have provided references to line numbers in Pseudocode Car(x) from Appendix. Case A: q ∈ B1 ∪ B2 The rank of q ∈ B1 , which is simply its position in B1 , can change in one of three different ways. 1. Some page x less recently used than q (i.e., PB1 [x] < PB1 [q]) was evicted (Line 7). In this case, it is clear that PB1 [q] decreases by at least 1. 2. The page q is the requested page and is moved to T2 (Line 16). In this case, q ∈ Opt and hence its rank cannot affect the potential function. 3. Some page x is added to MRU of B1 (Line 27). Since pages are ordered from LRU to MRU, the added page cannot affect the rank of q. Using identical arguments for q ∈ B2 , we conclude that a miss will not increase the rank of any page in B1 ∪ B2 . Case B: q ∈ T10 ∪ T20 The rank of page q ∈ T10 , defined as R[q] = 2PT10 [q] + b1 , may be affected in four different ways. 1. If page q is the head of T1 and gets moved to B1 (Line 27), by lemma 9, the change in rank of q is at most −1. 2. If an unmarked page x is added to the tail of T1 (Line 13), then since the ordering is from head to tail, it does not affect the position of page q. Since there was no change in b1 , it is clear that the change in R[q] is 0. 3. If the unmarked page x 6= q at the head of T1 is marked and moved to tail of T2 (Line 29), then P [q] decreases by at least 1. Since the content of B1 is unchanged, the change in R[q] = 2P [q] + b1 is at most -2. 4. If the unmarked page x 6= q at the head of T1 is moved to B1 (Line 29), then P [q] decreases by at least 1, and b1 increases by 1. Hence the change in R[q] = 2P [q] + b1 is at most -1. The arguments are identical for q ∈ T20 . In each case, we have shown that a miss will not increase the rank of any page in T10 ∪ T20 . Case C: q ∈ T11 The rank of page q ∈ T11 , defined as R[q] = 3N + 2PT11 [q] + b1 , may be affected in four different ways. 1. If an unmarked page x is added to the tail of T1 (Line 13), then since the ordering is from head to tail, it does not affect the position of page q. Since there was no change in b1 , it is clear that the change in R[q] is 0. 2. If the unmarked page x 6= q at the head of T1 is marked and moved to tail of T2 (Line 29), then P [q] decreases by at least 1. Since B1 is unchanged, the change in R[q] = 3N + 2P [q] + b1 is at most -2. 3. If the unmarked page x 6= q at the head of T1 is moved to B1 (Line 29), then P [q] decreases by at least 1, and b1 increases by 1. Hence the change in R[q] = 3N + 2P [q] + b1 is at most -1. 20 4. Next, we consider the case when the marked page q is the head of T1 and gets unmarked and moved to T2 (Line 29). Prior to the move, the rank of q is given by R[q] = 3N + 2PT11 [q] + b1 . Since B1 could be empty, we know that R[q] ≥ 3N + 2. After page q is unmarked and moved to T2 , its rank is given by R[q] = 2PT20 [q] + b2 . Since P [q] ≤ N and b2 ≤ N , we know that the new R[q] ≤ 3N . Thus, the rank of page q does not increase. In each case, we have shown that a miss will not increase the rank of any page in T11 . Case D: q ∈ T21 The rank of page q ∈ T21 , defined as R[q] = 3N + 2PT21 [q] + b2 , may be affected in four different ways. 1. If an unmarked page x is added to the tail of T2 (Lines 16, 19, or 29), and if b2 does not change, it is once again clear that the change in R[q] is 0. 2. If a marked page x 6= q at the head of T2 gets unmarked and moved to the tail of T2 (Line 36), the position of q will decrease by 1 and there is no change in b2 . Thus R[q] changes by at most -2. 3. If an unmarked page x at the head of T2 is moved to B2 (Line 34), P [q] decreases by 1 and b2 increases by 1. Thus R[q] changes by at most -1. 4. Finally, we consider the case when the marked page q is the head of T2 and gets unmarked and moved to the tail of T2 (Line 36). Prior to the move, the rank of q is given by R[q] = 3N + 2PT21 [q] + b2 . Even if B2 is empty, we know that R[q] ≥ 3N + 2. After page q is unmarked and moved to T2 , its rank is given by R[q] = 2PT20 [q] + b2 . Since P [q] ≤ N and b2 ≤ N , we know that the new R[q] ≤ 3N . Thus, the rank of page q does not increase. In each case, we have shown that a miss will not increase the rank of any page in T21 . The four cases (A through D) together complete the proof of Lemma 10. We continue with the remaining cases for the proof of Theorem 9. Case 2.2: Car has a page miss and the missing page is in B1 We consider the change in the potential function (defined in Eq. 13) by analyzing each of its three terms. 1. Value of p increases by 1, except when it is equal to N , in which case it remains unchanged. (See Line 15.) Thus, the first term increases by at most 1. 2. The call to Replace has no effect on the value of (t1 + b1 ) because an item is moved either from T1 to B1 or from T2 to B2 . Since the requested page in B1 is moved to T2 , (t1 + b1 ) decreases by 1. 3. By Lemma 10, we already know that the last term increases by at most 0. Since p increases by at most 1 and the term 2(t1 +b1 ) decreases by at least 2, the total change in the potential function, is at most -1. Case 2.3: Car has a page miss and the missing page is in B2 When the missing page is in B2 , Car makes a call to Replace (Line 5) and then executes Lines 18-19. Thus, p is decremented except if it is already equal to 0. We consider two subcases: ∆p < 0 and ∆p = 0. ∆p < 0: As in Case 2.2, the call to Replace has no effect on (t1 + b1 ). Since, Lines 18-19 do not affect T1 ∪ B1 , the second term does not change. By Lemma 10, we know that the last term increases by at most 0. Since ∆p ≤ −1, the total change in the potential function, ∆p + ∆2(t1 + b1 ) is at most -1. 21 ∆p = 0: Unlike the subcase above when p decreases by 1, the change in p cannot guarantee the required reduction in the potential. We therefore need a tighter argument. We know that there is a call to Replace. Three cases arise and are discussed below. • If T1 is empty, then T2 must have N pages, at least one of which must be in D. Also, Replace must act on T2 , eventually evicting an unmarked page from head of T2 , causing the rank of any page from T2 \ Opt to decrease by 1. • If T1 is not empty and has at least one page from D, then the condition in Line 24 passes and Replace must act on T1 , eventually evicting an unmarked page from head of T1 , causing the rank of at least one page from T1 \ Opt to decrease by 1. • Finally, if T1 is not empty and all its pages are in Opt, then T2 must have a page q ∈ D. Since the requested page x was found in B2 and is moved to the tail of T2 , even though the position of q in T2 does not change, b2 decreased by 1 and consequently the rank of q decreases by 1. Thus, in each case, even though neither p nor the quantity (t1 + b1 ) changed, the third term involving ranks, and consequently, the potential function decreased by at least 3. The following two lemmas are useful for Case 2.4, when the missing page is not in T1 ∪ T2 ∪ B1 ∪ B2 . Lemma 11. We make two claims: 1. If t1 + b1 = N and the LRU page of B1 is evicted from the cache on Line 7, then decrease by at least one. 2. If t2 + b2 > N , and the LRU page of B2 , is evicted from the cache on Line 9, then decrease by at least one. P q∈D P q∈D R[q] will R[q] will Proof. We tacke the first claim. Assume that y is the LRU page of B1 that is being evicted on Line 7. Then Car must have had a page miss on x 6∈ B1 ∪ B2 , and the requested page x is added to the tail of T1 . Since t1 + b1 = N , there is at least one page q ∈ T1 ∪ B1 that is not in Opt’s cache and whose rank contributes to the potential function. First, we assume that q ∈ T1 \ Opt, whose rank is given by: R[q] = 2 ∗ P [q] + b1 . For each of the three cases, we show that the potential function does decrease by at least 1. • If Replace acts on T1 and the unmarked head of T1 , different from q, is moved to B1 then the size of B1 remains the same (because a page gets added to B1 while another page is evicted) but the position of q in T1 decreases by one. Therefore R[q] decreases by 2. • If Replace acts on T1 and q itself is moved to B1 then by Lemma 9, R[q] decreases by at least 1. • If Replace acts on T2 , then we use the fact that a page is evicted from B1 , and the b1 term in R[q] must decrease by 1. Next, we assume that q ∈ B1 \ Opt. Since LRU (B1 ) is evicted, the position of the page q will decrease by one. Thus R[q] = PB1 [q] must decrease by at least 1, completing the proof of the first claim in the lemma. The proof of the second claim is very similar and only requires appropriate changes to the subscripts. Next we tackle the last case in the proof of Theorem 9. Case 2.4: Car has a page miss and the missing page is not in B1 ∪ B2 We assume that Car’s cache is full (i.e., l1 + l2 = 2N ). We consider two cases below – first, if l1 = N and the next when l1 < N . If l1 = t1 + b1 = N , Car will call Replace, evict LRU (B1 ) and then add the requested page to the tail of T1 . Below, we analyze the changes to the three terms in the potential function. • Since p is not affected, the first term does not change. • Since a page is added to T1 and a page is evicted from B1 , the net change in the second term is 0. • Since the conditions of Lemma 11 apply, the total rank will decrease by at least 1. 22 Adding up all the changes, we conclude that the potential function decreases by at least 3. If l1 < N , Car will call Replace, evict LRU (B2 ) and then add a page to the tail of T1 . As above, we analyze the changes to the three terms in the potential function. • Since p is not affected, the first term does not change. • A page is added to T1 and a page is evicted from B2 hence (t1 + b1 ) increases by 1. • Since l2 > N , the conditions of Lemma 11 apply, the total rank will decrease by at least 1. Adding up all the changes, we conclude that the potential function decreases by at least 1, thus completing Case 2.4. Wrapping up the proof of Theorem 9: Combining the four cases (2.1 through 2.4) proves that inequality (14) is satisfied when Car serves request σ. This completes the proof of Theorem 9, establishing that the upper bound on the competitiveness of Car is 18N . 5 Conclusions and Future Work Adaptive algorithms are tremendously important in situations where inputs are infinite online sequences and no single optimal algorithm exists for all inputs. Thus, different portions of the input sequence require different algorithms to provide optimal responses. Consequently, it is incumbent upon the algorithm to sense changes in the nature of the input sequence and adapt to these changes. Unfortunately, these algorithms are harder to analyze. We present the analysis of two important adaptive algorithms called Arc and Car and show that they are competitive along with proving good lower bounds on the competitiveness ratios. Two important open questions remain unanswered. Given that there is a gap between the lower and upper bounds on the competitiveness ratios of the two adaptive algorithms, Arc and Car, what is the true ratio? More importantly, is there an “expected” competitiveness ratio for request sequences that come from real applications? The second question would help explain why Arc and Car perform better in practice than Lru and Clock, respectively. Acknowledgments This work was partly supported by two NSF Grants (CNS-1018262 and CNS-1563883) and the NSF Graduate Research Fellowship (DGE-1038321). We are grateful to Kirk Pruhs for suggesting enhancing our results with the assumption of unequal cache sizes. 23 References [Alb96] S. Albers. Competitive online algorithms. Technical report, BRICS Lecture Series, Computer Science Department, University of Aarhus, 1996. [BM04] S. Bansal and D. S. Modha. CAR: CLOCK with adaptive replacement. In Proceedings of the 3rd USENIX Conference on File and Storage Technologies, FAST ’04, pages 187–200, Berkeley, CA, USA, 2004. USENIX Association. [Cor68] F. J. Corbato. A paging experiment with the MULTICS system. Technical report, DTIC Document, 1968. [Fri99] M. B. Friedman. Windows NT page replacement policies. In Proceedings of the Intl. CMG Conference, pages 234–244, 1999. [Hoc97] D. S. Hochbaum, editor. Approximation algorithms for NP-hard problems. PWS Publishing Co., Boston, MA, USA, 1997. [JCZ05] S. Jiang, F. Chen, and X. Zhang. CLOCK-Pro: An effective improvement of the CLOCK replacement. In USENIX Annual Technical Conference, General Track, pages 323–336, 2005. [JIPP10] A. Janapsatya, A. Ignjatovic, J. Peddersen, and S. Parameswaran. Dueling CLOCK: adaptive cache replacement policy based on the CLOCK algorithm. In Design, Automation & Test in Europe Conference & Exhibition (DATE), 2010, pages 920–925. IEEE, 2010. [JS94] T. Johnson and D. Shasha. 2Q: A low overhead high performance buffer management replacement algorithm. In Proc. of VLDB, pages 297–306, 1994. [JZ02] S. Jiang and X. Zhang. LIRS: An efficient low inter-reference recency set replacement policy to improve buffer cache performance. In Proc. ACM Sigmetrics Conf., pages 297–306. ACM Press, 2002. [LCK+ 01] D. Lee, J. Choi, J. H. Kim, S. H. Noh, S. L. Min, Y. Cho, and C. S. Kim. LRFU: A spectrum of policies that subsumes the least recently used and least frequently used policies. IEEE Trans. Comput., 50(12):1352–1361, December 2001. [MM03] N. Megiddo and D. S. Modha. ARC: A self-tuning, low overhead replacement cache. In Proceedings of the 2nd USENIX Conference on File and Storage Technologies, FAST ’03, pages 115–130, Berkeley, CA, USA, 2003. USENIX Association. [MM04] N. Megiddo and D. S. Modha. Outperforming LRU with an adaptive replacement cache algorithm. IEEE Computer, 37(4):58–65, 2004. [OOW93] E. J. O’Neil, P. E. O’Neil, and G. Weikum. The LRU-K page replacement algorithm for database disk buffering. SIGMOD Rec., 22(2):297–306, June 1993. [ST85] D. D. Sleator and R. E. Tarjan. Amortized efficiency of list update and paging rules. Commun. ACM, 28(2):202–208, February 1985. 24 6 Appendix We reproduce the pseudocode for Arc and Car below. Pseudocode: Arc(x) INPUT: The requested page x INITIALIZATION: Set p = 0 and set lists T1 , B1 , T2 , and B2 to empty 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: if (x is in T1 ∪ T2 ) then Move x to the top of T2 else if (x is in B1 ) then Adaptation: Update p = min{p + 1, N } Replace() Fetch x and move to the top of T2 else if (x is in B2 ) then Adaptation: Update: p = max{p − 1, 0} Replace() Fetch x and move to the top of T2 else if (t1 + b1 = N ) then if (t1 < N ) then Discard LRU item in B1 Replace() else Discard LRU page in T1 and remove from cache end if else if ((t1 + b1 < N ) and (t1 + t2 + b1 + b2 ≥ N )) then if (t1 + t2 + b1 + b2 = 2N ) then Discard LRU item in B2 end if Replace() end if Fetch x and move to the top of T1 end if Replace() 26: if ((t1 ≥ 1) and ((x ∈ B2 and t1 = p) or (t1 > p))) then 27: Discard LRU page in T1 and insert as MRU history item in B1 28: else 29: Discard LRU page in T2 and insert as MRU history item in B2 30: end if 25 ⊲ cache hit ⊲ cache history hit ⊲ learning rate = 1 ⊲ make space in T1 or T2 ⊲ cache history hit ⊲ learning rate = 1 ⊲ make space in T1 or T2 ⊲ cache and history miss ⊲ make space in T1 or T2 ⊲ make space in T1 or T2 Pseudocode: Car(x) INPUT: The requested page x INITIALIZATION: Set p = 0 and set lists T1 , B1 , T2 , and B2 to empty 1: 2: 3: 4: 5: 6: 7: 8: 9: 10: 11: 12: 13: 14: 15: 16: 17: 18: 19: 20: 21: if (x is in T1 ∪ T2 ) then ⊲ cache hit Mark page x else ⊲ cache miss if (t1 + t2 = N ) then ⊲ cache full, replace a page from cache Replace() ⊲ make space in T1 or T2 if ((x 6∈ B1 ∪ B2 ) and (t1 + b1 = N )) then Discard LRU page in B1 else if ((x 6∈ B1 ∪ B2 ) and (t1 + t2 + b1 + b2 = 2N )) then Discard LRU page in B2 . end if end if if (x 6∈ B1 ∪ B2 ) then ⊲ cache miss Insert x at the tail of T1 ; Unmark page x else if (x ∈ B1 ) then ⊲ cache history hit Adaptation: Update p = min{p + 1, N } ⊲ learning rate = 1 Move x to the tail of T2 ; Unmark page x else ⊲ cache history hit Adaptation: Update: p = max{p − 1, 0} ⊲ learning rate = 1 Move x to the tail of T2 ; Unmark page x end if end if Replace() 22: found = false 23: repeat 24: if (t1 ≥ max{1, p}) then 25: if (head page in T1 is unmarked) then 26: found = true 27: Discard head page in T1 and insert as MRU history item in B1 28: else 29: Unmark head page in T1 , move page as tail page in T2 , and move head of T1 clockwise 30: end if 31: else 32: if (head page in T2 is unmarked) then 33: found = true 34: Discard head page in T2 and insert as MRU history item in B2 35: else 36: Unmark head page in T2 , and move head of T2 clockwise 37: end if 38: end if 39: until (found) 26 

1–12 Further and stronger analogy between sampling and optimization: Langevin Monte Carlo and gradient descent Arnak S. Dalalyan ARNAK . DALALYAN @ ENSAE . FR arXiv:1704.04752v2 [math.ST] 28 Jul 2017 ENSAE/CREST/Université Paris Saclay Abstract 1 In this paper , we revisit the recently established theoretical guarantees for the convergence of the Langevin Monte Carlo algorithm of sampling from a smooth and (strongly) log-concave density. We improve the existing results when the convergence is measured in the Wasserstein distance and provide further insights on the very tight relations between, on the one hand, the Langevin Monte Carlo for sampling and, on the other hand, the gradient descent for optimization. Finally, we also establish guarantees for the convergence of a version of the Langevin Monte Carlo algorithm that is based on noisy evaluations of the gradient. Keywords: Markov Chain Monte Carlo, Approximate sampling, Rates of convergence, Langevin algorithm, Gradient descent 1. Introduction p Let R p be a positive integer and f : R → R be a measurable function such that the integral Rp exp{−f (θ)} dθ is finite. In various applications, one is faced with the problems of finding the minimum point of f or computing the average with respect to the probability density π(θ) = R e−f (θ) . −f (u) du Rp e In other words, one often looks for approximating the values θ ∗ and θ̄ defined as Z θ̄ = θ π(θ) dθ, θ ∗ ∈ arg minp f (θ). Rp θ∈R In most situations, the approximations of these values are computed using iterative algorithms which share many common features. There is a vast variety of such algorithms for solving both tasks, see for example (Boyd and Vandenberghe, 2004) for optimization and (Atchadé et al., 2011) for approximate sampling. The similarities between the task of optimization and that of averaging have been recently exploited in the papers (Dalalyan, 2014; Durmus and Moulines, 2016; Durmus et al., 2016) in order to establish fast and accurate theoretical guarantees for sampling from and averaging with respect to the density π using the Langevin Monte Carlo algorithm. The goal of the present work is to push further this study both by improving the existing bounds and by extending them in some directions. 1. This paper has been published in proceedings of COLT 2017. However, this version is more recent. We have corrected some typos (2/(m + M ) instead of 1/(m + M ) on pages 3-4) and slightly improved the upper bound of Theorem 3. c A.S. Dalalyan. DALALYAN We will focus on strongly convex functions f having a Lipschitz continuous gradient. That is, we assume that there exist two positive constants m and M such that  f (θ) − f (θ 0 ) − ∇f (θ 0 )> (θ − θ 0 ) ≥ (m/2)kθ − θ 0 k22 , ∀θ, θ 0 ∈ Rp , (1) k∇f (θ) − ∇f (θ 0 )k ≤ M kθ − θ 0 k , 2 2 where ∇f stands for the gradient of f and k · k2 is the Euclidean norm. We say that the density π(θ) ∝ e−f (θ) is log-concave (resp. strongly log-concave) if the function f satisfies the first inequality of (1) with m = 0 (resp. m > 0). The Langevin Monte Carlo (LMC) algorithm studied throughout this work is the analogue of the gradient descent algorithm for optimization. Starting from an initial point ϑ(0) ∈ Rp that may be deterministic or random, the iterations of the algorithm are defined by the update rule √ ϑ(k+1,h) = ϑ(k,h) − h∇f (ϑ(k,h) ) + 2h ξ (k+1) ; k = 0, 1, 2, . . . (2) where h > 0 is a tuning parameter, referred to as the step-size, and ξ (1) , . . . , ξ (k) , . . . is a sequence of mutually independent, and independent of ϑ(0) , centered Gaussian vectors with covariance matrices equal to identity. Under the assumptions imposed on f , when h is small and k is large (so that the product kh is large), the distribution of ϑ(k,h) is close in various metrics to the distribution with density π(θ), hereafter referred to as the target distribution. An important question is to quantify this closeness; this might be particularly useful for deriving a stopping rule for the LMC algorithm. The measure of approximation used in this paper is the Wasserstein-Monge-Kantorovich distance W2 . For two measures µ and ν defined on (Rp , B(Rp )), W2 is defined by Z  1/2 W2 (µ, ν) = inf kθ − θ 0 k22 dγ(θ, θ 0 ) , γ∈Γ(µ,ν) Rp ×Rp where the inf is with respect to all joint distributions γ having µ and ν as marginal distributions. This distance is perhaps more suitable for quantifying the quality of approximate sampling schemes than other metrics such as the total variation. Indeed, on the one hand, bounds on the Wasserstein distance—unlike the bounds on the total-variation distance—directly provide the level of approximating the first order moment. For instance, if µ and ν are two Dirac measures at the points θ and θ 0 , respectively, then the total-variation distance DTV (δθ , δθ0 ) equals one whenever θ 6= θ 0 , whereas W2 (δθ , δθ0 ) = kθ − θ 0 k2 is a smoothly increasing function of the Euclidean distance between θ and θ 0 . This seems to better correspond to the intuition on the closeness of two distributions. 2. Improved guarantees for the Wasserstein distance The rationale behind the LMC algorithm (2) is simple: the Markov chain {ϑ(k,h) }k∈N is the Euler discretization of a continuous-time diffusion process {Lt : t ∈ R+ }, known as Langevin diffusion, that has π as invariant density (Bhattacharya, 1978, Thm. 3.5). The Langevin diffusion is defined by the stochastic differential equation √ dLt = −∇f (Lt ) dt + 2 dW t , t ≥ 0, (3) where {W t : t ≥ 0} is a p-dimensional Brownian motion. When f satisfies condition (1), equation (3) has a unique strong solution which is a Markov process. Let νk be the distribution of the k-th iterate of the LMC algorithm, that is ϑ(k,h) ∼ νk . 2 F URTHER ANALOGY BETWEEN SAMPLING AND OPTIMIZATION Theorem 1 Assume that h ∈ (0, 2/M ). The following claims hold: (a) If h ≤ 2/(m+M ) then W2 (νK , π) ≤ (1 − mh)K W2 (ν0 , π) + 1.82(M/m)(hp)1/2 . (b) If h ≥ 2/(m+M ) then W2 (νK , π) ≤ (M h − 1)K W2 (ν0 , π) + 1.82 Mh (hp)1/2 . 2 − Mh The proof of this theorem is postponed to Section 6. We content ourselves here by discussing the relation of this result to previous work. Note that if the initial value ϑ(0) = θ (0) is deterministic then, according to (Durmus and Moulines, 2016, Theorem 1), we have Z 2 kθ (0) − θk22 π(dθ) W2 (ν0 , π) = Rp Z (0) 2 kθ̄ − θk22 π(dθ) = kθ − θ̄k2 + Rp ≤ kθ (0) − θ̄k22 + p/m. (4) First of all, let us remark that if we choose h and K so that h ≤ 2/(m+M ), e−mhK W2 (ν0 , π) ≤ ε/2, 1.82(M/m)(hp)1/2 ≤ ε/2, (5) then we have W2 (νK , π) ≤ ε. In other words, conditions (5) are sufficient for the density of the output of the LMC algorithm with K iterations to be within the precision ε of the target density when the precision is measured using the Wasserstein distance. This readily yields h≤ m 2 ε2 2 ∧ 2 14M p m + M and hK ≥  2(kθ (0) − θ̄k2 + p/m)1/2  1 2 log m ε Assuming m, M and kθ (0) − θ̄k22 /p to be constants, we can deduce from the last display that it suffices K = Cpε−2 log(p/ε) number of iterations in order to reach the precision level ε. This fact has been first established in (Dalalyan, 2014) for the LMC algorithm with a warm start and the total-variation distance. It was later improved by Durmus and Moulines (2016), who showed that the same result holds for any starting point and established similar bounds for the Wasserstein distance. In order to make the comparison easier, let us recall below the corresponding result from2 (Durmus and Moulines, 2016). It asserts that under condition (1), if h ≤ 2/(m+M ) then   mM h K 2 M hp m + M  M 2 h M 2 h2  W22 (νK , π) ≤ 2 1 − W2 (ν, π) + (m + M ) h + 2+ + . m+M m 2mM m 6 (6) When we compare this inequality with the claims of Theorem 1, we see that i) Theorem 1 holds under weaker conditions: h ≤ 2/M instead of h ≤ 2/(m+M ). ii) The analytical expressions of the upper bounds on the Wasserstein distance in Theorem 1 are not as involved as those of (6). 2. We slightly adapt the original result taking into account the fact that we are dealing with the LMC algorithm with a constant step. 3 DALALYAN 14 logK(p) 12 10 8 ε= ε= ε= ε= 6 4 2 0 20 40 60 0 .3 0 .3 0 .1 0 .1 nb nb nb nb 80 of of of of 100 p steps steps steps steps fro m fro m fro m fro m 120 the DM bound o ur b o und o ur b o und the DM bound 140 160 180 200 Figure 1: The curves of the functions p 7→ log K(p), where K(p) is the number of steps— derived either from our bound or from the bound (6) of (Durmus and Moulines, 2016)—sufficing for reaching the precision level ε (for ε = 0.1 and ε = 0.3). iii) If we take a closer look, we can check that when h ≤ 2/(m+M ), the upper bound in part (a) of Theorem 1 is sharper than that of (6). In order to better illustrate the claim in iii) above, we consider a numerical example in which m = 4, M = 5 and kθ (0) − θ̄k22 = p. Let Four (h, K, p) and FDM (h, K, p) be the upper bounds on W2 (νK , π) provided by Theorem 1 and (6). For different values of p, we compute  Kour (p) = min K : there exists h ≤ 2/(m+M ) such that Four (h, K, p) ≤ ε ,  KDM (p) = min K : there exists h ≤ 2/(m+M ) such that FDM (h, K, p) ≤ ε . The curves of the functions p 7→ log Kour (p) and p 7→ log KDM (p), for ε = 0.1 and ε = 0.3 are plotted in Figure 1. We can deduce from these plots that the number of iterations yielded by our bound is more than 5 times smaller than the number of iterations recommended by bound (6) of Durmus and Moulines (2016). Remark 2 Although the upper bound on W2 (ν0 , π) provided by (4) is relevant for understanding the order of magnitude of W2 (ν0 , π), it has limited applicability since the distance kθ 0 − θ̄k might 4 F URTHER ANALOGY BETWEEN SAMPLING AND OPTIMIZATION be hard to evaluate. An attractive alternative to that bound is the following3 : Z 2 kθ (0) − θk22 π(dθ) W2 (ν0 , π) = RpZ   2 f (θ 0 ) − f (θ) − ∇f (θ)> (θ 0 − θ) π(dθ) ≤ m Rp Z  2 = f (θ 0 ) − f (θ) π(dθ) + p . m Rp If f is lower bounded by some known constant, for instance
if f ≥ 0, the last inequality provides 2 the computable upper bound W2 (ν0 , π)2 ≤ m f (θ 0 ) + p . 3. Relation with optimization We have already mentioned that the LMC algorithm is very close to the gradient descent algorithm for computing the minimum θ ∗ of the function f . However, when we compare the guarantees of Theorem 1 with those available for the optimization problem, we remark the following striking difference. The approximate computation of θ ∗ requires a number of steps of the order of log(1/ε) to reach the precision ε, whereas, for reaching the same precision in sampling from π, the LMC algorithm needs a number of iterations proportional to (p/ε2 ) log(p/ε). The goal of this section is to explain that this, at first sight very disappointing behavior of the LMC algorithm is, in fact, continuously connected to the exponential convergence of the gradient descent. The main ingredient for the explanation is that the function f (θ) and the function fτ (θ) = f (θ)/τ have the same point of minimum θ ∗ , whatever the real number τ > 0. In addition, if we
define the density function πτ (θ) ∝ exp − fτ (θ) , then the average value Z θ̄ τ = θ πτ (θ) dθ Rp ∗ tends to the minimum point θ when τ goes to zero. Furthermore, the distribution πτ (dθ) tends to the Dirac measure at θ ∗ . Clearly, fτ satisfies (1) with the constants mτ = m/τ and Mτ = M/τ . Therefore, on the one hand, we can apply to πτ claim (a) of Theorem 1, which tells us that if we choose h = 1/Mτ = τ /M , then   M  pτ 1/2 m K W2 (νK , πτ ) ≤ 1 − W2 (δθ(0) , πτ ) + 2 . (7) M m M On the other hand, the LMC algorithm with the step-size h = τ /M applied to fτ reads as r 1 2τ (k+1) ϑ(k+1,h) = ϑ(k,h) − ∇f (ϑ(k,h) ) + ξ ; k = 0, 1, 2, . . . (8) M M When the parameter τ goes to zero, the LMC sequence (8) tends to the gradient descent sequence θ (k) . Therefore, the limiting case of (7) corresponding to τ → 0 writes as  m K (0) kθ (K) − θ ∗ k2 ≤ 1 − kθ − θ ∗ k2 , M which is a well-known result in Optimization. This clearly shows that Theorem 1 is a natural extension of the results of convergence from optimization to sampling. 3. RThe second line follows from R strong convexity whereas the third line is a consequence of the two identities ∇f (θ)π(dθ) = 0 and Rp θ > ∇f (θ)π(dθ) = p. These identities follow from the fundamental theorem of Rp calculus and the integration by parts formula, respectively. 5 DALALYAN 4. Guarantees for the noisy gradient version In some situations, the precise evaluation of the gradient ∇f (θ) is computationally expensive or practically impossible, but it is possible to obtain noisy evaluations of ∇f at any point. This is the setting considered in the present section. More precisely, we assume that at any point ϑ(k,h) ∈ Rp of the LMC algorithm, we can observe the value Y (k,h) = ∇f (ϑ(k,h) ) + σ ζ (k) , where {ζ (k) : k = 0, 1, . . .} is a sequence of independent zero mean random vectors such that E[kζ (k) k22 ] ≤ p and σ > 0 is a deterministic noise level. Furthermore, the noise vector ζ (k) is independent of the past states ϑ(1,h) , . . . , ϑ(k,h) . The noisy LMC (nLMC) algorithm is then defined as √ k = 0, 1, 2, . . . (9) ϑ(k+1,h) = ϑ(k,h) − hY (k,h) + 2h ξ (k+1) ; where h > 0 and ξ (k+1) are as in (2). The next theorem extends the guarantees of Theorem 1 to the noisy-gradient setting and to the nLMC algorithm. Theorem 3 Let ϑ(K,h) be the K-th iterate of the nLMC algorithm (9) and νK be its distribution. If the function f satisfies condition (1) and h ≤ 2/M then the following claims hold: (a) If h ≤ 2/(m+M ) then   2hp 1/2 n mh K 3.3M 2 o1/2 W2 (νK , π) ≤ 1 − σ2 + W2 (ν0 , π) + . 2 m m (10) (b) If h ≥ 2/(m+M ) then W2 (νK , π) ≤  M h K 2 W2 (ν0 , π) +  2h2 p 1/2 n 6.6M o1/2 σ2 + . 2 − Mh 2 − Mh To understand the potential scope of applicability of this result, let us consider a typical statistical problem in which f (θ) is the negative log-likelihood of n independent random variables X1 , . . . , Xn . Then, if `(θ, x) is the log-likelihood of one variable, we have f (θ) = n X `(θ, Xi ). i=1 In such a situation, if the Fisher information is not degenerated, both m and M are proportional to the sample size n. When the gradient of `(θ, Xi ) with respect to parameter θ is hard to compute, one can replace the evaluation of ∇f (ϑ(k,h) ) at each step k by that of Yk = n∇θ `(ϑ(k,h) , Xk ). Under suitable assumptions, this random vector satisfies the conditions of Theorem 3 with a σ 2 proportional to n. Therefore, if we analyze the expression between curly brackets in (10), we see that the additional term, σ 2 , due to the subsampling is of the same order of magnitude as the term 3.3M 2 /m. Thus, using the subsampled gradient in the LMC algorithm does not cause a significant deterioration of the precision while reducing considerably the computational burden. 6 F URTHER ANALOGY BETWEEN SAMPLING AND OPTIMIZATION 5. Discussion and outlook We have established simple guarantees for the convergence of the Langevin Monte Carlo algorithm under the Wasserstein metric. These guarantees are valid under strong convexity and Lipschitzgradient assumptions on the log-density function, for a step-size smaller than 2/M , where M is the constant in the Lipschitz condition. These guarantees are sharper than previously established analogous results and in perfect agreement with the analogous results in Optimization. Furthermore, we have shown that similar results can be obtained in the case where only noisy evaluations of the gradient are possible. There are a number of interesting directions in which this work can be extended. One relevant and closely related problem is the approximate computation of the volume of a convex body, or, the problem of sampling from the uniform distribution on a convex body. This problem has been analyzed by other Monte Carlo methods such as “Hit and Run” in a series of papers by Lovász and Vempala (2006b,a), see also the more recent paper (Bubeck et al., 2015). Numerical experiments reported in (Bubeck et al., 2015) suggest that the LMC algorithm might perform better in practice than “Hit and Run”. It would be interesting to have a theoretical result corroborating this observation. Other interesting avenues for future research include the possible adaptation of the Nesterov acceleration to the problem of sampling, extensions to second-order methods as well as the alleviation of the strong-convexity assumptions. We also plan to investigate in more depth the applications is high-dimensional statistics (see, for instance, Dalalyan and Tsybakov (2012)). Some results in these directions are already obtained in (Dalalyan, 2014; Durmus and Moulines, 2016; Durmus et al., 2016). It is a stimulating question whether we can combine ideas of the present work and the aforementioned earlier results to get improved guarantees. 6. Proofs The first part of the proofs of Theorem 1 and Theorem 3 is the same. We start this section by this common part and then we proceed with the proofs of the two theorems separately. √ Let W be a p-dimensional Brownian Motion such that W (k+1)h − W kh = h ξ (k+1) . We define the stochastic process L so that L0 ∼ π and Z t √ Lt = L0 − ∇f (Ls ) ds + 2 W t , ∀ t > 0. (11) 0 It is clear that this equation implies that Z (k+1)h L(k+1)h = Lkh − ∇f (Ls ) ds + kh Z (k+1)h = Lkh − ∇f (Ls ) ds + √ √ 2 (W (k+1)h − W kh ) 2h ξ (k+1) . kh Furthermore, {Lt : t ≥ 0} is a diffusion process having π as the stationary distribution. Since the initial value L0 is drawn from π, we have Lt ∼ π for every t ≥ 0. 7 DALALYAN Let us denote ∆k = Lkh − ϑ(k,h) and Ik = (kh, (k + 1)h]. We have Z ∇f (Lt ) dt ∆k+1 = ∆k + hY (k,h) − Ik Z

(k,h) (k,h) (k) = ∆k − h ∇f (ϑ ∇f (Lt ) − ∇f (Lkh ) dt . + ∆k ) − ∇f (ϑ ) + σhζ − | {z } I | k {z } :=U k :=V k In view of the triangle inequality, we get k∆k+1 k2 ≤ k∆k − hU k + σhζ (k) k2 + kV k k2 . (12) For the first norm in the right hand side, we can use the following inequalities: E[k∆k − hU k + σhζ (k) k22 ] = E[k∆k − hU k k22 ] + E[kσhζ (k) k22 ] = E[k∆k − hU k k22 ] + σ 2 h2 p. (13) We need now three technical lemmas the proofs of which are postponed to Section 6.3. Lemma 1 Let us introduce the constant γ that equals |1 − mh| if h ≤ 2/(m+M ) and |1 − M h| if h ≥ 2/(m+M ). (Since h ∈ (0, 2/M ), this value γ satisfies 0 < γ < 1). It holds that k∆k − hU k k2 ≤ γk∆k k2 . (14) Lemma 2 If the function f is continuously differentiable and the gradient of f is Lipschitz with constant M , then Z k∇f (x)k22 π(x) dx ≤ M p. Rp Lemma 3 If the function f has a Lipschitz-continuous gradient with
the Lipschitz constant M , L R a+h ∇f (Lt ) − ∇f (La ) dt for some a ≥ 0, then is the Langevin diffusion (11) and V (a) = a E[kV
1/2 (a)k22 ]  ≤ 1 4 3 h M p 3 1/2 + (h3 p)1/2 M. This completes the common part of the proof. We present below the proofs of the theorems. 6.1. Proof of Theorem 1 Using (12) with σ = 0 and Lemma 1, we get k∆k+1 k2 ≤ γk∆k k2 + kV k k2 , ∀k ∈ N. In view of the Minkowski inequality and Lemma 3, this yields (E[k∆k+1 k22 ])1/2 ≤ γ(E[k∆k k22 ])1/2 + (E[kV k k22 ])1/2 ≤ γ(E[k∆k k22 ])1/2 + 1.82(h3 M 2 p)1/2 , 8 F URTHER ANALOGY BETWEEN SAMPLING AND OPTIMIZATION where we have used the fact that h ≤ 2/M . Using this inequality iteratively with k − 1, . . . , 0 instead of k, we get (E[k∆k+1 k22 ])1/2 ≤ γ k+1 (E[k∆0 k22 ])1/2 + 1.82(h3 M 2 p)1/2 k X γj j=0 ≤γ k+1 (E[k∆0 k22 ])1/2 3 2 1/2 + 1.82(h M p) (1 − γ)−1 . (15) Since ∆k+1 = L(k+1)h − ϑ(k+1,h) and L(k+1)h ∼ π, we readily get the inequality W2 (νk+1 , π) ≤
1/2
1/2 E[k∆k+1 k22 ] . In addition, one can choose L0 so that W2 (ν0 , π) = E[k∆0 k22 ] . Using these relations and substituting γ by its expression in (15), we get the two claims of the theorem. 6.2. Proof of Theorem 3 Using (12), (13) and Lemma 1, we get (for every t > 0) E[k∆k+1 k22 ] = E[k∆k − hU k + V k k22 ] + E[kσhζ (k) k22 ] ≤ (1 + t)E[k∆k − hU k k22 ] + (1 + t−1 )E[kV k k22 ] + σ 2 h2 p ≤ (1 + t)γ 2 E[k∆k k22 ] + (1 + t−1 )E[kV k k22 ] + σ 2 h2 p. Since h ≤ 2/M , Lemma 3 implies that E[k∆k+1 k22 ] ≤ (1 + t)γ 2 E[k∆k k22 ] + (1 + t−1 )(1.82)2 h3 M 2 p + σ 2 h2 p 1+γ 2 2 2 for every t > 0. Let us choose t = ( 1+γ 2γ ) − 1 so that (1 + t)γ = ( 2 ) . By recursion, this leads to  1 + γ 2(k+1) o  2 n W22 (νk+1 , π) ≤ σ 2 h2 p + (1 + t−1 )(1.82)2 h3 M 2 p . W22 (ν0 , π) + 2 1−γ In the case h ≤ 2/(m + M ), γ = 1 − mh and we get (1 + t−1 )h3 M 2 p = 1+γ 2 = 1 − 21 mh. Furthermore, (1 + γ)2 h3 M 2 p h2 M 2 p ≤ . (1 − γ)(1 + 3γ) m This readily yields  2hp 1/2 n  mh k+1 3.3M 2 o1/2 W2 (νk+1 , π) ≤ 1 − W2 (ν0 , π) + σ2 + . 2 m m Similarly, in the case h ≥ 2/(m + M ), γ = M h − 1 and we get (1 + t−1 )h3 M 2 p = 1+γ 2 = 12 M h. Furthermore, (1 + γ)2 h3 M 2 p h3 M 2 p 2h2 M p ≤ ≤ . (1 − γ)(1 + 3γ) 2 − Mh 2 − Mh This implies the inequality W2 (νk+1 , π) ≤  M h k+1 2 W2 (ν0 , π) + which completes the proof. 9  2h2 p 1/2 n 6.6M o1/2 σ2 + , 2 − Mh 2 − Mh DALALYAN 6.3. Proofs of lemmas Proof [Proof of Lemma 1] Since f is m-strongly convex, it satisfies the inequality
mM 1 ∆> ∇f (ϑ + ∆) − ∇f (ϑ) ≥ k∆k22 + k∇f (ϑ + ∆) − ∇f (ϑ)k22 , m+M m+M for all ∆, ϑ ∈ Rp . Therefore, simple algebra yields 2 2 k∆k − hU k k22 = k∆k k22 − 2h∆> k U k + h kU k k2
(k,h) = k∆k k22 − 2h∆> + ∆k ) − ∇f (ϑ(k,h) ) + h2 kU k k22 k ∇f (ϑ 2h 2hmM ≤ k∆k k22 − k∆k k22 − kU k k22 + h2 kU k k22 m+M m+M   2  2hmM  k∆k k22 + h h − kU k k22 . = 1− m+M m+M (16) Note that, thanks to the strong convexity of f , the inequality kU k k2 = k∇f (ϑ(k,h) + ∆k ) − ∇f (ϑ(k,h) )k2 ≥ mk∆k k2 is true. If h ≤ 2/(m+M ), this inequality can be combined with (16) to obtain k∆k − hU k k22 ≤ (1 − hm)2 k∆k k22 . Similarly, when h ≥ 2/(m+M ), we can use the Lipschitz property of ∇f to infer that kU k k2 ≤ M k∆k k2 . Combining with (16), this yields k∆k − hU k k22 ≤ (hM − 1)2 k∆k k22 , h ≥ 2/(m+M ). if Thus, we have checked that (14) is true for every h ∈ (0, 2/M ). Proof [Proof of Lemma 2] To simplify notations, we prove the lemma for p = 1. The function x 7→ f 0 (x) being Lipschitz continuous is almost surely differentiable. Furthermore, it is clear that |f 00 (x)| ≤ M for every x for which this second derivative exists. The result of (Rudin, 1987, Theorem 7.20) implies that Z x f 0 (x) − f 0 (0) = f 00 (y) dy. 0 Therefore, using f 0 (x) π(x) = −π 0 (x), we get Z Z Z Z x  f 0 (x)2 π(x) dx = f 0 (0) f 0 (x) π(x) dx + f 00 (y) dy f 0 (x) π(x) dx R R Z Z RZ x 0  0 0 = −f (0) π (x) dx − f 00 (y) dy π 0 (x) dx R R 0 Z ∞Z x Z 0 Z 0 00 0 =− f (y) π (x) dy dx + f 00 (y) π 0 (x) dy dx. 0 −∞ 0 In view of Fubini’s theorem, we arrive at Z Z ∞ Z 0 2 00 f (x) π(x) dx = f (y) π(y) dy + R 0 −∞ 0 10 x f 00 (y) π(y) dy ≤ M. F URTHER ANALOGY BETWEEN SAMPLING AND OPTIMIZATION This completes the proof. Proof [Proof of Lemma 3] Since the process L is stationary, V (a) has the same distribution as V (0). For this reason, it suffices to prove the claim of the lemma for a = 0 only. Using the Lipschitz continuity of f , we get E[kV (0)k22 ] h h Z =E Z ≤h 2i
∇f (Lt ) − ∇f (L0 ) dt 2 0 h  E ∇f (Lt ) − ∇f (L0 ) Z h  2 E Lt − L0 2 dt. ≤ hM 2 0 2 2 dt 0 Combining this inequality with the stationarity of Lt , we arrive at  E[kV 1/2 (0)k22 ]  ≤ hM 2 h Z E  − ∇f (Ls ) ds + ≤ hM 2 h Z E 0  ≤  = √ 0 0  t Z  Z t ∇f (Ls ) ds 0 hM 2 E ∇f (L0 )  2 2 1 4 2 h M E ∇f (L0 ) 3  Z h 2 dt 2 2 2 W t 2 dt 1/2  Z 2 + 2hpM 2 h 1/2 t dt 0 1/2  Z t2 dt + 2hpM 2 h 1/2 t dt 0 0 2 1/2 1/2
1/2 + h3 M 2 p . To complete the proof, it suffices to apply Lemma 2. Acknowledgments The work of the author was partially supported by the grant Investissements d’Avenir (ANR-11IDEX-0003/Labex Ecodec/ANR-11-LABX-0047). The author would like to thank Nicolas Brosse, who suggested an improvement in Theorem 3. References Y. Atchadé, G. Fort, E. Moulines, and P. Priouret. Adaptive Markov chain Monte Carlo: theory and methods. In Bayesian time series models, pages 32–51. Cambridge Univ. Press, Cambridge, 2011. R. N. Bhattacharya. Criteria for recurrence and existence of invariant measures for multidimensional diffusions. Ann. Probab., 6(4):541–553, 08 1978. S. Boyd and L. Vandenberghe. Convex optimization. Cambridge University Press, Cambridge, 2004. 11 DALALYAN S. Bubeck, R. Eldan, and J. Lehec. Sampling from a log-concave distribution with Projected Langevin Monte Carlo. ArXiv e-prints, July 2015. A. S. Dalalyan. Theoretical guarantees for approximate sampling from smooth and log-concave densities. ArXiv e-prints, December 2014. A. S. Dalalyan and A. B. Tsybakov. Sparse regression learning by aggregation and Langevin MonteCarlo. J. Comput. System Sci., 78(5):1423–1443, 2012. A. Durmus and E. Moulines. High-dimensional Bayesian inference via the Unadjusted Langevin Algorithm. ArXiv e-prints, May 2016. Alain Durmus, Eric Moulines, and Marcelo Pereyra. Sampling from convex non continuously differentiable functions, when Moreau meets Langevin. February 2016. URL https://hal. archives-ouvertes.fr/hal-01267115. L. Lovász and S. Vempala. Hit-and-run from a corner. SIAM J. Comput., 35(4):985–1005 (electronic), 2006a. L. Lovász and S. Vempala. Fast algorithms for logconcave functions: Sampling, rounding, integration and optimization. In 47th Annual IEEE Symposium on Foundations of Computer Science (FOCS 2006), 21-24 October 2006, Berkeley, California, USA, Proceedings, pages 57–68, 2006b. Walter Rudin. Real and complex analysis. McGraw-Hill Book Co., New York, third edition, 1987. 12 

Model Accuracy and Runtime Tradeoff in Distributed Deep Learning: A Systematic Study Suyog Gupta1,∗ , Wei Zhang1,∗ , Fei Wang2 IBM T. J. Watson Research Center. Yorktown Heights. NY. ∗ Equal Contribution. Department of Healthcare Policy and Research, Weill Cornell Medical College. New York City. NY 1 arXiv:1509.04210v3 [stat.ML] 5 Dec 2016 2 Abstract—Deep learning with a large number of parameters requires distributed training, where model accuracy and runtime are two important factors to be considered. However, there has been no systematic study of the tradeoff between these two factors during the model training process. This paper presents Rudra, a parameter server based distributed computing framework tuned for training large-scale deep neural networks. Using variants of the asynchronous stochastic gradient descent algorithm we study the impact of synchronization protocol, stale gradient updates, minibatch size, learning rates, and number of learners on runtime performance and model accuracy. We introduce a new learning rate modulation strategy to counter the effect of stale gradients and propose a new synchronization protocol that can effectively bound the staleness in gradients, improve runtime performance and achieve good model accuracy. Our empirical investigation reveals a principled approach for distributed training of neural networks: the mini-batch size per learner should be reduced as more learners are added to the system to preserve the model accuracy. We validate this approach using commonly-used image classification benchmarks: CIFAR10 and ImageNet. I. I NTRODUCTION Deep neural network based models have achieved unparalleled accuracy in cognitive tasks such as speech recognition, object detection, and natural language processing [18]. For certain image classification benchmarks, deep neural networks have been touted to even surpass human-level performance [13, 11]. Such accomplishments are made possible by the ability to perform fast, supervised training of complex neural network architectures using large quantities of labeled data. Training a deep neural network translates into solving a non-convex optimization problem in a very high dimensional space, and in the absence of a solid theoretical framework to solve such problems, practitioners are forced to rely on trial-and-error empirical observations to design heuristics that help obtain a well-trained model[1]. Naturally, fast training of deep neural network models can enable rapid evaluation of different network architectures and facilitate a more thorough hyper-parameter optimization for these models. Recent years have seen a resurgence of interest in deploying large-scale computing infrastructure designed specifically for training deep neural networks. Some notable efforts in this direction include distributed computing infrastructure using thousands of CPU cores [3, 6], high-end graphics processors (GPUs)[16], or a combination of CPUs and GPUs [4]. The large-scale deep learning problem can hence be viewed as a confluence of elements from machine learning (ML) and high-performance computing (HPC). Much of the work in the ML community is focused on non-convex optimization, model selection, and hyper-parameter tuning to improve the neural network’s performance (measured as classification accuracy) while working under the constraints of the computational resources available in a single computing node (CPU with or without GPU acceleration). From a HPC perspective, prior work has addressed, to some extent, the problem of accelerating the neural network training by mapping the asynchronous version of mini-batch stochastic gradient descent (SGD) algorithm onto multiple computing nodes. Contrary to the popular belief that asynchrony necessarily improves model accuracy, we find that adopting the approach of scaleout deep learning using asynchronous-SGD, gives rise to complex interdependencies between the training algorithm’s hyperparameters and the distributed implementation’s design choices (synchronization protocol, number of learners), ultimately impacting the neural network’s accuracy and the overall system’s runtime performance. In this paper we present Rudra, a parameter server based deep learning framework to study these interdependencies and undertake an empirical evaluation with public image classification benchmarks. Our key contributions are: 1) A systematic technique (vector clock) for quantifying the staleness of gradient descent parameter updates. 2) An investigation of the impact of the interdependence of training algorithm’s hyperparameters (mini-batch size, learning rate (gradient descent step size)) and distributed implementation’s parameters (gradient staleness, number of learners) on the neural network’s classification accuracy and training time. 3) A new learning rate tuning strategy that reduces the effect of stale parameter updates. 4) A new synchronization protocol to reduce network bandwidth overheads while achieving good classification accuracy and runtime performance. 5) An observation that to maintain a given level of model accuracy, it is necessary to reduce the mini-batch size as the number of learners is increased. This suggests a hard limit on the amount of parallelism that can be exploited in training a given model. II. BACKGROUND A neural network computes a parametric, non-linear transformation fθ : X 7→ Y , where θ represents a set of adjustable parameters (or weights). In a supervised learning context (such as image classification), X is the input image and Y corresponds to the label assigned to the image. A deep neural network organizes the parameters θ into multiple layers, each of which consists of a linear transformation followed by a nonlinear function such as sigmoid, tanh, etc. In a feed-forward deep neural network, the layers are arranged hierarchically such that the output of the layer l − 1 feeds into the input of layer l. The terminal layer generates the network’s output Ŷ = fθ (X), corresponding to the input X. A neural network training algorithm seeks to find a set of parameters θ∗ that minimizes the discrepancy between Ỹ and the ground truth Y . This is usually accomplished by defining a differentiable cost function C(Ŷ , Y ) and iteratively updating each of the model parameters using some variant of the gradient descent algorithm:   1 Xm C Yˆs , Ys , s=1 m  ∇θ(k) (t) = ∂Em /∂θ(k) (t), (1b) θ(k) (t + 1) = θ(k) (t) − α(t)∇θ(k) (t) (1c) Em = unique mini-batch. The model parallelism approach augments this framework by splitting the neural network model across multiple learners. With model parallelism, each learner trains only a portion of the network; edges that cross learner boundaries must be synchronized before gradients can be computed for the entire model. Several different synchronization strategies are possible. The most commonly used one is the asynchronous protocol, in which the learners work completely independently of each other and the parameter server. Section III will discuss three different synchronization strategies, each associated with a unique tradeoff between model accuracy and runtime. III. D ESIGN AND I MPLEMENTATION (1a) where θ(k) (t) represents the k th parameter at iteration t, α is the step size (also known as the learning rate) and m is the batch size. The batch gradient descent algorithm sets m to be equal to the total number of training examples N . Due to the large amount of training data, deep neural networks are typically trained using the Stochastic Gradient Descent (SGD), where the parameters are updated with a randomly selected training example (Xs , Ys ). The performance of SGD can be improved by computing the gradients using a minibatch containing m = µ  N training examples. Deep neural networks are generally considered hard to train [1, 10, 23] and the trained model’s generalization error depends strongly on hyperparameters such as the initializations, learning rates, mini-batch size, network architecture, etc. In addition, neural networks are prone to overfit the data. Regularization methods (e.g., weight decay and dropout) [16] applied during training have been shown to combat overfitting and reduce the generalization error. Scale-out deep learning: A typical implementation of distributed training of deep neural networks consists of a master (parameter server) that orchestrates the work among one or more slaves (learners). Each learner does the followings: 1) getMinibatch: Select randomly a mini-batch of examples from the training data. 2) pullWeights: Request the parameter server for the current set of weights/parameters. 3) calcGradient: Compute gradients based on the training error for the current mini-batch (equation 1b). 4) pushGradient: Send the computed gradients to the parameter server The parameter server maintains a global view of the model weights and performs the following functions: 1) sumGradients: Receive and accumulate the gradients from the learners. 2) applyUpdate: Multiply the accumulated gradient by the learning rate and update the weights (equation 1c) Learners exploit data parallelism by each maintaining a copy of the entire model, and training independently over a A. Terminology Throughout the paper, we use the following definitions: • Parameter Server: a server that holds the model weights. [22] describes a typical parameter server using a distributed key-value store to synchronize state between processes. The parameter server collects gradients from learners and updates the weights accordingly. • Learner: A computing process that can calculate weight updates (gradients). • µ: mini-batch size. • α: learning rate. • λ: number of learners. • Epoch: a pass through the entire training dataset. • Timestamp: we use a scalar clock [20] to represent weights timestamp tsi , starting from i = 0. Each weight update increments the timestamp by 1. The timestamp of a gradient is the same as the timestamp of the weight used to compute the gradient. • σ: staleness of the gradient. A gradient with timestamp tsi is pushed to the parameter server with current weight timestamp tsj , where tsj ≥ tsi . We define the staleness of this gradient σ as j − i. • hσi, average staleness of gradients. The timestamps of the set of n gradients that triggers the advancement of weights timestamp from tsi−1 to tsi form a vector clock [17] htsi1 , tsi2 , ..., tsin i, where max{i1 , i2 , ..., in } < i. The average staleness of gradients hσi is defined as: hσi = (i − 1) − mean(i1 , i2 , ..., in ) • Hardsync protocol: To advance weights timestamp from tsi to tsi+1 , each learner calculates exactly one minibatch and sends its gradient ∇θl to the parameter server. The parameter server averages the gradients and updates the weights according to Equation (3), then broadcasts the new weights to all learners. Staleness in the hardsync protocol is always zero. 1 Xλ (k) ∇θl l=1 λ θ(k) (i + 1) = θ(k) (i) − α∇θ(k) (i) ∇θ(k) (i) = • (2) (3) Async protocol: Each learner calculates the gradients and asynchronously pushes/pulls the gradients/weights W ' = W − α f (ΔW,...) W@E   Parameter  Server   Stats  Server   W@MB   ΔW @MB   TrainErr@MB   Learner   …   Learner   Mini-­‐batch@MB   Data  Server   (a) Rudra-adv architecture W:  Model  Weights   @MB:  Per  Mini-­‐batch   @E:  Per  Epoch   (b) Rudra-adv∗ architecture Fig. 2. Rudra-adv architecture Fig. 1. Rudra-base architecture to/from parameter server. The Async weight update rule is given by: (k) ∇θ(k) (i) = ∇θl , Ll ∈ L1 , ..., Lλ θ(k) (i + 1) = θ(k) (i) − α∇θ(k) (i) • (4) Gradient staleness may be hard to control due to the asynchrony in the system. [6] describe Downpour SGD, an implementation of the Async protocol for a commodity scale-out system in which the staleness can be as large as hundreds. n-softsync protocol: Each learner pulls the weights from the parameter server, calculates the gradients and pushes the gradients to the parameter server. The parameter server updates the weights after collecting at least c = b(λ/n)c gradients. The splitting parameter n can vary from 1 to λ. The n-softsync weight update rule is given by: c = b(λ/n)c 1 Xc (k) ∇θl , Lj ∈ L1 , ..., Lλ ∇θ (i) = l=1 c θ(k) (i + 1) = θ(k) (i) − α∇θ(k) (i) (k) (5) In Section V-A we will show that in a homogeneous cluster where each learner proceeds at roughly the same speed, the staleness of the model can be empirically bounded at 2n. Note that when n is equal to λ, the weight update rule at the parameter server is exactly the same as in Async protocol. B. Rudra-base System Architecture Figure 1 illustrates the parameter server design that we use to study the interplay of hyperparameter tuning and system scale-out factor. This system implements both hardsync and nsoftsync protocols. The arrows between each entity represent a (group of) MPI message(s), except the communication between Learner and Data Server, which is achieved by a global file system. We describe each entity’s role and its implementation below. Learner is a single-process multithreaded SGD solver. Before training each mini-batch, a learner pulls the weights and the corresponding timestamp from the parameter server. A learner reduces the pullWeights traffic by first inquiring the timestamp from the parameter server: if the timestamp is as old as the local weights’, then this learner does not pull the weights. After training the mini-batch, learner sends the gradients along with gradients’ timestamp to parameter server. The size of pull and push messages is the same as the model size plus the size of scalar timestamp equal to one. Data Server is hosted on IBM GPFS, a global file system. Each learner has an I/O thread, which prefetches the minibatch via random sampling prior to training. Prefetching is completely overlapped with the computation. Parameter Server is a multithreaded process, that accumulates gradients from each learner and applies update rules according to Equations (3–5). In this study, we implemented hardsync protocol and n-softsync protocol. Learning rate is configured differently in either protocol. Inphardsync protocol, the learning rate is multiplied by a factor λµ/B, where B is the batch size of the reference model. In the n-softsync protocol, the learning rate is multiplied by the reciprocal of staleness. We demonstrate in Section V-A that this treatment of learning rate in n-softsync can significantly improve the model accuracy. Parameter server records the vector clock of each weight update to keep track of the the average staleness. When a specified number of epochs are trained, parameter server shuts down each learner. Statistics Server is a multithreaded process that receives the training error from each learner and receives the model from the parameter server at the end of each epoch and tests the model. It monitors the model training quality. This architecture is non-blocking everywhere except for pushing up gradients and pushing down weights, which are blocking MPI calls (e.g. MPI_Send). Parameter server handles each incoming message one by one (the message handling itself is multithreaded). In this way, we can precisely control how each learner’s gradients are received and handled by the parameter server. The purpose of Rudra-base is to control the noise of the system, so that we can study the interplay of scaleout factor and the hyperparameter selection. For a moderatelysized dataset like CIFAR-10, Rudra-base shows good scale-out factor (see Section V-B). C. Rudra-adv and Rudra-adv∗ System Architecture To achieve high classification accuracy, the required model size may be quite large (e.g. hundreds of MBs). In many cases, to achieve best possible model accuracy, mini-batch size µ must be small, as we will demonstrate in Section V-B. In order to meet these requirements with acceptable performance, we implemented Rudra-adv and Rudra-adv∗ . Rudra-adv system architecture. Rudra-base clearly is not a scalable solution when the model gets large. Under ideal circumstances (see Section IV-A for our experimental hardware system specification), a single learner pushing a model of 300 MB (size of a typical deep neural network, see section IV-B) would take more than 10 ms to transfer this data. If 16 tasks are sending 300 MB to the same receiver and there is link contention, it would take over a second for the messages to be delivered. To alleviate the network traffic to parameter server, we build a parameter server group that forms a tree. We colocate each tree leaf node on the same node as the learners for which it is responsible. Each node in the parameter server group is responsible for averaging the gradients sent from its learners and relaying the averaged gradient to its parent. The root node in the parameter server group is responsible for applying weight update and broadcasting the updated weights. Each non-leaf node pulls the weights from its parent and responds to its children’s weight pulling requests. Rudra-adv can significantly improve performance compared to Rudrabase and manage to scale out to large model and small µ, while maintaining the control of the gradients’ staleness. Figure 2(a) illustrates the system architecture for Rudra-adv. Red boxes represent the parameter server group, in which the gradients are pushed and aggregated upwards. Green boxes represent learners, each learner pushes the gradient to its parameter server parent and receives weights from its parameter server parent. The key difference between Rudra-adv and a sharded parameter server design (e.g., Distbelief [6] and Adam [3]) is that the weights maintained in Rudra-adv have the same timestamps whereas shared parameter servers maintain the weights with different timestamps. Having consistent weights makes the analysis of hyperparameter/scale-out interplay much more tractable. Rudra-adv∗ system architecture. We built Rudra-adv∗ to further improve the runtime performance in two ways: Broadcast weights within learners. To further reduce the traffic to the parameter server group, we form a tree within all learners and broadcast the weights down this tree. In this way the network links to/from learners are also utilized. Asynchronous pushGradient and pullWeights. Ideally, one would use MPI non-blocking send calls to asynchronously send gradients and weights. However, depending on the MPI implementation, it is difficult to guarantee if the non-blocking send calls make progress in the background [9]. Therefore we open additional communication threads and use MPI blocking send calls in the threads. Each learner process runs two additional communication threads: the pullWeights thread and pushGradient thread. In this manner, computing can continue without waiting for the communication. Note that since we need to control µ (the smaller µ is, the better model converges, as we demonstrate in Section V-B), we must guarantee that the learner pushes each calculated gradient to the server. Alternatively, one could locally accrue gradients and send the sum, as in [6], however that will effectively increase µ. For this Implementation Rudra-base Rudra-adv Rudra-adv∗ Communication overlap (%) 11.52 56.75 99.56 TABLE I C OMMUNICATION OVERLAP MEASURED IN RUDRA - BASE , RUDRA - ADV , RUDRA - ADV∗ FOR AN ADVERSARIAL SCENARIO , WHERE THE MINI - BATCH SIZE IS THE SMALLEST POSSIBLE FOR 4- WAY MULTI - THREADED LEARNERS , MODEL SIZE 300MB, AND THERE ARE ABOUT 60 LEARENERS . reason, the pushGradient thread cannot start sending the current gradient before the previous one has been delivered. As demonstrated in Table I that as long as we can optimize the use of network links, this constraint has no bearing on the runtime performance, even when µ is extremely small. In contrast, pullWeights thread has no such constraint – we maintain a computation buffer and a communication buffer for pullWeights thread, and the communication always happens in the background. To use the newly received weights only requires a pointer swap. Figure 2(b) illustrates the system architecture for Rudra-adv∗ . Different from Rudra-adv, each learner continuously receives weights from the weights downcast tree, which consists of the top level parameter server node and all the learners. We measure the communication overlap by calculating the ratio between computation time and the sum of computation and communication time. Table I records the the communication overlap for Rudra-base, Rudra-adv, and Rudra-adv∗ in an adversarial scenario. Rudra-adv∗ can almost completely overlap computation with communication. Rudra-adv∗ can scale out to very large model size and work with smallest possible size of mini-batch. In Section V-E, we demonstrate Rudra-adv∗ ’s effectiveness in improving runtime performance while achieving good model accuracy. IV. M ETHODOLOGY A. Hardware and software environment We deploy the Rudra distributed deep learning framework on a P775 supercomputer. Each node of this system contains four eight-core 3.84 GHz POWER7 processors, one optical connect controller chip and 128 GB of memory. A single node has a theoretical floating point peak performance of 982 Gflop/s, memory bandwidth of 512 GB/s and bi-directional interconnect bandwidth of 192 GB/s. The cluster operating system is Red Hat Enterprise Linux 6.4. To compile and run Rudra we used the IBM xlC compiler version 12.1 with the -O3 -q64 -qsmp options, ESSL for BLAS subroutines, and IBM MPI (IBM Parallel Operating Environment 1.2). B. Benchmark datasets and neural network architectures To evaluate Rudra’s scale-out performance we employ two commonly used image classification benchmark datasets: CIFAR10 [15] and ImageNet [8]. The CIFAR10 dataset comprises of a total of 60,000 RGB images of size 32 × 32 pixels partitioned into the training set (50,000 images) and the test set (10,000 images). Each image belongs to one of the V. E VALUATION In this section we present results of evaluation of our scaleout deep learning training implementation. For an initial design 2−Softsync 60 2 Staleness (σ) 50 1.5 1−Softsync 1 Probability 70 2.5 Average staleness 〈σ 〉 10 classes, with 6000 images per class. For this dataset, we construct a deep convolutional neural network (CNN) with 3 convolutional layers each followed by a subsampling/pooling layer. The output of the 3rd pooling layer connects, via a fully-connected layer, to a 10-way softmax output layer that generates a probability distribution over the 10 output classes. This neural network architecture closely mimics the CIFAR10 model (cifar10 full.prototxt) available as a part of the opensource Caffe deep learning package [14]. The total number of trainable parameters in this network are ∼ 90 K, resulting in the model size of ∼350 kB when using 32-bit floating point data representation. The neural network is trained using momentum-accelerated mini-batch SGD with a batch size of 128 and momentum set to 0.9. As a data preprocessing step, the per-pixel mean is computed over the entire training dataset and subtracted from the input to the neural network. The training is performed for 140 epochs and results in a model that achieves 17.9% misclassification error rate on the test dataset. The base learning rate α0 is set to 0.001 are reduced by a factor of 10 after the 120th and 130th epoch. This learning rate schedule proves to be quite essential in obtaining the low test error of 17.9%. Our second benchmark dataset is collection of natural images used as a part of the 2012 edition of the ImageNet Large Scale Visual Recognition Challenge (ILSVRC 2012). The training set is a subset of the hand-labeled ImageNet database and contains 1.2 million images. The validation dataset has 50,000 images. Each image maps to one of the 1000 non-overlapping object categories. The images are converted to a fixed resolution of 256×256 to be used input to a deep convolution neural network. For this dataset, we consider the neural network architecture introduced in [16] consisting of 5 convolutional layers and 3 fully-connected layers. The last layer outputs the probability distribution over the 1000 object categories. In all, the neural network has ∼72 million trainable parameters and the total model size is 289 MB. The network is trained using momentum-accelerated SGD with a batch size of 256 and momentum set to 0.9. Similar to the CIFAR10 benchmark, per-pixel mean computed over the entire training dataset is subtracted from the input image feeding into the neural network. To prevent overfitting, a weight regularization penalty of 0.0005 is applied to all the layers in the network and a dropout of 50% is applied to the 1st and 2nd fully-connected layers. The initial learning rate α0 is set equal to 0.01 and reduced by a factor of 10 after the 15th and 25th epoch. Training for 30 epochs results in a top-1 error of 43.95% and top-51 error of 20.55% on the validation set. 0.2 0.1 0 0 20 40 60 Staleness (σ) 40 30 20 0.5 10 0 0 10 1 10 2 10 3 10 4 10 0 0 10 1 10 (a) top-5 error corresponds to the fraction of samples where the correct label does not appear in the top-5 labels considered most probable by the model 3 10 4 10 5 10 (b) Fig. 3. Average staleness hσi of the gradients as a function of the weight update step at the parameter server when using (a) 1-softsync, 2-softsync and (b) λ-softsync protocol. Inset in (b) shows the distribution of the gradient staleness values for λ-softsync protocol. Number of learners λ is set to 30. space exploration, we use the CIFAR10 dataset and Rudrabase system architecture. Subsequently we extend our findings to the ImageNet dataset using the Rudra-adv and Rudra-adv∗ system architectures. A. Stale gradients In the hardsync protocol introduced in section III-A, the transition from θ(i) to θ(i + 1) involves aggregating the gradients calculated with θ(i). As a result, each of the gradients ∇θl carries with it a staleness σ equal to 0. However, departing from the hardsync protocol towards either the n-softsync or the Async protocol inevitably adds staleness to the gradients, as a subset of the learners contribute gradients calculated using weights with timestamp earlier than the current timestamp of the weights at the parameter server. To measure the effect of gradient staleness when using the n-softsync protocol, we use the CIFAR10 dataset and train the neural network described in section IV-B using λ = 30 learners. For the 1-softsync protocol, the parameter server updates the current set of weights when it has received a total of 30 gradients from the learners. Similarly, the 2-softsync protocol forces the parameter server to accumulate λ/2 = 15 gradient contributions from the learners before updating the weights. As shown in Figure 3(a) the average staleness hσi for the 1-softsync and 2-softsync protocols remains close to 1 and 2, respectively. In the 1-softsync protocol, the staleness σLl for the gradients computed by the learner Ll takes values 0, 1, or 2, whereas σLl ∈ {0, 1, 2, 3, 4} for the 2-softsync protocol. Figure 3(b) shows the gradient staleness for the n-softsync protocol where n = λ = 30. In this case, the parameter server updates the weights after receiving a gradient from any of the learners. A large fraction of the gradients have staleness close to 30, and only with a very low probability (< 0.0001) does σ exceed 2n = 60. These measurements show that, in general, σLl ∈ {0, 1, . . . , 2n} and hσi = n for our implementation of the n-softsync protocol. Modifying the learning rate for stale gradients: In our experiments with the n-softsync protocol we found it beneficial, and at times necessary, to modulate the learning rate α to take into account the staleness of the gradients. For the n-softsync protocol, we set the learning rate as: α = α0 /hσi = α0 /n 1 The 2 10 Gradient step Gradient step (6) where α0 is the learning rate used for the baseline (control) experiment: µ = 128, λ = 1. Figure 4 shows a set of 100 90 α ← α0 Test error(%) 80 70 60 50 40 30 0 α← n=30, λ=30 n=4, λ=30 20 40 60 α0 hσi 80 100 120 140 Training epoch Fig. 4. Effect of learning rate modulation strategy: Dividing the learning rate by the average staleness aids in better convergence and achieves lower test error when using the n-softsync protocol. Number of learners, λ = 30; mini-batch size µ = 128. H p r o t o c o l 8 12 λ = 30 = 2 0 µ T e s t e rro r (% ) a r d s y n c ( 0 , 1 2 8 , 3 0 ) 2 5 ( 0 , 1 2 8 , 1 ) ( 0 , 4 , 3 0 ) 1 5 1 0 3 λ=1 µ=4 T r a in in g t im e 1 0 ( s ) (a) (b) Fig. 6. (σ, µ, λ) tradeoff curves for (a) λ-softsync protocol and (b) 1-softsync protocol. Shaded region in shows the region bounded by µ = 128, λ = 30, and µ = 4 contours for the hardsync protocol. λ ∈ {1, 2, 4, 10, 18, 30} and µ ∈ {4, 8, 16, 32, 64, 128}. Note that for λ = 1, n-softsync protocol degenerates to the hardsync protocol ( 0 , 4 , 1 ) 4 Fig. 5. (σ, µ, λ) tradeoff curves for the hardsync protocol. The dashed black line represents the 17.9% test error achieved by the baseline model (σ, µ, λ) = (0, 128, 1) on the CIFAR10 dataset. representative results illustrating the benefits of adopting this learning rate modulation strategy. We show the evolution of the test error on the CIFAR10 dataset as a function of the training epoch for two different configurations of the nsoftsync protocol (n = 4, n = 30) and set the number of learners, λ = 30. In both these configurations, setting the learning rate in accordance with equation (6) results in lower test error as compared with the cases where the learning rate is set to α0 . Surprisingly, the configuration 30-softsync, λ = 30, α = α0 fails to converge and shows a constant high error rate of 90%. Reducing the learning rate by a factor hσi = n = 30 makes the model with much lower test error2 . B. (σ, µ, λ) tradeoff curves Hyperparameter optimization plays a central role in obtaining good accuracy from neural network models [2]. For the SGD training algorithm, this includes a search over the neural network’s training parameters such as learning rates, weight regularization, depth of the network, mini-batch size etc. in order to improve the quality of the trained neural network model (quantified as the error on the validation dataset). Additionally, when distributing the training problem across multiple learners, parameters such as the number of learners and the synchronization protocol enforced amongst the learners impact not only the runtime of the algorithm but also the quality of the trained model. An exhaustive search over the space defined by these parameters for joint optimization of the runtime performance and the model quality can prove to be a daunting task even 2 Although not explored as a part of this work, it is certainly possible to implement a finer-grained learning rate modulation strategy that depends on the staleness of each of gradients computed by the learners instead of the average staleness. Such a strategy should apply smaller learning rates to staler gradients for a small model such as that used for the CIFAR10 dataset, and clearly intractable for models and datasets the scale of ImageNet. To develop a better understanding of the interdependence among the various tunable parameters in the distributed deep learning problem, we introduce the notion of (σ, µ, λ) tradeoff curves. A (σ, µ, λ) tradeoff curve plots the error on the validation set (or test set) and the total time to train the model (wall clock time) for different configurations of average gradient staleness hσi, mini-batch size per learner µ, and the number of learners λ. The configurations (σ, µ, λ) that achieve the virtuous combination of low test error and small training time are preferred and form ideal candidates for further hyperparameter optimization. We run3 the CIFAR10 benchmark for λ ∈ {1, 2, 4, 10, 18, 30} and µ ∈ {4, 8, 16, 32, 64, 128}. Figure 5 shows a set of (σ, µ, λ) curves for the hardsync protocol i.e. σ = 0. The baseline configuration with λ = 1 learner and mini-batch size µ = 128 achieves a test error of 17.9%. With p the exception of modifying the learning rate as α = α0 µλ/128, all the other neural network’s hyperparameters were kept unchanged from the baseline configuration while generating the data points for different values of µ and λ. Note that it is possible to achieve test error lower than the baseline by reducing the mini-batch size from 128 to 4. However, this configuration (indicated on the plot as (σ, µ, λ) = (0, 4, 1)) increases training time compared with the baseline. This is primarily due to the fact that the dominant computation performed by the learners involves multiple calls to matrix multiplication (GEMM) to compute W X where samples in a mini-batch form columns of the matrix X. Reducing the mini-batch size cause a proportionate decrease in the GEMM throughput and slower processing of the mini-batch by the learner. In Figure 5, the contour labeled µ = 128 is the configurations with the mini-batch size per learner is kept constant at 128 and the number of learners is varied from λ = 1 to λ = 30. 3 The mapping between λ and the number of computing nodes η is (λ, η) = {(1, 1), (2, 1), (4, 2), (10, 4), (18, 4), (30, 8)} The training time reduces monotonically with λ, albeit at the expense of an increase in the test error. Traversing along the λ = 30 contour from configuration (σ, µ, λ) = (0, 128, 30) to (σ, µ, λ) = (0, 4, 30) (i.e. reducing the mini-batch size from 128 to 4) helps restore much of this degradation in the test error by partially sacrificing the speed-up obtained by the virtue of scaling out to 30 learners. 3 5 3 5 H a rd λ- S o 1 -S o lin e a 3 0 2 5 µ = 128 n c y n c y n c p e e d u p H a rd λ- S o 1 -S o lin e a 3 0 2 5 2 0 S p e e d u p S p e e d u p 2 0 s y fts fts r s 1 5 1 0 s y fts fts r s µ=4 n c y n c y n c p e e d u p 1 5 1 0 5 5 0 0 0 5 1 0 1 5 2 0 N u m b e r o f L e a rn e rs , λ (a) 2 5 3 0 0 5 1 0 1 5 2 0 2 5 3 0 N u m b e r o f L e a rn e rs , λ (b) Fig. 7. Speed-up in the training time for mini-batch size and (a) µ = 128 (b) µ = 4 for 3 different protocols: hardsync, λ-softsync, and 1-softsync. Speedup numbers in (a) and (b) are calculated relative to (σ, µ, λ) = (0, 128, 1) and (σ, µ, λ) = (0, 4, 1), respectively. Figure 6(a) shows (σ, µ, λ) tradeoff curves for the λsoftsync protocol. In this protocol, the parameter server updates the weights as soon as it receives a gradient from any of the learners. Therefore, as shown in section V-A the average gradient staleness hσi = λ and σmax ≤ 2λ with high probability. The learning rate is set in accordance with equation 6. All the other hyperparameters are left unchanged from the baseline configuration. Qualitatively, the (σ, µ, λ) tradeoff curves for λ-softsync look similar to those observed for the hardsync protocol. In this case, however, the degradation in the test error relative to the baseline for the (σ, µ, λ) = (30, 128, 30) configuration is much more pronounced. As observed previously, this increase in the test error can largely be mitigated by reducing the size of mini-batch processed by each learner (λ = 30 contour). Note, however, the sharp increase in the training time for the configuration (σ, µ, λ) = (30, 4, 30) as compared with (σ, µ, λ) = (30, 128, 30). The smaller mini-batch size not only reduces the computational throughput at each learner, but also increases the frequency of pushGradient and pullWeights requests at the parameter server. In addition, small mini-batch size increases the frequency of weight updates at the parameter server. Since in the Rudra-base architecture (section III-B), the learner does not proceed with the computation on the next mini-batch till it has received the updated gradients, the traffic at the parameter server and the more frequent weight updates can delay servicing the learner’s pullWeights request, potentially stalling the gradient computation at the learner. Interestingly, all the configurations along the µ = 4 contour show similar, if not better, test error as the baseline. For these configurations, the average staleness varies between 2 and 30. From this empirical observation, we infer that a small mini-batch size per learner confers upon the training algorithm a fairly high degree of immunity to stale gradients. The 1-softsync protocol shows (σ, µ, λ) tradeoff curves (Figure 6(b)) that appear nearly identical to those observed for the λ-softsync protocol. In this case, the average staleness is 1 irrespective of the number of learners. Since the parameter server waits for λ gradients to arrive before updating the weights, there is a net reduction in the pullWeights traffic at the parameter server (see section III-B). As a result, the 1-softsync protocol avoids the degradation in runtime observed in the λ-softsync protocol for the configuration with µ = 4 and λ = 30. The distinction in terms of the runtime performance between these two protocols becomes obvious when comparing the speed-ups obtained for different minibatch sizes (Figure 7). For µ = 128, the 1-softsync and λ-softsync protocol demonstrate similar speed-ups over the baseline configuration for upto λ = 30 learners. In this case, the communication between the learners and the parameter server is sporadic enough to prevent the learners from waiting on the parameter server for updated weights. However, as the number of learners is increased beyond 30, the bottlenecks at the parameter server are expected to diminish the speed-up obtainable using the λ-softsync protocol. The effect of frequent pushGradient and pullWeights requests due to smaller at the parameter manifest clearly as the mini-batch size is reduced to 4, in which case, the λ-softsync protocol shows subdued speed-up compared with 1-softsync protocol. In either scenario, the hardsync protocol fares the worst in terms of runtime performance improvement when scaling out to large number of learners. The upside of adopting the hardsync protocol, however, is that it achieves substantially lower test error, even for large mini-batch sizes. C. µλ = constant In the hardsync protocol, given a configuration with λ learners and mini-batch size µ per learner, the parameter server averages the λ number of gradients reported to it by the learners. Using equations (1) and (3): ∇θ(k) (i) = =    ˆ µ λ λ 1 X  1 X ∂C Ys , Ys  1X (k) ∇θl = λ λ µ s=1 ∂θ(k) (i) l=1 l=1 l   ˆ µλ 1 X ∂C Ys , Ys (7) µλ s=1 ∂θ(k) (i) The last step equation (7) is valid since each training example (Xs , Ys ) is drawn independently from the training set and also because the hardsync protocol ensures that all the learners compute gradients on identical set of weights (k) i.e. θl (i) = θ(k) (i) ∀ l ∈ {1, 2, . . . , λ}. According to equation (7), the configurations (σ, µ, λ) = (0, µ0 λ0 , 1) and (σ, µ, λ) = (0, µ0 , λ0 ) are equivalent from the perspective of stochastic gradient descent optimization. In general, hardsync configurations with the same µλ product are expected to give nearly4 the same test error. In Table II we report the test error at the end of 140 epochs for configurations with µλ = constant. Interestingly, 4 small differences in the final test error achieved may arise due to the inherent nondeterminism of random sampling in stochastic gradient descent and the random initialization of the weights. σ µ λ Test error Training time(s) σ µ λ Synchronization protocol Test error Training time(s) 1 30 18 10 4 2 4 4 8 16 32 64 30 30 18 10 4 2 18.09% 18.41% 18.92% 18.79% 18.82% 17.96% 1573 2073 2488 3396 7776 13449 1 0 30 0 18 4 8 4 4 8 30 30 30 30 18 1-softsync Hardsync 30-softsync Hardsync 18-softsyc 18.09% 18.56% 18.41% 18.15% 18.92% 1573 1995 2073 2235 2488 µλ ≈ 256 1 30 18 10 4 2 1 8 8 16 32 64 128 128 30 30 18 10 4 2 2 20.04% 19.65% 20.33% 20.82% 20.70% 19.52% 19.59% 1478 1509 2938 3518 6631 11797 11924 µλ ≈ 512 1 30 18 10 4 16 16 32 64 128 30 30 18 10 4 23.25% 22.14% 23.63% 24.08% 23.01% 1469 1502 2255 2683 7089 µλ ≈ 128 tiny 1 30 18 1 10 32 30 27.16% 1299 32 30 27.27% 1420 64 18 28.31% 1713 µλ ≈ 1024 128 10 29.83% 2551 128 10 29.90% 2626 TABLE II Results on CIFAR10 benchmark: T EST ERROR AT THE END OF 140 EPOCHS AND TRAINING TIME FOR (σ, µ, λ) CONFIGURATIONS WITH µλ = CONSTANT . we find that even for the n-softsync protocol, configurations that maintain µλ = constant achieve comparable test errors. At the same time, the test error turns out to be rather independent of the staleness in the gradients for a given µλ product. For instance, Table II shows that when µλ ≈ 128, but the (average) gradient staleness is allowed to vary between 1 and 30, the test error stays ∼18-19%. Although this result may seem counterintuitive, a plausible explanation emerges when considering the measurements shown earlier in Figure 3, that our implementation of the n-softsync protocol achieves an average gradient staleness of n while bounding the maximum staleness at 2n. Consequently, at any stage in the gradient descent algorithm, the weights being used by the different learners (θl (t)) do not differ significantly and can be considered to be approximately the same. The quality of this approximation improves when each update θ(k) (i + 1) = θ(k) (i) − α∇θ(k) (i) creates only a small displacement in the weight space. This can be controlled by suitably tuning the learning rate α. Qualitatively, the learning rate should decrease as the staleness in the system increases in order to reduce the divergence across the weights seen by the learners. The learning rate modulation of equation (6) achieves precisely this effect. These results help define a principled approach for distributed training of neural networks: the mini-batch size per learner should be reduced as more learners are added to the system in way that keeps µλ product constant. In addition, the learning rate should be modulated to account for stale gradients. In Table II, 1-softsync (σ = 1) protocol invariably TABLE III Results on CIFAR10 benchmark: T OP -5 BEST (σ, µ, λ) CONFIGURATIONS THAT ACHIEVE A COMBINATION OF LOW TEST ERROR AND SMALL TRAINING TIME . shows the smallest training time for any µλ. This is expected, since the 1-softsync protocol minimizes the traffic at the parameter server. Table II also shows that the test error increases monotonically with the µλ product. These results reveal the scalability limits under the constraints of preserving the model accuracy. Since the smallest possible mini-batch size is 1, the maximum number of learners is bounded. This upper bound on the maximum number of learners can be relaxed if an inferior model is acceptable. Alternatively, it may be possible to reduce the test error for higher µλ by running for more number of epochs. In such a scenario, adding more learners to the system may give diminishing improvements in the overall runtime. From machine learning perspective, this points to an interesting research direction on designing optimization algorithm and learning strategies that perform well with large mini-batch sizes. D. Summary of results on CIFAR10 benchmark Table III summarizes the results obtained on the CIFAR10 dataset using the Rudra-base system architecture. As a reference for comparison, the baseline configuration (σ, µ, λ) = (0, 128, 1) achieves a test error of 17.9% and takes 22,392 seconds to finish training 140 epochs. E. Results on ImageNet benchmark The large model size of the neural network used for the ImageNet benchmark and the associated computational cost of training this model prohibits an exhaustive state space exploration. The baseline configuration (µ = 256, λ = 1) takes 54 hours/epoch. Guided by the results of section V-C, we first consider a configuration with µ = 16, λ = 18 and employ the Rudra-base architecture with hardsync protocol (base-hardsync). This configuration performs training at the speed of ∼330 minutes/epoch and achieves a top-5 error of 20.85%, matching the accuracy of the baseline configuration (µ = 256, λ = 1, section IV-B). The synchronization overheads associated with the hardsync protocol deteriorate the runtime performance and the training speed can be further improved by switching over to the 1softsync protocol. Training using the 1-softsync protocol with mini-batch size of µ = 16 and 18 learners takes ∼270 minutes/epoch, reaching a top-1 (top-5) accuracy of 45.63% (22.08%) by the end of 30 epochs (base-softsync). For this particular benchmark, the training setup for the 1-softsync protocol differs from the hardsync protocol in certain subtle, Configuration Architecture µ λ Synchronization protocol Validation error(top-1) Validation error (top-5) Training time (minutes/epoch) base-hardsync base-softsync adv-softsync adv∗ -softsync Rudra-base Rudra-base Rudra-adv Rudra-adv∗ 16 16 4 4 18 18 54 54 Hardsync 1-softsync 1-softsync 1-softsync 44.35% 45.63% 46.09% 46.53% 20.85% 22.08% 22.44% 23.38% 330 270 212 125 TABLE IV Results on ImageNet benchmark: VALIDATION ERROR AT THE END OF 30 EPOCHS AND TRAINING TIME PER EPOCH FOR DIFFERENT CONFIGURATIONS . 8 5 b a s b a s a d v a d v 8 0 V a lid a tio n e r r o r ( to p - 1 ) ( % ) but important ways. First, we use an adaptive learning rate method (AdaGrad [7, 6]) to improve the stability of SGD when training using the 1-softsync protocol. Second, to improve convergence we adopt the strategy of warmstarting [21] the training procedure by initializing the network’s weights from a model trained with hardsync for 1 epoch. Further improvement in the runtime performance may be obtained by adding more learners to the system. However, as observed in the previous section, increase in the number of learners needs to be accompanied by a reduction in the mini-batch size to prevent degradation in the accuracy of the trained model. The combination of a large number of learners and a small mini-batch size represents a scenario where the Rudra-base architecture may suffer from a bottleneck at the parameter server due to the frequent pushGradient and pullWeights requests. These effects are expected to be more pronounced for large models such as ImageNet. The Rudra-adv architecture alleviates these bottlenecks, to some extent, by implementing a parameter server group organized in a tree structure. λ = 54 learners, each processing a mini-batch size µ = 4 trains at ∼212 minutes/epoch when using Rudraadv architecture and 1-softsync protocol (adv-softsync). As in the case of Rudra-base, the average staleness in the gradients is close to 1 and this configuration also achieves a top-1(top-5) error of 46.09%(22.44%). The Rudra-adv∗ architecture improves the runtime further by preventing the computation at the learner from stalling on the parameter server. However, this improvement in performance comes at the cost of increasing the average staleness in the gradients, which may deteriorate the quality of the trained model. The previous configuration runs at ∼125 minutes/epoch, but suffers an increase in the top-1 validation error (46.53%) when using Rudra-adv∗ architecture (adv∗ softsync). Table IV summarizes the results obtained for the 4 configurations discussed above. It is worth mentioning that the configuration µ = 8, λ = 54 performs significantly worse, producing a model that gives top-1 error of > 50% but trains at a speed of ∼96 minutes/epoch. This supports our observation that scaling out to large number of learners must be accompanied by reducing the mini-batch size per learner so the quality of the trained model can be preserved. Figure 8 compares the evolution of the top-1 validation error during training for the 4 different configuration summarized in Table IV. The training speed follows the order adv∗ -softsync > adv-softsync > base-softsync > base-hardsync. As a result, adv∗ -softsync is the first configuration to hit the 48% validation 7 5 7 0 6 5 6 0 a rd o fts fts y o fts s y n c y n c n c y n c + 5 5 5 0 4 8 .0 % 4 5 4 0 e -h e -s -s o *-s 0 4 3 .9 5 % 2 0 4 0 6 0 8 0 1 0 0 1 2 0 T r a in in g tim e ( h o u r s ) 1 4 0 1 6 0 1 8 0 Fig. 8. Results on ImageNet benchmark: Error on the validation set as a function of training time for the configurations listed in Table IV error mark. Configurations other than base-hardsync show marginally higher validation error compared with the baseline. As mentioned earlier, the experiments with 1-softsync protocol use AdaGrad to achieve stable convergence. It is welldocumented [24, 21] that AdaGrad is sensitive to the initial setting on the learning rates. We speculate that tuning the initial learning rate can help recover the slight loss of accuracy for the 1-softsync runs. VI. R ELATED W ORKS To accelerate training of deep neural networks and handle large dataset and model size, many researchers have adopted GPU-based solutions, either single-GPU [16] or multiGPU [26] GPUs provide enormous computing power and are particularly suited for the matrix multiplications which are the core of many deep learning implementations. However, the relatively limited memory available on GPUs may restrict their applicability to large model sizes. Distbelief [6] pioneered scale-out deep learning on CPUs. Distbelief is built on tens of thousands of commodity PCs and employs both model parallelism via dividing a single model between learners, and data parallelism via model replication. To reduce traffic to the parameter server, Distbelief shards parameters over a parameter server group. Learners asynchronously push gradients and pull weights from the parameter server. The frequency of communication can be tuned via npush and nfetch parameters. More recently, Adam [3] employs a similar system architecture to DistBelief, while improving on Distbelief in two respects: (1) better system tuning, e.g. customized concurrent memory allocator, better linear algebra library implementation, and passing activation and error gradient vector instead of the weights update; and (2) leveraging the recent improvement in machine learning, in particular convolutional neural network to achieve better model accuracy. In any parameter server based deep learning system [12], staleness will negatively impact model accuracy. Orthogonal to the system design, many researchers have proposed solutions to counter staleness in the system, such as bounding the staleness in the system [5] or changing optimization objective function, such as elastic averaging SGD [25]. In this paper, we empirically study how staleness affects the model accuracy and discover the heuristics that smaller mini-batch size can effectively counter system staleness. In our experiments, we derive this heuristics from a small problem size(e.g., CIFAR10) and we find this heuristic is applicable even to much larger problem size (e.g., ImageNet). Our finding coincides with a very recent theoretical paper [19], in which the authors prove that in order to achieve linear speedup using asynchronous protocol while maintaining good model accuracy, one needs to increase the number of weight updates conducted at the parameter server. In our system, this theoretical finding is equivalent to keeping constant number of training epochs while reducing the mini-batch size. To the best of our knowledge, our work is the first systematic study of the tradeoff between model accuracy and runtime performance for distributed deep learning. VII. C ONCLUSION In this paper, we empirically studied the interplay of hyperparameter tuning and scale-out in three protocols for communicating model weights in asynchronous stochastic gradient descent. We divide the learning rate by the average staleness of gradients, resulting in faster convergence and lower test error. Our experiments show that the 1-softsync protocol (in which the parameter server accumulates λ gradients before updating the weights) minimizes gradient staleness and achieves the lowest runtime for a given test error. We found that to maintain a model accuracy, it is necessary to reduce the mini-batch size as the number of learners is increased. This suggests an upper limit on the level of parallelism that can be exploited for a given model, and consequently a need for algorithms that permit training over larger batch sizes. ACKNOWLEDGEMENT The work of Fei Wang is partially supported by National Science Foundation under Grant Number IIS-1650723. R EFERENCES [1] Y. Bengio. Practical recommendations for gradient-based training of deep architectures. In Neural Networks: Tricks of the Trade, pages 437–478. Springer, 2012. [2] T. M. Breuel. The effects of hyperparameters on SGD training of neural networks. arXiv:1508.02788, 2015. [3] T. Chilimbi, Y. Suzue, J. Apacible, and K. Kalyanaraman. Project Adam: Building an efficient and scalable deep learning training system. OSDI’14, pages 571–582, 2014. [4] A. Coates, B. Huval, T. Wang, D. Wu, B. Catanzaro, and N. Andrew. Deep learning with cots hpc systems. In Proceedings of the 30th ICML, pages 1337–1345, 2013. [5] H. e. a. Cui. Exploiting bounded staleness to speed up big data analytics. In USENIX ATC’14, pages 37–48, 2014. [6] J. Dean, G. S. Corrado, R. Monga, K. Chen, M. Devin, Q. V. Le, M. Z. Mao, M. Ranzato, A. Senior, P. Tucker, K. Yang, and A. Y. Ng. Large scale distributed deep networks. In NIPS, 2012. [7] J. Duchi, E. Hazan, and Y. Singer. Adaptive subgradient methods for online learning and stochastic optimization. The Journal of Machine Learning Research, 12:2121–2159, 2011. [8] O. R. et al. ImageNet Large Scale Visual Recognition Challenge. IJCV, pages 1–42, 2015. [9] M. P. I. Forum. Mpi 3.0 standard. www.mpiforum.org/docs/mpi-3.0/mpi30-report.pdf, 2012. [10] X. Glorot and Y. Bengio. Understanding the difficulty of training deep feedforward neural networks. In AISTATS, pages 249–256, 2010. [11] K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. arXiv preprint arXiv:1502.01852, 2015. [12] Q. Ho, J. Cipar, H. Cui, S. Lee, J. K. Kim, P. B. Gibbons, G. A. Gibson, G. Ganger, and E. P. Xing. More effective distributed ML via a stale synchronous parallel parameter server. In NIPS 26, pages 1223–1231. 2013. [13] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167, 2015. [14] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrell. Caffe: Convolutional architecture for fast feature embedding. arXiv preprint arXiv:1408.5093, 2014. [15] A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Computer Science Department, University of Toronto, Tech. Rep, 1(4):7, 2009. [16] A. Krizhevsky, I. Sutskever, and G. E. Hinton. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, pages 1097–1105, 2012. [17] L. Lamport. Time, clocks, and the ordering of events in a distributed system. Commun. ACM, 21(7):558–565, 1978. [18] Y. LeCun, Y. Bengio, and G. Hinton. Deep learning. Nature, 521(7553):436–444, 2015. [19] X. Lian, Y. Huang, Y. Li, and J. Liu. Asynchronous Parallel Stochastic Gradient for Nonconvex Optimization. ArXiv eprints, June 2015. [20] M. Raynal and M. Singhal. Logical time: Capturing causality in distributed systems. Computer, 29(2):49–56, 1996. [21] A. Senior, G. Heigold, M. Ranzato, and K. Yang. An empirical study of learning rates in deep neural networks for speech recognition. In ICASSP, pages 6724–6728. IEEE, 2013. [22] A. Smola and S. Narayanamurthy. An architecture for parallel topic models. Proc. VLDB Endow., 3(1-2):703–710, 2010. [23] I. Sutskever, J. Martens, G. Dahl, and G. Hinton. On the importance of initialization and momentum in deep learning. In Proceedinge of the 30th ICML, pages 1139–1147, 2013. [24] M. D. Zeiler. Adadelta: An adaptive learning rate method. arXiv preprint arXiv:1212.5701, 2012. [25] S. Zhang, A. Choromanska, and Y. LeCun. Deep learning with elastic averaging SGD. CoRR, abs/1412.6651, 2014. [26] Y. Zou, X. Jin, Y. Li, Z. Guo, E. Wang, and B. Xiao. Mariana: Tencent deep learning platform and its applications. Proc. VLDB Endow., 7(13):1772–1777, 2014. 

The Likelihood Ratio Test in High-Dimensional Logistic Regression Is Asymptotically a Rescaled Chi-Square arXiv:1706.01191v1 [math.ST] 5 Jun 2017 Pragya Sur∗ Yuxin Chen† Emmanuel J. Candès∗‡ June 2017 Abstract Logistic regression is used thousands of times a day to fit data, predict future outcomes, and assess the statistical significance of explanatory variables. When used for the purpose of statistical inference, logistic models produce p-values for the regression coefficients by using an approximation to the distribution of the likelihood-ratio test. Indeed, Wilks’ theorem asserts that whenever we have a fixed number p of variables, twice the log-likelihood ratio (LLR) 2Λ is distributed as a χ2k variable in the limit of large sample sizes n; here, χ2k is a chi-square with k degrees of freedom and k the number of variables being tested. In this paper, we prove that when p is not negligible compared to n, Wilks’ theorem does not hold and that the chi-square approximation is grossly incorrect; in fact, this approximation produces p-values that are far too small (under the null hypothesis). Assume that n and p grow large in such a way that p/n → κ for some constant κ < 1/2. We prove that d for a class of logistic models, the LLR converges to a rescaled chi-square, namely, 2Λ → α(κ)χ2k , where the scaling factor α(κ) is greater than one as soon as the dimensionality ratio κ is positive. Hence, the LLR is larger than classically assumed. For instance, when κ = 0.3, α(κ) ≈ 1.5. In general, we show how to compute the scaling factor by solving a nonlinear system of two equations with two unknowns. Our mathematical arguments are involved and use techniques from approximate message passing theory, from non-asymptotic random matrix theory and from convex geometry. We also complement our mathematical study by showing that the new limiting distribution is accurate for finite sample sizes. Finally, all the results from this paper extend to some other regression models such as the probit regression model. Keywords. Logistic regression, likelihood-ratio tests, Wilks’ theorem, high-dimensionality, goodness of fit, approximate message passing, concentration inequalities, convex geometry, leave-one-out analysis 1 Introduction Logistic regression is by far the most widely used tool for relating a binary response to a family of explanatory variables. This model is used to infer the importance of variables and nearly all standard statistical softwares have inbuilt packages for obtaining p-values for assessing the significance of their coefficients. For instance, one can use the snippet of R code below to fit a logistic regression model from a vector y of binary responses and a matrix X of covariates: fitted <- glm(y ~ X+0, family = ‘binomial’) pvals <- summary(fitted)$coefficients[,4] ∗ Department of Statistics, Stanford University, Stanford, CA 94305, U.S.A. of Electrical Engineering, Princeton University, Princeton, NJ 08544, U.S.A. ‡ Department of Mathematics, Stanford University, Stanford, CA 94305, U.S.A. † Department 1 The vector pvals stores p-values for testing whether a variable belongs to a model or not, and it is well known that the underlying calculations used to produce these p-values can also be used to construct confidence intervals for the regression coefficients. Since logistic models are used hundreds of times every day for inference purposes, it is important to know whether these calculations—e.g. these p-values—are accurate and can be trusted. 1.1 Binary regression Imagine we have n samples of the form (yi , Xi ), where yi ∈ {0, 1} and Xi ∈ Rp . In a generalized linear model, one postulates the existence of a link function µ(·) relating the conditional mean of the response variable to the linear predictor Xi> β, E[yi |Xi ] = µ(Xi> β), (1) where β = [β1 , β2 , . . . , βp ]> ∈ Rp is an unknown vector of parameters. We focus here on the two most commonly used binary regression models, namely, the logistic and the probit models for which ( et /(1 + et ) in the logistic model, µ(t) := (2) Φ(t) in the probit model; here, Φ is the cumulative distribution function (CDF) of a standard normal random variable. In both cases, the Symmetry Condition µ(t) + µ(−t) = 1 (3) holds, which says that the two types yi = 0 and yi = 1 are treated in a symmetric fashion. Assuming that the observations are independent, the negative log-likelihood function is given by [1, Section 4.1.2]    n  X µi + log (1 − µi ) , µi := µ(Xi> β). yi log ` (β) := − 1 − µ i i=1 Invoking the symmetry condition, a little algebra reveals an equivalent expression Xn
` (β) := ρ − ỹi Xi> β , i=1 where ( 1 if yi = 1, ỹi := −1 if yi = 0, ( and ρ(t) := log (1 + et ) in the logistic case, − log Φ (−t) in the probit case. (4) (5) Throughout we refer to this function ρ as the effective link. 1.2 The likelihood-ratio test and Wilks’ phenomenon Researchers often wish to determine which covariates are of importance, or more precisely, to test whether the jth variable belongs to the model or not: formally, we wish to test the hypothesis Hj : βj = 0 versus βj 6= 0. (6) Arguably, one of the most commonly deployed techniques for testing Hj is the likelihood-ratio test (LRT), which is based on the log-likelihood ratio (LLR) statistic

Λj := ` β̂(−j) − ` β̂ . (7) Here, β̂ and β̂(−j) denote respectively the maximum likelihood estimates (MLEs) under the full model and the reduced model on dropping the jth predictor; that is, β̂ = arg minp `(β) β∈R β̂(−j) = arg and 2 min β∈Rp ,βj =0 `(β). Inference based on such log-likelihood ratio statistics has been studied extensively in prior literature [12, 37, 52]. Arguably, one of the most celebrated results in the large-sample regime is the Wilks’ theorem. To describe the Wilk’s phenomenon, imagine we have a sequence of observations (yi , Xi ) where yi ∈ {0, 1}, Xi ∈ Rp with p fixed. Since we are interested in the limit of large samples, we may want to assume that the covariates are i.i.d. drawn from some population with non-degenerate covariance matrix so that the problem is fully p-dimensional. As before, we assume a conditional logistic model for the response. In this setting, Wilks’ theorem [52] calculates the asymptotic distribution of Λj (n) when n grows to infinity: (Wilks’ phenomenon) Under suitable regularity conditions which, for instance, guarantee that the MLE exists and is unique,1 the LLR statistic for testing Hj : βj = 0 vs. βj 6= 0 has asymptotic distribution under the null given by d 2Λj (n) → χ21 , as n → ∞. (8) This fixed-p large-n asymptotic result, which is a consequence of asymptotic normality properties of the MLE [50, Theorem 5.14], applies to a much broader class of testing problems in parametric models; for instance, it applies to the probit model as well. We refer the readers to [34, Chapter 12] and [50, Chapter 16] for a thorough exposition and details on the regularity conditions under which Wilks’ theorem holds. Finally, there is a well-known extension which states that if we were to drop k variables from the model, then the LLR would converge to a chi-square distribution with k degrees of freedom under the hypothesis that the reduced model is correct. 1.3 Inadequacy of Wilks’ theorem in high dimensions The chi-square approximation to the distribution of the LLR statistic is used in standard statistical softwares to provide p-values for the single or multiple coefficient likelihood ratio tests. Here, we perform a simple experiment on synthetic data to study the accuracy of the chi-square approximation when p and n are both decently large. Specifically, we set β = 0 and test β1 = 0 vs. β1 6= 0 using the LRT in a setting i.i.d. where p = 1200. In each trial, n = 4000 observations are produced with yi ∼ Bernoulli(1/2), and X := [X1 , · · · , Xn ]> ∈ Rn×p is obtained by generating a random matrix composed of i.i.d. N (0, 1) entries. We fit a logistic regression of y on X using R, and extract the p-values for each coefficient. Figure 1 plots the pooled histogram that aggregates 4.8 × 105 p-values in total (400 trials with 1200 p-values obtained in each trial). If the χ21 approximation were true, then we would expect to observe uniformly distributed p-values. The histrogram from Fig. 1 is, however, far from uniform. This is an indication of the inadequacy of Wilks’ theorem when p and n are both large. The same issue was also reported in [11], where the authors observed that this discrepancy is highly problematic since the distribution is skewed towards smaller values. Hence, such p-values cannot be trusted to construct level-α tests and the problem is increasingly severe when we turn attention to smaller p-values as in large-scale multiple testing applications. 1.4 The Bartlett correction? A natural question that arises immediately is whether the observed discrepancy could be an outcome of a finite-sample effect. It has been repeatedly observed that the chi-square approximation does not yield accurate results with finite sample size. One correction to the LRT that is widely used in finite samples is the Bartlett correction, which dates back to Bartlett [5] and has been extensively studied over the past few decades (e.g. [8, 10, 13, 15, 33]). In the context of testing for a single coefficient in the logistic model, this correction can be described as follows [38]: compute the expectation of the LLR statistic up to terms of order 1/n2 ; that is, compute a parameter α such that   1 α , E[2Λj ] = 1 + + O n n2 1 Such conditions would also typically imply asymptotic normality of the MLE. 3 12500 30000 15000 10000 7500 Counts Counts Counts 20000 10000 10000 5000 5000 2500 0 0 0.00 0.25 0.50 P−Values 0.75 1.00 0 0.00 (a) 0.25 0.50 P−Values 0.75 1.00 (b) 0.00 0.25 0.50 P−Values 0.75 1.00 (c) Figure 1: Histogram of p-values for logistic regression under i.i.d. Gaussian design, when β = 0, n = 4000, p = 1200, and κ = 0.3: (a) classically computed p-values; (b) Bartlett-corrected p-values; (c) adjusted p-values. which suggests a corrected LLR statistic 2Λj 1 + αnn (9) with αn being an estimator of α. With a proper choice of αn , one can ensure     2Λj 1 E =1+O 1 + αnn n2 in the classical setting where p is fixed and n diverges. In expectation, this corrected statistic is closer to a χ21 distribution than the original LLR for finite samples. Notably, the correction factor may in general be a function of the unknown β and, in that case, must be estimated from the null model via maximum likelihood estimation. In the context of GLMs, Cordeiro [13] derived a general formula for the Bartlett corrected LLR statistic, see [14, 17] for a detailed survey. In the case where there is no signal (β = 0), one can compute αn for the logistic regression model following [13] and [38], which yields αn =

 n 2 Tr Dp2 − Tr Dp−1 . 2 (10)
−1 > > Here, Dp is the diagonal part of X(X > X)−1 X > and Dp−1 is that of X(−j) X(−j) X(−j) X(−j) in which X(−j) is the design matrix X with the jth column removed. Comparing the adjusted LLRs to a χ21 distribution yields adjusted p-values. In the setting of Fig. 1(a), the histogram of Bartlett corrected p-values is shown in Fig. 1(b). As we see, these p-values are still far from uniform. If the mismatch is not due to finite sample-size effects, what is the distribution of the LLR in high dimensions? Our main contribution is to provide a very precise answer to this question; below, we derive the high-dimensional asymptotic distribution of the log-likelihood ratios, i.e. in situations where the dimension p is not necessarily negligible compared to the sample size n. 4 2 Main results 2.1 Modelling assumptions In this paper, we focus on the high-dimensional regime where the sample size is not much larger than the number of parameters to be estimated—a setting which has attracted a flurry of activity in recent years. In particular, we assume that the number p(n) of covariates grows proportionally with the number n of observations; that is, p(n) = κ, (11) lim n→∞ n where κ > 0 is a fixed constant independent of n and p(n). In fact, we shall also assume κ < 1/2 for both the logistic and the probit models, as the MLE is otherwise at ∞; see Section 2.2. To formalize the notion of high-dimensional asymptotics when both n and p(n) diverge, we consider a sequence of instances {X(n), y(n)}n≥0 such that for any n, • X(n) ∈ Rn×p(n) has i.i.d. rows Xi (n) ∼ N (0, Σ), where Σ ∈ Rp(n)×p(n) is positive definite;
ind. • yi (n) | X(n) ∼ yi (n) | Xi (n) ∼ Bernoulli µ(Xi (n)> β(n)) , where µ satisfies the Symmetry Condition; • we further assume β(n) = 0. From the Symmetry Condition it follows that µ(0) = 1/2, which directly implies that y(n) is a vector with i.i.d Bernoulli(1/2) entries. The MLE is denoted by β̂(n) and there are p(n) LLR statistics Λj (n) (1 ≤ j ≤ p(n)), one for each of the p(n) regression coefficients. In the sequel, the dependency on n shall be suppressed whenever it is clear from the context. 2.2 When does the MLE exist? Even though we are operating in the regime where n > p, the existence of the MLE cannot be guaranteed for all p and n. Interestingly, the norm of the MLE undergoes a sharp phase transition in the sense that kβ̂k = ∞ if κ > 1/2 and kβ̂k < ∞ if κ < 1/2. Here, we develop some understanding about this phenomenon. Given that ρ(t) ≥ ρ(−∞) = 0 for both the logistic and probit models, each summand in (4) is minimized if ỹi Xi> β = ∞, which occurs when sign(Xi> β) = sign(ỹi ) and kβk = ∞. As a result, if there exists a nontrivial ray β such that Xi> β > 0 if ỹi = 1 and Xi> β < 0 if ỹi = −1 (12) for any 1 ≤ i ≤ n, then pushing kβk to infinity leads to an optimizer of (4). In other words, the solution to (4) becomes unbounded (the MLE is at ∞) whenever there is a hyperplane perfectly separating the two sets of samples {i | ỹi = 1} and {i | ỹi = −1}. Under the assumptions from Section 2.1, ỹi is independent of X and the distribution of X is symmetric. Hence, to calculate the chance that there exists a separating hyperplane, we can assume ỹi = 1 (1 ≤ i ≤ n) without loss of generality. In this case, the event (12) becomes {Xβ | β ∈ Rp } ∩ Rn++ 6= ∅, (13) where Rn++ is the positive orthant. Write X = ZΣ1/2 so that Z is an n × p matrix with i.i.d. standard Gaussian entries, and θ = Σ1/2 β. Then the event (13) is equivalent to {Zθ | θ ∈ Rp } ∩ Rn++ 6= ∅. 5 (14) Now the probability that (14) occurs is the same as that {Zθ | θ ∈ Rp } ∩ Rn+ 6= {0} (15) occurs, where Rn+ denotes the non-negative orthant. From the approximate kinematic formula [3, Theorem I] in the literature on convex geometry, the event (15) happens with high probability if and only if the total statistical dimension of the two closed convex cones exceeds the ambient dimension, i.e.
δ ({Zθ | θ ∈ Rp }) + δ Rn+ > n + o(n). (16) Here, the statistical dimension of a closed convex cone K is defined as   δ(K) := Eg∼N (0,I) kΠK (g) k2 (17) with ΠK (g) := arg minz∈K kg − zk the Euclidean projection. Recognizing that [3, Proposition 2.4] δ ({Zθ | θ ∈ Rp }) = p and δ(Rn+ ) = n/2, we reduce the condition (16) to p + n/2 > n + o(n) or p/n > 1/2 + o(1), thus indicating that kβ̂k = ∞ when κ = lim p/n > 1/2. This argument only reveals that kβ̂k = ∞ in the regime where κ > 1/2. If κ = p/n < 1/2, then kΣ1/2 β̂k = O(1) with high probability, a fact we shall prove in Section 5. In light of these observations we work with the additional condition κ < 1/2. (18) 2.3 The high-dimensional limiting distribution of the LLR In contrast to the classical Wilks’ result, our findings reveal that the LLR statistic follows a rescaled chisquare distribution with a rescaling factor that can be explicitly pinned down through the solution to a system of equations. 2.3.1 A system of equations We start by setting up the crucial system of equations. Before proceeding, we first recall the proximal operator   1 (19) proxbρ (z) := arg min bρ(x) + (x − z)2 x∈R 2 defined for any b > 0 and convex function ρ(·). As in [18], we introduce the operator Ψ(z; b) := bρ0 (proxbρ (z)), (20) which is simply the proximal operator of the conjugate (bρ)∗ of bρ.2 To see this, we note that Ψ satisfies the relation [18, Proposition 6.4] Ψ(z; b) + proxbρ (z) = z. (21) The claim that Ψ(·; b) = prox(bρ)∗ (·) then follows from the Moreau decomposition proxf (z) + proxf ∗ (z) = z, 2 The ∀z, conjugate f ∗ of a function f is defined as f ∗ (x) = supu∈dom(f ) {hu, xi − f (u)}. 6 (22) 25.0 ● ● 20.0 2.00 ● 15.0 ● Rescaling Constant Rescaling Constant ● 10.0 ● 1.75 ● ● ● ● ● ● 1.50 ● ● ● ● ● 1.25 1.00 ●● ●● ●● ●●● 0.00 0.05 ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● ● 5.0 ● ● ● ● 1.0 0.10 0.15 0.20 κ 0.25 0.30 0.35 0.40 ● 2.5 ● ●●● ● ● ●●●● ●●● ●●● ●●● ●● ●●● ●●● ●● ●● ●● ●● ● ● ●● ● ● ● ● ● ● 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 κ Figure 2: Rescaling constant τ∗2 /b∗ as a function of κ for the logistic model. Note the logarithmic scale in the right panel. The curves for the probit model are nearly identical. which holds for a closed convex function f [39, Section 2.5]. Interested readers are referred to [18, Appendix 1] for more properties of proxbρ and Ψ. We are now in position to present the system of equations that plays a crucial role in determining the distribution of the LLR statistic in high dimensions: i 1 h 2 E (Ψ (τ Z; b)) , κ  κ = E Ψ0 (τ Z; b) , τ2 = (23) (24) where Z ∼ N (0, 1), and Ψ0 (·, ·) denotes differentiation with respect to the first variable. The fact that this system of equations would admit a unique solution in R2+ is not obvious a priori. We shall establish this for the logistic and the probit models later in Section 6. 2.3.2 Main result Theorem 1. Consider a logistic or probit regression model under the assumptions from Section 2.1. If κ ∈ (0, 1/2), then for any 1 ≤ j ≤ p, the log-likelihood ratio statistic Λj as defined in (7) obeys d 2Λj → τ∗2 2 χ , b∗ 1 as n → ∞, (25) where (τ∗ , b∗ ) ∈ R2+ is the unique solution to the system of equations (23) and (24). Furthermore, the LLR statistic obtained by dropping k variables for any fixed k converges to (τ∗2 /b∗ )χ2k . Finally, these results extend to all binary regression models with links obeying the assumptions listed in Section 2.3.3. Hence, the limiting distribution is a rescaled chi-square with a rescaling factor τ∗2 /b∗ that only depends on the aspect ratio κ. Fig. 2 illustrates the dependence of the rescaling factor on the limiting aspect ratio κ for logistic regression. The figures for the probit model are similar as the rescaling constants actually differ by very small values. To study the quality of approximation for finite samples, we repeat the same numerical experiments as before but now obtain the p-values by comparing the LLR statistic with the rescaled chi-square suggested by 7 Theorem 1. For a particular run of the experiment (n = 4000, p = 1200, κ = 0.3), we compute the adjusted LLR statistic (2b∗ /τ∗2 )Λj for each coefficient and obtain the p-values based on the χ21 distribution. The pooled histogram that aggregates 4.8 × 105 p-values in total is shown in Fig. 1(c). As we clearly see, the p-values are much closer to a uniform distribution now. One can compute the chi-square goodness of fit statistic to test the closeness of the above distribution to uniformity. To this end, we divide the interval [0, 1] into 20 equally spaced bins of width 0.05 each. For each bin we compute the observed number of times a p-value falls in the bin out of the 4.8 × 105 values. Then a chi-square goodness of fit statistic is computed, noting that the expected frequency is 24000 for each bin. The chi-square statistic in this case is 16.049, which gives a p-value of 0.654 in comparison with a χ219 variable. The same test when performed with the Bartlett corrected p-values (Fig. 1(b)) yields a chi-square statistic 5599 with a p-value of 0. 3 Thus, our correction gives the desired uniformity in the p-values when the true signal β = 0. Practitioners would be concerned about the validity of p-values when they are small—again, think about multiple testing applications. In order to study whether our correction yields valid results for small p-values, we compute the proportion of times the p-values (in all the three cases) lie below 5%, 1%, 0.5%, 0.1% out of the 4.8 × 105 times. The results are summarized in Table 1. This further illustrates the deviation from uniformity for the classical and Bartlett corrected p-values, whereas the “adjusted” p-values obtained invoking Theorem 1 are still valid. P{p-value ≤ 5%} P{p-value ≤ 1%} P{p-value ≤ 0.5%} P{p-value ≤ 0.1%} P{p-value ≤ 0.05%} P{p-value ≤ 0.01%} Classical 11.1044%(0.0668%) 3.6383%(0.038%) 2.2477%(0.0292%) 0.7519%(0.0155%) 0.4669%(0.0112%) 0.1575%(0.0064%) Bartlett-corrected 6.9592%(0.0534%) 1.6975%(0.0261%) 0.9242%(0.0178%) 0.2306%(0.0078%) 0.124%(0.0056%) 0.0342%(0.0027%) Adjusted 5.0110%(0.0453%) 0.9944%(0.0186%) 0.4952%(0.0116%) 0.1008%(0.0051%) 0.0542%(0.0036%) 0.0104%(0.0014%) Table 1: Estimates of p-value probabilities with estimated Monte Carlo standard errors in parentheses under i.i.d. Gaussian design. 2.3.3 Extensions As noted in Section 1.1, the Symmetry Condition (3) allows to express the negative log-likelihood in the form (4), which makes use of the effective link ρ(·). Theorem 1 applies to any ρ(·) obeying the following properties: 1. ρ is non-negative, has up to three derivatives, and obeys ρ(t) ≥ t. 2. ρ0 may be unbounded but it should grow sufficiently slowly, in particular, we assume |ρ0 (t)| = O(|t|) and ρ0 (proxcρ (Z)) is a sub-Gaussian random variable for any constant c > 0 and any Z ∼ N (0, σ 2 ) for some finite σ > 0. 3. ρ00 (t) > 0 for any t which implies that ρ is convex, and supt ρ00 (t) < ∞. 4. supt |ρ000 (t)| < ∞. 5. Given any τ > 0 ,the equation (24) has a unique solution in b. 6. The map V(τ 2 ) as defined in (57) has a fixed point. 3 Note that the p-values obtained at each trial are not exactly independent. However, they are exchangeable, and weakly dependent (see the proof of Corollary 1 for a formal justification of this fact). Therefore, we expect the goodness of fit test to be an approximately valid procedure in this setting. 8 It can be checked that the effective links for both the logistic and the probit models (5) obey all of the above. The last two conditions are assumed to ensure existence of a unique solution to the system of equations (23) and (24) as will be seen in Section 6; we shall justify these two conditions for the logistic and the probit models in Section 6.1. 2.4 Reduction to independent covariates In order to derive the asymptotic distribution of the LLR statistics, it in fact suffices to consider the special case Σ = Ip . Lemma 1. Let Λj (X) be the LLR statistic based on the design matrix X, where the rows of X are i.i.d. N (0, Σ) and Λj (Z) that where the rows are i.i.d. N (0, Ip ). Then d Λj (X) = Λj (Z). Proof: Recall from (4) that the LLR statistic for testing the jth coefficient can be expressed as Λj (X) = min β n X ρ(−ỹi e> i Xβ) i=1 n X − min β:βj =0 ρ(−ỹi e> i Xβ). i=1 Write Z 0 = XΣ−1/2 so that the rows of Z 0 are i.i.d. N (0, Ip ) and set θ 0 = Σ1/2 β. With this reparameter0 p ization, we observe that the constraint βj = 0 is equivalent to a> j θ = 0 for some non-zero vector aj ∈ R . This gives n n X X > 0 0 0 0 ρ(−ỹ e Z θ ) − min ρ(−ỹi e> Λj (X) = min i i i Z θ ). 0 θ 0 θ 0 :a> j θ =0 i=1 i=1 p Now let Q be an orthogonal matrix mapping aj ∈ R into the vector kaj kej ∈ Rp , i.e. Qaj = kaj kej . 0 Additionally, set Z = Z 0 Q (the rows of Z are still i.i.d. N (0, Ip )) and θ = Qθ 0 . Since a> j θ = 0 occurs if and only if θj = 0, we obtain Λj (X) = min θ n X i=1 ρ(−ỹi e> i Zθ) − min θ:θj =0 n X ρ(−ỹi e> i Zθ) = Λj (Z), i=1  which proves the lemma. In the remainder of the paper we, therefore, assume Σ = Ip . 2.5 Proof architecture This section presents the main steps for proving Theorem 1. We will only prove the theorem for {Λj }, the LLR statistic obtained by dropping a single variable. The analysis for the LLR statistic obtained on dropping k variables (for some fixed k) follows very similar steps and is hence omitted for the sake of conciceness. As discussed before, we are free to work with any configuration of the yi ’s. For the two steps below, we will adopt two different configurations for convenience of presentation. 2.5.1 Step 1: characterizing the asymptotic distributions of β̂j Without loss of generality, we assume here that yi = 1 (and hence ỹi = 1) for all 1 ≤ i ≤ n and, therefore, the MLE problem reduces to Xn minimizeβ∈Rp ρ(−Xi> β). i=1 We would first like to characterize the marginal distribution of β̂, which is crucial in understanding the LLR statistic. To this end, our analysis follows by a reduction to the setup of [18–21], with certain modifications 9 that are called for due to the specific choices of ρ(·) we deal with here. Specifically, consider the linear model y = Xβ + w, (26) and prior work [18–21] investigating the associated M-estimator Xn minimizeβ∈Rp ρ(yi − Xi> β). i=1 (27) Our problem reduces to (27) on setting y = w = 0 in (27). When ρ(·) satisfies certain assumptions (e.g. strong convexity), the asymptotic distribution of kβ̂k has been studied in a series of works [19–21] using a leave-one-out analysis and independently in [18] using approximate message passing (AMP) machinery. An outline of their main results is described in Section 2.7. However, the function ρ(·) in our cases has vanishing curvature and, therefore, lacks the essential strong convexity assumption that was utilized in both the aforementioned lines of work. To circumvent this issue, we propose to invoke the AMP machinery as in [18], in conjunction with the following critical additional ingredients: • (Norm Bound Condition) We utilize results from the conic geometry literature (e.g. [3]) to establish that kβ̂k = O(1) with high probability as long as κ < 1/2. This will be elaborated in Theorem 4. • (Likelihood Curvature Condition) We establish some regularity conditions on the Hessian of the loglikelihood function, generalizing the strong convexity condition, which will be detailed in Lemma 4. • (Uniqueness of the Solution to (23) and (24)) We establish that for both the logistic and the probit case, the system of equations (23) and (24) admits a unique solution. We emphasize that these elements are not straightforward, require significant effort and a number of novel ideas, which form our primary technical contributions for this step. These ingredients enable the use of the AMP machinery even in the absence of strong convexity on ρ(·), finally leading to the following theorem: Theorem 2. Under the conditions of Theorem 1, lim kβ̂k2 =a.s. τ∗2 . n→∞ (28) This theorem immediately implies that the marginal distribution of β̂j is normal. Corollary 1. Under the conditions of Theorem 1, for every 1 ≤ j ≤ p, it holds that √ d pβ̂j → N (0, τ∗2 ), as n → ∞. (29) Proof: From the rotational invariance of our i.i.d. Gaussian design, it can be easily verified that β̂/kβ̂k is uniformly distributed on the unit sphere Sp−1 and is independent of kβ̂k. Therefore, β̂j has the same √ distribution as kβ̂kZj /kZk, where Z = (Z1 , . . . , Zp ) ∼ N (0, Ip ) independent of kβ̂k. Since pkβ̂k/kZk √ converges in probability to τ∗ , we have, by Slutsky’s theorem, that pβ̂j converges to N (0, τ∗2 ) in distribution.  2.5.2 Step 2: connecting Λj with β̂j Now that we have derived the asymptotic distribution of β̂j , the next step involves a reduction of the LLR statistic to a function of the relevant coordinate of the MLE. Before continuing, we note that the distribution of Λj is the same for all 1 ≤ j ≤ p due to exchangeability. As a result, going forward we will only analyze Λ1 without loss of generality. In addition, we introduce the following convenient notations and assumptions: 10 • the design matrix on dropping the first column is written as X̃ and the MLE in the corresponding reduced model as β̃; • write X = [X1 , · · · , Xn ]> ∈ Rn×p and X̃ = [X̃1 , · · · , X̃n ]> ∈ Rn×(p−1) ; • without loss of generality, assume that ỹi = −1 for all i in this subsection, and hence the MLEs under the full and the reduced models reduce to Xn β̂ = arg minβ∈Rp `(β) := ρ(Xi> β), (30) i=1 Xn ˜ β̃ = arg minβ∈Rp−1 `(β) := ρ(X̃i> β). (31) i=1 With the above notations in place, the LLR statistic for testing β1 = 0 vs. β1 6= 0 can be expressed as n n o X ˜ β̃) − `(β̂) = Λ1 := `( ρ(X̃i> β̃) − ρ(Xi> β̂) . (32) i=1 To analyze Λ1 , we invoke Taylor expansion to reach Λ1 = n  1X   2   ρ00 Xi> β̂ X̃i> β̃ − Xi> β̂ ρ0 Xi> β̂ X̃i> β̃ − Xi> β̂ + 2 i=1 i=1 {z } | n X :=Qlin n +  3 1 X 000 ρ (γi ) X̃i> β̃ − Xi> β̂ , 6 i=1 (33) where γi lies between X̃i> β̃ and Xi> β̂. A key observation is that the linear term Qlin in the above equation vanishes. To see this, note that the first-order optimality conditions for the MLE β̂ is given by Xn ρ0 (Xi> β̂)Xi = 0. (34) i=1   0 Replacing X̃i> β̃ with Xi> in Qlin and using the optimality condition, we obtain β̃   Xn   >  0  0 > Qlin = ρ Xi β̂ Xi − β̂ = 0. i=1 β̃ Consequently, Λ1 simplifies to the following form n n  3 2 1 X 1 X 00 >  > ρ (Xi β̂) X̃i β̃ − Xi> β̂ + ρ000 (γi ) X̃i> β̃ − Xi> β̂ . Λ1 = 2 i=1 6 i=1 (35) Thus, computing the asymptotic distribution of Λ1 boils down to analyzing Xi> β̂ − X̃i> β̃. Our argument is inspired by the leave-one-predictor-out approach developed in [19, 20]. We re-emphasize that our setting is not covered by that of [19,20], due to the violation of strong convexity and some other technical assumptions. We sidestep this issue by utilizing the Norm Bound Condition and the Likelihood Curvature Condition. In the end, our analysis establishes the equivalence of Λ1 and β̂1 up to some explicit multiplicative factors modulo negligible error terms. This is summarized as follows. Theorem 3. Under the assumptions of Theorem 1, p P 2Λ1 − β̂12 → 0, b∗ as n → ∞. (36) Theorem 3 reveals a simple yet surprising connection between the LLR statistic Λ1 and the MLE β̂. As we shall see in the proof of the theorem, the quadratic term in (35) is 12 bp∗ β̂12 + o(1), while the remaining third-order term of (35) is vanishingly small. Finally, putting Corollary 1 and Theorem 3 together directly establishes Theorem 1. 11 2.6 Comparisons with the classical regime We pause to shed some light on the interpretation of the correction factor τ∗2 /b∗ in Theorem 1 and understand the differences from classical results. Classical theory (e.g. [30,31]) asserts that when p is fixed and n diverges, the MLE for a fixed design X is asymptotically normal, namely, √ d n(β̂ − β) → N (0, Iβ−1 ), where  Iβ = 1 > X Dβ X n  with Dβ :=  ρ00 X1> β (37)
 .. . ρ 00 Xn> β
  (38) is the normalized Fisher information at the true value β. In particular, under the global null and i.i.d. Gaussian design, this converges to ( 1 I, for the logistic model EX [Iβ ] = 42 π I, for the probit model as n tends to infinity [50, Example 5.40]. The behavior in high dimensions is different. In particular, Corollary 1 states that under the global null, we have √ d p(β̂j − βj ) → N (0, τ∗2 ). (39) Comparing the variances in the logistic model, we have that ( √  4κ, in classical large-sample theory; lim Var pβ̂j = n→∞ τ∗2 , in high dimensions. Fig. 3 illustrates the behavior of the ratio τ∗2 /κ as a function of κ. Two observations are immediate: • First, in Fig. 3(a) we have τ∗2 ≥ 4κ for all κ ≥ 0. This indicates an inflation in variance or an “extra Gaussian noise” component that appears in high dimensions, as discussed in [18]. The variance of the “extra Gaussian noise” component increases as κ grows. • Second, as κ → 0, we have τ∗2 /4κ → 1 in the logistic model, which indicates that classical theory becomes accurate in this case. In other words, our theory recovers the classical prediction in the regime where p = o(n). Further, for the testing problem considered here, the LLR statistic in the classical setup can be expressed, through Taylor expansion, as 2Λ1 = n(β̂ − β̃)> Iβ (β̂ − β̃) + oP (1), (40) where β̃ is defined in (31). In the high-dimensional setting, we will also establish a quadratic approximation of the form 1 2Λ1 = n(β̂ − β̃)> G(β̂ − β̃) + oP (1), G = X > Dβ̂ X. n In Theorem 7, we shall see that b∗ is the limit of n1 Tr(G−1 ), the Stieltjes transform of the empirical spectral distribution of G evaluated at 0. Thus, this quantity in some sense captures the spread in the eigenvalues of G one would expect to happen in high dimensions. 12 10 ● ● 30 ● ● ● ● 7.5 ● ● ● ● 20 ● τ2 κ τ2 κ ● ● ● ● ● ● ● 5 ● ● ● ● ● ● ● ● 10 4 ●● ●● ●● ●● ●● ●● ● ● ●●● ●●● ● 0.00 0.05 0.10 0.15 ● ● ● ● ● ● ● ● 2.5 ● ●●● ● 1.57 0.20 κ 0.25 0.30 0.35 0.40 ●● ●● ●● ●● ● ● ●● ●● ● ● ● ● ● ● ● ● ● ● 0.00 0.05 0.10 0.15 0.20 0.25 0.30 0.35 0.40 κ (a) logistic regression (b) probit regression Figure 3: Ratio of asymptotic variance and dimensionality factor κ as a function of κ. 2.7 Prior art Wilks’ type of phenomenon in the presence of a diverging dimension p has received much attention in the past. For instance, Portnoy [43] investigated simple hypotheses in regular exponential families, and established the asymptotic chi-square approximation for the LLR test statistic as long as p3/2 /n → 0. This phenomenon was later extended in [46] to accommodate the MLE with a quadratic penalization, and in [53] to account for parametric models underlying several random graph models. Going beyond parametric inference, Fan et al. [22, 24] explored extensions to infinite-dimensional non-parametric inference problems, for which the MLE might not even exist or might be difficult to derive. While the classical Wilks’ phenomenon fails to hold in such settings, Fan et al. [22, 24] proposed a generalization of the likelihood ratio statistics based on suitable non-parametric estimators and characterized the asymptotic distributions. Such results have further motivated Boucheron and Massart [9] to investigate the non-asymptotic Wilks’ phenomenon or, more precisely, the concentration behavior of the difference between the excess empirical risk and the true risk, from a statistical learning theory perspective. The Wilks’ phenomenon for penalized empirical likelihood has also been established [48]. However, the precise asymptotic behavior of the LLR statistic in the regime that permits p to grow proportional to n is still beyond reach. On the other hand, as demonstrated in Section 2.5.1, the MLE here under the global null can be viewed as an M-estimator for a linear regression problem. Questions regarding the behavior of robust linear regression estimators in high dimensions—where p is allowed to grow with n—–were raised in Huber [30], and have been extensively studied in subsequent works, e.g. [36, 40–42]. When it comes to logistic regression, the behavior of the MLE was studied for a diverging number of parameters by [28], which characterized the squared estimation error of the MLE if (p log p)/n → 0. In addition, the asymptotic normality properties of the MLE and the penalized MLE for logistic regression have been established by [35] and [23], respectively. A very recent paper by Fan et al. [25] studied the logistic model under the global null β = 0, and investigated the classical asymptotic normality as given in (37). It was discovered in [25] that the convergence property (37) breaks down even in terms of the marginal distribution, namely, √ Iβ nβ̂i d
−1/2 9 N (0, 1) , i,i 13 Iβ = 1 > X X, 4n 12500 30000 15000 10000 7500 Counts Counts Counts 20000 10000 10000 5000 5000 2500 0 0 0.00 0.25 0.50 P−Values 0.75 1.00 (a) 0 0.00 0.25 0.50 P−Values (b) 0.75 1.00 0.00 0.25 0.50 P−Values 0.75 1.00 (c) Figure 4: Histogram of p-values for logistic regression under i.i.d. Bernoulli design, when β = 0, n = 4000, p = 1200, and κ = 0.3: (a) classically computed p-values; (b) Bartlett corrected p-values; (c) adjusted p-values. as soon as p grows at a rate exceeding n2/3 . This result, however, does not imply the asymptotic distribution of the likelihood-ratio statistic in this regime. In fact, our theorem implies that LLR statistic 2Λj goes to χ21 (and hence Wilks phenomenon remains valid) when κ = p/n → 0. The line of work that is most relevant to the present paper was initially started by El Karoui et al. [21]. Focusing on the regime where p is comparable to n, the authors uncovered, via a non-rigorous argument, that the asymptotic `2 error of the MLE could be characterized by a system of nonlinear equations. This seminal result was later made rigorous independently by Donoho et al. [18] under i.i.d. Gaussian design and by El Karoui [19, 20] under more general i.i.d. random design as well as certain assumptions on the error distribution. Both approaches rely on strong convexity on the function ρ(·) that defines the M-estimator, which does not hold in the models considered herein. 2.8 Notations We adopt the standard notation f (n) = O (g(n)) or f (n) . g(n) which means that there exists a constant c > 0 such that |f (n)| ≤ c|g(n)|. Likewise, f (n) = Ω (g(n)) or f (n) & g(n) means that there exists a constant c > 0 such that |f (n)| ≥ c |g(n)|, f (n)  g(n) means that there exist constants c1 , c2 > 0 such that (n) c1 |g(n)| ≤ |f (n)| ≤ c2 |g(n)|, and f (n) = o(g(n)) means that limn→∞ fg(n) = 0. Any mention of C, Ci , c, ci for i ∈ N refers to some positive universal constants whose value may change from line to line. For a square symmetric matrix M , the minimum eigenvalue is denoted by λmin (M ). Logarithms are base e. 3 3.1 Numerics Non-Gaussian covariates In this section we first study the sensitivity of our result to the Gaussianity assumption on the design matrix. To this end, we consider a high dimensional binary regression set up with a Bernoulli design matrix. We i.i.d. simulate n = 4000 i.i.d. observations (yi , Xi ) with yi ∼ Bernoulli(1/2), and Xi generated independent of yi , such that each entry takes on values in {1, −1} w.p. 1/2. At each trial, we fit a logistic regression model to the data and obtain the classical, Bartlett corrected and adjusted p-values (using the rescaling factor τ∗2 /b∗ ). Figure 4 plots the histograms for the pooled p-values, obtained across 400 trials. 14 It is instructive to compare the histograms to that obtained in the Gaussian case (Figure 1). The classical and Bartlett corrected p-values exhibit similar deviations from uniformity as in the Gaussian design case, whereas our adjusted p-values continue to have an approximate uniform distribution. We test for deviations from uniformity using a formal chi-squared goodness of fit test as in Section 2.3.2. For the Bartlett corrected p-values, the chi-squared statistic turns out to be 5885, with a p-value 0. For the adjusted p-values,the chi-squared statistic is 24.1024, with a p-value 0.1922.4 Once again, the Bartlett correction fails to provide valid p-values whereas the adjusted p-values are consistent with a uniform distribution. These findings indicate that the distribution of the LLR statistic under the i.i.d. Bernoulli design is in agreement to the rescaled χ21 derived under the Gaussian design in Theorem 1, suggesting that the distribution is not too sensitive to the Gaussianity assumption. Estimates of p-value probabilities for our method are provided in Table 2. P{p-value ≤ 5%} P{p-value ≤ 1%} P{p-value ≤ 0.5%} P{p-value ≤ 0.1%} P{p-value ≤ 0.05%} P{p-value ≤ 0.01%} Adjusted 5.0222%(0.0412%) 1.0048%(0.0174%) 0.5123%(0.0119%) 0.1108%(0.005%) 0.0521%(0.0033%) 0.0102%(0.0015%) Table 2: Estimates of p-value probabilities with estimated Monte Carlo standard errors in parentheses under i.i.d. Bernoulli design. 3.2 Quality of approximations for finite sample sizes In the rest of this section, we report some numerical experiments which study the applicability of our theory in finite sample setups. Validity of tail approximation The first experiment explores the efficacy of our correction for extremely small p-values. This is particularly important in the context of multiple comparisons, where practitioners care about the validity of exceedingly small p-values. To this end, the empirical cumulative distribution of the adjusted p-values is estimated under a standard Gaussian design with n = 4000, p = 1200 and 4.8 × 105 p-values. The range [0.1/p, 12/p] is divided into points which are equi-spaced with a distance of 1/p between any two consecutive points. The estimated empirical CDF at each of these points is represented in blue in Figure 5. The estimated CDF is in near-perfect agreement with the diagonal, suggesting that the adjusted p-values computed using the rescaled chi-square distribution are remarkably close to a uniform, even when we zoom in at very small resolutions as would be the case when applying Bonferroni-style corrections. Moderate sample sizes The final experiment studies the accuracy of our asymptotic result for moderately large samples. This is especially relevant for applications where the sample sizes are not too large. We repeat our numerical experiments with n = 200, p = 60 for i.i.d. Gaussian design, and 4.8 × 105 p-values. The empirical CDF for these p-values are estimated and Figure 6 shows that the adjusted p-values are nearly uniformly distributed even for moderate sample sizes such as n = 200. 4 Preliminaries This section gathers a few preliminary results that will be useful throughout the paper. We start by collecting some facts regarding i.i.d. Gaussian random matrices. 4 Recall our earlier footnote about the use of a χ2 test. 15 0.0100 ● ●● ● ● ●● ●● ● ●● ●● ● ●● ●● ● ●● ● ●● ●● ● ● ●● ●● ● ●● ●● ● ●● ●● ●● ●● ● ● ● ●● ●● ●● ●● ● ● ●● ● ● ●● ●● ● ● ●● ●● ●● ●● ● ●● ●● ●● ●● ● ● ●● ●● ● ●● ●● ● ● Empirical cdf 0.0075 0.0050 0.0025 0.0000 0.000 0.002 0.004 0.006 0.008 0.010 t Figure 5: Empirical CDF of adjusted pvalues for logistic regression when β = 0, n = 4000, p = 1200. Here, the blue points represent the empirical CDF (t vs. the fraction of p-values below t), and the red line is the diagonal. 1.00 Empirical cdf 0.75 0.50 0.25 0.00 ●● ●● ●● ●● ● ● ●● ●● ●● ●● ● ● ●● ●● ●● ●● ● ● ●● ●● ●● ●● ● ● ●● ●● ●● ●● ● ● ●● ●● ●● ●● ● ● ●● ●● ●● ●● ● ● ●● ●● ●● ●● ● ● ●● ●● ●● ●● ● ●● ●● ●● ●● ● ● ●● 0.00 0.25 0.50 0.75 1.00 t Figure 6: Empirical CDF of adjusted pvalues for logistic regression when β = 0, n = 200, p = 60. Here, the blue points represent the empirical CDF (t vs. the fraction of p-values below t), and the red line is the diagonal. 16 Lemma 2. Let X = [X1 , X2 , . . . Xn ]> be an n × p matrix with i.i.d. standard Gaussian entries. Then
P kX > Xk ≤ 9n ≥ 1 − 2 exp(−n/2); (41) √
√ P sup1≤i≤n kXi k ≤ 2 p ≥ 1 − 2n exp(−( p − 1)2 /2). Proof: This is a straighforward application of [51, Corollary 5.35] and the union bound. (42)  Lemma 3. Suppose X is an n × p matrix with pentries i.i.d N (0, 1), then there exists a constant 0 such √ that whenever 0 ≤  ≤ 0 and 0 ≤ t ≤ 1 −  − p/n, !  r 2 √ 1X p > λmin 1−− Xi Xi − t , ∀S ⊆ [n] with |S| = (1 − )n (43) ≥ n n i∈S     2 with probability exceeding 1 − 2 exp − (1−)t − H () n . Here, H() = − log  − (1 − ) log(1 − ). 2  Proof: See Appendix A.1. The above facts are useful in establishing an eigenvalue lower bound on the Hessian of the log-likelihood function. Specifically, recall that Xn
∇2 `(β) = ρ00 Xi> β Xi Xi> , (44) i=1 and the result is this: Lemma 4 (Likelihood Curvature Condition). Suppose that p/n < 1 and that ρ00 (·) ≥ 0. Then there exists a constant 0 such that whenever 0 ≤  ≤ 0 , with probability at least 1 − 2 exp (−nH ()) − 2 exp (−n/2), the matrix inequality !2 ! r r √ p H() 1 2 00 ∇ `(β)  inf −2 I (45) ρ (z) 1−− 3kβk n n 1− z:|z|≤ √ holds simultaneously for all β ∈ Rp .  Proof: See Appendix A.2. The message of Lemma 4 is this: take  > 0 to be a sufficiently small constant. Then 1 2 ∇ `(β)  ω(kβk) I n for some non-increasing and positive function ω(·) independent of n. This is a generalization of the strong convexity condition. 5 5.1 When is the MLE bounded? Phase transition In Section 2.2, we argued that the MLE is at infinity if we have less than two observations per dimension or κ > 1/2. In fact, a stronger version of the phase transition phenemonon occurs in the sense that kβ̂k = O(1) as soon as κ < 1/2. This is formalized in the following theorem. 17 Theorem 4 (Norm Bound Condition). Fix any small constant  > 0, and let β̂ be the MLE for a model with effective link satisfying the conditions from Section 2.3.3. (i) If p/n ≥ 1/2 + , then kβ̂k = ∞
with probability exceeding 1 − 4 exp −2 n/8 . (ii) There exist universal constants c1 , c2 , C2 > 0 such that if p/n < 1/2 − c1 3/4 , then5 kβ̂k < 4 log 2 2 with probability at least 1 − C2 exp(−c2 2 n). These conclusions clearly continue to hold if β̂ is replaced by β̃ (the MLE under the restricted model obtained on dropping the first predictor). The rest of this section is devoted to proving this theorem. As we will see later, the fact that kβ̂k = O(1) is crucial for utilizing the AMP machinery in the absence of strong convexity. 5.2 Proof of Theorem 4 As in Section 2.5.1, we assume ỹi ≡ 1 throughout this section, and hence the MLE reduces to Xn minimizeβ∈Rp `0 (β) := ρ(−Xi> β). i=1 5.2.1 (46) Proof of Part (i) Invoking [3, Theorem I] yields that if
δ ({Xβ | β ∈ Rp }) + δ Rn+ ≥ (1 + ) n, or equivalently, if p/n ≥ 1/2 + , then 
P {Xβ | β ∈ Rp } ∩ Rn+ 6= {0} ≥ 1 − 4 exp −2 n/8 . As is seen in Section 2.2, kβ̂k = ∞ when {Xβ | β ∈ Rp } ∩ Rn+ 6= {0}, establishing Part (i) of Theorem 4. 5.2.2 Proof of Part (ii) We now turn to the regime in which p/n ≤ 1/2 − O(3/4 ), where 0 <  < 1 is any fixed constant. Begin by observing that the least singular value of X obeys √ (47) σmin (X) ≥ n/4
2
with probability at least 1 − 2 exp − 21 34 − √12 n (this follows from Lemma 3 using  = 0). Then for any β ∈ Rp obeying Xn
`0 (β) = ρ −Xj> β ≤ n log 2 = `0 (0) (48) j=1 and 5 When kβk ≥ 4 log 2 , 2 Xi ∼ N (0, Σ) for a general Σ  0, one has kΣ1/2 β̂k . 1/2 with high probability. 18 (49) we must have n X  max −Xj> β, 0 =
(a) −Xj> β ≤ X j: Xj> β<0 j=1 X
(b) ρ −Xj> β ≤ n log 2; j: Xj> β<0 (a) follows since t ≤ ρ(t) and (b) is a consequence of (48). Continuing, (47) and (49) give √ kXβk √ n log 2 ≤ 4 n log 2 ≤ 2 nkXβk. kβk This implies the following proposition: if the solution β̂—which necessarily satisfies `0 (β̂) ≤ `0 (0)—has 2 norm exceeding kβ̂k ≥ 4 log 2 , then X β̂ must fall within the cone n o Xn √ max {−uj , 0} ≤ 2 nkuk . A := u ∈ Rn (50) j=1 2 Therefore, if one wishes to rule out the possibility of having kβ̂k ≥ 4 log 2 , it suffices to show that with high probability, {Xβ | β ∈ Rp } ∩ A = {0} . (51) This is the content of the remaining proof. We would like to utilize tools from conic geometry [3] to analyze the probability of the event (51). Note, however, that A is not convex, while the theory developed in [3] applies only to convex cones. To bypass the non-convexity issue, we proceed in the following three steps:
1. Generate a set of N = exp 22 p closed convex
cones {Bi | 1 ≤ i ≤ N } such that it forms a cover of A with probability exceeding 1 − exp −Ω(2 p) .  √ 3 √  2. Show that if p < 21 − 2 2 4 − 2H(2 ) n and if n is sufficiently large, then ( 1 P {{Xβ | β ∈ R } ∩ Bi = 6 {0}} ≤ 4 exp − 8 p  √ 3 √ p 1 − 2 2 4 − 10H(2 ) − 2 n 2 ) n for each 1 ≤ i ≤ N . 3. Invoke the union bound to reach P {{Xβ | β ∈ Rp } ∩ A 6= {0}} ≤ P {{Bi | 1 ≤ i ≤ N } does not form a cover of A} + N X P {{Xβ | β ∈ Rp } ∩ Bi 6= {0}} i=1
≤ exp −Ω(2 p) , where we have used the fact that  2 ) √ √ 3 1 1 p P {{Xβ | β ∈ Rp } ∩ Bi 6= {0}} ≤ 4N exp − − 2 2 4 − 10H(2 ) − n 8 2 n i=1 ( ! )  2 √ 3 √ 1 1 p 2 < 4 exp − − 2 2 4 − 10H(2 ) − − 2 n 8 2 n  < 4 exp −2 n .  2 √ 3 √ Here, the last inequality holds if 21 − 2 2 4 − 10H(2 ) − np > 242 , or equivalently, np < 12 − √ 3 √ √ 2 2 4 − 10H(2 ) − 24. ( N X 19 √ √ 3 √ Taken collectively, these steps establish the following claim: if np < 12 − 2 2 4 − 10H(2 ) − 24, then    4 log 2 P kβ̂k > < exp −Ω(2 n) , 2  thus establishing Part (ii) of Theorem 4. We defer the complete details of the preceding steps to Appendix D. Asymptotic `2 error of the MLE 6 This section aims to establish Theorem 2, which characterizes precisely the asymptotic squared error of the MLE β̂ under the global null β = 0. As described in Section 2.5.1, it suffices to assume that β̂ is the solution to the following problem Xn minimizeβ∈Rp ρ(−Xi> β). (52) i=1 In what follows, we derive the asymptotic convergence of kβ̂k under the assumptions from our main theorem. Theorem 5. Under the assumptions of Theorem 1, the solution β̂ to (52) obeys lim kβ̂k2 =a.s. τ∗2 . n→∞ (53) Theorem 5 is derived by invoking the AMP machinery [6, 7, 32]. The high-level idea is the following: in order to study β̂, one introduces an iterative algorithm (called AMP) where a sequence of iterates β̂ t is formed at each time t. The algorithm is constructed so that the iterates asymptotically converge to the MLE in the sense that lim lim kβ̂ t − β̂k2 =a.s. 0. (54) t→∞ n→∞ On the other hand, the asymptotic behavior (asymptotic in n) of β̂ t for each t can be described accurately by a scalar sequence {τt }—called state evolution (SE)—following certain update equations [6]. This, in turn, provides a characterization of the `2 loss of β̂. Further, in order to prove Theorem 2, one still needs to justify (a) the existence of a solution to the system of equations (23) and (24), (b) and the existence of a fixed point for the iterative map governing the SE sequence updates. We will elaborate on these steps in the rest of this section. 6.1 State evolution We begin with the SE sequence {τt } introduced in [18]. Starting from some initial point τ0 , we produce two sequences {bt } and {τt } following a two-step procedure. • For t = 0, 1, . . .: – Set bt to be the solution in b to   κ = E Ψ0 (τt Z; b) ; (55) – Set τt+1 to be 2 τt+1 =  1  2 E (Ψ (τt Z; bt )) . κ 20 (56) 10.0 7.5 7.5 ν(τ2) ν(τ2) 10.0 5.0 2.5 0.0 0.0 5.0 2.5 2.5 5.0 τ2 0.0 7.5 (a) logistic regression 0 1 2 τ2 3 4 (b) probit regression Figure 7: The variance map for both the logistic and the probit models when κ = 0.3: (blue line) variance map V(τ 2 ) as a function of τ 2 ; (red line) diagonal. Suppose that for given any τ > 0, the solution in b to (55) with τt = τ exists and is unique, then one can denote the solution as b(τ ), which in turn allows one to write the sequence {τt } as 2 τt+1 = V(τt2 ) with the variance map  1  2 E Ψ (τ Z; b(τ )) . (57) κ As a result, if there exists a fixed point τ∗ obeying V(τ∗2 ) = τ∗2 and if we start with τ0 = τ∗ , then by induction, V(τ 2 ) = τt ≡ τ∗ and bt ≡ b∗ := b(τ∗ ), t = 0, 1, . . . Notably, (τ∗ , b∗ ) solves the system of equations (23) and (24). We shall work with this choice of initial condition throughout our proof. The preceding arguments hold under two conditions: (i) the solution to (24) exists and is unique for any τt > 0; (ii) the variance map (57) admits a fixed point. To verify these two conditions, we make two observations. • Condition (i) holds if one can show that the function G(b) := E [Ψ0 (τ Z; b)] , b>0 (58) is strictly monotone for any given τ > 0, and that limb→0 G(b) < κ < limb→∞ G(b). • Since V(·) is a continuous function, Condition (ii) becomes self-evident once we show that V(0) > 0 and that there exists τ > 0 obeying V(τ 2 ) < τ 2 . The behavior of the variance map is illustrated in Figure 7 for the logistic and probit regression when κ = 0.3. One can in fact observe that the fixed point is unique. For other values of κ, the variance map shows the same behavior. In fact, the aforementioned properties can be proved for a certain class of effective links, as summarized in the following lemmas. In particular, they can be shown for the logistic and the probit models. 21 Lemma 5. Suppose the effective link ρ satisfies the following two properties: (a) ρ0 is log-concave. (b) For any fixed τ > 0 and any fixed z, bρ00 (proxbρ (τ z)) → ∞ when b → ∞. Then for any τ > 0, the function G(b) defined in (58) is an increasing function in b (b > 0), and the equation G(b) = κ has a unique positive solution.  Proof: See Appendix B. Lemma 6. Suppose that 0 < κ < 1/2 and that ρ = log(1 + et ) or ρ = − log Φ(−t). Then (i) V(0) > 0; (ii) V(τ 2 ) < τ 2 for some sufficiently large τ 2 .  Proof: See Appendix C and the supplemental material [47]. Remark 1. A byproduct of the proof is that the following relations hold for any constant 0 < κ < 1/2: • In the logistic case,  2  limτ →∞ V (τ2 ) τ  lim τ →∞ b(τ ) τ = x2 P{Z>x}+E[Z 2 1{0<Z<x} ] P{0<Z<x} ; x=Φ−1 (κ+0.5) = Φ−1 (κ + 0.5). • In the probit case, lim b(τ ) = τ →∞ 2κ 1 − 2κ and V(τ 2 ) = 2κ. τ →∞ τ 2 lim (59) Remark 2. Lemma 6 is proved for the two special effective link functions, the logistic and the probit cases. However, the proof sheds light on general conditions on the effective link that suffice for the lemma to hold. Such general sufficient conditions are also discussed in the supplemental material [47]. 6.2 AMP recursion In this section, we construct the AMP trajectory tracked by two sequences {β̂ t (n) ∈ Rp } and {η t (n) ∈ Rn } for t ≥ 0. Going forward we suppress the dependence on n to simplify presentation. Picking β̂ 0 such that lim kβ̂ 0 k2 = τ02 = τ∗2 n→∞ and taking η −1 = 0 and b−1 = 0, the AMP path is obtained via Algorithm 1, which is adapted from the algorithm in [18, Section 2.2]. 22 Algorithm 1 Approximate message passing. For t = 0, 1, · · · : 1. Set
η t = X β̂ t + Ψ η t−1 ; bt−1 ; (60) κ = E [Ψ0 (τt Z; b)] , (61) 2. Let bt be the solution to where τt is the SE sequence value at that time. 3. Set
1 β̂ t+1 = β̂ t − X > Ψ η t ; bt . p (62) Here, Ψ(·) is applied in an entrywise manner, and Ψ0 (., .) denotes derivative w.r.t the first variable. As asserted by [18], the SE sequence {τt } introduced in Section 6.1 proves useful as it offers a formal procedure for predicting operating characteristics of the AMP iterates at any fixed iteration. In particular it assigns predictions to two types of observables: observables which are functions of the β̂ t sequence and those which are functions of η t . Repeating identical argument as in [18, Theorem 3.4], we obtain lim kβ̂ t k2 =a.s. τt2 ≡ τ∗2 , n→∞ 6.3 t = 0, 1, . . . . (63) AMP converges to the MLE We are now in position to show that the AMP iterates {β̂ t } converge to the MLE in the large n and t limit. Before continuing, we state below two properties that are satisfied under our assumptions. • The MLE β̂ obeys lim kβ̂k < ∞ n→∞ (64) almost surely. • And there exists some non-increasing continuous function 0 < ω (·) < 1 independent of n such that   1 2 P ∇ ` (β)  ω (kβk) · I, ∀β ≥ 1 − c1 e−c2 n . (65) n In fact, the norm bound (64) follows from Theorem 4 together with Borel-Cantelli, while the likelihood curvature condition (65) is an immediate consequence of Lemma 4. With this in place, we have: Theorem 6. Suppose (64) and (65) hold. Let (τ∗ , b∗ ) be a solution to the system (23) and (24), and assume that limn→∞ kβ̂ 0 k2 = τ∗2 . Then the AMP trajectory as defined in Algorithm 1 obeys lim lim kβ̂ t − β̂k =a.s. 0. t→∞ n→∞ Taken collectively, Theorem 6 and Eqn. (63) imply that lim kβ̂k =a.s. lim lim kβ̂ t k =a.s. τ∗ , n→∞ t→∞ n→∞ thus establishing Theorem 5. In addition, an upshot of these theorems is a uniqueness result: Corollary 2. The solution to the system of equations (23) and (24) is unique. 23 (66) Proof: When the AMP trajectory β̂ t is started with the initial condition from Theorem 6, limn→∞ kβ̂k2 =a.s. τ∗2 . This holds for any τ∗ such that (τ∗ , b∗ ) is a solution to (23) and (24). However, since the MLE problem is strongly convex and hence admits a unique solution β̂, this implies that τ∗ must be unique, which together with the monotonicity of G(·) (cf. (58)) implies that b∗ is unique as well.  Proof of Theorem 6: To begin with, repeating the arguments in [18, Lemma 6.9] we reach lim lim kβ̂ t+1 − β̂ t k2 =a.s. 0; t→∞ n→∞ (67) 1 t+1 kη − η t k2 =a.s. 0. (68) n To show that the AMP iterates converge to the MLE, we shall analyze the log-likelihood function. Recall from Taylor’s theorem that >    E 1 D β̂ − β̂ t ∇2 ` β̂ t + λ(β̂ − β̂ t ) β̂ − β̂ t `(β̂) = `(β̂ t ) + ∇`(β̂ t ), β̂ − β̂ t + 2 holds for some 0 < λ < 1. To deal with the quadratic term, we would like to control the Hessian of the likelihood at a point between β̂ and β̂ t . Invoking the likelihood curvature condition (65), one has D E 1  n o `(β̂ t ) ≥ `(β̂) ≥ `(β̂ t ) + ∇`(β̂ t ), β̂ − β̂ t + nω max kβ̂k, kβ̂ t k kβ̂ − β̂ t k2 (69) 2 with high probability. Apply Cauchy-Schwarz to yield that with exponentially high probability, lim lim t→∞ n→∞ kβ̂ − β̂ t k ≤ 1 1 2 2 n o
∇`(β̂ t ) ≤ ∇`(β̂ t ) ,

t n n t ω kβ̂k ω kβ̂ k ω max kβ̂k, kβ̂ k where the last inequality follows since 0 < ω(·) < 1 and ω(·) is non-decreasing. It remains to control k∇`(β̂ t )k. The identity Ψ(z; b∗ ) = z − proxb∗ ρ (z) and (60) give
proxb∗ ρ η t−1 = X β̂ t + η t−1 − η t . (70) 0 In addition, substituting Ψ (z; b) = bρ (proxρb (z)) into (62) yields  
p t (β̂ − β̂ t−1 ) = −X > ρ0 proxb∗ ρ (η t−1 ) = −X > ρ0 X β̂ t + η t−1 − η t . b∗ We are now ready to bound k∇`(β̂ t )k. Recalling that      ∇`(β̂ t ) = X > ρ0 (X > β̂ t ) = X > ρ0 X β̂ t + η t−1 − η t + X > ρ0 (X > β̂ t ) − ρ0 X β̂ t + η t−1 − η t and that supz ρ00 (z) < ∞, we have     ∇`(β̂ t ) ≤ −X > ρ0 X β̂ t + η t−1 − η t + kXk ρ0 X β̂ t + η t−1 − η t − ρ0 (X β̂ t )   p t t−1 00 kβ̂ − β̂ k + kXk sup ρ (z) kη t−1 − η t k. ≤ b∗ z This establishes that with probability at least 1 − c1 e−c2 n ,     2 p 1    kβ̂ − β̂ t k ≤  kβ̂ t − β̂ t−1 k + sup ρ00 (z) kXkkη t−1 − η t k . (71) b∗ n n z ω kβ̂k ω kβ̂ t k √ Using (42) together with Borel-Cantelli yields limn→∞ kXk/ n < ∞ almost surely. Further, it follows from (63) that limn→∞ kβ̂ t k is finite almost surely as τ∗ < ∞. These taken together with (64), (67) and (68) yield lim lim kβ̂ − β̂ t k =a.s. 0 (72) t→∞ n→∞  as claimed. 24 7 Likelihood ratio analysis This section presents the analytical details for Section 2.5.2, which relates the log-likelihood ratio statistic Λi with β̂i . Recall from (35) that the LLR statistic for testing β1 = 0 vs. β1 6= 0 is given by n Λ1 = >  3   1X 1 X̃ β̃ − X β̂ Dβ̂ X̃ β̃ − X β̂ + ρ000 (γi ) X̃i> β̃ − Xi> β̂ , 2 6 i=1 where    ρ00 X1> β̂ (73)    Dβ̂ :=   .. .  ρ00 Xn> β̂      (74) and γi lies between Xi> β̂ and X̃i> β̃. The asymptotic distribution of Λ1 claimed in Theorem 3 immediately follows from the result below, whose proof is the subject of the rest of this section. Theorem 7. Let (τ∗ , b∗ ) be the unique solution to the system of equations (23) and (24), and define G̃ = 1 > X̃ Dβ̃ X̃ n and α̃ = 1 Tr(G̃−1 ). n (75) Suppose p/n → κ ∈ (0, 1/2) . Then (a) the log-likelihood ratio statistic obeys P 2Λ1 − pβ̂12 /α̃ → 0; (76) (b) and the scalar α̃ converges, P α̃ → b∗ . 7.1 (77) More notations and preliminaries Before proceeding, we introduce some notations that will be used throughout. For any matrix X, denote by Xij and X·j its (i, j)-th entry and jth column, respectively. We denote an analogue r = {ri }1≤i≤n (resp. r̃ = {r̃i }1≤i≤n ) of residuals in the full (resp. reduced) model by

ri := −ρ0 Xi> β̂ and r̃i := −ρ0 X̃i> β̃ . (78) As in (74), set   Dβ̃ :=  ρ00 X̃1> β̃
.. . ρ00 X̃n> β̃      and Dβ̂,b̃ :=  ρ00 (γ1∗ )
 .. . ρ 00 (γn∗ ),  , (79) where γi∗ is between Xi> β̂ and Xi> b̃, and b̃ is to be defined later in Section 7.2. Further, as in (75), introduce the Gram matrices G := 1 > X Dβ̂ X n and Gβ̂,b̃ = 1 > X Dβ̂,b̃ X. n (80) Let G̃(i) denote the version of G̃ without the term corresponding to the ith observation, that is, G̃(i) = 1X ρ00 (X̃j> β̃)X̃j X̃j> . j:j6=i n 25 (81) Additionally, let β̂[−i] be the MLE when the ith observation is dropped and let G[−i] be the corresponding Gram matrix, G[−i] = 1X ρ00 (Xj> β̂[−i] )Xj Xj> . j:j6=i n (82) Further, let β̃[−i] be the MLE when the first predictor and ith observation are removed, i.e. X β̃[−i] := arg min ρ(X̃j> β). β∈Rp−1 j:j6=i Below G̃[−i] is the corresponding version of G̃, G̃[−i] = 1X ρ00 (X̃j> β̃[−i] )X̃j X̃j> . j:j6=i n (83) For these different versions of G, their least eigenvalues are all bounded away from 0, as asserted by the following lemma. Lemma 7. There exist some absolute constants λlb , C, c > 0 such that P(λmin (G) > λlb ) ≥ 1 − Ce−cn . Moreover, the same result holds for G̃, Gβ̂,b̃ , G̃(i) , G[−i] and G̃[−i] for all i ∈ [n]. Proof: This result follows directly from Lemma 2, Lemma 4, and Theorem 4.  Throughout the rest of this section, we restrict ourselves (for any given n) to the following event: An := {λmin (G̃) > λlb } ∩ {λmin (G) > λlb } ∩ {λmin (Gβ̂,b̃ ) > λlb } ∩ {∩ni=1 λmin (G̃(i) ) > λlb } ∩ {∩ni=1 λmin (G̃[−i] ) > λlb } ∩ {∩ni=1 λmin (G[−i] ) > λlb }. (84) By Lemma 7, An arises with exponentially high probability, i.e. P(An ) ≥ 1 − exp(−Ω(n)). 7.2 (85) A surrogate for the MLE In view  of (73), the main step in controlling Λ1 consists of characterizing the differences X β̂ − X̃ β̃ or 0 β̂ − . Since the definition of β̂ is implicit and not amenable to direct analysis, we approximate β̂ by β̃ a more amenable surrogate b̃, an idea introduced in [19–21]. We collect some properties of the surrogate which will prove valuable in the subsequent analysis. To begin with, our surrogate is     0 1 b̃ = + b̃1 , (86) −G̃−1 w β̃ where G̃ is defined in (80), w := 1 Xn 1 ρ00 (X̃i> β̃)Xi1 X̃i = X̃ > Dβ̃ X·1 , i=1 n n and b̃1 is a scalar to be specified later. This vector is constructed in the hope that ( β̂1 ≈ b̃1 , β̂ ≈ b̃, or equivalently, β̂2:p − β̃ ≈ −b̃1 G̃−1 w, 26 (87) (88) where β̂2:p contains the 2nd through pth components of β̂. Before specifying b̃1 , we shall first shed some insights into the remaining terms in b̃. By definition, " #     > > > X D X X D X̃ X·1 Dβ̃ X·1 nw> 0 ·1 ·1 ·1 2 > β̃ β̃ ∇ ` = X Dβ̃ X = . = β̃ X̃ > Dβ̃ X·1 X̃ > Dβ̃ X̃ nw nG̃ Employing the first-order approximation of ∇`(·) gives        0 0 0 ∇2 ` β̂ − ≈ ∇`(β̂) − ∇` . β̃ β̃ β̃ (89)   0 are β̂ also very close to each other. Therefore, taking the 2nd through pth components of (89) and approximating them by zero give   h i 0 w, G̃ β̂ − ≈ 0. β̃ Suppose β̂2:p is well approximated by β̃. Then all but the first coordinates of ∇`(β̃) and ∇` This together with a little algebra yields β̂2:p − β̃ ≈ −β̂1 G̃−1 w ≈ −b̃1 G̃−1 w, which coincides with (88). In fact, for all but the 1st entries, b̃ is constructed by moving β̃ one-step in the direction which takes it closest to β̂. Next, we come to discussing the scalar b̃1 . Introduce the projection matrix H := I − and define b̃1 as b̃1 := 1 1/2 1/2 D X̃ G̃−1 X̃ > Dβ̃ , n β̃ > X·1 r̃ , 1/2 1/2 > X·1 Dβ̃ HDβ̃ X·1 (90) (91) where r̃ comes from (78). In fact, the expression b̃1 is obtained through similar (but slightly more complicated) first-order approximation as for b̃2:p , in order to make sure that b1 ≈ β̂1 ; see [21, Pages 14560-14561] for a detailed description. We now formally justify that the surrogate b̃ and the MLE β̂ are close to each other. Theorem 8. The MLE β̂ and the surrote b̃ (86) obey kβ̂ − b̃k . n−1+o(1) , (92) |b̃1 | . n−1/2+o(1) , (93) sup |Xi> b̃ − X̃i> β̃| . n−1/2+o(1) (94) and 1≤i≤n with probability tending to one as n → ∞.  Proof: See Section 7.4. The global accuracy (92) immediately leads to a coordinate-wise approximation result between β̂1 and b̃1 . 27 Corollary 3. With probability tending to one as n → ∞, √ n|b̃1 − β̂1 | . n−1/2+o(1) . (95) Another consequence from Theorem 8 is that the value Xi> β̂ in the full model and its counterpart X̃i> β̃ in the reduced model are uniformly close. Corollary 4. The values Xi> β̂ and X̃i> β̃ are uniformly close in the sense that sup Xi> β̂ − X̃i> β̃ . n−1/2+o(1) (96) 1≤i≤n holds with probability approaching one as n → ∞. Proof: Note that sup Xi> β̂ − X̃i> β̃ ≤ 1≤i≤n sup Xi> (β̂ − b̃) + sup Xi> b̃ − X̃i> β̃ . 1≤i≤n 1≤i≤n The second term in the right-hans side is upper bounded by n−1/2+o(1) with probability 1 − o(1) according to Theorem 8. Invoking Lemma 2 and Theorem 8 and applying Cauchy-Schwarz inequality yield that the first term is O(n−1/2+o(1) ) with probability 1 − o(1). This establishes the claim.  7.3 Analysis of the likelihood-ratio statistic We are now positioned to use our surrogate b̃ to analyze the likelihood-ratio statistic. In this subsection we establish Theorem 7(a). The proof for Theorem 7(b) is deferred to Appendix I. Recall from (35) that n 2Λ1 = (X̃ β̃ − X β̂)> Dβ̂ (X̃ β̃ − X β̂) + 1 X 000 ρ (γi )(X̃i> β̃ − Xi> β̂)3 . 3 i=1 {z } | :=I3 000 To begin with, Corollary 4 together with the assumption supz ρ (z) < ∞ implies that I3 . n−1/2+o(1) with probability 1 − o(1). Hence, I3 converges to zero in probability. Reorganize the quadratic term as follows:  2 X (X̃ β̃ − X β̂)> Dβ̂ (X̃ β̃ − X β̂) = ρ00 (Xi> β̂) Xi> β̂ − X̃i> β̃ i = X = X h i2 ρ00 (Xi> β̂) Xi> (β̂ − b̃) + (Xi> b̃ − X̃i> β̃) i ρ00 (Xi> β̂)(Xi> (β̂ − b̃))2 + 2 X ρ00 (Xi> β̂)Xi> (β̂ − b̃)(Xi> b̃ − X̃i> β̃) i i + X ρ 00 (Xi> β̂) Xi> b̃
2 − X̃i> β̃ . (97) i We control each of the three terms in the right-hand side of (97). • Since supz ρ00 (z) < ∞, the first term in (97) is bounded by X X ρ00 (Xi> β̂)(Xi> (β̂ − b̃))2 . ||β̃ − b̃||2 Xi Xi> i i with probability 1 − o(1), by an application of Theorem 8 and Lemma 2. 28 . n−1+o(1) • From the definition of b̃, the second term can be upper bounded by   p X X 1 ≤ |b̃1 | · kβ̂ − b̃k · Xi Xi> · 1 + w> G̃−2 w 2 ρ00 (Xi> β̂)(β̂ − b̃)> Xi Xi> b̃1 −1 i −G̃ w i 1 . n− 2 +o(1) with probability 1 − o(1), where the last line follows from a combination of Theorem 8, Lemma 2 and the following lemma. Lemma 8. Let G̃ and w be as defined in (80) and (87), respectively. Then   P w> G̃−2 w . 1 ≥ 1 − exp(−Ω(n)). (98)  Proof: See Appendix E. • The third term in (97) can be decomposed as X ρ00 (Xi> β̂)(Xi> b̃ − X̃i> β̃))2 i = X = X  X ρ00 (Xi> β̂) − ρ00 (X̃i> β̃) (Xi> b̃ − X̃i> β̃))2 + ρ00 (X̃i> β̃)(Xi> b̃ − X̃i> β̃)2 i i  ρ000 (γ̃i )(Xi> β̂ − X̃i> β̃) Xi> b̃ − X̃i> β̃ i 2 + X  2 ρ00 (X̃i> β̃) Xi> b̃ − X̃i> β̃ (99) i for some γ̃i between Xi> β̂ and X̃i> β̃. From Theorem 8 and Corollary 4, the first term in (99) is O(n−1/2+o(1) ) with probability 1 − o(1). Hence, the only remaining term is the second. In summary, we have 2Λ1 − X  2 P → 0, ρ00 (X̃i> β̃) Xi> b̃ − X̃i> β̃ (100) i | {z =v > X > Dβ̃ Xv }  where v := b̃1  1 according to (86). On simplification, the quadratic form reduces to −G̃−1 w  >   v > X > Dβ̃ Xv = b̃21 X·1 − X̃ G̃−1 w Dβ̃ X·1 − X̃ G̃−1 w   > > = b̃21 X·1 Dβ̃ X·1 − 2X·1 Dβ̃ X̃ G̃−1 w + w> G̃−1 X̃ > Dβ̃ X̃ G̃−1 w   > = b̃21 X·1 Dβ̃ X·1 − nw> G̃−1 w ! 1/2 > 1/2 2 1 = nb̃1 X·1 Dβ̃ HDβ̃ X·1 , |n {z } :=ξ recalling the definitions (80), (87), and (90). Hence, the log-likelihood ratio 2Λ1 simplifies to nb̃21 ξ + oP (1) on An . Finally, rewrite v > X > Dβ̃ Xv as n(b̃21 − β̂12 )ξ + nβ̂12 ξ. To analyze the first term, note that 1 n|b̃21 − β̂12 | = n|b̃1 − β̂1 | · |b̃1 + β̂1 | ≤ n|b̃1 − βˆ1 |2 + 2n|b̃1 | · |b̃1 − β̂1 | . n− 2 +o(1) 29 (101) with probability 1 − o(1) in view of Theorem 8 and Corollary 3. It remains to analyze ξ. Recognize that X·1 1/2 1/2 is independent of Dβ̃ HDβ̃ . Applying the Hanson-Wright inequality [27, 44] and the Sherman-MorrisonWoodbury formula (e.g. [26]) leads to the following lemma: Lemma 9. Let α̃ = 1 −1 ), n Tr(G̃ where G̃ = 1 > n X̃ Dβ̃ X̃. Then one has 1 > 1/2 p−1 1/2 − α̃ X·1 Dβ̃ HDβ̃ X·1 . n−1/2+o(1) n n (102) with probability approaching one as n → ∞.  Proof: See Appendix F. In addition, if one can show that α̃ is bounded away from zero with probability 1 − o(1), then it is seen from Lemma 9 that p P ξ− → 0. (103) nα̃ To justify the above claim, we observe that since ρ00 is bounded, λmax (G̃) . λmax (X̃ > X̃)/n . 1 with exponentially high probability (Lemma 2). This yields α̃ = Tr(G̃−1 )/n & p/n with probability 1 − o(1). On the other hand, on An one has α̃ ≤ p/(nλmin (G̃)) . p/n. Hence, it follows that ξ = Ω(1) with probability 1 − o(1). Putting this together with (101) gives the approximation v > X > Dβ̃ Xv = nβ̂12 ξ + o(1). (104) Taken collectively (100), (103) and (104) yields the desired result P 2Λ1 − pβ̂12 /α̃ → 0. 7.4 Proof of Theorem 8 This subsection outlines the main steps for the proof of Theorem 8. To begin with, we shall express the difference β̂ − b̃ in terms of the gradient of the negative log-likelihood function. Note that ∇`(β̂) = 0, and hence Xn ∇`(b̃) = ∇`(b̃) − ∇`(β̂) = Xi [ρ0 (Xi> b̃) − ρ0 (Xi0 β̂)] i=1 Xn = ρ00 (γi∗ )Xi Xi> (b̃ − β̂), i=1 where γi∗ is between Xi> β̂ and Xi> b̃. Recalling the notation introduced in (80), this can be rearranged as b̃ − β̂ = 1 −1 G ∇`(b̃). n β̂,b̃ Hence, on An , this yields kβ̂ − b̃k ≤ k∇`(b̃)k . λlb n   0 The next step involves expressing ∇`(b̃) in terms of the difference b̃ − . β̃ 30 (105) Lemma 10. On the event An (84), the negative log-likelihood evaluated at the surrogate b̃ obeys    n X  00 ∗  0 00 > > ∇`(b̃) = ρ (γi ) − ρ (X̃i β̃) Xi Xi b̃ − , β̃ i=1 where γi∗ is some quantity between Xi> b̃ and X̃i> β̃. Proof: The proof follows exactly the same argument as in the proof of [20, Proposition 3.11], and is thus omitted.    0 The point of expressing ∇`(b̃) in this way is that the difference b̃ − is known explicitly from the β̃ definition of b̃. Invoking Lemma 10 and the definition (86) allows one to further upper bound (105) as kβ̂ − b̃k . 1 ∇`(b̃) n . sup ρ00 (γi∗ ) − ρ00 (X̃i> β̃) i . sup Xi> b̃ − X̃i> β̃ |ρ000 |∞ i   n 1X 0 Xi Xi> b̃ − β̃ n i=1 p 1 Xn Xi Xi> · |b̃1 | 1 + w> G̃−2 w i=1 n . |b̃1 | sup Xi> b̃ − X̃i> β̃ (106) i with probability at least 1 − exp(−Ω(n)). The last inequality here comes from our assumption that supz |ρ000 (z)| < ∞ together with Lemmas 2 and 8. In order to bound (106), we first make use of the definition of b̃ to reach sup Xi> b̃ − X̃i> β̃ = |b̃1 | sup |Xi1 − X̃i> G̃−1 w|. i (107) i The following lemma provides an upper bound on supi |Xi1 − X̃i> G̃−1 w|. Lemma 11. With G̃ and w as defined in (80) and (87),   > −1 o(1) P sup Xi1 − X̃i G̃ w ≤ n ≥ 1 − o(1). (108) 1≤i≤n  Proof: See Appendix G. In view of Lemma 11, the second term in the right-hand side of (107) is bounded above by no(1) with high probability. Thus, in both the bounds (106) and (107), it only remains to analyze the term b̃1 . To this end, we control the numerator and the denominator of b̃1 separately. > • Recall from the definition (91) that the numerator of b̃1 is given by X·1 r̃ and that r̃ is independent > of X·1 . Thus, conditional on X̃, the quantity X·1 r̃ is distributed as a Gaussian with mean zero and variance Xn
2 σ2 = ρ0 (X̃i> β̃) . i=1 0 Since |ρ (x)| = O(|x|), the variance is bounded by Xn  σ 2 . β̃ > X̃i X̃i> β̃ . nkβ̃k2 . n i=1 (109) with probability at least 1 − exp(−Ω(n))), a consequence from Theorem 4 and Lemma 2. Therefore, with probability 1 − o(1), we have 1 > √ X·1 r̃ . no(1) . (110) n 31 • We now move on to the denominator of b̃1 in (91). In the discussion following Lemma 9 we showed 1/2 1 > 1/2 n X·1 Dβ̃ HDβ̃ X·1 = Ω(1) with probability 1 − o(1). Putting the above bounds together, we conclude   1 P |b̃1 | . n− 2 +o(1) = 1 − o(1). (111) Substitution into (106) and (107) yields kβ̂ − b̃k . n−1+o(1) and sup Xi> b̃ − X̃i> β̃ . n−1/2+o(1) i with probability 1 − o(1) as claimed. 8 Discussion In this paper, we derived the high-dimensional asymptotic distribution of the LLR under our modelling assumptions. In particular, we showed that the LLR is inflated vis a vis the classical Wilks’ approximation and that this inflation grows as the dimensionality κ increases. This inflation is typical of high-dimensional problems, and one immediate practical consequence is that it explains why classically computed p-values are completely off since they tend to be far too small under the null hypothesis. In contrast, we have shown in our simulations that our new limiting distribution yields reasonably accurate p-values in finite samples. Having said this, our work raises a few important questions that we have not answered and we conclude this paper with a couple of them. • We expect that our results continue to hold when the covariates are not normally distributed, see Section 3 for some numerical evidence in this direction. To be more precise, we expect the same limiting distribution to hold when the variables are simply sub-Gaussian. If this were true, then this would imply that our rescaled chi-square has a form of universal validity. • The major limitation of our work is arguably the fact that our limiting distribution holds under the global null; that is, under the assumption that all the regression coefficients vanish. It is unclear to us how the distribution would change in the case where the coefficients are not all zero. In particular, would the limiting distribution depend upon the unknown values of these coefficients? Are there assumptions under which it would not? Suppose for instance that we model the regression coefficients as i.i.d. samples from the mixture model (1 − )δ0 + Π? , where 0 <  < 1 is a mixture parameter, δ0 is a point mass at zero and Π? is a distribution with vanishing mass at zero. Then what would we need to know about  and Π? to compute the asymptotic distribution of the LLR under the null? Acknowledgements E. C. was partially supported by the Office of Naval Research under grant N00014-16-1-2712, and by the Math + X Award from the Simons Foundation. Y. C. and P. S. are grateful to Andrea Montanari for his help in understanding AMP and [18]. Y. C. thanks Kaizheng Wang and Cong Ma for helpful discussion about [20], and P. S. thanks Subhabrata Sen for several helpful discussions regarding this project. E. C. would like to thank Iain Johnstone for a helpful discussion as well. 32 A A.1 Proofs for Eigenvalue Bounds Proof of Lemma 3 p √ Fix  ≥ 0 sufficiently small. For any given S ⊆ [n] obeying |S| = (1 − )n and 0 ≤ t ≤ 1 −  − p/n it follows from [51, Corollary 5.35] that ! r 2  √ √ 2 1 p p 1X √ > λmin < Xi Xi |S| − p − t n = 1−− −t n n n i∈S    2  2 n . Taking the union bound over all possible holds with probability at most 2 exp − t 2|S| = 2 exp − (1−)t 2 subsets S of size (1 − )n gives ( 1 λmin n P ∃S ⊆ [n] with |S| = (1 − )n s.t. ! X Xi Xi> <  √ r 1−− i∈S p −t n 2 )    n (1 − ) t2 n 2 exp − (1 − )n 2   2 (1 − ) t n , ≤ 2 exp nH () − 2  ≤ where the last line is a consequence of the inequality A.2 n (1−)n
≤ enH() [16, Example 11.1.3]. Proof of Lemma 4 Define  SB (β) := i : |Xi> β| ≤ Bkβk for any B > 0 and any β. Then n X
ρ00 Xi> β Xi Xi> X  i=1
ρ00 Xi> β Xi Xi>  i∈SB (β) inf z:|z|≤Bkβk ρ00 (z) X Xi Xi> . i∈SB (β) If one also has |SB (β) | ≥ (1 − )n (for  ≥ 0 sufficiently small), then this together with Lemma 3 implies that r  2 n √
1 X 00 p ρ Xi> β Xi Xi>  inf ρ00 (z) 1−− −t I n i=1 n z:|z|≤Bkβk     2 with probability at least 1 − 2 exp − (1−)t − H () n . 2 Thus if we can ensure that with high probability, |SB (β) | ≥ (1 − )n holds simultaneously for all β, then we are done. From Lemma 2 we see that n1 X > X ≤ 9 with probability exceeding 1 − 2 exp (−n/2). On this event, 2 kXβk ≤ 9nkβk2 , ∀β. (112) On the other hand, the definition of SB (β) gives 2 kXβk ≥ X 2 Xi> β i∈S / B (β)   |SB (β)| B 2 kβk2 . ≥ n − |SB (β)| (Bkβk) = n 1 − n 2
Taken together, (112) and (113) yield  |SB (β)| ≥ 1− 9 B2 33  n, ∀β (113) with probability at least 1−2 exp(−n/2). Therefore, with probability 1−2 exp(−n/2), S3/√ (β) ≥ (1 − ) n q holds simultaneously for all β. Putting the above results together and setting t = 2 H() 1− give n X ρ 00 Xi> β
Xi Xi>  00 ρ (z) inf z:|z|≤ i=1 √ r 1−− 3kβk √  !2 r p H() −2 I n 1− simultaneously for all β with probability at least 1 − 2 exp (−nH ()) − 2 exp (−n/2). B Proof of Lemma 5 Applying an integration by parts leads to Z ∞ 1 E [Ψ0 (τ Z; b)] = Ψ0 (τ z; b)φ(z)dz = Ψ(τ z; b)φ(z) τ −∞ Z 1 ∞ Ψ(τ z; b)φ0 (z)dz = − τ −∞ with φ(z) = √1 2π ∞ −∞ − 1 τ Z ∞ Ψ(τ z; b)φ0 (z)dz −∞ exp(−z 2 /2). This reveals that 0 G (b) = =
Z Z ρ0 proxbρ (τ z) 1 ∞ ∂Ψ(τ z; b) 0 1 ∞
φ0 (z)dz − φ (z)dz = − τ −∞ ∂b τ −∞ 1 + bρ00 proxbρ (τ z)

! Z ρ0 proxbρ (−τ z) ρ0 proxbρ (τ z) 1 ∞
−
φ0 (z)dz, τ 0 1 + xρ00 proxbρ (−τ z) 1 + xρ00 proxbρ (τ z) (114) where the second identity comes from [18, Proposition 6.4], and the last identity holds since φ0 (z) = −φ0 (−z). Next, we claim that (a) The function h (z) := ρ0 (z) 1+bρ00 (z) is increasing in z; (b) proxbρ (z) is increasing in z. These two claims imply that ρ0 proxbρ (−τ z)
1 + bρ00 proxbρ (−τ z)
−
ρ0 proxbρ (τ z)
< 0, 1 + bρ00 proxbρ (τ z) ∀z > 0, which combined with the fact φ0 (z) < 0 for z > 0 reveals sign ρ0 proxbρ (−τ z)

ρ0 proxbρ (τ z)
−
1 + bρ00 proxbρ (−τ z) 1 + bρ00 proxbρ (τ z) ! ! 0 φ (z) = 1, ∀z > 0. In other words, the integrand in (114) is positive, which allows one to conclude that G0 (b) > 0. We then move on to justify (a) and (b). For the first, the derivative of h is given by h0 (z) = ρ00 (z) + b(ρ00 (z))2 − bρ0 (z)ρ000 (z) 2 (1 + bρ00 (z)) . Since ρ0 is log concave, this directly yields (ρ00 )2 −ρ0 ρ000 > 0. As ρ00 > 0 and b ≥ 0, the above implies h0 (z) > 0 ∂proxbρ (z) for all z. The second claim follows from ≥ 1+bkρ1 00 k∞ > 0 (cf. [18, Equation (56)]). ∂z 34 It remains to analyze the behavior of G in the limits when b → 0 and b → ∞. From [18, Proposition 6.4], G(b) can also be expressed as   1 G(b) = 1 − E . 1 + bρ00 (proxbρ (τ Z)) Since ρ00 is bounded and the integrand is at most 1, the dominated convergence theorem gives lim G(b) = 0. b→0 When b → ∞, bρ00 (proxbρ (τ z)) → ∞ for a fixed z. Again by applying the dominated convergence theorem, lim G(b) = 1. b→∞ It follows that limb→0 G(b) < κ < limb→∞ G(b) and, therefore, G(b) = κ has a unique positive solution. Remark 3. Finally, we show that the logistic and the probit effective links obey the assumptions of Lemma 5. We work with a fixed τ > 0. • A direct computation shows that ρ0 is log-concave for the logistic model. For the probit, it is well-known that the reciprocal of the hazard function (also known as Mills’ ratio) is strictly log-convex [4]. • To check the other condition, recall that the proximal mapping operator satisfies bρ0 (proxbρ (τ z)) + proxbρ (τ z) = τ z. (115) For a fixed z, we claim that if b → ∞, proxbρ (τ z) → −∞. To prove this claim, we start by assuming that this is not true. Then either proxbρ (τ z) is bounded or diverges to ∞. If it is bounded, the LHS above diverges to ∞ while the RHS is fixed, which is a contradiction. Similarly if proxbρ (τ z) diverges to ∞, the left-hand side of (115) diverges to ∞ while the right-hand side is fixed, which cannot be true as well. Further, when b → ∞, we must have proxbρ (τ z) → −∞, bρ0 (proxbρ (τ z)) → ∞, such that the difference of these two is τ z. Observe that for the logistic, ρ00 (x) = ρ0 (x)(1 − ρ0 (x)) and for the probit, ρ00 (x) = ρ0 (x)(ρ0 (x) − x) [45]. Hence, combining the asymptotic behavior of proxbρ (τ z) and bρ0 (proxbρ (τ z)), we obtain that bρ00 (proxbρ (τ z)) diverges to ∞ in both models when b → ∞. C C.1 Proof of Lemma 6 Proof of Part (i) Recall from [18, Proposition 6.4] that " # 1
. κ = E [Ψ (τ Z; b(τ ))] = 1 − E 1 + b(τ )ρ00 proxb(τ )ρ (τ Z) 0 If we denote c := proxbρ (0), then b(0) is given by the following relation: 1−κ= 1 1 + b(0)ρ00 (c) =⇒ b(0) = κ >0 ρ00 (c)(1 − κ) as ρ00 (c) > 0 for any given c > 0. In addition, since ρ0 (c) > 0, we have V(0) = Ψ(0, b(0))2 (a) b(0)2 ρ0 (c)2 = > 0, κ κ where (a) comes from (20). 35 (116) C.2 Proof of Part (ii) We defer the proof of this part to the supplemental materials [47]. D Proof of Part (ii) of Theorem 4 As discussed in Section 5.2.2, it suffices to (1) construct a set {Bi | 1 ≤ i ≤ N } that forms a cover of the cone A defined in (50), and (2) upper bound P {{Xβ | β ∈ Rp } ∩ Bi 6= {0}}. In what follows, we elaborate on these two steps.
• Step 1. Generate N = exp 22 p i.i.d. points z (i) ∼ N (0, p1 Ip ), 1 ≤ i ≤ N , and construct a collection of convex cones     z (i) p u, (i) Ci := u ∈ R ≥ kuk , 1 ≤ i ≤ N. kz k In words, Ci consists of all directions that have nontrivial positive correlation with z (i) . With high probability, this collection {Ci | 1 ≤ i ≤ N } forms a cover of Rp , a fact which is an immediate consequence of the following lemma.
Lemma 12. Consider any given constant 0 <  < 1, and let N = exp 22 p . Then there exist
some positive universal constants c5 , C5 > 0 such that with probability exceeding 1 − C5 exp −c5 2 p , N X 1{hx,z(i) i≥kxkkz(i) k} ≥ 1 i=1 holds simultaneously for all x ∈ Rp . With our family {Ci | 1 ≤ i ≤ N } we can introduce     n  (i) X √ z Bi := Ci ∩ u ∈ Rn | max {−uj , 0} ≤  n u, (i) ,  kz k  1 ≤ i ≤ N, (117) j=1 which in turn forms a cover of the nonconvex cone A defined in (50). DTo justifyE this, note that for (i) any u ∈ A, one can find i ∈ {1, · · · , N } obeying u ∈ Ci , or equivalently, u, kzz(i) k ≥ kuk, with high probability. Combined with the membership to A this gives n X   √ √ z (i) max {−uj , 0} ≤ 2 nkuk ≤  n u, (i) , kz k j=1 indicating that u is contained within some Bi . • Step 2. We now move on to control P {{Xβ | β ∈ Rp } ∩ Bi 6= {0}}. If the statistical dimensions of the two cones obey δ (Bi ) < n − δ ({Xβ | β ∈ Rp }) = n − p, then an application of [3, Theorem I] gives (  2 ) 1 n − δ ({Xβ | β ∈ Rp }) − δ (Bi ) p √ P {{Xβ | β ∈ R } ∩ Bi 6= {0}} ≤ 4 exp − 8 n ( ) 2 (n − p − δ(Bi )) ≤ 4 exp − . (118) 8n It then comes down to upper bounding δ(Bi ), which is the content of the following lemma. 36 Lemma 13. Fix  > 0. When n is sufficiently large, the statistical dimension of the convex cone Bi defined in (117) obeys   √ 3 √ 1 4 δ(Bi ) ≤ (119) + 2 2 + 10H(2 ) n, 2 where H(x) := −x log x − (1 − x) log(1 − x). Substitution into (118) gives P {{Xβ | β ∈ Rp } ∩ Bi 6= {0}}   2  √ 3 √    1 4 − 10H(2   − 2 2 ) n − p 2 ≤ 4 exp −   8n   (  2 ) √ 3 √ p 1 1 4 − 2 2 − 10H(2 ) − n . = 4 exp − 8 2 n (120) Finally, we prove Lemmas 12-13 in the next subsections. These are the only remaining parts for the proof of Theorem 4. D.1 Proof of Lemma 12 To begin with, it is seen that all kz (i) k concentrates around 1. Specifically, apply [29, Proposition 1] to get r   2t t (i) 2 P kz k > 1 + 2 + ≤ e−t , p p and set t = 32 p to reach o n o n √ 2 P kz (i) k2 > 1 + 10 ≤ P kz (i) k2 > 1 + 2 3 + 62 ≤ e−3 p . Taking the union bound we obtain n o P ∃1 ≤ i ≤ N s.t. kz (i) k2 > 1 + 10 2 ≤ N e−3 p 2 = e− p . (121) Next, we note that it suffices to prove Lemma 12 for all unit vectors x. The following lemma provides a bound on z (i) , x for any fixed unit vector x ∈ Rp . Lemma
14. Consider any fixed unit vector x ∈ Rp and any given constant 0 <  < 1, and set N = exp 22 p . There exist positive universal constants c5 , c6 , C6 > 0 such that P )    7 2 7 1{hz(i) ,xi≥ 1 } ≤ exp (1 − o (1))  p ≤ exp −2 exp (1 − o (1)) 2 p . 2 4 4 i=1 (N X  (122) Recognizing that Lemma 12 is a uniform result, we need to extend Lemma 14 to all  x simultaneously, which we achieve via the standard covering argument. Specifically, one can find a set C := x(j) ∈ Rp | 1 ≤ j ≤ K
p of unit vectors with cardinality K = 1 + 2p2 to form a cover of the unit ball of resolution p−2 [51, Lemma 5.2]; that is, for any unit vector x ∈ Rp , there exists a x(j) ∈ C such that kx(j) − xk ≤ p−2 . 37 Apply Lemma 14 and take the union bound to arrive at   N X 7 1{hz(i) ,x(j) i≥ 1 } ≥ exp (1 − o(1)) 2 p > 1, 1≤j≤K (123) 2 4 i=1

 with probability exceeding 1−K exp −2 exp (1 − o(1)) 47 2 p ≥ 1−exp −2 (1 − o (1)) exp (1 − o(1)) 47 2 p . This guarantees that for each x(j) , one can find at least one z (i) obeying D E 1 z (i) , x(j) ≥ . 2
This result together with (121) yields that with probability exceeding 1 − C exp −c2 p , for some universal constants C, c > 0. E E D E D D E D ≥ z (i) , x(j) − kz (i) k · kx(j) − xk z (i) , x ≥ z (i) , x(j) − z (i) , x(j) − x 1  1 1 1 kz (i) k − 2 kz (i) k  − 2 kz (i) k ≥ √ 2 2 p p 1 + 10 1 kz (i) k ≥ 30 holds simultaneously for all unit vectors x ∈ Rp . Since  > 0 can be an arbitrary constant, this concludes the proof. ≥ Proof of Lemma 14: Without loss of generality, it suffices to consider x = e1 = [1, 0, · · · , 0]> . For any t > 0 and any constant ζ > 0, it comes from [2, Theorem A.1.4] that ( ) N
1 X √ √ 1{hz(i) ,e1 i<ζ } > (1 + t) Φ (ζ p) ≤ exp −2t2 Φ2 (ζ p) N . P N i=1 √
Setting t = 1 − Φ ζ p gives ( ) N   1 X √ √ √ 2 √ P 1{hz(i) ,e1 i<ζ } > (2 − Φ (ζ p)) Φ (ζ p) ≤ exp −2 (1 − Φ (ζ p)) Φ2 (ζ p) N . N i=1 Recall that for any t > 1, one has (t−1 − t−3 )φ(t) ≤ 1 − Φ(t) ≤ t−1 φ(t) which implies that   (1 + o (1)) ζ 2 p √ 1 − Φ (ζ p) = exp − . 2 Taking ζ = 12 , we arrive at √ √ (2 − Φ (ζ p)) Φ (ζ p) = √ 2 √ (1 − Φ (ζ p)) Φ2 (ζ p) =  
1 1 − exp − (1 + o (1)) ζ 2 p = 1 − exp − (1 + o (1)) 2 p , 4  
1 1 exp − (1 + o (1)) ζ 2 p = exp − (1 + o (1)) 2 p  . 4 N This justifies that (N ( )  ) N X 1 2 1 X √ √ P 1{hz(i) ,e1 i≥ 1 } ≤ N exp − (1 + o (1))  p =P 1 (i) > (2 − Φ (ζ p)) Φ (ζ p) 2 4 N i=1 {hz ,e1 i<ζ } i=1     1 2 ≤ exp −2 exp − (1 + o (1))  p N 4    7 = exp −2 exp (1 − o (1)) 2 p 4  as claimed. 38 D.2 Proof of Lemma 13 First of all, recall from the definition (17) that     h i 2 2 2 2 δ(Bi ) = E kΠBi (g)k = E kgk − min kg − uk = n − E min kg − uk u∈Bi u∈Bi   2 ≤ n − E min kg − uk , u∈Di where g ∼ N (0, In ), and Di is a superset of Bi defined by n o Xn √ Di := u ∈ Rn | max {−uj , 0} ≤  nkuk . (124) j=1 Recall from the triangle inequality that kg − uk ≥ kuk − kgk > kgk = kg − 0k, ∀u : kuk > 2kgk. Since 0 ∈ Di , this implies that arg min kg − uk ≤ 2kgk, u∈Di revealing that    2 E min kg − uk = E u∈Di 2  kg − uk min . u∈Di ,kuk≤2kgk In what follows, it suffices to look at the set of u’s within Di obeying kuk ≤ 2kgk, which verify Xn √ √ max {−uj , 0} ≤  nkuk ≤ 2 nkgk. (125) j=1 It is seen that kg − uk2 ≥ X 2 (gi − ui ) = ≥ gi2 + i:gi <0,ui ≥0 ≥ X i:gi <0, − i:gi <0, − X i:gi <0, ui >− √ X + i:gi <0, − i:gi <0, ui ≤− √ n (gi − ui ) n kgk<ui <0 gi2 − 2ui gi √
n kgk<ui <0 X i:gi <0, − 2ui gi . √ (126) n kgk<ui <0 1. Regarding the first term of (126), we first recognize that r Pn   P √ max {−ui , 0}  i: ui <0 |ui | i | ui ≤ − kgk ≤ p  = i=1 p  ≤ 2 n, n n kgk n kgk where the last inequality follows from the constraint (125). As a consequence, X X X gi2 ≥ gi2 − gi2 √ √  i:gi <0 i:gi <0, ui >− 2 (gi − ui )   kgk 2 √ gi2 − n kgk n kgk<ui <0    X + √ X gi2 + i:gi <0,ui ≥0 ≥ X  i:gi <0,ui ≥0 X i:gi <0 X    n kgk i:ui ≤− X ≥ i:gi <0 39 gi2 − n kgk max √ S⊆[n]: |S|=2 n X i∈S gi2 . 2. Next, we turn to the second term of (126), which can be bounded by v  u u X X u  ui gi ≤ u u2i   t √ √ i:gi <0, − n kgk<ui <0 i:gi <0, − ≤ v u u t √ i:gi <0, − ! i:− ≤ n kgk<ui <0  X n kgk<ui <0 v ur u  t kgk n · kgk2 |ui | i:ui <0 ! X n kgk<ui <0 ! X |ui | √ max  gi2  · kgk2 ≤ |ui | √ 3 2 4 kgk2 , i:ui <0 where the last inequality follows from the constraint (125). Putting the above results together, we have X 2 kg − uk ≥ gi2 − i:gi <0 √ S⊆[n]: |S|=2 n for any u ∈ Di obeying kuk ≤ 2kgk, whence    X 2 ≥ E gi2 − E min kg − uk u∈Di i:gi <0  = X max √ 3 1 − 2 2 4 2 √ 3 gi2 − 2 2 4 kgk2 i∈S max √ S⊆[n]: |S|=2 n X  √ 3 gi2 − 2 2 4 kgk2  i∈S # "  n−E max √ S⊆[n]: |S|=2 n X gi2 . (127) i∈S √ Finally, it follows from [29, Proposition 1] that for any t > 2 n, ( ) ) ( X X p 2 2 P gi ≥ 5t ≤ P gi ≥ |S| + 2 |S|t + 2t ≤ e−t , i∈S i∈S which together with the union bound gives ( ) X 2 P max √ gi ≥ 5t ≤ S⊆[n]: |S|=2 n i∈S ( X √ S⊆[n]: |S|=2 n P ) X gi2 ≥ 5t  √
≤ exp H 2  n − t . i∈S This gives " E # max √ S⊆[n]: |S|=2 n X gi2 Z = P 0 i∈S ≤ ( ∞ ) max √ S⊆[n]: |S|=2 n √
5H 2  n + √
< 10H 2  n, Z X gi2 ≥ t dt i∈S ∞   √
1 exp H 2  n − t dt √ 5 5H (2 )n for any given  > 0 with the proviso that n is sufficiently large. This combined with (127) yields     √ 3 √ 1 2 4 E min kg − uk ≥ − 2 2 − 10H(2 ) n u∈Di 2 as claimed. 40 (128) E Proof of Lemma 8 Throughout, we shall restrict ourselves on the event An as defined in (84), on which G̃  λlb I. Recalling the definitions of G̃ and w from (80) and (87), we see that w> G̃−2 w = ≤ −2 1 > X̃ Dβ̃ X̃ X̃ > Dβ̃ X·1 n  −2 1 1 > D X̃ X̃ Dβ̃ X̃ X̃ > Dβ̃ . n β̃ n 1 > X D X̃ n2 ·1 β̃ > X·1 n 2  1/2 If we let the singular value decomposition of √1n Dβ̃ X̃ be U ΣV > , then a little algebra gives Σ  and  −2 1 1/2 1 0 1/2 D X̃ X̃ Dβ̃ X̃ X̃ > Dβ̃ = U Σ−2 U >  λ−1 lb I. n β̃ n (129) √ λlb I Substituting this into (129) and using the fact kX·1 k2 . n with high probability (by Lemma 2), we obtain w> G̃−2 w . 1 kX·1 k2 . 1 nλL with probability at least 1 − exp(−Ω(n)). F Proof of Lemma 9 Throughout this and the subsequent sections, we consider Hn and Kn to be two diverging sequences with the following properties:

Hn = o (n ) , Kn = o (n ) , n2 exp −c1 Hn2 = o(1), n exp −c2 Kn2 = o(1), (130) for any constants ci > 0, i = 1, 2 and any  > 0. This lemma is an analogue of [20, Proposition 3.18]. We modify and adapt the proof ideas to establish the result in our setup. Throughout we shall restrict ourselves to the event An , on which G̃  λlb I. Due to independence between X·1 and {Dβ̃ , H}, one can invoke the Hanson-Wright inequality [44, Theorem 1.1] to yield P !  1 > 1/2 1  1/2 1/2 1/2 > t H, Dβ̃ X D HDβ̃ X·1 − Tr Dβ̃ HDβ̃ n ·1 β̃ n      2 t t  ≤ 2 exp −c min ,  K24 D 1/2 HD 1/2 2 K 2 D 1/2 HD 1/2  n n F β̃ β̃ β̃ β̃      2 t t , ≤ 2 exp −c min ,  K 4 D 1/2 HD 1/2 2 K 2 D 1/2 HD 1/2  n n β̃ β̃ β̃ β̃ √ 1/2 1/2 where k.kF denotes the Frobenius norm. Choose t = C 2 Dβ̃ HDβ̃ Hn / n with C > 0 a sufficiently large constant, and take Hn to be as in (130). Substitution into the above inequality and unconditioning 41 give  P   1 > 1/2 1  1/2 1 2 1/2 1/2 1/2 1/2 √ C Hn kDβ̃ HDβ̃ k X D HDβ̃ X·1 − Tr Dβ̃ HDβ̃ > n ·1 β̃ n n    4 2 √
C Hn C 2 nHn , = C exp −cHn2 = o(1), ≤ 2 exp −c min K4 K2 (131) for some universal constants C, c > 0. 1/2 1/2
We are left to analyzing Tr Dβ̃ HDβ̃ . Recall from the definition (90) of H that 1/2 1/2 Dβ̃ HDβ̃ = Dβ̃ − 1 D X̃ G̃−1 X̃ > Dβ̃ , n β̃ and, hence, Tr   1/2 1/2 Dβ̃ HDβ̃ = n X ρ 00 (X̃i> β̃) i=1 ρ00 (X̃i> β̃)2 > −1 X̃i G̃ X̃i − n ! . (132) This requires us to analyze G̃−1 carefully. To this end, recall that the matrix G̃(i) defined in (81) obeys G̃(i) = G̃ − 1 00 > ρ (X̃ β̃)X̃i X̃i> . n Invoking Sherman-Morrison-Woodbury formula (e.g. [26]), we have G̃ −1 = G̃−1 (i) − ρ00 (X̃i> β̃) −1 G̃(i) X̃i X̃i> G̃−1 (i) n ρ00 (X̃i> β̃) X̃i> G̃−1 (i) X̃i n 1+ It follows that X̃i> G̃−1 X̃i = X̃i> G̃−1 (i) X̃i − . ρ00 (X̃i> β̃) 2 (Xi> G̃−1 (i) X̃i ) n 1+ ρ00 (X̃i> β̃) X̃i> G̃−1 (i) X̃i n (133) , which implies that X̃i> G̃−1 (i) X̃i X̃i> G̃−1 X̃i = 1+ ρ00 (X̃i> β̃) X̃i> G̃−1 (i) X̃i n . (134) The relations (132) and (134) taken collectively reveal that n  1X 1  1/2 1/2 = Tr Dβ̃ HDβ̃ n n i=1 1 + ρ00 (X̃i β̃) ρ00 (X̃i> β̃) X̃i> G̃−1 (i) X̃i n . (135) We shall show that the trace above is close to Tr(I − H) up to some factors. For this purpose we analyze the latter quantity in two different ways. To begin with, observe that 1/2 ! 1/2 Tr(I − H) = Tr Dβ̃ X̃ G̃−1 X̃ > Dβ̃ n = Tr(G̃G̃−1 ) = p − 1. (136) On the other hand, it directly follows from the definition of H and (134) that the ith diagonal entry of H is given by 1 Hi,i = . ρ00 (X̃i> β̃) > G̃−1 X̃ 1+ X̃ i i (i) n 42 Applying this relation, we can compute Tr(I − H) analytically as follows: Tr(I − H) = X i = 1+ X ρ00 (X̃i> β̃) X̃i> G̃−1 (i) X̃i n 1+ 1 n Tr  − ρ00 (X̃i> β̃)α̃ ρ00 (X̃i> β̃) X̃i> G̃−1 (i) X̃i n ρ00 (X̃i> β̃)α̃Hi,i + X ρ00 (X̃i> β̃) i where α̃ := (137) ρ00 (X̃i> β̃) X̃i> G̃−1 (i) X̃i n X ρ00 (X̃i> β̃)α̃ + i = ρ00 (X̃i> β̃) X̃i> G̃−1 (i) X̃i n 1+ i  1 > −1 n X̃i G̃(i) X̃i − α̃ ρ00 (X̃i> β̃) X̃i> G̃−1 (i) X̃i n  , (138)  G̃−1 . 1/2 1/2
Observe that the first quantity in the right-hand side above is simply α̃Tr Dβ̃ HDβ̃ denote 1 ηi = X̃i> G̃−1 (i) X̃i − α̃. n . For simplicity, (139) Note that G̃(i)  0 on An and that ρ00 > 0. Hence the denominator in the second term in (138) is greater than 1 for all i. Comparing (136) and (138), we deduce that  p−1 1 X 00 > 1  1/2 1/2 − Tr Dβ̃ HDβ̃ α̃ ≤ sup |ηi | · |ρ (X̃i β̃)| . sup |ηi | (140) n n n i i i on An . It thus suffices to control supi |ηi |. The above bounds together with Lemma (85) and the proposition below complete the proof. Proposition 1. Let ηi be as defined in (139). Then there exist universal constants C1 , C2 , C3 > 0 such that  

C1 Kn2 Hn √ ≥ 1 − C2 n2 exp −c2 Hn2 − C3 n exp −c3 Kn2 P sup |ηi | ≤ n i − exp (−C4 n (1 + o(1))) = 1 − o(1), where Kn , Hn are diverging sequences as specified in (130) Proof of Proposition 1: Fix any index i. Recall that β̃[−i] is the MLE when the 1st predictor and ith observation are removed. Also recall the definition of G̃[−i] in (83). The proof essentially follows three steps. First, note that X̃i and G̃[−i] are independent. Hence, an application of the Hanson-Wright inequality [44] yields that    !   2 1  −1  t 1 > −1 t  X̃i G̃[−i] X̃i − Tr G̃[−i] > t G̃[−i] ≤ 2 exp −c min , P  K24 G̃−1 2 K 2 G̃−1  n n [−i] n [−i] F n      2 t t . ≤ 2 exp −c min ,  K 4 G̃−1 2 K 2 G̃−1  n [−i] n [−i] √ We choose t = C 2 G̃−1 [−i] Hn / n, where C > 0 is a sufficiently large constant. Now marginalizing gives  P  1 > −1 1  H √n X̃i G̃[−i] X̃i − Tr G̃−1 > C 2 G̃−1 [−i] [−i] n n n 43    ≤ 2 exp −c min
≤ 2 exp −C 0 Hn2 ,  √ C 4 Hn2 C 2 nHn , K4 K2 where C 0 > 0 is a sufficiently large constant. On An , the spectral norm G̃−1 (i) is bounded above by λlb for all i. Invoking (85) we obtain that there exist universal constants C1 , C2 , C3 > 0 such that  
1 > −1 1  −1  Hn P sup X̃i G̃[−i] X̃i − Tr G̃[−i] > C1 √ ≤ C2 n exp −C3 Hn2 . (141) n n n i

−1 > −1 > −1 The next step consists of showing that Tr G̃−1 [−i] (resp. X̃i G̃[−i] X̃i ) and Tr G̃(i) (resp. X̃i G̃(i) X̃i ) are uniformly close across all i. This is established in the following lemma. Lemma 15. Let G̃(i) and G̃[−i] be defined as in (81) and (83), respectively. Then there exist universal constants C1 , C2 , C3 , C4 , c2 , c3 > 0 such that  Kn2 Hn 1 > −1 1 > −1 P sup X̃i G̃(i) X̃i − X̃i G̃[−i] X̃i ≤ C1 √ n n n i

= 1 − C2 n2 exp −c2 Hn2 − C3 n exp −c3 Kn2 − exp (−C4 n (1 + o(1))) = 1 − o(1), (142)   
1
Kn2 Hn 1 −1 √ − ≤ C Tr G̃ P sup Tr G̃−1 1 (i) [−i] n n n i

2 2 = 1 − C2 n exp −c2 Hn − C3 n exp −c3 Kn2 − exp (−C4 n (1 + o(1))) = 1 − o(1), (143) where Kn , Hn are diverging sequences as defined in (130). This together with (141) yields that   1 Kn2 Hn 1 −1 √ X̃ − > C Tr( G̃ ) P sup X̃i> G̃−1 i 1 (i) (i) n n n i

≤ C2 n2 exp −c2 Hn2 + C3 n exp −c3 Kn2 + exp (−C4 n (1 + o(1))) . (144)

1 −1 The final ingredient is to establish that n1 Tr G̃−1 are uniformly close across i. (i) and n Tr G̃ Lemma 16. Let G̃ and G̃(i) be as defined in (80) and (81), respectively. Then one has  

1 −1 − Tr G̃ ≤ P Tr G̃−1 ≥ 1 − exp (−Ω(n)) . (i) λlb (145)  This completes the proof. Proof of Lemma 15: For two invertible matrices A and B of the same dimensions, the difference of their inverses can be written as A−1 − B −1 = A−1 (B − A)B −1 . Applying this identity, we have   −1 −1 G̃−1 − G̃ = G̃ G̃ − G̃ G̃−1 [−i] (i) (i) [−i] (i) [−i] . From the definition of these matrices, it follows directly that

 1 X  00 ρ X̃j> β̃[−i] − ρ00 X̃j> β̃ X̃j X̃j> . G̃[−i] − G̃(i) = n (146) j:j6=i As ρ000 is bounded, by the mean-value theorem, it suffices to control the differences Xj> β̃[−i] − X̃j> β̃ uniformly across all j. This is established in the following lemma, the proof of which is deferred to Appendix H. 44 Lemma 17. Let β̂ be the full model MLE and β̂[−i] be the MLE when the ith observation is dropped. Let qi be as described in Lemma 18 and Kn , Hn be as in (130). Then there exist universal constants C1 , C2 , C3 , C4 , c2 , c3 > 0 such that ! Kn2 Hn > > P sup Xj β̂[−i] − Xj β̂ ≤ C1 √ n j6=i

≥ 1 − C2 n exp −c2 Hn2 − C3 exp −c3 Kn2 − exp (−C4 n (1 + o(1))) = 1 − o(1), (147)   Kn2 Hn P sup |Xi> β̂ − proxqi ρ (Xi> β̂[−i] )| ≤ C1 √ n i

2 ≥ 1 − C2 n exp −c2 Hn − C3 exp −c3 Kn2 − exp (−C4 n (1 + o(1))) = 1 − o(1). (148) Invoking this lemma, we see that the spectral norm of (146) is bounded above by some constant times Kn2 Hn X √ X̃j X̃j> /n n j:j6=i with high probability as specified in (147). From Lemma 2, the spectral norm here is bounded by some constant with probability at least 1 − c1 exp(−c2 n). These observations together with (85) and the fact that on An the minimum eigenvalues of G̃(i) and G̃[−i] are bounded by λlb yield that  

Kn2 Hn −1 √ ≤ C − G̃ P G̃−1 ≥ 1 − C2 n exp −c2 Hn2 − C3 exp −c3 Kn2 − exp (−C4 n (1 + o(1))) . 1 (i) [−i] n This is true for any i. Hence, taking the union bound we obtain   Kn2 Hn −1 √ ≤ C − G̃ P sup G̃−1 1 (i) [−i] n i

≥ 1 − C2 n2 exp −c2 Hn2 − C3 n exp −c3 Kn2 − exp (−C4 n (1 + o(1))) . (149) In order to establish the first result, note that sup i 1 kX̃i k2 −1 > −1 sup kG̃−1 X̃i> G̃−1 X̃ − X̃ G̃ X̃ ≤ sup i i i (i) − G̃[−i] k. (i) [−i] n n i i To obtain the second result, note that sup i p−1 1 1 −1 Tr(G̃−1 ) − Tr(G̃−1 ) ≤ sup kG̃−1 (i) [−i] (i) − G̃[−i] k. n n n i  Therefore, combining (149) and Lemma 2 gives the desired result. Proof of Lemma 16: We restrict ourselves to the event An throughout. Recalling (133), one has −1 Tr(G̃−1 )= (i) ) − Tr(G̃ ρ00 (X̃i> β̃) n 1+ X̃i> G̃−2 (i) X̃i ρ00 (X̃i> β̃) X̃i> G̃−1 (i) X̃i n . In addition, on An we have   1 1 > −2 > −1 X̃i> G̃−1 X̃ − X̃ G̃ X̃ = X̃ G̃ G̃ − λ I G̃−1 i i lb (i) i (i) (i) (i) X̃i ≥ 0. λlb λlb i (i) Combining these results and recognizing that ρ00 > 0, we get −1 Tr(G̃−1 ) ≤ (i) ) − Tr(G̃ ρ00 (X̃i> β̃) n 1+ 1 > −1 λlb X̃i G̃(i) X̃i ρ00 (X̃i> β̃) X̃i> G̃−1 (i) X̃i n ≤ 1 λlb (150)  as claimed. 45 G Proof of Lemma 11 Again, we restrict ourselves to the event An on which G̃  λlb I. Note that X̃i> G̃−1 w = 1 > −1 > X̃ G̃ X̃ Dβ̃ X·1 . n i Note that {G̃, X̃} and X·1 are independent. Conditional on X̃, the left-hand side is Gaussian with mean zero and variance n12 X̃i> G̃−1 X̃ > Dβ̃2 X̃ G̃−1 X̃i . The variance is bounded above by 1 > −1 > 2 1 X̃ G̃ X̃ Dβ̃ X̃ G̃−1 X̃i ≤ sup ρ00 (X̃i> β̃) · 2 X̃i> G̃−1 X̃ > Dβ̃ X̃ G̃−1 X̃i n2 i n i 1 1 = sup ρ00 (X̃i> β̃) · X̃i> G̃−1 X̃i . kX̃i k2 n i n 2 σX := (151) In turn, Lemma 2 asserts that n−1 kX̃i k2 is bounded by a constant with high probability. As a result, applying Gaussian concentration results [49, Theorem 2.1.12] gives |X̃i> G̃−1 w| . Hn
with probability exceeding 1 − C exp −cHn2 , where C, c > 0 are universal constants.
In addition, supi |Xi1 | . Hn holds with probability exceeding 1 − C exp −cHn2 . Putting the above results together, applying the triangle inequality |Xi1 − X̃i> G̃−1 w| ≤ |Xi1 | + |X̃i> G̃−1 w|, and taking the union bound, we obtain
P( sup |Xi1 − X̃i> G̃−1 w| . Hn ) ≥ 1 − Cn exp −cHn2 = 1 − o(1). 1≤i≤n H Proof of Lemma 17 The goal of this section is to prove Lemma 17, which relates the full-model MLE β̂ and the MLE β̂[−i] . To this end, we establish the key lemma below. Lemma 18. Suppose β̂[−i] denote the MLE when the ith observation is dropped. Further let G[−i] be as in (82), and define qi and b̂ as follows: 1 > −1 X G Xi ; n i [−i]  
 1 0 > b̂ = β̂[−i] − G−1 X ρ prox X β̂ . i [−i] q ρ i i n [−i] qi = (152) Suppose Kn , Hn are diverging sequences as in (130). Then there exist universal constants C1 , C2 , C3 > 0 such that   Kn2 Hn P kβ̂ − b̂k ≤ C1 ≥ 1 − C2 n exp(−c2 Hn2 ) − C3 exp(−c3 Kn2 ) − exp(−C4 n(1 + o(1))); (153) n P sup j6=i Xj> β̂[−i] − Xj> b̂ Kn Hn ≤ C1 √ n !

≥ 1 − C2 n exp −c2 Hn2 − C3 exp −c3 Kn2 − exp (−C4 n (1 + o(1))) . (154) 46 The proof ideas are inspired by the leave-one-observation-out approach of [20]. We however emphasize once more that the adaptation of these ideas to our setup is not straightforward and crucially hinges on Theorem 4, Lemma 7 and properties of the effective link function. Proof of Lemma 18: Invoking techniques similar to that for establishing Lemma 7, it can be shown that n 1 X 00 ∗ ρ (γi )Xi Xi>  λlb I n i=1 (155) with probability at least 1 − exp(Ω(n)), where γi∗ is between Xi> b̂ and Xi> β̂. Denote by Bn the event where (155) holds. Throughout this proof, we work on the event Cn := An ∩ Bn , which has probability 1 − exp (−Ω(n)). As in (105) then, 1 ∇`(b̂) . (156) kβ̂ − b̂k ≤ nλlb Next, we simplify (156). To this end, recall the defining relation of the proximal operator bρ0 (proxbρ (z)) + proxbρ (z) = z, which together with the definitions of b̂ and qi gives   Xi> b̂ = proxqi ρ Xi> β̂[−i] . (157) Now, let `[−i] denote the negative log-likelihood function when the ith observation is dropped, and hence

∇`[−i] β̂[−i] = 0. Expressing ∇`(b̂) as ∇`(b̂) − ∇`[−i] β̂[−i] , applying the mean value theorem, and using the analysis similar to that in [20, Proposition 3.4], we obtain  i  1 1 X h 00 ∗ ∇`(b̂) = ρ (γj ) − ρ00 (Xj> β̂[−i] ) Xj Xj> b̂ − β̂[−i] , (158) n n j:j6=i where γj∗ is between Xj> b̂ and Xj> β̂[−i] . Combining (156) and (158) leads to the upper bound kβ̂ − b̂k ≤ 1 λlb  
1 X 1 −1 Xj Xj> · sup ρ00 (γj∗ ) − ρ00 Xj> β̂[−i] · G[−i] Xi · ρ0 proxqi ρ (Xi> β̂[−i] ) . (159) n n j6=i j:j6=i We need to control each term in the right-hand side. To start with, the first term is bounded by a universal constant with probability 1 − exp(−Ω(n)) (Lemma 2). For the second term, since γj∗ is between Xj> b̂ and Xj> β̂[−i] and kρ000 k∞ < ∞, we get sup ρ00 (γj∗ ) − ρ00 (Xj> β̂[−i] ) ≤ kρ000 k∞ kXj> b̂ − Xj> β̂[−i] k (160) j6=i 
 1 > −1 Xj G[−i] Xi ρ0 proxqi ρ Xi> β̂[−i] n   1 0 > ≤ kρ000 k∞ sup Xj> G−1 X · ρ prox (X β̂ ) . i [−i] qi ρ i [−i] n j6=i ≤ kρ000 k∞ (161) (162) Given that {Xj , G[−i] } and Xi are independent for all j 6= i, conditional on {Xj , G[−i] } one has   > −2 Xj> G−1 X ∼ N 0, X G X . i j j [−i] [−i] In addition, the variance satisfies |Xj> G−2 [−i] Xj | ≤ 47 kXj k2 .n λ2lb (163) with probability at least 1 − exp(−Ω(n)). Applying standard Gaussian concentration results [49, Theorem 2.1.12], we obtain  
1 ≥ C H ≤ C2 exp −c2 Hn2 + exp (−C3 n (1 + o(1))) . P √ Xj> G−1 X (164) 1 n i [−i] p By the union bound P 1 √ sup Xj> G−1 [−i] Xi ≤ C1 Hn p j6=i !
≥ 1 − nC2 exp −c2 Hn2 − exp (−C3 n (1 + o(1))) . (165) Consequently,   1 sup ρ00 (γj∗ ) − ρ00 (Xj> β̂[−i] ) . sup kXj> b̂ − Xj> β̂[−i] k . √ Hn ρ0 proxqi ρ (Xi> β̂[−i] ) . n j6=i j6=i In addition, the third term in the right-hand side of (159) can be upper bounded as well since 1 1 1 q > −2 |Xi G[−i] Xi | . √ kG−1 X k = i [−i] n n n (166) (167) with high probability.   It remains to bound ρ0 proxqi ρ (Xi> β̂[−i] ) . To do this, we begin by considering ρ0 (proxcρ (Z)) for any constant c > 0 (rather than a random variable qi ). Recall that for any constant c > 0 and any Z ∼ N (0, σ 2 ) with finite variance, the random variable ρ0 (proxcρ (Z)) is sub-Gaussian. Conditional on β̂[−i] , one has
Xi> β̂[−i] ∼ N 0, kβ̂[−i] k2 . This yields !# "     C32 Kn2 0 > P ρ proxcρ (Xi β̂[−i] ) ≥ C1 Kn ≤ C2 E exp − kβ̂[−i] k2
≤ C2 exp −C3 Kn2 + C4 exp (−C5 n) (168) for some constants C1 , C2 , C3 , C4 , C5 > 0 sincekβ̂[−i] k is bounded with high probability (see Theorem 4). ∂prox (z) bρ ≤ 0 by [18, Proposition 6.3]. Hence, in order to move over from the above concenNote that ∂b tration result established for a fixed constant c to the random variables qi , it suffices to establish a uniform lower bound for qi with high probability. Observe that for each i, qi ≥ kXi k2 1 n G[−i] ≥ C∗ with probability 1 − exp(−Ω(n)), where C ∗ is some universal constant. On this event, one has       ρ0 proxqi ρ Xi> β̂[−i] ≤ ρ0 proxC ∗ ρ Xi> β̂[−i] . This taken collectively with (168) yields     P ρ0 (proxqi ρ (Xi> β̂[−i] )) ≤ C1 Kn ≥ P ρ0 (proxC ∗ ρ (Xi> β̂[−i] )) ≤ C1 Kn
≥ 1 − C2 exp −C3 Kn2 − C4 exp (−C5 n) . (169) (170) This controls the last term. To summarize, if {Kn } and {Hn } are diverging sequences satisfying the assumptions in (130), combining (159) and the bounds for each term in the right-hand side finally gives (153). On the other hand, combining (165) and (170) yields (154).  48 With the help of Lemma 18 we are ready to prove Lemma 17. Indeed, observe that Xj> (β̂[−i] − β̂) ≤ Xj> (b̂ − β̂) + Xj> (β̂[−i] − b̂) , and hence by combining Lemma 2 and Lemma 18, we establish the first claim (147). The second claim (148) follows directly from Lemmas 2, 18 and (157). I Proof of Theorem 7(b)
This section proves that the random sequence α̃ = Tr G̃−1 /n converges in probability to the constant b∗ defined by the system of equations (23) and (24). To begin with, we claim that α̃ is close to a set of auxiliary random variables {q̃i } defined below. Lemma 19. Define q̃i to be q̃i = 1 > −1 X̃ G̃ X̃i , n i [−i] where G̃[−i] is defined in (83). Then there exist universal constants C1 , C2 , C3 , C4 , c2 , c3 > 0 such that   Kn2 Hn P sup |q̃i − α̃| ≤ C1 √ n i

≥ 1 − C2 n2 exp c2 Hn2 − C3 n exp −c3 Kn2 − exp (−C4 n (1 + o(1))) = 1 − o(1), where Kn , Hn are as in (130). Proof: This result follows directly from Proposition 1 and equation (142).   A consequence is that proxq̃i ρ Xi> β̂[−i]    becomes close to proxα̃ρ Xi> β̂[−i] . Lemma 20. Let q̃i and α̃ be as defined earlier. Then one has       Kn3 Hn P sup proxq̃i ρ Xi> β̂[−i] − proxα̃ρ Xi> β̂[−i] ≤ C1 √ n i

2 2 ≥ 1 − C2 n exp −c2 Hn − C3 n exp −c3 Kn2 − exp (−C4 n (1 + o(1))) = 1 − o(1), (171) where Kn , Hn are as in (130).   The key idea behind studying proxα̃ρ Xi> β̂[−i] is that it is connected to a random function δn (·) defined below, which happens to be closely related to the equation (24). In fact,  we will show that δn (α̃) converges in probability to 0; the proof relies on the connection between proxα̃ρ Xi> β̂[−i] and the auxiliary quantity   proxq̃i ρ Xi> β̂[−i] . The formal results is this: Proposition 2. For any index i, let β̂[−i] be the MLE obtained on dropping the ith observation. Define δn (x) to be the random function n δn (x) := p 1X 1    . −1+ n n i=1 1 + xρ00 prox > xρ Xi β̂[−i] P Then one has δn (α̃) → 0. 49 (172) Furthermore, the random function δn (x) converges to a deterministic function ∆(x) defined by   1 ∆(x) = κ − 1 + EZ , 1 + xρ00 (proxxρ (τ∗ Z)) (173) where Z ∼ N (0, 1), and τ∗ is such that (τ∗ , b∗ ) is the unique solution to (23) and (24). P Proposition 3. With ∆(x) as in (173), ∆(α̃) → 0. In fact, one can easily verify that   ∆(x) = κ − E Ψ0 (τ∗ Z; x) , (174) and hence by Lemma 5, the solution to ∆(x) = 0 is exactly b∗ . As a result, putting the above claims together, we show that α̃ converges in probability to b∗ . It remains to formally prove the preceding lemmas and propositions, which is the goal of the rest of this section. Proof of Lemma 20: By [18, Proposition 6.3], one has ∂proxbρ (z) ρ0 (x) =− ∂b 1 + bρ00 (x) , x=proxbρ (z) which yields     sup proxq̃i ρ Xi> β̂[−i] − proxα̃ρ Xi> β̂[−i] i   0 ρ (x) = sup 
· |q̃i − α̃| 1 + qα̃,i ρ00 (x) x=prox > β̂ i X [−i] qα̃,i ρ i   ≤ sup ρ0 proxqα̃,i (Xi> β̂[−i] ) · sup |q̃i − α̃|, i (175) i where qα̃,i is between q̃i and α̃. Here, the last inequality holds since qα̃,i , ρ00 ≥ 0. In addition, just as in the proof of Lemma 18, one can show that qi is bounded below by some constant √ C ∗ > 0 with probability 1 − exp(−Ω(n)). Since qα̃,i ≥ min{q̃i , α̃}, on the event supi |q̃i − α̃| ≤ C1 Kn2 Hn / n, which happens with high probability (Lemma 19), qα̃,i ≥ Cα for some universal constant Cα > 0. Hence, by an argument similar to that establishing (170), we have      P sup ρ0 proxqα̃,i Xi> β̂[−i] ≥ C1 Kn i

≤ C2 n2 exp −c2 Hn2 + C3 n exp −c3 Kn2 + exp (−C4 n (1 + o(1))) .  This together with (175) and Lemma 19 concludes the proof. Proof of Proposition 2: To begin with, recall from (136) and (137) that on An , n ρ00 (X̃i> β̃) X̃i> G̃−1 (i) X̃i n p−1 X = n i=1 1 + ρ00 (X̃i> β̃) n n X̃i> G̃−1 (i) X̃i =1− 50 1X n i=1 1 + 1 ρ00 (X̃i> β̃) X̃i> G̃−1 (i) X̃i n . (176) Using the fact that 1 1+x − 1 1+y ≤ |x − y| for x, y ≥ 0, we obtain n 1X n i=1 1 + n 1 ρ00 (X̃i> β̃) X̃i> G̃−1 (i) X̃i n − 1X 1 00 n i=1 1 + ρ (X̃i> β̃)α̃ n ≤ 1 X 00 > 1 1 > −1 00 ρ (X̃i β̃) X̃i> G̃−1 (i) X̃i − α̃ ≤ kρ k∞ sup n X̃i G̃(i) X̃i − α̃ n i=1 n i Kn2 Hn = kρ00 k∞ sup |ηi | ≤ C1 √ , n i with high probability (Proposition 1). This combined with (176) yields ! n p−1 Kn2 Hn 1X 1 ≥ C1 √ P −1+ n n i=1 1 + ρ00 (X̃i> β̃)α̃ n

≤ C2 n2 exp −c2 Hn2 + C3 n exp −c3 Kn2 + exp (−C4 n (1 + o(1))) . Pn Pn 1 1 1 The above bound concerns n1 i=1 1+ρ00 (X̃ > β̃)α̃ , and it remains to relate it to n i=1 1+ρ00 (prox (X̃ > β̃ ))α̃ . α̃ρ i i To this end, we first get from the uniform boundedness of ρ000 and Lemma 17 that     Kn2 Hn P sup ρ00 (X̃i> β̃) − ρ00 proxq̃i ρ (X̃i> β̃[−i] ) ≥ C1 √ n i ≤ C2 n exp(−c2 Hn2 ) + C3 exp(−c3 Kn2 ) + exp(−C4 n(1 + o(1))). (177) Note that n n 1 1 1X 1X − > 00 00 n i=1 1 + ρ (X̃i β̃)α̃ n i=1 1 + ρ (proxα̃ρ (X̃i> β̃[−i] ))α̃   ≤ |α̃| sup ρ00 (X̃i> β̃) − ρ00 proxα̃ρ (X̃i> β̃[−i] ) i n        o ≤ |α̃| sup ρ00 X̃i> β̃ − ρ00 proxq̃i ρ (X̃i> β̃[−i] ) + ρ00 proxq̃i ρ (X̃i> β̃[−i] ) − ρ00 proxα̃ρ (X̃i> β̃[−i] ) . i By the bound (177), an application of Lemma 20, and the fact that α̃ ≤ p/(nλlb ) (on An ), we obtain ! n 1X 1 Kn3 Hn p −1+ P
≥ C1 √ n n i=1 1 + ρ00 proxα̃ρ (Xi> β̂[−i] ) α̃ n


≤ C2 n2 exp − c2 Hn2 + C3 n exp − c3 Kn2 + exp − C4 n(1 + o(1)) . P This establishes that δn (α̃) → 0.  Proof of Proposition 3: Note that since 0 < α ≤ p/(nλlb ) := B on An , it suffices to show that P sup |δn (x) − ∆(x)| → 0. x∈[0,B] We do this by following three steps. Below, M > 0 is some sufficiently large constant. 1. First we truncate the random function δn (x) and define δ̃n (x) = n X p  −1+ n 00 prox i=1 1 + xρ 51 1  xρ Xi> β̂[−i] 1{kβ̂[−i] k≤M }  . P The first step is to show that supx∈[0,B] δ̃n (x) − δn (x) → 0. We stress that this truncation does not arise in [20], and we keep track of the truncation throughout the rest of the proof.   P 2. Show that supx∈[0,B] δ̃n (x) − E δ̃n (x) → 0.   P 3. Show that supx∈[0,B] E δ̃n (x) − ∆(x) → 0. 1 1 − 1+z | ≤ |y−z| for any y, z > 0 and that To argue about the first step, observe that | 1+y Proposition 6.3]. Then ∂proxcρ (x) ∂x ≤ 1 [18, |δn (x) − δ̃n (x)| ≤ |x| · kρ000 k∞ · sup Xi> β̂[−i] − Xi> β̂[−i] 1{kβ̂[−i] k≤M } . i For a sufficiently large constant M > 0, P(kβ̂[−i] k ≥ M ) ≤ exp(−Ω(n)) by Theorem 4. Hence, for any  > 0, !    > P sup |δn (x) − δ̃n (x)| ≥  ≤ P sup Xi β̂[−i] 1{kβ̂[−i] k≥M } ≥ Bkρ000 k∞ i x∈[0,B] ≤ n   X P kβ̂[−i] k ≥ M = o(1), (178) i=1 establishing Step 1. To argue about Step 2, note that for any x and z, n δ̃n (x) − δ̃n (z) ≤ 1 1X 1      −  n i=1 1 + xρ00 prox > 1 + zρ00 proxzρ Xi> β̂[−i] 1{kβ̂[−i] k≤M } xρ Xi β̂[−i] 1{kβ̂[−i] k≤M } n       1X xρ00 proxxρ Xi> β̂[−i] 1{kβ̂[−i] k≤M } − zρ00 proxzρ Xi> β̂[−i] 1{kβ̂[−i] k≤M } n i=1   n 0 ρ (x) 1 X  00 kρ k∞ |x − z| + |z| · kρ000 k∞ ≤
|x − z| n i=1 1 + z̃ρ00 (x) x=prox X > β̂[−i] 1 z̃ρ i {kβ̂[−i] k≤M } ! n    1X 0 > 00 000 ρ proxz̃ρ Xi β̂[−i] 1{kβ̂[−i] k≤M } ≤ |x − z| kρ k∞ + |z| · kρ k∞ , n i=1 ≤ where z̃ ∈ (x, z). Setting Yn := kρ00 k∞ + Bkρ000 k∞    1 Xn ρ0 proxz̃ρ Xi> β̂[−i] 1{kβ̂[−i] k≤M } , i=1 n then for any , η > 0 we have ! P sup x,z∈(0,B],|x−z|≤η |δ̃n (x) − δ̃n (z)| ≥    η  ≤ E[Yn ] ≤ ηC1 (), ≤ P Yn ≥ η  (179) where C1 () is some function independent of n. The inequality (179) is an analogue of [20, Lemma 3.24]. We remark that the truncation is particularly important here in guaranteeing that E[Yn ] < ∞. Set   Gn (x) := E δ̃n (x) , 52 and observe that   i  h  . Gn (x) − Gn (z) ≤ |x − z| kρ00 k∞ + |z|kρ000 k∞ E ρ0 proxz̃ρ Xi> β̂[−i] 1{kβ̂[−i] k≤M } A similarly inequality applies to ∆(x) in which Xi> β̂[−i] 1{kβ̂[−i] k≤M } is replaced by τ∗ Z. In either case, |Gn (x) − Gn (z)| ≤ C2 η sup |∆(x) − ∆(z)| ≤ C3 η sup and x,z∈(0,B],|x−z|≤η (180) x,z∈(0,B],|x−z|≤η for any η > 0. For any 0 > 0, set K = max{C1 (0 ), C2 }. Next, divide [0, B] into finitely many segments [0, x1 ), [x1 , x2 ), . . . , [xK−1 , xK := B] such that the length of each segment is η/K for any η > 0. Then for every x ∈ [0, B], there exists l such that |x − xl | ≤ η/K. As a result, for any x ∈ [0, B], sup sup δ̃n (x) − δ̃n (xl ) + sup δ̃n (xl ) − Gn (xl ) . x,xl ∈(0,B],|x−xl |≤η/K 1≤l≤k δ̃n (x) − Gn (x) ≤ η + x∈(0,B] Now fix δ > 0,  > 0. Applying the above inequality gives ! P sup |δ̃n (x) − Gn (x)| ≥ δ x∈(0,B] ! δ−η ≤P sup |δ̃n (x) − δ̃n (xl )| ≥ 2 x,xl ∈(0,B],|x−xl |≤η/K   δ−η + P sup |δ̃n (xl ) − Gn (xl )| ≥ 2 1≤l≤k     η δ−η δ−η ≤ C1 + P sup |δ̃n (xl ) − Gn (xl )| ≥ . K 2 2 1≤l≤k (181) Choose η < min{/2, δ}, K = max{C1 ( δ−η 2 ), C2 }. Then the first term in the right-hand side is at most /2. Furthermore, suppose one can establish that for any fixed x, P |δ̃n (x) − Gn (x)| → 0. (182) Since the second term in the right-hand side is a supremum over finitely many points, there exists an integer N such that for all n ≥ N , the second term is less than or equal to /2. Hence, for all n ≥ N , right-hand side is at most , which proves Step 2. Thus, it remains to prove (182) for any fixed x ≥ 0. We do this after the analysis of Step 3. We argue about Step 3 in a similar fashion and letting K = max{C2 , C3 }, divide (0, B] into segments of length η/K. For every x ∈ (0, B] there exists xl such that |x − xl | ≤ η/K. Then |Gn (x) − ∆(x)| ≤ |Gn (x) − Gn (xl )| + |Gn (xl ) − ∆(xl )| + |∆(xl ) − ∆(x)| ≤ 2η + |Gn (xl ) − ∆(xl )|, =⇒ sup |Gn (x) − ∆(x)| ≤ 2η + sup |Gn (xl ) − ∆(xl )|. 1≤l≤k [0,B] Hence, it suffices to show that for any fixed x, |Gn (x) − ∆(x)| → 0. To this end, observe that for any fixed x, |Gn (x) − ∆(x)|  p  ≤ − κ + EX  n 1 + xρ00 prox  1  xρ X1> β̂[−1] 1{kβ̂[−1] k≤M } 53   − EZ " 1 + xρ00 1
proxxρ (τ∗ Z) # . Additionally, X1> β̂[−1] 1{kβ̂[−1] k≤M } = kβ̂[−1] k1{kβ̂[−1] k≤M } Z̃, P where Z̃ ∼ N (0, 1). Since kβ̂[−1] k1{kβ̂[−1] k≤M } → τ∗ , by Slutsky’s theorem, X1> β̂[−1] 1{kβ̂[−1] k≤M } converges weakly to τ∗ Z̃. Since t 7→ 1/(1 + xρ”(proxxρ (t))) is bounded, one directly gets P Gn (x) − ∆(x) → 0 for every x. Finally, we establish (182). To this end, we will prove instead that for any given x, L δn (x) − Gn (x) →2 0. Define Mi := Xi> β̂[−i] 1{kβ̂[−i] k≤M } Then δn (x) − Gn (x) = 1 n Pn i=1 and f (Mi ) := 1 + xρ00 " # 1 1
−E
. proxxρ (Mi ) 1 + xρ00 proxxρ (Mi ) f (Mi ). Hence, for any x ∈ [0, B], n   1 X  2 1 X  E f (Mi ) + 2 E f (Mi )f (Mj ) 2 n i=1 n i6=j  2   E f (M1 ) n(n − 1)  = + E f (M1 )f (M2 ) . n n2 Var[δn (x)] =   The first term in the right-hand side is at most 1/n. Hence, it suffices to show that E f (M1 )f (M2 ) → 0. Let β̂[−12] be the MLE when the 1st and 2nd observations are dropped. Define  1 X 00  > 1 ρ Xj β̂[−12] Xj Xj> , G[−12] := q2 := X2> G−1 [−12] X2 , n n j6=1,2     1 0 > b̂[−1] := β̂[−12] + G−1 X −ρ prox X β̂ . 2 [−12] q2 ρ 2 n [−12] By an application of Lemma 18,  

K 2 Hn ≤ C2 n exp −c2 Hn2 − C3 exp −c3 Kn2 − exp (−C4 n(1 + o(1))) . (183) P β̂[−1] − b̂[−1] ≥ C1 n n Also, by the triangle inequality, X1> (β̂[−1] − β̂[−12] ) ≤ X1> (β̂[−1] − b̂[−1] ) +     1 0 > X1> G−1 −ρ prox X β̂ . X q2 ρ 2 [−12] [−12] 2 n Invoking Lemma 2, (183), and an argument similar to that leading to (170) and (164), we obtain     Kn2 Hn > P X1 β̂[−1] − β̂[−12] ≥ C1 √ n

≤ C2 n exp −c2 Hn2 − C3 exp −c3 Kn2 − exp (−C4 n (1 + o(1))) . The event {kβ̂[−1] k ≤ M } ∩ {kβ̂[−12] k ≤ M } occurs with probability at least 1 − C exp(−cn). Hence, one obtains     Kn2 Hn > P X1 β̂[−1] 1kβ̂[−1] k≤M − β̂[−12] 1kβ̂[−12] k≤M ≤ C1 √ n

≥ C2 n exp −c2 Hn2 − C3 exp −c3 Kn2 − exp (−C4 n (1 + o(1))) . (184) 54 A similar statement continues to hold with X1 replaced by X2 and β̂[−1] replaced by β̂[−2] . Some simple computation yields that kf 0 k∞ is bounded by some constant times |x|. By the mean value theorem and the fact that kf k∞ ≤ 1,     f (M1 )f (M2 ) − f X1> β̂[−12] 1{kβ̂[−12] k≤M } f X2> β̂[−12] 1{kβ̂[−12] k≤M }   n ≤ kf k∞ f (M1 ) − f X1> β̂[−12] 1{kβ̂[−12] k≤M }  o + f (M2 ) − f X2> β̂[−12] 1{kβ̂[−12] k≤M } ≤ C|x| · X1> β̂[−1] 1{kβ̂[−1] k≤M } − X1> β̂[−12] 1{kβ̂[−12] k≤M } + |x| · X2> β̂[−2] 1{kβ̂[−2] k≤M } − X2> β̂[−12] 1{kβ̂[−12] k≤M } . Consequently,     P f (M1 )f (M2 ) − f X1> β̂[−12] 1{kβ̂[−12] k≤M } f X2> β̂[−12] 1{kβ̂[−12] k≤M } → 0. As kf k∞ ≤ 1, this implies convergence in L1 . Thus, it simply suffices to show that h    i E f X1> β̂[−12] 1{kβ̂[−12] k≤M } f X2> β̂[−12] 1{kβ̂[−12] k≤M } → 0. Denote the design matrix on dropping the first and second row as X[−12] . Note that conditional on X[−12] , X1> β̂[−12] 1{kβ̂[−12] k≤M } and X2> β̂[−12] 1{ kβ̂[−12] k ≤ M } are independent and have distribution   N 0, kβ̂[−12] k2 1{kβ̂[−12] k≤M } . Using this and by arguments similar to [20, Lemma 3.23], one can show that " E e  i tX1> β̂[−12] 1{kβ̂ [−12] k≤M } +wX2> β̂[−12] 1{kβ̂  # [−12] k≤M } itX1> β̂[−12] 1{kβ̂ −E e [−12] k≤M }    iwX2> β̂[−12] 1{kβ̂ k≤M } [−12] E e → 0. (185) On repeated application of the multivariate inversion theorem for obtaining densities from characteristic functions, we get that h    i E f X1> β̂[−12] 1{kβ̂[−12] k≤M } f X2> β̂[−12] 1{kβ̂[−12] k≤M } h  i h  i − E f X1> β̂[−12] 1{kβ̂[−12] k≤M } E f X2> β̂[−12] 1{kβ̂[−12] k≤M } → 0. Since f is centered, this completes the proof.  References [1] Alan Agresti and Maria Kateri. Categorical data analysis. Springer, 2011. [2] Noga Alon and Joel H Spencer. The probabilistic method (3rd edition). John Wiley & Sons, 2008. [3] Dennis Amelunxen, Martin Lotz, Michael B McCoy, and Joel A Tropp. Living on the edge: Phase transitions in convex programs with random data. Information and Inference, page iau005, 2014. 55 [4] Árpád Baricz. Mills’ ratio: monotonicity patterns and functional inequalities. Journal of Mathematical Analysis and Applications, 340(2):1362–1370, 2008. [5] Maurice S Bartlett. Properties of sufficiency and statistical tests. Proceedings of the Royal Society of London. Series A, Mathematical and Physical Sciences, pages 268–282, 1937. [6] Mohsen Bayati and Andrea Montanari. The dynamics of message passing on dense graphs, with applications to compressed sensing. IEEE Transactions on Information Theory, 57(2):764–785, 2011. [7] Mohsen Bayati and Andrea Montanari. The LASSO risk for Gaussian matrices. IEEE Transactions on Information Theory, 58(4):1997–2017, 2012. [8] Peter J Bickel and JK Ghosh. A decomposition for the likelihood ratio statistic and the bartlett correction–a bayesian argument. The Annals of Statistics, pages 1070–1090, 1990. [9] Stéphane Boucheron and Pascal Massart. A high-dimensional Wilks phenomenon. Probability theory and related fields, 150(3-4):405–433, 2011. [10] George Box. A general distribution theory for a class of likelihood criteria. Biometrika, 36(3/4):317–346, 1949. [11] Emmanuel Candès, Yingying Fan, Lucas Janson, and Jinchi Lv. Panning for gold: Model-free knockoffs for high-dimensional controlled variable selection. arXiv preprint arXiv:1610.02351, 2016. [12] Herman Chernoff. On the distribution of the likelihood ratio. The Annals of Mathematical Statistics, pages 573–578, 1954. [13] Gauss M Cordeiro. Improved likelihood ratio statistics for generalized linear models. Journal of the Royal Statistical Society. Series B (Methodological), pages 404–413, 1983. [14] Gauss M Cordeiro and Francisco Cribari-Neto. An introduction to Bartlett correction and bias reduction. Springer, 2014. [15] Gauss M Cordeiro, Franciso Cribari-Neto, Elisete CQ Aubin, and Silvia LP Ferrari. Bartlett corrections for one-parameter exponential family models. Journal of Statistical Computation and Simulation, 53(34):211–231, 1995. [16] Thomas M Cover and Joy A Thomas. Elements of information theory. John Wiley & Sons, 2012. [17] Francisco Cribari-Neto and Gauss M Cordeiro. On bartlett and bartlett-type corrections francisco cribari-neto. Econometric reviews, 15(4):339–367, 1996. [18] David Donoho and Andrea Montanari. High dimensional robust M-estimation: Asymptotic variance via approximate message passing. Probability Theory and Related Fields, pages 1–35, 2013. [19] Noureddine El Karoui. Asymptotic behavior of unregularized and ridge-regularized high-dimensional robust regression estimators: rigorous results. arXiv preprint arXiv:1311.2445, 2013. [20] Noureddine El Karoui. On the impact of predictor geometry on the performance on high-dimensional ridge-regularized generalized robust regression estimators. Probability Theory and Related Fields, pages 1–81, 2017. [21] Noureddine El Karoui, Derek Bean, Peter J Bickel, Chinghway Lim, and Bin Yu. On robust regression with high-dimensional predictors. Proceedings of the National Academy of Sciences, 110(36):14557– 14562, 2013. [22] Jianqing Fan and Jiancheng Jiang. Nonparametric inference with generalized likelihood ratio tests. Test, 16(3):409–444, 2007. 56 [23] Jianqing Fan and Jinchi Lv. Nonconcave penalized likelihood with NP-dimensionality. IEEE Transactions on Information Theory, 57(8):5467–5484, 2011. [24] Jianqing Fan, Chunming Zhang, and Jian Zhang. Generalized likelihood ratio statistics and Wilks phenomenon. Annals of statistics, pages 153–193, 2001. [25] Yingying Fan, Emre Demirkaya, and Jinchi Lv. Nonuniformity of p-values can occur early in diverging dimensions. https://arxiv.org/abs/1705.03604, May 2017. [26] William W Hager. Updating the inverse of a matrix. SIAM review, 31(2):221–239, 1989. [27] David L. Hanson and Farroll T. Wright. A bound on tail probabilities for quadratic forms in independent random variables. The Annals of Mathematical Statistics, 42(3):1079–1083, 1971. [28] Xuming He and Qi-Man Shao. On parameters of increasing dimensions. Journal of Multivariate Analysis, 73(1):120–135, 2000. [29] Daniel Hsu, Sham Kakade, and Tong Zhang. A tail inequality for quadratic forms of subgaussian random vectors. Electron. Commun. Probab, 17(52):1–6, 2012. [30] Peter J Huber. Robust regression: asymptotics, conjectures and Monte Carlo. The Annals of Statistics, pages 799–821, 1973. [31] Peter J Huber. Robust statistics. Springer, 2011. [32] Adel Javanmard and Andrea Montanari. State evolution for general approximate message passing algorithms, with applications to spatial coupling. Information and Inference, page iat004, 2013. [33] DN Lawley. A general method for approximating to the distribution of likelihood ratio criteria. Biometrika, 43(3/4):295–303, 1956. [34] Erich L Lehmann and Joseph P Romano. Testing statistical hypotheses. Springer Science & Business Media, 2006. [35] Hua Liang, Pang Du, et al. Maximum likelihood estimation in logistic regression models with a diverging number of covariates. Electronic Journal of Statistics, 6:1838–1846, 2012. [36] Enno Mammen. Asymptotics with increasing dimension for robust regression with applications to the bootstrap. The Annals of Statistics, pages 382–400, 1989. [37] Peter McCullagh and James A Nelder. Generalized linear models. Monograph on Statistics and Applied Probability, 1989. [38] Lawrence H Moulton, Lisa A Weissfeld, and Roy T St Laurent. Bartlett correction factors in logistic regression models. Computational statistics & data analysis, 15(1):1–11, 1993. [39] Neal Parikh and Stephen Boyd. 1(3):127–239, 2014. Proximal algorithms. Foundations and Trends in Optimization, [40] Stephen Portnoy. Asymptotic behavior of M-estimators of p regression parameters when p2 /n is large. i. consistency. The Annals of Statistics, pages 1298–1309, 1984. [41] Stephen Portnoy. Asymptotic behavior of M-estimators of p regression parameters when p2 /n is large; ii. normal approximation. The Annals of Statistics, pages 1403–1417, 1985. [42] Stephen Portnoy. Asymptotic behavior of the empiric distribution of m-estimated residuals from a regression model with many parameters. The Annals of Statistics, pages 1152–1170, 1986. 57 [43] Stephen Portnoy et al. Asymptotic behavior of likelihood methods for exponential families when the number of parameters tends to infinity. The Annals of Statistics, 16(1):356–366, 1988. [44] Mark Rudelson, Roman Vershynin, et al. Hanson-Wright inequality and sub-gaussian concentration. Electron. Commun. Probab, 18(82):1–9, 2013. [45] Michael R Sampford. Some inequalities on mill’s ratio and related functions. The Annals of Mathematical Statistics, 24(1):130–132, 1953. [46] Vladimir Spokoiny. Penalized maximum likelihood estimation and effective dimension. arXiv preprint arXiv:1205.0498, 2012. [47] Pragya Sur, Yuxin Chen, and Emmanuel Candès. Supplemental materials for “the likelihood ratio test in high-dimensional logistic regression is asymptotically a rescaled chi-square”. http: // statweb. stanford. edu/ ~candes/ papers/ supplement_ LRT. pdf , 2017. [48] Cheng Yong Tang and Chenlei Leng. Penalized high-dimensional empirical likelihood. Biometrika, pages 905–919, 2010. [49] Terence Tao. Topics in random matrix theory, volume 132. American Mathematical Society Providence, RI, 2012. [50] A. W. Van der Vaart. Asymptotic statistics, volume 3. Cambridge university press, 2000. [51] Roman Vershynin. Introduction to the non-asymptotic analysis of random matrices. Compressed Sensing: Theory and Applications, pages 210 – 268, 2012. [52] Samuel S Wilks. The large-sample distribution of the likelihood ratio for testing composite hypotheses. The Annals of Mathematical Statistics, 9(1):60–62, 1938. [53] Ting Yan, Yuanzhuang Li, Jinfeng Xu, Yaning Yang, and Ji Zhu. High-dimensional Wilks phenomena in some exponential random graph models. arXiv preprint arXiv:1201.0058, 2012. 58 

Visual Data Augmentation through Learning arXiv:1801.06665v1 [cs.CV] 20 Jan 2018 Grigorios G. Chrysos1 , Yannis Panagakis1,2 , Stefanos Zafeiriou1 1 Department of Computing, Imperial College London, UK 2 Department of Computer Science, Middlesex University London, UK {g.chrysos, i.panagakis, s.zafeiriou}@imperial.ac.uk Abstract The rapid progress in machine learning methods has been empowered by i) huge datasets that have been collected and annotated, ii) improved engineering (e.g. data pre-processing/normalization). The existing datasets typically include several million samples, which constitutes their extension a colossal task. In addition, the state-ofthe-art data-driven methods demand a vast amount of data, hence a standard engineering trick employed is artificial data augmentation for instance by adding into the data cropped and (affinely) transformed images. However, this approach does not correspond to any change in the natural 3D scene. We propose instead to perform data augmentation through learning realistic local transformations. We learn a forward and an inverse transformation that maps an image from the high-dimensional space of pixel intensities to a latent space which varies (approximately) linearly with the latent space of a realistically transformed version of the image. Such transformed images can be considered two successive frames in a video. Next, we utilize these transformations to learn a linear model that modifies the latent spaces and then use the inverse transformation to synthesize a new image. We argue that the this procedure produces powerful invariant representations. We perform both qualitative and quantitative experiments that demonstrate our proposed method creates new realistic images. Figure 1: (Preferably viewed in color) We want to augment arbitrary images (left column) by learning local transformations. We find a low-dimensional space and learn the forward and inverse transformations from the image to the representation space. Then, we can perform a simple linear transformation in the (approximately) linear lowdimensional space and acquire a new synthesized image (the middle column). The same procedure can be repeated with the latest synthesized image (e.g. from the middle to the right columns). work, we propose a new data augmentation technique that finds a low-dimensional space in which performing a simple linear change results in a nonlinear change in the image space. Data augmentation methods are used as label-preserving transformations with twofold goals: a) avoid over-fitting, b) ensure that enough samples have been provided to the network for learning. A plethora of label-preserving transformations have been proposed, however the majority is classified into a) either model-based methods, b) or generic augmentations. The model-based demand an elaborate model to augment the data, e.g. the 3DMM-based [2] face profiling of [29], the novel-view synthesis from 3D models [26]. Such models are available for only a small number of classes and the realistic generation from 3D models/synthetic data is still an open problem [3]. The second augmentation category is comprised of methods defined artificially; these methods do not correspond to any natural 1. Introduction The lack of training data has till recently been an impediment for training machine learning methods. The latest breakthroughs of Neural Networks (NNs) can be partly attributed to the increased amount of (public) databases with massive number of labels/meta-data. Nevertheless, the state-of-the-art networks include tens or hundreds of millions of parameters [8, 37], i.e. they require more labelled examples than we have available. To ameliorate the lack of sufficient labelled examples, different data augmentation methods have become commonplace during training. In this 1 movement in the scene/object. For instance, a 2D image rotation does not correspond to any actual change in the 3D scene space; it is purely a computational method for encouraging rotation invariance. We argue that a third category of augmentations consists of local transformations. We learn a nonlinear transformation that maps the image to a low-dimensional space that is assumed to be (approximately) linear. This linear property allows us to perform a linear transformation and map the original latent representation to the representation of a slightly transformed image (e.g. a pair of successive frames in a video). If we can learn the inverse transform, i.e. mapping from the low-dimensional space to the transformed image, then we can modify the latent representation of the image linearly and this results in a nonlinear change in the image domain. We propose a three-stage approach that learns a forward transformation (from image to low-dimensional representation) and an inverse transformation (from latent to image representation) so that a linear change in the latent space results in a nonlinear change in the image space. The forward and the inverse learned transformations are approximated by an Adversarial Autoencoder and a GAN respectively. In our work, we learn object-specific transformations while we do not introduce any temporal smoothness. Even though learning a generic model for all classes is theoretically plausible, we advocate that with the existing methods, there is not sufficient capacity to learn such generic transformations for all the objects. Instead we introduce objectspecific transformations. Even though we have not explicitly constrained our low-dimensional space to be temporally smooth, e.g. by using the cosine distance, we have observed that the transformations learned are powerful enough to linearize the space. As a visual illustration, we have run TSNE [17] with the latent representations of the first video of 300VW [28] against the rest 49 videos of the published training set; Fig. 2 validates our hypothesis that the latent representations of that video reside in a discrete cluster over the rest of the representations. In a similar experiment with the collected videos of cats, the same conclusion is reached, i.e. the representations of the first video form a discrete cluster. We have opted to report the results in the facial space that is highly nonlinear, while the representations are quite rich. To assess further our approach, we have used two ad-hoc objects, i.e. cat faces, dog faces, that have far less data labelled available online. Additionally, in both ad-hoc objects the shape/appearance presents greater variation than that of human faces, hence more elaborate transformations should be learned. In the following Sections we review the neural networks based on which we have developed our method (Sec. 2.1), introduce our method in Sec. 3. Sequentially, we demon- strate our experimental results in Sec. 4. Due to the restricted space, additional visualizations are deferred to the supplementary material, including indicative figures of the cats’, dogs’ videos, additional (animated) visual results of our method, an experiment illustrating that few images suffice to learn object-specific deformable models. We strongly encourage the reviewers to check the supplementary material. Notation: A small (capital) bold letter represents a vector (matrix); a plain letter designates a scalar number. A vectorized image of a dynamic scene at time t is denoted as (t ) i(t) , while ik k refers to the k th training sample. 2. Background The following lines of research are related with our proposed method: Model-based augmentation for faces: The methods in this category utilize 2D/3D geometric information. In T-CNN [32] the authors introduce an alignment-sensitive method tailored to their task. Namely, they warp a face from its original shape (2D landmarks) to a similar shape (based on their devised clustering). Recently, Zhu et al. [36] use a 3D morphable model (3DMM) [2] to simulate the effect of profiling for synthesizing images in large poses. Tran et al. in [29] fit a 3DMM to estimate the facial pose and learn a GAN conditioned on the pose. During inference, new facial images are synthesized by sampling different poses. The major limitation of the model-based methods is that they require elaborate 2D/3D models. Such models have been studied only for the human face1 or the human body, while the rest objects, e.g. animals faces, have not attracted such attention yet. On the contrary, our method is not limited to any object (we have learned models with cats’ faces and 1 18 years since the original 3DMM model and the problem is not solved for all cases. Figure 2: (Preferably viewed in color) T-SNE [17] in the latent representations of a) 300VW [28] (left Fig.), b) cats’ videos. In both cases the representations of the first video (red dots) are compared against the rest videos (blue dots)). To avoid cluttering the graphs every second frame is skipped (their representation is similar to the previous/next frame). For further emphasis, a green circle is drawn around the red points. dogs’ faces) and does not require elaborate 3D/2D shape models. Unconditional image synthesis: The successful application of GANs [5] in a variety of tasks including photorealistic image synthesis[14], style transfer [34], inpainting [25], image-to-image mapping tasks [11] has led to a proliferation of works on unconditional image synthesis [1, 35]. Even though unconditional image generation has significant applications, it cannot be used for conditional generation when labels are available. Another line of research is directly approximating the conditional distribution over pixels [23]. The generation of a single pixel is conditioned on all the previously generated pixels. Even though realistic samples are produced, it is costly to sample from them; additionally such models do not provide access to the latent representation. Video frames’ prediction: The recent (experimental) breakthroughs of generative models have accelerated the progress in video frames prediction. In [30] the authors learn a model that captures the scene dynamics and synthesizes new frames. To generalize the deterministic prediction of [30], the authors of [33] propose a probabilistic model, however they show only a single frame prediction in low-resolution objects. In addition, the unified latent code z (learned for all objects) does not allow particular motion patterns, e.g. of an object of interest in the video, to be distinguished. Lotter et al. [15] approach the task as a conditional generation. They employ a Recurrent Neural Network (RNN) to condition future frames on previously seen ones, which implicitly imposes temporal smoothness. A core differentiating factor of these approaches from our work is that they i) impose temporal smoothness, ii) make simplifying assumptions (e.g. stationary camera [30]); these restrictions constrain their solution space and allow for realistic video frames’ prediction. In addition, the techniques for future prediction often result in blurry frames, which can be attributed to the multimodal distributions of unconstrained natural images, however our end-goal consists in creating realistic images for highly-complex images, e.g. animals’ faces. The work of [6] is the most similar to our work. The authors construct a customized architecture and loss to linearize the feature space and then perform frame prediction to demonstrate that they have successfully achieved the linearization. Their highly customized architecture (in comparison to our off-the-shelves networks) have not been applied to any highly nonlinear space, in [6] mostly synthetic, simple examples are demonstrated. Apart from the highly nonlinear objects we experiment with, we provide several experimental indicators that our proposed method achieves this linearization in challenging cases. An additional differentiating factor from the aforementioned works is that, to the best of our knowledge, this three- stage approach has not been used in the past for a related task. 2.1. cGAN and Adversarial Autoencoder Let us briefly describe the two methods that consist our workhorse for learning the transformations. These are the conditional GAN and the Adversarial Autoencoder. A Generative Adversarial Network (GAN) [5] is a generative network that has been very successfully employed for learning probability distributions [14]. A GAN is comprised of a generator G and a discriminator D network, where the generator samples from a pre-defined distribution in order to approximate the probability distribution of the training data, while the discriminator tries to distinguish between the samples originating from the model distribution to those from the data distribution. Conditional GAN (cGAN) [20] extends the formulation by conditioning the distributions with additional labels. More formally, if we denote with pd the true distribution of the data, with pz the distribution of the noise, with s the conditioning label and y the data, then the objective function is: LcGAN (G, D) = Es,y∼pd (s,y) [log D(s, y)]+ Es∼pd (s),z∼pz (z) [log(1 − D(s, G(s, z)))] (1) This objective function is optimized in an iterative manner, as min max LcGAN (G, D) = Es,y∼pd (s,y) [log D(s, y; wD )]+ wG wD Es∼pd (s),z∼pz (z) [log(1 − D(s, G(s, z; wG ); wD ))] where wG , wD denote the generator’s, discriminator’s parameters respectively. An Autoencoder (AE) [9, 19] is a neural network with two parts (an encoder and a decoder) and aims to learn a latent representation z of their input y. Autoencoders are mostly used in an unsupervised learning context [12] with the loss being the reconstruction error. On the other hand, an Adversarial Autoencoder (AAE) [18] consists of two sub-networks: i) a generator (an AE network), ii) a discriminator. The discriminator, which is motivated by GAN’s discriminator, accepts the latent vector (generated by the encoder) and tries to match the latent space representation with a pre-defined distribution. 3. Method The core idea of our approach consists in finding a lowdimensional space that is (approximately) linear with respect to the projected representations. We aim to learn the (forward and inverse) transformations from the image space to the low-dimensional space. We know that an image i(t) is an instance of a dynamic scene at time t, hence the difference between the representations of two temporally close Figure 3: The architectures used in (a) separate training per step (the network for Stage I is on the top, for Stage III on the bottom), (b) fine-tuning of the unified model, (c) prediction. The ‘[]’ symbol denotes concatenation. moments should be small and linear. We can learn the linear transitions of the representations and transform our image to i(t+x) . We perform this linearization in 2-steps; an additional step is used to synthesize images of the same object with slightly different representations. The synthesized image can be thought of as a locally transformed image, e.g. the scene at t + x moment with x sufficiently small. 3.1. Stage I: Latent image representation Our goal consists in learning the transformations to the linearized space, however for the majority of the objects there are not enough videos annotated that can express a sufficient percent of the variation. For instance, it is not straightforward to find long videos of all breeds of dogs where the full body is visible. However, there are far more static images available online, which are faster to collect and can be used to learn the transformation from the image space to the latent space. In an unsupervised setting a single image i(t) (per step) suffices for learning latent representations, no additional labels are required, which is precisely the task that Autoencoders were designed for. The latent vector of the Autoencoder lies in the latent space we want to find. We experimentally noticed that the optimization converged faster if we used an adversarial learning procedure. We chose an Adversarial Autoencoder (AAE) [18] with a customized loss function. The encoder feI accepts an image i(t) , encodes it to d(t) ; the decoder fdI reconstructs i(t) . We modify the discriminator to accept both the latent representation and the reconstructed image as input (fake example) and try to distinguish those from the distribution sample and the input image respectively. Moreover, we add a loss term that captures the reconstruction loss, which in our case consists of i) an `1 norm and ii) `1 in the image gradients. Con- sequently, the final loss function is comprised of the following two terms: i) the adversarial loss, ii) the reconstruction loss or: LI = Ladver + λI LIrec (2) with LIrec = ||fdI (feI (y)) − y||`1 + ||∇fdI (feI (y)) − ∇y||`1 (3) (t ) The vector y in this case is a training sample ik k , while λI is a hyper-parameter. 3.2. Stage II: Linear Model Learning In this stage the latent representation d(t) of an image i(t) (as learned from stage I) is used to learn a mapping to the latent representation d(t+x) of the image i(t+x) ; the simple method of linear regression is chosen as a very simple transformation we can perform in a linear space. Given N pairs of images2 (t ) (t +x ) (t ) (t +x ) (t ) (t +x ) {(i1 1 , i1 1 1 ), (i2 2 , i2 2 2 ), . . . , (iNN , iNN N )}, the set of the respective latent representations D = (t ) (t +x ) (t +x ) (t ) (t +x ) (t ) {(d1 1 , d1 1 1 ), (d2 2 , d2 2 2 ), . . . , (dNN , dNN N )}; the set D is used to learn the linear mapping: d(tj +xj ) = A · [d(tj ) ; 1] +  (4) where  is the noise; the Frobenius norm of the residual consists the error term: L = ||d(tj +xj ) − A · [d(tj ) ; 1]||2F (5) To ensure the stability of the linear transformation we add a Tikhonov regularization term (i.e, Frobenius norm) 2 Each pair includes two highly correlated images, i.e. two nearby frames from a video sequence. 3.4. End-to-end fine-tuning on Eq. 5. That is, II L (tj +xj ) = ||d (tj ) − A · [d ; 1]||2F + λII ||A||2f , (6) with λII a regularization hyper-parameter. The closed-form solution to Eq. 6 is A = Y · X T · (X · X T + λII · I)−1 , (7) where I denotes an identity matrix, X, Y two matrices that contain column-wise the initial and target representations (t ) respectively, i.e. for the k th sample X(:, k) = [dk k ; 1], (t +x ) Y (:, k) = dk k k . 3.3. Stage III: Latent representation to image In this step, we want to learn a transformation from the latent space to the image space, i.e. the inverse transformation of Stage I. In particular, we aim to map the regressed representation dˆ(t+x) to the image i(t+x) . Our prior distribution consists of a low-dimensional space, which we want to map to a high-dimensional space; GANs have experimentally proven very effective in such mappings [14, 25]. A conditional GAN is employed for this step; we condition GAN in both the (regressed) latent representation dˆ(t+x) and the original image i(t) . Conditioning on the original image has experimentally resulted in faster convergence and it might be a significant feature in case of limited amount of training samples. Inspired by the work of [11], we form the generator as an autoencoder denoting the encoder as feIII , the decoder as fdIII . Skip connections are added from the second and fourth layers of the encoder to the respective layers in the decoder with the purpose of allowing the low-level features of the original images to be propagated to the result. In conjunction with [11] and Sec. 3.1, we add a reconstruction loss term as III III III III LIII rec = ||fd (fe (y))−s||`1 +||∇fd (fe (y))−∇s||`1 (8) (t ) where y is a training sample ik k and s is the conditioning (tk−x ) label (original image) ik−x . In addition, we add a loss term that encourages the features of the real/fake samples to be similar. Those features are extracted from the penultimate layer of the AAE’s discriminator. Effectively, this leads the fake (i.e. synthesized) images to have representations that are close to the original image. The final objective function for this step includes three terms, i.e. the adversarial, the reconstruction and the feature loss: LIII = LcGAN + λIII LIII rec + λIII,f eat Lf eat Even though the training in each of the aforementioned three stages is performed separately, all the components are differentiable with respect to their parameters. Hence, Stochastic Gradient Descent (SGD) can be used to fine-tune the pipeline. Not all of the components are required for the finetuning, for instance the discriminator of the Adversarial Autoencoder is redundant. From the network in Stage I, only the encoder is utilized for extracting the latent representations, then linear regression (learned matrix A) can be thought of as a linear fully-connected layer. From network in Stage III, all the components are kept. The overall architecture for fine-tuning is depicted in Fig. 3. 3.5. Prediction The structure of our three-stage pipeline is simplified for performing predictions. The image i(t) is encoded (only the encoder of the network in Stage I is required); the resulting representation d(t) is multiplied by A to obtain dˆ(t+x) , which is fed into the conditional GAN to synthesize a new image î(t+x) . This procedure is visually illustrated in Fig. 3, while more formally: î(t+x) = fdIII (feIII (A · [feI (i(t) ; 1)], i(t) ))) (10) 3.6. Network architectures Our method includes two networks, i.e. an Adversarial Autoencoder for Stage I and a conditional GAN for Stage III. The encoder/decoder of both networks share the same architecture, i.e. 8 convolutional layers followed by batch normalization [10] and LeakyRELU [16]. The discriminator consists of 5 layers in both cases, while the dimensionality of the latent space is 1024 for all cases. Please refer to the table in the supplementary material for further details about the layers. 4. Experiments In this Section we provide the details of the training procedure along with the dedicated qualitative and quantitative results for all three objects, i.e. human faces, cats’ faces and dogs’ faces. Our objective is to demonstrate that this augmentation leads to learning invariances, e.g. deformations, not covered by commonly used techniques. (9) where LcGAN is defined in Eq.1, Lf eat represents the similarity cost imposed on the features from the discriminator’s penultimate layer and λIII , λIII,f eat are scalar hyperparameters. To reduce the amount of hyper-parameters in our work, we have set λIII = λI . 4.1. Implementation details The pairs of images required by the second and third stages, were obtained by sequential frames of that object. Different sampling of x was allowed per frame to increase the variation. To avoid the abrupt changes between pairs (a) Human faces (b) Cats’ faces (c) Dogs’ faces Figure 4: Average variance in the dynamics representation per video for (a) the case of human faces, (b) cats’ faces, (c) dogs’ faces. Figure 5: (Preferably viewed in color) Conditional, iterative prediction from our proposed method. The images on the left are the original ones; then from the left to the right the ith column depicts the (i − 1)th synthesized image (iteration (i − 1)). In both rows, the image on the left is animated, hence if opened with Adobe Acrobat reader the transitions will be auto-played. of frames, the structural similarity (SSIM) of a pair was required to lie in an interval, i.e. the frames with i) zero, ii) excessive movement were omitted. Each of the aforementioned stages was trained separately; after training all of them, we have performed finetuning in the combined model (all stages consist of convolutions). However, as is visually illustrated in figures in the supplementary material there are minor differences in the two models. The results of the fine-tuned model are marginally more photo-realistic, which consists fine-tuning optional. 4.2. Datasets A brief description of the databases utilized for training is provided below: Human faces: The recent dataset of MS Celeb [7] was employed for Stage I (Sec. 3.1). MS Celeb includes 8,5 million facial images of 100 thousand celebrities consisting it one of the largest public datasets for static facial images. In our case, the grayscale images were excluded, while from the remaining images a subset of 2 million random images was sampled. For the following two stages that require pairs of images the dataset of 300 Videos in-the-wild (300VW) [28] was employed. This dataset includes 114 videos with approximately 1 minute duration each. The to- Figure 6: (Preferably viewed in color) Visual results of the synthesized images. There are four columns from the left to the right (split into left and right parts) which depict: (a) the original image, (b) the linear model (PCA + regression), (c) our proposed method, (d) the difference in intensities between the proposed method and the original image. The difference does not depict accurately the pose variation; the gif images in the supplementary material demonstrate the animated movement. Nevertheless, some noticeable changes are the following: a) in the left part in the second, and fifth images there is a considerable 3D rotation (pose variation), b) in the first, third and sixth in the left split there are several deformations (eyes closing, mouth opening etc.), c) in the second image on the right part, the person has moved towards the camera. tal amount of frames sampled for Stage II (Sec. 3.2) is 13 thousand frames; 10 thousand frames are sampled for validation, while the rest are used for training the network in Stage III (Sec. 3.3). Cat faces: The pet dataset of [24] was employed for learning representations of cats’ faces. The dataset includes 37 different breeds of cats and dogs (12 for cats) with approximately 200 images each3 . In addition to those, we collected 1000 additional images, for a total of 2000 images. For the subsequent stages of our pipeline, pairs of images were required, hence we have collected 20 videos with an average duration of 200 frames. The head was detected with the DPM detector of [4] in the first frame and the rest 3 Each image is annotated with a head bounding box. tracked with the robust MDNET tracker of [22]. Since the images of cats are limited, the prior weights learned for the (human) facial experiment were employed (effectively the pre-trained model includes a prior which we adapt for cats). Dog faces: The Stanford dog dataset [13] includes 20 thousand images of dogs from 120 breeds. The annotations are in the body-level, hence the DPM detector was utilized to detect a bounding box of the head. The detected images, i.e. 8 thousand, consisted the input for Stage I of our pipeline. Similarly to the procedure for cats, 30 videos (with average duration of 200 frames) were collected and tracked for Stages II and III. 4.3. Variance in the latent space A quantitative self-evaluation experiment was to measure the variance of latent representations per video. The latent representations of sequential frames should be highly correlated; hence the variance in a video containing the same object should be low. A PCA was learned per video and the cumulative eigenvalue ratio was computed. We repeated the same procedure for all the videos (per object) and then averaged the results. The resulting plots with the average cumulative ratio are visualized in Fig. 4. In the videos of the cats and the dogs, we observe that the first 30 components express 90% of the variance. In the facial videos that are longer (over 1500 frames) the variance is greater, however the first 50 components explain over 90% of the variance. 4.4. Qualitative assessment Considering the sub-space defined by PCA as the latent space and learning a linear regression there is the linear counterpart of our proposed method. To demonstrate the complexity of the task, we have learned a PCA per object4 ; the representations of each pair were extracted, linear regression was performed and then the regressed representations were used to create the new sample. In Fig. 6, we have visualized some results for all three cases (human, cats’ and dogs’ faces). In all cases the images were not seen during the training with the cats’ and dogs’ images being downloaded from the web (all were recently uploaded), while the faces are from WIKI-DB dataset [27]. The visualizations verify our claims that a linear transformation in the latent space, can produce a realistic non-linear transformation in the image domain. In all of the facial images there is a deformation of the mouth, while in the majority of them there is a 3D movement. On the contrary, on the dogs’ and the cats’ faces, the major source of deformation seems to be the 3D rotation. An additional remark is that the linear model, i.e. regressing the components of PCA, does not result in realistic new images, which can be attributed to the linear assumptions of PCA. Aside of the visual assessment of the synthesized images, we have considered whether the new synthesized image is realistic enough to be considered as input itself to the pipelne. Hence, we have run an iterative procedure of applying our method, i.e. the outcome of iteration k becomes the input to iteration k + 1. Such an iterative procedure essentially creates a collection of different images (constrained to include the same object of interest but with slightly different latent representations). Two such collections are depicted in Fig. 5, where the person in the first row performs a 3D movement, while in the second different 4 To provide a fair comparison PCA received the same input as our method (i.e. there was no effort to provide further (geometric) details about the image, the pixel values are the only input). deformations of the mouth are observed. The image on the left is animated, hence if opened with Adobe Acrobat reader the transitions will be auto-played. We strongly encourage the reviewers to view the animated images and check the supplementary animations. 4.5. Age estimation with augmented data To ensure that a) our method did not reproduce the input to the output, b) the images are close enough (small change in the representations) we have validated our method by performing age estimation with the augmented data. We utilized as a testbed the AgeDB dataset of [21], which includes 16 thousand manually selected images. As the authors of [21] report, the annotations of AgeDB are accurate to the year, unlike the semi-automatic IMDB-WIKI dataset of [27]. For the aforementioned reasons, we selected AgeDB to perform age estimation with i) the original data, ii) the original plus the new synthesized samples. The first 80% of the images was used as training set and the rest as test-set. We augmented only the training set images with our method by generating one new image for every original one. We discarded the examples that have a structural similarity (SSIM) [31] of less than 0.4 with the original image; this resulted in synthesizing 6 thousand new frames (approximately 50% augmentation). We trained a Resnet-50 [8] with i) the original training images, ii) the augmented images and report here the Mean Absolute Error (MAE). The pre-trained DEX [27] resulted in a MAE of 12.8 years in our test subset [21], the Resnet with the original data in MAE of 11.4 years, while with the augmented data resulted in a MAE of 10.3 years, which is a 9.5% relative decrease in the MAE. That dictates that our proposed method can generate new samples that are not trivially replicated by affine transformations. 5. Conclusion In this work, we have introduced a method that finds a low-dimensional (approximately) linear space. We have introduced a three-stage approach that learns the transformations from the hihgly non-linear image space to the latent space along with the inverse transformation. This approach enables us to make linear changes in the space of representations and these result in non-linear changes in the image space. The first transformation was approximated by an Advervarsial Autoencoder, while a conditional GAN was employed for learning the inverse transformation and acquiring the synthesized image. The middle step consists of a simple linear regression to transform the representations. We have visually illustrated that i) the representations of a video form a discrete cluster (T-SNE in Fig. 2) ii) the representations of a single video are highly correlated (average cumulative eigenvalue ratio for all videos). References [1] M. Arjovsky, S. Chintala, and L. Bottou. Wasserstein gan. arXiv preprint arXiv:1701.07875, 2017. 3 [2] V. Blanz and T. Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, 1999. 1, 2 [3] K. Bousmalis, N. Silberman, D. Dohan, D. Erhan, and D. Krishnan. Unsupervised pixel-level domain adaptation with generative adversarial networks. CVPR, 2017. 1 [4] P. F. Felzenszwalb, R. B. Girshick, D. McAllester, and D. Ramanan. Object detection with discriminatively trained partbased models. T-PAMI, 32(9):1627–1645, 2010. 7 [5] I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio. Generative adversarial nets. In NIPS, 2014. 3 [6] R. Goroshin, M. F. Mathieu, and Y. LeCun. Learning to linearize under uncertainty. In NIPS, pages 1234–1242, 2015. 3 [7] Y. Guo, L. Zhang, Y. Hu, X. He, and J. Gao. Ms-celeb-1m: A dataset and benchmark for large-scale face recognition. In ECCV, pages 87–102, 2016. 6 [8] K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. In CVPR. IEEE, 2016. 1, 8 [9] G. E. Hinton and R. S. Zemel. Autoencoders, minimum description length and helmholtz free energy. In NIPS, 1994. 3 [10] S. Ioffe and C. Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 5 [11] P. Isola, J.-Y. Zhu, T. Zhou, and A. A. Efros. Image-to-image translation with conditional adversarial networks. CVPR, 2017. 3, 5 [12] M. Kan, S. Shan, H. Chang, and X. Chen. Stacked progressive auto-encoders (spae) for face recognition across poses. In CVPR, 2014. 3 [13] A. Khosla, N. Jayadevaprakash, B. Yao, and F.-F. Li. Novel dataset for fine-grained image categorization: Stanford dogs. In CVPR Workshops, volume 2, page 1, 2011. 7 [14] C. Ledig, L. Theis, F. Huszár, J. Caballero, A. Cunningham, A. Acosta, A. Aitken, A. Tejani, J. Totz, Z. Wang, et al. Photo-realistic single image super-resolution using a generative adversarial network. CVPR, 2017. 3, 5 [15] W. Lotter, G. Kreiman, and D. Cox. Deep predictive coding networks for video prediction and unsupervised learning. ICLR, 2017. 3 [16] A. L. Maas, A. Y. Hannun, and A. Y. Ng. Rectifier nonlinearities improve neural network acoustic models. In ICML, 2013. 5 [17] L. v. d. Maaten and G. Hinton. Visualizing data using t-sne. JMLR, 9(Nov):2579–2605, 2008. 2 [18] A. Makhzani, J. Shlens, N. Jaitly, I. Goodfellow, and B. Frey. Adversarial autoencoders. arXiv preprint arXiv:1511.05644, 2015. 3, 4 [19] J. Masci, U. Meier, D. Cireşan, and J. Schmidhuber. Stacked convolutional auto-encoders for hierarchical feature extraction. ICANN, 2011. 3 [20] M. Mirza and S. Osindero. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784, 2014. 3 [21] S. Moschoglou, A. Papaioannou, C. Sagonas, J. Deng, I. Kotsia, and S. Zafeiriou. Agedb: the first manually collected, in-the-wild age database. In CVPR Workshops, 2017. 8 [22] H. Nam and B. Han. Learning multi-domain convolutional neural networks for visual tracking. In CVPR, 2016. 7 [23] A. v. d. Oord, N. Kalchbrenner, and K. Kavukcuoglu. Pixel recurrent neural networks. arXiv preprint arXiv:1601.06759, 2016. 3 [24] O. M. Parkhi, A. Vedaldi, A. Zisserman, and C. Jawahar. Cats and dogs. In CVPR, 2012. 7 [25] D. Pathak, P. Krahenbuhl, J. Donahue, T. Darrell, and A. A. Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536–2544, 2016. 3, 5 [26] K. Rematas, T. Ritschel, M. Fritz, and T. Tuytelaars. Imagebased synthesis and re-synthesis of viewpoints guided by 3d models. In CVPR, 2014. 1 [27] R. Rothe, R. Timofte, and L. Van Gool. Deep expectation of real and apparent age from a single image without facial landmarks. International Journal of Computer Vision, pages 1–14, 2016. 8 [28] J. Shen, S. Zafeiriou, G. Chrysos, J. Kossaifi, G. Tzimiropoulos, and M. Pantic. The first facial landmark tracking inthe-wild challenge: Benchmark and results. In 300-VW in ICCV-W, December 2015. 2, 7 [29] L. Tran, X. Yin, and X. Liu. Disentangled representation learning gan for pose-invariant face recognition. In CVPR, volume 4, page 7, 2017. 1, 2 [30] C. Vondrick, H. Pirsiavash, and A. Torralba. Generating videos with scene dynamics. In NIPS, pages 613–621, 2016. 3 [31] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli. Image quality assessment: from error visibility to structural similarity. TIP, 13(4):600–612, 2004. 8 [32] Y. Wu, T. Hassner, K. Kim, G. Medioni, and P. Natarajan. Facial landmark detection with tweaked convolutional neural networks. arXiv preprint arXiv:1511.04031, 2015. 2 [33] T. Xue, J. Wu, K. Bouman, and B. Freeman. Visual dynamics: Probabilistic future frame synthesis via cross convolutional networks. In NIPS, pages 91–99, 2016. 3 [34] D. Yoo, N. Kim, S. Park, A. S. Paek, and I. S. Kweon. Pixellevel domain transfer. In European Conference on Computer Vision, pages 517–532. Springer, 2016. 3 [35] J. Zhao, M. Mathieu, and Y. LeCun. Energy-based generative adversarial network. arXiv preprint arXiv:1609.03126, 2016. 3 [36] X. Zhu, Z. Lei, X. Liu, H. Shi, and S. Z. Li. Face alignment across large poses: A 3d solution. In CVPR, pages 146–155, 2016. 2 [37] B. Zoph, V. Vasudevan, J. Shlens, and Q. V. Le. Learning transferable architectures for scalable image recognition. arXiv preprint arXiv:1707.07012, 2017. 1 

The power of sum-of-squares for detecting hidden structures arXiv:1710.05017v1 [cs.DS] 13 Oct 2017 Samuel B. Hopkins∗ Prasad Raghavendra Pravesh K. Kothari † Tselil Schramm‡ Aaron Potechin David Steurer§ October 31, 2017 Abstract We study planted problems—finding hidden structures in random noisy inputs—through the lens of the sum-of-squares semidefinite programming hierarchy (SoS). This family of powerful semidefinite programs has recently yielded many new algorithms for planted problems, often achieving the best known polynomial-time guarantees in terms of accuracy of recovered solutions and robustness to noise. One theme in recent work is the design of spectral algorithms which match the guarantees of SoS algorithms for planted problems. Classical spectral algorithms are often unable to accomplish this: the twist in these new spectral algorithms is the use of spectral structure of matrices whose entries are low-degree polynomials of the input variables. We prove that for a wide class of planted problems, including refuting random constraint satisfaction problems, tensor and sparse PCA, densest-k-subgraph, community detection in stochastic block models, planted clique, and others, eigenvalues of degree-d matrix polynomials are as powerful as SoS semidefinite programs of size roughly n d . For such problems it is therefore always possible to match the guarantees of SoS without solving a large semidefinite program. Using related ideas on SoS algorithms and low-degree matrix polynomials (and inspired by recent work on SoS and the planted clique problem [BHK+ 16]), we prove new nearly-tight SoS lower bounds for the tensor and sparse principal component analysis problems. Our lower bounds are the first to suggest that improving upon the signal-to-noise ratios handled by existing polynomial-time algorithms for these problems may require subexponential time. ∗ Cornell University, samhop@cs.cornell.edu Partially supported by an NSF GRFP under grant no. 1144153, by a Microsoft Research Graduate Fellowship, and by David Steurer’s NSF CAREER award. † Princeton University and IAS, kothari@cs.princeton.edu ‡ UC Berkeley, tscrhamm@cs.berkeley.edu. Supported by an NSF Graduate Research Fellowship (1106400). § Cornell University, dsteurer@cs.cornell.edu. Supported by a Microsoft Research Fellowship, a Alfred P. Sloan Fellowship, an NSF CAREER award, and the Simons Collaboration for Algorithms and Geometry. Contents 1 Introduction 1.1 SoS and spectral algorithms for robust inference . . . . . . 1.2 SoS and information-computation gaps . . . . . . . . . . . 1.3 Exponential lower bounds for sparse PCA and tensor PCA 1.4 Related work . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.5 Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2 Distinguishing Problems and Robust Inference 3 Moment-Matching Pseudodistributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 2 4 5 8 9 9 12 4 Proof of Theorem 2.6 15 4.1 Handling Inequalities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22 5 Applications to Classical Distinguishing Problems 23 6 Exponential lower bounds for PCA problems 29 6.1 Predicting sos lower bounds from low-degree distinguishers for Tensor PCA . . . . . 29 6.2 Main theorem and proof overview for Tensor PCA . . . . . . . . . . . . . . . . . . . . 31 6.3 Main theorem and proof overview for sparse PCA . . . . . . . . . . . . . . . . . . . . 33 References 35 A Bounding the sum-of-squares proof ideal term 40 B Lower bounds on the nonzero eigenvalues of some moment matrices 43 C From Boolean to Gaussian lower bounds 45 1 Introduction Recent years have seen a surge of progress in algorithm design via the sum-of-squares (SoS) semidefinite programming hierarchy. Initiated by the work of [BBH+ 12], who showed that polynomial time algorithms in the hierarchy solve all known integrality gap instances for Unique Games and related problems, a steady stream of works have developed efficient algorithms for both worst-case [BKS14, BKS15, BKS17, BGG+ 16] and average-case problems [HSS15, GM15, BM16, RRS16, BGL16, MSS16a, PS17]. The insights from these works extend beyond individual algorithms to characterizations of broad classes of algorithmic techniques. In addition, for a large class of problems (including constraint satisfaction), the family of SoS semidefinite programs is now known to be as powerful as any semidefinite program (SDP) [LRS15]. In this paper we focus on recent progress in using Sum of Squares algorithms to solve averagecase, and especially planted problems—problems that ask for the recovery of a planted signal perturbed by random noise. Key examples are finding solutions of random constraint satisfaction problems (CSPs) with planted assignments [RRS16] and finding planted optima of random polynomials over the n-dimensional unit sphere [RRS16, BGL16]. The latter formulation captures a wide range of unsupervised learning problems, and has led to many unsupervised learning algorithms with the best-known polynomial time guarantees [BKS15, BKS14, MSS16b, HSS15, PS17, BGG+ 16]. In many cases, classical algorithms for such planted problems are spectral algorithms—i.e., using the top eigenvector of a natural matrix associated with the problem input to recover a planted solution. The canonical algorithms for the planted clique [AKS98], principal components analysis (PCA) [Pea01], and tensor decomposition (which is intimately connected to optimizaton of polynomials on the unit sphere) [Har70] are all based on this general scheme. In all of these cases, the algorithm employs the top eigenvector of a matrix which is either given as input (the adjacency matrix, for planted clique), or is a simple function of the input (the empirical covariance, for PCA). Recent works have shown that one can often improve upon these basic spectral methods using SoS, yielding better accuracy and robustness guarantees against noise in recovering planted solutions. Furthermore, for worst case problems—as opposed to the average-case planted problems we consider here—semidefinite programs are strictly more powerful than spectral algorithms.1 A priori one might therefore expect that these new SoS guarantees for planted problems would not be achievable via spectral algorithms. But curiously enough, in numerous cases these stronger guarantees for planted problems can be achieved by spectral methods! The twist is that the entries of these matrices are low-degree polynomials in the input to the algorithm . The result is a new family of low-degree spectral algorithms with guarantees matching SoS but requriring only eigenvector computations instead of general semidefinite programming [HSSS16, RRS16, AOW15a]. This leads to the following question which is the main focus of this work. Are SoS algorithms equivalent to low-degree spectral methods for planted problems? We answer this question affirmatively for a wide class of distinguishing problems which includes refuting random CSPs, tensor and sparse PCA, densest-k-subgraph, community detection in stochastic block models, planted clique, and more. Our positive answer to this question implies 1For example, consider the contrast between the SDP algorithm for Max-Cut of Goemans and Williamson, [GW94], and the spectral algorithm of Trevisan [Tre09]; or the SDP-based algorithms for coloring worst-case 3-colorable graphs [KT17] relative to the best spectral methods [AK97] which only work for random inputs. 1 that a light-weight algorithm—computing the top eigenvalue of a single matrix whose entries are low-degree polynomials in the input—can recover the performance guarantees of an often bulky semidefinite programming relaxation. To complement this picture, we prove two new SoS lower bounds for particular planted problems, both variants of component analysis: sparse principal component analysis and tensor principal component analysis (henceforth sparse PCA and tensor PCA, respectively) [ZHT06, RM14]. For both problems there are nontrivial low-degree spectral algorithms, which have better noise tolerance than naive spectral methods [HSSS16, DM14b, RRS16, BGL16]. Sparse PCA, which is used in machine learning and statistics to find important coordinates in high-dimensional data sets, has attracted much attention in recent years for being apparently computationally intractable to solve with a number of samples which is more than sufficient for brute-force algorithms [KNV+ 15, BR13b, MW15a]. Tensor PCA appears to exhibit similar behavior [HSS15]. That is, both problems exhibit information-computation gaps. Our SoS lower bounds for both problems are the strongest yet formal evidence for informationcomputation gaps for these problems. We rule out the possibility of subexponential-time SoS algorithms which improve by polynomial factors on the signal-to-noise ratios tolerated by the known low degree spectral methods. In particular, in the case of sparse PCA, it appeared possible prior to this work that it might be possible in quasipolynomial time to recover a k-sparse unit vector v in p dimensions from O(k log p) samples from the distribution N(0, Id +vv ⊤ ). Our lower bounds suggest that this is extremely unlikely; in fact this task probably requires polynomial SoS degree and hence exp(n Ω(1) ) time for SoS algorithms. This demonstrates that (at least with regard to SoS algorithms) both problems are much harder than the planted clique problem, previously used as a basis for reductions in the setting of sparse PCA [BR13b]. Our lower bounds for sparse and tensor PCA are closely connected to the failure of low-degree spectral methods in high noise regimes of both problems. We prove them both by showing that with noise beyond what known low-degree spectral algorithms can tolerate, even low-degree scalar algorithms (the result of restricting low-degree spectral algorithms to 1 × 1 matrices) would require subexponential time to detect and recover planted signals. We then show that in the restricted settings of tensor and sparse PCA, ruling out these weakened low-degree spectral algorithms is enough to imply a strong SoS lower bound. 1.1 SoS and spectral algorithms for robust inference We turn to our characterization of SoS algorithms for planted problems in terms of low-degree spectral algorithms. First, a word on planted problems. Many planted problems have several formulations: search, in which the goal is to recover a planted solution, refutation, in which the goal is to certify that no planted solution is present, and distinguishing, where the goal is to determine with good probability whether an instance contains a planted solution or not. Often an algorithm for one version can be parlayed into algorithms for the others, but distinguishing problems are often the easiest, and we focus on them here. A distinguishing problem is specified by two distributions on instances: a planted distribution supported on instances with a hidden structure, and a uniform distribution, where samples w.h.p. contain no hidden structure. Given an instance drawn with equal probability from the planted or the uniform distribution, the goal is to determine with probability greater than 12 whether or not 2 the instance comes from the planted distribution. For example: Planted clique Uniform distribution: G(n, 12 ), the Erdős-Renyi distribution, which w.h.p. contains no clique of size ω(log n). Planted distribution: The uniform distribution on graphs containing a n ε -size clique, for some ε > 0. (The problem gets harder as ε gets smaller, since the distance between the distributions shrinks.) Planted 3xor Uniform distribution: a 3xor instance on n variables and m > n equations x i x j x k  a i jk , where all the triples (i, j, k) and the signs a i jk ∈ {±1} are sampled uniformly and independently. No assignment to x will satisfy more than a 0.51-fraction of the equations, w.h.p. Planted distribution: The same, except the signs a i jk are sampled to correlate with b i b j b k for a randomly chosen b i ∈ {±1}, so that the assignment x  b satisfies a 0.9-fraction of the equations. (The problem gets easier as m/n gets larger, and the contradictions in the uniform case become more locally apparent.) We now formally define a family of distinguishing problems, in order to give our main theorem. Let I be a set of instances corresponding to a product space (for concreteness one may think of n I to be the set of graphs on n vertices, indexed by {0, 1}( 2 ) , although the theorem applies more broadly). Let ν, our uniform distrbution, be a product distribution on I. With some decision problem P in mind (e.g. does G contain a clique of size > n ε ?), let X be a set of solutions to P; again for concreteness one may think of X as being associated with cliques in a graph, so that X ⊂ {0, 1} n is the set of all indicator vectors on at least n ε vertices. For each solution x ∈ X, let µ |x be the uniform distribution over instances I ∈ I that contain x. For example, in the context of planted clique, if x is a clique on vertices 1, . . . , n ε , then µ |x would be the uniform distribution on graphs containing the clique 1, . . . , n ε . We define the planted distribution µ to be the uniform mixture over µ x , µ  U x∼X µ |x . The following is our main theorem on the equivalence of sum of squares algorithms for distinguishing problems and spectral algorithms employing low-degree matrix polynomials. Theorem 1.1 (Informal). Let N, n ∈ N, and let A, B be sets of real numbers. Let I be a family of instances over A N , and let P be a decision problem over I with X  B n the set of possible solutions to P over I. Let {1 j (x, I)} be a system of n O(d) polynomials of degree at most d in the variables x and constant degree in the variables I that encodes P, so that • for I ∼ν I, with high probability the system is unsatisfiable and admits a degree-d SoS refutation, and • for I ∼µ I, with high probability the system is satisfiable by some solution x ∈ X, and x remains feasible even if all but an n −0.01 -fraction of the coordinates of I are re-randomized according to ν. n n Then there exists a matrix whose entries are degree-O(d) polynomials Q : I → ’(6 d)×(6 d) such that I∼ν   λ +max (Q(I)) 6 1, while I∼µ where λ +max denotes the maximum non-negative eigenvalue.   λ +max (Q(I)) > n 10d , The condition that a solution x remain feasible if all but a fraction of the coordinates of I ∼ µ |x are re-randomized should be interpreted as a noise-robustness condition. To see an example, in the context of planted clique, suppose we start with a planted distribution over graphs with a clique x of size n ε+0.01 . If a random subset of n 0.99 vertices are chosen, and all edges not entirely 3 contained in that subset are re-randomized according to the G(n, 1/2) distribution, then with high probability at least n ε of the vertices in x remain in a clique, and so x remains feasible for the problem P: G has a clique of size > n ε ? 1.2 SoS and information-computation gaps Computational complexity of planted problems has become a rich area of study. The goal is to understand which planted problems admit efficient (polynomial time) algorithms, and to study the information-computation gap phenomenon: many problems have noisy regimes in which planted structures can be found by inefficient algorithms, but (conjecturally) not by polynomial time algorithms. One example is the planted clique problem, where the goal find a large clique in a sample from the uniform distribution over graphs containing a clique of size n ε for a small constant ε > 0. While the problem is solvable for any ε > 0 by a brute-force algorithm requiring n Ω(log n) time, polynomial time algorithms are conjectured to require ε > 21 . A common strategy to provide evidence for such a gap is to prove that powerful classes of efficient algorithms are unable to solve the planted problem in the (conjecturally) hard regime. SoS algorithms are particularly attractive targets for such lower bounds because of their broad applicability and strong guarantees. In a recent work, Barak et al. [BHK+ 16] show an SoS lower bound for the planted clique problem, demonstrating that when ε < 12 , SoS algorithms require n Ω(log n) time to solve planted clique. Intriguingly, they show that in the case of planted clique that SoS algorithms requiring ≈ n d time can distinguish planted from random graphs only when there is a scalar-valued degree ≈ d · log n polynomial p(A) : ’n×n → ’ (here A is the adjacency matrix of a graph) with G(n,1/2) p(A)  0, planted p(A) > n Ω(1) ·  – G(n,1/2) p(A)  1/2 . That is, such a polynomial p has much larger expectation in under the planted distribution than its standard deviation in uniform distribution. (The choice of n Ω(1) is somewhat arbitrary, and could be replaced with Ω(1) or n Ω(d) with small changes in the parameters.) By showing that as long as ε < 21 any such polynomial p must have degree Ω(log n)2 , they rule out efficient SoS algorithms when ε < 12 . Interestingly, this matches the spectral distinguishing threshold—the spectral algorithm of [AKS98] is known to work when ε > 21 . This stronger characterization of SoS for the planted clique problem, in terms of scalar distinguishing algorithms rather than spectral distinguishing algorihtms, may at first seem insignificant. To see why the scalar characterization is more powerful, we point out that if the degree-d moments of the planted and uniform distributions are known, determining the optimal scalar distinguishing polynomial is easy: given a planted distribution µ and a random distribution ν over instances I, one just solves a linear algebra problem in the n d log n coefficients of p to maximize the expectation over µ relative to ν: max [p 2 (I)] s.t. [p 2 (I)]  1 . p I∼µ I∼ν It is not difficult to show that the optimal solution to the above program has a simple form: it is the projection of the relative density of ν with respect to µ projected to the degree-d log n polynomials. So given a pair of distributions µ, ν, in n O(d log n) time, it is possible to determine whether there 4 exists a degree-d log n scalar distinguishing polynomial. Answering the same question about the existence of a spectral distinguisher is more complex, and to the best of our knowledge cannot be done efficiently. Given this powerful theorem for the case of the planted clique problem, one may be tempted to conjecture that this stronger, scalar distinguisher characterization of the SoS algorithm applies more broadly than just to the planted clique problem, and perhaps as broadly as Theorem 1.1. If this conjecture is true, given a pair of distributions ν and µ with known moments, it would be possible in many cases to efficiently and mechanically determine whether polynomial-time SoS distinguishing algorithms exist! Conjecture 1.2. In the setting of Theorem 1.1, the conclusion may be replaced with the conclusion that there exists a scalar-valued polynomial p : I → ’ of degree O(d · log n) so that uniform p(I)  0 and planted p(I) > n Ω(1)  2 uniform p(I)  1/2 To illustrate the power of this conjecture, in the beginning of Section 6 we give a short and self-contained explanation of how this predicts, via simple linear algebra, our n Ω(1) -degree SoS lower bound for tensor PCA. As evidence for the conjecture, we verify this prediction by proving such a lower bound unconditionally. We also note why Theorem 1.1 does not imply Conjecture 1.2. While, in the notation of that theorem, the entries of Q(I) are low-degree polynomials in I, the function M 7→ λ +max (M) is not (to the best of our knowledge) a low-degree polynomial in the entries of M (even approximately). (This stands in contrast to, say the operator norm or Frobenious norm of M, both of which are exactly or approximately low-degree polynomials in the entries of M.) This means that the final output of the spectral distinguishing algorithm offered by Theorem 1.1 is not a low-degree polynomial in the instance I. 1.3 Exponential lower bounds for sparse PCA and tensor PCA Our other main results are strong exponential lower bound on the sum-of-squares method (specifΩ(1) ically, against 2n time or n Ω(1) degree algorithms) for the tensor and sparse principal component analysis (PCA). We prove the lower bounds by extending the techniques pioneered in [BHK+ 16]. In the present work we describe the proofs informally, leaving full details to a forthcoming full version. Tensor PCA. We start with the simpler case of tensor PCA, introduced by [RM14]. Problem 1.3 (Tensor PCA). Given an order-k tensor in (’n )⊗k , determine whether it comes from: • Uniform Distribution: each entry of the tensor sampled independently from N(0, 1). • Planted Distribution: a spiked tensor, T  λ · v ⊗k + G where v is sampled uniformly from “n−1 , and where G is a random tensor with each entry sampled independently from N(0, 1). Here, we think of v as a signal hidden by Gaussian noise. The parameter λ is a signal-to-noise ratio. In particular, as λ grows, we expect the distinguishing problem above to get easier. 5 Tensor PCA is a natural generalization of the PCA problem in machine learning and statistics. Tensor methods in general are useful when data naturally has more than two modalities: for example, one might consider a recommender system which factors in not only people and movies but also time of day. Many natural tensor problems are NP hard in the worst-case. Though this is not necessarily an obstacle to machine learning applications, it is important to have average-case models to in which to study algorithms for tensor problems. The spiked tensor setting we consider here is one such simple model. Turning to algorithms: consider first the ordinary PCA problem in a spiked-matrix model. Given an n × n matrix M, the problem is to distinguish between the case where every entry of M is independently drawn from the standard Gaussian distribution N(0, 1) and the case when M is drawn from a distribution as above with an added rank one shift λvv ⊤ in a uniformly random direction v. A natural and well-studied algorithm, which solves this problem to informationtheoretic optimality is to threshold on the largest singular value/spectral norm of the input matrix. Equivalently, one thresholds on the maximizer of the degree two polynomial hx, Mxi in x ∈ “n−1 . A natural generalization of this algorithm to the tensor PCA setting (restricting for simplicity k  3 for this discussion) is the maximum of the degree-three polynomial hT, x ⊗3i over the unit sphere—equivalently, the (symmetric) injective tensor norm of T. This maximum can be shown √ to be much larger in case of the planted distribution so long as λ ≫ n. Indeed, this approach to distinguishing between planted and uniform distributions is information-theoretically optimal [PWB16, BMVX16]. Since recovering the spike v and optimizing the polynomial hT, x ⊗3 i on the sphere are equivalent, tensor PCA can be thought of as an average-case version of the problem of optimizing a degree-3 polynomial on the unit sphere (this problem is NP hard in the worst case, even to approximate [HL09, BBH+12]). Even in this average-case model, it is believed that there is a gap between which signal strengths λ allow recovery of v by brute-force methods and which permit polynomial time algorithms. This is quite distinct from the vanilla PCA setting, where eigenvector algorithms solve the spike-recovery problem to information-theoretic optimality. Nevertheless, the best-known algorithms for tensor PCA arise from computing convex relaxations of this degree-3 polynomial optimization problem. Specifically, the SoS method captures the state of the art algorithms for the problem; it is known to recover the vector v to o(1) error in polynomial time whenever λ ≫ n 3/4 [HSS15]. A major open question in this direction is to understand the complexity of the problem for λ 6 n 3/4−ε . O(ε) Algorithms (again captured by SoS) are known which run in 2n time [RRS16, BGG+ 16]. We show the following theorem which shows that the sub-exponential algorithm above is in fact nearly optimal for SoS algorithm. Theorem 1.4. For a tensor T, let SoS d (T)  max ˜ [hT, x ⊗k i] such that ˜ is a degree d pseudoexpectation and satisfies {kx k 2  1}2 ˜ For every small enough constant ε > 0, if T ∈ ’n×n×n has iid Gaussian or {±1} entries, n k/4−ε , for every d 6 n c·ε for some universal c > 0. T SoS d (T) > In particular for third order tensors (i.e k  3), since degree n Ω(ε) SoS is unable to certify that a 2For definitions of pseudoexpectations and related matters, see the survey [BS14]. 6 random 3-tensor has maximum value much less than n 3/4−ε , this SoS relaxation cannot be used to distinguish the planted and random distributions above when λ ≪ n 3/4−ε .3 Sparse PCA. We turn to sparse PCA, which we formalize as the following planted distinguishing problem. Problem 1.5 (Sparse PCA (λ, k)). Given an n × n symmetric real matrix A, determine whether A comes from: • Uniform Distribution: each upper-triangular entry of the matrix A is sampled iid from N(0, 1); other entries are filled in to preserve symmetry. √ • Planted Distribution: a random k-sparse unit vector v with entries {±1/ k, 0} is sampled, and B is sampled from the uniform distribution above; then A  B + λ · vv⊺ . We defer significant discussion to Section 6, noting just a few things before stating our main theorem on sparse PCA. First, the planted model above is sometimes called the spiked Wigner model—this refers to the independence of the entries of the matrix B. An alternative model for Í sparse PCA is the spiked Wishart model: A is replaced by i 6 m x i x i⊺ , where each x i ∼ N(0, Id +βvv⊺ ), for some number m ∈ Ž of samples and some signal-strength β ∈ ’. Though there are technical differences between the models, to the best of our knowledge all known algorithms with provable guarantees are equally applicable to either model; we expect that our SoS lower bounds also apply in the spiked Wishart model. We generally think of k, λ as small powers of n; i.e. n ρ for some ρ ∈ (0, 1); this allows us to generally ignore logarithmic factors in our arguments. As in the tensor PCA setting, a natural and information-theoretically optimal algorithm for sparse PCA is to maximize the quadratic form hx, Axi, this time over k-sparse unit vectors. For A from the uniform distribution √ standard techniques (ε-nets and union bounds) show that the maximum value achievable is O( k log√n) with high probability, while for A from the planted model of course hv, Avi ≈ λ. So, when λ ≫ k one may distinguish the two models by this maximum value. However, this maximization problem is NP hard for general quadratic forms A [CPR16]. So, efficient algorithms must use some other distinguisher which leverages the randomness in the instances. Essentially only two polynomial-time-computable distinguishers are known.4 If λ ≫ √ n then the maximum eigenvalue of A distinguishes the models. If λ ≫ k then the planted model can be distinguished by the presence of large diagonal entries of A. Notice both of these √ √ distinguishers fail for some choices of λ (that is, k ≪ λ ≪ n, k) for which brute-force methods (optimizing hx, Axi over sparse x) could successfully distinguish planted from uniform A’s. √ The theorem below should be interpreted as an impossibility result for SoS algorithms in the k ≪ √ λ ≪ n, k regime. This is the strongest known impossibility result for sparse PCA among those ruling out classes of efficient algorithms (one reduction-based result is also know, which shows sparse PCA is at least as hard as the planted clique problem [BR13a]. It is also the first evidence that the problem may require subexponential (as opposed to merely quasi-polynomial) time. 3In fact, our proof for this theorem will show somewhat more: that a large family of constraints—any valid constraint which is itself a low-degree polynomial of T—could be added to this convex relaxation and the lower bound would still obtain. 4If one studies the problem at much finer granularity than we do here, in particular studying λ up to low-order additive terms and how precisely it is possible to estimate the planted signal v, then the situation is more subtle [DM14a]. 7 Theorem 1.6. If A ∈ ’n×n , let  SoS d,k (A)  max ˜ hx, Axi s.t. ˜ is degree d and satisfies x 3i  x i , kx k 2  k . ˜ There are absolute constants c, ε ∗ > 0 so that for every ρ ∈ (0, 1) and ε ∈ (0, ε ∗ ), if k  n ρ , then for d 6 n c·ε , SoS d,k (A) > min(n 1/2−ε k, n ρ−ε k) . n A∼{±1} ( 2 ) For more thorough discussion of the theorem, see Section 6.3. 1.4 Related work On interplay of SoS relaxations and spectral methods. As we have already alluded to, many prior works explore the connection between SoS relaxations and spectral algorithms, beginning with the work of [BBH+ 12] and including the followup works [HSS15, AOW15b, BM16] (plus many more). Of particular interest are the papers [HSSS16, MS16b], which use the SoS algorithms to obtain fast spectral algorithms, in some cases running in time linear in the input size (smaller even than the number of variables in the associated SoS SDP). In light of our Theorem 1.1, it is particularly interesting to note cases in which the known SoS lower bounds matching the known spectral algorithms—these problems include planted clique (upper bound: [AKS98], lower bound:5 [BHK+ 16]), strong refutations for random CSPs (upper bound:6 [AOW15b, RRS16], lower bounds: [Gri01b, Sch08, KMOW17]), and tensor principal components analysis (upper bound: [HSS15, RRS16, BGG+ 16], lower bound: this paper). We also remark that our work applies to several previously-considered distinguishing and average-case problems within the sum-of-squares algorithmic framework: block models [MS16a] , densest-k-subgraph [BCC+ 10]; for each of these problems, we have by Theorem 1.1 an equivalence between efficient sum-of-squares algorithms and efficient spectral algorithms, and it remains to establish exactly what the tradeoff is between efficiency of the algorithm and the difficulty of distinguishing, or the strength of the noise. To the best of knowledge, no previous work has attempted to characterize SoS relaxations for planted problems by simpler algorithms in the generality we do here. Some works have considered characterizing degree-2 SoS relaxations (i.e. basic semidefinie programs) in terms of simpler algorithms. One such example is recent work of Fan and Montanari [FM16] who showed that for some planted problems on sparse random graphs, a class of simple procedures called local algorithms performs as well as semidefinite programming relaxations. On strong SoS lower bounds for planted problems. By now, there’s a large body of work that establishes lower bounds on SoS SDP for various average case problems. Beginning with the work of Grigoriev [Gri01a], a long line work have established tight lower bounds for random constraint satisfaction problems [Sch08, BCK15, KMOW17] and planted clique [MPW15, DM15, HKP15, RS15, 5SDP lower bounds for the planted clique problem were known for smaller degrees of sum-of-squares relaxations and for other SDP relaxations before; see the references therein for details. 6There is a long line of work on algorithms for refuting random CSPs, and 3SAT in particular; the listed papers contain additional references. 8 BHK+ 16]. The recent SoS lower bound for planted clique of [BHK+ 16] was particularly influential to this work, setting the stage for our main line of inquiry. We also draw attention to previous work on lower bounds for the tensor PCA and sparse PCA problems in the degree-4 SoS relaxation [HSS15, MW15b]—our paper improves on this and extends our understanding of lower bounds for tensor and sparse PCA to any degree. Tensor principle component analysis was introduced by Montanari and Richard [RM14] who indentified information theoretic threshold for recovery of the planted component and analyzed the maximum likelihood estimator for the problem. The work of [HSS15] began the effort to analyze the sum of squares method for the problem and showed that it yields an efficient algorithm for recovering the planted component with strength ω̃(n 3/4 ). They also established that this threshold is tight for the sum of squares relaxation of degree 4. Following this, Hopkins et al. [HSSS16] showed how to extract a linear time spectral algorithm from the above analysis. Tomioka and Suzuki derived tight information theoretic thresholds for detecting planted components by establishing tight bounds on the injective tensor norm of random tensors [TS14]. Finally, very recently, Raghavendra et. al. and Bhattiprolu et. al. independently showed sub-exponential time algorithms for tensor pca [RRS16, BGL16]. Their algorithms are spectral and are captured by the sum of squares method. 1.5 Organization In Section 2 we set up and state our main theorem on SoS algorithms versus low-degree spectral algorithms. In Section 5 we show that the main theorem applies to numerous planted problems— we emphasize that checking each problem is very simple (and barely requires more than a careful definition of the planted and uniform distributions). In Section 3 and Section 4 we prove the main theorerm on SoS algorithms versus low-degree spectral algorithms. In section 7 we get prepared to prove our lower bound for tensor PCA by proving a structural theorem on factorizations of low-degree matrix polynomials with well-behaved Fourier transforms. In section 8 we prove our lower bound for tensor PCA, using some tools proved in section 9. def Notation. For two matrices A, B, let hA, Bi  Tr(AB). Let kAkFr denote the Frobenius norm, and kAk its spectral norm. For matrix valued functions A, B over I and a distribution ν over I ∼ I, def we will denote hA, Biν  I∼ν hA(I), B(I)i and by kAkFr,ν  ( I∼ν hA(I), A(I)i)1/2 . For a vector of formal variables x  (x1 , . . . , x n ), we use x 6 d to denote the vector consisting of all def monomials of degree at most d in these variables. Furthermore, let us denote X 6 d  (x 6 d )(x 6 d )T . 2 Distinguishing Problems and Robust Inference In this section, we set up the formal framework within which we will prove our main result. Uniform vs. Planted Distinguishing Problems We begin by describing a class of distinguishing problems. For A a set of real numbers, we will use I  A N denote a space of instances indexed by N variables—for the sake of concreteness, it 9 
will be useful to think of I as {0, 1} N ; for example, we could have N  n2 and I as the set of all graphs on n vertices. However, the results that we will show here continue to hold in other contexts, where the space of all instances is ’N or [q]N . Definition 2.1 (Uniform Distinguishing Problem). Suppose that I is the space of all instances, and suppose we have two distributions over I, a product distribution ν (the “uniform” distribution), and an arbitrary distribution µ (the “planted” distribution). In a uniform distinguishing problem, we are given an instance I ∈ I which is sampled with probability 21 from ν and with probability 12 from µ, and the goal is to determine with probability greater than 21 + ε which distribution I was sampled from, for any constant ε > 0. Polynomial Systems In the uniform distinguishing problems that we are interested in, the planted distribution µ will be a distribution over instances that obtain a large value for some optimization problem of interest (i.e. the max clique problem). We define polynomial systems in order to formally capture optimization problems. Program 2.2 (Polynomial System). Let A, B be sets of real numbers, let n, N ∈ Ž, and let I  A N be a space of instances and X ⊆ B n be a space of solutions. A polynomial system is a set of polynomial equalities 1 j (x, I)  0 ∀ j ∈ [m], where {1 j } m are polynomials in the program variables {x i }i∈[n] , representing x ∈ X, and in the j1 instance variables {Ij } j∈[N] , representing I ∈ I. We define degprog (1 j ) to be the degree of 1 j in the program variables, and deginst (1 j ) to be the degree of 1 j in the instance variables. Remark 2.3. For the sake of simplicity, the polynomial system Program 2.2 has no inequalities. Inequalities can be incorporated in to the program by converting each inequality in to an equality with an additional slack variable. Our main theorem still holds, but for some minor modifications of the proof, as outlined in Section 4. A polynomial system allows us to capture problem-specific objective functions as well as problem-specific constraints. For concreteness, consider a quadtratic program which checks if a graph on n vertices contains a clique of size k. We can express this with the polynomial system n over program variables x ∈ ’n and instance variables I ∈ {0, 1}( 2 ) , where Ii j  1 iff there is an edge from i to j, as follows: nÕ i∈[n] o x i − k  0 ∪ {x i (x i − 1)  0}i∈[n] ∪ {(1 − Ii j )x i x j  0}i, j∈( [n]) . 2 Planted Distributions We will be concerned with planted distributions of a particular form; first, we fix a polynomial system of interest S  {1 j (x, I)} j∈[m] and some set X ⊆ B n of feasible solutions for S, so that the 10 program variables x represent elements of X. Again, for concreteness, if I is the set of graphs on n vertices, we can take X ⊆ {0, 1} n to be the set of indicators for subsets of at least n ε vertices. For each fixed x ∈ X, let µ |x denote the uniform distribution over I ∈ I for which the polynomial system {1 j (x, I)} j∈[m] is feasible. The planted distribution µ is given by taking the uniform mixture over the µ |x , i.e., µ ∼ U x∼X [µ |x ]. SoS Relaxations If we have a polynomial system {1 j } j∈[m] where degprog (1 j ) 6 2d for every j ∈ [m], then the degree-2d sum-of-squares SDP relaxation for the polynomial system Program 2.2 can be written as, Program 2.4 (SoS Relaxation for Polynomial System). Let S  {1 j (x, I)} j∈[m] be a polynomial system in instance variables I ∈ I and program variables x ∈ X. If degprog (1 j ) 6 2d for all j ∈ [m], then an SoS relaxation for S is hG j (I), Xi  0 ∀ j ∈ [m] X0 6d 6d where X is an [n]6 d ×[n]6 d matrix containing the variables of the SDP and G j : I → ’[n] ×[n] are matrices containing the coefficients of 1 j (x, I) in x, so that the constraint hG j (I), Xi  0 encodes the constraint 1 j (x, I)  0 in the SDP variables. Note that the entries of G j are polynomials of degree at most deginst (1 j ) in the instance variables. Sub-instances Suppose that I  A N is a family of instances; then given an instance I ∈ I and a subset S ⊆ [N], let IS denote the sub-instance consisting of coordinates within S. Further, for a distribution Θ over subsets of [N], let IS ∼Θ I denote a subinstance generated by sampling S ∼ Θ. Let I↓ denote the set of all sub-instances of an instance I, and let I↓ denote the set of all sub-instances of all instances. Robust Inference Our result will pertain to polynomial systems that define planted distributions whose solutions to sub-instances generalize to feasible solutions over the entire instance. We call this property “robust inference.” Definition 2.5. Let I  A N be a family of instances, let Θ be a distribution over subsets of [N], let S be a polynomial system as in Program 2.2, and let µ be a planted distribution over instances feasible for S. Then the polynomial system S is said to satisfy the robust inference property for probability distribution µ on I and subsampling distribution Θ, if given a subsampling IS of an instance I from µ, one can infer a setting of the program variables x ∗ that remains feasible to S for most settings of IS . Formally, there exists a map x : I↓ → ’n such that  I∼µ,S∼Θ,Ĩ∼ν|IS [x(IS ) is a feasible for S on IS ◦ Ĩ] > 1 − ε(n, d) 11 for some negligible function ε(n, d). To specify the error probability, we will say that polynomial system is ε(n, d)-robustly inferable. Main Theorem We are now ready to state our main theorem. Theorem 2.6. Suppose that S is a polynomial system as defined in Program 2.2, of degree at most 2d in the program variables and degree at most k in the instance variables. Let B > d · k ∈ Ž such that 1. The polynpomial system S is n18B -robustly inferable with respect to the planted distribution µ and the sub-sampling distribution Θ. 2. For I ∼ ν, the polynomial system S admits a degree-d SoS refutation with numbers bounded by n B with probability at least 1 − n18B . Let D ∈ Ž be such that for any subset α ⊆ [N] with |α| > D − 2dk,  [α ⊆ S] 6 S∼Θ 1 n 8B There exists a degree 2D matrix polynomial Q : I → ’[n] + I∼µ [λ max (Q(I))] + I∼ν [λ max (Q(I))] 6 d ×[n] 6 d such that, > n B/2 Remark 2.7. Our argument implies a stronger result that can be stated in terms of the eigenspaces of the subsampling operator. Specifically, suppose we define def Sε   α |  {α ⊆ S} 6 ε S∼Θ  Then, the distinguishing polynomial exhibited by Theorem 2.6 satisfies Q ∈ span{ monomials Iα |α ∈ Sε }. This refinement can yield tighter bounds in cases where all monomials of a certain degree are not equivalent to each other. For example, in the Planted Clique problem, each monomial consists of a subgraph and the right measure of the degree of a sub-graph is the number of vertices in it, as opposed to the number of edges in it. In Section 5, we will make the routine verifications that the conditions of this theorem hold for a variety of distinguishing problems: planted clique (Lemma 5.2), refuting random CSPs (Lemma 5.4, stochastic block models (Lemma 5.6), densest-k-subgraph (Lemma 5.8), tensor PCA (Lemma 5.10), and sparse PCA (Lemma 5.12). Now we will proceed to prove the theorem. 3 Moment-Matching Pseudodistributions We assume the setup from Section 2: we have a family of instances I  A N , a polynomial system S  {1 j (x, I)} j∈[m] with a family of solutions X  B n , a “uniform” distribution ν which is a product distribution over I, and a “planted” distribution µ over I defied by the polynomial system S as described in Section 2. 12 The contrapositive of Theorem 2.6 is that if S is robustly inferable with respect to µ and a distribution over sub-instances Θ, and if there is no spectral algorithm for distinguishing µ and ν, then with high probability there is no degree-d SoS refutation for the polynomial system S (as defined in Program 2.4). To prove the theorem, we will use duality to argue that if no spectral algorithm exists, then there must exist an object which is in some sense close to a feasible solution to the SoS SDP relaxation. Since each I in the support of µ is feasible for S by definition, a natural starting point is the 6d 6d SoS SDP solution for instances I ∼µ I. With this in mind, we let Λ : I → (’[n] ×[n] )+ be an arbitrary function from the support of µ over I to PSD matrices. In other words, we take Λ(I)  µ̂(I) · M(I) where µ̂ is the relative density of µ with respect to ν, so that µ̂(I)  µ(I)/ν(I), and M is some matrix valued function such that M(I)  0 and kM(I)k 6 B for all I ∈ I. Our goal is to find a PSD matrix-valued function P that matches the low-degree moments of Λ in the variables I, while being supported over most of I (rather than just over the support of µ). 6d 6d The function P : I → (’[n] ×[n] )+ is given by the following exponentially large convex program over matrix-valued functions, Program 3.1 (Pseudodistribution Program). min s.t. 2 kP kFr,ν (3.1) hQ, Piν  hQ, Λ′iν P0 Λ′  Λ + η · Id, [n]6 d ×[n]6 d ∀Q : I → ’ 2n 2−2 , deginst (Q) 6 D >η>0 (3.2) (3.3) The constraint (3.2) fixes Tr(P), and so the objective function (3.1) can be viewied as minimizing Tr(P 2 ), a proxy for the collision probability of the distribution, which is a measure of entropy. Remark 3.2. We have perturbed Λ in (3.3) so that we can easily show that strong duality holds in the proof of Claim 3.4. For the remainder of the paper we ignore this perturbation, as we can accumulate the resulting error terms and set η to be small enough so that they can be neglected. The dual of the above program will allow us to relate the existence of an SoS refutation to the existence of a spectral algorithm. Program 3.3 (Low-Degree Distinguisher). max s.t. hΛ, Qiν Q : I → ’[n] 6 d ×[n] 6 d 2 kQ + kFr,ν 6 1, , deginst (Q) 6 D where Q + is the projection of Q to the PSD cone. Claim 3.4. Program 3.3 is a manipulation of the dual of Program 3.1, so that if Program 3.1 has √ optimum c > 1, Program 3.3 as optimum at least Ω( c). 13 Before we present the proof of the claim, we summarize its central consequence in the following theorem: if Program 3.1 has a large objective value (and therefore does not provide a feasible SoS solution), then there is a spectral algorithm. [n]6 d ×[n]6 d be such that Id  M  0. Let λ +max (·) be the Theorem 3.5. Fix a function M : I → ’+ function that gives the largest non-negative eigenvalue of a matrix. Suppose Λ  µ · M then the optimum of Program 3.1 is equal to opt > 1 only if there exists a low-degree matrix polynomial Q such that, I∼µ p [λ+max (Q(I))] > Ω( opt/n d ) while, I∼ν [λ +max (Q(I))] 6 1 . Proof. By Claim 3.4, if the value of Program 3.1 is opt > 1, then there is a polynomial Q achieves a √ value of Ω( opt) for the dual. It follows that I∼µ 1 nd [λ +max (Q(I))] > while I∼ν [λ+max (Q(I))] 6 I∼µ q [hId, Q(I))i] > I∼ν [λ +max (Q(I))2 ] p 1 hΛ, Qi  Ω( opt/n d ), ν nd 6 r I∼ν 2 6 1. kQ + (I)kFr  It is interesting to note that the specific structure of the PSD matrix valued function M plays no role in the above argument—since M serves as a proxy for monomials in the solution as represented by the program variables x ⊗d , it follows that the choice of how to represent the planted solution is not critical. Although seemingly counterintuitive, this is natural because the property of being distinguishable by low-degre distinguishers or by SoS SDP relaxations is a property of ν and µ. We wrap up the section by presenting a proof of the Claim 3.4. Proof of Claim 3.4. We take the Lagrangian dual of Program 3.1. Our dual variables will be some combination of low-degree matrix polynomials, Q, and a PSD matrix A: 2 L(P, Q, A)  kP kFr,ν − hQ, P − Λ′iν − hA, Piν s.t. A  0. It is easy to verify that if P is not PSD, then A can be chosen so that the value of L is ∞. Similarly if there exists a low-degree polynomial upon which P and Λ differ in expectation, Q can be chosen as a multiple of that polynomial so that the value of L is ∞. Now, we argue that Slater’s conditions are met for Program 3.1, as P  Λ′ is strictly feasible. Thus strong duality holds, and therefore min max L(P, Q, A) 6 max min L(P, Q, A). P A0,Q A0,Q P Taking the partial derivative of L(P, Q, A) with respect to P, we have ∂ L(P, Q, A)  2 · P − Q − A. ∂P 14 6d 6d where the first derivative is in the space of functions from I → ’[n] ×[n] . By the convexity of ∂ L as a function of P, it follows that if we set ∂P L  0, we will have the minimizer. Substituting, it follows that 1 1 1 2 kA + Q kFr,ν − hQ, A + Q − Λ′iν − hA, A + Qiν 4 2 2 1 2  max hQ, Λ′iν − kA + Q kFr,ν 4 A0,Q min max L(P, Q, A) 6 max P A0,Q A0,Q (3.4) Now it is clear that the maximizing choice of A is to set A  −Q − , the negation of the negativesemi-definite projection of Q. Thus (3.4) simplifies to min max L(P, Q, A) 6 max hQ, Λ′iν − P A0,Q Q 1 2 kQ + kFr,ν 4 1 4 2 , 6 max hQ, Λiν + η Trν (Q + ) − kQ + kFr,ν Q (3.5) def where we have used the shorthand Trν (Q + )  I∼ν Tr(Q(I)+ ). Now suppose that the low-degree matrix polynomial Q ∗ achieves a right-hand-side value of 1 2 > c. hQ ∗ , Λiν + η · Trν (Q +∗ ) − kQ +∗ kFr,ν 4 Consider Q ′  Q ∗ /kQ +∗ kFr,ν . Clearly kQ +′ kFr,ν  1. Now, multiplying the above inequality through by the scalar 1/kQ +∗ kFr,ν , we have that Trν (Q +∗ ) 1 c − η · + kQ +∗ kFr,ν ∗ ∗ kQ + kFr,ν kQ + kFr,ν 4 1 c − η · n d + kQ +∗ kFr,ν . > ∗ kQ + kFr,ν 4 hQ ′ , Λiν > √ Therefore hQ ′ , Λiν is at least Ω(c 1/2 ), as if kQ +∗ kFr,ν > c then the third term gives the lower bound, and otherwise the first term gives the lower bound. Thus by substituting Q ′, the square root of the maximum of (3.5) within an additive ηn d lower-bounds the maximum of the program max s.t. hQ, Λiν Q : I → ’[n] 2 kQ + kFr,ν 6 d ×[n] 6 d 6 1. This concludes the proof. , deginst (Q) 6 D  4 Proof of Theorem 2.6 We will prove Theorem 2.6 by contradiction. Let us assume that there exists no degree-2D matrix polynomial that distinguishes ν from µ. First, the lack of distinguishers implies the following fact about scalar polynomials. 15 Lemma 4.1. Under the assumption that there are no degree-2D distinguishers, for every degree-D scalar polynomial Q, 2 2 kQ kFr,µ 6 n B kQ kFr,ν Proof. Suppose not, then the degree-2D 1 × 1 matrix polynomial Tr(Q(I)2 ) will be a distinguisher between µ and ν.  Constructing Λ. First, we will use the robust inference property of µ to construct a pseudodistribution Λ. Recall again that we have defined µ̂ to be the relative density of µ with respect to ν, so that µ̂(I)  µ(I)/ν(I). For each subset S ⊆ [N], define a PSD matrix-valued function 6d 6d ΛS : I → (’[n] ×[n] )+ as, ΛS (I)  I′ S [ µ̂(IS ◦ I ′)] · x(IS )6 d (x(IS )6 d )T S where we use IS to denote the restriction of I to S ⊂ [N], and IS ◦ I ′ to denote the instance given by S completing the sub-instance IS with the setting I ′. Notice that ΛS is a function depending only on S def IS —this fact will be important to us. Define Λ  function that satisfies hΛ∅,∅ , 1iν  I∼ν S∼Θ I ′ ∼ν S S∼Θ ΛS . [ µ̂(IS ◦ I ′)]  S Observe that Λ is a PSD matrix-valued S IS IS ◦I ′ ∼ν S [ µ̂(IS ◦ I ′)]  1 S (4.1) Since Λ(I) is an average over ΛS (I), each of which is a feasible solution with high probability, Λ(I) is close to a feasible solution to the SDP relaxation for I. The following Lemma formalizes this intuition. def Define G  span{χS · G j | j ∈ [m], S ⊆ [N]}, and use ΠG to denote the orthogonal projection into G. Lemma 4.2. Suppose Program 2.2 satisfies the ε-robust inference property with respect to planted distribution µ and subsampling distribution Θ and if kx(IS )6 d k22 6 K for all IS then for every G ∈ G, we have ! 1/2 √ hΛ, Giν 6 ε · K · kG(IS ◦ IS )k22 S∼Θ Ĩ ∼ν I∼µ S Proof. We begin by expanding the left-hand side by substituting the definition of Λ. We have hΛ, Giν  S∼Θ I∼ν hΛS (IS ), G(I)i  S∼Θ I∼ν I ′ ∼ν S µ̂(IS ◦ I ′) · hx(IS )6 d (x(IS )6 d )T , G(I)i S And because the inner product is zero if x(IS ) is a feasible solution, 6 6 S∼Θ I∼ν I ′ ∼ν S S∼Θ I∼ν I ′ ∼ν S µ̂(IS ◦ I ′) · ‰[x(IS ) is infeasible for S(I)] · x(IS )6 d S 2 2 · kG(I)kFr µ̂(IS ◦ I ′) · ‰[x(IS ) is infeasible for S(I)] · K · kG(I)kFr S 16 And now letting ĨS denote the completion of IS to I, so that IS ◦ ĨS  I, we note that the above is like sampling I ′ , ĨS independently from ν and then reweighting by µ̂(IS ◦ I ′ ), or equivalently S S taking the expectation over IS ◦ I ′  I ′ ∼ µ and ĨS ∼ ν: S  S∼Θ I ′ ∼µ Ĩ ∼ν S · ‰[x(IS ) is infeasible for S(IS ◦ ĨS )] · K · kG(IS ◦ ĨS )kFr and by Cauchy-Schwarz, 6K· S∼Θ I ′ ∼µ Ĩ ∼ν S · ‰[x(IS ) is infeasible for S(IS ◦ ĨS )] ! 1/2 · S∼Θ I ′ ∼µ Ĩ ∼ν S 2 kG(IS ◦ ĨS )kFr ! 1/2 The lemma follows by observing that the first term in the product above is exactly the nonrobustness of inference probability ε.  Corollary 4.3. If G ∈ G is a degree-D polynomial in I, then under the assumption that there are no degree-2D distinguishers for ν, µ, hΛ, Giν 6 √ ε · K · n B · kGkFr,ν Proof. For each fixing of ĨS , kG(IS ◦ ĨS )k22 is a degree-2D-scalar polynomial in I. Therefore by Lemma 4.1 we have that, I∼µ 2 kG(IS ◦ ĨS )kFr 6 nB · I∼ν 2 kG(IS ◦ ĨS )kFr . Substituting back in the bound in Lemma 4.2 the corollary follows.  Now, since there are no degree-D matrix distinguishers Q, for each S in the support of Θ we can apply reasoning similar to Theorem 3.5 to conclude that there is a high-entropy PSD matrix-valued function PS that matches the degree-D moments of ΛS . Lemma 4.4. If there are no degree-D matrix distinguishers Q for µ, ν, then for each S ∼ Θ, there exists a solution PS to Program 3.1 (with the variable Λ : ΛS ) and kPS kFr,ν 6 n (B+d)/4 6 n B/2 (4.2) This does not follow directly from Theorem 3.5, because a priori a distinguisher for some specific S may only apply to a small fraction of the support of µ. However, we can show that Program 3.1 has large value for ΛS only if there is a distinguisher for µ, ν. Proof. By Claim 3.4, it suffices for us to argue that there is no degree-D matrix polynomial Q which has large inner product with ΛS relative to its Frobenius norm. So, suppose by way of contradiction that Q is a degree-D matrix that distinguishes ΛS , so that hQ, ΛS iν > n B+d but kQ kFr,ν 6 1. It follows by definition of ΛS that n B+d 6 hQ, ΛS iν  I∼ν I ′ ∼ν S µ̂(IS ◦ I ′) · hQ(I), x(IS )6 d (x(IS )6 d )⊤ i S 17   IS ◦I ′ ∼µ IS ∼ν 6 µ  S λ +max  Q(IS ◦ IS ), x(IS ) IS ∼ν Q(IS ◦ IS )  6d (x(IS ) 6d ⊤ · x(IS )6 d ) 2 2  . So, we will show that Q S (I)  I ′∼ν Q(IS ◦ I ′) is a degree-D distinguisher for µ. The degree of Q S S S is at most D, since averaging over settings of the variables cannot increase the degree. Applying our assumption that kx(IS )6 d k22 6 K 6 n d , we already have µ λ +max (Q S ) > n B . It remains to show that ν λ +max (Q S ) is bounded. For this, we use the following fact about the trace. Fact 4.5 (See e.g. Theorem 2.10 in [CC09]). For a function f : ’ → ’ and a symmetric matrix A with Í Í eigendecomposition λ · vv ⊤ , define f (A)  f (λ) · vv ⊤ . If f : ’ → ’ is continuous and convex, then the map A → Tr( f (A)) is convex for symmetric A. The function f (t)  (max{0, t})2 is continuous and convex over ’, so the fact above implies 2 is convex for symmetric A. We can take Q S to be symmetric without loss that the map A → kA+ kFr of generality, as in the argument above we only consider the inner product of Q S with symmetric matrices. Now we have that k(Q S (I))+ k 2Fr  I′ S h Q(IS ◦ I ′) S i ! 2 6 I′ S + Fr  Q(IS ◦ I ′) S  2 + Fr , where the inequality is the definition of convexity. Taking the expectation over I ∼ ν gives us that 2 2 k(Q S )+ kFr,ν 6 kQ + kFr,ν 6 1, which gives us our contradiciton.  def Now, analogous to Λ, set P  S∼Θ PS . Random Restriction. We will exploit the crucial property that Λ and P are averages over functions that depend on subsets of variables. This has the same effect as a random restriction, in that hP, Riν essentially depends on the low-degree part of R. Formally, we will show the following lemma. Lemma 4.6. (Random Restriction) Fix D, ℓ ∈ Ž. For matrix-valued functions R : I → ’ℓ×ℓ and a family of functions {PS : IS → ’ℓ×ℓ }S⊆[N] , and a distribution Θ over subsets of [N], I∼ν S∼Θ hPS (IS ), R(I)i > S∼Θ I∼ν hPS (IS ), R <D S (IS )i 1/2 − ρ(D, Θ) ·  2 kPS kFr,ν S∼Θ where ρ(D, Θ)  max  [α ⊆ S]. α,|α| > D S∼Θ Proof. We first re-express the left-hand side as I∼ν S∼Θ hPS (IS ), R(I)i  S∼Θ I∼ν 18 hPS (IS ), R S (IS )i  12 kRkFr,ν def where R S (IS )  IS [R(I)] obtained by averaging out all coordinates outside S. Splitting the function R S into its low-degree and high-degree parts, R S  R S6D + R >D , then applying a CauchyS Schwartz inequality we get S∼Θ I∼ν hPS (IS ), R S (IS )i > S∼Θ I∼ν hPS (IS ), R <D S (IS )i −  2 kPS kFr,ν S∼Θ  1/2  · >D 2 S∼Θ kR S kFr,ν  1/2 . Expressing R >D (I) in the Fourier basis, we have that over a random choice of S ∼ Θ, S∼Θ 2 kR S>D kFr,ν  Õ α,|α| > D 2  [α ⊆ S] · R̂ 2α 6 ρ(D, Θ) · kRkFr S∼Θ Substituting into the above inequality, the conclusion follows.  Equality Constraints. Since Λ is close to satisfying all the equality constraints G of the SDP, the function P approximately satisfies the low-degree part of G. Specifically, we can prove the following. Lemma 4.7. Let k > deginst (G j ) for all G j ∈ S. With P defined as above and under the conditions of Theorem 2.6 for any function G ∈ G, hP, G 6D iν 6 2 kGkFr,ν n 2B Proof. Recall that G  span{χS · G j | j ∈ [m], S ⊆ [N]} and let ΠG be the orthogonal projection into G. Now, since G ∈ G, G 6D  (ΠG G)6D  (ΠG G 6D−2k )6D + (ΠG G >D−2k )6D . (4.3) Now we make the following claim regarding the effect of projection on to the ideal G, on the degree of a polynomial. Claim 4.8. For every polynomial Q, deg(ΠG Q) 6 deg(Q) + 2k. Furthermore for all α, ΠG Q >α has no monomials of degree 6 α − k Proof. To establish the first part of the claim it suffices to show that ΠG Q ∈ span{χS · G j | |S| 6 deg(Q) + k}, since deg(G j ) 6 k for all j ∈ [m]. To see this, observe that ΠG Q ∈ span{χS · G j | |S| 6 deg(Q) + k} and is orthogonal to every χS · G j with |S| > deg(Q) + k: hΠG Q, χS · G j iν  hQ, ΠG χS · G j iν  hQ, χS · G j iν  hQG j , χS iν  0, where the final equality is because deg(χS ) > deg(G j ) + deg(Q). On the other hand, for every subset S with deg(χS ) 6 α − k, hΠG Q >α , χS · G j i  hQ >α , ΠG χS · G j i  hQ >α , χS · G j i  0, since α > deg(G j ) + deg(χS ) This implies that ΠG Q >α ∈ span{χS · G j | |S| > α − k} which implies that ΠG Q >α has no monomials of degree 6 α − k.  19 Incorporating the above claim into (4.3), we have that G 6D  ΠG G 6D−2k + (ΠG G >D−2k )[D−3k,D] , where the superscript [D − 3k, D] denotes the degree range. Now, hP, G 6D iν  hP, ΠG G 6D−2k iν + hP, (ΠG G >D−2k )[D−3k,D] iν And since ΠG G 6D−2k is of degree at most D we can replace P by Λ,  hΛ, ΠG G 6D−2k iν + hP, (ΠG G >D−2k )[D−3k,D] iν Now bounding the first term using Corollary 4.3 with a n B bound on K,  1 6 8B n  1/2 6 D−2k · n B · (n B · kΠG G∅,∅ kFr,ν ) + hP, (ΠG G >D−2k )[D−3k,D] i And for the latter term we use Lemma 4.6, 1 1 6 D−2k 6 2B kΠG G∅,∅ kFr,ν + 4B n n  2 kPS kFr,ν S  1/2 kGkFr,ν , where we have used the fact that (ΠG G >D−2k )[D−3k,D] is high degree. By property of orthogonal projections, kΠG G >D−2k kFr,ν 6 kG >D−2k kFr,ν 6 kGkFr,ν . Along with the bound on kPS kFr,ν from (4.2), this implies the claim of the lemma.  Finally, we have all the ingredients to complete the proof of Theorem 2.6. Proof of Theorem 2.6. Suppose we sample an instance I ∼ ν, and suppose by way of contradiction this implies that with high probability the SoS SDP relaxation is infeasible. In particular, this implies that there is a degree-d sum-of-squares refutation of the form, −1  a I (x) + Õ j∈[m] 1 Ij (x) · q Ij (x), where a I is a sum-of-squares of polynomials of degree at most 2d in x, and deg(q Ij ) + deg(1 Ij ) 6 2d. 6d 6d Let AI ∈ ’[n] ×[n] be the matrix of coefficients for a I (c) on input I, and let G I be defined similarly Í for j∈[m] 1 j (x) · q j (x). We can rewrite the sum-of-squares refutation as a matrix equality, −1  hX 6 d , AI i + hX 6 d , G I i, where G I ∈ G, the span of the equality constraints of the SDP. Define s : I → {0, 1} as def s(I)  ‰[∃ a degree-2d sos-refutation for S(I)] By assumption, setting, I∼ν [s(I)]  1− 1 . n 8B Define matrix valued functions A, G : I → ’[n] def A(I)  s(I) · AI 20 6 d ×[n] 6 d by def G(I)  s(I) · G I With this notation, we can rewrite the sos-refutation identity as a polynomial identity in X and I, −s(I)  hX 6 d , A(I)i + hX 6 d , G(I)i . Let e∅,∅ denote the [n]6 d × [n]6 d matrix with the entry corresponding to (∅, ∅) equal to 1, while the remaining entries are zero. We can rewrite the above equality as, −hX 6 d , s(I) · e∅,∅ i  hX 6 d , A(I)i + hX 6 d , G(I)i . for all I and formal variables X. Now, let P  S∼Θ PS where each PS is obtained by from the Program 3.1 with ΛS . Substituting 6 d X with P(I) and taking an expectation over I, hP, s(I) · e∅,∅ iν  hP, Aiν + hP, Giν (4.4) > hP, Giν (4.5) where the inequality follows because A, P  0. We will show that the above equation is a contradiction by proving that LHS is less than −0.9, while the right hand side is at least −0.5. First, the right hand side of (4.4) can be bounded by Lemma 4.7 hP, Giν  > I∼ν S∼Θ hPS (IS ), G(I)i I∼ν S∼Θ hPS (IS ), G 6D 1 (I)i − 4B · n 1 2 > − 2B · kGkFr,ν − 4B n n 1 >− 2  2 kPS kFr,ν S  2 kPS kFr,ν S  1/2 kGkFr,ν  1/2 · kGkFr,ν (random restriction Lemma 4.6) (using Lemma 4.7) where the last step used the bounds on kPS kFr,ν from (4.2) and on kGkFr,ν from the n B bound assumed on the SoS proofs in Theorem 2.6. Now the negation of the left hand side of (4.4) is I∼ν hP(I), s(I) · e∅,∅ i > I∼ν [P∅,∅ (I) · 1] − [(s − 1)2 ]1/2 · kP kFr,ν The latter term can be simplified by noticing that the expectation of the square of a 0,1 indicator is equal to the expectation of the indicator, which is in this case n18B by assumption. Also, since 1 is a constant, P∅,∅ and Λ∅,∅ are equivalent:  I∼ν [Λ∅,∅ (I) · 1] − 1 · kP kFr,ν n 4B 1 · kP kFr,ν ( using (4.1)) n 4B 1 (using (4.2))  1 − 3B n  1− We have the desired contradiction in (4.4).  21 4.1 Handling Inequalities Suppose the polynomial system Program 2.2 includes inequalities of the form h(I, x) > 0, then a natural approach would be to introduce a slack variable z and set h(I, x) − z 2  0. Now, we can view the vector (x, z) consisting of the original variables along with the slack variables as the hidden planted solution. The proof of Theorem 2.6 can be carried out as described earlier in this section, with this setup. However, in many cases of interest, the inclusion of slack variables invalidates the robust inference property. This is because, although a feasible solution x can be recovered from a subinstance IS , the value of the corresponding slack variables could potentially depend on IS . For instance, in a random CSP, the value of the objective function on the assignment x generated from IS depends on all the constraints outside of S too. The proof we described is to be modified as follows. • As earlier, construct ΛS using only the robust inference property of original variables x, and the corresponding matrix functions PS . • Convert each inequality of the form h i (I, x) > 0, in to an equality by setting h i (I, x)  z 2i . • Now we define a pseudo-distribution Λ̃S (IS ) over original variables x and slack variables z as follows. It is convenient to describe the pseudo-distribution in terms of the corresponding pseudo-expectation operator. Specifically, if x(IS ) is a feasible solution for Program 2.2 then define ( if σi odd for some i def 0 Ẽ[z σ x α ]  Î σ i /2 · x(I ) otherwise S α i∈σ (h i (I, x(IS ))) Intuitively, the pseudo-distribution picks the sign for each z i uniformly at random, independent of all other variables. Therefore, all moments involving an odd power of z i are zero. On the other hand, the moments of even powers of z i are picked so that the equalities h i (I, x)  z i are satisfied. It is easy to check that Λ̃ is psd matrix valued, satisfies (4.1) and all the equalities. • While ΛS in the original proof was a function of IS , Λ̃S is not. However, the key observation is that, Λ̃S is degree at most k · d in the variables outside of S. Each function h i (I, x(IS )) is degree at most k in IS , and the entries of Λ̃S (IS ) are a product of at most d of these polynomials. • The main ingredient of the proof that is different from the case of equalities is the random restriction lemma which we outline below. The error in the random restriction is multiplied by D dk/2 6 n B/2 ; however this does not substantially change our results, since Theorem 2.6 requires ρ(D, Θ) < n −8B , which leaves us enough slack to absorb this factor (and in every application ρ(D, Θ)  p O(D) for some p < 1 sufficiently small that we meet the requirement that D dk ρ(D − dk, Θ) is monotone non-increasing in D). Lemma 4.9. [Random Restriction for Inequalities] Fix D, ℓ ∈ Ž. Consider a matrix-valued function R : I → ’ℓ×ℓ and a family of functions {PS : I → ’ℓ×ℓ }S⊆[N] such that each PS has degree at most dk in IS . If Θ is a distribution over subsets of [N] with ρ(D, Θ)  max  [α ⊆ S], α,|α| > D S∼Θ 22 and the additional requirement that D dk · ρ(D − dk, Θ) is monotone non-increasing in D, then I∼ν S∼Θ hPS (IS ), R(I)i > S∼Θ I∼ν hPS (IS ), R̃ <D S (IS )i −D 1/2 dk/2 · ρ(D − dk, Θ) ·  2 kPS k2,ν S∼Θ  12 kRkFr,ν Proof. I∼ν S∼Θ hPS (IS ), R(I)i  S∼Θ I∼ν hPS (IS ), R̃ S (I)i where R̃ S (I) is now obtained by averaging out the values for all monomials whose degree in S is > dk. Writing R̃ S  R̃ S6D + R̃ >D and applying a Cauchy-Schwartz inequality we get, S S∼Θ I∼ν hPS (IS ), R̃ S (I)i > S∼Θ I∼ν hPS (IS ), R̃ <D S (I)i −  2 kPS kFr,ν S∼Θ  1/2  · >D S∼Θ k R̃ S kFr,ν  1/2 Over a random choice of S, S∼Θ 2  k R̃ S>D kFr,ν Õ α,|α| > D 2  [|α ∩ S| 6 dk] · R̂ 2α 6 D dk · ρ(D − dk, Θ) · kRkFr , S∼Θ where we have used that D dk ρ(D − dk, Θ) is a monotone non-increasing function of D. Substituting this in the earlier inequality the Lemma follows.  5 Applications to Classical Distinguishing Problems In this section, we verify that the conditions of Theorem 2.6 hold for a variety of canonical distinguishing problems. We’ll rely upon the (simple) proofs in Appendix A, which show that the ideal term of the SoS proof is well-conditioned. Problem 5.1 (Planted clique with clique of size n δ ). Given a graph G  (V, E) on n vertices, determine whether it comes from: • Uniform Distribution: the uniform distribution over graphs on n vertices (G(n, 21 )). • Planted Distribution: the uniform distribution over n-vertex graphs with a clique of size at least n δ The usual polynomial program for planted clique in variables x1 , . . . , x n is: obj 6 Õ xi i x 2i  x i ∀i ∈ [n] x i x j  0 ∀(i, j) ∈ E Lemma 5.2. Theorem 2.6 applies to the above planted clique program, so long as obj 6 n δ−ε for any c·d ε > D−6d for a fixed constant c. Proof. For planted clique, for our notion of “instance degree”, rather than the multiplicity of instance variables, the “degree” of Iα will be the number of distinct vertices incident on the edges 23 in α. The proof of Theorem 2.6 proceeds identically with this notion of degree, but we will be able to achieve better bounds on D relative to d. In this case, the instance degree of the SoS relaxation is k  2. We have from Corollary A.3 that the degree-d SoS refutation is well-conditioned, with numbers bounded by n c1 ·d for some constant c 1 /2. Define B  c 1 d > dk. Our subsampling distribution Θ is the distribution given by including every vertex with probability ρ, producing an induced subgraph of ≈ ρn vertices. For any set of edges α of instance degree at most D − 6d,  [α ⊆ S] 6 ρ D−6d , S∼Θ since the instance degree corresponds to the number of vertices incident on α. This subsampling operation satisfies the subsample inference condition for the clique constraints with probability 1, since a clique in any subgraph of G is also a clique in G. Also, if there is a clique of size n δ in G, then by a Chernoff bound β 2 ρn δ  [∃ clique of size > (1 − β)ρn ∈ S] > 1 − exp(− ). S∼Θ 2 δ q 10B log n , this gives us that Θ gives n −10B -robust inference for the planted clique Choosing β  ρn δ problem, so long as obj 6 ρn/2. Choosing ρ  n −ε for ε so that ρ D−6d 6 n −8B ⇒ ε > c2 d , D − 6d for some constant c 2 , all of the conditions required by Theorem 2.6 now hold.  Problem 5.3 (Random CSP Refutation at clause density α). Given an instance of a Boolean k-CSP with predicate P : {±1} k → {±1} on n variables with clause set C, determine whether it comes from: • Uniform Distribution: m ≈ αn constraints are generated as follows. Each k-tuple of variables S ∈ [n]k is independently with probability p  αn −k+1 given the constraint P(x S ◦ z S )  b S (where ◦ is the entry-wise multiplication operation) for a uniformly random z S ∈ {±1} k and b S ∈ {±1}. • Planted Distribution: a planted solution y ∈ {±1} n is chosen, and then m ≈ αn constraints are generated as follows. Each k-tuple of variables S ∈ [n]k is independently with probability p  αn −k+1 given the constraint P(x S ◦ z S )  b S for a uniformly random z S ∈ {±1} k , but b S  P(yS ◦ z S ) with probability 1 − δ and b S is uniformly random otherwise. The usual polynomial program for random CSP refutation in variables x1 , . . . , x n is: obj 6 Õ S∈[n]  1 + P(x S ◦ z S ) · b S ‰[∃ constraint on S] · 2 k  x 2i  1 ∀ i ∈ [n] Lemma 5.4. If α > 1, then Theorem 2.6 applies to the above random k-CSP refutation problem, so long c·d log n as obj 6 (1 − δ − ε)m for any ε > D−3d , where c is a fixed constant. 24 Proof. In this case, the instance degree of the SoS relaxation k  1. We have from Corollary A.3 that the degree-d SoS refutation is well-conditioned, with numbers bounded by n c1 d for some constant c 1 . Define B  c 1 d. Our subsampling distribution Θ is the distribution given by including each constraint independently with probability ρ, producing an induced CSP instance on n variables with approximately ρm constraints. Since each constraint survives the subsampling with probability ρ, for any C
, α ∈ D−3d  [α ⊆ S] 6 ρ D−3d . S∼Θ The subsample inference property clearly holds for the boolean constraints {x 2i  1}i∈[n] , as a Boolean assignment to the variables is valid regardless of the number of constraints. Before subsampling there are at least (1 − δ)m satisfied constraints, and so letting OS be the number of constraints satisfied in sub-instance S, we have by a Chernoff bound   β 2 ρ(1 − δ)m .  [OS > (1 − β) · ρ(1 − δ)m] > 1 − exp − 2 S∼Θ Choosing β  q 10B log n ρ(1−δ)m  o(1) (with overwhelming probability since we have α > 1 ⇒ n −10B -robust inference [m] > n), we have that Θ gives us for the random CSP refutation problem, so long as obj 6 (1 − o(1))ρ(1 − δ)m. Choosing ρ  (1 − ε) so that ρ D−3d 6 n −8B ⇒ ε > c 2 d log n , D − 3d for some constant c 2 . The conclusion follows (after making appropriate adjustments to the constant).  Problem 5.5 (Community detection with average degree d (stochastic block model)). Given a graph G  (V, E) on n vertices, determine whether it comes from: • Uniform Distribution: G(n, b/n), the distribution over graphs in which each edge is included independently with probability b/n. • Planted Distribution: the stochastic block model—there is a partition of the vertices into two equally-sized sets, Y and Z, and the edge (u, v) is present with probability a/n if u, v ∈ Y or u, v ∈ Z, and with probability (b − a)/n otherwise. Letting x1 , . . . , x n be variables corresponding to the membership of each vertex’s membership, and let A be the adjacency of the graph. The canonical polynomial optimization problem is obj 6 x ⊤ Ax Õ x 2i  1 ∀i ∈ [n] x i  0. i Lemma 5.6. Theorem 2.6 applies to the community detection problem so long as obj 6 (1 − ε) for ε > c·d log n D−3d where c is a fixed constant. 25 (2a−b) 4 n, Proof. The degree of the SoS relaxation in the instance is k  1. Since we have only hypercube and balancedness constraints, we have from Corollary A.3 that the SoS ideal matrix is well-conditioned, with no number in the SoS refutation larger than n c1 d for some constant c 1 . Let B  c 1 d. Consider the solution x which assigns x i  1 to i ∈ Y and x i  −1 to i ∈ Z. Our subsampling operation is to remove every edge independently with probability 1−ρ. The resulting distribution Θ and the corresponding restriction of x clearly satisfies the Booleanity and balancedness constraints with probability 1. Since each edge is included independently with probability ρ, for any α ∈ E
D−3d ,  [α ⊆ S] 6 ρ D−3d . S∼Θ In the sub-instance, the expected value (over the choice of planted instance and over the choice of sub-instance) of the restricted solution x is ρa · n    |Z| |Y| + 2 2  −ρ (2a − b)ρn b−a · |Y| · |Z|  − ρa, n 4 and by a Chernoffqbound, the value in the sub instance is within a (1 − β)-factor with probability 10B log n 1 − n −10B for β  . On resampling the edges outside the sub-instance from the uniform n distribution, this value can only decrease by at most (1 − ρ)(1 + β)nb/2 w.h.p over the choice of the outside edges. c (2a−b) log n If we set ρ  (1 − ε(2a − b)/10b), then ρ D−3d 6 n −8B for ε > 2 D−3d . for some constant c 2 , while the objective value is at least (1 − ε) adjustments to the constant). (2a−b)n . 4 The conclusion follows (after making appropriate  Problem 5.7 (Densest-k-subgraph). Given a graph G  (V, E) on n vertices, determine whether it comes from: • Uniform Distribution: G(n, p). • Planted Distribution: A graph from G(n, p) with an instance of G(k, q) planted on a random subset of k vertices, p < q. Letting A be the adjacency matrix, the usual polynomial program for densest-k-subgraph in variables x1 , . . . , x n is: obj 6 x ⊤ Ax Õ x 2i  x i ∀i ∈ [n] xi  k i Lemma 5.8. When k 2 (p + q) ≫ d log n, Theorem 2.6 applies to the densest-k-subgraph problem with
c·d log n obj 6 (1 − ε)(p + q) 2k for any ε > D−3d for a fixed constant c. Proof. The degree of the SoS relaxation in the instance is k  1. We have from Corollary A.3 that the SoS proof has no values larger than n c1 d for a constant c 1 ; fix B  c 1 d. Our subsampling operation is to include each edge independently with probability ρ, and take the subgraph induced by the included edges. Clearly, the Booleanity and sparsity constraints are 26 preserved by this subsampling distribution Θ. Since each edge is included independently with E
, probability ρ, for any α ∈ D−3d  [α ⊆ S] 6 ρ D−3d . S∼Θ
Now, the expected objective value (over the instance and the sub-sampling) is at least ρ(p + q) 2k , and applying a Chernoff bound, we hace that the probability the r sub-sampled instance has value less than (1 − β)ρ(p + q) k
2 is at most n −10B if we choose β  10B log n ρ(p+q)(2k) (which is valid since we assumed that d log n ≪ (p + q)k 2 ). Further, a dense subgraph on a subset of the edges is still dense when more edges are added back, so we have the n −10B -robust inference property. Thus, choosing ρ  (1 − ε) and setting ρ D−3d 6 n −8B ⇒ ε > c 2 d log n , D − 3d for some constant c 2 , which concludes the proof (after making appropriate adjustments to the constant).  Problem 5.9 (Tensor PCA). Given an order-k tensor in (’n )⊗k , determine whether it comes from: • Uniform Distribution: each entry of the tensor sampled independently from N(0, 1). • Planted Distribution: a spiked tensor, T  λ · v ⊗k + G where v is sampled uniformly from {± √1n } n , and where G is a random tensor with each entry sampled independently from N(0, 1). Given the tensor T, the canonical program for the tensor PCA problem in variables x1 , . . . , x n is: obj 6 hx ⊗k , Ti kx k22  1 Lemma 5.10. For λn −ε ≫ log n, Theorem 2.6 applies to the tensor PCA problem with obj 6 λn −ε for c·d for a fixed constant c. any ε > D−3d Proof. The degree of the SoS relaxation in the instance is k  1. Since the entries of the noise component of the tensor are standard normal variables, with exponentially good probability over the input tensor T we will have no entry of magnitude greater than n d . This, together with Corollary A.3, gives us that except with exponentially small probability the SoS proof will have no values exceeding n c1 d for a fixed constant c 1 . Our subsampling operation is to set to zero every entry of T independently with probability 1 − ρ, obtaining a sub-instance T′ on the nonzero entries. Also, for any α ∈  [α ∈ S] 6 ρ D−3d . [n]k
, D−3d S∼Θ This subsampling operation clearly preserves the planted solution unit sphere constraint. Additionally, let R be the operator that restricts a tensor to the nonzero entries. We have that hR(λ · v ⊗k ), v ⊗k i has expectation λ · ρ, since every entry of v ⊗k has magnitude n −k/2 . Applying a Chernoff bound, q we have that this quantity will be at least (1 − β)λρ with probability at least n −10B if we choose β  10B log n . λρ 27 It remains to address the noise introduced by GT′ and resampling all the entries outside of the subinstance T′. Each of these entries is a standard normal entry. The quantity h(Id −R)(N), v ⊗k i is a sum over at most n k i.i.d. Gaussian entries each with standard deviation n −k/2 (since that is the magnitude of (v ⊗k )α . The entire quantity is thus a Gaussian random variable with mean 0 and p −10B variance 1, and therefore with probability at least n this quantity will not exceed 10B log n. p So long as 10B log n ≪ λρ, the signal term will dominate, and the solution will have value at least λρ/2. Now, we set ρ  n −ε so that ρ D−3d 6 n −8B ⇒ ε > 2c 1 d , D − 3d which concludes the proof (after making appropriate adjustments to the constant c 1 ).  Problem 5.11 (Sparse PCA). Given an n × m matrix M in ’n , determine whether it comes from: • Uniform Distribution: each entry of the matrix sampled independently from √ N(0, 1). • Planted Distribution: a random vector with k non-zero entries v ∈ {0, ±1/ k} n is chosen, and then the ith column of the matrix is sampled independently by taking s i v + γi for a uniformly random sign s i ∈ {±1} and a standard gaussian vector γi ∼ N(0, Id). The canonical program for the sparse PCA problem in variables x1 , . . . , x n is: obj 6 kM ⊤ x k22 x 2i  x i kx k22 k ∀i ∈ [n] Lemma 5.12. For kn −ε/2 ≫ log n, Theorem 2.6 applies to the sparse PCA problem with obj 6 k 2−ε m for c·d any ε > D−6d for a fixed constant c. Proof. The degree of the SoS relaxation in the instance is 2. Since the entries of the noise are standard normal variables, with exponentially good probability over the input matrix M we will have no entry of magnitude greater than n d . This, together with Corollary A.3, gives us that except with exponentially small probability the SoS proof will have no values exceeding n c1 d for a fixed constant c 1 . Our subsampling operation is to set to zero every entry of M independently with probability M
, 1 − ρ, obtaining a sub-instance M on the nonzero entries. Also, for any α ∈ D−6d  [α ∈ S] 6 ρ D−6d . S∼Θ This subsampling operation clearly preserves the constraints on the solution variables. √ We take our subinstance solution y  kv, which is feasible. Let R be the subsampling operator that zeros out a set of columns. On subsampling, and then resampling the zeroed out columns from the uniform distribution, we can write the resulting M̃ as M̃ ⊤  R(sv T ) + G ⊤ T where √ G is a random Gaussian matrix. Therefore, the objective value obtained by the solution y  kv is 28 M̃ ⊤ y  √ √ k · R(sv ⊤ )v + k · G ⊤ v The first term is a vector u si1nal with m entries, each of which is a sum of k Bernoulli random variables, all of the same sign, with probability ρ of being nonzero. The second term is a vector u noise with m entries, each of them an independent Gaussian variable with variance bounded by k. We have that [ku si1nal k22 ]  (ρk)2 m, Θ and by Chernoff bounds qwe have that this concentrates within a (1 − β) factor with probability 10B log n . 1 − n −10B if we take β  (ρk)2 m The expectation of hu si1nal , u noise i is zero, and applying similar concentration arguments we have that with probability 1 − n 10B , |hu si1nal , u noise i| 6 (1 + β)ρk. Taking the union bound over these events and applying Cauchy-Schwarz, we have that kR(M)y k22 > (ρk)2 m − 2km  ρ 2 k 2 m − 2km. so long as ρk ≫ 1, the first term dominates. Now, we set ρ  n −ε for ε < 1 so that ρ D−6d 6 n −8B ⇒ ε > c2 d , D − 6d for some constant c 2 , which concludes the proof.  Remark 5.13. For tensor PCA and sparse PCA, the underlying distributions were Gaussian. Applying Theorem 2.6 in these contexts yields the existence of distinguishers that are low-degree in a non-standard sense. Specifically, the degree of a monomial will be the number of distinct variables in it, irrespective of the powers to which they are raised. 6 Exponential lower bounds for PCA problems In this section we give an overview of the proofs of our SoS lower bounds for the tensor and sparse PCA problems. We begin by showing how Conjecture 1.2 predicts such a lower bound in the tensor PCA setting. Following this we state the key lemmas to prove the exponential lower bounds; since these lemmas can be proved largely by techniques present in the work of Barak et al. on planted clique [BHK+ 16], we leave the details to a forthcoming full version of the present paper. 6.1 Predicting sos lower bounds from low-degree distinguishers for Tensor PCA In this section we demonstrate how to predict using Conjecture 1.2 that when λ ≪ n 3/4−ε for ε > 0, SoS algorithms cannot solve Tensor PCA. This prediction is borne out in Theorem 1.4. Theorem 6.1. Let µ be the distribution on ’n⊗n⊗n which places a standard Gaussian in each entry. Let ν be the density of the Tensor PCA planted distribution with respect to µ, where we take the planted vector v 29 to have each entry uniformly chosen from {± √1n }.7 If λ 6 n 3/4−ε , there is no degree n o(1) polynomial p with p(A)  0, µ p(A) > n planted Ω(1)  · – p(A) µ  1/2 . We sketch the proof of this theorem. The theorem follows from two claims. Claim 6.2. ν max deg p 6 n o(1) , µ p(T)0 µ p(T)  ( (ν 6 d (T) − 1)2 )1/2
1/2 µ 2 (6.1) p(T) where ν 6 d is the orthogonal projection (with respect to µ) of the density ν to the degree-d polynomials. Note that the last quantity is just the 2 norm, or the variance, of the truncation to low-degree polynomials of the density ν of the planted distribution. Claim 6.3. ( µ (v 6 d (T) − 1)2 )1/2 ≪ 1 when λ 6 n 3/4−ε for ε > Ω(1) and d  n o(1) . The theorem follows immediately. We sketch proofs of the claims in order. Sketch of proof for Claim 6.2. By definition of ν, the maximization is equivalent to maximizing o(1) and with 2 µ ν(T) · p(T) among all p of degree d  n µ p(T)  1 and µ p(T)  0. Standard Fourier analysis shows that this maximum is achieved by the orthogonal projection of ν − 1 into the span of degree d polynomials. To make this more precise, recall that the Hermite polynomials provide an orthonormal basis for real-valued polynomials under the multivariate Gaussian distribution. (For an introduction to Hermite polynomials, see the book [O’D14].) The tensor T ∼ µ is an n 3 -dimensional multivariate Gaussian. For a (multi)-set W ⊆ [n]3 , let HW be the W-th Hermite polynomial, so that µ HW (T)HW ′ (T)  ‰WW ′ . Then the best p (ignoring normalization momentarily) will be the function p(A)  ν 6 d (A) − 1  Õ 16 |W | 6 d ( T∼µ ν(T)HW (T)) · HW (A) Here µ ν(T)HW (T)  b ν (W) is the W-th Fourier coefficient of ν. What value for (6.1) is achieved by this p? Again by standard Fourier analysis, in the numerator we have, ν p(T)  ν (ν 6D (T) − 1)  µ ν(T) · (ν 6D (T) − 1)  µ (ν 6 d (T) − 1)2 Comparing this to the denominator, the maximum value of (6.1) is ( µ (v 6 d (T) − 1)2 )1/2 . This is nothing more than the 2-norm of the projection of ν − 1 to degree-d polynomials!  The following fact, used to prove Claim 6.3, is an elementary computation with Hermite polynomials. Fact 6.4. Let W ⊆ [n]3 . Then b ν (W)  λ |W | n −3|W |/2 if W, thought of as a 3-uniform hypergraph, has all even degrees, and is 0 otherwise. 7This does not substantially modify the problem but it will make calculations in this proof sketch more convenient. 30 š  µ ν(T)HW (T)  ν HW (T), To see that this calculation is straightforward, note that ν(W) so it is enough to understand the expectations of the Hermite polynomials under the planted distribution. Sketch of proof for Claim 6.3. Working in the Hermite basis (as described above), we get µ (v 6 d (T) − Í 1)2  16 |W | 6 d b ν (W)2 . For the sake of exposition, we will restrict attention in the sum to W in which no element appears with multiplicity larger than 1 (other terms can be treated similarly). Í What is the contribution to 16 |W | 6 d b ν (W)2 of terms W with |W |  t? By the fact above, to contribute a nonzero term to the sum, W,considered as a 3-uniform hypergraph must have even degrees. So, if it has t hyperedges, it contains at most 3t/2 nodes. There are n 3t/2 choices for these nodes, and having chosen them, at most t O(t) 3-uniform hypergraphs on those nodes. Hence, Õ 16 |W | 6 d 2 b ν (W) 6 d Õ n 3t/2 t O(t) λ 2t n −3t . t1 So long as λ 2 6 n 3/2−ε for some ε  Ω(1) and t 6 d 6 n O(ε) , this is o(1).  6.2 Main theorem and proof overview for Tensor PCA In this section we give an overview of the proof of Theorem 1.4. The techniques involved in proving the main lemmas are technical refinements of techniques used in the work of Barak et al. on SoS lower bounds for planted clique [BHK+ 16]; we therefore leave full proofs to a forthcoming full version of this paper. To state and prove our main theorem on tensor PCA it is useful to define a Boolean version of the problem. For technical convenience we actually prove an SoS lower bound for this problem; then standard techniques (see Section C) allow us to prove the main theorem for Gaussian tensors. Problem 6.5 (k-Tensor PCA, signal-strength λ, boolean version). Distinguish the following two n def distributions on Ωk  {±1}( k ) . • the uniform distribution: A ∼ Ω chosen uniformly at random. • the planted distribution: Choose v ∼ {±1} n and let B  v ⊗k . Sample A by rerandomizing every coordinate of B with probability 1 − λn −k/2 . We show that the natural SoS relaxation of this problem suffers from a large integrality gap, when λ is slightly less than n k/4 , even when the degree of the SoS relaxation is n Ω(1) . (When O(ε) are known for k  O(1) [RM14, HSS15, HSSS16, λ ≫ n k/4−ε , algorithms with running time 2n BGL16, RRS16].) Theorem 6.6. Let k  O(1). For A ∈ Ωk , let def SoS d (A)  max ˜ hx ⊗k , Ai s.t. ˜ is a degree-d pseudoexpectation satisfying {kx k 2  1} . ˜ There is a constant c so that for every small enough ε > 0, if d 6 n c·ε , then for large enough n,  {SoS d (A) > n k/4−ε } > 1 − o(1) A∼Ω 31 and A∼Ω SoS d (A) > n k/4−ε . Moreover, the latter also holds for A with iid entries from N(0, 1).8 To prove the theorem we will exhibit for a typical sample A from the uniform distribution a degree n Ω(ε) pseudodistribution ˜ which satisfies {kx k 2  1} and has ˜ hx ⊗k , Ai > n k/4−ε . The following lemma ensures that the pseudo-distribution we exhibit will be PSD. Í Lemma 6.7. Let d ∈ Ž and let Nd  s 6 d n(n − 1) · · · (n − (s − 1)) be the number of 6 d-tuples with unique entries from [n]. There is a constant ε ∗ independent of n such that for any ε < ε ∗ also independent of n, the following is true. Let λ  n k/4−ε . Let µ(A) be the density of the following distribution (with respect n to the uniform distribution on Ω  {±1}( k ) ). The Planted Distribution: Choose v ∼ {±1} n uniformly. Let B  v ⊗k . Sample A by • replacing every coordinate of B with a random draw from {±1} independently with probability 1 − λn −k/2 , • then choosing a subset S ⊆ [n] by including every coordinate with probability n −ε , • then replacing every entry of B with some index outside S independently with a uniform draw from {±1}. Let Λ : Ω → ’Nd ×Nd be the following function Λ(A)  µ(A) · v|A v ⊗ 62d Here we abuse notation and denote by x 6 ⊗2d the matrix indexed by tuples of length 6 d with unique entries from [n]. For D ∈ Ž, let Λ6D be the projection of Λ into the degree-D real-valued polynomials on n {±1}( k ) . There is a universal constant C so that if Cd/ε < D < n ε/C , then for large enough n  {Λ6D (A)  0} > 1 − o(1) . A∼Ω For a tensor A, the moment matrix of the pseudodistribution we exhibit will be Λ6D (A). We will need it to satisfy the constraint {kx k 2  1}. This follows from the following general lemma. (The lemma is much more general than what we state here, and uses only the vector space structures of space of real matrices and matrix-valued functions.) n Lemma 6.8. Let k ∈ Ž. Let V be a linear subspace of ’N×M . Let Ω  {±1}( k ) . Let Λ : Ω → V. Let Λ6D be the entrywise orthogonal projection of Λ to polynomials of degree at most D. Then for every A ∈ Ω, the matrix Λ6D (A) ∈ V. Proof. The function Λ is an element of the vector space ’N×M ⊗ ’Ω . The projection ΠV : ’N×M → V and the projection Π6D from ’Ω to the degree-D polynomials commute as projections on ’N×M ⊗ ’Ω , since they act on separate tensor coordinates. It follows that Λ6D ∈ V ⊗ (’Ω )6D takes values in V.  8For technical reasons we do not prove a tail bound type statement for Gaussian A, but we conjecture that this is also true. 32 Last, we will require a couple of scalar functions of Λ6D to be well concentrated. Lemma 6.9. Let Λ, d, ε, D be as in Lemma 6.7. The function Λ6D satisfies 6D • A∼Ω {Λ∅,∅ (A)  1 ± o(1)} > 1 − o(1) (Here Λ∅,∅  1 is the upper-left-most entry of Λ.) • A∼Ω {hΛ6D (A), Ai  (1 ± o(1)) · n 3k/4−ε } > 1 − o(1) (Here we are abusing notation to write hΛ6D (A), Ai for the inner product of the part of Λ6D indexed by monomials of degree k and A.) The Boolean case of Theorem 6.6 follows from combining the lemmas. The Gaussian case can be proved in a black-box fashion from the Boolean case following the argument in Section C. The proofs of all the lemmas in this section follow analogous lemmas in the work of Barak et al. on planted clique [BHK+ 16]; we defer them to the full version of the present work. 6.3 Main theorem and proof overview for sparse PCA In this section we prove the following main theorem. Formally, the theorem shows that with high probability for a random n × n matrix A, even high-degree SoS relaxations are unable to certify that no sparse vector v has large quadratic form hv, Avi. Theorem 6.10 (Restatement of Theorem 1.6). If A ∈ ’n×n , let  SoS d,k (A)  max ˜ hx, Axi s.t. ˜ is degree d and satisfies x 3i  x i , kx k 2  k . ˜ There are absolute constants c, ε ∗ > 0 so that for every ρ ∈ (0, 1) and ε ∈ (0, ε ∗ ), if k  n ρ , then for d 6 n c·ε ,  n {SoS d,k (A) > min(n 1/2−ε k, n ρ−ε k)} > 1 − o(1) A∼{±1} ( 2 ) and n A∼{±1} ( 2 ) SoS d,k (A) > min(n 1/2−ε k, n ρ−ε k) . Furthermore, the latter is true also if A is symmetric with iid entries from N(0, 1).9 We turn to some discussion of the theorem statement. First of all, though it is technically convenient for A in the theorem statement above to be a ±1 matrix, the entries may be replaced by standard Gaussians (see Section C). Remark 6.11 (Relation to the spiked-Wigner model of sparse principal component analysis). To get some intuition for the theorem statement, it is useful to return to a familiar planted problem: the spiked-Wigner model of sparse principal component analysis. Let W be a symmetric matrix with √ iid entries from N(0, 1), and let v be a random k-sparse unit vector with entries {±1/ k, 0}. Let B  W + λvv⊺ . The problem is to distinguish between a single sample from B and a sample from W. There are two main algorithms for this problem, both captured by the SoS hierarchy. The √ first, applicable when λ ≫ n, is vanilla PCA: the top eigenvalue of B will be larger than the top eigenvalue of W. The second, applicable when λ ≫ k, is diagonal thresholding: the diagonal 9For technical reasons we do not prove a tail bound type statement for Gaussian A, but we conjecture that this is also true. 33 entries of B which corresponds to nonzero coordinates will be noticeably large. The theorem statement above (transferred to the Gaussian setting, though this has little effect) shows that once λ is well outside these parameter regimes, i.e. when λ < n 1/2−ε , k 1−ε for arbitrarily small ε > 0, even degree n Ω(ε) SoS programs do not distinguish between B and W. Remark 6.12 (Interpretation as an integrality gap). A second interpretation of the theorem statement, independent of any planted problem, is as a strong integrality gap for random instances for the problem of maximizing a quadratic form over k-sparse vectors. Consider the actual maximum of hx, Axi for random ({±1} or Gaussian) A over k-sparse unit vectors x. There are roughly 2k log n points in a 21 -net for such vectors, meaning that by standard arguments, max kxk1,x is k-sparse √ hx, Axi 6 O( k log n) . With the parameters of the theorem, this means that the integrality gap of the degree n Ω(ε) SoS relaxation is at least min(n ρ/2−ε , n 1/2−ρ/2−ε ) when k  n ρ . Remark 6.13 (Relation to spiked-Wishart model). Theorem 1.6 most closely concerns the spikedWigner model of sparse PCA; this refers to independence of the entries of the matrix A. Often, sparse PCA is instead studied in the (perhaps more realistic) spiked-Wishart model, where the input is m samples x1 , . . . , x m from an n-dimensional Gaussian vector N(0, Id +λ · vv ⊤ ), where v is a unit-norm k-sparse vector. Here the question is: as a function of the sparsity k, the ambient dimension n, and the signal strength λ, how many samples m are needed to recover the vector v? The natural approach to recovering v in this setting is to solve a convex relaxation of the problem Í of maximizing he quadratic form of the empirical covariance M  i 6 m x i x i⊺ over k-sparse unit vectors (the maximization problem itself is NP-hard even to approximate [CPR16]). Theoretically, one may apply our proof technique for Theorem 1.6 directly to the spiked-Wishart model, but this carries the expense of substantial technical complication. We may however make intelligent guesses about the behavior of SoS relaxations for the spiked-Wishart model on the basis of Theorem 1.6 alone. As in the spiked Wigner model, there are essentially two known algorithms to recover a planted sparse vector v in the spiked Wishart model: vanilla PCA and diagonal thresholding [DM14b]. We conjecture that, as in the spiked Wigner model, the SoS hierarchy requires n Ω(1) degree to improve the number of samples required by these algorithms by any polynomial factor. Concretely, considering the case λ  1 for simplicity, we conjecture that there are constants c, ε ∗ such that for every ε ∈ (0, ε ∗ ) if m 6 min(k 2−ε , n 1−ε ) and x1 , . . . , x m ∼ N(0, Id) are iid, then with high probability for every ρ ∈ (0, 1) if k  n ρ , SoS d,k Õ i6m ! x i x i⊺ > min(n 1−ε k, k 2−ε ) for all d 6 n c·ε . Lemmas for Theorem 1.6. Our proof of Theorem 1.6 is very similar to the analogous proof for Tensor PCA, Theorem 6.6. We state the analogues of Lemma 6.7 and Lemma 6.9. Lemma 6.8 can be used unchanged in the sparse PCA setting. The main lemma, analogous to Lemma 6.7 is as follows. 34 Í Lemma 6.14. Let d ∈ Ž and let Nd  s 6 d n(n − 1) · · · (n − (s − 1)) be the number of 6 d-tuples with unique entries from [n]. Let µ(A) be the density of the following distribution on n × n matrices A with n respect to the uniform distribution on {±1}( 2) . Planted distribution: Let k  k(n) ∈ Ž and λ  λ(n) ∈ ’, and γ > 0, and assume λ 6 k. Sample a uniformly random k-sparse vector v ∈ ’n with entries ±1, 0. Form the matrix B  vv ⊤ . For each nonzero entry of B independently, replace it with a uniform draw from {±1} with probability 1 − λ/k (maintaining the symmetry B  B ⊤ ). For each zero entry of B, replace it with a uniform draw from {±1} (maintaining the same symmetry). Finally, choose every i ∈ [n] with probability n −γ independently; for those indices that were not chosen, replace every entry in the corresponding row and column of B with random ±1 entries.10 Output the resulting matrix A. (We remark that this matrix is a Boolean version of the more standard spiked-Wigner model B + λvv√⊤ where B has iid standard normal entries and v is a random k-sparse unit vector with entries from {±1/ k, 0}.) n Let Λ : {±1}( 2) → ’Nd ×Nd be the following function Λ(A)  µ(A) · v|A v ⊗ 62d where the expectation is with respect to the planted distribution above. For D  D(n) ∈ Ž, let Λ6D be the entrywise projection of Λ into the Boolean functions of degree at most D. There are constants C, ε ∗ > 0 such that for every γ > 0 and ρ ∈ (0, 1) and every ε ∈ (0, ε ∗ ) (all independent of n), if k  n ρ and λ 6 min{n ρ−ε , n 1/2−ε }, and if Cd/ε < D < n ε/C , then for large enough n  n {Λ6D (A)  0} > 1 − o(1) . A∼{±1} ( 2 ) Remark 6.15. We make a few remarks about the necessity of some of the assumptions above. A useful intuition is that the function Λ6D (A) is (with high probability) positive-valued when the parameters ρ, ε, γ of the planted distribution are such that there is no degree-D polynomial n f : {±1}( 2) → ’ whose values distinguish a typical sample from the planted distribution from a null model: a random symmetric matrix with iid entries. At this point it is useful to consider a more familiar planted model, which the lemma above n mimics. Let W be a n × n symmetric √ matrix with iid entries from N(0, 1). Let v ∈ ’ be a k-sparse ⊺ unit vector, with entries in {±1/ k, 0}. Let A  W + λvv . Notice that if λ ≫ k, then diagonal thresholding on the matrix W identifies the nonzero coordinates of v. (This is the analogue of the covariance-thresholding algorithm in the spiked-Wishart version of sparse PCA.) On the other √ √ hand, if λ ≫ n then (since typically kW k ≈ n), ordinary PCA identifies v. The lemma captures computational hardness for the problem of distinguishing a single sample from A from a sample from the null model W both diagonal thresholding and ordinary PCA fail. Next we state the analogue of Lemma 6.9. Lemma 6.16. Let Λ, d, k, λ, γ, D be as in Lemma 6.14. The function Λ6D satisfies 6D • A∼{±1}(nk) {Λ∅,∅ (A)  1 ± o(1)} > 1 − o(1). • A∼{±1}(nk) {hΛ6D (A), Ai  (1 ± o(1)) · λn Θ(−γ) } > 1 − o(1). 10This additional n −γ noising step is a technical convenience which has the effect of somewhat decreasing the number of nonzero entries of v and decreasing the signal-strength λ. 35 References [AK97] Noga Alon and Nabil Kahale, A spectral technique for coloring random 3-colorable graphs, SIAM J. Comput. 26 (1997), no. 6, 1733–1748. 1 [AKS98] Noga Alon, Michael Krivelevich, and Benny Sudakov, Finding a large hidden clique in a random graph, Random Struct. Algorithms 13 (1998), no. 3-4, 457–466. 1, 4, 8 [AOW15a] Sarah R. Allen, Ryan O’Donnell, and David Witmer, How to refute a random CSP, 2015 IEEE 56th Annual Symposium on Foundations of Computer Science—FOCS 2015, IEEE Computer Soc., Los Alamitos, CA, 2015, pp. 689–708. MR 3473335 1 [AOW15b] Sarah R. Allen, Ryan O’Donnell, and David Witmer, How to refute a random CSP, FOCS, IEEE Computer Society, 2015, pp. 689–708. 8 [BBH+ 12] Boaz Barak, Fernando G. S. L. Brandão, Aram Wettroth Harrow, Jonathan A. Kelner, David Steurer, and Yuan Zhou, Hypercontractivity, sum-of-squares proofs, and their applications, STOC, ACM, 2012, pp. 307–326. 1, 6, 8 [BCC+ 10] Aditya Bhaskara, Moses Charikar, Eden Chlamtac, Uriel Feige, and Aravindan Vijayaraghavan, Detecting high log-densities—an O(n 1/4 ) approximation for densest k-subgraph, STOC’10—Proceedings of the 2010 ACM International Symposium on Theory of Computing, ACM, New York, 2010, pp. 201–210. MR 2743268 8 [BCK15] Boaz Barak, Siu On Chan, and Pravesh K. Kothari, Sum of squares lower bounds from pairwise independence [extended abstract], STOC’15—Proceedings of the 2015 ACM Symposium on Theory of Computing, ACM, New York, 2015, pp. 97–106. MR 3388187 8 [BGG+ 16] Vijay V. S. P. Bhattiprolu, Mrinal Kanti Ghosh, Venkatesan Guruswami, Euiwoong Lee, and Madhur Tulsiani, Multiplicative approximations for polynomial optimization over the unit sphere, Electronic Colloquium on Computational Complexity (ECCC) 23 (2016), 185. 1, 6, 8 [BGL16] Vijay V. S. P. Bhattiprolu, Venkatesan Guruswami, and Euiwoong Lee, Certifying random polynomials over the unit sphere via sum of squares hierarchy, CoRR abs/1605.00903 (2016). 1, 2, 9, 31 [BHK+ 16] Boaz Barak, Samuel B. Hopkins, Jonathan A. Kelner, Pravesh Kothari, Ankur Moitra, and Aaron Potechin, A nearly tight sum-of-squares lower bound for the planted clique problem, FOCS, IEEE Computer Society, 2016, pp. 428–437. 1, 4, 5, 8, 9, 29, 31, 33 [BKS14] Boaz Barak, Jonathan A. Kelner, and David Steurer, Rounding sum-of-squares relaxations, STOC, ACM, 2014, pp. 31–40. 1 [BKS15] , Dictionary learning and tensor decomposition via the sum-of-squares method, STOC, ACM, 2015, pp. 143–151. 1 36 [BKS17] Boaz Barak, Pravesh Kothari, and David Steurer, Quantum entanglement, sum of squares, and the log rank conjecture, CoRR abs/1701.06321 (2017). 1 [BM16] Boaz Barak and Ankur Moitra, Noisy tensor completion via the sum-of-squares hierarchy, COLT, JMLR Workshop and Conference Proceedings, vol. 49, JMLR.org, 2016, pp. 417– 445. 1, 8 [BMVX16] Jess Banks, Cristopher Moore, Roman Vershynin, and Jiaming Xu, Information-theoretic bounds and phase transitions in clustering, sparse pca, and submatrix localization, CoRR abs/1607.05222 (2016). 6 [BR13a] Quentin Berthet and Philippe Rigollet, Complexity theoretic lower bounds for sparse principal component detection, COLT, JMLR Workshop and Conference Proceedings, vol. 30, JMLR.org, 2013, pp. 1046–1066. 7 [BR13b] Quentin Berthet and Philippe Rigollet, Computational lower bounds for sparse pca, COLT (2013). 2 [BS14] Boaz Barak and David Steurer, Sum-of-squares proofs and the quest toward optimal algorithms, CoRR abs/1404.5236 (2014). 6 [CC09] Eric Carlen and ERIC CARLEN, Trace inequalities and quantum entropy: An introductory course, 2009. 18 [CPR16] Siu On Chan, Dimitris Papailliopoulos, and Aviad Rubinstein, On the approximability of sparse PCA, COLT, JMLR Workshop and Conference Proceedings, vol. 49, JMLR.org, 2016, pp. 623–646. 7, 34 [DM14a] Yash Deshpande and Andrea Montanari, Information-theoretically optimal sparse PCA, ISIT, IEEE, 2014, pp. 2197–2201. 7 [DM14b] , Sparse PCA via covariance thresholding, NIPS, 2014, pp. 334–342. 2, 34 [DM15] , Improved sum-of-squares lower bounds for hidden clique and hidden submatrix problems, COLT, JMLR Workshop and Conference Proceedings, vol. 40, JMLR.org, 2015, pp. 523–562. 9 [DX13] Feng Dai and Yuan Xi, Spherical harmonics, arXiv preprint arXiv:1304.2585 (2013). 44 [Fil16] Yuval Filmus, An orthogonal basis for functions over a slice of the boolean hypercube, Electr. J. Comb. 23 (2016), no. 1, P1.23. 44 [FM16] Zhou Fan and Andrea Montanari, How well do local algorithms solve semidefinite programs?, CoRR abs/1610.05350 (2016). 8 [GM15] Rong Ge and Tengyu Ma, Decomposing overcomplete 3rd order tensors using sum-of-squares algorithms, APPROX-RANDOM, LIPIcs, vol. 40, Schloss Dagstuhl - Leibniz-Zentrum fuer Informatik, 2015, pp. 829–849. 1 37 [Gri01a] Dima Grigoriev, Complexity of positivstellensatz proofs for the knapsack, Computational Complexity 10 (2001), no. 2, 139–154. 8 [Gri01b] , Linear lower bound on degrees of positivstellensatz calculus proofs for the parity, Theor. Comput. Sci. 259 (2001), no. 1-2, 613–622. 8 [GW94] Michel X. Goemans and David P. Williamson, .879-approximation algorithms for MAX CUT and MAX 2sat, STOC, ACM, 1994, pp. 422–431. 1 [Har70] Richard A Harshman, Foundations of the parafac procedure: Models and conditions for an" explanatory" multi-modal factor analysis. 1 [HKP15] Samuel B. Hopkins, Pravesh K. Kothari, and Aaron Potechin, Sos and planted clique: Tight analysis of MPW moments at all degrees and an optimal lower bound at degree four, CoRR abs/1507.05230 (2015). 9 [HL09] Christopher J. Hillar and Lek-Heng Lim, Most tensor problems are NP hard, CoRR abs/0911.1393 (2009). 6 [HSS15] Samuel B. Hopkins, Jonathan Shi, and David Steurer, Tensor principal component analysis via sum-of-square proofs, COLT, JMLR Workshop and Conference Proceedings, vol. 40, JMLR.org, 2015, pp. 956–1006. 1, 2, 6, 8, 9, 31 [HSSS16] Samuel B. Hopkins, Tselil Schramm, Jonathan Shi, and David Steurer, Fast spectral algorithms from sum-of-squares proofs: tensor decomposition and planted sparse vectors, STOC, ACM, 2016, pp. 178–191. 1, 2, 8, 9, 31 [KMOW17] Pravesh K. Kothari, Ryuhei Mori, Ryan O’Donnell, and David Witmer, Sum of squares lower bounds for refuting any CSP, CoRR abs/1701.04521 (2017). 8 [KNV+ 15] Robert Krauthgamer, Boaz Nadler, Dan Vilenchik, et al., Do semidefinite relaxations solve sparse pca up to the information limit?, The Annals of Statistics 43 (2015), no. 3, 1300–1322. 2 [KT17] Ken-ichi Kawarabayashi and Mikkel Thorup, Coloring 3-colorable graphs with less than n1/5 colors, J. ACM 64 (2017), no. 1, 4:1–4:23. 1 [LRS15] James R. Lee, Prasad Raghavendra, and David Steurer, Lower bounds on the size of semidefinite programming relaxations, STOC, ACM, 2015, pp. 567–576. 1 [MPW15] Raghu Meka, Aaron Potechin, and Avi Wigderson, Sum-of-squares lower bounds for planted clique [extended abstract], STOC’15—Proceedings of the 2015 ACM Symposium on Theory of Computing, ACM, New York, 2015, pp. 87–96. MR 3388186 9 [MS16a] Andrea Montanari and Subhabrata Sen, Semidefinite programs on sparse random graphs and their application to community detection, STOC, ACM, 2016, pp. 814–827. 8 [MS16b] Andrea Montanari and Nike Sun, Spectral algorithms for tensor completion, CoRR abs/1612.07866 (2016). 8 38 [MSS16a] Tengyu Ma, Jonathan Shi, and David Steurer, Polynomial-time tensor decompositions with sum-of-squares, CoRR abs/1610.01980 (2016). 1 [MSS16b] , Polynomial-time tensor decompositions with sum-of-squares, FOCS, IEEE Computer Society, 2016, pp. 438–446. 1 [MW15a] Tengyu Ma and Avi Wigderson, Sum-of-squares lower bounds for sparse PCA, NIPS, 2015, pp. 1612–1620. 2 [MW15b] , Sum-of-squares lower bounds for sparse PCA, CoRR abs/1507.06370 (2015). 9 [O’D14] Ryan O’Donnell, Analysis of boolean functions, Cambridge University Press, 2014. 30 [Pea01] Karl Pearson, On lines and planes of closes fit to systems of points in space, Philosophical Magazine 2 (1901), 559–572. 1 [PS17] Aaron Potechin and David Steurer, Exact tensor completion with sum-of-squares, CoRR abs/1702.06237 (2017). 1 [PWB16] Amelia Perry, Alexander S. Wein, and Afonso S. Bandeira, Statistical limits of spiked tensor models, CoRR abs/1612.07728 (2016). 6 [RM14] Emile Richard and Andrea Montanari, A statistical model for tensor PCA, NIPS, 2014, pp. 2897–2905. 2, 5, 9, 31 [RRS16] Prasad Raghavendra, Satish Rao, and Tselil Schramm, Strongly refuting random csps below the spectral threshold, CoRR abs/1605.00058 (2016). 1, 2, 6, 8, 9, 31 [RS15] Prasad Raghavendra and Tselil Schramm, Tight lower bounds for planted clique in the degree-4 SOS program, CoRR abs/1507.05136 (2015). 9 [RW17] Prasad Raghavendra and Benjamin Weitz, On the bit complexity of sum-of-squares proofs, CoRR abs/1702.05139 (2017). 40, 42 [Sch08] Grant Schoenebeck, Linear level lasserre lower bounds for certain k-csps, FOCS, IEEE Computer Society, 2008, pp. 593–602. 8 [Tre09] Luca Trevisan, Max cut and the smallest eigenvalue, STOC, ACM, 2009, pp. 263–272. 1 [TS14] Ryota Tomioka and Taiji Suzuki, Spectral norm of random tensors, arXiv preprint arXiv:1407.1870 (2014). 9 [Wei17] Benjamin Weitz, Polynomial proof systems, effective derivations, and their applications in the sum-of-squares hierarchy, Ph.D. thesis, UC Berkeley, 2017. 42 [ZHT06] Hui Zou, Trevor Hastie, and Robert Tibshirani, Sparse principal component analysis, Journal of Computational and Graphical Statistics 15 (2006), no. 2, 265–286. 2 39 A Bounding the sum-of-squares proof ideal term We give conditions under which sum-of-squares proofs are well-conditioned, using techniques similar to those that appear in [RW17] for bounding the bit complexity of SoS proofs. We begin with some definitions. Definition A.1. Let P be a polynomial optimization problem and let D be the uniform distribution over the set of feasible solutions S for P. Define the degree-2d moment matrix of D to be X D  s∼D [ ŝ ⊗2d ], where ŝ  [1 s]⊤ . • We say that P is k-complete on up to degree 2d if every zero eigenvector of X D has a degree-k derivation from the ideal constraints of P. Theorem A.2. Let P be a polynomial optimization problem over variables x ∈ ’n of degree at most 2d, with objective function f (x) and ideal constraints {1 j (x)  0} j∈[m] . Suppose also that P is 2d-complete up to degree 2d. Let G be the matrix of ideal constraints in the degree-2d SoS proof for P. Then if • the SDP optimum value is bounded by n O(d) • the coefficients of the objective function are bounded by n O(d) , • there is a set of feasible solutions S ⊆ ’n with the property that for each α ⊆ [n]d , |α| 6 d for which χα is not identically zero over the solution space, there exists some s ∈ S such that the square monomial χα (s)2 > n −O(d) , it follows that the SoS certificate for the problem is well-conditioned, with no value larger than n O(d). To prove this, we essentially reproduce the proof of the main theorem of [RW17], up to the very end of the proof at which point we slightly deviate to draw a different conclusion. Proof. Following our previous convention, the degree-2d sum-of-squares proof for P is of the form sdpOpt − f (x)  a(x) + 1(x), where the 1(x) is a polynomial in the span of the ideal constraints, and A is a sum of squares of polynomials. Alternatively, we have the matrix characterization, sdpOpt −hF, x̂ ⊗2d i  hA, x̂ ⊗2d i + hG, x̂ ⊗2d i, where x̂  [1 x]⊤ , F, A, and G are matrix polynomials corresponding to f , a, and 1 respectively, and with A  0. Now let s ∈ S be a feasible solution. Then we have that sdpOpt −hF, s ⊗2d i  hA, s ⊗2d i + hG, s ⊗2d i  hA, s ⊗2d i, where the second equality follows because each s ∈ S is feasible. By assumption the left-hand-side is bounded by n O(d). We will now argue that the diagonal entries of A cannot be too large. Our first step is to argue that A cannot have nonzero diagonal entries unless there is a solution element in the solution Let X D  [x ⊗2d ] be the 2d-moment matrix of the uniform distribution of feasible solutions to 40 P. Define Π to be the orthogonal projection into the zero eigenspace of X D . By linearity and orthonormality, we have that hX D , Ai  X D , (Π + Π⊥ )A(Π + Π⊥ )  X D , Π⊥ AΠ⊥ + X D , ΠAΠ⊥ + X D , Π⊥ AΠ + hX D , ΠAΠi . By assumption P is 2d-complete on D up to degree 2d, and therefore Π is derivable in degree 2d from the ideal constraints {1 j } j∈[m] . Therefore, the latter three terms may be absorbed into G, or more formally, we can set A′  Π⊥ AΠ⊥ , G′  G + (Π + Π⊥ )A(Π + Π⊥ ) − Π⊥ AΠ⊥ , and re-write the original proof sdpOpt −hF, x̂ ⊗2d i  hA′ , x̂ ⊗2d i + hG′ , x̂ ⊗2d i. (A.1) The left-hand-side remains unchanged, so we still have that it is bounded by n O(d) for any feasible solution s ∈ S. Furthermore, the nonzero eigenspaces of X D and A′ are identical, and so A′ cannot be nonzero on any diagonal entry which is orthogonal to the space of feasible solutions. Now, we argue that every diagonal entry of A′ is at most n O(d) . To see this, for each diagonal term χ2α , we choose the solution s ∈ S for which χα (s)2 > n −O(d) . We then have by the PSDness of A′ that A′α,α · χα (s)2 6 hs ⊗2d , A′i 6 n O(d) , which then implies that A′α,α 6 n O(d) . It follows that Tr(A′) 6 n O(d) , and again since A′ is PSD, kA′ kF 6 p Tr(A′) 6 n O(d) . (A.2) Putting things together, we have from our original matrix identity (A.1) that kG′ kF  k sdpOpt −A′ − F kF 6 k sdpOpt kF + kA′ kF + kF kF 6 k sdpOpt kF + n O(d) + kF kF (triangle inequality) (from (A.2)). Therefore by our assumptions that k sdpOpt k, kF kF  n O(d) , the conclusion follows.  We now argue that the conditions of this theorem are met by several general families of problems. Corollary A.3. The following problems have degree-2d SoS proofs with all coefficients bounded by n O(d) : 1. The hypercube: Any polynomial optimization problem with the only constraints being {x 2i  x i }i∈[n] or {x 2i  1}i∈[n] and objective value at most n O(d) over the set of integer feasible solutions. (Including max k-csp). 2. The hypercube with balancedness constraints: Any polynomial optimization problem with the only Í constraints being {x 2i − 1}i∈[n] ∪ { i x i  0}. (Including community detection). Í 3. The unit sphere: Any polynomial optimization problem with the only constraints being { i∈[n] x 2i  1} and objective value at most n O(d) over the set of integer feasible solutions. (Including tensor PCA). 41 4. The sparse hypercube: As long as 2d 6 k, any polynomial optimization problem with the only Í Í constraints being {x 2i  x i }i∈[n] ∪ { i∈[n] x i  k}, or {x 3i  x i }i∈[n] ∪ { i∈[n] x 2i  k}, and objective value at most n O(d) over the set of integer feasible solutions. (Including densest k-subgraph and the Boolean version of sparse PCA). 5. The max clique problem. We prove this corollary below. For each of the above problems, it is clear that the objective value is bounded and the objective function has no large coefficients. To prove this corollary, we need to verify the completeness of the constraint sets, and then demonstrate a set of feasible solutions so that each square term receives non-negligible mass from some solution. A large family of completeness conditions were already verified by [RW17] and others (see the references therein): Proposition A.4 (Completeness of canonical polynomial optimization problems (from Corollary 3.5 of [RW17])). The following pairs of polynomial optimization problems P and distributions over solutions D are complete: 1. If the feasible set is x ∈ ’n with {x 2i  1}i∈[n] or {x 2i  x i }i∈[n] , P is d-complete up to degree d Í (e.g. if P is a CSP). This is still true of the constraints {x 2i  1}i∈[n] ∪ { i x i  0} (e.g. if P is a community detection problem). 2. If the feasible set is x ∈ ’n with is the tensor PCA problem). Í i∈[n] x 2i  α, then P is d-complete on D up to degree d (e.g. if P 3. If P is the max clique problem with feasible set x ∈ ’n with {x 2i  x i }i∈[n] ∪ {x i x j  0}(i, j)∈E , then P is d-complete on D up to degree d. A couple of additional examples can be found in the upcoming thesis of Benjamin Weitz [Wei17]: Proposition A.5 (Completeness of additional polynomial optimization problems) [Wei17]). The following pairs of polynomial optimization problems P and distributions over solutions D are complete: 1. If P is the densest k-subgraph relaxation, with feasible set x ∈ ’n with {x 2i  x i }i∈[n] ∪ { k}, P is d-complete on D up to degree d 6 k. Í i∈[n] xi  2. If P is the sparse PCA relaxation with sparsity k, with feasible set x ∈ ’n with {x 3i  x i }i∈[n] ∪ Í { i∈[n] x 2i  k}, P is d-complete up to degree d 6 k/2. Proof of Corollary A.3. We verify the conditions of Theorem A.2 separately for each case. 1. The hypercube: the completeness conditions are satisfied by Proposition A.4. We choose the ® for which χ2α (s)  1 always. set of feasible solutions to contain a single point, s  1, 2. The hypercube with balancedness constraints: the completeness conditions are satisfied by Proposition A.4. We choose the set of feasible solutions to contain a single point, s, some perfectly balanced vector, for which χ2α (s)  1 always. 42 3. The unit sphere: the completeness conditions are satisfied by Proposition A.4. We choose the set of feasible solutions to contain a single point, s  √1n · ®1, for which χ2α (s) > n −d as long as |α| 6 d, which meets the conditions of Theorem A.2. 4. The sparse hypercube: the completeness conditions are satisfied by Proposition A.5. Here, Í we choose the set of solutions S  {x ∈ {0, 1} n | i x i  k}. as long as k > d, for any |α| 6 d we have that χS (x)2  1 when s is 1 on α. 5. The max clique problem: the completeness conditions are satisfied by Proposition A.4. We choose the solution set S to be the set of 0, 1 indicators for cliques in the graph. Any α that corresponds to a non-clique in the graph has χα identically zero in the solution space. Otherwise, χα (s)2  1 when s ∈ S is the indicator vector for the clique on α. This concludes the proof.  B Lower bounds on the nonzero eigenvalues of some moment matrices In this appendix, we prove lower bounds on the magnitude of nonzero eigenvalues of covariance matrices for certain distributions over solutions. Many of these bounds are well-known, but we re-state and re-prove them here for completeness. We first define the property we want: Definition B.1. Let P be a polynomial optimization problem and let D be the uniform distribution over the set of feasible solutions S for P. Define the degree-2d moment matrix of D to be X D  x∼D [ x̂ ⊗2d ], where x̂  [1 x]⊤ . • We say that D is δ-spectrally rich up to degree 2d if every nonzero eigenvalue of X D is at least δ. Proposition B.2 (Spectral richness of polynomial optimization problems). The following distributions over solutions D are polynomially spectrally rich: 1. If D is the uniform distribution over {±1} n , then D is polynomially spectrally rich up to degree d 6 n. 2. If D is the uniform distribution over α · Sn−1 , then D is polynomially spectrally rich up to degree d 6 n. 3. If D is the uniform distribution over x ∈ {1, 0} n with kx k0  k, then if 2d 6 k, D is polynomially spectrally rich up to degree d. 4. If D is the uniform distribution over x ∈ {±1, 0} n with kx k0  k, then if 2d 6 k, D is polynomially spectrally rich up to degree d. def Proof. In the proof of each statement, denote the 2dth moment matrix of D by X D  x∼D [x ⊗2d ]. Because X D is a sum of rank-1 outer-products, an eigenvector of X D has eigenvalue 0 if and only if it is orthogonal to every solution in the support of D, and therefore the zero eigenvectors correspond exactly to the degree at most d constraints that can be derived from the ideal constraints. 43 Now, let p1 (x), . . . , p r (x) be a basis for polynomials of degree at most 2d in x which is orthonormal with respect to D, so that x∼D [p i (x)p j (x)]  ( 1 ij 0 otherwise If p̂ i is the representation of p i in the monomial basis, we have that (p̂ i )⊤ X D p̂ j  Therefore, the matrix R  Í ⊤ i e i ( p̂ i ) x∼D [p i (x)p j (x)]. diagonalizes X D , RX D R ⊤  Id . It follows that the minimum non-zero eigenvalue of X D is equal to the smallest eigenvalue of (RR ⊤ )−1 , which is in turn equal to σ 1(R)2 where σmax (R) is the largest singular value of R. Therefore, max for each of these cases it suffices to bound the singular values of the change-of-basis matrix between the monomial basis and an orthogonal basis over D. We now proceed to handle each case separately. 1. D uniform over hypercube: In this case, the monomial basis is an orthogonal basis, so R is the identity on the space orthogonal to the ideal constraints, and σmax (R)  1, which completes the proof. 2. D uniform over sphere: Here, the canonical orthonormal basis the spherical harmonic polynomials. Examining an explicit characterization of the spherical harmonic polynomials (given for example in [DX13], Theorem 5.1), we have that when expressing p i in the monomial basis, no coefficient of a monomial (and thus no entry of p̂ i ) exceeds n O(d) , and since there
Í are at most n d polynomials each with di0 nd 6 n d coefficients, employing the triangle inequality we have that σmax (R) 6 n O(d) , which completes the proof. 3. D uniform over {x ∈ {0, 1} k | kx k0  k}: In this case, the canonical orthonormal basis is the correctly normalized Young’s basis (see e.g. [Fil16] Theorems 3.1,3.2 and 5.1), and agan we have that when expressing an orthonormal basis polynomial p i in the monomial basis, no coefficient exceeds n O(d) . As in the above case, this implies that σmax (R) 6 n O(d) and completes the proof. 4. D uniform over {x ∈ {±1, 0} k | kx k0  k}: Again the canonical orthonormal basis is Young’s basis with a correct normalization. We again apply [Fil16] Theorems 3.1,3.2, but this time we calculate the normalization by hand: we have that in expressing each p i , no element of the monomial basis has coefficient larger than n O(d) multiplied by the quantity x∼D " d Ö i1 2 (x2i−1 − x2i ) #  O(1). This gives the desired conclusion.  44 C From Boolean to Gaussian lower bounds In this section we show how to prove our SoS lower bounds for Gaussian PCA problems using the lower bounds for Boolean problems in a black-box fashion. The techniques are standard and more broadly applicable than the exposition here but we prove only what we need. The following proposition captures what is needed for tensor PCA; the argument for sparse PCA is entirely analogous so we leave it to the reader. n Proposition C.1. Let k ∈ Ž and let A ∼ {±1}( k ) be a symmetric random Boolean tensor. Suppose that for n every A ∈ {±1}( k ) there is a degree-d pseudodistribution ˜ satisfying {kx k 2  1} such that A ˜ hx ⊗k , Ai  C . n Let T ∼ N(0, 1)( k ) be a Gaussian random tensor. Then T max ˜ hx ⊗k , Ti > Ω(C) ˜ where the maximization is over pseudodistributions of degree d which satisfy {kx k 2  1}. Proof. For a tensor T ∈ (’n )⊗k , let A(T) have entries A(T)α  sign(Tα ). Now consider T ˜ A(T) hx ⊗k , Ti  Õ α T ˜ A(T) x α Tα where α ranges over multi-indices of size k over [n]. We rearrange each term above to A(T) where 1 ∼ N(0, 1). Since ( ˜ A(T) x α ) · Tα | A(T) Tα  ( ˜ A(T) x α ) · A(T)α · A(T) |1 | |1 | is a constant independent of n, all of this is Ω(1) · Õ α A ˜ A x α · Aα  C . 45  

arXiv:1709.08605v1 [cs.CV] 25 Sep 2017 Muon Trigger for Mobile Phones M Borisyak1,2 , M Usvyatsov2,3,4 , M Mulhearn5 , C Shimmin6,7 and A Ustyuzhanin1,2 1 National Research University Higher School of Economics, 20 Myasnitskaya st., Moscow 101000, Russia 2 Yandex School of Data Analysis, 11/2, Timura Frunze st., Moscow 119021, Russia 3 Moscow Institute of Physics and Technology, 9 Institutskiy per., Dolgoprudny, Moscow Region, 141700, Russia 4 Skolkovo Institute of Science and Technology, Skolkovo Innovation Center, Building 3, Moscow 143026, Russia 5 University of California, Davis, 1 One Shields Avenue, Davis, CA 95616, USA 6 University of California, Irvine, 4129 Frederick Reines Hall, Irvine, CA 92697-4575, USA 7 Yale University, 217 Prospect Street, New Haven, CT 06520, USA E-mail: mborisyak@hse.ru Abstract. The CRAYFIS experiment proposes to use privately owned mobile phones as a ground detector array for Ultra High Energy Cosmic Rays. Upon interacting with Earth’s atmosphere, these events produce extensive particle showers which can be detected by cameras on mobile phones. A typical shower contains minimally-ionizing particles such as muons. As these particles interact with CMOS image sensors, they may leave tracks of faintly-activated pixels that are sometimes hard to distinguish from random detector noise. Triggers that rely on the presence of very bright pixels within an image frame are not efficient in this case. We present a trigger algorithm based on Convolutional Neural Networks which selects images containing such tracks and are evaluated in a lazy manner: the response of each successive layer is computed only if activation of the current layer satisfies a continuation criterion. Usage of neural networks increases the sensitivity considerably comparable with image thresholding, while the lazy evaluation allows for execution of the trigger under the limited computational power of mobile phones. 1. Introduction The problem of pattern detection over a set of images arises in many applications. The CRAYFIS experiment is dedicated to observations of Ultra-High-Energy Cosmic Rays (UHECR) by a distributed network of mobile phones provided by volunteers. In the process of interaction with the Earth’s atmosphere, UHECRs produce cascades of particles called Extensive Air Showers (EAS). Some of the particles reach the ground, affecting areas of up to several hundreds of meters in radius. These particles can be detected by cameras on mobile phones, and a localized coincidence of particle detection by several phones can be used to observe very rare UHECR events [1]. This approach presents a number of challenges. In order to observe an EAS, each active smartphone needs to continuously monitor its camera output by scanning megapixel-scale images at rates of 15-60 frames per second. This generates a vast amount of raw data, which presents problems both for volunteers1 and experimenters if transmitted to data processing servers for later analysis. However, the recorded data contains almost entirely random camera noise, as signals from cosmic ray interactions are expected to occur in fewer than 1 in 10,000 image frames. As there would be potentially millions of smartphones operating simultaneously, it is critical to utilize the local processing power available on each device to select only the most interesting data. Hence, a trigger algorithm is required to filter out background data and identify possible candidates for cosmic rays traces. It is also important that the camera monitoring is subject to negligible dead time; therefore any trigger must operate with an average evaluation response time on the order of 30ms to track with the raw data rate. Some constituents of an EAS, such as electrons and gamma-ray photons, leave bright traces in the camera [1]. In this case, the simplest trigger strategy is cut on brightness (if there are bright pixels in a fragment then this fragment is considered interesting). This usually enough to provide acceptable background rejection rate in the case of bright traces, and given a target background rejection rate it is possible to automatically determine the threshold value for decision making. However, this strategy is much less effective against another component of the shower, comprising minimally-ionizing particles such as high-energy muons. These particles may leave relatively faint signals in the pixels they traverse, possibly at a level comparable to the sensor’s intrinsic noise. Nevertheless, these minimally-ionizing particles traverse the sensor’s pixels in distinctively straight lines. If these tracks are long enough in the plane of the sensor, there is a low probability of the same pattern emerging from intrinsic random camera noise. Thus it is still possible to discriminate even these faintly-interacting particles from background. In this work, we propose a novel approach for fast visual pattern detection, realized here as a trigger designed for fast identification of muon traces on mobile phone’s camera. This method is based on Convolutional Neural Networks and does not contain any specific assumptions for identification of muon traces, hence, in principle, it can be applied to any visual pattern detection problem. 2. Related Work Minimally-ionizing particles are characterized by the pattern of activated pixels they leave over a small region of an exposure. Hence the problem of minimally-ionizing particle detection can be transformed to the problem of pattern detection over the image. Several attempts were performed to solve pattern detection problem in different setups. The solution proposed in works [2] and [3] utilizes Convolutional Neural Networks (CNNs). Certain properties of CNNs such as translation invariance, locality, and correspondingly few weights to be learned, make them particularly well suited to extracting features from images. The performance of even simple CNNs on image classification tasks have been shown to be very good compared to other methods [4]. However, the training and evaluation of CNNs requires relatively intense computation to which smartphones are not particularly well suited. Viola and Jones in [5] introduced the idea of a simple feature-based cascading classifier. They proposed using small binary filters as feature detectors and to increase computational power from cascade to cascade. Bagherinzhad et al. in [6] enumerated a wide range of methods which have been proposed to address efficient training and inference in deep neural networks. Although these methods are orthogonal to our approach, they may be incorporated with the method described here in order to improve efficiency in related tasks. 1 For example, it may quickly exceed smartphone’s storage capacity or introduce a considerable load on networks. 3. CNN trigger The key insight of the proposed method is to view a Deep Convolutional Neural Network (CNN) as a chain of triggers, or cascades: each trigger filters its input stream of data before passing it further down the chain. The main feature of such chains is that amount of data passing successfully through the chain gradually decreases, while the complexity of triggers gradually increases, allowing finer selection with each subsequent trigger. This architecture allows one to effectively control the computational complexity of the overall chain, and usually, to substantially decrease the amount of computational resources required [5]. Convolutional Neural Networks are particularly well suited for adopting this approach as instead of passing an image itself throughout the chain, the CNN computes a series of intermediate representations (activations of hidden layers) of the image [4]. Following the same reasoning as in Deeply Supervised Nets (DSN) [7], one can build a network for which discriminative power of intermediate representations grows along the network, making it possible to treat such CNN as a progressive trigger chain 2 . In order to build a trigger chain from a CNN, we propose a method similar to DSN: each layer of the CNN is extended with a binary trigger based on the image representation obtained by this layer. In the present work we use logistic regression as model for the trigger, although, any binary classifier with a differentiable model would be sufficient. The trigger is applied to each region of the output image to determine if that region should proceed further down the trigger chain. We call these layers with their corresponding triggers a convolutional cascade, by analogy with Viola-Jones cascades [5]. The output of the trigger at each stage produces what we refer to as an activation map, to be passed to the next layer, as illustrated in Fig. 1a. From another perspective, this approach can be seen as an extension of the CNN architecture in which network computations concerning regions of the input image are halted as soon as a negative outcome for a region becomes evident3 . This is effectively accomplished by generalizing the convolution operator to additionally accept as input an activation map indicating the regions to be computed. In the case where sparse regions of activation are expected, this lazy application can result in much fewer computations being performed. After each application of the lazy convolution, the activation map for the subsequent layer is furnished by the trigger associated with the current layer. The whole chain is illustrated in Fig. 1b. Training of the CNN trigger may be problematic for gradient methods, since prediction is no longer a continuous function of network parameters. This is because the lazy convolution, described above, is in general non-differentiable. In order to overcome this limitation, we propose to use a slightly different network architecture during training by substituting a differentiable approximation of the lazy convolution operator. The basic idea is that instead of skipping the evaluation of unlikely regions, we simply ensure that any region which has low activation on a given cascade will continue to have low activation on all subsequent cascades. In this scheme, the evaluation is no longer lazy, but since training may be performed on much more powerful hardware, this is not a concern. To accomplish this, we first replace the binary activation maps (which result from the trigger classification) with continuous probability estimates. Secondly, we introduce intermediate activation maps, which are computed by the trigger function at each layer. The intermediate map is multiplied by the previous layer’s activation map to produce a refined activation map4 . In this way, the activation probability for any region is nonincreasing with each cascade layer. 2 However, in contrast to DSN, the growth of discriminative power is a requirement for an effective trigger chain rather than for network regularization. 3 Lazy application can be viewed as a variation of attention mechanisms recently proposed in Deep Learning literature, see e.g. [8]. cascade input cascade output convolution convolutions, subsampling activation, lazy application lazy application activation maps input activation map output activation map (b) CNN trigger structure (a) convolutional cascade Figure 1: Fig. 1a shows the building block of the CNN trigger, an individual convolutional cascade. In contrast to conventional convolutional layers, the convolutional cascade has an additional input, the activation map, which indicates regions to which the convolutional operator should be applied (lazy application, denoted by dashed lines). The activation map is updated by the associated trigger (represented by the s-shaped node), which may be passed on to the subsequent cascade or interpreted as the final output indicating regions of interest. Fig. 1b shows the full structure of CNN trigger as a sequence of convolutional cascades. Initially the whole image is activated (red areas). As the image proceeds through the chain, the activated area becomes smaller as each cascade refines the activation map. The process is depicted schematically in Fig. 2. This differentiable version of the lazy application operation for the ith cascade is described by the following equations: Ii Âix,y i A A0x,y = = hi (I i−1 ); i σ (Ix,y ); i i−1 (1) i = Â ⊗ A := 1, (2) ; (3) (4) where I i is the intermediate representation of the input image I 0 after successive applications of CNN layers, and hi represents the transformation associated with the ith layer of the CNN (typically this would be convolution, nonlinearity, and pooling). σ i is the function associated with the trigger (in our case, logistic regression), and its result Âi is the intermediate activation map of the ith cascade. Finally, Ai is the differentiable version of the activation map, given by the element-wise product of the intermediate activation with the previous layer’s activation. That elements of the initial activation map A0 are set to 1, and the subscripts x, y denote region position. Note that the dimensions of the activation map define the granularity of regions-of-interest that may be triggered, and may in general be smaller than the dimensions of the input image. In this case, the trigger function σ should incorporate some downsampling such as max-pooling. We also note that since similar but still technically different networks are used for training and prediction, special care should be taken while transitioning from probability estimations to binary classification. In particular, additional adjustment of classifier thresholds may be 4 If layers of underlying CNN contains pooling, i.e. change size of the image, pooling should be applied to intermediate activation maps as well. convolutions, subsampling activation intermediate activation maps x combination x x activation maps Figure 2: Schematic of the CNN trigger used for training. To make the network differentiable, lazy application is replaced by its approximation, that does not involve any “laziness”. Activation maps are approximated as the elementwise product of unconditional trigger response (intermediate activation maps) and the previous activation map. required. Nevertheless, in the present work no significant differences in the two networks’ behaviors were found (see Fig. 4). To train the network we utilize cross-entropy loss. Since activation maps are also intermediate results, and the activation map An of the last cascade is the final result for a network of n cascades, the loss can be written as: Ln = − X 1 n n Yx,y log Ix,y + γ n (1 − Yx,y ) log(1 − Ix,y ) W × H x,y (5) where Y ∈ RW ×H denotes the ground truth map with width W and height H. The truth map is defined with Yx,y = 1 if the region at coordinates (x, y) contains a target object, otherwise Yx,y = 0. The coefficient γ n is introduced to provide control over the penalty for triggering on background. If the cross-entropy term (5) is the only component of the loss, the network will have no particular incentive to assign regions a low activation on early cascades, limiting the benefit of the lazy evaluation architecture. One approach to force network to produce good results on intermediate layers is to directly introduce penalty term C for unnecessary computations: C= n X ci i=1 X (1 − Yx,y ) Ai−1 x,y (6) x,y where ci represents the per-region cost of performing convolution and trigger in the ith cascade. We use a naive estimation of coefficients the ci , assuming, for simplicity, that convolution is performed by elementwise multiplication of the convolutional kernel with a corresponding image patch. In this case, for l filters of size (k, k) applied to image with m channels: ci = 2 mlk | {z } multiplications + l(mk 2 − 1) + |2l{z − 1} | {z } summations (7) trigger Combining these terms, the resulting total loss function is given by: L = Ln + βC (8) where the parameter β is introduced to regulate the trade-off between computational efficiency and classification quality. Another approach is to apply a DSN technique: L = Ln + n−1 X αi Li . (9) i=1 Here, Li is the loss associated with ith cascade (i.e. companion loss in DSN terminology) defined by analogy with (5). The coefficients αi regulate the trade-off between losses on different cascades5 . In the present work, we find that the objectives defined by (8) and (9) are highly correlated. However, (9) seems to propagate gradients more effectively, resulting in faster training. 4. Experiments 4.1. Dataset As of this writing, no labeled dataset of CMOS images containing true muon tracks is available6 . Instead, an artificial dataset was constructed with properties similar to those expected from real data, in order to assess the CNN trigger concept. (a) original traces (b) composition Figure 3: Test dataset creation steps: 3a selection of bright photon tracks, 3b track brightness is lowered and superimposed on noisy background. To construct the artificial dataset, images were taken from a real mobile phone exposed to radioactive 226 Ra, an X-ray photon source. These photons interact in the sensor primarily via compton scattering, producing energetic electrons which leave tracks as seen in Fig. 3a. These tracks are similar to those expected by muons, the main difference being that the electron tracks tend to be much brighter than the background noise, rendering the classification problem almost trivial. Therefore, the selected particle tracks are renormalized such that their average brightness is approximately at the level of the camera’s intrinsic noise. Gloom traces are than superimposed on the background with some Gaussian noise to modeling intrinsic camera sensor noise. An example of the resulting artificial data is shown in Fig. 3b. After these measures, the dataset better emulates the case of low-brightness muons, and also forces the classifier to use more sophisticated (geometric) features for classification. 5 One may find αi ∼ ci to be a relatively good heuristic. To obtain real data and fully validate performance of the algorithm, an experimental setup with muon scintillators is scheduled this year. 6 (a) intermediate activation maps and ground truth map (b) activation maps and ground truth map (c) binary activation maps and ground truth map Figure 4: Evaluation of the trigger CNN (using the input image from Fig. 3b). Figs. 4a and 4b are activation maps for training regime; Fig. 4c are binary activation maps for the application regime. The resolution of the map is reduced after each cascade to match the downsampling of the internal image representation. 4.2. CNN trigger evaluation To evaluate the performance of the method, we consider the case of a CNN trigger with 4 cascades. The first cascade has a single filter of size 1 × 1, equivalent to simple thresholding. The second, third, and fourth cascades have 1, 3, and 6 filters of size 3 × 3, respectively. Within each cascade, convolutions are followed by 2 × 2 max-pooling. Due to the simple structure of the first cascade, its coefficient c1 from (6) is set to 1. As motivated in Sec. 1, a successful trigger must run rapidly on hardware with limited abilities. One of the simplest algorithms that satisfies this restriction, thresholding by brightness, is chosen as baseline for comparison.7 This strategy yields a background rejection rate of around 10−2 (mainly due to assumptions built in dataset) with perfect signal efficiency. Two versions of the CNN trigger with average computational costs8 of 1.4 and 2.0 operations per pixel were trained. This computational cost is controlled by varying coefficients in the loss function (9). For each, signal efficiency and background rejection rates at three working points 7 In order to obtain comparable results, the output of thresholding was max-pooled to match the size of the CNN trigger output. 8 As estimated by (6). are presented in Table 1. Fig. 4 shows some typical examples of activation maps for different network regimes. complexity signal efficiency background rejection 0.90 0.60 1.4 op. per pixel 0.95 0.39 0.99 0.12 0.90 0.65 2.0 op. per pixel 0.95 0.44 0.99 0.15 Table 1: CNN trigger performance for two models with computational costs 1.4 and 2.0 operations per pixel. Different points for signal efficiency and background rejection were obtained by varying threshold on output of the CNN trigger (i.e. activation map of the last cascade). These results indicate a significant improvement of background rejection rate relative to the baseline strategy, even for nearly perfect signal efficiency. Another performance metric that is interesting to consider is, the normalized computational complexity: " n #−1 X X Ĉ = C · ci (1 − Yx,y ) . (10) i=1 x,y For the models described above, the normalized computational complexity, Ĉ, is around 4-5 percent, which indicates that a significant amount of computational resources is saved due to lazy application, as compared to a conventional CNN with the same structure. 5. Conclusion We have introduced a novel approach to construct a CNN trigger for fast visual pattern detection of rare events, designed particularly for the use case of fast identification of muon tracks on mobile phone cameras. Nevertheless, the proposed method does not contain any application-specific assumptions and can be, in principle, applied to a wide range of problems. The method extends Convolutional Neural Networks by introducing lazy application of convolutional operators, which can achieve comparable performance with lower computational costs. The CNN trigger was evaluated on an artificial dataset with properties similar to those expected from real data. Our results show significant improvement of background rejection rate relative to a simple baseline strategy with nearly perfect signal efficiency, while the per-pixel computational cost of the algorithm is increased by less than a factor of 2. The effective computational cost is equivalent to 4-5 percent of the cost required by a conventional CNN of the same size. Therefore the method can enable the evaluation of powerful CNNs in instances where time and resources are limited, or where the network is very large. This is a promising result for CNNs in many other possible applications, such as very fast triggering with radiation-hard electronics, or power-efficient realtime processing of high resolution sensors. References [1] Whiteson D, Mulhearn M, Shimmin C, Cranmer K, Brodie K and Burns D 2016 Astroparticle Physics 79 1–9 [2] Li H, Lin Z, Shen X, Brandt J and Hua G 2015 A convolutional neural network cascade for face detection Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition pp 5325–5334 [3] Ren S, He K, Girshick R and Sun J 2015 Faster r-cnn: Towards real-time object detection with region proposal networks Advances in neural information processing systems pp 91–99 [4] LeCun Y and Bengio Y 1995 The handbook of brain theory and neural networks 3361 1995 [5] Viola P and Jones M 2001 Rapid object detection using a boosted cascade of simple features Computer Vision and Pattern Recognition, 2001. CVPR 2001. Proceedings of the 2001 IEEE Computer Society Conference on vol 1 (IEEE) pp I–511 [6] Hessam B, Mohammad R and Ali F 2016 arXiv preprint arXiv:1611.06473 [7] Lee C Y, Xie S, Gallagher P, Zhang Z and Tu Z 2015 Deeply-supervised nets. AISTATS vol 2 p 6 [8] Xu K, Ba J, Kiros R, Cho K, Courville A C, Salakhutdinov R, Zemel R S and Bengio Y 2015 Show, attend and tell: Neural image caption generation with visual attention. ICML vol 14 pp 77–81 

arXiv:1708.09047v1 [math.GR] 29 Aug 2017 Computable groups which do not embed into groups with decidable conjugacy problem Arman Darbinyan 1 Abstract We show the existence of finitely generated torsion-free groups with decidable word problem that cannot be embedded into groups with decidable conjugacy problem. This answers a well-known question of Collins from the early 1970’s. 2 Introduction Two of the most central decision problems associated with finitely generated groups are word and conjugacy problems. One of the important questions about these problems is concerning about the relation between them. For example, if the conjugacy problem is decidable for a finitely generated group G, then the word problem is decidable as well. However, in general, the inverse is far from being true, [16, 6, 17, 15, 8]. If G is a finitely generated group and H ≤ G is a subgroup of finite index, then the word problem in G is decidable if and only if it is decidable for H. However, it is shown by Goryaga-Kirkinskii, [10], and independently by CollinsMiller, [7], that subgroups of index 2 of some specific finitely generated groups have decidable (respectively, undecidable) conjugacy problem, while the groups themselves have undecidable (respectively, decidable) conjugacy problem. Another important type of questions about word and conjugacy problems in groups is the following: Is it true that every finitely generated group with decidable word problem (respectively, conjugacy problem) embeds in a finitely presented group with decidable word problem (respectively, conjugacy problem)? Both of these questions have positive answer. The answer for the word problem is obtained by Clapham in 1967, [5], based on the classical embedding theorem of Higman (see [13]), while the analogous question for the conjugacy problem was a long-standing open problem until it got positive answer in 2004 by a work of Olshanskii and Sapir. See [18] and also [19]. In light of the aforementioned, a natural question about the connection of word and conjugacy problems in groups is the following question, asked by 1 Donald Collins in the early 1970s. Question 1. Can every torsion-free group with solvable word problem be embedded in a group with solvable conjugacy problem? This question appears in the 1976 edition of The Kourovka Notebook as Problem 5.21, [12]. Probably, the first source where this problem was posed in a written form is [3]. For yet another source, see [2]. It was mentioned by Collins in [12] that due to an example by A. Macintyre, there exists a group with torsions which cannot be embedded into a finitely generated group with decidable conjugacy problem. However, the case for torsion-free groups remained open until now. Indeed, one of the reasons why the torsion and torsion-free cases are different is based on the observation that conjugate elements in a group must have the same order, and since in a torsion-free group all non trivial elements have the same (infinite) order, one cannot make use of this observation in order to answer Question 1. In [19], Olshanskii and Sapir showed the following theorem which gives a positive answer to Question 1 under the stronger assumption of decidability of the power problem. Theorem 1 (Olshanskii-Sapir, [19]). Every countable group with solvable power and order problems is embeddable into a 2-generated finitely presented group with solvable conjugacy and power problems. Note that as it is defined in [19], for a given group G the power problem is said to be decidable, if there exists an algorithm which for any given pair (g, h) of elements from G decides whether or not g is equal to some power of h in G. The order problem is decidable in G if there exists an algorithm which for each input g ∈ G computes the order of g. The main result of the current work is the negative answer to Question 1 in the general case. Theorem 2. There exists a finitely presented torsion-free group G with decidable word problem such that G cannot be embedded into a group with decidable conjugacy problem. A remarkable theorem of Osin (see [20]) says that every torsion-free countable group can be embedded into a two generated group with exactly two conjugacy classes. In the context of this theorem, it is very natural to ask whether or not every torsion-free countable group with decidable word problem (= computable group) can be embedded into a group with exactly two conjugacy classes and with decidable word problem. A more relaxed version of this question would be whether or not every torsion-free countable group with decidable word problem can be embedded in a finitely generated recursively presented group with finitely many conjugacy classes. It turns out that a direct consequence of Theorem 2 gives negative answer to both of these questions. 2 In fact, the decidability of the conjugacy problem for groups with exactly two conjugacy classes is equivalent to the decidability of the word problem. Even more, as it is shown in a recent paper of Miasnikov and Schupp [15], a finitely generated recursively presented group with finitely many conjugacy classes has decidable conjugacy problem. Therefore, a direct corollary from Theorem 2 is the following. Theorem 3. There exists a torsion-free finitely presented group with decidable word problem that cannot be embedded into a finitely generated recursively presented group with finitely many conjugacy classes. Proof. Just take the group G from Theorem 2. Remark 1. In fact, the mentioned result of Miasnikov and Schupp is true not only for finitely generated recursively presented groups, but for all recursively presented groups in general. Therefore, Theorem 3 stays true after dropping the assumption that the group in which the initial group is embedded is finitely generated. (The exact definition of recursive presentations of groups is given in the next section.) Acknowledgements. I would like to thank Alexander Olshanskii for his thoughtful comments on this work. 3 3.1 Preliminaries Groups with decidable word problem A countable group G is said to have recursive presentation, if G can be presented as G = hX | Ri such that X and R are enumerable by some algorithm (i.e. Turing machine). See [11]. If in addition, there is an algorithm which for each pair of words (w, w′ ) from (X ∪ X −1 )∗ verifies whether or not w and w′ represent the same element of G, then the presentation G = hX | Ri is called computable and in case G possesses such a presentation, the group G itself is called computable as well. Modulo some slight variances, the original definition of the concept of computable groups is due to Rabin [21] and Mal’cev [14]. In case the group G is finitely generated (i.e. |X| < ∞) computability property of G is the same as saying that G has decidable word problem. It is not hard to notice that decidability of the word problem does not depend on the finite generating sets. From the computability perspective, the last observation is one of the main advantages of finitely generated groups over countably generated ones, because in case of finitely generated groups decidability of the word problem is an intrinsic property of a group, rather than of its presentation. However, in this paper, to keep the notations as uniform as possible, we say that G has decidable word problem if it is given by a computable presentation. 3 Let G = hx1 , x2 , . . . | r1 , r2 , . . .i, where {x1 , x2 , . . .} and {r1 , r2 , . . .} are recursive enumerations of X and R, respectively. Then, the embedding constructions of [9] and [19] imply the following theorem. Theorem 4. If G = hx1 , x2 , . . . | r1 , r2 , . . .i has decidable word problem, then there exists an embedding Φ : G → H of G into a two generated group H such that the following holds. (1). The word problem is decidable in H; (2). The map i 7→ Φ(xi ) is computable; (3). An element of H is of finite order if and only if it is conjugate to an image under Φ of an element of finite order in G. 3.2 HNN-extensions In the proof of the existence of the group G from Theorem 2 we use some group theoretical constructions based on HNN-extensions. Therefore, in this subsection we would like to recall some well-known basic facts about HNNextensions. The basics of the theory of HNN-extensions can also be found in [13]. Suppose that A, B ≤ H and φ : A → B is a group isomorphism from A to B. Then the HNN-extension H ′ of H with respect to A and B (and φ) and with stable letter t is defined as H ′ = hH, t | t−1 at = φ(a), a ∈ Ai. In the current text, the isomorphism φ will be clear from the context, hence we will simply use the notation H ′ = hH, t | t−1 At = Bi. Clearly, every element h′ of H ′ can be decomposed as a product h ′ = h 0 tǫ 1 h 1 . . . t ǫ n h n , (1) where ǫi ∈ {±1}, hj ∈ H for 1 ≤ i ≤ n, 0 ≤ j ≤ n. The decomposition (1) is said to be in reduced form, if it does not contain subproduct of one of the forms t−1 gi t, gi ∈ A or tgi t−1 , gi ∈ B, for 1 ≤ i ≤ n. Analogously, if H = hXi, then the word u′ ∈ (X ∪ X −1 ∪ {t±1 })∗ given by u ′ = u 0 tǫ 1 u 1 tǫ 2 . . . t ǫ n u n , where ǫi ∈ {±1}, uj ∈ (X ∪ X −1 )∗ , is said to be a reduced word with respect to the HNN-extension H ′ if the decomposition h0 tǫ1 h1 . . . tǫn hn is in reduced form, where hi corresponds to the word ui in H. The following well-known lemma is attributed to Britton in [13]. Lemma 1 (Britton’s Lemma). If the decomposition (1) is reduced and n ≥ 1, then h′ 6= 1 in H ′ . 4 Lemma 2 (See Theorem 2.1 in [13]). Let H ′ = hH, t | t−1 At = Bi be an HNNextension with respect to isomorphic subgroups A and B. Then H embeds in H ′ by the maps h 7→ h, h ∈ H. Lemma 3 (The Torsion Theorem for HNN-extensions. See Theorem 2.4 in [13]). Let H ′ = hH, t | t−1 At = Bi be an HNN-extension. Then every element of finite order in H ′ is a conjugate of an element of finite order in the base H. Thus H ′ has elements of finite order n if an only if H has elements of order n. 4 Proof of Theorem 2 In order to show the existence of G from Theorem 2, first, we will construct a special countable group Ġ with decidable word problem, then G will be defined as a group in which Ġ embeds in a certain way. Two disjoint sets of natural numbers S1 , S2 ⊂ N are called recursively inseparable if there is no recursive set T ⊂ N such that S1 ⊆ T and S2 ⊆ N \ T . The set T is called separating set. Clearly, if two disjoint sets are recursively inseparable, then they cannot be recursive. Indeed, if, say, S1 and S2 are disjoint and, say, S1 is recursive, then as a recursive separating set one could simply take S1 . Nevertheless, it is a well-known fact that there exist disjoint recursively enumerable and recursively inseparable sets. See, for example, [22]. Let us fix two disjoint recursively enumerable and recursively inseparable sets N = {n1 , n2 , . . .} ⊂ N and M = {m1 , m2 , . . .} ⊂ N such that the maps i 7→ ni and i 7→ mi are computable. Now, for all n ∈ N, define An as a torsion-free abelian additive group of rank two with basis {an,0 , an,1 }, i.e. An = han,0 i ⊕ han,1 i and such that the groups A1 , A2 , . . . are disjoint. For all n ∈ N, define the groups Ȧn as follows.  if n ∈ / N ∪M ,  An Ȧn = An / ≪ an,1 = 2i an,0 ≫ if n = ni ∈ N ,  An / ≪ an,1 = 3i an,0 ≫ if n = mi ∈ M. For all n ∈ N and m ∈ {0, 1}, let us denote the images of an,m under the natural homomorphisms An → Ȧn by ȧn,m . Convention. In this text, whenever we deal with an additive group, say, A, ∗ with finite generating set, say, {a1 , . . . , ak }, by P {±a1 , . . . , ±ak } we denote the set of formal finite sums of the form w = λi aji , where λi ∈ Z and aji ∈ 5 {a1 , . . . , ak }, and we say that w is a word formed by letters {±a1 , . . . , ±ak }. Note that this is the additive analogue of the central in combinatorial group theory concept of words, where the alphabet composing P the words is a set of group generators. This is why the finite formal sums w = λi aji we call words from {±a1 , . . . , ±ak }∗ . Before moving forward, we prove the following important lemma. Lemma 4. There exists an algorithm such that for each input n ∈ N and w ∈ {±ȧn,0 , ±ȧn,1 }∗ , it decides whether or not w represents the trivial element in the group Ȧn . Proof. Indeed, since Ȧn is abelian with generating set {ȧn,0 , ȧn,1 }, each word w from {±ȧn,0 , ±ȧn,1 }∗ can be effectively transformed to a word of the form w′ = λ0 ȧn,0 + λ1 ȧn,1 which represents the same element in Ȧn as the initial word w, where λ0 , λ1 ∈ Z. Now, assuming that λ0 6= 0, λ1 6= 0, in order w′ to represent the trivial element in Ȧn it must be that n ∈ N ∪ M, because otherwise, by definition, the group Ȧn is torsion-free abelian of rank 2 with basis {ȧn,0 , ȧn,1 }. In case n ∈ N , by definition we have that ȧn,1 = 2x ȧn,0 , where x is the index of n in N , i.e. n = nx . Similarly, in case n ∈ M, by definition we have that ȧn,1 = 3x ȧn,0 , where x is the index of n in M, i.e. n = mx . Now, if λ0 = 0 and λ1 = 0, then clearly w′ (hence also w) represents the trivial element in Ȧn . Therefore, without loss of generality we can assume that at least one of λ0 and λ1 is not 0. Then, if we treat x as an unknown variable, depending on whether n = nx or n = mx , the equality w′ = 0 would imply one of the following equations: λ0 + λ1 2x = 0 (2) λ0 + λ1 3x = 0, (3) or respectively. This observation suggests that in case λ0 6= 0 or λ1 6= 0, in order to verify whether or not w′ = 0 in Ȧn , we can first try to find x satisfying (2) or (3), and in case such an x does not exist, conclude that w′ (hence, also w) does not represent the trivial element in Ȧn . Otherwise, if x is the root of the equation (2), we can check whether or not n = nx (since N is recursively enumerable, this checking can be done algorithmically). Similarly, if x is the root of the equation (3), we can check whether or not n = mx . If as a result of this checking, we get n = nx (respectively, n = mx ), then the conclusion will be that w′ (hence, also w) represents the trivial element in 6 Ȧn , otherwise, if n 6= nx (respectively, n 6= mx ), then the conclusion will be that w′ (hence, also w) does not represent the trivial element in Ȧn . Now, for all n ∈ N, define the group Bn as a torsion-free additive abelian group of rank 2, that is Bn = hbn,0 i ⊕ hbn,1 i such that B1 , B2 , . . . are disjoint. Now, for all n ∈ N, define the groups Ḃn as follows.  Bn if n ∈ / N ∪M , Ḃn = Bn / ≪ bn,1 = 2i bn,0 ≫ if n = ni ∈ N or n = mi ∈ M. For all n ∈ N, m ∈ {0, 1}, let us denote the images of bn,m under the natural homomorphism Bn → Ḃn by ḃn,m . It follows from the definitions of Ȧn and Ḃn that for all n ∈ N, these groups are infinite and torsion free. Lemma 5. There exists an algorithm such that for each input n ∈ N and w ∈ {±ḃn,0 , ±ḃn,1 }∗ , it decides whether or not w represents the trivial element in the group Ḃn . Proof. Follows from the repetition of arguments of the proof of Lemma 4. Lemma 6. The map ȧn,0 7→ ḃn,0 , ȧn,1 7→ ḃn,1 induces a group isomorphism between the groups hȧn,0 , ȧn,1 i = Ȧn and hḃn,0 , ḃn,1 i = Ḃn if and only if n ∈ N \ M. Proof. Indeed, in case n ∈ N , by definition, hȧn,0 , ȧn,1 i = hȧn,0 i and ȧn,1 = 2i ȧn,0 , where i is the index of n in N . Also hḃn,0 , ḃn,1 i = hḃn,0 i and ḃn,1 = 2i ḃn,0 . Therefore, in case n ∈ N , the map ȧn,0 7→ ḃn,0 , ȧn,1 7→ ḃn,1 induces a group isomorphism between the groups hȧn,0 , ȧn,1 i and hḃn,0 , ḃn,1 i. In case n ∈ N \ (N ∪ M), the groups Ȧn and Ḃn are torsion-free and abelian of rank 2 with generating sets {ȧn,0 , ȧn,1 } and {ḃn,0 , ḃn,1 }, respectively. Therefore, if n ∈ N \ (N ∪ M), the map ȧn,0 7→ ḃn,0 , ȧn,1 7→ ḃn,1 induces a group isomorphism between the groups hȧn,0 , ȧn,1 i and hḃn,0 , ḃn,1 i as well. Now suppose that n ∈ M. Then, hȧn,0 , ȧn,1 i = hȧn,0 i and hḃn,0 , ḃn,1 i = hḃn,0 i, however, by definition, ȧn,1 = 3i ȧn,0 while ḃn,1 = 2i ḃn,0 . Therefore, the map ȧn,0 7→ ḃn,0 , ȧn,1 7→ ḃn,1 does not induce a group isomorphism between the groups hȧn,0 , ȧn,1 i and hḃn,0 , ḃn,1 i when n ∈ M. 7 Now, let T = F (t1 , t2 , . . .) be a free group with countable free basis {t1 , t2 , . . .}. Denote the infinite free products Ȧ1 ∗ Ȧ1 ∗ . . . and Ḃ1 ∗ Ḃ1 ∗ . . . by ∗∞ n=1 Ȧn and ∗∞ Ḃ , respectively. Then define n=1 n ∞ Ġ = (∗∞ n=1 Ȧn ) ∗ (∗n=1 Ḃn ) ∗ T / ≪ R ≫, (4) where the set of defining relators R is defined as  R = t−1 i ȧni ,0 ti = ḃni ,0 | i ∈ N . Define ∞ Ġ0 = (∗∞ n=1 Ȧn ) ∗ (∗n=1 Ḃn ), and for all k ∈ N, define Ġk as ∞ Ġk = (∗∞ n=1 Ȧn ) ∗ (∗n=1 Ḃn ) ∗ F (t1 , . . . , tk )/ ≪ Rk ≫, where the set of defining relators Rk is defined as  Rk = t−1 i ȧni ,0 ti = ḃni ,0 | 1 ≤ i ≤ k . Then, clearly the group Ġ is the direct limit of the sequence of group {Ġk }∞ k=0 connected by homomorphisms ǫk : Ġk → Ġk+1  such that ǫk are the homomorphisms induced by the identity maps from ȧn,0 , ȧn,1 , ḃn,0 , ḃn,1 , ti | n ∈ N, i ∈ {1, 2, . . . , k} to themselfs for all k ∈ N. Let us denote  S0 = ± ȧn,m , ± ḃn,m | n ∈ N, m ∈ {0, 1} and for k ∈ N, Sk =  ±1 ± ȧn,m , ± ḃn,m , t±1 | n ∈ N, m ∈ {0, 1} . 1 , . . . , tk Note that since the sets N and M are recursively enumerable, the groups Ġ and Ġk have recursive presentations with respect to the generating sets S0 ∪ {t1 , t2 , . . .} and Sk , k ∈ N ∪ {0}, respectively. Lemma 7. There exists an algorithm such that for each input w ∈ S0∗ it decides whether or not w = 1 in Ġ0 . Moreover, there exists an algorithm such that for each input (w, i), w ∈ S0∗ , i ∈ N, it decides whether or not w represents an element from hȧni ,0 i, and in case it represents such an element, the algorithm returns λȧni ,0 , λ ∈ Z, such that w = λȧni ,0 in Ġ0 . Analogous statement remains true when we replace ȧni ,0 with ḃni ,0 . 8 Proof. Indeed, these properties immediately follow from the basic properties of the direct products of groups combined with Lemmas 4 and 5. Lemma 8. For all k ∈ N ∪ {0} and n ∈ N, the following holds. (i). The groups Ȧn and Ḃn embed into Ġk under the maps induced by ȧn,m 7→ ȧn,m and ḃn,m 7→ ḃn,m for m ∈ {0, 1}, respectivley; (ii). The group Ġk+1 is an HNN-extension of the group Ġk . More precisely, Ġk+1 = hĠk , tk+1 | t−1 k+1 ȧnk+1 ,0 tk+1 = ḃnk+1 ,0 i. Proof. Indeed, if k = 0, then (i) and (ii) are obvious. Now, let us apply induction with respect to k. Suppose that for all 0 ≤ l < k, the statements of (i) and (ii) are true. Then, since by the inductive assumption, Ġk is obtained from Ġk−1 as an HNNextension with respect to the isomorphic subgroups hȧnk ,0 i ⋍ hḃnk ,0 i, by the basic properties of HNN-extensions (see Lemma 2), we get that the statement of (i) holds for Ġk . Therefore, since the subgroups hȧnk+1 ,0 i ≤ Ġk and hḃnk+1 ,0 i ≤ Ġk are isomorphic, and in the definition of Ġk+1 the only defining relation which −1 involves the letters t±1 k+1 is the relation tk+1 ȧnk+1 ,0 tk+1 = ḃnk+1 ,0 , we get that the statement of (ii) holds as well. Corollary 1. If k < l, then the group Ġk embeds into the group Ġl under the map induced by ȧn,m 7→ ȧn,m , ḃn,m 7→ ḃn,m for n ∈ N and m ∈ {0, 1} and t1 7→ t1 , . . . , tk 7→ tk . Proof. Indeed, by Lemma 8, the group Ġl is obtained from the group Ġk by (multiple) HNN-extensions. Therefore, the statement follows from the basic properties of HNN-extensions, namely, by Lemma 2. Corollary 2. The map ȧn,0 7→ ḃn,0 , ȧn,1 7→ ḃn,1 induces a group isomorphism between the subgroups hȧn,0 , ȧn,1 i = Ȧn and hḃn,0 , ḃn,1 i = Ḃn of Ġ if and only if n ∈ N \ M. Proof. By Corollary 1, Ġ0 embeds in Ġ by the map induced by ȧn,0 7→ ȧn,0 , ȧn,1 7→ ȧn,1 , ḃn,0 7→ ḃn,0 , ḃn,1 7→ ḃn,1 for n ∈ N. Therefore, the statement of the corollary follows from Lemma 6. Definition 1 (Reduced words over Sk∗ ). Let k ∈ N. Then, for a given word w ∈ Sk∗ , we say that w is a reduced word over Sk∗ if the following properties hold. 9 (0). w is freely reduced, i.e. w does not contain subwords of the form xx−1 , x ∈ Sk ; (1). For all 1 ≤ i ≤ k, w does not contain subwords of the form t−1 i uti , where u ∈ S0∗ is such that u = λȧni ,0 in Ġ0 for some λ ∈ Z; (2). For all 1 ≤ i ≤ k, w does not contain subwords of the form ti vt−1 i , where v ∈ S0∗ is such that v = λḃni ,0 in Ġ0 for some λ ∈ Z. ∗ Lemma 9. For all k ∈ N, if w ∈ Sk∗ \ Sk−1 is a reduced word over Sk∗ , then ∗ w 6= 1 in Ġk . Moreover, w 6= u in Ġk for any word u ∈ Sk−1 . Proof. Let us prove by induction on k. If k = 1, then the group Ġ1 = hĠ0 , t1 | t−1 1 ȧn1 ,0 t1 = ḃn1 ,0 i is an HNN-extension of Ġ0 with respect to the isomorphic subgroups hȧn1 ,0 i ≤ Ġ0 and hḃn1 ,0 i ≤ Ġ0 . Therefore, by Britton’s Lemma (see Lemma 1), w 6= 1 in Ġ1 provided that it is a reduced word over S1∗ . Also for any u ∈ S0∗ , the word wu−1 is a reduced word with respect to the HNN-extension Ġ1 = hĠ0 , t1 | t−1 1 ȧn1 ,0 t1 = ḃn1 ,0 i. Therefore, by Britton’s Lemma (see Lemma 1), wu−1 6= 1 in Ġ1 or, in other words, w 6= u in Ġ1 . ∗ Now assume that k > 1 and w ∈ Sk∗ \ Sk−1 is a reduced word over Sk∗ . Also, suppose that the statement of the lemma is true for all l < k. Then, first of all, note that from the definition of the reduced words over Sk∗ it follows that if ∗ ∗ . , then v is a reduced word over Sk−1 v is a subword of w such that v ∈ Sk−1 −1 −1 Consequently, by the inductive hypothesis, if tk utk (or tk utk ) is a subword of ∗ w such that u ∈ Sk−1 and u represents an element from the image of Ȧnk (or Ḃnk ) in Ġk , then u ∈ S0∗ . However, this contradicts the assumption that w is a reduced word over Sk∗ . Therefore, since Ġk = hĠk−1 , tk | t−1 k ȧnk ,0 tk = ḃnk ,0 i is an HNN-extension of Ġk−1 with respect to the isomorphic subgroups hank ,0 i = Ȧnk ≤ Ġk−1 and hbnk ,0 i = Ḃnk ≤ Ġk−1 , we get that if w is a reduced word over Sk∗ , then w is a reduced word over this HNN-extension. Hence, by Britton’s Lemma, we get that w 6= 1 in Ġk . Similarly, for any u ∈ S0∗ , again by Britton’s Lemma, we get that wu−1 6= 1 in Ġk or, in other words, w 6= u in Ġk . Lemma 10. There exists an algorithm such that for each input (k, w), k ∈ N ∪ {0}, w ∈ Sk∗ , it decides whether or not w = 1 in Ġk . Proof. Let (k, w) be a fixed input. Without loss of generality assume that w is a freely reduced word in Sk∗ . If k = 0, then one can apply the word problem algorithm for the group Ġ0 = hS0∗ i. See Lemma 7. Otherwise, if k ≥ 1, for each k1 ≤ k such that w contains a letter from {tk1 , t−1 k1 }, do the following: Find all subwords of w which are of one of the forms −1 ∗ ut t−1 k1 or tk1 vtk1 , where u, v ∈ S0 and u = λȧnk1 ,0 , v = λḃnk1 ,0 in Ġ0 for some k1 λ ∈ Z. (By Lemma 7, subwords of these form can be found algorithmically.) Then, if, say, a subword of the form t−1 k1 utk1 is found, replace it with λḃnk1 ,0 . 10 Thanks to the identity t−1 k1 λȧnk1 ,0 tk1 = λḃnk1 ,0 , the newly obtained word is equal to w in Ġk . Then repeat this procedure on the newly obtained word until there is no more subwords of the mentioned forms. Let w1 be the word obtained as a result of this procedure. Then, by Lemma 9, either w1 ∈ S0∗ or for some k0 ≥ 1, w1 ∈ Sk∗0 \ Sk∗0 −1 . Then, in the last case, by Lemma 9, w1 is a reduced word over Sk∗0 . Also in the first case (i.e. when w1 ∈ S0∗ ), w1 = 1 in Ġk if and only if w1 = 1 in Ġ0 , hence by Lemma 7, in this case, the identity w1 = 1 can be checked algorithmically. In the second case, by Lemma 9, w1 6= 1 in Ġk . Lemma 11. The word problem in Ġ is decidable with respect to the presentation (4). Proof. Suppose that w is a finite word with letters from  ±1 Sk = ± ȧn,m , ± ḃn,m , t±1 | n ∈ N, m ∈ {0, 1} , 1 , . . . , tk where k is some natural number. Also suppose that w represents the trivial element in Ġ. Then, since Ġ is a direct limit of the groups {Ġi }∞ i=1 , there exists a minimal integer N ≥ 0 such that w represents the trivial element in ĠN . We claim that N ≤ k. Indeed, if N > k, then since N was chosen as the minimal index such that w = 1 in ĠN , we get w 6= 1 in Ġk . However, by Corollary 1, Ġk embeds into ĠN under the map induces by ȧn,m 7→ ȧn,m and t1 7→ t1 , . . . , tk 7→ tk , for n ∈ N, m ∈ {0, 1}, which implies that if w 6= 1 in Ġk , then w 6= 1 in ĠN . A contradiction. Thus, if w ∈ Sk∗ represents the trivial element in Ġ, then it represents the trivial element in Ġk as well. In other words, in order to check whether or not w represents the trivial element in Ġ it is enough to check its triviality in Ġk . Therefore, since for each w ∈ S ∗ one can algorithmically find (the minimal) k ∈ N such that w ∈ Sk∗ , the decidability of the word problem in Ġ follows from Lemma 10. Lemma 12. The group Ġ is torsion-free. Proof. First of all, notice that by the properties of the groups Ȧk , Ḃk , k ∈ N, and by the basic properties of direct products, the group Ġ0 is torsion free. Now, suppose that u ∈ S ∗ is such that it represents a torsion element of Ġ. Then, since Ġ is a direct limit of the groups {Ġi }∞ i=1 , there exists k ∈ N such that u ∈ Sk∗ and u represents a torsion element in Ġk as well. Since Ġk is obtained from Ġ0 by multiple HNN-extensions, then, by Lemma 3, Ġk is a torsion free group. Therefore, u represents the trivial element in Ġk as well as in Ġ. 11 Now suppose that Φ : Ġ ֒→ G̈ is an embedding of the group Ġ into a finitely generated torsion-free group G̈ such that the maps φ1 : (n, m) 7→ Φ(ȧn,m ), φ2 : (n, m) 7→ Φ(ḃn,m ), and φ3 : n 7→ Φ(tn ), where n ∈ N, m ∈ {0, 1}, are computable, and G̈ has decidable word problem. Then the next lemma shows that the group G̈ has the desirable properties we were looking for. Lemma 13. The group G̈ cannot be embedded in a group with decidable conjugacy problem. Proof. By contradiction, let us assume that G̈ embeds in a group Ḡ which has decidable conjugacy problem. Then, for the purpose of convenience, without loss of generality let us assume that G̈ is a subgroup of the group Ḡ. Below we show that the decidability of the conjugacy problem in Ḡ contradicts the assumption that N and M are disjoint and recursively inseparable. Let us define C ⊆ N as  C = n ∈ N | Φ(ȧn,0 ) is conjugate to Φ(ḃn,0 ) in Ḡ . Then, the decidability of the conjugacy problem in Ḡ implies that the set C is recursive, because, since the group Ḡ has decidable conjugacy problem, and since by our assumptions, the above mentioned maps φ1 , φ2 and φ3 are computable, for any input n ∈ N one can algorithmically verify whether or not Φ(ȧn,0 ) is conjugate to Φ(ḃn,0 ) in Ḡ. Therefore, since for groups with decidable conjugacy problem one can algorithmically find conjugator element for each pair of conjugate elements of the group, we also get that there exists a computable map f : C → Ḡ such that for all n ∈ C we have f (n)−1 Φ(ȧn,0 )f (n) = Φ(ḃn,0 ). For n ∈ C, let us denote f (n) = gn ∈ Ḡ. Now let us define  A = n ∈ C | gn−1 Φ(ȧn,1 )gn = Φ(ḃn,1 ) ⊆ N. 12 Since the word problem in Ḡ is decidable, the sets C is recursive and the maps Φ and f are computable, we get that the set A is a recursive subset of N. Also since the following identities ȧni ,1 = 2i ȧni ,0 , ḃni ,1 = 2i ḃni ,0 and ti−1 ȧni ,0 ti = ḃni ,0 , for i ∈ N, hold in Ġ, we get that in Ḡ the following identities hold i i Φ(ȧni ,1 ) = Φ(ȧni ,0 )2 , Φ(ḃni ,1 ) = Φ(ḃni ,0 )2 and Φ(ti )−1 Φ(ȧni ,0 )Φ(ti ) = Φ(ḃni ,0 ) for all ni ∈ N . Therefore, we get that N ⊆ A. On the other hand, Corollary 2 implies that for any n ∈ M, the pairs of elements

Φ(ȧn,0 ), Φ(ḃn,0 ) and Φ(ȧn,1 ), Φ(ḃn,1 ) cannot be conjugate in Ḡ by the same conjugator. Therefore, we get that A ∩ M = ∅. Thus we get that N ⊆ A and A ∩ M = ∅, which implies that A ⊂ N is a recursive separating set for N and M, which contradicts the assumption that N and M are recursively inseparable. Finally, the embedding Φ : Ġ ֒→ G̈ with the prescribed properties exists, thanks to Theorem 4. Therefore, the group G̈ with the above mentioned properties exists. Also by a version of Higman’s embedding theorem described by Aanderaa and Cohen in [1], the group G̈ can be embedded into a finitely presented group G with decidable word problem. By a recent result of Chiodo and Vyas, [4], the group G defined this way will also inherit the property of torsion-freeness from the group G̈. Clearly, since G̈ cannot be embedded into a group with decidable conjugacy problem, this property will be inherited by G. Thus Theorem 2 is proved. 13 References [1] S. Aanderaa, D. E. Cohen, Modular machines I, II, in [2], Stud. Logic Found. Math. 95 (1980), 1-18, 19-28. [2] G. Baumslag, A. Myasnikov, V. Shpilrain et al., Open problems in Combinatorial and Geometric Group Theory, http://www.grouptheory.info/. [3] W.W. Boone, F.B. Cannonito, and R.C. Lyndon, Word Problems: Decision Problems and the Burnside Problem in Group Theory, Studies in Logic and the Foundations of Mathematics, vol. 71, North-Holland, Amsterdam, 1973. [4] M. Chiodo, R. Vyas, Torsion length and finitely presented groups, arXiv:1604.03788, 2016. [5] Clapham C.R.J., An embedding theorem for finitely generated groups. Proc. London Math. Soc. (1967) 17:419-430. [6] D.J. Collins, Representation of Turing reducibility by word and conjugacy problems in finitely presented groups, Acta Math. 128 , (1972) no. 1-2,7390. [7] D.J. Collins, C.F. Miller III , The conjugacy problem and subgroups of finite index, Proc. London Math. Soc. (3) 34 (1977), no. 3, 535-556. [8] A. Darbinyan, Word and conjugacy problems in lacunary hyperbolic groups, arXiv:1708.04591. [9] A. Darbinyan, Group embeddings with algorithmic properties, Communications in Algebra, 43:11 (2015), 4923-4935. [10] A.V. Gorjaga, A.S. Kirkinski , The decidability of the conjugacy problem cannot be transferred to finite extensions of groups. (Russian) Algebra i Logika 14 (1975), no. 4, 393-406. [11] G. Higman. Subgroups of finitely presented groups. Proc. Roy. Soc. Ser. A, 262:455-475, 1961. [12] Kourovka Notebook. Unsolved Problems in Group Theory. 5th edition, Novosibirsk, 1976. [13] R.C. Lyndon, P.E. Schupp, Combinatorial group theory, Springer, Berlin, 1977. [14] A. Mal’cev, Constructive algebras. I, Uspehi Mat. Nauk, vol. 16 (1961), no. 3 (99), pp. 3-60. [15] A. Miasnikov, P. Schupp, Computational complexity and the conjugacy problem. Computability, 2016. 14 [16] C. F. Miller III. On group-theoretic decision problems and their classification, volume 68 of Annals of Mathematics Studies. Princeton University Press, 1971. [17] C.F. Miller III, Decision Problems for Groups Survey and Reflections. Algorithms and Classification in Combinatorial Group Theory (1992) 23:159. [18] A.Yu. Olshanskii, M. Sapir, The conjugacy problem and Higman embeddings. Mem. Amer. Math. Soc. 170 (2004), no. 804 [19] A. Yu. Olshanskii, M.V. Sapir, Subgroups of finitely presented groups with solvable conjugacy problem. Internat. J. Algebra Comput., 15(5-6):10751084, 2005. [20] D. Osin, Small cancellations over relatively hyperbolic groups and embedding theorems, Ann. Math. 172 (2010), no. 1, 1-39. [21] M. Rabin, Computable algebra, general theory and theory of computable fields., Trans. Amer. Math. Soc., vol. 95 (1960), pp. 341-360. [22] J. R. Shoenfield, Mathematical logic. Addison Wesley, 1967. A. Darbinyan, Department of Mathematics, Vanderbilt University, 1326 Stevenson Center, Nashville, TN 37240 E-mail address: arman.darbinyan@vanderbilt.edu 15 

arXiv:1510.04156v1 [math.AC] 14 Oct 2015 A CHANGE OF RINGS RESULT FOR MATLIS REFLEXIVITY DOUGLAS J. DAILEY AND THOMAS MARLEY Abstract. Let R be a commutative Noetherian ring and E the minimal injective cogenerator of the category of R-modules. An R-module M is (Matlis) reflexive if the natural evaluation map M −→ HomR (HomR (M, E), E) is an isomorphism. We prove that if S is a multiplicatively closed subset of R and M is a reflexive R-module, then M is a reflexive RS -module. The converse holds when S is the complement of the union of finitely many nonminimal primes of R, but fails in general. 1. Introduction Let R be a commutative Noetherian L ring and E the minimal injective cogenerator of the category of R-modules; i.e., E = m∈Λ ER (R/m), where Λ denotes the set of maximal ideals of R and ER (−) denotes the injective hull. An R-module M is said to be (Matlis) reflexive if the natural evaluation map M −→ HomR (HomR (M, E), E) is an isomorphism. In [1], the authors assert the following “change of rings” principal for Matlis reflexivity ([1, Lemma 2]): Let S be a multiplicatively closed subset of R and suppose M is an RS -module. Then M is reflexive as an R-module if and only if M is reflexive as an RS -module. However, the proof given in [1] is incorrect (see Examples 3.1-3.3) and in fact the “if” part is false in general (cf. Proposition 3.4). In this note, we prove the following: Theorem 1.1. Let R be a Noetherian ring, S a multiplicatively closed subset of R, and M an RS -module. (a) If M is reflexive as an R-module then M is reflexive as an RS -module. (b) If S = R \ (p1 ∪ . . . ∪ pr ) where each pi is a maximal ideal or a nonminimal prime ideal, then the converse to (a) holds. 2. Main results Throughout this section R will denote a Noetherian ring and S a multiplicatively closed set of R. We let ER (or just E if the ring is clear) denote the minimal injective cogenerator of the category of R-modules as defined in the introduction. A semilocal ring is said to be complete if it is complete with respect to the J-adic topology, where J is the Jacobson radical. We will make use of the main result of [1]: Date: October 14, 2015. 2010 Mathematics Subject Classification. Primary 13C05; Secondary 13C13. Key words and phrases. Matlis reflexive, minimal injective cogenerator. The first author was partially supported by U.S. Department of Education grant P00A120068 (GAANN). 1 2 DOUGLAS J. DAILEY AND THOMAS MARLEY Theorem 2.1. ([1, Theorem 12]) Let R be a Noetherian ring, M an R-module, and I = AnnR M . Then M is reflexive if and only if R/I is a complete semilocal ring and there exists a finitely generated submodule N of M such that M/N is Artinian. We remark that the validity of this theorem does not depend on [1, Lemma 2], as the proof of [1, Theorem 12] uses this lemma only in a special case where it is easily seen to hold. (See the proof of [1, Theorem 9], which is the only instance [1, Lemma 2] is used critically.) Lemma 2.2. ([1, Lemma 1] Let M be an R-module and I an ideal of R such that IM = 0. Then M is reflexive as an R-module if and only if M is reflexive as an R/I-module. Proof. Since ER/I = HomR (R/I, ER ), the result follows readily by Hom-tensor adjunction.  Lemma 2.3. Let R = R1 × · · · × Rk be a product of Noetherian local rings. Let M = M1 × · · · × Mk be an R-module. Then M is reflexive as an R-module if and only if Mi is reflexive as an Ri -module for all i. Proof. Let ρi : R−→Ri be the canonical projections for i = 1, . . . , k. Let ni be the maximal ideal of Ri and mi = ρ−1 i (ni ) the corresponding maximal ideal of R. Then mi Ri = ni and mi Rj = Rj for all j 6= i. Note that Rmi ∼ = Ri and Ei := ER (R/mi ) ∼ = ERi (Ri /ni ) for all i. Then ER = E1 ⊕ · · · ⊕ Ek . It is easily seen that HomR (HomR (M, ER ), ER ) ∼ = k M HomRi (HomRi (Mi , Ei ), Ei ), i=1 and that this isomorphism commutes with the evaluation maps. The result now follows.  Theorem 2.4. Let S be a multiplicatively closed set of R and M an RS -module which is reflexive as an R-module. Then M is reflexive as an RS -module. Proof. By Lemma 2.2, we may assume AnnR M = AnnRS M = 0. Thus, R is semilocal and complete by Theorem 2.1. Hence, R = R1 × · · · × Rk where each Ri is a complete local ring. Then RS = (R1 )S1 × · · · × (Rk )Sk where Si is the image of S under the canonical projection R−→Ri . Write M = M1 × · · · × Mk , where Mi = Ri M . As M is reflexive as an R-module, Mi is reflexive as an Ri -module for all i. Thus, it suffices to show that Mi is reflexive as an (Ri )Si -module for all i. Hence, we may reduce the proof to the case (R, m) is a complete local ring with AnnR M = 0 by passing to R/ AnnR M , if necessary. As M is reflexive as an R-module, we have by Theorem 2.1 that there exists an exact sequence 0−→N −→M −→X−→0 where N is a finitely generated R-module and X is an Artinian R-module. If S ∩ m = ∅, the RS = R and there is nothing to prove. Otherwise, as SuppR X ⊆ {m}, we have XS = 0. Hence, M ∼ = NS , a finitely generated RS -module. To see that M is RS -reflexive, it suffices to show that RS is Artinian (hence semilocal and complete). Since AnnR NS = AnnR M = 0, we have that AnnR N = 0. Thus, dim R = dim N . Since M is an RS -module and S ∩m 6= ∅, i (M ) ∼ H i i (X) = 0 for i ≥ 1. we have Hm = mRS (M ) = 0 for all i. Further, as X is Artinian, Hm i (N ) = 0 for Thus, from the long exact sequence on local cohomology, we conclude that Hm A CHANGE OF RINGS RESULT FOR MATLIS REFLEXIVITY 3 i ≥ 2. Thus, dim R = dim N ≤ 1, and hence, dim RS = 0. Consequently, RS is Artinian, and M is a reflexive RS -module.  To prove part (b) of Theorem 1.1, we will need the following result on Henselian local rings found in [2] (in which the authors credit it to F. Schmidt). As we need a slightly different version of this result than what is stated in [2] and the proof is short, we include it for the convenience of the reader: Proposition 2.5. ([2, Satz 2.3.11]) Let (R, m) be a local Henselian domain which is not a field and F the field of fractions of R. Let V be a discrete valuation ring with field of fractions F . Then R ⊆ V . Proof. Let k be the residue field of R and a ∈ m. As R is Henselian, for every positive integer n not divisible by the characteristic of k, the polynomial xn − (1 + a) has a root b in R. Let v be the valuation on F associated to V . Then nv(b) = v(1 + a). If v(a) < 0 then v(1 + a) < 0 which implies v(b) ≤ −1. Hence, v(1 + a) ≤ −n. As n can be arbitrarily large, this leads to a contradiction. Hence, v(a) ≥ 0 and a ∈ V . Thus, m ⊆ V . Now let c ∈ R be arbitrary. Choose d ∈ m, d 6= 0. If v(c) < 0 then v(cℓ d) < 0 for ℓ sufficiently large. But this contradicts that cℓ d ∈ m ⊆ V for every ℓ. Hence v(c) ≥ 0 and R ⊆ V .  For a Noetherian ring R, let Min R and Max R denote the set of minimal and maximal primes of R, respectively. Let T(R) = (Spec R \ Min R) ∪ Max R. Lemma 2.6. Let R be a Noetherian ring and p ∈ T(R). If Rp is Henselian then the natural map ϕ : R−→Rp is surjective; i.e., R/ ker ϕ ∼ = Rp . Proof. By replacing R with R/ ker ϕ, we may assume ϕ is injective. Then p contains every minimal prime of R. Let u ∈ R, u 6∈ p. It suffices to prove that the image of u in R/q is a unit for every minimal prime q of R. Hence, we may assume that R is a domain. (Note that (R/q)p = Rp /qRp is still Henselian.) If Rp is a field, then, as p ∈ T(R), we must have R is a field (as p must be both minimal and maximal in a domain). So certainly u 6∈ p = (0) is a unit in R. Thus, we may assume Rp is not a field. Suppose u is not a unit in R. Then u ∈ n for some maximal ideal n of R. Now, there exists a discrete valuation ring V with same field of fractions as R such that mV ∩ R = n ([5, Theorem 6.3.3]). As Rp is Henselian, Rp ⊆ V by Proposition 2.5. But as u ∈ / p, u is a unit in Rp , hence in V , contradicting u ∈ n ⊆ mV . Thus, u is a unit in R and R = Rp .  Proposition 2.7. Let R be a Noetherian ring and S = R \ (p1 ∪ · · · ∪ pr ) where p1 , . . . , pr ∈ T(R). Suppose RS is complete with respect to its Jacobson radical. Then the natural map ϕ : R−→RS is surjective. S Proof. First, we may assume that pj * i6=j pi for all j. Also, by passing to the ring R/ ker ϕ, we may assume ϕ is injective. (We note that if pi1 , . . . , pit are the ideals in the set {p1 , . . . , pr } containing ker ϕ, it is easily seen that (R/ ker ϕ)S = (R/ ker ϕ)T where T = R\(pi1 ∪· · ·∪pit ). Hence, we may assume each pi contains ker ϕ.) As RS is semilocal and complete, the map ψ : RS −→Rp1 × · · · × Rpr given by ψ(u) = ( u1 , . . . , u1 ) is an isomorphism. For each i, let ρi : R−→Rpi be the natural map. Since R−→RS is an injection, ∩i ker ρi = (0). It suffices to prove that u is a unit in R for every u ∈ S. As Rpi is complete, hence Henselian, we have that ρi is surjective for each i by Lemma 2.6. Thus, u is a unit in 4 DOUGLAS J. DAILEY AND THOMAS MARLEY R/ ker ρi for every i; i.e., (u) + ker ρi = R for i = 1, . . . , r. Then (u) = (u) ∩ (∩i ker ρi ) = R. Hence, u is a unit in R.  We now prove part (b) of the Theorem 1.1: Theorem 2.8. Let R be a Noetherian ring and M a reflexive RS -module, where S is the complement in R of the union of finitely many elements of T(R). Then M is reflexive as an R-module. Proof. We may assume M 6= 0. Let S = R \ (p1 ∪ · · · ∪ pr ), where p1 , . . . , pr ∈ T(R) Let I = AnnR M , whence IS = AnnRS M . As in the proof of Proposition 2.7, we may assume each pi contains I. Then by Lemma 2.2, we may reduce to the case AnnR M = AnnRS M = 0. Note that this implies the natural map R−→RS is injective. As M is RS -reflexive, RS is complete with respect to its Jacobson radical by Theorem 2.1. By Proposition 2.7, we have that R ∼  = RS and hence M is R-reflexive. 3. Examples The following examples show that HomR (RS , ER ) need not be the minimal injective cogenerator for the category of RS -modules, contrary to what is stated in the proof of [1, Lemma 2]: Example 3.1. Let (R, m) be a local ring of dimension at least two and p any prime which is not maximal or minimal. By [3, Lemma 4.1], every element of Spec Rp is an associated prime of the Rp -module HomR (Rp , ER ). In particular, HomR (Rp , ER ) ∼ 6 ERp . = Example 3.2. ([3, p. 127]) Let R be a local domain such that the completion of R has a nonminimal prime contracting to (0) in R. Let Q be the field of fractions of R. Then HomR (Q, ER ) is not Artinian. Example 3.3. Let R be a Noetherian domain which is not local. Let m 6= n be maximal ideals of R. By a slight modification of the proof of [3, Lemma 4.1], one obtains that (0) is an associated prime of HomR (Rm , ER (R/n)), which is a direct summand of HomR (Rm , ER ). Hence, HomR (Rm , ER ) 6∼ = ERm . We now show that the converse to part (a) of Theorem 1.1 does not hold in general. Let R be a domain and Q its field of fractions. Of course, Q is reflexive as a Q = R(0) -module. But as the following theorem shows, Q is rarely a reflexive R-module. Proposition 3.4. Let R be a Noetherian domain and Q the field of fractions of R. Then Q is a reflexive R-module if and only if R is a complete local domain of dimension at most one. Proof. We first suppose R is a one-dimensional complete local domain with maximal ideal m. Let E = ER (R/m). By [4, Theorem 2.5], HomR (Q, E) ∼ = Q. Since the evaluation map of the Matlis double dual is always injective, we obtain that Q−→ HomR (HomR (Q, E), E) is an isomorphism. Conversely, suppose Q is a reflexive R-module. By Theorem 2.1, R is a complete semilocal domain, hence local. It suffices to prove that dim R ≤ 1. Again by Theorem 2.1, there exists a finitely generated R-submodule N of Q such that Q/N is Artinian. Since AnnR N = 0, i (N ) = 0 for i ≥ 2. But this follows dim R = dim N . Thus, it suffices to prove that Hm A CHANGE OF RINGS RESULT FOR MATLIS REFLEXIVITY 5 i (Q) = 0 for all i and H i (Q/N ) = 0 for i ≥ 1 (as Q/N is readily from the facts that Hm m Artinian).  Acknowledgments: The authors would like to thank Peder Thompson for many helpful discussions on this topic. They are also very grateful to Bill Heinzer for pointing out the existence of Proposition 2.5. References 1. R. Belshoff, E. Enochs, and J. Garcı́a-Rozas, Generalized Matlis duality. Proc. Amer. Math. Soc. 128 (1999), no. 5, 1307-1312. 2. R. Berger, R. Kiehl, E. Kunz, and H.-J. Nastold, Differentialrechnung in der analytischen Geometrie. Lecture Notes in Mathematics 38, Springer-Verlag, Berlin-New York, 1967. 3. L. Melkersson and P. Schenzel, The co-localization of an Artinian module. Proc. Edinburgh Math. Soc. 38 (1995), 121–131. 4. P. Schenzel, A note on the Matlis dual of a certain injective hull. J. Pure Appl. Algebra 219 (2015), no. 3, 666-671. 5. I. Swanson and C. Huneke, Integral closures of ideals, rings and modules. London Mathematical Society Lecture Note Series 336, Cambridge University Press, Cambridge, 2006. Department of Mathematics, University of Nebraska-Lincoln, Lincoln, NE 68588-0130 E-mail address: ddailey2@math.unl.edu Department of Mathematics, University of Nebraska-Lincoln, Lincoln, NE 68588-0130 E-mail address: tmarley1@unl.edu 

JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 1 CUDAMPF++: A Proactive Resource Exhaustion Scheme for Accelerating Homologous Sequence Search on CUDA-enabled GPU arXiv:1707.09683v1 [cs.CE] 30 Jul 2017 Hanyu Jiang, Student Member, IEEE, Narayan Ganesan, Senior Member, IEEE, and Yu-Dong Yao, Fellow, IEEE Abstract—Genomic sequence alignment is an important research topic in bioinformatics and continues to attract significant efforts. As genomic data grow exponentially, however, most of alignment methods face challenges due to their huge computational costs. HMMER, a suite of bioinformatics tools, is widely used for the analysis of homologous protein and nucleotide sequences with high sensitivity, based on profile hidden Markov models (HMMs). Its latest version, HMMER3, introdues a heuristic pipeline to accelerate the alignment process, which is carried out on central processing units (CPUs) with the support of streaming SIMD extensions (SSE) instructions. Few acceleration results have since been reported based on HMMER3. In this paper, we propose a five-tiered parallel framework, CUDAMPF++, to accelerate the most computationally intensive stages of HMMER3’s pipeline, multiple/single segment Viterbi (MSV/SSV), on a single graphics processing unit (GPU). As an architecture-aware design, the proposed framework aims to fully utilize hardware resources via exploiting finer-grained parallelism (multi-sequence alignment) compared with its predecessor (CUDAMPF). In addition, we propose a novel method that proactively sacrifices L1 Cache Hit Ratio (CHR) to get improved performance and scalability in return. A comprehensive evaluation shows that the proposed framework outperfroms all existig work and exhibits good consistency in performance regardless of the variation of query models or protein sequence datasets. For MSV (SSV) kernels, the peak performance of the CUDAMPF++ is 283.9 (471.7) GCUPS on a single K40 GPU, and impressive speedups ranging from 1.x (1.7x) to 168.3x (160.7x) are achieved over the CPU-based implementation (16 cores, 32 threads). Index Terms—GPU, CUDA, SIMD, L1 cache, hidden Markov model, HMMER, MSV, SSV, Viterbi algorithm. F 1 I NTRODUCTION T YPICAL algorithms and applications in bioinformatics, computational biology and system biology share a common trait that they are computationally challenging and demand more computing power due to the rapid growth of genomic data and the need for high fidelity simulations. As one of the most important branches, the genomic sequence analysis with various alignment methods scales the abstraction level from atoms to RNA/DNA molecules and even whole genomes, which aims to interpret the similarity and detect homologous domains amongst sequences [1]. For example, the protein motif detection is key to identify conserved protein domains within a known family of proteins. This paper addresses HMMER [2], [3], a widely used toolset designed for the analysis of homologous protein and nucleotide sequences with high sensitivity, which is carried out on central processing units (CPUs) originally. HMMER is built on the basis of probabilistic inference methods with profile hidden Markov models (HMMs) [3]. Particularly, the profile HMM used in HMMER is Plan7 architecture that consists of five main states (Match(M), Insert(I), Delete(D), Begin(B) and End(E)) as well as five special states (N, C, J, S and T). The M, I and D states which are in the same position form a node, and the number of • H. Jiang, N. Ganesan, Y. Yao are with the Department of Electrical and Computer Engineering, Stevens Institute of Technology, Hoboken, NJ 07030. E-mail: {hjiang5, nganesan, Yu-Dong.Yao}@stevens.edu. Manuscript received xxxx xx, xxxx; revised xxxx xx, xxxx. nodes included in a profile HMM indicates its length. The digital number “7” in Plan-7 refers to the total of seven transitions per node, which exist in the architecture and each has a transition probability. In addtion, some states also have emission probabilities. This architecute is a little bit different from the original one proposed by Krogh et al. [4] which contains extra I -D and D-I transitions. The profile HMMs employ position-specific Insert or Delete probabilities rather than gap penalties, which enables HMMER to outperform BLAST [5] on senstivity [3]. However, the previous version of HMMER, HMMER2, suffers the computational expense and gains less utilization than BLAST. Due to well-designed heuristics, BLAST is in the order of 100x to 1000x faster than HMMER2 [3]. Therefore, numerous acceleration efforts have been made for HMMER2, such as [6], [7], [8], [9], [10]. Most of them employ application accelerators and co-processors, like fieldprogrammable gate array (FPGA), graphics processing unit (GPU) and other parallel infrastructures, which provide good performence improvement. To popularize HMMER for standard commodity processors, Eddy et al. propose new versions of HMMER, HMMER3 (v3.0) and its subsequent version (v3.1), which achieve the comparable performance as BLAST [3]. As the main contribution, HMMER3 implements a heuristic pipeline in hmmsearch which aligns a query model with the whole sequence dataset to find out significantly similar sequence matches. The heuristic acceleration pipeline is highly optimized on CPU-based systems with the support of streaming SIMD extensions (SSE) instructions, JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 and hence only few acceleration attempts, including [11], [12], [13], [14], [15], [16] and [17], report further speedups. Our previous work [18], CUDAMPF, proposes a multitiered parallel framework to accelerate HMMER3 pipeline on a single GPU, which was shown to exceed the current state-of-the-art. However, the performance evaluation shows that the thoughput of the computational kernel depends on the model length, especially for small models, which implies underutilization of the GPU. This inspires us to exploit finer-grained parallelism, compared with the framework presented in [18]. In this paper, we describe another tier of parallelization that aims to fully take advantage of the hardware resources provided by single GPU. A novel optimization strategy that proactively utilizes onchip cache system is proposed to further boosts kernel throughput and improves scalability of the framework. A comprehensive evaluation indicates that our method exhibits good consistency of performance regardless of query models and protein sequence datasets. The generalization of the proposed framework as well as performance-oriented suggestions are also discussed. The rest of the paper is organized as follows. Section 2 presents background of HMMER3 pipeline, GPU architecture and highlighted CUDA features, followed by a review of CUDAMPF implementation. In Section 3, the in-depth description of our proposed framework is presented. Then, we give comprehensive evaluations and analyses in Section 4. Related works and discussions are presented in Section 5 and 6, respectively. Finally, we present the conclusion of this paper. 2 BACKGROUND In this section, we go through the new heuristic pipeline of HMMER3 and highlights its computationally intensive stages. An overview of the GPU architecture and the CUDA programming model is also presented. For better understanding of subsequent ideas, we briefly review our previous work, CUDAMPF, at end of this section. 2.1 Heuristic Pipeline in HMMER3 The main contribution that accelerates the HMMER3 is a new algorithm, multiple segment Viterbi (MSV) [3], which is derived from the standard Viterbi algoritm. The MSV model is a kind of ungapped local alignment model with multiple hits, as shown in Fig. 1, and it is achieved by pruning Delete and Insert states as well as their transitions in the original profile HMMs. The M -M transitions are also treated as constants of 1. In addition to the MSV algorithm, another simpler algorithm, single segment Viterbi (SSV), is also introduced to boost the overall performance further. Given that the J state is the bridge between two matched alignments, the SSV model assumes that there is rarely a matched alignment with a score that is higher than the cost of going through the J state, and hence it speculatively removes the J state to gain a significant speedup [19]. However, in order to avoid false negatives, the SSV model is followed by regular MSV processing to re-calculate suspected sequences. Fig. 1 illustrates profiles of P7Viterbi, MSV and SSV models with an example of 4 nodes. The solid 2 arrows indicate transitions between different types of states whereas dashed arrows represent the self-increase of a state. I1 S N B M1 I2 I3 M2 M3 M4 D2 D3 D4 E C T E C T E C T J P7Viterbi: Plan-7 architecture based Viterbi S N B M1 M2 M3 M4 J MSV: Multiple ungapped local alignment Segment Viterbi S N B M1 M2 M3 M4 SSV: Single ungapped local alignment Segment Viterbi Fig. 1. Profiles of P7Viterbi, MSV and SSV models. In the pipeline, SSV and MSV models work as heuristic filters (stages) that filter out nonhomologous sequences. All sequences are scored during SSV and MSV stages, and only about 2.2% of sequences are passed to the next stage, given a threshold. The second stage consists of the P7Viterbi model which only allows roughly 0.1% of sequences pass, and resulting sequences are then scored with the the full Forward algorithm [3]. These four stages mentioned above form the main part of HMMER3’s pipeline. However, SSV and MSV stages consumes more than 70% of the overall execution time [17], [18], and hence they are prime targets of optimization. Fig. 2 illustrates the dynamic programming (DP) matrix of the P7Viterbi stage, corresponding to Fig. 1. A lattice of the middle region contains three scores for Match, Insert and Delete states, respectively, whereas flanking lattices only have one. To complete the alignment of a sequence, we need to calculate every lattice of the DP matrix starting from the left-top corner to the right-bottom corner (green) in a row-by-row order, which is a computationally intensive process. In the middle region of the DP matix, each lattice (red) depends on four cells (blue) directly, denoted by solid arrows, which can be formulated as: JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 N B Query model (states) 1 j M 3 E J C Sequence (residues) 0 i L Fig. 2. Dynamic programming matrix of the P7Viterbi stage.   VM [i − 1, j − 1] + T (Mj−1 , Mj ),  V [i − 1, j − 1] + T (I I j−1 , Mj ), VM [i, j] = M + max  VD [i − 1, j − 1] + T (Dj−1 , Mj ),    B[i − 1] + T (B, Mj ) ( VM [i − 1, j] + T (Mj , Ij ), VI [i, j] = I + max VI [i − 1, j] + T (Ij , Ij ) ( VM [i, j − 1] + T (Mj−1 , Dj ), VD [i, j] = max VD [i, j − 1] + T (Dj−1 , Dj ) 2.3 (1) where  denotes emission scores. V and T represent scores of M /I /D states and transitions, respectively. As for MSV and SSV stages, the mathematical formula can be simplified via removing VI and VD , which results in moderate dependencies and fewer amount of computation than the P7Viterbi stage. However, in order to exceed the performance of the highly optimized CPU-based implementation with the SIMD vector parallelization, it is imperative to go beyond general methods and exploit more parallelism on other multi/many-core processor architectures. 2.2 datapath from the existing L1 and shared memory datapath, and the maximum amount of available registers for each thread is increased to 255 instead of prior 63 per thread. Moreover, a set of Shuffle instructions that enables a warp of threads to share data without going through shared memory are also introduced in Kepler architecture. This new feature is heavily used in our proposed framework. The CUDA programming model is designed for NVIDIA GPUs, and it provides users with a development environment to easily leverage horsepower of GPUs. In CUDA, a kernel is usually defined as a function that is executed by all CUDA threads concurrently. Both grid and block are vitural units that form a thread hierarchy with some restrictions. Although CUDA allows users to launch thousands of threads, only a warp of threads (32 threads, currently) guarantee that they advance exectuions in lockstep, which is scheduled by a warp scheduler. Hence, the full efficiency is achieved only if all threads within a warp have the same execution path. Traditionally, in order to make sure that threads keep the same pace, a barrier synchronization has to be called explicitly, which imposes additional overhead. GPU Architecture and CUDA Programming Model As parallel computing engines, CUDA-enabled GPUs are built around a scalable array of multi-threaded streaming multiprocessors (SMs) for large-scale data and task parallelism, which are capable of executing thousands of threads in the single-instruction multiple-thread (SIMT) pattern [20]. Each generation of GPU introduces more hardware resources and new features, which aims to deal with the everincreasing demand for computing power in both industry and academia. In this paper, we implement our design on Tesla K40 GPU of Kepler GK110 architecture which equips with 15 powerful streaming multiprocessors, also known as SMXs. Each SMX consists of 192 single-precision CUDA cores, 64 double-precision units, 32 special function units and load/store units [21]. The architecture offers another 48KB on-chip read-only texture cache with an independent CUDAMPF In [18], we proposed a four-tiered parallel framework, CUDAMPF, implemented on single GPU to accelerate SSV, MSV and P7Viterbi stages of hmmsearch pipeline. The framework describes a hierarchical method that parallelizes algorithms and distributes the computational workload considering available hardware resources. CUDAMPF is completely warp-based that regards each resident warp as a compute unit to handle the exclusive workload, and hence the explict thread-synchronization is eliminated. Instead, the built-in warp-synchronism is fully utilized. A warp of threads make the alignment of one protein sequence one time and then pick up next scheduled sequence. Given that 8-bit or 16bit values are sufficient to the precision of algorithms, we couple SIMT execution mechanism with SIMD video instructions to achieve 64 and 128-fold parallelism within each warp. In addition, the runtime compilation (NVRTC), first appeared in CUDA v7.0, was also incorporated into the framework, which enabled swichable kernels and innermost loop unrolling to boost the performance further. CUDAMPF yields upto 440, 277 and 14.3 GCUPS (giga cells updates per second) with strong scalability for SSV, MSV and P7Viterbi kernels, respectively, and it has been proved to exceed all existing work. 3 P ROPOSED F RAMEWORK : CUDAMPF++ This section presents detailed implementations of the proposed framework, CUDAMPF++, that is designed to gain more parallelism based on CUDAMPF. We first introduce a new tier of parallelism followed by a data reformatting scheme for protein sequence data, and then in-depth explanations of kernel design are presented. Finally, we discuss the optimizations of the proposed framework. 3.1 Five-tiered Parallelism In CUDAMPF, the four-tiered parallel framework is proposed to implement MSV, SSV and P7Viterbi kernels. Although the performance improvement is observed on all JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 accelerated kernels, the speedup on P7Viterbi kernel is very limited whereas MSV/SSV kernel yields significant improvement. Given the profiling information [18], we are able to gain additional insights into the behaviors: (a) L1 Cache Hit Ratio (CHR) of the P7Viterbi kernel degrades rapidly as model size increases, and (b) its register usage always exceed the maximum pre-allocation for each thread, which indicates the exhaustion of on-chip memory resources and serious register spill. As for MSV/SSV kernels, however, the on-chip memory resources are sufficient. A large amount of low-latency registers, especially when aligning with small models, are underutilized. This can also be proved by performance curves in [18] in which only upward slopes are observed without any flat or downward trends as model size increases from 100 to 2405. The underutilization leaves an opportunity to exploit further parallelism that can fully take advantage of hardware resources on GPUs. In addition to original four-tiered structure, another tier of parallelism, inserted between 3rd and 4th tiers, is proposed to enable each warp handle multiple alignments with different protein sequences in parallel while the design of the CUDAMPF only allow single-sequence aligment per warp. This scheme aims to exhaust on-chip memory resources, regardless of the model length, to maximize the throughput of MSV/SSV kernels. Fig. 3 illustrates the fivetiered parallel framework. The first tier is based on multiple SMXs that possess plenty of computing and memory resources individually. Given the inexistence of data and execution dependency between each protein sequence, it is straightforward to partition whole sequence database into several chunks and distribute them to SMXs, as the most basic data parallelism. Tier 2 describes the parallelism between multiple warps that reside on each SMX. Since our implementation still applies the warp-synchronous execution that all warps are assigned to process different sequences without interwarp interactions, explicit synchronizations are eliminated completely. Unlike the CUDAMPF in which warps move to their next scheduled task once the sequence at hand is done, the current design allocates a sequence data block to each warp in advance. A data block may contain thousands of sequences, less or more, depending on the total size of protein sequence dataset and available warps. For multi-sequence alignment, all sequences within each data block need to be reformatted as the striped layout, which enables coalesced access to the global memory. Details of data reformatting will be discussed in Sec. 3.2. The number of resident warps per SMX is still fixed to 32 due to the complexity of MSV and SSV algorithms, which ensures that every thread obtains enough registers to handle complex execution dependencies and avoids excessive register spill. Besides, each SMX contains 4 warp schedulers, each with dual instruction dispath units [21], and hence it is able to dispatch upto 8 independent instructions each cycle. Those hardware resources have the critical influence on the parallelism of tier 2 and the performance of warp-based CUDA kernels. Tier 3 is built on the basis of warps. A warp of threads update different model states simultaneously and iterate over remaining model states in batches. The number of iterations depends on the query model size. Once the alignment 4 SMX SMX SMX Block Shared Memory L1 Cache Block Readonly Cache Shared Memory L1 Cache Block Readonly Cache Shared Memory Readonly Cache L1 Cache L2 Cache Global Memory Tier 1 sequence data blocks Warp Scheduler Block 2 Instruction Dispatch Units Warp Warp Scheduler Warp 2 Instruction Dispatch Units Warp Warp Scheduler Warp 2 Instruction Dispatch Units Tier 2 32 threads in one warp iterate through all states of the query model Seq. A R striped layout of query model states Tier 3 32 threads in one warp Group 1 L Group 2 Seq. 1 A R Group S Seq. 2 alignment A C C L H striped model states Seq. S V V L D S R H D W L H E W W Tier 4 MSV / SSV (4-lane per thread) Thread 1 P7Viterbi (2-lane per thread) Thread 2 Thread 1 8 bits 32-bit register Thread 2 16 bits 32-bit register Tier 5 Fig. 3. Five-tiered parallel framework based on the NVIDIA Kepler architecture. of an amino-acid residue is done, such as the ’A’ and its alignment scores (marked as a row of blue lattices) shown in Fig. 3 (Tier 3), the warp moves to next residue ’R’ and start over the alignment until the end of current sequence. On the basis of tier 3, tier 4 illustrates the multi-sequence JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 alignment that a warp of threads are evenly partitioned into several groups, each has L threads, to make alignments with S different sequences in parallel. For example, the first residues of S sequences, like ’A’ of Seq. 1, ’V’ of Seq. 2 and ’V’ of Seq. S, are extracted together for the firstround alignment. Each group of threads update W scores per iteration, and H iterations are required to finish one alignment. The model states are formatted as a rectangle with W ×H lattices. Considering two models, a large model Ml and a small model Ms where the size of Ml is S times larger than the size of Ms , we are able to get Ws = S1 Wl given Hs = Hl , which provides S -lane parallelism and roughly keeps register utilization of Ms as same as Ml . The tier 5 remains unchanged as the fine-grained data parallelism: every thread can operate on four 8-bit values and two 16-bit values simutaneously, using single SIMD video instruction [22], for MSV/SSV and P7Viterbi algorithms, respectively. With the support of tier 5, the parallelism of tier 4 is further extended because each thread takes charge of four different sequences at most. The value of S , as the number of sequences processed in parallel by a warp, is defined as below: ŝŵr {S | S = 2i , i ∈ Z ∩ [1, log2 ]}, (2) ŵv where ŝ is the warp size, ŵr and ŵv represent the width of registers and participant values, respectively. With Eq. 2, the rest of values can be also formulated as: ŝŵr , W = ŵv S ŝ (3) L=d e S m̂ H = max{2, d e}, W where m̂ represents the size of query model. W , L and H can be regarded as functions of S . 3.2 Warp-based Sequence Data Blocks Due to the introduction of the multi-sequence alignment, loading sequence data in the sequential layout is highly inefficient. Originally, in [18], residues of each sequence are stored in contiguous memory space, and warps always read 128 residues of one sequence by one coalesced global memory transaction. As for current design, however, the sequential data layout may lead to 128 transactions per memory request while extracting residues from 128 different sequences. Hence, a striped data layout is the most straightforward solution. Given that warps have exclusive tasks, we propose a data reformatting method that arranges sequences in a striped layout to (a) achieve fully coalesced memory access and (b) balance workload amongst warps. The proposed method partitions large sequence dataset into blocks based on the number of resident warps. As shown in Fig. 4, all protein sequences are divided into N blocks, each consists of Mi × 128 residues, where N is the total number of resident warps, and Mi represents the height of each block with i = [1, 2, 3...N ]. The number 128, as the width of each block, is pre-fixed for two reasons: (a) MSV/SSV kernels only need participant values with the width of ŵv = 8 bits, and hence a warp can handle up 5 Warp 1 Addr. (Bytes) 0 1 3 Warp 2 4 5 Warp 3 Warp N 126 127 128 M M M M M M C A E P S V L C A P E A M V M T L T E R K H S P A L Q L V D G S F L A Y F @ Q K S R @ G T Y A T D T G V L Q K D G H Mi=1 rows 2 P G Q D H @ R G @ S @ K M M V H A R W K C M N R D P E S R @ N D L P D P E V @ I @ D A E @ Q S V A P @ # @ T @ @ G # # # @ # # Q # # # # # # @ # # # # # 128 cols 128 columns (128 Bytes / row) 128 cols 128 cols Fig. 4. Warp-based sequence data blocks. to S = 128 sequences simultaneously. (b) 128 residues, each occupies 1 byte, achieves aligned memory address for coalescing access. Marked as different colors, residues consist of three types: regular (green), padding (blue) and ending (red). In each block, sequences are concatenated and filled up 128 columns. A ending residue ’@’ is inserted at the end of each sequence, which is used to trigger an intra-wrap branch to calculate and record the score of current sequence in the kernel. The value of Mi is equal to the length of longest column within each block, such as second column of the data block for warp 1. As for the rest of columns whose length is less than Mi , padding residue ’#’s are attached to make them all aligned. Residues that belong to the same warp are stored together in row-major order. Algorithm 1 shows the pseduo-code of reformatting protein sequence data. N is determined by the number of SMXs and the number of resident warps per SMX, denoted by Nsmx and Nwarp , respectively (line 1). Given S , N ∗ S containers Cx are created to hold and shape sequences as we need. One container corresponds to one column as shown in Fig. 4. We load first N ∗ S seqeunces as the basis (line 4), and the for loop (line 7 to 19) iterates over the rest to distribute them into Cx . To avoid serious imbalance, the maxLen (line 6) is employed to monitor the longest container as the upper limit. Every new sequence searches for a suitable position based on the accumulated length of each container and maxLen (line 13). We force the sequence to be attached to the last container (line 12) if no position is found. maxLen always be checked and updated by checkM ax() function once a new sequence is attached successfully (line 17). Cx are evenly divdied into N blocks when all sequences are attached, and the length of longest container within each block Mi is recorded by getM ax(S) function (line 20). The padding process, as the last step (line 21), shapes warp-based sequence data blocks Bi into rectangles. Implementation of this method and the evaluation of workload balancing are presented in Sec. 4.1. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 Algorithm 1 Protein sequence data reformatting Input: all sequences sorted by length in descending order Seq , and some parameters, such as S . Output: warp-based sequence data blocks Bi . 1: N ← Nsmx ∗ Nwarp 2: lines ← N ∗ S 3: create lines containers Cx where x ∈ Z ∩ [0, lines). 4: Cx ← Seq[x] . load first lines sequences 5: ptr ← lines . start from last container 6: maxLen ← max{Cx .len()} 7: for all Seq[y] where y ≥ lines do 8: repeat 9: ptr ← ptr − 1 10: if ptr < 0 then . position is not found 11: ptr ← lines . start over 12: C(ptr−1) .attach(Seq[y]) 13: else if Seq[y].len() + Cptr .len() ≤ maxLen then 14: Cptr .attach(Seq[y]) 15: end if 16: until Seq[y] is attached. 17: maxLen.checkM ax() 18: ptr ← lines if ptr < 0. . refresh pointer normally 19: end for . all sequences are attached 20: Mi ← Cx .divide(N ).getM ax(S) where i = [1, ..., N ] 21: Bi ← Cx .padding(Mi ) 22: return Bi 3.3 Kernel Design Since the number of sequences that are processed in parallel is within the range of {2, 4, 8, 16, 32, 64, 128}, the proposed algorithms are designed to cover all these cases. Therefore, 14 different types of kernels are generated, named as S -lane MSV/SSV kernels, and their implementations slightly vary with value S . Algorithm 2 outlines the S -lane MSV kernels that is more complex than the implementation of single-sequence alignment in [18]. Some features are inherited, like (a) using local memory to hold intermediate values as well as register spill (line 1), (b) loading scores through read-only cache instead of shared memory to avoid weak scalability with low occupancy (line 16) and (c) fully unrolling the innermost loop for maximizing registers usage to reside high frequency values on the on-chip memory (line 14). In order to assign different threads of a warp to work on different sequences without mutual interference, we label group ID gid and the offset in group oig on each thread (line 3 and 4). Threads that work on the same sequence are grouped with a unique gid, and they are assigned to different in-group tasks based on the oig . Inter-thread collaborations are only allowed within each thread group. The outer loop (line 5) iterates over columns of the warpbased sequence data block B while the middle loop (line 11) takes charge of each row. The cycle times of outer loop is directly affected by the query model size: the larger model results in the more cycles. This is because on-chip memory resources are limited when making alignment with large models, and it further leads to the kernel selection with small S . For example, a model with length of 45 can be handled by 128-lane kernels whereas a model of 1000-length 6 Algorithm 2 MSV kernels with multi-sequence alignment Input: emission score E , sequence data block B , height of data block M , sequence length Len, offset of sequence Oseq , offset of sequence length Olen and other parameters, such as L, W , H , S and dbias, etc. Output: P-values of all sequences Pi . 1: local memory Γ[H] 2: wid ← blockIdx.y ∗ blockDim.y + threadIdx.y 3: gid ← bthreadIdx.y/Lc . group id 4: oig ← threadIdx.x % L . offset in group 5: for k ← 0 to W − 1 do 6: R ← M [wid] 7: count ← 0 8: mask , scE ← 0x00000000 9: Iseq ← Oseq [wid] + gid ∗ L + bk/4c 10: scB ← initialize it based on Len, Olen [wid] and k . 11: while count < R do 12: r ← extract res(B, Iseq , k, S) 13: γ ← inter or intra reorder(L, S, v, oig) 14: #pragma unroll H 15: for h ← 0 to H − 1 do 16: σ ← load emission score(E, S, L, oig, h, r) 17: v ← vmaxu4(γ, scB ) 18: v ← vaddus4(v, dbias) 19: v ← vsubus4(v, σ) 20: scE ← vmaxu4(scE , v) 21: γ ← Γ[h] & mask . load old scores 22: Γ[h] ← v . store new scores 23: end for 24: scE ← max reduction(L, S, scE ) 25: scJ , scB ← update special states, given scE . 26: mask ← 0xffffffff 27: if r contains ending residue @ then . branch 28: Pi ← calculate P-value and record it. 29: mask ← set affected bits to 0x00. 30: scJ , scE , scB ← update or reset special states. 31: end if 32: count, Iseq ← step up. 33: end while 34: end for may only select 4-lane kernels. Given that, in one warp, k is used to index S columns of residues simultaneously during each iteration, and the Iseq always points to residues that are being extracted from global memory. The details of residue extraction are shown in Algorithm 3. For S -lane kernels with S ≤ 32, only one 8-bit value (one residue) per thread group is extracted and be ready to make alignment though a warp always have the fully coalesced access to 128 residues assembled in 128-byte memory space. Instead, 64lane kernels extract two 8-bit values, and 128-lane kernels are able to handle all of them. These residues are then used in the function load emission score (line 16) to load corresponding emission scores of “Match” states (line 3, 5 and 11 in Algorithm 4). The total number of amino acids is extended to 32, and the extra states are filled with invalid scores, which aims to cover the newly introduced residues (ending and padding). 64 and 128-lane kernels are treated in a specical way as shown in Algorithm 4 (line 3-9 and line 1215) due to the demand of score assembly. In this case, each JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 thread assembles two or four scores of different residues into a 32-bit register to be ready for subsequent SIMD instructions. All emission scores are loaded through read-only cache to keep shared/L1 cache path from overuse, and the score sharing is done via inter-thread shuffle instructions. Algorithm 3 Extract residues from data block - extract res Input: B , Iseq , k and S . Output: a 32-bit value r that contains one, two or four residues, given S . 1: if S ∈ {2, 4, 8, 16, 32} then 2: r ← (B[Iseq ] >> 8 ∗ (k % 4)) & 0x000000ff 3: else if S = 64 then 4: r ← (B[Iseq ] >> 16 ∗ k) & 0x0000ffff 5: else if S = 128 then 6: r ← B[Iseq ] 7: end if Algorithm 4 Get “Match” scores - load emission score Input: E , S , L, oig , h and r. Output: a 32-bit value σ that contains four emission scores, each is 8-bit, in striped layout. 1: Na ← 32 . amino acids 2: if S ∈ {2, 4, 8, 16, 32} then 3: σ ← E[h ∗ Na ∗ L + r ∗ L + oig] . ldg 4: else if S = 64 then 5: sc ← E[h ∗ Na + threadIdx.x] & 0x0000ffff 6: res ← r & 0x000000ff . assembly 7: σ ← σ k ( shf l(sc, res)) 8: res ← (r >> 8) & 0x000000ff 9: σ ← σ k ( shf l(sc, res) << 16) . assembly 10: else if S = 128 then 11: sc ← E[h ∗ Na + threadIdx.x] & 0x000000ff 12: for bits ∈ {0, 8, 16, 24} do 13: res ← (r >> bits) & 0x000000ff . assembly 14: σ ← σ k ( shf l(sc, res) << bits) 15: end for 16: end if Algorithm 5 and 6 detail two crucial steps of MSV/SSV kernels, inter or intra reorder and max reduction, (line 13 and 24 in Algorithm 2) via the PTX assembly to expose internal mechanisms of massive bitwise operaions for the multi-sequence alignment. They aim to reorder 8-bit values and get the maximum value amongst each thread group in parallel, and meanwhile, noninterference between thread groups is guaranteed. Our design still avoids to use shared memory since available L1 cache is the key factor on performance when unrolling the innermost loop. Therefore, all intermediate or temporary values are held by private memory space of each thread, such as registers or local memory. The shuffle instruction, shfl, is employed again to achieve the inter-thread communication but the difference is that a mask is specified to split a warp into sub-segments (line 8 in Algorithm 5). Each sub-segment represents a thread group. In Algorithm 6, two reduction phases are required for S -lane kernels with S ≤ 16. Line 5 to 12 presents interthread reductions by using shfl and vmax to get top four values of 8-bit within each thread group. The following lines are intra-thread reductions which only happen inside 7 each thread and eventually works out the maximum value. As an example, Fig. 5 illustrates the reordering and maxreduction for 16-lane kernels. W = 8 is the width of striped model states, and it can also be regarded as the scope of each thread group. For S = 16, a warp is partitioned into 16 thread groups, and each handles eight values of 8-bit. Yellow lattices in Fig. 5(a) are values that need to be exchanged betweem two threads. Arrows indicate the reordering direction. In Fig. 5(b), assuming digital numbers (0 to 127) labeled inside lattices represent the values held by threads, the yellow lattices always track the maximum value of each thread group. Three pairs of shuffle and SIMD instructions are used to calculate the maximum value and broadcast it to all members of the thread group. Algorithm 5 Reorder 8-bit value inter-thread or intra-thread - inter or intra reorder Input: L, S , v and oig . Output: a 32-bit value γ that contain four 8-bit values. 1: if S ∈ {2, 4, 8, 16} then . inter-thread 2: lane ← (oig + L − 1) % L . source lane 3: asm{shr.u32 x, v, 24} 4: asm{mov.b32 mask, 0x1f} 5: asm{sub.b32 mask, mask, L(S)-1} 6: asm{shl.b32 mask, mask, 16} 7: asm{or.b32 mask, mask, 0x1f} 8: asm{shfl.idx.b32 x, x, lane, mask} 9: asm{shl.b32 y, v, 8} 10: γ←x|y 11: else if S = 32 then . intra-thread 12: asm{shr.u32 x, v, 24} 13: asm{shl.b32 y, v, 8} 14: γ←x|y 15: else if S = 64 then . intra-thread 16: asm{shr.u32 x, v, 8} 17: asm{and.b32 x, x, 0x00ff00ff} 18: asm{shl.b32 y, v, 8} 19: asm{and.b32 y, y, 0xff00ff00} 20: γ←x|y 21: else if S = 128 then 22: γ ← 0x00000000 or 0x80808080 . not reorder 23: end if The potential divergent execution only happens in the branch for recording P-value of ended sequences, as shown in Algorithm 2, line 27-31. Threads which find the existence of ending residues keep active and move into the branch whereas others are inactive during this period. A mask is introduced to mark the position of ended sequences within 32-bit memory space and set those affected bits to 0. This is particularly helpful to 64 and 128-lane kernels because it only cleans up corresponding lanes for new sequence while keeping data of other lanes unchanged. Moreover, bitwise operations with mask minimize the number of instructions needed inside the innermost loop, which is also beneficial to the overall performance. As for the SSV kernel shown in Algorithm 7, it shares the same framework with the MSV kernel but has less computational workload. Besides, one more mask value is added to reset affected bits since the -inf of SSV kernel is 0x80 rather than 0x00. mask1 cleans up outdated scores as the first step followed by a bitwise JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 8 Algorithm 6 Get and broadcast maximum value through reduction operations - max reduction Input: L, S and scE . Output: a 32-bit value scE that contains four 8-bit or two 16-bit values. 1: if S = 128 then 2: do nothing but Return scE . 3: else 4: x ← scE , y ← 0, z ← 0 5: if S ∈ {2, 4, 8, 16} then . inter-thread reduction 6: i ← log2 L − 1 7: for lm ← 20 to 2i do 8: asm{shfl.bfly.b32 y, x, lm, 0x1f} 9: asm{and.b32 m, m, 0x00000000} 10: asm{vmax4.u32.u32.u32 x, x, y, m} 11: end for 12: end if 13: asm{shr.u32 y, x, 8} 14: asm{and.b32 y, y, 0x00ff00ff} 15: asm{shl.b32 z, x, 8} 16: asm{and.b32 z, z, 0xff00ff00} 17: asm{or.b32 y, y, z} 18: asm{and.b32 m, m, 0x00000000} 19: asm{vmax4.u32.u32.u32 x, x, y, m} 20: Return scE ← x if S = 64. 21: asm{shr.u32 y, x, 16} 22: asm{shl.b32 z, x, 16} 23: asm{or.b32 y, y, z} 24: asm{and.b32 m, m, 0x00000000} 25: asm{vmax4.u32.u32.u32 x, x, y, m} 26: Return scE ← x if S ∈ {1, 2, 4, 8, 16, 32}. 27: end if Thread #31 Thread #30 Thread #3 Thread #2 Thread #1 Thread #0 32-bit reg. 32-bit reg. 32-bit reg. 32-bit reg. 32-bit reg. 32-bit reg. Reorder 8-bit value W=8 W=8 W=8 (a) Reordering for MSV/SSV kernels with S = 16 Thread #31 Thread #30 Thread #3 Thread #2 32-bit reg. Thread #1 32-bit reg. 32-bit reg. 32-bit reg. 123 122 121 120 15 14 13 12 11 10 8 7 6 5 4 3 2 1 0 123 122 121 120 127 126 125 124 11 10 8 15 14 13 12 3 2 1 0 7 6 5 4 127 126 125 124 127 126 125 124 15 14 13 12 15 14 13 12 7 6 5 4 7 6 5 4 126 127 124 125 126 127 124 125 14 15 12 13 14 15 12 13 6 7 4 5 6 7 4 5 127 127 125 125 127 127 125 125 15 15 13 13 15 15 13 13 7 7 5 5 7 7 5 5 125 125 127 127 125 125 127 127 13 13 15 15 13 13 15 15 5 5 7 7 5 5 7 7 127 127 127 127 127 127 127 127 15 15 15 15 15 15 15 15 7 7 7 7 7 7 7 7 9 32-bit reg. Thread #0 127 126 125 124 9 32-bit reg. Inter-thread shuffle vmax4 Intra-thread shuffle vmax4 Intra-thread shuffle vmax4 (b) Max-reduction for MSV/SSV kernels with S = 16 Fig. 5. An example of Reordering and Max-reduction for kernels that handle 16 sequences simutaneously. disjunction with mask2 to reset local memory Γ (line 18). 3.4 Algorithm 7 SSV kernel with multi-sequence alignment Input: emission score E , sequence data block B , height of data block M , sequence length Len, offset of sequence Oseq , offset of sequence length Olen and other parameters, such as L, W , H , S and dbias, etc. Output: P-values of all sequences Pi . 1: local memory Γ[H] 2: wid ← blockIdx.y ∗ blockDim.y + threadIdx.y 3: gid ← bthreadIdx.y/Lc . group id 4: oig ← threadIdx.x % L . offset in group 5: for k ← 0 to W − 1 do 6: T ← M [wid] 7: count ← 0 8: mask1, mask2, scE ← 0x80808080 9: Iseq ← Oseq [wid] + gid ∗ L + bk/4c 10: while count < T do 11: r ← extract res(B, Iseq , k, S) 12: γ ← inter or intra reorder(L, S, v, oig) 13: #pragma unroll H 14: for h ← 0 to H − 1 do 15: σ ← load emission score(E, S, L, oig, h, r) 16: v ← vsubus4(v, σ) 17: scE ← vmaxu4(scE , v) 18: γ ← Γ[h] & mask1 k mask2 . load old scores 19: Γ[h] ← v . store new scores 20: end for 21: scE ← max reduction(L, S, scE ) 22: mask1 ← 0xffffffff, mask2 ← 0x00000000 23: if r contains ending residue @ then . branch 24: Pi ← calculate P-value and record it. 25: mask1 ← set affected bits to 0x00. 26: mask2 ← set affected bits to 0x80. 27: scE ← update or reset special states. 28: end if 29: count, Iseq ← step up. 30: end while 31: end for Kernel Optimization The kernel performance of CUDAMPF shows that H = 19 is able to cover the longest query model whose length is 2405, for MSV and SSV algorithms [18]. In current design, however, 2-lane kernels can only handle models with the length of 1216 at most, given the same H . In addition, we recall that L1 Cache-Hit-Ratio (CHR) is employed as a metric to evaluate register spill in CUDAMPF, and MSV/SSV kernels with maximum 64 registers per thread have no spill to local memory. This enables us to push up H to hold larger models for the multi-sequence alignment. Two optimization schemes are proposed to improve overall performance and address the concern about scalability. 3.4.1 CHR Sacrificed Kernel It is well-known that L1 cache shares the same block of on-chip memory with shared memory physically, and it is used to cache accesses to local memory as well as register spill. The L1 cache has low latency on data retrieval and storage with a cache hit, which can be further utilized to increase the throughput of kernels based on the proposed framework. We treat L1 cache as secondary registers, and the usage is measured by CHR for local loads and stores. By increasing up H , more model states can reside in registers and cached local memory. The moderate loss of performance JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 9 due to uncached register spills are acceptable, which is attributed to highly optimized task and data parallelism in the current framework. However, it is impossible to increase H unboundedly due to the limited capacity of L1 cache. Overly large H leads to the severe register spill that causes low CHR and stalls warps significantly. Hence, there always is a trade-off between CHR and H , and the goal is to find a reasonable point where kernel performance (GCUPS) begins fall off. The decline in performance indicates that the latency, caused by excessive communications between on and off-chip memory, starts to overwhelm the benefits of parallelism. The corresponding H at the turning point is considered to be the maximum one, denoted by Hmax . TABLE 1 Benchmarks of the maximum H via innermost loop unrolling Once the kernel type (S -lane) is determined, the H of every query model located in the coverage area can be obtained via: argmin Φ = {H | ∀H ∈ [d H where Φ represents the function of H . Eq. (5) minimizes H to fit query models perfectly, which thereby avoid redundant computation and memory allocation. In summary, this optimization scheme aims to fully leverage the speedy on-chip memory, including L1 cache via sacrificing CHR proactively, to further boost kernel throughput, and in the meanwhile, it extends coverage of the proposed framework to larger query models. 3.4.2 H reg. per thread stack frame spill stores spill loads GCUPS L1 CHR (%) 258.7 261.5 272.1 279.3 283.9 280.6 263.9 240.8 99.97 99.97 99.97 75.65 75.41 75.49 25.61 14.95 269.5 314.0 329.0 347.9 361.5 375.9 375.9 340.9 99.96 99.96 99.96 99.96 99.94 80.44 60.30 17.81 MSV kernels 20 25 30 35 40 45 50 55 63 63 64 64 64 64 64 64 8 8 8 40 48 48 96 128 20 25 30 35 40 45 50 55 62 62 62 61 64 64 64 64 8 8 8 8 16 48 56 80 0 0 0 40 52 56 152 208 0 0 0 44 56 52 96 140 SSV kernels 0 0 0 0 4 44 64 116 0 0 0 0 4 40 44 72 *** Data collections on 32-lane kernels compiled with nvcc 8.0. Use env nr [23] as the sequence dataset. TABLE 1 lists a benchmark result that shows the relationship between H , kernel performance and CHR. Starting from H = 20 with a step of 5 , intuitively, CHR is being consumed after on-chip registers are exhausted, and the kernel performance increases first and falls back eventually as expected. We choose 45 and 50 as the Hmax for MSV and SSV kernels, respectively. Larger H results in rapid degradation of both performance and L1 CHR. Besides, the difference of Hmax indicates that MSV kernels have more instructions than SSV kernels within the innermost loop, and hence more registers or local memory are used while unrolling the loop. The Hmax is therefore algorithmdependent. Given Eq. (2), (3) and Hmax , we formulate the selection of S as below: argmax f = {S | f = WS Hmax , ∀m̂ ≤ W2 Hmax : f ≥ m̂}, S (4) where f is the function of S , and W2 Hmax indicates the maximum length of query models that 2-lane kernels can handle. The CUDAMPF implementation will be used instead if any larger model is applicable. Eq. (4) describes a rule of kernel selection that always prefer to use kernels with more lanes if they are able to cover the model length. m̂ e, Hmax ] ∩ Z : Φ = WS H, WS Φ ≥ m̂}, (5) Performance-oriented Kernel Although the proposed framework achieves a significant improvement in performance, it certainly introduces overhead due to the implementation of multi-sequence alignment, compared with CUDAMPF. This downside becomes more apparent as the model length increases (S decreases). Therefore, it is expected that CUDAMPF with singlesequence alignment may exceed CUDAMPF++ for large enough query models. In order to pursue optimal performance consistently, we also merge CUDAMPF implemention into the proposed framework as a special case with S = 1. The maximum model length m̂max on which CUDAMPF++ still outperforms is defined as the threshold of kernel switch. Similar to m̂max , another threshold m̂min can also be employed to optimize 128 and 64-lane kernels for small models. We recall that 128 and 64-lane kernels need extra operations to load emission scores in Algorithm 4. Thus, they have more overhead than other kernels within the innermost loop, which may counteract their advantages on the number of parallel lanes. We extend the coverage of 32lane kernels to handle small models owned by 128 and 64lane kernels previously, and the evaluation is presented in Sec. 4.3. 4 E XPERIMENTAL R ESULTS In this section, we present several performance evaluations on the proposed CUDAMPF++, such as workload balancing, kernel throughput and scalability. The comparison targets consist of CUDAMPF, CPU-based hmmsearch of latest HMMER v3.1b2 and other related acceleration attempts. Both CUDAMPF++ and CUDAMPF are evaluated on a NVIDIA Tesla K40 GPU and compiled with CUDA v8.0 compiler. Tesla K40 is built with the Kepler GK110 architecture that contains 15 SMXs (2880 CUDA cores) and 12 GB off-chip memory [24]. One of NVIDIA profiling tools, nvprof [25], is also used to track metrics like L1/tex CHR, register usage and spill. For hmmsearch, two types of CPUs are employed to collect performance results: Intel Xeon E5620 (4 physical cores with maximum 8 threads) and dual Intel Xeon E5-2650 (16 physical cores and 32 threads in total). All programs are executed in the 64-bit Linux operating system. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 Unlike CUDAMPF implementation, the NVRTC is deprecated in current design due to its unstability in compiling kernels with high usage of on-chip registers. Even with latest compiler nvcc v8.0, runtime compilation with the NVRTC library still generates unexpected binary files or report the error of resources exhaustion, especially when unrolling large loops and register spill happens. Thus, we choose the just-in-time (JIT) compilation instead. All kernels are pre-compiled in offline mode and stored as .ptx files, each with an unique kernel name inside. Given different query models, the corresponding .ptx file is loaded and further compiled to binary code at runtime. The load time and overhead of compilation are negligible. Two protein sequence databsets [23] are chosen for experiments: (a) env nr (1.9 GB) and (b) est human (5.6 GB). As for query models, we still use Pfam 27.0 [26] that contains 34 thousand HMMs with different sizes ranging from 7 to 2405. The overall performance is measured in kernel throughput (GCUPS) which is directly calculated by the total number of residues contained in each database, model length and kernel execution time. 4.1 Evaluation of Workload Balancing To avoid time overhead of data reformatting introduced in Sec. 3.2, we incorporate Redis [27], a high performance inmemory database, into the proposed framework. Redis is written in ANSI C and able to work with CUDA seamlessly. It currently works as an auxiliary component to hold warpbased sequence data blocks and query models in memory, which offers blazing fast speed for data retrieval. Given the single K40 GPU, each protein sequence dataset is partitioned into 61,440 blocks which are then ingested into Redis database separately as key-value pairs. The quantity of data blocks resided in Redis database should be integral multiple of the number of available warps. Table 2 summarizes the evaluation result of workload balancing for both protein sequence datasets. The “avg.” and “sd.” represent average value and standard deviation across all blocks, respectively. We recall that M is the height of data block which serves as the metrics of computational workload for each warp, and the number of ending residues is another impact factor of performance because the ending residue may lead to thread idling. It is clear to see that both sd. M and sd. ending residues are trivial, and the last two columns show that average multiprocessor efficiency approaches 100%, which are strong evidences of balanced workload over all warps on GPU. Besides, the Padding-toReal Ratio (PRR) that compares the level of invalid computation to the level of desired computation is investigated to assess the negative effect of padding residues, and it is also proved to be negligible. 4.2 Performance Evaluation and Analysis In order to demonstrate the outstanding performance of proposed method and its correlation with the utilization of memory resources, we make an in-depth comparison between CUDAMPF++ and CUDAMPF via profiling both MSV and SSV kernels, reported in Table 3 and 4, respectively. A total of 27 query models are selected to investigate the impact of H on the performance of S -lane kernels. Each 10 kernel type is evaluated with two models that correspond −1 + 1e, except for the 2-lane to H = Hmax and H = d Hmax 2 kernel since the 2405 is the largest model length in [26] with corresponding H = 38. For the MSV kernels, the maximum speedup listed on Table 3 is 17.0x when m̂ = 23, and the trend of speedup is descending as model length increases. This is because memory resources, like on-chip registers, L1 cache and even local memory, are significantly underutilized in CUDAMPF when making alignment with small models whereas CUDAMPF++ always intends to fully take advantage of them. Given 64 as the maximum number of register per thread, only about half the amount of registers are occupied in CUDAMPF till m̂ = 735, and other resources are not utilized at all. In contrast, the CUDAMPF++ not only keeps high usage of registers but also utilizes L1 cache and local memory to hold more data, which results in a near constant performance regardless of the model length. The texture CHR is dominated by model length since we only use texture cache for loading emission scores. Larger model leads to lower texture CHR. Comparing the performance of S -lane kernels in CUDAMPF++, the cases of H = 45 outperform the cases of H = 23 though more local memory are allocated with register spill. One exception is the 128lane kernel due to its higher complexity of innermost loop, which can be optimized via using the 32-lane kernel instead. As shown in Table 4, SSV kernels have similar evaluation results with MSV kernels but higher throughput. Starting from m̂ = 832, nevertheless, CUDAMPF outperforms and eventually yields upto 468.9 GCUPS which is 1.5x faster than CUDAMPF++. The case that peak performance of two frameworks are not comparable is due to the overhead of extra instructions introduced for the multi-sequence alignment in CUDAMPF++. The kernel profiling indicates that both MSV and SSV kernels are bounded by computation and memory bandwidth (texture). However, unlike MSV kernels, SSV kernels have fewer operations within the innermost loop, which makes them more “sensitive”. In other words, newly added operations (i.e., bitwise operations for mask) within the innermost loop, compared with CUDAMPF, have more negative effect on SSV kernels than MSV kernels. Therefore, an upto 50% performance gap is observed only in SSV kernels. 4.3 Scalability Evaluation In order to demonstrate the scalability of the proposed framework, a total of 57 query models with different sizes ranging from 10 to 2450 are investigated. The interval of model length is fixed to 50. Fig. 6 and 7 show the performance comparison between CUDAMPF++ and CUDAMPF for MSV and SSV kernels, respectively. The coverage area of model length for each kernel type is highlighted. Right subfigure depicts the performance of 128 and 64-lane kernels while others are shown in the left one. Overall, CUDAMPF++ achieves near constant performance and significantly outperform CUDAMPF with small models. The S -lane MSV (SSV) kernel yields the maximum speedup of 30x (23x) with respect to CUDAMPF. It is worth mentioning that, in CUDAMPF++, SSV kernels have larger fluctuation margin of performance than MSV kernels. This is caused JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 11 TABLE 2 Evaluation of workload balancing for the warp-based sequence data blocks * DB name DB size (GB) total seq. total residues avg. M sd. M avg. ending residues sd. ending residues avg. PRR SMX eff. MSV (%)* SMX eff. SSV (%)* env nr 1.9 6,549,721 1,290,247,663 21,109 85 13,645 71 1.14E-4 97.13 96.7 est human 5.6 8,704,954 4,449,477,543 72,563 174 18,135 60 3.22E-5 97.97 97.76 Data collection with 32-lane kernels. TABLE 3 Performance comparison of MSV kernel between the proposed CUDAMPF++ and CUDAMPF CUDAMPF++ vs. CUDAMPF *** S -lane kernels model length m̂ acc. ID H reg. per thread stack frame spill stores spill loads L1 CHR (%) Tex. CHR (%) GCUPS speedup 128 128 64 64 32 32 16 16 8 8 4 4 2 2 23 45 46 90 92 180 184 360 368 720 735 1439 1471 2405 PF13823.1 PF05931.6 PF09501.5 PF05777.7 PF00207.17 PF02737.13 PF00596.16 PF01117.15 PF05208.8 PB000053 PF03971.9 PF12252.3 PB006678 PB003055 23 / 2 45 / 2 23 / 2 45 / 2 23 / 2 45 / 2 23 / 2 45 / 3 23 / 3 45 / 6 23 / 6 45 / 12 23 / 12 38 / 19 63 / 29 64 / 29 64 / 29 64 / 29 64 / 29 64 / 29 63 / 29 64 / 30 63 / 30 64 / 34 63 / 34 64 / 51 63 / 51 64 / 64 24 / 0 48 / 0 16 / 0 48 / 0 8/0 48 / 0 8/0 56 / 0 8/0 56 / 0 8/0 56 / 0 8/0 40 / 0 0/0 48 / 0 0/0 64 / 0 0/0 56 / 0 0/0 60 / 0 0/0 60 / 0 0/0 60 / 0 0/0 32 / 0 0/0 24 / 0 0/0 32 / 0 0/0 52 / 0 0/0 60 / 0 0/0 60 / 0 0/0 60 / 0 0/0 44 / 0 99.94 / unused 51.93 / unused 99.96 / unused 50.88 / unused 99.97 / unused 75.49 / unused 99.99 / unused 80.11 / unused 99.99 / unused 80.10 / unused 100 / unused 80.08 / unused 100 / unused 100 / unused 100 / 100 100 / 100 100 / 100 100 / 100 100 / 100 100 / 100 100 / 100 100 / 100 100 / 100 78.48 / 92.37 75.3 / 92.37 67.09 / 72.66 63.38 / 72.6 60.15 / 64.06 168.6 / 9.9 144.4 / 19.4 225.7 / 19.8 231.3 / 38.7 274.1 / 39.6 278.8 / 77.4 266.7 / 78.9 277.0 / 130.9 266.7 / 133.6 271.6 / 183.8 262.4 / 187.4 271.5 / 231.6 259.7 / 237.3 264.0 / 271.2 17.0x 7.5x 11.4x 6.0x 7.0x 3.6x 3.4x 2.1x 2.0x 1.5x 1.4x 1.2x 1.1x -1.0x The env nr [23] is used in data collection. TABLE 4 Performance comparison of SSV kernel between the proposed CUDAMPF++ and CUDAMPF CUDAMPF++ vs. CUDAMPF *** S -lane kernels model length acc. ID H reg. per thread stack frame spill stores spill loads L1 CHR (%) Tex. CHR (%) GCUPS speedup 128 128 64 64 32 32 16 16 8 8 4 4 2 2 26 50 52 100 104 200 208 400 416 800 832 1600 1630 2405 PF02822.9 PF03869.9 PF02770.14 PB000229 PF14807.1 PF13087.1 PF15420.1 PF13372.1 PF06808.7 PF02460.13 PB001474 PB000744 PB000663 PB003055 26 / 2 50 / 2 26 / 2 50 / 2 26 / 2 50 / 2 26 / 2 50 / 4 26 / 4 50 / 7 26 / 7 50 / 13 26 / 13 38 / 19 62 / 33 64 / 33 63 / 33 64 / 33 61 / 33 64 / 33 62 / 33 64 / 37 62 / 37 64 / 42 62 / 42 64 / 56 62 / 56 64 / 63 24 / 0 56 / 0 16 / 0 72 / 0 8/0 56 / 0 8/0 56 / 0 8/0 56 / 0 8/0 56 / 0 8/0 16 / 0 0/0 60 / 0 0/0 100 / 0 0/0 64 / 0 0/0 72 / 0 0/0 72 / 0 0/0 72 / 0 0/0 0/0 0/0 32 / 0 0/0 52 / 0 0/0 44 / 0 0/0 52 / 0 0/0 52 / 0 0/0 52 / 0 0/0 0/0 99.93 / unused 54.29 / unused 99.96 / unused 30.65 / unused 99.96 / unused 60.3 / unused 99.98 / unused 58.57 / unused 99.99 / unused 58.53 / unused 99.99 / unused 58.49 / unused 100 / unused 100 / unused 100 / 100 100 / 100 100 / 100 100 / 100 100 / 100 100 / 100 100 / 100 95.66 / 100 98.92 / 100 75.79 / 87.24 76.90 / 87.24 65.68 / 70.80 66.57 / 70.81 59.69 / 64.05 178.4 / 15.9 144.6 / 30.7 276.1 / 31.8 239.8 / 61.4 307.8 / 63.6 365.0 / 122.4 305.0 / 127.2 347.8 / 199.3 302.6 / 209.1 344.8 / 306.9 302.1 / 319.1 349.3 / 403.5 293.1 / 411.1 318.8 / 468.9 11.2x 4.7x 8.7x 3.9x 4.8x 3.0x 2.4x 1.7x 1.4x 1.1x -1.1x -1.2x -1.4x -1.5x The env nr [23] is used in data collection. by the overhead of using read-only cache (texture cache) to load emission scores. Although both MSV and SSV kernels have only one ldg() function inside innermost loop, the texture cache read in SSV kernels, as one of kernel limits, has a higher proportion of negative effect on performance than that in MSV kernels, which results in such obvious fluctuation. A simple evidence is that the performance curve will be smooth and regular if replacing texture cache reads with a constant value. Besides, 32-lane kernels are also tested to compare with 128 and 64-lane kernels. By decreasing H , the 32-lane kernel is able to cover smaller query models, and it outperforms 128 and 64-lane kernels until the model length is slightly larger than 20. We simply set m̂min = 20 for both MSV and SSV kernels in terms of evaluation results. As for m̂max , the model lengths of 2450 and 1000 are selected for MSV and SSV kernels, respectively. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 Coverage Area of S-lane Kernels 8 4 2 250 200 150 100 CUDAMPF++ CUDAMPF 50 0 500 1000 1500 Model Length 2000 Giga Cells Update Per Second (GCUPS) Giga Cells Update Per Second (GCUPS) 32 16 12 250 200 150 100 50 0 2500 Coverage Area of S-lane Kernels 128 64 32-lane kernel CUDAMPF++ CUDAMPF 10 20 30 40 50 60 Model Length 70 80 90 Fig. 6. Performance comparison between CUDAMPF++ and CUDAMPF for the MSV kernel. 32 16 Coverage Area of S-lane Kernels 8 4 2 350 400 350 300 250 200 150 CUDAMPF++ CUDAMPF 100 0 500 1000 1500 Model Length 2000 2500 Giga Cells Update Per Second (GCUPS) Giga Cells Update Per Second (GCUPS) 450 300 Coverage Area of S-lane Kernels 128 64 32-lane kernel CUDAMPF++ CUDAMPF 250 200 150 100 50 0 20 40 60 Model Length 80 100 Fig. 7. Performance comparison between CUDAMPF++ and CUDAMPF for the SSV kernel. 4.4 Performance Comparison: CUDAMPF++ vs. Others A comprehensive performance comparison is also made between optimized CUDAMPF++ and other implementations. Fig 8 and 9 present results of comparison between CUDAMPF++ and CPU-based MSV/SSV stages with two datasets. The CUDAMPF++ achieves upto 282.6 (283.9) and 465.7 (471.7) GCUPS for MSV and SSV kernels, respectively, given the env nr (est human) dataset. Compared with the best performance achieved by dual Xeon E5-2650 CPUs, a maximum speedup of 168.3x (160.7x) and a minimum speedup of 1.8x (1.7x) are observed for the MSV (SSV) kernel of CUDAMPF++. In the original HMMER3 paper [3], Eddy reports 12 GCUPS for MSV stage, achieved by a single CPU core. Several acceleration efforts exist and report higher performance: (a) an FPGA-based implementation [11] yields upto 81 GCUPS for MSV stage; (b) Lin [13] inherits and modifies a GPU-based implementation of HMMER2 [6] to accelerate MSV stage of HMMER3, which achieves upto 32.8 GCUPS on a Quadro K4000 GPU; (c) [16] claims the first acceleration work on SSV stage of latest HMMER v3.1b2 and reports the maximum performance of 372.1 GCUPS on a GTX570 GPU. To sum up, as shown in Fig. 8 and 9, the proposed JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 HMMER3,env,E5620(8 threads) HMMER3,human,E5620(8 threads) 250 Giga Cells Update Per Second (GCUPS) Giga Cells Update Per Second (GCUPS) CUDAMPF++,env,K40 CUDAMPF++,human,K40 13 200 150 100 50 0 HMMER3,env,E5-2650(32 threads) HMMER3,human,E5-2650(32 threads) 250 200 150 100 50 0 0 500 1000 1500 Model Length 2000 2500 10 20 30 40 50 60 Model Length 70 80 90 Fig. 8. Performance comparison between CUDAMPF++ and HMMER3’s CPU-based implementation for the MSV kernel (stage). HMMER3,env,E5620(8 threads) HMMER3,human,E5620(8 threads) 400 300 200 100 0 HMMER3,env,E5-2650(32 threads) HMMER3,human,E5-2650(32 threads) 350 400 Giga Cells Update Per Second (GCUPS) Giga Cells Update Per Second (GCUPS) CUDAMPF++,env,K40 CUDAMPF++,human,K40 300 250 200 150 100 50 0 500 1000 1500 Model Length 2000 2500 20 40 60 Model Length 80 100 Fig. 9. Performance comparison between CUDAMPF++ and HMMER3’s CPU-based implementation for the SSV kernel (stage). framework, CUDAMPF++, exceeds all existing work and exhibits strong consistency in performance regardless of either the model length or the amount of protein sequences. 5 R ELATED W ORK As one of the most popular tool for the analysis of homologous protein and nucleotide sequences, HMMER at- tracts many acceleration attempts. The previous version, HMMER2, is based on Viterbi algorithm that has proved to be the computational bottleneck. The initial effort of GPUbased implementation for HMMER2 is ClawHMMER [28] which introduces a streaming version of Viterbi algorithm for GPUs. They also demonstrate the implementation running on a 16-node GPU cluster, each equipped with a JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 Radeon 9800 Pro GPU. Another early GPU-based implementation is proposed by Walters et al. [6] who properly fit the Viterbi algorithm into the CUDA-enabled GPU with several optimizations, like memory coalescing, proper kernel occupancy and shared/constant memory usage, which outperforms the ClawHMMER substantially. Yao et al. [29] present a CPU-GPU cooperative pattern to accelerate HMMER2. Ganesan et al. [7] re-design the aligment process of a single sequence across multiple threads to partially break the sequential dependency in computation of Viterbi scores. This helps building a hybrid task and data-level parallelism that eliminates the overhead due to unbalanced sequence lengths. However, with the heuristic pipeline, HMMER3 achieves about 100x to 1000x speedups over its predecessor [3], which hence renders any acceleration effort of HMMER2 obsolete. There are only few existing work that aim to accelerate SSV, MSV and P7Viterbi stages of hmmsearch pipeline in HMMER3. Abbas et al. [11] re-writes mathematical formulas of MSV and Viterbi algorithms to expose reduction and prefix scan computation patterns which are fitted into the FPGA architecture. In [12], a speculative method is proposed to reduce the number of global memory access on the GPU, which aims to accelerate the MSV stage. Lin et al. [13], [14] also focus on MSV stage but incorporate SIMD video instructions provied by the CUDA-enabled GPU into their method. Like the strategy of [6], they assign each thread to handle a whole sequence. A CPU-based implementation of P7Viterb stage is done by Ferreira et al. [15] who propose a cache-oblivious parallel SIMD Viterbi algorithm that offsets cache miss penalties of original HMMER3 work. Neto et al. [16] accelerate the SSV stage via a set of optimizations on the GPU, such as model tiling, outer loop unrolling, coalesced and vectorized memory access. 6 D ISCUSSION While we have shown that the proposed framework with hierarchical parallelism achieves impressive performance based on Kepler architecture, we believe that more advanced GPU architectures, like Maxwell, Pascal and Volta, could also benefit from it because of its hardware-based design. It is easy to port the framework to run on advanced GPUs and gain better performance given more available hardware resources, such as on-chip registers, cache capacity, memory bandwidth and SMs. Also, the framework naturally has linear scalability when distributing protein sequences to multiple GPUs. To handle large models that exceed the carrying capability of single GPU, however, one potential solution is the model partitioning that distributes different segments of model to different GPUs while introducing inter-device communication (i.e., max-reduction, reordering). The multi-GPU implementation of the proposed framework is being investigated. As for the general applicability, not only is the framework suitable for accelerating analogous algorithms of genomic sequence analysis, other domain-specified applications with some features may also benefit from it. The highlight features, for example, may include data irregularity, large-scaled working set and relatively complex logic with execution dependency. In the contrary, for some agent-based 14 problems that usually investigate the behavior of millions of individuals, such as molecular dynamics or simulation of spatio-temporal dynamics, our framework may not be the preferred choice. Actually, the key performance factor is the innermost loop, corresponding to 3rd, 4th and 5th tiers of the proposed framework, in which we should only put necessary operations. In general, assuming that kernels are bound by the innermost loop, there are several suggestions related to minimizing the cost inside the innermost loop: (a) try to hold repeatedly used values in registers to avoid high-frequency communications between on and off-chip memory; (b) pre-load data needed by the innermost loop in outer loops; (c) use either L1 or texture cache to reduce the overhead of load/store operations; (d) try to use highthroughput arithmetic instructions. (e) use shuffle instructions rather than shared memory, if applicable. Ultimately, this work sheds light a strategy to amplify the horsepower of individual GPU in an architecture-aware way while other acceleration efforts usually aim to exploit performance scaling with muliple GPUs. 7 C ONCLUSION In this paper, we propose a five-tiered parallel framework, CUDAMPF++, to accelerate computationally intensive tasks of the homologous protein sequence search with profile HMMs. This framework is based on CUDA-enabled GPUs, and it aims to fully utilize hardware resources of the GPU via exploiting finer-grained parallelism (multi-sequence alignment) compared with its predecessor. In addition, we introduce a novel idea that improves the performance and scalability of the proposed framework by sacrificing L1 CHR proactively. As shown by experimental results, the optimized framework outperforms all existing work, and it exhibits good consistency in performance regardless of the variation of query models or protein sequence datasets. For MSV (SSV) kernels, the peak performance of the CUDAMPF++ is 283.9 (471.7) GCUPS on single K40 GPU, and impressive speedups ranging from 1.8x (1.7x) to 168.3x (160.7x) are achieved over the CPU-based implementation (16 cores, 32 threads). Moreover, further generalization of the proposed framework is also discussed. ACKNOWLEDGMENTS The authors would like to thank the NVIDIA-Professor partnership for generous donations in carrying out this research. R EFERENCES [1] [2] [3] [4] [5] N. Marco S., C. Paolo, T. Andrea, and B. Daniela, “Graphics processing units in bioinformatics, computational biology and systems biology,” Briefings in Bioinformatics, p. 116, 2016. S. Eddy, “Profile hidden markov models,” Bioinformatics, vol. 14, pp. 755–763, 1998. ——, “Accelerated profile HMM searches,” PLoS Comput Biol, vol. 7, no. 10, 2011, doi:10.1371/journal.pcbi.1002195. K. Anders, B. Michael, M. I. Saira, S. Kiminen, and H. David, “Hidden Markov models in computational biology: Applications to protein modeling,” Journal of Molecular Biology, vol. 235, pp. 1501–1531, 1994. S. Altschul, W. Gish, W. Miller, E. Myers, and D. Lipman, “Basic local alignment search tool,” Journal of Molecular Biology, vol. 215, no. 3, pp. 403–410, 1990. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2015 [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] [26] [27] [28] J. P. Walters, V. Balu, S. Kompalli, and V. Chaudhary, “Evaluating the use of gpus in liver image segmentation and hmmer database searches,” in International Symposium on Parallel & Distributed Processing (IPDPS); Rome. IEEE, 2009, pp. 1–12. N. Ganesan, R. D. Chamberlain, J. Buhler, and M. Taufer, “Accelerating HMMER on GPUs by implementing hybrid data and task parallelism,” in Proceedings of the First ACM Int. Conf. on Bioinformatics and Computational Biology (ACM-BCB); Buffalo. ACM, 2010, pp. 418–421. R. Maddimsetty, J. Buhler, R. Chamberlain, M. Franklin, and B. Harris, “Accelerator design for protein sequence HMM search,” in Proc. 20th ACM International Conference on Supercomputing, 2006. T. Oliver, L. Y. Yeow, and B. Schmidt, “Integrating FPGA acceleration into HMMer,” Parallel Computing, vol. 34, no. 11, pp. 681–691, 2008. T. Takagi and T. Maruyama, “Accelerating HMMER search using FPGA,” in International Conference on Field Programmable Logic and Applications (FPL); Prague. IEEE, 2009, pp. 332–337. N. Abbas, S. Derrien, S. Rajopadye, and P. Quinton, “Accelerating HMMER on FPGA using Parallel Prefixes and Reductions,” in International Conference on Field-Programmable Technology (FPT) : 2810 Dec. 2010; Beijing. IEEE., 2010, pp. 37–44. X. Li, W. Han, G. Liu, H. An, M. Xu, W. Zhou, and Q. Li, “A speculative HMMER search implementation on GPU,” in 26th IPDPS Workshop and PhD Forum; Shanghai. IEEE, 2012, pp. 73– 74. L. Cheng and G. Butler, “Implementing and Accelerating HMMER3 Protein Sequence Search on CUDA-Enabled GPU,” Ph.D. dissertation, Concordia University, The Department of Computer Science and Software Engineering, 2014. ——, “Accelerating search of protein sequence databases using CUDA-enabled GPU,” in 20th International Conference on Database Systems for Advanced Applications (DASFAA) : April 20-23. 2015; Hanoi. IEEE., 2015, pp. 279–298. M. Ferreira, N. Roma, and L. Russo, “Cache-Oblivious parallel SIMD Viterbi decoding for sequence search in HMMER,” BMC Bioinformatics, vol. 15, no. 165, 2014. A. C. de Arajo Neto and N. Moreano, “Acceleration of Singleand Multiple-Segment Viterbi Algorithms for Biological SequenceProfile Comparison on GPU,” in 21st International Conference on Parallel and Distributed Processing Techniques and Applications (PDPTA) : July 27-30. 2015; Las Vegas. WORLDCOMP, 2015, pp. 65–71. H. Jiang and G. Narayan, “Fine-Grained Acceleration of HMMER 3.0 via Architecture-aware Optimization on Massively Parallel Processors,” in 14th IEEE International Workshop on High Performance Computational Biology (HiCOMB) in IPDPSW : May 25-29. 2015; Hyderabad. IEEE., 2015. H. Jiang and N. Ganesan, “CUDAMPF: a multi-tiered parallel framework for accelerating protein sequence search in HMMER on CUDA-enabled GPU,” BMC Bioinformatics, vol. 17, no. 1, p. 106, 2016. K. Bjarne. (2015, Mar.) The ssv filter implementation. NVIDIA, “CUDA C programming guide,” June 2017. [Online]. Available: http://docs.nvidia.com/pdf/CUDA C Programming Guide.pdf ——, “NVIDIAs next generation CUDA compute architecture: Kepler GK110/210,” 2014, nVIDIA Corporation Whitepaper. [Online]. Available: http://international.download.nvidia.com/pdf/kepler/ NVIDIA-Kepler-GK110-GK210-Architecture-Whitepaper.pdf ——, “Parallel thread execution ISA,” June 2017. [Online]. Available: http://docs.nvidia.com/pdf/ptx isa 5.0.pdf Fasta sequence databases. sep. 2014. [Online]. Available: ftp://ftp.ncbi.nlm.nih.gov/blast/db/FASTA/ NVIDIA, “NVIDIA Tesla GPU accelerators,” 2013. [Online]. Available: http://www.nvidia.com/content/tesla/pdf/ NVIDIA-Tesla-Kepler-Family-Datasheet.pdf ——, “Profiler user’s guide,” June 2017. [Online]. Available: http://docs.nvidia.com/pdf/CUDA Profiler Users Guide.pdf Pfam: Protein family database. may 2013. [Online]. Available: ftp://ftp.ebi.ac.uk/pub/databases/Pfam/releases/Pfam27.0/ S. Karl and N. Perry, “The little redis book,” 2012. [Online]. Available: http://openmymind.net/redis.pdf D. Horn, M. Houston, and P. Hanrahan, “ClawHMMER: A streaming HMMer-search implementation,” in Proceedings of the ACM/IEEE Supercomputing Conference. IEEE, 2005. 15 [29] Y. Ping, A. Hong, X. Mu, L. Gu, L. Xiaoqiang, W. Yaobin, and H. Wenting, “CuHMMer: A load-balanced CPU-GPU cooperative bioinformatics application,” in International Conference on High Performance Computing and Simulation (HPCS) : June. 2010; Caen, France. IEEE., 2010, pp. 24–30. Hanyu Jiang Biography text here. PLACE PHOTO HERE Narayan Ganesan Biography text here. PLACE PHOTO HERE Yu-Dong Yao Biography text here. PLACE PHOTO HERE 

arXiv:1506.02361v1 [cs.NE] 8 Jun 2015 June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model Julien Chevallier Laboratoire J. A. Dieudonné, UMR CNRS 6621, Université de Nice Sophia-Antipolis, Parc Valrose 06108 Nice Cedex 2, France julien.chevallier@unice.fr Marı́a José Cáceres Departamento de Matemática Aplicada , Universidad de Granada, Campus de Fuentenueva E-18071 Granada, Spain caceresg@ugr.es Marie Doumic UPMC University of Paris 6, JL Lions Lab., 4 place Jussieu 75005 Paris, France Patricia Reynaud-Bouret Laboratoire J. A. Dieudonné, UMR CNRS 6621, Université de Nice Sophia-Antipolis, Parc Valrose 06108 Nice Cedex 2, France patricia.reynaud-bouret@unice.fr The spike trains are the main components of the information processing in the brain. To model spike trains several point processes have been investigated in the literature. And more macroscopic approaches have also been studied, using partial differential equation models. The main aim of the present article is to build a bridge between several point processes models (Poisson, Wold, Hawkes) that have been proved to statistically fit real spike trains data and age-structured partial differential equations as introduced by Pakdaman, Perthame and Salort. Keywords: Hawkes process; Wold process; renewal equation; neural network AMS Subject Classification:35F15, 35B10, 92B20, 60G57, 60K15 Introduction In Neuroscience, the action potentials (spikes) are the main components of the realtime information processing in the brain. Indeed, thanks to the synaptic integration, the membrane voltage of a neuron depends on the action potentials emitted by some others, whereas if this membrane potential is sufficiently high, there is production of action potentials. 1 June 9, 2015 2 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret To access those phenomena, schematically, one can proceed in two ways: extracellularly record in vivo several neurons, at a same time, and have access to simultaneous spike trains (only the list of events corresponding to action potentials) or intracellularly record the whole membrane voltage of only one neuron at a time, being blind to the nearby neurons. Many people focus on spike trains. Those data are fundamentally random and can be modelled easily by time point processes, i.e. random countable sets of points on R+ . Several point processes models have been investigated in the literature, each of them reproducing different features of the neuronal reality. The easiest model is the homogeneous Poisson process, which can only reproduce a constant firing rate for the neuron, but which, in particular, fails to reproduce refractory periodsa . It is commonly admitted that this model is too poor to be realistic. Indeed, in such a model, two points or spikes can be arbitrary close as soon as their overall frequency is respected in average. Another more realistic model is the renewal process 37 , where the occurrence of a point or spike depends on the previous occurrence. More precisely, the distribution of delays between spikes (also called inter-spike intervals, ISI) is given and a distribution, which provides small weights to small delays, is able to mimic refractory periods. A deeper statistical analysis has shown that Wold processes is showing good results, with respect to goodness-of-fit test on real data sets 38 . Wold processes are point processes for which the next occurrence of a spike depends on the previous occurrence but also on the previous ISI. From another point of view, the fact that spike trains are usually non stationary can be easily modelled by inhomogeneous Poisson processes 43 . All those models do not reflect one of the main features of spike trains, which is the synaptic integration and there has been various attempts to catch such phenomenon. One of the main model is the Hawkes model, which has been introduced in 13 and which has been recently shown to fit several stationary data 40 . Several studies have been done in similar directions (see for instance 5 ). More recently a vast interest has been shown to generalized linear models 36 , with which one can infer functional connectivity and which are just an exponential variant of Hawkes models. There has also been several models of the full membrane voltage such as Hodgkin-Huxley models. It is possible to fit some of those probabilistic stochastic differential equations (SDE) on real voltage data 22 and to use them to estimate meaningful physiological parameters 18 . However, the lack of simultaneous data (voltages of different neurons at the same time) prevent these models to be used as statistical models that can be fitted on network data, to estimate network parameters. A simple SDE model taking synaptic integration into account is the well-known Integrate-and-Fire (IF) model. Several variations have been proposed to describe several features of real neural networks such as oscillations 7,8 . In particular, there exists hybrid IF models including inhomogeneous voltage driven Poisson process 21 that are able to mimic real membrane potential data. However up to our knowledge a Biologically, a neuron cannot produce two spikes too closely in time. June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 3 and unlike point processes models, no statistical test have been applied to show that any of the previous variations of the IF model fit real network data. Both, SDE and point processes, approaches are microscopic descriptions, where random noise explains the intrinsic variability. Many authors have argued that there must be some more macroscopic approach describing huge neural networks as a whole, using PDE formalism 15,42 . Some authors have already been able to perform link between PDE approaches as the macroscopic system and SDE approach (in particular IF models) as the microscopic model 39,30,26 . Another macroscopic point of view on spike trains is proposed by Pakdaman, Perthame and Salort in a series of articles 31,32,33 . It uses a nonlinear age-structured equation to describe the spikes density. Adopting a population view, they aim at studying relaxation to equilibrium or spontaneous periodic oscillations. Their model is justified by a qualitative, heuristic approach. As many other models, their model shows several qualitative features such as oscillations that make it quite plausible for real networks, but once again there is no statistical proof of it, up to our knowledge. In this context, the main purpose of the present article is to build a bridge between several point processes models that have been proved to statistically fit real spike trains data and age structured PDE of the type of Pakdaman, Perthame and Salort. The point processes are the microscopic models, the PDE being their mesomacroscopic counterpart. In this sense, it extends PDE approaches for IF models to models that statistically fit true spike trains data. In the first section, we introduce Pakdaman, Perthame and Salort PDE (PPS) via its heuristic informal and microscopic description, which is based on IF models. Then, in Section 2, we develop the different point process models, quite informally, to draw the main heuristic correspondences between both approaches. In particular, we introduce the conditional intensity of a point process and a fundamental construction, called Ogata’s thinning 29 , which allows a microscopic understanding of the dynamics of a point process. Thanks to Ogata’s thinning, in Section 3, we have been able to rigorously derive a microscopic random weak version of (PPS) and to propose its expectation deterministic counterpart. An independent and identically distributed (i.i.d) population version is also available. Several examples of applications are discussed in Section 4. To facilitate reading, technical results and proofs are included in two appendices. The present work is clearly just a first to link point processes and PDE: there are much more open questions than answered ones and this is discussed in the final conclusion. However, we think that this can be fundamental to acquire a deeper understanding of spike train models, their advantages as well as their limitations. 1. Synaptic integration and (PPS) equation Based on the intuition that every neuron in the network should behave in the same way, Pakdaman, Perthame and Salort proposed in 31 a deterministic PDE denoted (PPS) in the sequel. The origin of this PDE is the classical (IF) model. In this section we describe the link between the (IF) microscopic model and the mesoscopic (PPS) June 9, 2015 4 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret model, the main aim being to show thereafter the relation between (PPS) model and other natural microscopic models for spike trains: point processes. 1.1. Integrate-and-fire Integrate-and-fire models describe the time evolution of the membrane potential, V (t), by means of ordinary differential equations as follows dtV = −gL (V − VL ) + I(t), (1.1) dt where Cm is the capacitance of the membrane, gL is the leak conductance and VL is the leak reversal potential. If V (t) exceeds a certain threshold θ, the neuron fires / emits an action potential (spike) and V (t) is reset to VL . The synaptic current I(t) takes into account the fact that other presynaptic neurons fire and excite the neuron of interest, whose potential is given by V (t). As stated in 31 , the origin of (PPS) equation comes from 35 , where the explicit solution of a classical IF model as (1.1) has been discussed. To be more precise the membrane voltage of one neuron at time t is described by: Z t V (t) = Vr + (VL − Vr )e−(t−T )/τm + h(t − u)Ninput (du), (1.2) Cm T where Vr is the resting potential satisfying VL < Vr < θ, T is the last spike emitted by the considered neuron, τm is the time constant of the system (normally τm = gL /Cm ), h is the excitatory post synaptic potential (EPSP) and Ninput is the sum of Dirac masses at each spike of the presynaptic neurons. Since after firing, V (t) is reset to VL < Vr , there is a refractory period when the neuron is less excitable than at rest. The constant time τm indicatesR whether the next spike can occur t more or less rapidly. The other main quantity, T h(t − u)Ninput (du), is the synaptic integration term. In 35 , they consider a whole random network of such IF neurons and look at the behavior of this model, where the only randomness is in the network. In many other studies 7,8,9,11,26,42,30 IF models as (1.1) are considered to finally obtain other systems of partial differential equations (different to (PPS)) describing neural networks behavior. In these studies, each presynaptic neuron is assumed to fire as an independent Poisson process and via a diffusion approximation, the synaptic current is then approximated by a continuous in time stochastic process of Ornstein-Uhlenbeck. 1.2. The (PPS) equation The deterministic PDE proposed by Pakdaman, Perthame and Salort, whose origin is also the microscopic IF model (1.2), is the following: ( ∂n(s,t) + ∂n(s,t) + p (s, X (t)) n (s, t) = 0 ∂t ∂s (PPS) R +∞ m (t) := n (0, t) = 0 p (s, X (t)) n (s, t) ds. In this equation, n(s, t) represents a probability density of neurons at time t having discharged at time t − s. Therefore, s represents the time elapsed since the last June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 5 discharge. The fact that the equation is an elapsed time structured equation is natural, because the IF model (1.2) clearly only depends on the time since the last spike. More informally, the variable s represents the ”age” of the neuron. The first equation of the system (PPS) represents a pure transport process and means that as time goes by, neurons of age s and past given by X(t) are either aging linearly or reset to age 0 with rate p (s, X (t)). The second equation of (PPS) describes the fact that when neurons spike, the age (the elapsed time) returns to 0. Therefore, n(0, t) depicts the density of neurons undergoing a discharge at time t and it is denoted by m(t). As a consequence of this boundary condition, for n at s = 0, the following conservation law is obtained: Z ∞ Z ∞ n (s, t) ds = n (s, 0) ds 0 0 This means that if n (·, 0) is a probabilistic density then n (·, t) can be interpreted as a density at each time t. Denoting by dt the Lebesgue measure and since m(t) is the density of firing neurons at time t in (PPS), m(t)dt can also be interpreted as the limit of Ninput (dt) in (1.2) when the population of neurons becomes continuous. The system (PPS) is nonlinear since the rate p (s, X(t)) depends on n(0, t) by means of the quantity X(t): Zt Zt h(u)m(t − u)du = X(t) = 0 h(u)n(0, t − u)du. (1.3) 0 The quantity X(t) represents the interactions between neurons. It ”takes into account the averaged propagation time for the ionic pulse in this network” 31 . More precisely with respect to the IF models (1.2), this is the synaptic integration term, once the population becomes continuous. The only difference is that in (1.2) the memory is cancelled once the last spike has occurred and this is not the case here. However informally, both quantities have the same interpretation. Note nevertheless, that in 31 , the function h can be much more general than the h of the IF models which clearly corresponds to EPSP. From now on and in the rest of the paper, h is just a general non negative function without forcing the connection with EPSP. The larger p (s, X(t)) the more likely neurons of age s and past X(t) fire. Most of the time (but it is not a requisite), p is assumed to be less than 1 and is interpreted as the probability that neurons of age s fire. However, as shown in Section 3 and as interpreted in many population structured equation 14,19,34 , p(s, X(t)) is closer to a hazard rate, i.e. a positive quantity such that p (s, X(t)) dt is informally the probability to fire given that the neuron has not fired yet. In particular, it could be not bounded by 1 and does not need to integrate to 1. A toy example is obtained if p (s, X(t)) = λ > 0, where a steady state solution is n(s, t) = λe−λs 1s≥0 : this is the density of an exponential variable with parameter λ. However, based on the interpretation of p (s, X(t)) as a probability bounded by 1, one of the main model that Pakdaman, Perthame and Salort consider is p (s, X(t)) = 1s≥σ(X(t)) . This again can be easily interpreted by looking at (1.2). June 9, 2015 6 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret Indeed, since in the IF models the spike happens when the threshold θ is reached, one can consider that p (s, X(t)) should be equal to 1 whenever V (t) = Vr + (VL − Vr )e−(t−T )/τm + X(t) ≥ θ, and 0 otherwise. Since VL − Vr < 0, p (s, X(t)) = 1 is indeed equivalent to s = t − T larger than some decreasing function of X(t). This has the double advantage to give a formula for the refractory period (σ(X(t))) and to model excitatory systems: the refractory period decreases when the whole firing rate increases via X(t) and this makes the neurons fire even more. This is for this particular case that Pakdaman, Perthame and Salort have shown existence of oscillatory behavior 32 . Another important parameter in the (PPS) model and introduced in 31 is J, R which can be seen with our formalism as h and which describes the network connectivity or the strength of the interaction. In 31 it has been proved that, for highly or weakly connected networks, (PPS) model exhibits relaxation to steady state and periodic solutions have also been numerically observed for moderately connected networks. The authors in 32 have quantified the regime where relaxation to a stationary solution occurs in terms of J and described periodic solution for intermediate values of J. Recently, in 33 , the (PPS) model has been extended including a fragmentation term, which describes the adaptation and fatigue of the neurons. In this sense, this new term incorporates the past activity of the neurons. For this new model, in the linear case there is exponential convergence to the steady states, while in the weakly nonlinear case a total desynchronization in the network is proved. Moreover, for greater nonlinearities, synchronization can again been numerically observed. 2. Point processes and conditional intensities as models for spike trains We first start by quickly reviewing the main basic concepts and notations of point processes, in particular, conditional intensities and Ogata’s thinning 29 . We refer the interested reader to 3 for exhaustiveness and to 6 for a much more condensed version, with the main useful notions. 2.1. Counting processes and conditional intensities We focus on locally finite point processes on R, equipped with the borelians B(R). Definition 2.1 (Locally finite point process). A locally finite point process N on R is a random set of points such that it has almost surely (a.s.) a finite number of points in finite intervals. Therefore, associated to N there is an ordered sequence of extended real valued random times (Tz )z∈Z : · · · ≤ T−1 ≤ T0 ≤ 0 < T1 ≤ · · · . For a measurable set A, NA denotes the number of points of N in A. This is a random variable with values in N ∪ {∞}. Definition 2.2 (Counting process associated to a point process). The process on R+ defined by t 7→ Nt := N(0,t] is called the counting process associated to the point process N . June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 7 The natural and the predictable filtrations are fundamental for the present work. Definition 2.3 (Natural
filtration of a point process). The natural filtration of N is the family FtN t∈R of σ-algebras defined by FtN = σ (N ∩ (−∞, t]). Definition 2.4 (Predictable filtration of a point
process). The NpreN defined by Ft− = dictable filtration of N is the family of σ-algebra Ft− t∈R σ (N ∩ (−∞, t)). The intuition behind this concept is that FtN contains all the information given by the point process at time t. In particular, it contains the information whether N only contains the information given t is a point of the process or not while Ft− by the point process strictly before t. Therefore, it does not contain (in general) N represents (the the information whether t is a point or not. In this sense, Ft− information contained in) the past. Under some rather classical conditions 3 , which are always assumed to be satN ), which is isfied here, one can associate to (Nt )t≥0 a stochastic intensity λ(t, Ft− N a non negative random quantity. The notation λ(t, Ft− ) for the intensity refers to the Rpredictable version of the intensity associated to the natural filtration and t N N (Nt − 0 λ(u, Fu− )du)t≥0 forms a local martingale 3 . Informally, λ(t, Ft− )dt represents the probability to have a new point in interval [t, t + dt) given the past. Note N ) should not be understood as a function, in the same way as density that λ(t, Ft− is for random variables. It is a ”recipe” explaining how the probability to find a new point at time t depends on the past configuration: since the past configuration depends on its own past, this is closer to a recursive formula. In this respect, the intensity should obviously depend on N ∩(−∞, t) and not on N ∩(−∞, t] to predict the occurrence at time t, since we cannot know whether t is already a point or not. The distribution of the point process N on R is completely characterized by the N ) on R+ and the distribution of N− = N ∩ R− , knowledge of the intensity λ(t, Ft− which is denoted by P0 in the sequel. The information about P0 is necessary since each point of N may depend on the occurrence of all the previous points: if for all N t > 0, one knows the ”recipe” λ(t, Ft− ) that gives the probability of a new point at time t given the past configuration, one still needs to know the distribution of N− to obtain the whole process. Two main assumptions are used depending on the type of results we seek:  1 RT ,a.s. N ALλ,loc for any T ≥ 0, 0 λ(t, Ft− )dt is finite a.s.  1  hR i T ,exp N ALλ,loc for any T ≥ 0, E 0 λ(t, Ft− )dt is finite.  1   1   1  Clearly ALloc,exp implies ALloc,a.s. . Note that ALloc,a.s. implies non-explosion in finite time for the counting processes (Nt ). Definition 2.5 (Point measure associated to a point process). The point P measure associated to N is denoted by N (dt) and defined by N (dt) = i∈Z δTi (dt), where δu is the Dirac mass in u. By analogy with (PPS), and since points of point processes correspond to spikes June 9, 2015 8 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret (or times of discharge) for the considered neuron in spike train analysis, N (dt) is the microscopic equivalent of the distribution of discharging neurons m(t)dt. Following this analogy, and since TNt is the last point less or equal to t for every t ≥ 0, the age St at time t is defined by St = t − TNt . In particular, if t is a point of N , then St = 0. Note that St is FtN measurable for every t ≥ 0 and therefore, S0 = −T0 is F0N measurable. To define an age at time t = 0, one assumes that (AT0 ) There exists a first point before 0 for the process N− , i.e. −∞ < T0 . As we have remarked before, conditional intensity should depend on N ∩ (−∞, t). Therefore, it cannot be function of St , since St informs us if t is a point or not. N That is the main reason for considering this Ft− measurable variable St− = t − TNt− , (2.1) where TNt− is the last point strictly before t (see Figure 1). Note also that knowing (St− )t≥0 or (Nt )t≥0 is completely equivalent given F0N . The last and most crucial equivalence between (PPS) and the present point N process set-up, consists in noting that the quantities p(s, X(t)) and λ(t, Ft− ) have informally the same meaning: they both represent a firing rate, i.e. both give the rate of discharge as a function of the past. This dependence is made more explicit N ). in p(s, X(t)) than in λ(t, Ft− 2.2. Examples Let us review the basic point processes models of spike trains and see what kind of analogy is likely to exist between both models ((PPS) equation and point processes). These informal analogies are possibly exact mathematical results (see Section 4). N Homogeneous Poisson process This is the simplest case where λ(t, Ft− ) = λ, with λ a fixed positive constant representing the firing rate. There is no dependence in time t (it is homogeneous) and no dependence with respect to the past. This case should be equivalent to p(s, X(t)) = λ in (PPS). This can be made even more explicit. Indeed in the case where the Poisson process exists on the whole real line (stationary case), it is easy to see that
P (St− > s) = P N[t−s,t) = 0 = exp(−λs), meaning that the age St− obeys an exponential distribution with parameter λ, i.e. the steady state of the toy example developed for (PPS) when p(s, X(t)) = λ. Inhomogeneous Poisson process To model non stationarity, one can use N λ(t, Ft− ) = λ(t), which only depends on time. This case should be equivalent to the replacement of p(s, X(t)) in (PPS) by λ(t). Renewal process This model is very useful to take refractory period into account. It corresponds to the case where the ISIs (delays between spikes) are independent and identically distributed (i.i.d.) with a certain given density ν on R+ . The asso- June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 9 ciated hazard rate is ν(s) f (s) = R +∞ , ν(x)dx s R +∞ when s ν(x)dx > 0. Roughly speaking, f (s)ds is the probability that a neuron spikes with age s given that its age is larger than s. In this case, considering the set of spikes as the point process N , it is easy to show (see the Appendix B.1) that its N corresponding intensity is λ(t, Ft− ) = f (St− ) which only depends on the age. One can also show quite easily that the process (St− )t>0 , which is equal to (St )t>0 almost everywhere (a.e.), is a Markovian process in time. This renewal setting should be equivalent in the (PPS) framework to p(s, X(t)) = f (s). Note that many people consider IF models (1.2) with Poissonian inputs with or without additive white noise. In both cases, the system erases all memory after each spike and therefore the ISIs are i.i.d. Therefore as long as we are only interested by the spike trains and their point process models, those IF models are just a particular case of renewal process 8,10,17,35 . Wold process and more general structures Let A1t be the delay (ISI) between the last point and the occurrence just before (see also Figure 1), A1t = TNt− −TNt− −1 . N ) = f (St− , A1t ). This model A Wold process 24,16 is then characterized by λ(t, Ft− has been matched to several real data thanks to goodness-of-fit tests 38 and is therefore one of our main example with the next discussed Hawkes process case. One can show in this case that the successive ISI’s form a Markov chain of order 1 and that the continuous time process (St− , A1t ) is also Markovian. This case should be equivalent to the replacement of p(s, X(t)) in (PPS) by f (s, a1 ), with a1 denoting the delay between the two previous spikes. Naturally in this case, one should expect a PDE of higher dimension with third variable a1 . More generally, one could define Akt = TNt− −(k−1) − TNt− −k , (2.2) N and point processes with intensity λ(t, Ft− ) = f (St− , A1t , ..., Akt ). Those processes satisfy more generally that their ISI’s form a Markov chain of order k and that the continuous time process (St− , A1t , ..., Akt ) is also Markovian (see the Appendix B.2). Remark 2.1. The dynamics of the successive ages is pretty simple. On the one hand, the dynamics of the vector of the successive ages (St− , A1t , ..., Akt )t>0 is deterministic between two jumping times. The first coordinate increases with rate 1. On the other hand, the dynamics at any jumping time T is given by the following shift:    the age process goes to 0, i.e. ST = 0, (2.3) the first delay becomes the age, i.e. A1T + = ST − ,   i−1 i the other delays are shifted, i.e. AT + = AT for all i ≤ k. June 9, 2015 10 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret Hawkes processes The most classical setting is the linear (univariate) Hawkes process, which corresponds to N λ(t, Ft− ) t− Z h(t − u)N (du), =µ+ −∞ where the positive parameter µ is called the spontaneous rate and the non negative function h, with support in R+ , is called the interaction function, which is generally R assumed to satisfy R+ h < 1 to guarantee the existence of a stationary version 16 . This model has also been matched to several real neuronal data thanks to goodnessof-fit tests 40 . Since it can mimic synaptic integration, as explained below, this represents the main example of the present work. In the case where T0 tends to −∞, this is equivalent to say that there is no point on the negative half-line and in this case, one can rewrite N λ(t, Ft− ) Z t− h(t − u)N (du). =µ+ 0 R t− By analogy between N (dt) and m(t)dt, one sees that 0 h(t − u)N (du) is indeed the analogous of X(t) the synaptic integration in (1.3). So one could expect that the PDE analogue is given by p(s, X(t)) = µ + X(t). In Section 4, we show that this does not hold stricto sensu, whereas the other analogues work well. Note that this model shares also some link with IF models. Indeed, the formula for the intensity is close to the formula for the voltage (1.2), with the same flavor for the synaptic integration term. The main difference comes from the fact that when the voltage reaches a certain threshold, it fires deterministically for the IF model, whereas the higher the intensity, the more likely is the spike for the Hawkes model, but without certainty. In this sense Hawkes models seem closer to (PPS) since as we discussed before, the term p(s, X(t)) is closer to a hazard rate and never imposes deterministically the presence of a spike. To model inhibition (see 41 for instance), one can use functions h that may take R t− N negative values and in this case λ(t, Ft− ) = µ + −∞ h(t − u)N (du) , which + N should correspond to p(s, X(t)) = (µ + X(t))+ . Another possibility is λ(t, Ft− )=   R t− exp µ + −∞ h(t − u)N (du) , which is inspired by the generalized linear model as used by 36 and which should correspond to p(s, X(t)) = exp (µ + X(t)). Note finally that Hawkes models in Neuroscience (and their variant) are usually multivariate meaning that they model interaction between spike trains thanks to interaction functions between point processes, each process representing a neuron. To keep the present analogy as simple as possible, we do not deal with those multivariate models in the present article. Some open questions in this direction are presented in conclusion. June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 11 2.3. Ogata’s thinning algorithm N To turn the analogy between p(s, X(t)) and λ(t, Ft− ) into a rigorous result on the PDE level, we need to understand the intrinsic dynamics of the point process. This dynamics is often not explicitly described in the literature (see e.g. the reference book by Brémaud 3 ) because martingale theory provides a nice mathematical setting in which one can perform all the computations. However, when one wants to simulate point processes based on the knowledge of their intensity, there is indeed a dynamics that is required to obtain a practical algorithm. This method has been described at first by Lewis in the Poisson setting 25 and generalized by Ogata in 29 . If there is a sketch of proof in 29 , we have been unable to find any complete mathematical proof of this construction in the literature and we propose a full and mathematically complete version of this proof with minimal assumptions in the Appendix B.4. Let us just informally describe here, how this construction works. The principle consists in assuming that is given an external homogeneous Poisson process Π of intensity 1 in R2+ and with associated point measure Π (dt, dx) = P (T,V )∈Π δ(T,V ) (dt, dx). This means in particular that E [Π(dt, dx)] = dt dx. (2.4) Once a realisation of N− fixed, which implies that F0N is known and which can be seen as an initial condition for the dynamics, the construction of the process N on R+ only depends on Π. N ) in the sense of the ”recipe” More precisely, if we know the intensity λ(t, Ft− that explicitly depends on t and N ∩(−∞, t), then once a realisation of Π and of N− N ) for t ∈ R+ is fixed, the dynamics to build a point process N with intensity λ(t, Ft− is purely deterministic. It consists (see also Figure 1) in successively projecting on N the abscissa axis the points that are below the graph of λ(t, Ft− ). Note that a point N projection may change the shape of λ(t, Ft− ), just after the projection. Therefore the N graph of λ(t, Ft− ) evolves thanks to the realization of Π. For a more mathematical description, see Theorem B.11 in the Appendix B.4.  Note in particular that the construction ends on any finite interval [0, T ] a.s. if A1,a.s λ,loc holds. Then the point process N , result of Ogata’s thinning, is given by the union of N N− on R− and the projected points on R+ . It admits the desired intensity λ(t, Ft− ) on R+ . Moreover, the point measure can be represented by 1t>0 N (dt) = X (T,X)∈Π / X≤λ(T,FTN− ) N Z λ(t,Ft− ) δT (dt) = ! Π (dt, dx) . (2.5) x=0 NB: The last equality comes from the following convention. If δ(c,d) is a Dirac mass Rb in (c, d) ∈ R2+ , then x=a δ(c,d) (dt, dx), as a distribution in t, is δc (dt) if d ∈ [a, b] and 0 otherwise. June 9, 2015 12 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret Fig. 1. Example of Ogata’s thinning algorithm on a linear Hawkes process with interaction function h(u) = e−u and no point before 0 (i.e. N− = ∅). The crosses represent a realization of Π, Poisson N ), process of intensity 1 on R2+ . The blue piecewise continuous line represents the intensity λ(t, Ft− which starts in 0 with value µ and then jumps each time a point of Π is present underneath it. N )) is given by the blue circles. Age S The resulting Hawkes process (with intensity λ(t, Ft− t− at 1 time t and the quantity At are also represented. 3. From point processes to PDE Let us now present our main results. Informally, we want to describe the evolution of the distribution in s of the age St according to the time t. Note that at fixed time t, St− = St a.s. and therefore it is the same as the distribution of St− . We prefer to study St− since its predictability, i.e. its dependence in N ∩ (−∞, t), makes all definitions proper from a microscopic/random point of view. Microscopically, the interest lies in the evolution of δSt− (ds) as a random measure. But it should also be seen as a distribution in time, for equations like (PPS) to make sense. Therefore, we need to go from a distribution only in s to a distribution in both s and t. Then one can either focus on the microscopic level, where the realisation of Π in Ogata’s thinning construction is fixed or focus on the expectation of such a distribution. 3.1. A clean setting for bivariate distributions in age and time In order to obtain, from a point process, (PPS) system we need to define bivariate distributions in s and t and marginals (at least in s), in such a way that weak solutions of (PPS) are correctly defined. Since we want to possibly consider more than two variables for generalized Wold processes, we consider the following definitions. In the following, < ϕ, ν > denotes the integral of the integrable function ϕ with respect to the measure ν. Let k ∈ N. For every bounded measurable function ϕ of (t, s, a1 , ..., ak ) ∈ Rk+2 + , one can define (1) ϕt (s, a1 , ..., ak ) = ϕ(t, s, a1 , ..., ak ) and ϕ(2) s (t, a1 , ..., ak ) = ϕ(t, s, a1 , ..., ak ). Let us now define two sets of regularities for ϕ. June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model Mc,b (Rk+2 + ) 13 The function ϕ belongs to Mc,b (Rk+2 + ) if and only if • ϕ is a measurable bounded function, (1) • there exists T > 0 such that for all t > T , ϕt = 0. ∞ The function ϕ belongs to Cc,b (Rk+2 + ) if and only if • ϕ is continuous, uniformly bounded, ∞ Cc,b (Rk+2 + ) • ϕ has uniformly bounded derivatives of every order, (1) • there exists T > 0 such that for all t > T , ϕt = 0. Let (ν1t )t≥0 be a (measurable w.r.t. t) family of positive measures on Rk+1 + , and k+1 s (ν2 )s≥0 be a (measurable w.r.t. s) family of positive measures R+ . Those families satisfy the Fubini property if R (1) t R (2) (PF ubini ) for any ϕ ∈ Mc,b (Rk+2 hϕt , ν1 idt = hϕs , ν2s ids. + ), k+2 In this case, one can define ν, measure on Rk+2 + , by the unique measure on R+ such that for any test function ϕ in Mc,b (Rk+2 + ), Z Z (1) s < ϕ, ν >= hϕt , ν1t idt = hϕ(2) s , ν2 ids. To simplify notations, for any such measure ν(t, ds, da1 , ..., dak ), we define ν(t, ds, da1 , ..., dak ) = ν1t (ds, da1 , ..., dak ), ν(dt, s, da1 , ..., dak ) = ν2s (dt, da1 , ..., dak ). In the sequel, we need in particular a measure on R2+ , ηx , defined for any real x by its marginals that satisfy (PF ubini ) as follows ∀ t, s ≥ 0, ηx (t, ds) = δt−x (ds)1t−x>0 and ηx (dt, s) = δs+x (dt)1s≥0 . (3.1) It represents a Dirac mass ”travelling” on the positive diagonal originated in (x, 0). 3.2. The microscopic construction of a random PDE For a fixed realization of Π, we therefore want to define a random distribution U (dt, ds) in terms of its marginals, thanks to (PF ubini ), such that, U (t, ds) represents the distribution at time t > 0 of the age St− , i.e. ∀ t > 0, U (t, ds) = δSt− (ds) (3.2) and satisfies similar equations as (PPS). This is done in the following proposition.
N Proposition 3.1. Let Π, F0N and an intensity λ(t, Ft− ) t>0 be given as in Section  1  ,a.s. 2.3 and satisfying (AT0 ) and ALλ,loc . On the event Ω of probability 1, where Ogata’s thinning is well defined, let N be the point process on R that is constructed thanks to Ogata’s thinning with associated predictable age process (St− )t>0 and whose points are denoted (Ti )i∈Z . Let the (random) measure U and its corresponding marginals be defined by U (dt, ds) = +∞ X i=0 ηTi (dt, ds) 10≤t≤Ti+1 . (3.3) June 9, 2015 14 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret Then, on Ω, U satisfies (PF ubini ) and U (t, ds) = δSt− (ds). Moreover, on Ω, U is a solution in the weak sense of the following system ! N Z λ(t,Ft− ) ∂ ∂ U (dt, ds) + U (dt, ds) + Π (dt, dx) U (t, ds) = 0, ∂t ∂s x=0 ! N Z Z λ(t,Ft− ) U (dt, 0) = Π (dt, dx) U (t, ds) + δ0 (dt)1T0 =0 , s∈R+ (3.4) (3.5) x=0 U (0, ds) = δ−T0 (ds)1T0 <0 = U in (ds)1s>0 , (3.6) ∞ where U in (ds) = δ−T0 (ds). The weak sense means that for any ϕ ∈ Cc,b (R2+ ),   Z ∂ ∂ ϕ (t, s) + ϕ (t, s) U (dt, ds) + ∂t ∂s R+ ×R+ ! N Z Z λ(t,Ft− ) [ϕ (t, 0) − ϕ (t, s)] Π (dt, dx) U (t, ds) + ϕ(0, −T0 ) = 0. (3.7) R+ ×R+ x=0 The proof of Proposition 3.1 is included in Appendix A.1. Note also that thanks to the Fubini property, the boundary condition (3.5) is satisfied also in a strong sense. System (3.4)–(3.6) is a random microscopic version of (PPS) if T0 < 0, where n(s, t) the density of the age at time t is replaced by U (t, ·) = δSt− , the Dirac mass in the age at time t. The assumption T0 < 0 is satisfied a.s. if T0 has a density, but this may not be the case for instance if the experimental device gives an impulse at time zero (e.g. 38 studied Peristimulus time histograms (PSTH), where the spike trains are locked on a stimulus given at time 0). This result may seem rather poor from a PDE point of view. However, since this equation is satisfied at a microscopic level, we are able to define correctly all the important quantities at a macroscopic level. Indeed, the analogy between p(s, X(t)) N and λ(t, Ft− ) is actually on the random microscopic scale a replacement of p(s, X(t)) N R λ(t,Ft− ) Π (dt,
dx), whose expectancy given the past is, heuristically speaking, by x=0 N equal to λ t, Ft− because the mean behaviour of Π is given by the Lebesgue measure (see (2.4)). Thus, the main question at this stage is : can we make this argument valid by taking the expectation of U ? This is addressed in the next section. The property (PF ubini ) and the quantities ηTi mainly allows to define U (dt, 0) as well as U (t, ds). As expected, with this definition, (3.2) holds as well as U (dt, 0) = 1t≥0 N (dt), (3.8) i.e. the spiking measure (the measure in time with age 0) is the point measure. Note also that the initial condition is given by F0N , since F0N fixes in particular the value of T0 and (AT0 ) is required to give sense to the age at time 0. To understand the initial condition, remark that if T0 = 0, then U (0, ·) = 0 6= limt→0+ U (t, ·) = δ0 by definitions of ηTi but that if T0 < 0, U (0, ·) = limt→0+ U (t, ·) = δ−T0 . June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 15 R∞ The conservativeness (i.e. for all t ≥ 0, 0 U (t, ds) = 1) is obtained by using (a sequence of test functions converging to) ϕ = 1t≤T . Proposition 3.1 shows that the (random) measure U , defined by (3.3), in terms of a given point process N , is a weak solution of System (3.4)-(3.6). The study of the well-posedness of this system could be addressed following, for instance, the ideas given in 12 . In this case U should be the unique solution of system (3.4)–(3.6). As last comment about Proposition 3.1, we analyse the particular case of the linear Hawkes process, in the following remark. R t− N ) = µ + −∞ h(t − z)N (dz). Remark 3.1. In the linear Hawkes process, λ(t, Ft− Thanks to (3.8) one decomposes the intensity into a term given Rby the initial condit− N ) = µ+F0 (t)+ 0 h(t−z)U (dz, 0), tion plus a term given by the measure U , λ(t, Ft− R0 where F0 (t) = −∞ h(t − z)N− (dz) is (F0N )-measurable and considered as an initial N ) is condition. Hence, (3.4)–(3.6) becomes a closed system in the sense that λ(t, Ft− now an explicit function of the solution of the system. This is not true in general. 3.3. The PDE satisfied in expectation In this section, we want to find the system satisfied by the expectation of the random measure U . First, we need to give a proper definition of such an object. The construction is based on the construction of U and is summarized in the following proposition. (The proofs of all the results of this subsection are in Appendix A.1).
N Proposition 3.2. Let Π, F0N and an intensity λ(t, Ft− ) t>0 be given as in Section  1  ,exp 2.3 and satisfying (AT0 ) and ALλ,loc . Let N be the process resulting of Ogata’s thinning and let U be the random measure defined by (3.3). Let E denote the expectation with respect to Π and F0N . R  2 RThen for any test function ϕ in Mc,b (R+ ), E ϕ(t, s)U (t, ds) and E ϕ(t, s)U (dt, s) are finite and one can define u(t, ds) and u(dt, s) by  Z  Z   ϕ(t, s)u(t, ds) = E ϕ(t, s)U (t, ds) , ∀ t ≥ 0, Z  R   ϕ(t, s)u(dt, s) = E ϕ(t, s)U (dt, s) . ∀ s ≥ 0, Moreover, u(t, ds) and u(dt, s) satisfy (PF ubini ) and one can define u(dt, ds) = u(t, ds)dt = u(dt, s)ds on R2+ , such that for any test function ϕ in Mc,b (R2+ ), Z  Z ϕ(t, s)u(dt, ds) = E ϕ(t, s)U (dt, ds) , quantity which is finite. R  R In particular, since ϕ(t, s)u(t, ds) = E ϕ(t, s)U (t, ds) = E [ϕ(t, St− )], u(t, ·) is therefore the distribution of St− , the (predictable version of the) age at time t. Now let us show that as expected, u satisfies a system similar to (PPS).
N Theorem 3.3. Let Π, F0N and an intensity λ(t, Ft− ) t>0 be given as in Section June 9, 2015 16 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret 2.3 and satisfying (AT0 ) and   L1 ,exp Aλ,loc . If N is the process resulting of Ogata’s thinning, (St− )t>0 its associated predictable age process, U its associated random measure, defined by (3.3), and u its associated mean measure, defined in Proposition 3.2, then, there exists a bivariate measurable function ρλ,P0 satisfying  Z TZ   ∀ T ≥ 0, ρλ,P0 (t, s)u(dt, ds) < ∞, (3.9) 0 s   
 N u(dt, ds)- a.e ρλ,P0 (t, s) = E λ t, Ft− St− = s and such that u is solution in the weak sense of the following system ∂ ∂ u (dt, ds) + u (dt, ds) + ρλ,P0 (t, s)u (dt, ds) = 0, ∂t ∂s Z ρλ,P0 (t, s)u (t, ds) dt + δ0 (dt)uin ({0}), u (dt, 0) = (3.10) (3.11) s∈R+ u (0, ds) = uin (ds)1s>0 , (3.12) ∞ where uin is the law of −T0 . The weak sense means here that for any ϕ ∈ Cc,b (R2+ ),   Z ∂ ∂ + ϕ (t, s) u (dt, ds) + ∂t ∂s R+ ×R+ Z Z [ϕ(t, 0) − ϕ(t, s)]ρλ,P0 (t, s)u(dt, ds) + ϕ(0, s)uin (ds) = 0, (3.13) R+ ×R+ R+ Comparing this system to (PPS), one first sees that n(·, t), the density of the age at time t, is replaced by the mean measure u(t, ·). If uin ∈ L1 (R+ ) we have uin ({0}) = 0 so we get an equation which is exactly of renewal type, as (PPS). In the general case where uin is only a probability measure, the difference with (PPS) lies in the term δ0 (dt)uin ({0}) in the boundary condition for s = 0 and in the term 1s>0 in the initial condition for t = 0. Both these extra terms are linked to the possibility for the initial measure uin to charge zero. This possibility is not considered in 31 - else, a similar extra term would be needed in the setting of 31 as well. As said above in the comment of Proposition 3.1, we want to keep this term here since it models the case where there is a specific stimulus at time zero 38 . In general and without more assumptions on λ, it is not clear that u is not only a measure satisfying (PF ubini ) but also absolutely continuous wrt to dt ds and that the equations can be satisfied in a strong sense. Concerning p(s, X(t)), which has always been thought of as the equivalent of N N λ(t, Ft− ), it is not replaced by λ(t, Ft− ), which would  have no meaning in general N since this is a random quantity, nor by E λ(t, F ) t− which would have been a first  N possible guess; it is replaced by E λ(t, Ft− )|St− = s . Indeed intuitively, since "Z # N λ(t,Ft− )
N N = λ t, Ft− dt, E Π (dt, dx) Ft− x=0 the corresponding weak term can be interpreted as, for any test function ϕ, June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model N Z λ(t,Ft− ) "Z E ϕ (t, s) ! # Z Π (dt, dx) U (t, ds) = E ϕ (t, s) λ N t, Ft−
17  δSt− (ds)dt x=0 Z = Zt = 
 N E ϕ (t, St− ) λ t, Ft− dt  
 N E ϕ (t, St− ) E λ t, Ft− |St− dt, t R which is exactly ϕ(t, s)ρλ,P0 (t, s)u(dt, ds). This conditional expectation makes dependencies particularly complex, but this also enables to derive equations even in non-Markovian setting (as Hawkes processes for instance, see Section 4). More explicitly, ρλ,P0 (t, s) is a function of the time t, of the age s, but it also depends on λ, the shape of the intensity of the underlying process and on the distribution of the initial condition N− , that is P0 . As explained in Section 2, it is both the knowledge of P0 and λ that characterizes the distribution of the process and in general the conditional expectation cannot be reduced to something depending on less than that. In Section 4, we discuss several examples of point processes where one can (or cannot) reduce the dependence. Note that here again, we can prove that the equation is conservative by taking (a sequence of functions converging to) ϕ = 1t≤T as a test function. A direct corollary of Theorem 3.3 can be deduced thanks to the law of large numbers. This can be seen as the interpretation of (PPS) equation at a macroscopic level, when the population of neurons is i.i.d..
∞ Corollary 3.4. Let N i i=1 be some i.i.d. point processes with intensity given by  1  ,exp Ni λ(t, Ft− ) on (0, +∞) satisfying ALλ,loc and associated predictable age processes i (St− )t>0 . Suppose furthermore that the distribution of N 1 on (−∞, 0] is given by 1 P0 which is such that P0 (N− = ∅) = 0. Then there exists a measure u satisfying (PF ubini ), weak solution of Equations (3.10) and (3.11), with ρλ,P0 defined by h   i N1 1 ρλ,P0 (t, s) = E λ t, Ft− |St− = s , u(dt, ds) − a.e. ∞ and with uin distribution of the age at time 0, such that for any ϕ ∈ Cc,b (R2+ ) ! Z Z n 1X a.s. δ i (ds) −−−−→ ϕ(t, s)u(t, ds), (3.14) ∀ t > 0, ϕ(t, s) n→∞ n i=1 St In particular, informally, the fraction of neurons at time t with age in [s, s + ds) in this i.i.d. population of neurons indeed tends to u(t, ds). 4. Application to the various examples Let us now apply these results to the examples presented in Section 2.2. June 9, 2015 18 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret 4.1. When the intensity only depends on time and age
N = f (t, St− ) (homogeneous and inhomogeneous Poisson processes and If λ t, Ft− renewal processes are particular examples) then the intuition giving that p(s, X(t))
N is analogous
to λ t, F works. Let us assume that f (t, s) ∈ L∞ (R2+ ). We have t−  N E λ t, Ft− |St− = s = f (t, s). Under this assumption, we may apply Theorem 3.3, so that we know that the mean measure u associated to the random process is a solution of System (3.10)–(3.12). Therefore the mean measure u satisfies a completely explicit PDE of the type (PPS) with ρλ,P0 (t, s) = f (t, s) replacing p(s, X(t)). In particular, in this case ρλ,P0 (t, s) does not depend on the initial condition. As already underlined, in general, the distribution of the process is characterized by
N = f (t, St− ) and by the distribution of N− . Therefore, in this special λ t, Ft− case, this dependence is actually reduced to the function f and the distribution of
−T0 . Since f (·, ·) ∈ L∞ [0, T ] × R+ , assuming also uin ∈ L1 (R+ ), it is well-known
that there exists a unique solution u such that (t 7→ u(t, ·)) ∈ C [0, T ], L1 (R+ ) , see for instance 34 Section 3.3. p.60. Note that following 12 uniqueness for measure solutions may also be established, hence the mean measure u associated to the random process is the unique solution of System (3.10)–(3.12), and it is in
1 C [0, T ], L (R+ ) : the PDE formulation, together with existence and uniqueness, has provided a regularity result on u which is obtained under weaker assumptions than through Fokker-Planck / Kolmogorov equations. This is another possible application field of our results: using the PDE formulation to gain regularity. Let us now develop the Fokker-Planck / Kolmogorov approach for renewal processes.
N Renewal processes The renewal process, i.e. when λ t, Ft− = f (St− ), with f a continuous function on R+ , has particular properties. As noted in Section 2.2, the renewal age process (St− )t>0 is an homogeneous Markovian process. It is known for a long time that it is easy to derive PDE on the corresponding density through Fokker-Planck / Kolmogorov equations, once the variable of interest (here the age) is Markovian (see for instance 1 ). Here we briefly follow this line to see what kind of PDE can be derived through the Markovian properties and to compare the equation with the (PPS) type system derived in Theorem 3.3. Since f is continuous, the infinitesimal generatorb of (St )t>0 is given by (Gφ)(x) = φ0 (x) + f (x) (φ(0) − φ(x)) , (4.1) for all φ ∈ C 1 (R+ ) (see 2 ). Note that, since for every t > 0 St− = St a.s., the process (St− )t>0 is also Markovian with the same infinitesimal generator. b The infinitesimal generator of an homogeneous Markov process (Zt )t≥0 is the operator G which is defined to act on every function φ : Rn → R in a suitable space D by Gφ(x) = lim t→0+ E [φ(Zt )|Z0 = x] − φ(x) . t June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 19 Let us now define for all t > 0 and all φ ∈ C 1 (R+ ), Z Pt φ(x) = E [φ(St− )|S0 = x] = φ(s)ux (t, ds), where x ∈ R+ and ux (t, ·) is the distribution of St− given that S0 = x. Note that ux (t, ds) corresponds to the marginal in the sense of (PF ubini ) of ux given by Theorem 3.3 with ρλ,P0 (t, s) = f (s) and initial condition δx , i.e. T0 = −x a.s. In this homogeneous Markovian case, the forward Kolmogorov equation gives ∂ Pt = Pt G. ∂t ∞ Let ϕ ∈ Cc,b (R2+ ) and let t > 0. This implies that ∂ ∂ (Pt ϕ(t, s)) = Pt Gϕ(t, s) + Pt ϕ(t, s) ∂t ∂t   ∂ ∂ = Pt ϕ(t, s) + f (s) (ϕ(t, 0) − ϕ(t, s)) + ϕ(t, s) . ∂s ∂t Since ϕ is compactly supported in time, an integration with respect to t yields   Z Z ∂ ∂ −P0 ϕ(0, s) = Pt + ϕ(t, s)dt + Pt f (s) (ϕ(t, 0) − ϕ(t, s)) dt, ∂t ∂s or equivalently  Z Z  ∂ ∂ + ϕ (t, s) ux (t, ds) dt − (ϕ(t, s) − ϕ(t, 0))f (s)ux (t, ds)dt, − ϕ(0, x) = ∂t ∂s (4.2) in in terms of ux . This is exactly Equation (3.13) with u = δx . The result of Theorem 3.3 is stronger than the application of the forward Kolmogorov equation on homogeneous Markovian systems since the result of Theorem 3.3 never used the Markov assumption and can be applied to non Markovian processes (see Section 4.3). So the present work is a general set-up where one can deduce PDE even from non Markovian microscopic random dynamics. Note also that only boundedness assumptions and not continuity ones are necessary to directly obtain (4.2) via Theorem 3.3: to obtain the classical Kolmogorov theorem, one would have assumed f ∈ C 0 (R2+ ) rather than f ∈ L∞ (R2+ ). 4.2. Generalized Wold process
N = f (St− , A1t , ..., Akt ), with f being a non-negative funcIn the case where λ t, Ft− tion, one can define in a similar way uk (t, s, a1 , . . . , ak ) which is informally the distribution at time t of the processes with age s and past given by a1 , ...ak for the last k ISI’s. We want to investigate this case not for its Markovian properties, which are nevertheless presented in Proposition B.2 in the appendix for sake of completeness, but because this is the first basic example where the initial condition is indeed impacting ρλ,P0 in Theorem 3.3. June 9, 2015 20 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret To do so, the whole machinery applied on u(dt, ds) is first extended in the next result
1 k to uk dt, ds, da , . . . , da which represents the dynamics of the age and the last k ISI’s. This could have been done in a very general way by an easy generalisation of Theorem 3.3. However to avoid too cumbersome equations, we express it only for generalized Wold processes to provide a clean setting to illustrate the impact of the initial conditions on ρλ,P0 . Hence, we similarly define a random distribution Uk (dt, ds, da1 , . . . , dak ) such that its evaluation at any given time t exists and is Uk (t, ds, da1 , . . . , dak ) = δ(St− ,A1t ,...,Akt ) (ds, da1 , . . . , dak ). (4.3) The following result states the PDE satisfied by uk = E [Uk ]. Proposition 4.1. Let k be a positive integer and f be some non negative function on Rk+1 + . Let N be a generalized Wold process with predictable age process N ) = f (St− , A1t , ..., Akt ) sat(St− )t>0 points (Ti )i∈Z and intensity λ(t, Ft−  , associated  1 ,exp isfying ALλ,loc , where A1t , . . . , Akt are the successive ages defined by (2.2). Suppose that P0 is such that P0 (T−k > −∞) = 1. Let Uk be defined by Uk (dt, ds, da1 , . . . , dak ) = +∞ X i=0 ηTi (dt, ds) k Y j=1 δAj (daj ) 10≤t≤Ti+1 , (4.4) Ti If N is the result of Ogata’s thinning on the Poisson process Π, then Uk satisfies (4.3) and (PF ubini ) a.s. in Π and F0N . Assume that the initial condition uin k , defined 1 k k+1 as the distribution of (−T0 , A0 , . . . , A0 ) which is a random vector in R , is such k ) = 0. Then U admits a mean measure u which also satisfies ({0} × R that uin k k + k , (PF ubini ) and the following system in the weak sense: on R+ × Rk+1 + ∂ ∂ + uk (dt, ds, da1 , ..., dak )+f (s, a1 , ..., ak )uk (dt, ds, da1 , ..., dak ) = 0, ∂t ∂s Z∞ uk (dt, 0, ds, da1 , ..., dak−1 ) = f (s, a1 , ..., ak ) uk (t, ds, da1 , ..., dak ) dt, (4.5) (4.6) ak =0 uk (0, ds, da1 , . . . , dak ) = uin k (ds, da1 , . . . , dak ) . (4.7) k We have assumed uin k ({0}×R+ ) = 0 (i.e. T0 6= 0 a.s.) for the sake of simplicity, but this assumption may of course be relaxed and Dirac masses at 0 should then be added in a similar way as in Theorem 3.3. If f ∈ L∞ (Rk+1 + ), we may apply Proposition 4.1, so that the mean measure k+1 1 uk satisfy System (4.5)–(4.7). Assuming an initial condition uin k ∈ L (R+ ), we can prove exactly as for the renewal equation (with a Banach fixed point argument for instance)
34that there exists a unique solution uk such that (t 7→ uk (t, ·)) ∈ C R+ , L1 (Rk+1 ) to the generalized Wold case, the boundary assumption on the + kth penultimate point before time 0 being necessary to give sense to the successive ages at time 0. By uniqueness, this proves
that the mean measure uk is this solution, so that it belongs to C R+ , L1 (Rk+1 : Proposition 4.1 leads to a regularity result ) + on the mean measure. June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 21 Now that we have clarified the dynamics of the successive ages, one can look at this system from the point of view of Theorem 3.3, that is when only two variables s and t are considered. In this respect, let us note that U defined by (3.3) is R such that U (dt, ds) = a1 ,...,ak Uk (dt, ds, da1 , . . . , dak ). Since the integrals and the expectations are exchangeable in the weak R sense, the mean measure u defined in Proposition 3.2 is such that u(dt, ds) = a1 ,...,ak uk (dt, ds, da1 , . . . , dak ). But (4.5) ∞ in the weak sense means, for all ϕ ∈ Cc,b (Rk+2 ),  Z  ∂ ∂ + ϕ(t, s, a1 , ..., ak )uk (dt, ds, da1 , . . . , dak ) ∂t ∂s Z + [ϕ (t, 0, a1 , . . . , ak ) − ϕ(t, s, a1 , . . . , ak )] f (s, a1 , . . . , ak ) uk (dt, ds, da1 , . . . , dak ) Z + ϕ (0, s, a1 , . . . , ak ) uin k (ds, da1 , . . . , dak ) = 0. (4.8) ∞ ∞ Letting ψ ∈ Cc,b (R2 ) and ϕ ∈ Cc,b (Rk+2 ) being such that ∀ t, s, a1 , . . . , ak , ϕ(t, s, a1 , . . . , ak ) = ψ(t, s), we end up proving that the function ρλ,P0 defined in Theorem 3.3 satisfies Z ρλ,P0 (t, s)u (dt, ds) = f (s, a1 , . . . , ak ) uk (dt, ds, da1 , . . . , dak ) , (4.9) a1 ,...,ak u(dt, ds)−almost everywhere (a.e.). Equation (4.9) means exactly from a probabilistic point of view that   ρλ,P0 (t, s) = E f (St− , A1t , ..., Akt )|St− = s , u(dt, ds) − a.e. Therefore, in the particular case of generalized Wold process, the quantity ρλ,P0 depends on the shape of the intensity (here the function f ) and also on uk . But, by Proposition 4.1, uk depends on its initial condition given by the distribution of (−T0 , A10 , . . . , Ak0 ), and not only −T0 as in the initial condition for u. That is, as announced in the remarks following Theorem 3.3, ρλ,P0 depends in particular on the whole distribution of the underlying process before time 0, namely P0 and not only on the initial condition for u. Here, for generalized Wold processes, it only depends on the last k points before time 0. For more general non Markovian settings, the integration cannot be simply described by a measure uk in dimension (k + 2) being integrated with respect to da1 ...dak . In general, the integration has to be done on N all the ”randomness” hidden behind the dependence of λ(t, Ft− ) with respect to the past once St− is fixed and in this sense it depends on the whole distribution P0 of N− . This is made even clearer on the following non Markovian example: the Hawkes process. 4.3. Hawkes process As we have seen in Section 2.2, there are many different examples of Hawkes proR 
t− N cesses that can all be expressed as λ t, Ft− = φ −∞ h (t − x) N (dx) , where the main case is φ(θ) = µ + θ, for µ some positive constant, which is the linear case. June 9, 2015 22 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret R 
t− N When there is no point before 0, λ t, Ft− = φ 0 h (t − x) N (dx) . In this case, the interpretation is so close to (PPS) that the first guess, which is wrong, would be that the analogous in (PPS) is p(s, X(t)) = φ(X(t)), (4.10) hR i Rt t− where X(t) = E 0 h (t − x) N (dx) = 0 h (t − x) u(dx, 0). This is wrong, even
N in the linear case since λ t, Ft− depends on all the previous points. Therefore ρλ,P0 defined by (3.9) corresponds to a conditioning given only the last point. By looking at this problem through the generalized Wold approach, one can hope that for h decreasing fast enough:

N λ t, Ft− ' φ h(St− ) + h(St− + A1t ) + ... + h(St− + A1t + ... + Akt ) . In this sense and with respect to generalized Wold processes described in the previous section, we are informally integrating on ”all the previous points” except the last one and not integrating over all the previous points. This is informally why (4.10) is wrong even in the linear case. Actually, ρλ,P0 is computableR for linear t Hawkes processes R: we show in the next section that ρλ,P0 (t, s) 6= φ( −∞ h(t − ∞ x)u(dx, 0)) = µ + 0 h(t − x)u(dx, 0) and that ρλ,P0 explicitly depends on P0 . 4.3.1. Linear Hawkes process We are interested in Hawkes processes with a past before time 0 given by F0N , which is not necessarily the past given by a stationary Hawkes process. To illustrate the fact  that  the past is impacting the value of ρλ,P0 , we focus on two particular cases: A1N N− = {T0 } a.s. and T0 admits a bounded density f0 on R−  − A2N− N− is an homogeneous Poisson process with intensity α on R− Before stating the main result, we need some technical definitions. Indeed the proof is based on the underlying branching structure of the linear Hawkes process described in Section B.3.1 of the appendix and the following functions (Ls , Gs ) are naturally linked to this branching decomposition (see Lemma B.7). Lemma 4.2. Let h ∈ L1 (R+ ) such that khkL1 < 1. For all s ≥ 0, there exist a unique solution (Ls , Gs ) ∈ L1 (R+ ) × L∞ (R+ ) of the following system Z (x−s)∨0 Z Gs (x − w)h(w)dw − log(Gs (x)) = 0 x h(w)dw, (4.11) (h (w) + Ls (w)) Gs (w)h(x − w) dw, (4.12) 0 Z x Ls (x) = s∧x where a ∨ b (resp. a ∧ b) denotes the maximum (resp. minimum) between a and b. Moreover, Ls (x ≤ s) ≡ 0, Gs : R+ → [0, 1], and Ls is uniformly bounded in L1 . June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 23 This result allows to define two other important quantities, Ks and q, by, for all s, t ≥ 0, z ∈ R, Z (t−s)∨0 [h(t − x) + Ls (t − x)] Gs (t − x)h(x − z)dx, Ks (t, z) := 0 Z t log(q(t, s, z)) := − Z (t−s)∨0 h(x − z)dx − (t−s)∨0 [1 − Gs (t − x)] h(x − z)dx. (4.13) 0 Finally, the following result is just an obvious remark that helps to understand the resulting system. Remark 4.1. For a non negative Φ ∈ L∞ (R+ ) and v in ∈ L∞ (R+ ), there exists a unique solution v ∈ L∞ (R2+ ) in the weak sense to the following system, ∂ ∂ v(t, s) + v(t, s) + Φ(t, s)v(t, s) = 0, ∂t ∂s v(t, 0) = 1 v(t = 0, s) = v in (s) (4.14) (4.15) Moreover t 7→ v(t, .) is in C(R+ , L1loc (R+ )). If v in is a survival function (i.e. non increasing from 0 to 1), then v(t, .) is a survival function and −∂s v is a probability measure for all t > 0. Proposition 4.3. Using the notations of Theorem 3.3,   let  N be  a Hawkes process with past before 0 given by N− satisfying either A1N− or A2N− and with intensity on R+ given by Z t− N λ(t, Ft− )=µ+ h(t − x)N (dx), −∞ where µ is a positive realRnumber and h ∈ L∞ (R+ ) is a non-negative function with support in R+ such that h < 1. Then, the mean measure u defined in Proposition 3.2 satisfies Theorem 3.3 and R∞ moreover its integral v(t, s) := u(t, dσ) is the unique solution of the system (4.14)– s ∞ (4.15) where v in is the survival function of −T0 , and where Φ = Φµ,h P0 ∈ L (R+ ) is defined by µ,h h Φµ,h P0 = Φ+ + Φ−,P0 , (4.16) where for all non negative s, t  Z Φµ,h (t, s) = µ 1 + + t  (h(x) + Ls (x))Gs (x)dx , (4.17) s∧t   and where under Assumption A1N− , R 0∧(t−s) Φh−,P0 (t, s) = −∞ (h(t − t0 ) + Ks (t, t0 )) q(t, s, t0 )f0 (t0 )dt0 , R 0∧(t−s) q(t, s, t0 )f0 (t0 )dt0 −∞ (4.18) June 9, 2015 24 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret   or, under Assumption A2N− , Φh−,P0 (t, s) = α Z 0∧(t−s) (h(t − z) + Ks (t, z)) q(t, s, z)dz. (4.19) −∞ In these formulae, Ls , Gs , Ks and q are given by Lemma 4.2 and (4.13). Moreover Z +∞ Z +∞ ∀ s ≥ 0, ρλ,P0 (t, x)u(t, dx) = Φµ,h (t, s) u(t, dx). (4.20) P0 s s The proof is included in Appendix B.3. Proposition 4.3 givesa purely  analytical   1 definition for v, and thus for u, in two specific cases, namely AN− or A2N− . In the general case, treated in Appendix B (Proposition B.5), there remains a dependence with respect to the initial condition P0 , via the function Φh−,P0 . Remark 4.2. Contrarily to the general result in Theorem 3.3, Proposition 4.3 R +∞ focuses on the equation satisfied by v(dt, s) = s u(dt, dx) because in Equation (4.14) the function parameter Φ = Φµ,h P0 may be defined independently of the definitions of v or u, which is not the case for the rate ρλ,P0 appearing in Equation (3.10). Thus, it is possible to depart from the system of equations defining v, study it, prove existence, uniqueness and regularity for v under some assumptions on the initial distribution uin as well as on the birth function h, and then deduce regularity or asymptotic properties for u without any previous knowledge on the underlying process. In Sections 4.1 and 4.2, we were able to use the PDE formulation to prove that the distribution of the ages u has a density. Here, since we only obtain a closed formula for v and not for u, we would need to derive Equation (4.14) in s to obtain a similar µ,h result, so that we need to prove more regularity on Φµ,h P0 . Such regularity for ΦP0 is not obvious since it depends strongly on the assumptions on N− . This paves the way for future research, where the PDE formulation would provide regularity on the distribution of the ages, as done above for renewal and Wold processes.     Remark 4.3. These two cases A1N− and A2N− highlight the dependence with respect to all the past before time 0 (i.e. P0 ) and not only the initial condition (i.e. in the  age at time 0). In fact, they can give the same initial condition u : for instance, A1N− with −T0 exponentially distributed with parameter α > 0 gives the same   law for −T0 as A2N− with parameter α. However, if we fix some non-negative real number s, one can show that Φh−,P0 (0, s) is different in those two cases. It is clear from the definitions that for every real number z, q(0, s, z) = 1 and Ks (0, z) = 0. Thus, in the first case, R −s R∞ h(−t0 )αeαt0 dt0 h(z)αe−αz dz −∞ h Φ−,P0 (0, s) = = s R ∞ −αz , R −s αe dz αeαt0 dt0 s −∞ June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 25 R −s R∞ while in the second case, Φh−,P0 (0, s) = α −∞ h(−z)dz = α s h(w)dw. Therefore Φh−,P0 clearly depends on P0 and not just on the distribution of the last point before 0, and so is ρλ,P0 . Rt Remark 4.4. If we follow our first guest, ρλ,P0 would be either µ + 0 h(t − Rt x)u(dx, 0) or µ+ −∞ h(t−x)u(dx, 0). In particular, it would not depend on the age s. Therefore by (4.20), so would Φµ,h P0 . But for instance at time t = 0, when N− is R +∞ an homogeneous Poisson process of parameter α, Φµ,h h(w)dw, P0 (0, s) = µ + α s which obviously depends on s. Therefore the intuition linking Hawkes processes and (PPS) does not apply. 4.3.2. Linear Hawkes process with no past before time 0   A classical framework in point processes theory is the case in A1N− where T0 → R t− N ) = µ + 0 h(t − x)N (dx). The −∞, or equivalently, when N has intensity λ(t, Ft− problem in this case is that the age at time 0 is not finite. The age is only finite for times greater than the first spiking time T1 . Here again, the quantity v(t, s) reveals more informative and easier to use: having the distribution of T0 going to −∞ means that Supp(uin ) goes to +∞, so that the initial condition for v tends to value uniformly 1 for any 0 ≤ s < +∞. If we can prove that the contribution of Φh−,P0 vanishes, the following system is a good candidate to be the limit system: ∂ ∞ ∂ ∞ ∞ v (t, s) + v (t, s) + Φµ,h + (t, s) v (t, s) = 0, ∂t ∂s v ∞ (t, 0) = 1, v ∞ (0, s) = 1, (4.21) (4.22) where Φµ,h + is defined in Proposition 4.3. This leads us to the following proposition. Proposition 4.4. Under the assumptions and notations of Proposition 4.3, consider for all M ≥ 0, vM theunique solution of system (4.14)-(4.15) with Φ given by Proposition 4.3, case A1N− , with T0 uniformly distributed in [−M −1, −M ]. Then, as M goes to infinity, vM converges uniformly on any set of the type (0, T ) × (0, S) towards the unique solution v ∞ of System (4.21)-(4.22). Conclusion We present in this article a bridge between univariate point processes, that can model the behavior of one neuron through its spike train, and a deterministic age structured PDE introduced by Pakdaman, Perthame and Salort, named (PPS). More precisely Theorem 3.3 present a PDE that is satisfied by the distribution u of the age s at time t, where the age represents the delay between time t and the last spike before t. This is done in a very weak sense and some technical structure, namely (PF ubini ), is required. June 9, 2015 26 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret The main point is that the ”firing rate” which is a deterministic quantity written as p(s, X(t)) in (PPS) becomes the conditional expectation of the intensity given the age at time t in Theorem 3.3. This first makes clear that p(s, X(t)) should be interpreted as a hazard rate, which gives the probability that a neuron fires given that it has not fired yet. Next, it makes clearly rigorous several ”easy guess” bridges between both set-ups when the intensity only depends on the age. But it also explained why when the intensity has a more complex shape (Wold, Hawkes), this term can keep in particular the memory of all that has happened before time 0. One of the main point of the present study is the Hawkes process, for which what was clearly expected was a legitimation of the term X(t) in the firing rate p(s, X(t)) of (PPS), which models the synaptic integration. This is not the case, and the interlinked equations that have been found for the cumulative distribution function v(t, ·) do not have a simple nor direct deterministic interpretation. However one should keep in mind that the present bridge, in particular in the population wide approach, has been done for independent neurons. This has been done to keep the complexity of the present work reasonable as a first step. But it is also quite obvious that interacting neurons cannot be independent. So one of the main question is: can we recover (PPS) as a limit with precisely a term of the form X(t) if we consider multivariate Hawkes processes that really model interacting neurons ? Acknowledgment This research was partly supported by the European Research Council (ERC Starting Grant SKIPPERAD number 306321), by the french Agence Nationale de la Recherche (ANR 2011 BS01 010 01 projet Calibration) and by the interdisciplanary axis MTC-NSC of the University of Nice Sophia-Antipolis. MJC acknowledges support from the projects MTM2014-52056-P (Spain) and P08-FQM-04267 from Junta de Andalucı́a (Spain). We warmly thank François Delarue for very fruitful discussions and ideas. A. Proofs linked with the PDE A.1. Proof of Proposition 3.1 First, let us verify that U satisfies Equation (3.2). For any t > 0, U (t, ds) = X ηTi (t, ds)10≤t≤Ti+1 , i≥0 by definition of U . Yet, ηTi (t, ds) = δt−Ti (ds)1t>Ti , and the only i ∈ N such that Ti < t ≤ Ti+1 is i = Nt− . So, for all t > 0, U (t, ds) = δt−TNt− (ds) = δSt− (ds). Secondly, let us verify that U satisfies (PF ubini ). Let ϕ ∈ Mc,b (R2+ ), and let T be June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 27 (1) P+∞ = 0. Then since U (t, ds) = i=0 ηTi (t, ds)10≤t≤Ti+1 ,   ! Z Z X  ϕ(t, s)U (t, ds) dt ≤ |ϕ(t, s)| ηTi (t, ds)10≤t≤Ti+1  dt such that for all t > T , ϕt Z Z R+ = R+ R+ XZ i≥0 R+ |ϕ(t, t − Ti )|1t>Ti 10≤t≤Ti+1 dt = R+ i≥0 XZ i≥0 Z T1 0 |ϕ(t, t − Ti )|dt max(0,Ti ) Z X |ϕ(t, t − T0 )| + = Ti+1 Ti+1 |ϕ(t, t − Ti )|dt. i/0<Ti <T Ti Since there is a finite number of points of NR between 0 and T , on Ω, this quantity P +∞ R +∞ is finite and one can exchange i≥0 and t=0 s=0 . Therefore, since all the ηTi satisfy (PF ubini ) and ϕ(t, s)10≤t≤Ti+1 is in Mc,b (R2+ ), so does U . ∞ For the dynamics of U , similar computations lead for every ϕ ∈ Cc,b (R+ 2 ) to Z X Z Ti+1 −Ti ϕ (t, s) U (dt, ds) = ϕ (s + Ti , s) ds. i≥0 max(0,−Ti ) We also have   Z  X Z Ti+1 −Ti  ∂ ∂ ∂ ∂ + ϕ (t, s) U (dt, ds) = + ϕ (s + Ti , s) ds ∂t ∂s ∂t ∂s i≥0 max(0,−Ti ) X = [ϕ (Ti+1 , Ti+1 − Ti ) − ϕ (Ti , 0)] + ϕ(T1 , T1 − T0 ) − ϕ(0, −T0 ). (A.1) i≥1 R λ(t,F N ) P It remains to express the term with x=0 t− Π (dt, dx) = i≥0 δTi+1 (dt), that is X Z Z Z X ϕ (t, s) U (t, ds) δTi+1 (dt) = ϕ (t, s) U (t, ds) δTi+1 (dt) i≥0 i≥0 Z = ϕ (t, St− ) X i≥0 and, since R δTi+1 (dt) = X ϕ (Ti+1 , Ti+1 − Ti ) , (A.2) i≥0 U (t, ds) = 1 for all t > 0, Z Z X X ϕ (t, 0) U (t, ds) δTi+1 (dt) = ϕ (Ti+1 .0) , i≥0 (A.3) i≥0 Identifying all the terms in the right-hand side of Equation (A.1), this lead to Equation (3.7), which is the weak formulation of System (3.4)–(3.6). A.2. Proof of Proposition 3.2 (1) Let ϕ ∈ Mc,b (R2+ ), and let T be such that for all t > T , ϕt = 0. Then, Z |ϕ(t, s)|U (t, ds) ≤ ||ϕ||L∞ 10≤t≤T , (A.4) June 9, 2015 28 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret R since R at any fixed time t > 0, U (t, ds) = 1. Therefore, the expectation E ϕ(t, s)U (t, ds) is well-defined and finite and so u(t, .) is well-defined. On the other hand, at any fixed age s, Z ∞ X |ϕ(t, s)|U (dt, s) = |ϕ(s + Ti , s)|10≤s≤Ti+1 −Ti i=0 = X |ϕ(s + Ti , s)|10≤s+Ti ≤T 10≤s≤Ti+1 −Ti , i≥0 (1) because for all t > T , ϕt Z |ϕ(t, s)|U (dt, s) = 0. Then, one can deduce the following bound ≤ |ϕ(s + T0 , s)|1−T0 ≤s≤T −T0 10≤s≤T1 −T0 + X |ϕ(s + Ti , s)|10≤s≤T 1Ti ≤T i≥1 Since the intensity is L1loc Z E ≤ ||ϕ||L∞ (1−T0 ≤s≤T −T0 + NT 10≤s≤T ) . hR i T N in expectation, E [NT ] = E 0 λ(t, Ft− )dt <∞ and  |ϕ(t, s)|U (dt, s) ≤ ||ϕ||L∞ (E [1−T0 ≤s≤T −T0 ] + E [NT ] 10≤s≤T ) , (A.5) so the expectation is well-defined and finite and so u(·, s) is well-defined. Now, let us show (PF ubini ). First Equation (A.4) implies Z  Z E |ϕ(t, s)|U (t, ds) dt ≤ T ||ϕ||L∞ , and Fubini’s theorem implies that the following integrals are well-defined and that the following equality holds, Z  Z Z  Z E ϕ(t, s)U (t, ds) dt = E ϕ(t, s)U (t, ds)dt . (A.6) Secondly, Equation (A.5) implies Z  Z E |ϕ(t, s)|U (dt, s) ds ≤ ||ϕ||L∞ (T + T E [NT ]) , by exchanging the integral with the expectation and Fubini’s theorem implies that the following integrals are well-defined and that the following equality holds, Z  Z Z  Z E ϕ(t, s)U (dt, s) ds = E ϕ(t, s)U (dt, s)ds . (A.7) Now, it only remains to use (PF ubini ) for U to deduce that the right members of Equations (A.6) and (A.7) Rare  (PF ubini ) for U tells that these two R equal. Moreover, quantities are equal to E ϕ(t, s)U (dt, ds) . This concludes the proof. June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model A.3. Proof of Theorem 3.3  N E λ(t,Ft− )1|S t− −s|≤ε 29  , for every t > 0 and s ≥ 0. Since Let ρλ,P0 (t, s) := lim inf ε↓0 P(|St− −s|≤ε) N (λ(t, Ft− ))t>0 and (St− )t>0 are predictable processes, and a fortiori progressive processes (see page 9 in 3 ), ρλ,P0 is a measurable function of (t, s).  N For every t > 0, let µt be the measure defined by µ (A) = E λ(t, F )1 (S ) t A t− t−  1  ,exp for all measurable set A. Since Assumption ALλ,loc implies that dt-a.e.   N E λ(t, Ft− ) < ∞ and since u(t, ds) is the distribution of St− , µt is absolutely continuous with respect to u(t, ds) for dt-almost every t. Let ft denote the Radon Nikodym derivative  of µt with respect to u(t, ds). N ) St− = s by definition of the conditional For u(t, ds)-a.e. s, ft (s) = E λ(t, Ft− expectation. Moreover, a Theorem of Besicovitch 27 claims that for u(t, ds)-a.e.   N ) St− = s holds s, ft (s) = ρλ,P0 (t, s). Hence, the equalityρλ,P0 (t, s) = E λ(t, Ft− u(t, ds)dt = u(dt, ds)-almost everywhere. Next, in order to use (PF ubini ), let us note that for any T, K > 0, 2 ρK,T (A.8) λ,P0 : (t, s) 7→ (ρλ,P0 (t, s) ∧ K) 10≤t≤T ∈ Mc,b (R+ )  R R K,T R R K,T Hence, ρλ,P0 (t, s)u(dt, ds) = ρλ,P0 (t, s)u(t, ds) dt which is always upper
 RT R RT RT  N bounded by 0 ρλ,P0 (t, s)u(t, ds) dt = 0 µt (R+ )dt = 0 E λ(t, Ft− ) dt < ∞. RT R Letting K → ∞, one has that 0 ρλ,P0 (t, s)u(dt, ds) is finite for all T > 0. Once ρλ,P0 correctly defined, the proof of Theorem 3.3 is a direct consequence of Proposition 3.1. More precisely, let us show that (3.7) implies (3.13). Taking the expectation ∞ of (3.7) gives that for all ϕ ∈ Cc,b (R2+ ), N Z λ(t,Ft− ) "Z E [ϕ (t, s) − ϕ(t, 0)] ! # Z Π (dt, dx) U (t, ds) − ϕ (0, s) uin (ds) x=0 Z − (∂t + ∂s ) ϕ (t, s) u (dt, ds) = 0. (A.9) denote ψ(t,s) := ϕ(t, s) − ϕ(t, 0). Due to Ogata’s thinning construction,  Let us N R λ(t,Ft− ) Π (dt, dx) = N (dt)1t>0 where N is the point process constructed by x=0 thinning, and so, ! # "Z N Z Z λ(t,Ft− ) E ψ (t, s) Π (dt, dx) U (t, ds) = E x=0  ψ (t, St− ) N (dt) . t>0 (A.10) But ψ(t, St− ) is a (FtN )-predictable process and Z E t>0 N |ψ(t, St− )|λ(t, Ft− )dt "Z  ≤ kψkL∞ E 0 # T N λ(t, Ft− )dt < ∞, June 9, 2015 30 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret hence, using the martingale property of the predictable intensity, Z  Z 
N E ψ (t, St− ) N (dt) = E ψ (t, St− ) λ t, Ft− dt . t>0 (A.11) t>0 Moreover, thanks to Fubini’s Theorem, the right-hand term is finite and equal to R N E[ψ (t, St− ) λ(t, Ft− )]dt, which can also be seen as Z Z E [ψ (t, St− ) ρλ,P0 (t, St− )] dt = ψ(t, s)ρλ,P0 (t, s)u(t, ds)dt. (A.12) For all K > R0, ((t, s) 7→ ψ(t, s) (ρλ,P0 (t, s) ∧ K)) ∈RMc,b (R2+ ) and, from (PF ubini ), it is clear that ψ(t, s) (ρλ,P0 (t, s) ∧ K) u(t, ds)dt = ψ(t, s) (ρλ,P0 (t, s) ∧ K) u(dt, ds). Since RoneR can always upper-bound this quantity in absolute value by T kψkL∞ 0 s ρλ,P0 (t, s)u(dt, ds), this is finite. Letting K → ∞ one can show that Z Z ψ(t, s)ρλ,P0 (t, s)u(t, ds)dt = ψ(t, s)ρλ,P0 (t, s)u(dt, ds). (A.13) Gathering (A.10)-(A.13) with (A.9) gives (3.13). A.4. Proof of Corollary 3.4 i i For all i ∈ N∗ , let us denote N+ = N i ∩ (0, +∞) and N− = N i ∩ R− . Thanks i can be seen as constructed via thinning to Proposition B.12, the processes N+ 2 of independent Poisson processes on R+ . Let (Πi )i∈N be the sequence of point measures associated to independent Poisson processes of intensity 1 on R2+ given i . In particular, by Proposition B.12. Let T0i denote the closest point to 0 in N− i (T0 )i∈N∗ is a sequence of i.i.d. random variables. For each i, let U i denote the solution of the microscopic equation corresponding to Πi and T0i as defined in Proposition 3.1 by (3.3). Using (3.2), it is clear that Pn Pn i ∞ 2 i (ds) = i=1 δSt− i=1 U (t, ds) for all t > 0. Then, for every ϕ ∈ Cc,b (R+ ), ! Z n n Z 1X 1X δ i (ds) = ϕ(t, s)U i (t, ds). ϕ(t, s) n i=1 St n i=1 R The right-hand side is a sum n i.i.d. random variables with mean ϕ(t, s)u(t, ds), so (3.14) clearly follows from the law of large numbers. B. Proofs linked with the various examples B.1. Renewal process Proposition B.1. With the notations of Section 2, let N be a point process on R, with predictable age process (St− )t>0 , such that T0 = 0 a.s. The following statements are equivalent: (i) N+ = N ∩ (0, +∞) is a renewal process with ISI’s distribution given by some density ν : R+ → R+ . June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 31
N N (ii) N admits λ(t, Ft− ) = f (St− ) as an intensity on (0, +∞) and λ(t, Ft− ) t>0  1  ,a.s. satisfies ALλ,loc , for some f : R+ → R+ . In such a case, for all x ≥ 0, f and ν satisfy Z x • ν(x) = f (x) exp(− f (y)dy) with the convention exp(−∞) = 0, 0 Z ∞ ν(x) R • f (x) = ∞ if ν(y)dy 6= 0, else f (x) = 0. ν(y)dy x x (B.1) (B.2) Proof. For (ii) ⇒ (i). Since T0 = 0 a.s., Point (2) of Proposition B.2 given later on for the general Wold case implies that the ISI’s of N forms a Markov chain of order 0 i.e. they are i.i.d. with density given R ∞ by (B.1). For (i) ⇒ (ii). Let x0 = inf{x ≥ 0, x ν(y)dy = 0}. It may be infinite. Let us define f by (B.2) for every 0 ≤ x < x0 and let Ñ be a point process on R such Ñ Ñ that Ñ− = N− and Ñ admits λ(t, Ft− ) = f (St− ) as an intensity on (0, +∞) where Ñ (St− )t>0 is the predictable age process associated to Ñ . Applying (ii) ⇒ (i) to Ñ gives that the ISI’s of Ñ are i.i.d. with density given by ! Z x ν(x) ν(y) R∞ ν̃(x) = R ∞ exp − dy , ν(y)dy ν(z)dz 0 x y for every 0 ≤ x < x 0 and ν̃(x) = 0 for x ≥ x0 . It is clear that ν = ν̃ since the function Rx ν(y) 1 exp − 0 R ∞ dy is differentiable with derivative equal to 0. x 7→ R ∞ x ν(y)dy y ν(z)dz Since N and Ñ are renewal processes with same density ν and same first point T0 = 0, they have the same distribution. Since the intensity characterizes a point N N process, N also admits λ(t, Ft− ) = f (St− ) as an intensity on (0, +∞). Moreover,
N since N is a renewal process, it is non-explosive in finite time and so λ(t, Ft− ) t>0  1  ,a.s. satisfies ALλ,loc . B.2. Generalized Wold processes In this Section, we suppose that there exists k ≥ 0 such that the underlying point process N has intensity
N λ t, Ft− = f (St− , A1t , ..., Akt ), (B.3) where f is a function and the Ai ’s are defined by Equation (2.2). B.2.1. Markovian property and the resulting PDE Let N be a point process of intensity given by (B.3). If T−k > −∞, its associated age process (St )t can be defined up to t > T−k . Then let, for any integer i ≥ −k, Ai = Ti+1 − Ti = STi+1 − (B.4) June 9, 2015 32 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret and denote (FiA )i≥−k the natural filtration associated to (Ai )i≥−k . For any t ≥ 0, and point process Π on R2+ , let us denote Π≥t (resp. Π>t ) the restriction to R2+ (resp. (0, +∞) × R+ ) of the point process Π shifted t time units to the left on the first coordinate. That is, Π≥t (C × D) = Π((t + C) × D) for all C ∈ B(R+ ), D ∈ B(R+ ) (resp. C ∈ B((0, +∞))). Proposition B.2. Let consider k a non-negative integer, f some non negative function on Rk+1 + and N a generalized Wold process of intensity given by (B.3).Supposethat P0 is
1 ,a.s. N ) t>0 satisfies ALλ,loc . Then, such that P0 (T−k > −∞) = 1 and that λ(t, Ft−
(1) If (Xt )t≥0 = (St− , A1t , ..., Akt ) t≥0 , then for any finite non-negative stopping time τ , (Xtτ )t≥0 = (Xt+τ )t≥0 is independent of FτN− given Xτ . (2) the process (Ai )i≥1 given by (B.4) forms a Markov chain of order k with transition measure given by  Z x  ν(dx, y1 , ..., yk ) = f (x, y1 , ..., yk ) exp − f (z, y1 , ..., yk )dz dx. (B.5) 0 If T0 = 0 a.s., this holds for (Ai )i≥0 . If f is continuous then G, the infinitesimal generator of (Xt )t≥0 , is given by ∀φ ∈ C 1 (Rk+1 (Gφ)(s, a1 , ..., ak ) = + ), ∂ φ(s, a1 , ..., ak ) + f (s, a1 , ..., ak ) (φ(0, s, a1 , ..., ak−1 ) − φ(s, a1 , ..., ak )) . (B.6) ∂s Proof. First, let us show the first point of the Proposition. Let Π be such that N is the process resulting of Ogata’s thinning with Poisson measure Π. The existence of such a measure is assured by Proposition B.12. We show that for any finite stopping time τ , the process (Xtτ )t≥0 can be expressed as a function of Xτ and Π≥τ which is the restriction to R2+ of the Poisson process Π shifted τ time units to the left on the first coordinate. Let e1 = (1, 0, . . . , 0) ∈ Rk+1 . For all t ≥ 0, let Yt = Xτ + te1 and define ( ) Z Z f (Yw ) R0 = inf t ≥ 0, Π≥τ (dw, dx) = 1 . [0,t] x=0 Note that R0 may be null, in particular when τ is a jumping time of the underlying point process N . It is easy to check that R0 can be expressed as a measurable τ function of Xτ and Π≥τ . Moreover, it is clear that Xt∧R = Yt∧R0 for all t ≥ 0. 0 So, R0 can be seen as the delay until the first point of the underlying process N after time τ . Suppose that Rp , the delay until the (p + 1)th point, is constructed for some p ≥ 0 and let us show how Rp+1 can be constructed. For t ≥ Rp , let τ Zt = θ(XR )+te1 , where θ : (x1 , . . . , xk+1 ) 7→ (0, x1 , . . . , xk ) is a right shift operator p June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 33 modelling the dynamics described by (2.3). Let us define ( Rp+1 = inf Z Z f (Zw ) t > Rp , ) Π≥τ (dw, dx) = 1 . (Rp ,Rp +t] (B.7) x=0 Note that for any p ≥ 0, Rp+1 cannot be null. It is coherent with the fact that the counting process (Nt )t>0 only admits jumps with height 1. It is easy to check that τ Rp+1 can be expressed as a measurable function of θ(XR ) and Π>τ +Rp . It is also p τ clear that Xt∧Rp+1 = Zt∧Rp+1 for all t ≥ Rp . So, Rp+1 can be seen as the delay τ can be until the (p + 2)th point of the process N after time τ . By induction, XR p τ expressed as a function of Xτ and Π≥τ , and this holds for Rp+1 and XRp+1 too. To conclude, remark that the process (Xtτ )t≥0 is a measurable function of Xτ and all the Rp ’s for p ≥ 0. Thanks to the independence of the Poisson measure Π, FτN− is independent of Π≥τ . Then, since (Xtτ )t≥0 is a function of Xτ and Π≥τ , (Xtτ )t≥0 is independent of FτN− given Xτ which concludes the first point. For Point (2), fix i ≥ 1 and apply Point (1) with τ = Ti . It appears that in this case, R0 = 0 and R1 = Ai . Moreover, R1 = Ai can be expressed as a A function of θ(Xτ ) and Π>τ . However, θ(Xτ ) = (0, Ai−1 , . . . , Ai−k ) and Fi−1 ⊂ FTNi . A given Since τ = Ti , Π>τ is independent of FTNi and so Ai is independent of Fi−1 (Ai−1 , . . . , Ai−k ). That is, (Ai )i≥1 forms a Markov chain of order k. Note that if T0 = 0 a.s. (in particular it is non-negative), then one can use the previous argumentation with τ = 0 and conclude that the Markov chain starts one time step earlier, i.e. (Ai )i≥0 forms a Markov chain of order k. For (B.5),R1 = Ai , defined by (B.7), has the same distribution as the first point of a Poisson process with intensity λ(t) = f (t, Ai−1 , . . . , Ai−k ) thanks to the thinning Theorem. Hence, the transition measure of (Ai )i≥1 is given by (B.5). Now that (Xt )t≥0 is Markovian, one can compute its infinitesimal generator. Suppose that f is continuous and let φ ∈ Cb1 (Rk+1 + ), The generator of (Xt )t≥0 is defined by Gφ(s, a1 , . . . , ak ) = limh→0+ Phh−Id φ, where Ph φ (s, a1 , . . . , ak ) = E [φ (Xh )|X0 = (s, a1 , . . . , ak )]   = E φ (Xh ) 1{N ([0,h])=0} X0 = (s, a1 , . . . , ak )   +E φ (Xh ) 1{N ([0,h])>0} X0 = (s, a1 , . . . , ak ) = E0 + E>0 . The case with no jump is easy to compute, E0 = φ (s + h, a1 , . . . , ak ) (1 − f (s, a1 , . . . , ak ) h) + o(h), (B.8) thanks to the continuity of f . When h is small, the probability to have more than two jumps in [0, h] is a o(h), so the second case can be reduced to the case with June 9, 2015 34 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret exactly one random jump (namely T ),   E>0 = E φ (Xh ) 1{N ([0,h])=1} X0 = (s, a1 , . . . , ak ) + o(h)   = E φ (θ(X0 + T ) + (h − T )e1 ) 1{N ∩[0,h]={T }} X0 = (s, a1 , . . . , ak ) + o(h)   = E (φ (0, s, a1 , . . . , ak−1 ) + o(1)) 1{N ∩[0,h]={T }} X0 = (s, a1 , . . . , ak ) + o(h) = φ (0, s, a1 . . . , ak−1 ) (f (s, a1 , . . . , ak ) h) + o (h) , (B.9) thanks to the continuity of φ and f . Gathering (B.8) and (B.9) with the definition of the generator gives (B.6). B.2.2. Sketch of proof of Proposition 4.1 Let N be the point process construct by Ogata’s thinning of the Poisson process Π and Uk be as defined in Proposition 4.1. By an easy generalisation of Proposition 3.1, one can prove that on the event Ω of probability 1, where Ogata’s thinning is well defined, and where T0 < 0, Uk satisfies (PF ubini ), (4.3) and on R+ × Rk+1 + , the following system in the weak sense !   Z f (s,a1 ,...,ak ) ∂ ∂ + Uk (dt, ds, da) + Π (dt, dx) Uk (t, ds, da) = 0, ∂t ∂s x=0 ! Z Z f (s,a1 ,...,ak ) Uk (dt, 0, ds, da1 , ..., dak−1 ) = Π (dt, dx) Uk (t, ds, da) , ak ∈R x=0 with da = da1 × ... × dak and initial condition U in = δ(−T0 ,A10 ,...,Ak0 ) . Similarly to R Proposition 3.2, one can ϕ in  also prove R that for any test function  Mc,b (Rk+2 ), E ϕ(t, s, a)U (t, ds, da) and E ϕ(t, s, a)U (dt, s, da) are finite k k + and one can define uk (t, ds, da) and uk (dt, s, da) by, for all ϕ in Mc,b (Rk+2 + ), Z  Z ϕ(t, s, a)uk (t, ds, da) = E ϕ(t, s, a)Uk (t, ds, da) , for all t ≥ 0, and Z  Z ϕ(t, s, a)uk (dt, s, da) = E ϕ(t, s, a)Uk (dt, s, da) , for all s ≥ 0. Moreover, uk (t, ds, da) and uk (dt, s, da) satisfy (PF ubini ) and one can define uk (dt, ds, da) = uk (t, ds, da)dt = uk (dt, s, da)ds on R2+ , such that for any test function ϕ in Mc,b (Rk+2 + ), Z  Z ϕ(t, s, a)uk (dt, ds, da) = E ϕ(t, s, a)Uk (dt, ds, da) , quantity which is finite. The end of the proof is completely analogous to the one of Theorem 3.3. June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 35 B.3. Linear Hawkes processes B.3.1. Cluster decomposition Proposition B.3. Let g be a non negative L1loc (R+ ) function and h a non negative L1 (R+ ) function such that khk1 < 1. Then the branching point process N is defined as ∪∞ k=0 Nk the set of all the points in all generations constructed as follows: • Ancestral points are Nanc distributed as a Poisson process of intensity g; N0 := Nanc can be seen as the points of generation 0. • Conditionally to Nanc , each ancestor a ∈ Nanc gives birth, independently of anything else, to children points N1,a according to a Poisson process of intensity h(. − a); N1 = ∪a∈Nanc N1,a forms the first generation points. Then the construction is recursive in k, the number of generations: • Denoting Nk the set of points in generation k, then conditionally to Nk , each point x ∈ Nk gives birth, independently of anything else, to children points Nk+1,x according to a Poisson process of intensity h(. − x); Nk+1 = ∪x∈Nk Nk+1,x forms the points of the (k + 1)th generation. This construction ends almost surely in every finite interval. Moreover the intensity of N exists and is given by Z t− N λ(t, Ft− ) = g(t) + h(t − x)N (dx). 0 This is the cluster representation of the Hawkes process. When g ≡ ν, this has been proved in 20 . However up to our knowledge this has not been written for a general function g. Proof. First, let us fix some A > 0. The process ends up almost surely in [0, A] because there is a.s. a finite number of ancestors in [0, A]: if we consider the family of points attached to one particular ancestor, the number of points in each generation form a sub-critical Galton Watson process with reproduction distribution, a Poisson R variable with mean h < 1 and whose extinction is consequently almost sure. Next, to prove that N has intensity Z t− H(t) = g(t) + h(t − x)N (dx), 0 we exhibit a particular thinning construction, where on one hand, N is indeed a branching process by construction as defined by the proposition and, which, on the other hand, guarantees that Ogata’s thinning project the points below H(t). We can always assume that h(0) = 0, since changing the intensity of Poisson process in the Rbranching structure at one particular point has no impact. Hence H(t) = t g(t) + 0 h(t − x)N (dx). The construction is recursive in the same way. Fix some realisation Π of a Poisson process on R2+ . June 9, 2015 36 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret For Nanc , project the points below the curve t → g(t) on [0, A]. By construction, Nanc is a Poisson process of intensity g(t) on [0, A]. Note that for the identification (see Theorem B.11) we just need to do it on finite intervals and that the ancestors that may be born after time A do not have any descendants in [0, A], so we can discard them, since they do not appear in H(t), for t ≤ A. Enumerate the points in Nanc ∩ [0, A] from T1 to TN0,∞ . • The children of T1 , N1,T1 , are given by the projection of the points of Π whose ordinates are in the strip t 7→ (g(t), g(t)+h(t−T1 )]. As before, by the property of spatial independence of Π, this is a Poisson process of intensity h(. − T1 ) conditionally to Nanc . • Repeat until TN0,∞ , where N1,TN0,∞ are given by the projection of the points PN0,∞ −1 of Π whose ordinates are in the strip t 7→ (g(t) + i=1 h(t − Ti ), g(t) + PN0,∞ h(t − T )]. As before, by the property of independence of Π, this is a i i=1 Poisson process of intensity h(. − TN0,∞ ) conditionally to Nanc and because the consecutive strips do not overlap, this process is completely independent of the previous processes (N1,Ti )’s that have been constructed. Note that at the end of this first generation, N1 = ∪T ∈Nanc N1,T consists of the PN0,∞ projection of points of Π in the strip t 7→ (g(t), g(t) + i=1 R h(t − Ti )]. They PN0,∞ therefore form a Poisson process of intensity i=1 h(t − Ti ) = h(t − u)Nanc (du), conditionally to Nanc . For generation k + 1 replace in the previous construction Nanc by Nk and g(t) Pk−1 R by g(t) + j=0 h(t − u)dNj (u). Once again we end up for each point x in Nk with a process of children Nk+1,x which is a Poisson process of intensity h(t − x) conditionally to Nk and which is totally independent of the other Nk+1,y ’s. R Note also that as before, Nk+1 = ∪x∈Nk Nk+1,x is a Poisson process of intensity h(t − u)Nk (du), conditionally to N0 , ..., Nk . Hence we are indeed constructing a branching process as defined by the proposition. Because the underlying Galton Watson process ends almost surely, as shown before, it means that there exists a.s. one generation Nk∗ which will be completely empty and our recursive contruction ends up too. The main point is to realize that at the end the points in N = ∪∞ k=0 Nk are exactly the projection of the points in Π that are below ∞ Z ∞ Z t X X t 7→ g(t) + h(t − u)Nk (du) = g(t) + h(t − u)Nk (du) k=0 k=0 0 hence below Z t 7→ g(t) + t h(t − u)N (du) = H(t). 0 Moreover H(t) is FtN predictable. Therefore by Theorem B.11, N has intensity H(t), which concludes the proof. June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 37 A cluster process Nc , is aR branching process, as defined before, which admits t− Nc intensity λ(t, Ft− ) = h(t) + 0 h(t − z)Nc (dz). Its distribution only depends on the function h. It corresponds to the family generated by one ancestor at time 0 in Proposition B.3. Therefore, by Proposition B.3, a Hawkes process with empty past R t− N (N− = ∅) of intensity λ(t, Ft− ) = g(t) + 0 h(t − z)N (dz) can always be seen as the union of Nanc and of all the a + Nca for a ∈ Nanc where the Nca are i.i.d. cluster processes. For a Hawkes process N with non empty past, N− , this
is more technical. Let Nanc be a Poisson process of intensity g on R+ and NcV V ∈Nanc be a sequence of i.i.d. cluster processes associated to h. Let also ! [ V N>0 = Nanc ∪ V + Nc . (B.10) V ∈Nanc As we prove below, this represents the points in N that do not depend on N− . The points that are depending on N− are constructed as follows independently of N>0 .
T Given N− , let N1 T ∈N− denote a sequence of independent Poisson processes with
respective intensities λT (v) = h(v − T )1(0,∞) (v). Then, given N− and N1T T ∈N− ,
let NcT,V V ∈N T ,T ∈N be a sequence of i.i.d. cluster processes associated to h. The − 1 points depending on the past N− are given by the following formula as proved in the next Proposition:    [ [ N≤0 = N− ∪  N1T ∪  V + NcT,V  . (B.11) T ∈N− V ∈N1T Proposition B.4. Let N = N≤0 ∪ N>0 , where N>0 and N≤0 are given by (B.10) and (B.11). Then N is a linear HawkesR process with past given by N− and intensity t− N ) = g(t) + −∞ h(t − x)N (dx), where g and h are as in on (0, ∞) given by λ(t, Ft− Proposition B.3. Proof. Proposition B.3 yields that N>0 has intensity Z t− N>0 λN>0 (t, Ft− ) = g(t) + h(t − x)N>0 (dx), (B.12) 0 T and that, given N− , for any T ∈ N− , NH = N1T ∪ NT λNHT (t, Ft−H ) S V ∈N1T V + NcT,V  has intensity t− Z T h(t − x)NH (dx), = h(t − T ) + (B.13) 0 Moreover, all these processes are Windependent  given N− . For any t ≥ 0, one can T N≤0 NH N− note that Ft ⊂ Gt := F0 ∨ , and so N≤0 has intensity T ∈N− Ft λN≤0 (t, Gt− ) = X T ∈N− NT λNHT (t, Ft−H ) Z t− h(t − x)N≤0 (dx) = −∞ (B.14) June 9, 2015 38 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret N on (0, +∞). Since this last expression is Ft ≤0 -predictable, by page 27 in 3 , this is N also λN≤0 (t, Ft−≤0 ). Moreover, N≤0 and N>0 are independent by construction and, N≤0 for any t ≥ 0, FtN ⊂ Ft given by ∨ FtN>0 . Hence, as before, N has intensity on (0, +∞) N N>0 N ) = g(t) + λ(t, Ft− ) = λ(t, Ft−≤0 ) + λ(t, Ft− Z t− h(t − x)N (dx). −∞ B.3.2. A general result for linear Hawkes processes The following proposition is a consequence of Theorem 3.3 applied to Hawkes processes with general past N− . Proposition B.5. Using the notations of Theorem 3.3, let N be a Hawkes process with past before 0 given by N− of distribution P0 and with intensity on R+ given by Z t− N λ(t, Ft− ) = µ + h(t − x)N (dx), −∞ where µ is a positive real number and h is a non-negative function with support in R R+ such that h < 1. Suppose that P0 is such that Z 0  sup E h(t − x)N− (dx) < ∞. (B.15) t≥0 −∞ Then, the mean measure u defined in Proposition 3.2 satisfies Theorem 3.3 and R∞ moreover its integral v(t, s) := u(t, dσ) is a solution of the system (4.14)–(4.15) s µ,h where v in is the survival function of −T0 , and where Φ = Φµ,h P0 is given by ΦP0 = µ,h µ,h µ,h Φµ,h + + Φ−,P0 , with Φ+ given by (4.17) and Φ−,P0 given by, Z t−  ∀ s, t ≥ 0, Φµ,h (t, s) = E h(t − z)N (dz) N ([t − s, t)) = 0 . (B.16) ≤0 ≤0 −,P0 −∞ Moreover, (4.20) holds. B.3.3. Proof of the general result of Proposition B.5 Before proving Proposition B.5, we need some technical preliminaries. Events of the type {St− ≥ s} are equivalent to the fact that the underlying process has no point between t − s and t. Therefore, for any point process N and any real numbers t, s ≥ 0, let Et,s (N ) = {N ∩ [t − s, t) = ∅}. (B.17) Various sets Et,s (N ) are used in the sequel and the following lemma, whose proof is obvious and therefore omitted, is applied several times to those sets. June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 39 Lemma B.6. Let Y be some random variable and I(Y ) some countable set of indices depending on Y . Suppose that (Xi )i∈I(Y ) is a sequence of random variables which are independent conditionally on Y . Let A(Y ) be some event depending on Y and ∀ j ∈ I(Y ), Bj = Bj (Y, Xj ) be some event depending on Y and Xj . Then, for any i ∈ I(Y ), and for all sequence of measurable functions (fi )i∈I(Y ) such that the following quantities exist,     X X E [fi (Y, Xi )| Y, Bi ] A#B  , E fi (Y, Xi ) A#B  = E  i∈I(Y ) i∈I(Y ) E[fi (Y,Xi )1Bi | Y ] P(Bi | Y ) where E [fi (Y, Xi )| Y, Bi ] = and A#B = A(Y ) ∩ T  B . j j∈I(Y ) The following lemma is linked to Lemma 4.2. Lemma B.7. Let N be a linear Hawkes process with no pastR before time 0 (i.e. t− N ) = g(t) + 0 h(t − x)N (dx), N− = ∅) and intensity on (0, ∞) given by λ(t, Ft− where g and h are as in Proposition B.3 and let for any x, s ≥ 0   Z x   Lg,h (x) = E h(x − z)N (dz) Ex,s (N ) s 0   Gg,h (x) = P (E x,s (N )) , s Then, for any x, s ≥ 0, Lg,h s (x) = x Z
h,h h (z) + Lh,h s (z) Gs (z)g(x − z) dz, (B.18) s∧x and log(Gg,h s (x)) Z (x−s)∨0 = Gh,h s (x Z − z)g(z)dz − 0 x g(z)dz. (B.19) 0 h,h 1 ∞ In particular, (Lh,h and is a solution of (4.11)-(4.12). s , Gs ) is in L × L Proof. The statement only depends on the distribution of N . Hence, thanks to
V Proposition B.4, it is sufficient to consider N = Nanc ∪ ∪ V + N . c P V ∈Nanc  Let us show (B.18). First, let us write Lg,h s (x) = E X∈N h(x − X) Ex,s (N ) . and note that Lg,h s (x) = 0 if x ≤ s. The following decomposition holds     X X  h(x − V ) + Lg,h h(x − V − W ) Ex,s (N ) . s (x) = E V ∈Nanc W ∈NcV According to Lemma B.6 and the following decomposition, ! Ex,s (N ) = Ex,s (Nanc ) ∩ \ V ∈Nanc Ex−V,s (NcV ) , (B.20) June 9, 2015 40 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret let us denote Y = Nanc , XV = NcV and BV = Ex−V,s (NcV ) for all V ∈ Nanc . Let us fix V ∈ Nanc and compute the conditional expectation of the inner sum with respect to the filtration of Nanc which is   # " X X   h((x − V ) − W ) Ex−V,s (Nc ) E h(x − V − W ) Y, BV = E W ∈Nc W ∈NcV = Lh,h s (x − V ), (B.21) since, conditionally on Nanc , NcV has the same distribution as N c which is a linear R t− Nc Hawkes process with conditional intensity λ(t, Ft− ) = h(t) + 0 h(t − z)Nc (dz). Using the conditional independence of the cluster processes with respect to Nanc , one can apply Lemma B.6 and deduce that " # X
g,h h,h Ls (x) = E h(x − V ) + Ls (x − V ) Ex,s (N ) V ∈Nanc The following argument is inspired by Moller 28 . For every V ∈ Nanc , we say that V has mark 0 if V has no descendant or himself in [x − s, x) and mark 1 otherwise. Let 0 1 0 us denote Nanc the set of points with mark 0 and Nanc = Nanc \ Nanc . For any V ∈
h,h 0 Nanc , we have P V ∈ Nanc Nanc = Gs (x−V )1[x−s,x)c (V ), and all the marks are 1 0 are independent Poisson and Nanc chosen independently given Nanc . Hence, Nanc 0 h,h processes and the intensity of N is given by λ(v) = g(v)G anc s (x − v)1[x−s,x)c (v).  1 Moreover, the event Nanc = ∅ can be identified to Ex,s (N )and   X
1   Lg,h h(x − V ) + Lh,h s (x) = E s (x − V ) Nanc = ∅ 0 V ∈Nanc Z = x−
h,h h (x − w) + Lh,h s (x − w) g(w)Gs (x − w)1[x−s,x)c (w)dw −∞ (x−s)∨0 Z
h,h h (x − w) + Lh,h s (x − w) Gs (x − w)g(w) dw, = 0 where we used the independence between the two Poisson processes. It suffices to substitute w by z = x − w in the integral to get the desired formula. Since Gh,h is s h,h 1 bounded, it is obvious that Ls is L . Then, let us show (B.19). First note that if x < 0, Gg,h s (x) = 1. Next, following Q (B.20) one has Gg,h (x) = E 1 1 Ex,s (Nanc ) s X∈Nanc Ex−X,s (NcX ) . This is also   Y  Gg,h 1Ex−V,s (NcV )  , s (x) = E 1Nanc ∩[x−s,x)=∅ V ∈Nanc ∩[x−s,x)c  = E 1Nanc ∩[x−s,x)=∅  Y V ∈Nanc ∩[x−s,x)c  Gh,h s (x − V ) , June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 41 by conditioning with respect to Nanc . Since Nanc ∩ [x − s, x) is independent of Nanc ∩ [x − s, x)c , this gives " !# Z Z x h,h g,h g(z)dz)E exp log(Gs (x − z))Nanc (dz) . Gs (x) = exp(− [x−s,x)c x−s Rx R h,h This leads to log(Gg,h s (x)) = − x−s g(z)dz+ [x−s,x)c (Gs (x−z)−1)g(z)dz, thanks to Campbell’s Theorem 23 . Then, (B.19) clearly follows from the facts that if z > x > 0 then Gh,h s (x − z) = 1 and g(z) = 0 as soon as z < 0. Proof of Lemma 4.2 In turn, we use a Banach fixed point argument to prove that for all s ≥ 0 there exists a unique couple (Ls , Gs ) ∈ L1 (R+ )×L∞ (R+ ) solution to these equations. To do so, let us first study Equation (4.11) and define TG,s:  R Rx (x−s)∨0 L∞ (R+ ) → L∞ (R+ ) by TG,s (f )(x) := exp 0 f (x − z)h(z)dz − 0 h(z)dz . The right-hand side is well-defined since h ∈ L1 and f ∈ L∞ . Moreover we have R  Rx (x−s)∨0 R (x−s)∨0 kf kL∞ h(z)dz− h(z)dz (kf kL∞ −1) h(z)dz 0 0 0 TG,s (f )(x) ≤ e ≤e . This shows that TG,s maps the ball of radius 1 of L∞ into itself, and more precisely into the intersection of the positive cone and the ball. We distinguish two cases: Rx − If x < s, then TG,s (f )(x) = exp(− h(z)dz) for any f , thus, the unique fixed 0 Rx point is given by Gs : x 7→ exp(− h(z)dz), which does not depend on s > x. 0 − And if x > s, the functional TG,s is a k−contraction in {f ∈ L∞ (R+ ), kf kL∞ ≤ R∞ 1}, with k ≤ h(z)dz < 1, by convexity of the exponential. More precisely, using 0 that for all x, y, |ex − ey | ≤ emax(x,y) |x − y| we end up with, for kf k, kgkL∞ ≤ 1, − TG,s (f )(x) − TG,s (g)(x) ≤ e Rx 0 x−s R h(z)dz h(z)dz e0 Z ≤ kf − gkL∞ Z kf − gkL∞ (x−s) h(z)dz 0 h(z)dz. R+ Hence there exists only one fixed point Gs that we can identify with Gh,h given s in Proposition B.7 and Gh,h being a probability, G takes values in [0, 1]. s s 1 1 Analogously, we define the functional T : L (R L,s + ) → L (R+ ) by TL,s (f )(x) := Rx (h (z) + f (z)) Gs (z)h(x − z) dz, and it is easy to check that TL,s is well-defined s∧x as well. We similarly distinguish the two cases: − If x < s, then the unique fixed point is given by Ls (x) = 0. R∞ − And if x > s, thus TL,s is a k−contraction with k ≤ h(y)dy < 1 in L1 ((s, ∞)) 0 since kGs kL∞ ≤ 1 : June 9, 2015 42 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret kTL,s (f ) − TL,s (g)kL1 = R∞ Rx s
f (z) − g(z) Gs (z)h(x − z)dz dx s ≤ kGs kL∞ R∞ R∞ f (z) − g(z) h(x − z)dxdz s v = kGs kL∞ kf − gkL1 ((s,∞)) R∞ h(y)dy. 0 In the same way, there exists only one fixed point Ls = Lh,h given by Proposition s B.7. In particular Ls (x ≤ s) ≡ 0. Finally, as a consequence of Equation R (4.12) we find that if Ls is the unique fixed point of TL,s , then kLs kL1 (R+ ) ≤ ( ∞ 0 1− bounded in L1 with respect to s. h(y) dy)2 R∞ 0 h(y) dy and therefore Ls is uniformly Lemma B.8. Let N be a linear Hawkes process withR past before time 0 given by t− N N− and intensity on (0, ∞) given by λ(t, Ft− ) = µ + −∞ h(t − x)N (dx), where µ is a positive real number and h is a non-negative function with support in R+ , such 1 ,exp that ||h||L1 < 1. If the distribution of N− satisfies (B.15) then (ALλ,loc ) is satisfied.   N ) . By iProposition B.4, λ(t) = λ(t) = E λ(t, F Proof. For all t > 0, let t− h i hR R t− t− E µ + 0 h(t − x)N>0 (dx) + E −∞ h(t − x)N≤0 (dx) which is possibly infinite. Let us apply Proposition B.7 with g ≡ µ and s = 0, the choice s = 0 implying that Et,0 (N>0 ) is of probability 1. Therefore t−  Z E µ+    Z t h(t − x)N>0 (dx) = µ 1 + (h(x) + L0 (x))dx , 0 0 where (L0 , G0 = 1) is the solution of Lemma 4.2 fori s = 0, by identification of h R t− Proposition B.7. Hence E µ + 0 h(t − x)N>0 (dx) ≤ µ(1 + ||h||L1 + ||L0 ||L1 ). On the other hand, thanks to Lemma B.9, we have Z t− E −∞     Z t X  h(t − x)N≤0 (dx) = E  h(t − T ) + [h(t − x) + L0 (t − x)] h(x − T )dx  . 0 T ∈N− Since all the quantities are non negative, one can exchange all the integrals and deduce that Z t−  h(t − x)N≤0 (dx) ≤ M (1 + ||h||L1 + ||L0 ||L1 ), E −∞ with M = supt≥0 E hR 0 −∞ i h(t − x)N− (dx) which is finite by assumption. Hence, 1 ,exp λ(t) ≤ (µ + M )(1 + ||h||L1 + ||L0 ||L1 ), and therefore (ALλ,loc ) is satisfied. June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 43 Proof of Proposition B.5 First, by Proposition B.4   N E λ(t, Ft− ) St− ≥ s =   Z t− Z t− h(t − z)N≤0 (dz) Et,s (N ) h(t − z)N>0 (dz) Et,s (N ) + E µ+E Z = µ+E 0 t−  Z h(t − z)N>0 (dz) Et,s (N>0 ) +E −∞ t−  h(t − z)N≤0 (dz) Et,s (N≤0 ) −∞ 0   h N By Lemma B.7, we obtain E λ(t, Ft− ) St− ≥ s = µ + Lµ,h s (t) + Φ−,P0 (t, s). Idenand Gs = Gh,h tifying by Lemma 4.2, Ls = Lh,h s , we obtain s   h N E λ(t, Ft− ) St− ≥ s = Φµ,h + (t, s) + Φ−,P0 (t, s).   N Hence Φµ,h P0 (t, s) = E λ(t, Ft− ) St− ≥ s . Lemma B.8 ensures that the assumptions of Theorem 3.3 are fulfilled. Let u and ρµ,h 3.3. With respect to the P0 = ρλ,P0 be defined accordingly as in Theorem   N )1{St− ≥s} . The first PDE system, there are two possibilities to express E λ(t, Ft− h i one involves ρλ,P0 and is E ρµ,h (t, S )1 , whereas the second one involves t− S ≥s t− P0 µ,h Φµ,h P0 and is ΦP0 (t, s)P (St− ≥ s) . R +∞ R +∞ µ,h This leads to s ρµ,h u(t, dx), since u(t, ds) is P0 (t, x)u(t, dx) = ΦP0 (t, s) s R +∞ the distribution of St− . Let us denote v(t, s) = s u(t, dx): this relation, together with Equation (3.10) for u, immediately gives us that v satisfies Equation (4.14) R +∞ with Φ = Φµ,h . Moreover, u(t, dx) = 1, which gives us the boundary condition P0 0 in (4.15). B.3.4. Study of the general case for Φh−,P0 in Proposition B.5 Lemma B.9. Let consider h a non-negative function with support in R+ such that R h < 1, N− a point hR process on R− with distributioniP0 and N≤0 defined by (B.11). t− h If Φ−,P0 (t, s) := E −∞ h(t − z)N≤0 (dz) Et,s (N≤0 ) , for all s, t ≥ 0, then,   X h Φ−,P0 (t, s) = E  (h(t − T ) + Ks (t, T )) Et,s (N≤0 ) , (B.22) T ∈N− where Ks (t, u) is given by (4.13). Proof. Following the decomposition given in Proposition B.4, one has  X Φh−,P0 (t, s) = E  h(t − T ) T ∈N− !! + X V ∈N1T h(t − V ) + X W ∈NcT ,V h(t − V − W )  Et,s (N≤0 ) , June 9, 2015 44 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret where Et,s (N≤0 ) = Et,s (N− ) T  T 0 ∈N− Et,s (N1T ) T V 0 ∈N1T  0 Et−V 0 ,s (NcV ) . Let us fix T ∈ N− , V ∈ N1T and compute the conditional expectation of the inner sum with respect to N− and N1T . In the same way as for (B.21) we end up with   X E h(t − V − W ) N− , N1T , Et−V,s (NcT,V ) = Lh,h s (t − V ), W ∈NcT ,V since, conditionally on N− and N1T , NcT,V has the same distribution as Nc . Using the conditional independence of the cluster processes (NcT,V )V ∈N1T with respect

to N− , (N1T )T ∈N− , one can apply Lemma B.6 with Y = N− , (N1T )T ∈N− and X(T,V ) = NcT,V and deduce that     X X
h(t − T ) +  Et,s (N≤0 ) . Φh−,P0 (t, s) = E  h(t − V ) + Lh,h s (t − V ) T ∈N− V ∈N1T Let us fix T ∈ N− and compute the conditional expectation of the inner sum with respect to N− which is   X
T  Γ := E  h(t − V ) + Lh,h (B.23) s (t − V ) N− , At,s , V ∈N1T where ATt,s = Et,s (N1T ) ∩ T V 0 ∈N1T  0 Et−V 0 ,s (NcT,V ) . For every V ∈ N1T , we say that V has mark 0 if V has no descendant or himself in [t − s, t) and mark 1 otherwise. T,1 T,0 T Let us denote N1T,0 theset of points with  mark 0 and N1 = N1 \ N1 . For any V ∈ N1T , P V ∈ N1T,0 N1T = Gh,h s (t−V )1[t−s,t)c (V ) and all the marks are chosen independently given N1T . Hence, N1T,0 and N1T,1 are independent Poisson h,h processes and the intensity of N1T,0 is given n by λ(v)o= h(v − T )1[0,∞) (v)Gs (t − v)1[t−s,t)c (v). Moreover, ATt,s is the event N1T,1 = ∅ , so   n o X
T,1 h(t − V ) + Lh,h =∅  Γ = E s (t − V ) N− , N1 V ∈N1T ,0 Z t− =   h,h h(t − v) + Lh,h s (t − v) h(v − T )1[0,∞) (v)Gs (t − v)1[t−s,t)c (v)dv −∞ = Ks (t, T ). Using the independence of the cluster processes, one can apply Lemma B.6 with  T Y = N− and XT = N1 , (NcT,V )V ∈N1T and (B.22) clearly follows. Lemma B.10. Under the assumptions and notations of Proposition B.5 and Lemma 4.2, the function Φh−,P0 of Proposition B.5 can be identified with (4.18) under (A1N− ) and with (4.19) under (A2N− ) and (B.15) is satisfied in those two cases. June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 45 Proof. Using i hP E (N ) . Lemma B.9, we have Φh−,P0 (t, s) = E (h(t − T ) + K (t, T )) t,s ≤0 s T ∈N−   Under A1N− . On the one hand, for every t ≥ 0, Z 0  h(t − x)N− (dx) = E [h(t − T0 )] E −∞ Z 0 Z h(t − t0 )f0 (t0 )dt0 ≤ ||f0 ||L∞ = −∞ ∞ h(y)dy, 0 hence P0 satisfies (B.15). On to one point T0 ,  the other hand, since N− is reduced  1 h Φ−,P0 (t, s) = P E (N ) E (h(t − T0 ) + Ks (t, T0 )) 1Et,s (N≤0 ) , using the definition ( t,s ≤0 ) of the conditional expectation. First, we compute P(Et,s (N≤0 |T0 ). To do so, we use  T T0 ,V T0 Et−V,s (Nc ) the decomposition Et,s (N≤0 ) = {T0 < t − s} ∩ Et,s (N1T0 ) ∩ V ∈N 1 and the fact that, conditionally on N1T0 , for all V ∈ N1T0 , NcT0 ,V has the same distribution as Nc to deduce that   h i Y     Gs (t − V ) T0  , E 1Et,s (N≤0 ) T0 = 1T0 <t−s E 1Et,s (N T0 ) T0 E  1 T V ∈N1 0 ∩[t−s,t)c because the event Et,s (N1T0 ) involves N1T0 ∩ [t − s, t) whereas the product involves N1T0 ∩ [t − s, t)c , both of those processes being two independent Poisson processes. Their respective intensities are λ(x) = h(x − T0 )1[(t−s)∨0,t) (x) and λ(x) = h(x − T0 )1[0,(t−s)∨0) (x), so we end up with  h i  R  t  T0 E 1 T = exp − h(x − T )1 (x)dx  0 0 [0,∞)  t−s   Et,s (N1 )  R  Q    = exp − (t−s)∨0 [1 − Gs (t − x)] h(x − T0 )dx . E G (t − V ) T  s 0  0  T V ∈N 0 ∩[t−s,t)c 1 The product of these two last quantities is exactly q(t, s, T0 ) given by (4.13). Note that q(t, s, T0 ) is exactly the probability that T0 has no descendant in [t − s, t) given R 0∧(t−s) T0 . Hence, P (Et,s (N≤0 )) = −∞ q(t, s, t0 )f0 (t0 )dt0 and (4.18) clearly follows.   2 Under AN− . On the one hand, for any t ≥ 0, Z 0 E −∞  Z h(t − x)N− (dx) = E 0 −∞  Z h(t − x)αdx ≤ α ∞ h(y)dy, 0 hence P0 satisfies (B.15). On the other hand, since we are dealing with a Poisson process, we can use the same argumentation of marked Poisson processes as in the proof of Lemma B.7. For every T ∈ N− , we say that T has mark 0 if T has no 0 descendant or himself in [t − s, t) and mark 1 otherwise. Let us denote N− the set 1 0 of points with mark 0 and N− = N− \ N− . For any T ∈ N− , we have
0 P T ∈ N− N− = q(t, s, T )1[t−s,t)c (T ), June 9, 2015 46 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret 0 1 and all the marks are chosen independently given N− . Hence, N− and N− are 0 independent Poisson processes and the intensity of N− is given by λ(z) = α1z≤0 q(t, s, z)1[t−s,t)c (z)  1 Moreover, Et,s (N≤0 ) = N− = ∅ . Hence,   X 1 = ∅ (h(t − T ) + Ks (t, T )) N− Φh−,P0 (t, s) = E  0 T ∈N− 1 0 . and N− which gives (4.19) thanks to the independence of N− B.3.5. Proof of Propositions 4.3 and 4.4 Since we already proved Proposition B.5 and Lemma B.10, to obtain Proposi∞ 2 tion 4.3 it only remains to prove that Φµ,h P0 ∈ L (R+ ), to ensure uniqueness of the solution by Remark 4.1. To do so, it is easy to see that the assumption h ∈ L∞ (R+ ) combined with Lemma 4.2 giving that Gs ∈ [0, 1] and Ls ∈ L1 (R+ ) ensures that ∞ h Φµ,h + , q and Ks are in L (R+ ). In turn, this implies that Φ−,P0 in both (4.18) and ∞ (4.19) is in L (R+ ), which concludes the proof of Proposition 4.3. Proof of Proposition 4.4 The method of characteristics leads us to rewrite the solution v of (4.14)–(4.15) by defining f in ≡ v in on R+ , f in ≡ 1 on R− such that Rt  f in (s − t)e− (t−s)∨0 Φ(y,s−t+y) dy , when s ≥ t Rs v(t, s) = (B.24) − Φ(y+t−s,y) dy  in f (s − t)e (s−t)∨0 , when t ≥ s.   1 Let PM 0 be the distribution of the past given by AN− and T0 ∼ U([−M −1, −M ]). By Proposition 4.3, let vM be the solution of System (4.14)–(4.15) with Φ = Φµ,h PM 0 in and v in = vM , (i.e. the survival function of a uniform variable on [−M − 1, −M ]). ∞ Let also vM be the solution of System (4.14)–(4.15) with Φ = Φµ,h and v in ≡ 1, PM 0 and v∞ the solution of (4.21)-(4.22). Then ∞ ∞ kvM − v ∞ kL∞ ((0,T )×(0,S)) ≤ kvM − vM kL∞ ((0,T )×(0,S)) + kvM − v ∞ kL∞ ((0,T )×(0,S)) . in in By definition of vM , it is clear that vM (s) = 1 for s ≤ M, so that Formula (B.24) im∞ ∞ plies that vM (t, s) = vM (t, s) as soon as s−t ≤ M and so kvM −vM kL∞ ((0,T )×(0,S)) = 0 as soon as M ≥ S. ∞ To evaluate the distance kvM − v ∞ kL∞ ((0,T )×(0,S)) , it remains to prove that Rt h − Φ M (y,s−t+y) dy e 0 −,P0 → 1 uniformly on (0, T ) × (0, S) for any T > 0, S > 0. For this, it suffices to prove that Φh−,PM (t, s) → 0 uniformly on (0, T ) × (0, S). Since q 0 given by (4.13) takes values in [exp(−2||h||L1 ), 1], (4.18) implies R 0∧(t−s) (h(t − t0 ) + Ks (t, t0 )) 1[−M −1,−M ] (t0 )dt0 h Φ−,PM (t, s) ≤ −∞R 0∧(t−s) . 0 exp(−2||h||L1 )1[−M −1,−M ] (t0 )dt0 −∞ June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 47 Since ||Gs ||L∞ ≤ 1, Ls and h are non-negative, it is clear that Z +∞ [h(t − x) + Ls (t − x)] h(x − t0 )dx, Ks (t, t0 ) ≤ 0 and so Z −M Z +∞ Ks (t, t0 )dt0 ≤ −M −1 Z [h(t − x) + Ls (t − x)] h(x − t0 )dt0 dx −M −1 0 Z ! −M ∞ ≤ Z +∞ [h(t − x) + Ls (t − x)] dx h(y)dy ZM∞ 0 h(y)dy [||h||L1 + ||Ls ||L1 ] . ≤ M R∞ Hence, for M large enough Φh−,PM (t, s) 0 ≤ M h(y)dy [||h||L1 +||Ls ||L1 ] exp(−2||h||L1 ) → 0, uniformly 1 in (t, s) since Ls is uniformly bounded in L , which concludes the proof. B.4. Thinning The demonstration of Ogata’s thinning algorithm uses a generalization of point processes, namely the marked point processes. However, only the basic properties of simple and marked point processes are needed (see 3 for a good overview of point processes theory). Here (Ft )t>0 denotes a general filtration such that FtN ⊂ Ft for all t > 0, and not necessarily the natural one, i.e. (FtN )t>0 . Theorem B.11. Let Π be a (Ft )-Poisson process with intensity 1 on R2+ . Let 1 λ(t, Ft− ) be a non-negative (F R t )-predictable process which is Lloc a.s. and define the point process N by N (C) = C×R+ 1[0,λ(t,Ft− )] (z) Π (dt × dz) , for all C ∈ B (R+ ). Then
N admits λ(t, Ft− ) as a (Ft )-predictable intensity. Moreover, if λ is in fact
N N ), then N admits λ(t, Ft− ) as a FtN FtN -predictable, i.e. λ(t, Ft− ) = λ(t, Ft− predictable intensity. Proof. The goal is to apply the martingale characterization of the intensity (Chapter II, Theorem 9 in 3 ). We cannot consider Π as a point process on R+ marked in R+ (in particular, the point with the smallest abscissa cannot be defined). However, for every k ∈ N, we can define RΠ(k) , the restriction of Π to the points with ordinate smaller than k, by Π(k) (C) = C Π (dt × dz) for all C ∈ B (R+ × [0, k]). Then Π(k) can be seen as a point process on R+ marked in Ek := [0, k] with intensity kernel 1.dz with respect to (Ft ). In the same way, we define N (k) by Z N (k) (C) = 1z∈[0,λ(t,Ft− )] Π(k) (dt × dz) for all C ∈ B (R+ ) . C×R+ Let P(Ft ) be the predictable σ-algebra (see page 8 of 3 ). Let us denote Ek = B ([0, k]) and P̃k (Ft ) = P (Ft ) ⊗ Ek the associated marked predictable σ-algebra. June 9, 2015 48 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret For any fixed z in E, {(u, ω) ∈ R+ × Ω such that λ(u, Fu− ) (ω) ≥ z} ∈ P (Ft ) since λ is predictable. If Γk = {(u, ω, z) ∈ R+ × Ω × Ek , λ(u, Fu− ) (ω) ≥ z}, then    1 Γk = ∩ ∗ ∪ {(u, ω) ∈ R+ × Ω, λ(u, Fu− ) (ω) ≥ q} × 0, q + ∩ Ek . n∈N q∈Q+ n So, Γk ∈ P̃k (Ft ) and 1z∈[0,λ(u,Fu− )]∩Ek is P˜k (Ft )-measurable. Hence, one can apply the Integration Theorem (Chapter VIII, Corollary 4 in 3 ). So, Z t Z  1z∈[0,λ(u,Fu− )] M̄ (k) (du × dz) (Xt )t≥0 := is a (Ft )-local martingale 0 where M̄ (k) Ek t≥0 (k) (du × dz) = Π (du × dz) − dzdu. In fact, Z t (k) Xt = Nt − min (λ(u, Fu− ), k) du. 0 Rt (k) Nt Yet, (respectively 0 min (λ(u, Fu− ), k) du) is non-decreasingly converging Rt towards Nt (resp. 0 λ(u, Fu− )du). Both of the limits are finite a.s. thanks to the local integrability of the (see page  27 of 3 ). Thanks to monotone conver intensity Rt gence we deduce that Nt − 0 λ(u, Fu− )du is a (Ft )-local martingale. Then, t≥0 thanks to the martingale characterization of the intensity, Nt admits λ(t, Ft− ) as an (Ft )-intensity. The last point of the Theorem
is a reduction of the filtration. N ), it is a fortiori FtN -progressive and the desired result Since λ(t, Ft− ) = λ(t, Ft− follows (see page 27 in 3 ). This final result can be found in 4 . Proposition B.12 (Inversion Theorem).
Let N = {Tn }n>0 be a non explosive point process on R+ with FtN -predictable N intensity λt = λ(t, Ft− ). Let {Un }n>0 be a sequence of i.i.d. random variables with N uniform distribution on [0, 1]. Moreover, suppose that they are independent of F∞ . Denote Gt = σ (Un , Tn ≤ t). Let N̂ be an homogeneous Poisson process with intensity 1 on R2+ independent of F∞ ∨ G∞ . Define a point process N̄ on R2+ by Z Z X
N̄ ((a, b] × A) = 1(a,b] (Tn ) 1A Un λ(Tn , FTNn − ) + N̂ (dt × dz) (a,b] n>0 N )] A−[0,λ(t,Ft− for every 0 ≤ a < b and A ⊂ R+ . Then, N̄ is an homogeneous Poisson process on R2+ with intensity 1 with respect   to the filtration (Ht )t≥0 = Ft ∨ Gt ∨ FtN̂ . t≥0 References 1. M. Bossy, N. Champagnat, et al. Markov processes and parabolic partial differential equations. Encyclopedia of Quantitative Finance, pages 1142–1159, 2010. 2. O. Boxma, D. Perry, W. Stadje, and S. Zacks. A markovian growth-collapse model. Advances in applied probability, pages 221–243, 2006. June 9, 2015 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 Microscopic approach of a time elapsed neural model 49 3. P. Brémaud. Point processes and queues. Springer-Verlag, New York, 1981. Martingale dynamics, Springer Series in Statistics. 4. P. Brémaud and L. Massoulié. Stability of nonlinear Hawkes processes. The Annals of Probability, 24(3):1563–1588, 1996. 5. D.R. Brillinger, H.L. Bryant, and J.P. Segundo. Identification of synaptic interactions. Biol. Cybernetics, 22:213–228, 1976. 6. E.N. Brown, R. Barbieri, V. Ventura, R.E. Kass, and L.M. Frank. The time rescaling theorem and its application to neural spike train analysis. Neural Computation, 14(2):325–346, 2002. 7. N. Brunel. Dynamics of sparsely connected networks of excitatory and inhibitory spiking neurons. Journal of Computational Neuroscience, 8:183–208, 2000. 8. N. Brunel and V. Hakim. Fast global oscillations in networks of integrate-and-fire neurons with low firing rates. Neural Computation, 11:1621–1671, 1999. 9. M. J Cáceres, J. A Carrillo, and Benoı̂t Perthame. Analysis of nonlinear noisy integrate&fire neuron models: blow-up and steady states. The Journal of Mathematical Neuroscience, 1(1):1–33, 2011. 10. M. J. Cáceres, J. A. Carrillo, and L. Tao. A numerical solver for a nonlinear fokker-planck equation representation of neuronal network dynamics. J. Comp. Phys., 230:1084–1099, 2011. 11. M. J Cáceres and B. Perthame. Beyond blow-up in excitatory integrate and fire neuronal networks: refractory period and spontaneous activity. Journal of theoretical biology, 350:81–89, 2014. 12. José A. Cañizo, José A. Carrillo, and Sı́lvia Cuadrado. Measure solutions for some models in population dynamics. Acta Appl. Math., 123:141–156, 2013. 13. E.S. Chornoboy, L.P. Schramm, and A.F. Karr. Maximum likelihood identification of neural point process systems. Biol. Cybernetics, 59:265–275, 1988. 14. B. Cloez. Limit theorems for some branching measure-valued processes. arXiv:1106.0660v2, 2012. 15. A. Compte, N. Brunel, P. S. Goldman-Rakic, and X.-J. Wang. Synaptic mechanisms and network dynamics underlying spatial working memory in a cortical network model. Cerebral Cortex 10, 10:910–923, 2000. 16. D. J. Daley and D. Vere-Jones. An introduction to the theory of point processes, volume 2. Springer, 1988. 17. F. Delarue, J. Inglis, S. Rubenthaler, and E Tanré. Global solvability of a networked integrate-and-fire model of McKean-Vlasov type. Annals of Applied Probability, 2012. to appear. 18. S. Ditlevsen and A. Samson. Estimation in the partially observed stochastic morrislecar neuronal model with particle filter and stochastic approximation methods. Annals of Applied Statistics, to appear. 19. M. Doumic, M. Hoffmann, N. Krell, and L. Robert. Statistical estimation of a growthfragmentation model observed on a genealogical tree. Bernoulli, in press, 2014. 20. A. G. Hawkes and D. Oakes. A cluster process representation of a self-exciting process. Journal of Applied Probability, 11(3):493–503, 1974. 21. P. Jahn, R. W. Berg, J. Hounsgaard, and S. Ditlevsen. Motoneuron membrane potentials follow a time inhomogeneous jump diffusion process. J. Comput. Neurosci., 31:563–579, 2011. 22. P. Jahn, R.W. Berg, J. Hounsgaard, and S. Ditlevsen. Motoneuron membrane potentials follow a time inhomogeneous jump diffusion process. Journal of Computational Neuroscience, 31:563–579, 2011. 23. J.F.C. Kingman. Poisson processes. Oxford Science Publications, 1993. June 9, 2015 50 0:43 WSPC/INSTRUCTION FILE PDE˙Hawkes˙Marie11 J. Chevallier, M. Cáceres, M. Doumic, P. Reynaud-Bouret 24. C.D. Lai. An example of Wold’s Point Processes with Markov-Dependent Intervals. Journal of Applied Probability, 15(4):748–758, 1978. 25. P. A. W. Lewis and G. S. Shedler. Simulation of nonhomogeneous Poisson processes by thinning. Naval Research Logistics Quarterly, 26(3):403–413, 1979. 26. M. Mattia and P. Del Giudice. Population dynamics of interacting spiking neurons. Phys. Rev. E, 66:051917, 2002. 27. P. Mattila. Geometry of sets and measures in Euclidean spaces: fractals and rectifiability. Number 44 in Cambridge studies in advanced mathematics. Cambridge University Press, 1999. 28. J. Møller and J. G. Rasmussen. Perfect simulation of Hawkes processes. Advances in Applied Probability, pages 629–646, 2005. 29. Y. Ogata. On Lewis’ simulation method for point processes. IEEE Transactions on Information Theory, 27(1):23–31, 1981. 30. A Omurtag, Knight B. W., and L. Sirovich. On the simulation of large populations of neurons. J. Comp. Neurosci., 8:51–63, 2000. 31. K. Pakdaman, B. Perthame, and D. Salort. Dynamics of a structured neuron population. Nonlinearity, 23(1):55–75, 2010. 32. K. Pakdaman, B. Perthame, and D. Salort. Relaxation and self-sustained oscillations in the time elapsed neuron network model. SIAM Journal on Applied Mathematics, 73(3):1260–1279, 2013. 33. K. Pakdaman, B. Perthame, and D. Salort. Adaptation and fatigue model for neuron networks and large time asymptotics in a nonlinear fragmentation equation. The Journal of Mathematical Neuroscience, 4(14):1–26, 2014. 34. B. Perthame. Transport equations in biology. Frontiers in Mathematics. Birkhäuser Verlag, Basel, 2007. 35. J. Pham, K. Pakdaman, J. Champagnat, and J.-F. Vibert. Activity in sparsely connected excitatory neural networks: effect of connectivity. Neural Networks, 11(3):415– 434, 1998. 36. J.W. Pillow, J. Shlens, L. Paninski, A. Sher, A.M. Litke, E.J. Chichilnisky, and E.P. Simoncelli. Spatio-temporal correlations and visual signalling in a complete neuronal population. Nature, 454:995–999, 2008. 37. G. Pipa, S. Grün, and C. van Vreeswijk. Impact of spike train autostructure on probability distribution of joint spike events. Neural Computation, 25:1123–1163, 2013. 38. C. Pouzat and A. Chaffiol. Automatic spike train analysis and report generation. an implementation with R, R2HTML and STAR. Journal of Neuroscience Methods, pages 119–144, 2009. 39. A. Renart, N. Brunel, and X.-J. Wang. Mean-field theory of irregularly spiking neuronal populations and working memory in recurrent cortical networks. In Jianfeng Feng, editor, Computational Neuroscience: A comprehensive approach. Chapman & Hall/CRC Mathematical Biology and Medicine Series, 2004. 40. P. Reynaud-Bouret, V. Rivoirard, F. Grammont, and C. Tuleau-Malot. Goodness-offit tests and nonparametric adaptive estimation for spike train analysis. The Journal of Mathematical Neuroscience, 4(3), 2014. 41. P. Reynaud-Bouret, V. Rivoirard, and C. Tuleau-Malot. Inference of functional connectivity in neurosciences via Hawkes processes. In 1st IEEE Global Conference on Signal and Information Processing, 2013, Austin Texas. 42. L. Sirovich, A Omurtag, and K. Lubliner. Dynamics of neural populations: Stability and synchrony. Network: Computation in Neural Systems, 17:3–29, 2006. 43. V. Ventura, R. Carta, R.E. Kass, S.N. Gettner, and C.R. Olson. Statistical analysis of temporal evolution in single-neuron firing rates. Biostatistics, 3(1):1–20, 2002. 

F -PURE THRESHOLDS OF HOMOGENEOUS POLYNOMIALS arXiv:1404.3772v1 [math.AC] 14 Apr 2014 DANIEL J. HERNÁNDEZ, LUIS NÚÑEZ-BETANCOURT, EMILY E. WITT, AND WENLIANG ZHANG A BSTRACT. In this article, we investigate F -pure thresholds of polynomials that are homogeneous under some N-grading, and have an isolated singularity at the origin. We characterize these invariants in terms of the base p expansion of the corresponding log canonical threshold. As an application, we are able to make precise some bounds on the difference between F -pure and log canonical thresholds established by Mustaţă and the fourth author. We also examine the set of primes for which the F -pure and log canonical threshold of a polynomial must differ. Moreover, we obtain results in special cases on the ACC conjecture for F -pure thresholds, and on the upper semi-continuity property for the F -pure threshold function. 1. I NTRODUCTION The goal of this article is to investigate F -pure thresholds, and further study their relation with log canonical thresholds. The F -pure threshold, first defined in [TW04], is an invariant of singularities in positive characteristic defined via splitting conditions and the Frobenius (or pth -power) endomorphism. Though F -pure thresholds may be defined more generally, we will only consider F -pure thresholds of polynomials over fields of prime characteristic, and thus follow the treatment given in [MTW05]. Given such a polynomial f , the F -pure threshold of f , denoted by fpt (f ), is always a rational number in (0, 1], with smaller values corresponding to “worse” singularities [BMS08, BMS09, BSTZ09]. The log canonical threshold of a polynomial fQ over Q, denoted lct (fQ ), is an important invariant of singularities of fQ , and can be defined via integrability conditions, or more generally, via resolution of singularities. Like the F -pure threshold, lct (fQ ) is also a rational number contained in (0, 1]; see [BL04] for more on this (and related) invariants. In fact, the connections between F -pure and log canonical thresholds run far deeper: As any ab ∈ Q determines a well-defined element of Fp whenever p ∤ b, we may reduce the coefficients of fQ modulo p ≫ 0 to obtain polynomials fp over Fp . Amazingly, the F -pure thresholds of these so-called characteristic p models of fQ are related to the log canonical threshold of fQ as follows [MTW05, Theorems 3.3 and 3.4]: (1.0.1) fpt (fp ) ≤ lct (fQ ) for all p ≫ 0 and lim fpt (fp ) = lct (fQ ) . p→∞ In this article, we will not need to refer to the definition of lct (fQ ) via resolutions of singularities, and instead take the limit appearing in (1.0.1) as our definition of lct (fQ ). Via reduction to characteristic p > 0, one may reduce polynomials (and more generally, ideals of finite type algebras) over any field of characteristic zero to characteristic p ≫ 0 (e.g., see [Smi97b]). Moreover, the relations in (1.0.1) are just two of many deep connections between invariants of characteristic p models defined via the Frobenius endomorphism, and invariants of the original characteristic zero object that are often defined via resolution of singularities. For more in this direction, see, for example, [MTW05, BMS06, Smi00, Smi97a, Har98, HW02, HY03, Tak04, Sch07, BST, STZ12]. Motivated by the behavior exhibited when fQ defines an elliptic curve over Q, it is conjectured that for any polynomial fQ over Q, there exist infinitely many primes for which fpt (fp ) equals 1 lct (fQ ) [MTW05, Conjecture 3.6]. This conjecture, along with other characteristic zero considerations, has fueled interests in understanding various properties of fpt (fp ). In particular, arithmetic properties of the denominator of fpt (fp ) have recently been investigated, most notably by Schwede (e.g., see [Sch08]). Assuming fpt (fp ) 6= lct (fQ ), Schwede has asked when p must divide the denominator of fpt (fp ), and the first author has asked when the denominator of fpt (fp ) must be a power of p, and more specifically, when fpt (fp ) must be a truncation of lct (fQ ).1 Recall that, P given the unique non-terminating (base p) expansion lct (fQ ) = e≥1 λ(e) · p−e ∈ (0, 1], we call P (e) · p−e for some L ≥ 1. fpt (fp ) a truncation of lct (fQ ) (base p) if fpt (fp ) = L e=1 λ In this paper, we study F -pure thresholds associated to polynomials that are homogeneous under some (possibly, non-standard) N-grading, and that have an isolated singularity. The F purity of such polynomials was originally investigated by Fedder (e.g., see [Fed83, Lemma 2.3 and Theorem 2.5]), and more recently, by Bhatt and Singh, who showed the following: Given a (standard-graded) homogeneous polynomial f over Fp of degree n in n variables with an isolated singularity at the origin, if p ≫ 0, then fpt (fp ) = 1 − Ap for some integer 0 ≤ A ≤ n − 2. Bhatt and Singh also show that, if f is (standard-graded) homogeneous of arbitrary degree with an isolated singularity at the origin, and if fpt (fp ) 6= lct (fQ ), then the denominator of fpt (fp ) is a power of p whenever p ≫ 0 [BS, Theorem 1.1 and Proposition 5.4]. We combine a generalization of the methods in [BS] with a careful study of base p expansions to obtain our main result, Theorem 3.5, which characterizes F -pure thresholds of polynomials with an isolated singularity at the origin that are homogeneous under some N-grading. Our result states that such F -pure thresholds must have a certain (restrictive) form; in particular, it confirms that the denominator of fpt (fp ) is a power of p whenever fpt (fp ) 6= lct (fQ ) for this larger class of polynomials. Notably, the result also gives a bound for the power of p appearing in the denominator of fpt (fp ) for p ≫ 0. To minimize technicalities, we omit the statement of Theorem 3.5, and instead discuss the two variable case, where our main result takes the following concrete form; note that in what follows, we use Jac (f ) to denote the ideal generated by the partial derivatives of a polynomial f , and ord(p, b) to denote the least positive integer k such that pk ≡ 1 mod b. Theorem A p (cf. Theorem 4.4). Fix an N-grading on Fp [x, y], and consider a homogeneous polynoxy mial f with Jac (f ) = (x, y) such that deg f ≥ deg xy. If p ∤ deg f and fpt (f ) 6= deg deg f , then q L y p deg xy % deg f deg xy fpt (f ) = − for some integer 1 ≤ L ≤ ord(p, deg f ), deg f pL deg f q y where apL % b denotes the least positive residue of apL modulo b. In fact, we are able to give a slightly more refined description of the F -pure threshold, even in the two variable case; we refer the reader to Theorem 4.4 for the detailed statement. Moreover, we may recast Theorem A as a theorem relating F -pure and log canonical thresholds: If fQ ∈ Q[x, y] is a homogenous and satisfies the conditions appearing in Theorem A (i.e., deg fQ ≥ deg xy and p xy (x, y) = Jac (fQ )), then it is well-known (e.g., see Theorem 6.2) that lct (fQ ) = deg deg f . Substituting this identity into Theorem A leads to a description of fpt (fp ) in terms of lct (fQ ), and in fact is enough to show that fpt (fp ) is a truncation of lct (fQ ) (e.g., see Lemma 2.5). Though the situation is more subtle, many of the properties highlighted by Theorem A and the subsequent discussion hold in general (after some slight modifications); we refer the reader to Theorem 3.5 for a detailed description of F -pure thresholds in higher dimensions. Moreover, motivated by (the bounds for L appearing in) Theorem A, one may ask whether there always exists a (small) finite list of possible values for F -pure thresholds, say, as a function of the class of p modulo deg f . This question turns out to have a positive answer for homogeneous polynomials with 1https://sites.google.com/site/computingfinvariantsworkshop/open-questions 2 isolated singularities. Furthermore, these lists can be minimal, and strikingly, can even precisely determine fpt (fp ). For examples of such lists, see Examples 4.6, 4.7, and 4.9. The remaining results in this article are all applications of our description of F -pure thresholds. The first such application concerns uniform bounds for the difference between log canonical and F -pure thresholds. We recall the following result, due to Mustaţă and the fourth author: Given a polynomial fQ over Q, there exist constants C ∈ R>0 and N ∈ N (depending only on fQ ) such that C 1 ≤ lct (fQ ) − fpt (fp ) ≤ N p p whenever fpt (fp ) 6= lct (fQ ) and p ≫ 0 [MZ, Corollaries 3.5 and 4.5]. We stress that the preceding result applies to an arbitrary polynomial, and that the constants C and N are not explicitly stated as functions of fQ . In the special case of a homogeneous polynomial with an isolated singularity at the origin, we give a new proof of this result that makes explicit one optimal choice of constants. Theorem B ( cf. Theorem 6.2). Suppose fQ ∈ Q[x1 , · · · , xn ] is homogeneous under some N-grading with an isolated singularity at the origin, and write the rational number lct (fQ ) = ab in lowest terms. If p ≫ 0, then either fpt (fp ) = lct (fQ ), or b−1 ≤ lct (fQ ) − fpt (fp ) ≤ pord(p,b) Moreover, these bounds are sharp (see Remark 6.4). n − 1 − b−1 . p Much of the focus of this article is on studying the form of the F -pure threshold when it differs from the log canonical threshold. In Section 6.2, we give a simple criterion that, when satisfied, guarantees that the F -pure and log canonical threshold must differ. The main result of this section, Proposition 6.7, holds quite generally; that is, it can be applied to polynomials that are neither homogeneous, nor have an isolated singularity. Moreover, the proof of this result is elementary, and is based upon the fact that the base p expansion of an F -pure threshold must satisfy certain rigid conditions, as was observed in [BMS09, Her12] Theorem C (cf. Proposition 6.7). Let fQ denote any polynomial over Q, and write lct (fQ ) = ab in lowest terms. If a 6= 1, then the set of primes for which lct (fQ ) is not an F -pure threshold (of any polynomial) is infinite, and contains all primes p such that pe · a ≡ 1 mod b for some e ≥ 1. In 1 , particular, the density of the set of primes {p : fpt (fp ) 6= lct (fQ )} is greater than or equal to φ(b) where φ denotes Euler’s phi function. As a further application of our main theorem, we are also able to construct a large class of polynomials fQ over Q for which the density of the set {p : fpt (fp ) 6= lct (fQ )} is larger than any prescribed bound between zero and one. Theorem D (cf. Example 6.8). For every ε > 0, there exists an integer n with the following property: If fQ ∈ Q[x1 , . . . , xn−1 ] is homogeneous (under the standard grading) of degree n with an isolated singularity at the origin, then the density of the set of primes {p : fpt (fp ) 6= lct (fQ )} is greater than 1 − ε. The remaining applications deal with another connection between F -pure and log canonical thresholds: Motivated by results in characteristic zero, it was conjectured in [BMS09, Conjecture 4.4] that the set of all F -pure thresholds of polynomials in a (fixed) polynomial ring over a field of characteristic p > 0 satisfies the ascending chain condition (ACC), i.e., contains no strictly increasing sequences. In Proposition 7.3, we prove that a restricted set of F -pure thresholds satisfies ACC. Though the characteristic zero analog of Proposition 7.3 (that is, the statement obtained by replacing “Fp ” with “Q” and “F -pure threshold” with “log canonical threshold,” as appropriate) is obvious, our result relies strongly on the description of F -pure thresholds given in Theorem 3.5. 3 Finally, as detailed in [BMS09, Remark 4.5], the ACC conjecture for F -pure thresholds predicts that for any polynomial f ∈ Fp [x1 , · · · , xn ], there exists an integer N (which may depend on f ) such that fpt (f ) ≥ fpt (f + g) for all g ∈ (x1 , · · · , xn )N . In our final application, we are able to confirm this prediction in the following special case. Theorem E (cf. Proposition 7.10). Suppose that f ∈ Fp [x1 , · · · , xn ] is homogeneous under some p P deg xi . Then fpt (f ) = fpt (f + g) for N-grading such that Jac (fP) = (x1 , . . . , xn ) and deg f ≥ n deg f − deg x +1 i each g ∈ (x1 , · · · , xn ) . Notation. Throughout this article, p denotes a prime number and Fp denotes the field with p e elements. For every ideal I of a ring of characteristic p > 0, and every e ≥ 1, I [p ] denotes the  e eth Frobenius power of I, the ideal generated by the set gp : g ∈ I . For a real number a, ⌈a⌉ (respectively, ⌊a⌋) denotes the least integer that is greater than or equal to (respectively, greatest integer less or equal to) a. Acknowledgements. The authors are indebted to Bhargav Bhatt and Anurag Singh; the latter shared ideas on their joint work during the Midwest Commutative Algebra and Geometry Conference at Purdue University in 2011 that would eventually form the foundation of our approach. We would also like to thank Benjamin Weiss and Karen Smith for their comments on an earlier draft. The first author gratefully acknowledges support from the Ford Foundation (FF) through a FF Postdoctoral Fellowship. The second author thanks the National Council of Science and Technology (CONACyT) of Mexico for support through Grant 210916. The fourth author was partially supported by NSF grants DMS #1247354 and DMS #1068946, and a Nebraska EPSCoR First Award. This collaboration began during visits supported by a travel grant from the AMS Mathematical Research Communities 2010 Commutative Algebra program. Finally, much of the authors’ collaborations took place at the University of Michigan, the University of Minnesota, and the Mathematical Sciences Research Institute; we thank these institutions for their hospitality. 2. B ASICS OF BASE p EXPANSIONS Definition 2.1. Given α ∈ (0, 1], there exist unique integers α(e) for every e ≥ 1 such that 0 ≤ P α(e) ≤ p − 1, α = e≥1 α(e) · p−e , and such that the integers α(e) are not all eventually zero. We call P α(e) the eth digit of α (base p), and we call the expression α = e≥1 α(e) · p−e the non-terminating expansion of α (base p). Definition 2.2. Let α ∈ (0, 1], and fix e ≥ 1. We call hαie := α(1) · p−1 + · · · + α(e) · p−e the eth truncation of α (base p). We adopt the convention that hαi0 = 0 and hαi∞ = α. Notation 2.3. We adopt notation analogous to the standard decimal notation, using “ : ” to distinguish between consecutive digits. For example, we often write hαie = . α(1) : α(2) : · · · : α(e) (base p). Convention 2.4. Given a natural number b > 0 and an integer m, Jm % bK denotes the least positive residue of m modulo b. In particular, we have  that 1 q≤ kJm %ybK ≤ b for all m ∈ Z. Moreover, if p and b are relatively prime, ord(p, b) = min k ≥ 1 : p % b = 1 , which we call the order of p modulo b. In particular, note that ord(p, 1) = 1. Lemma 2.5. Fix λ ∈ (0, 1] ∩ Q. If we write λ = ab , not necessarily in lowest terms, then q e−1 y ap % b · p − Jape % bK Jape % bK (e) and hλie = λ − . λ = b bpe Note that it is important to keep in mind Convention 2.4 when interpreting these identities. 4 Proof. Since λ(e) = pe (hλie − hλie−1 ), the first identity follows from the second. Setting δ = λ − hλie and multiplying both sides of the equality ab = λ = hλie + δ by bpe shows that ape = bpe hλie + bpe δ. As 0 < δ ≤ p−e and pe hλie ∈ N, it follows that bpe δ is the least positive residue of ape modulo b. Finally, substituting δ = λ − hλie into bpe δ = Jape % bK establishes the second identity.  We gather some of the important basic properties of base p expansions below. Lemma 2.6. Fix α and β in [0, 1]. (1) α ≤ β if and only if hαie ≤ hβie for all e ≥ 1; if α < β, then these inequalities are strict for e ≫ 0. (2) If (ps − 1) · α ∈ N, then the base p expansion of α is periodic, with period dividing s. In particular, if λ = ab with p ∤ b, then the base p expansion of λ is periodic with period equal to ord(p, b). (3) Suppose λ = ab with p ∤ b. If s = ord(p, b), then for all k ≥ 1, pks · hλiks = (pks − 1) · λ. Proof. (1) follows by definition; (2) follows immediately from Lemma 2.5; (3) follows from (2).  Lemma 2.7. Consider α < β in (0, 1], and set ∆e := pe hβie − pe hαie . Note that, by Lemma 2.6, the integer ℓ = min {e : ∆e ≥ 1} is well-defined. Moreover, the following hold: (1) The sequence {∆e }e≥1 is non-negative, non-decreasing, and unbounded. (2) Suppose β = ab with p ∤ b. If s = ord(p, b), then ∆ℓ+s+k ≥ pk + 1 for every k ≥ 0. Proof. We first observe that the following recursion holds. ∆e+1 = p · ∆e + β (e+1) − α(e+1) for every e ≥ 0. (2.0.2) Setting e = ℓ in (2.0.2) and noting that ∆ℓ ≥ 1 shows that ∆ℓ+1 = p · ∆ℓ + β (ℓ+1) − α(ℓ+1) = (p − 1) · ∆ℓ + ∆ℓ + β (ℓ+1) − α(ℓ+1) ≥ (p − 1) · 1 + ∆ℓ + β (ℓ+1) − α(ℓ+1) ≥ ∆ℓ + β (ℓ+1) . Furthermore, an induction on e ≥ ℓ shows that ∆e+1 ≥ ∆e + β (e+1) for every e ≥ ℓ. (2.0.3) Thus, {∆e }e≥1 is non-decreasing, and as we consider non-terminating expansions, β (e) 6= 0 for infinitely many e, so that (2.0.3) also shows that ∆e+1 > ∆e for infinitely many e. We conclude that {∆e }e≥1 is unbounded, and it remains to establish (2). By definition, β (ℓ) − α(ℓ) = ∆ℓ ≥ 1, and hence β (ℓ) ≥ 1. In fact, setting s = ord(p, b), Lemma 2.6 states that β (ℓ+s) = β (ℓ) ≥ 1, and applying (2.0.3) with e = ℓ + s − 1 then shows that ∆ℓ+s ≥ ∆ℓ+s−1 + β (ℓ+s) ≥ 2. Hence, (2) holds for k = 0. Utilizing (2.0.2), an induction on k completes the proof. 3. F - PURE THRESHOLDS OF HOMOGENEOUS POLYNOMIALS : A  DISCUSSION We adopt the following convention from this point onward. Convention 3.1. Throughout this article, L will denote a field of characteristic p > 0, and m will denote the ideal generated by the variables in R = L[x1 , · · · , xn ]. 5 Definition 3.2. Consider a polynomial f ∈ m, and for every e ≥ 1, set o n e νf (pe ) = max N : f N ∈ / m[p ] . An important property of these integers is that {p−e · νf (pe )}e≥1 is a non-decreasing sequence contained in the open unit interval [MTW05]. Consequently, the limit νf (pe ) ∈ (0, 1] e→∞ pe fpt (f ) := lim exists, and is called the F -pure threshold of f . The following illustrates important properties of F -pure thresholds; we refer the reader to [MTW05, Proposition 1.9] or [Her12, Key Lemma 3.1] for a proof of the first, and [Her12, Corollary 4.1] for a proof of the second. Proposition 3.3. Consider a polynomial f contained in m. (1) The base p expansion of the F -pure threshold determines {νf (pe )}e≥1 ; more precisely, νf (pe ) = pe · hfpt (f )ie for every e ≥ 1. (2) The F -pure threshold is bounded above by the rational numbers determined by its trailing digits (base p); more precisely, fpt (f ) is less than or equal to . fpt (f )(s) : fpt (f )(s+1) : · · · : fpt (f )(s+k) : · · · (base p) for every s ≥ 1. 3.1. A discussion of the main results. In this subsection, we gather the main results of this article. Note that the proofs of these results appear in Section 5. Convention 3.4. Given a polynomial f , we use Jac (f ) to denote the ideal of R generated by the partial derivatives of f . If f is homogeneous under some N-grading on R, each partial derivative ∂i (f ) of f is also homogeneous, and if ∂i (f ) 6= 0, then deg ∂i (f ) = deg f − deg xi . Furthermore, if p ∤ deg f , then Euler’s relation X deg f · f = deg xi · xi · ∂i (f ) shows that f ∈ Jac (f ). Thus, if p ∤ deg(f ) and L is perfect, the Jacobian criterion states that p Jac (f ) = m if and only if f has an isolated singularity at the origin. Theorem 3.5. Fix an N-grading on L[x1 , · · · ,oxn ]. Consider a homogeneous polynomial f with nP p deg xi a Jac (f ) = m, and write λ := min deg f , 1 = b in lowest terms. (1) If fpt (f ) 6= λ, then fpt (f ) = λ − q = hλiL − for some pair (L, E) ∈ N2 ! y apL % b + bE bpL E pL with L ≥ 1 and 0 ≤ E ≤ n − 1 −  J apL % b K+a b  . (2) If p > (n − 2) · b and p ∤ b, then 1 ≤ L ≤ ord(p, b); note that ord(p, 1) = 1. (3) If p > (n − 2) · b and p > b, then a < Jape % bK for all 1 ≤ e ≤ L − 1. (4) If p > (n − 1) · b, then there exists a unique pair (L, E) satisfying the conclusions of (1). We postpone the proof of Theorem 3.5 to Subsection 5.2. The remainder of this subsection is focused on parsing the statement of Theorem 3.5, and presenting some related results. The reader interested in seeing examples should consult Section 4. 6 Remark 3.6 (Two points of view). Each of the two descriptions of fpt (f ) in Theorem 3.5, which are equivalent by Lemma 2.5, are useful in their own right. For example, the first description plays a key role in Section 4. On the other hand, the second description makes it clear that either fpt (f ) = λ, or fpt (f ) is a rational number whose denominator is a power of p, and further, describes how “far” fpt (f ) is from being a truncation of λ; these observations allow us to address the questions of Schwede and of the first author noted in the introduction. The second point of Theorem 3.5 also immediately gives a bound on the power of p appearing in the denominator of fpt (f ) whenever fpt (f ) 6= λ and p ≫ 0. For emphasis, we record this bound in the following corollary. Corollary 3.7. In the context of Theorem 3.5, if fpt (f ) 6= λ, and both p > (n − 2) · b and p ∤ b, then pord(p,b) · fpt (f ) ∈ N. In particular, for all such primes, pφ(b) · fpt (f ) ∈ N, where φ denotes Euler’s phi function. Using the techniques of the proof of Theorem 3.5, we can analogously find a bound for the power of p appearing in the denominator of fpt (f ) whenever fpt (f ) 6= λ and p is not large, which we record here. Corollary 3.8. In the setting of Theorem 3.5, if fpt (f ) 6= λ and p ∤ b, then pM · fpt (f ) ∈ N, where M := 2 · φ(b) + ⌈log2 (n − 1)⌉ , and φ denotes Euler’s phi function. Remark 3.9. We emphasize that the constant M in Corollary 3.8 depends only on the number of P deg(xi ) variables n and the quotient deg f = ab , but not on the particular values of deg xi and deg f ; this subtle point will play a key role in Proposition 7.3. Remark 3.10 (Towards minimal lists). For p ≫ 0, the bounds for L and E appearing in Theorem 3.5 allows one to produce a finite list of possible values of fpt (f ) for each class of p modulo deg f . We refer the reader to Section 4 for details and examples. The uniqueness statement in point (4) of the Theorem 3.5 need not hold in general. Example 3.11 (Non-uniqueness in low characteristic). If p = 2 and f ∈ L[x1 , x2 , x3 ] is any L∗ linear combination of x71 , x72 , x73 , then f satisfies the hypotheses of Theorem 3.5, under the standard grading. Using [Hera], one can directly compute that fpt (f ) = 41 . On the other hand, the identities !   q y 3 · 22 % 7 + 7 · 0 1 3 3 = − = 2 4 7 7·2 7 2 !   q y 3 · 23 % 7 + 7 · 1 3 1 3 = − 3 = − 3 7 7·2 7 3 2 show that the pairs (L, E) = (2, 0) and (L, E) = (3, 1) both satisfy the conclusions in Theorem 3.5. We point out that the proof of Theorem 3.5, being somewhat constructive, predicts the choice of (L, E) = (2, 0), but does not “detect” the choice of (L, E) = (3, 1). Before concluding this section, we present the following related result; like Theorem 3.5 and Corollary 3.8, its proof relies heavily on Proposition 5.6. However, in contrast to these results, its P focus is on showing that fpt (f ) = min {( deg xi ) / deg f, 1} for p ≫ 0 in a very specific setting, as opposed to describing fpt (f ) when it differs from this value. P P deg xi Theorem 3.12. In the context of Theorem 3.5, suppose that deg xi > deg f , so that ρ := deg f is greater than 1. If p > n−3 , then fpt (f ) = 1. ρ−1 As we see below, Theorem 3.12 need not hold in low characteristic. 7 Example 3.13 (Illustrating the necessity of p ≫ 0 in Theorem 3.12). Set f = xd1 + · · · + xdn . If e−1 e n > d > p, then f ∈ m[p] , and hence f p ∈ m[p ] for all e ≥ 1. Consequently, νf (pe ) ≤ pe−1 − 1, and therefore fpt (f ) = lim p−e · νf (pe ) ≤ p−1 . e→∞ 4. F - PURE THRESHOLDS OF HOMOGENEOUS POLYNOMIALS : E XAMPLES In this section, we illustrate, via examples, how Theorem 3.5 may be used to produce “short,” or even minimal, lists P of possible values for F -pure thresholds. We begin with the most transparent case: If deg f = deg xi , then the statements in Theorem 3.5 become less technical. Indeed, in this case, a = b = 1, and hence ord(p, b) = 1 = Jm % bK for every m ∈ N. In this context, substituting these values into Theorem 3.5 recovers the following identity, originally discovered by Bhatt and Singh under the standard grading. Example 4.1. [BS, Theorem 1.1] Fix an N-grading on L[x1 , · · · , xn ]. Consider a homogeneous polyp P nomial f with d := deg f = deg xi and Jac (f ) = m. If p > n − 2 and fpt (f ) 6= 1, then fpt (f ) = 1 − A · p−1 for some integer 1 ≤ A ≤ d − 2. P Next, we consider the situation when deg f = deg xi + 1; already, we see that this minor modification leads to a more complex statement. Corollary 4.2. Fix an N-grading p on L[x1 , · · · , xn ]. Consider a homogeneous polynomial f with P d := deg f = deg xi + 1 and Jac (f ) = m, and suppose that p > (n − 2) · d. (1) If fpt (f ) = 1 − d1 , then p ≡ 1 mod d.   (2) If fpt (f ) 6= 1 − d1 , then fpt (f ) = 1 − 1d − A − (a) 1 ≤ A ≤ d − 2 if p ≡ −1 mod d, and (b) 1 ≤ A ≤ d − 3 otherwise. Jp % dK d · p−1 for some integer A satisfying
(s)
(1) for s ≥ 1, and hence that ≤ d−1 Proof. We begin with (1): Lemma 2.5 implies that d−1



(s) (1) (1) (s) (4.0.1) 1 − d−1 = p − 1 − d−1 ≥ p − 1 − d−1 = 1 − d−1
(1) for every s ≥ 1. However, if fpt (f ) = 1 − d−1 , Proposition 3.3 implies that 1 − d−1 ≤
(s)
(1) −1 −1 1−d for every s ≥ 1. Consequently, equality holds throughout (4.0.1), and hence d =
(s) −1 d for every s ≥ 1, which by Lemma 2.5 occurs if and only if p ≡ 1 mod d. We now address the second point: In this setting, Theorem 3.5 states that fpt (f ) ∈ p−L · N for some integer L ≥ 1. We will now show that L must equal one: Indeed, otherwise L ≥ 2, which allows us to set e = 1 in the third point Theorem 3.5 to deduce that 1 ≤ d − 1 < Jp(d − 1) % dK = d − Jp % dK , and hence that Jp % dK < 1, which is impossible, as Jp % dK is always a positive integer. We conclude that L = 1, and the reader may verify that substituting L = 1, Jp(d − 1) % dK = d −  Jp % dK, and A := E + 1 into Theorem 3.5 produces the desired description of fpt (f ). 4.1. The two variable case. We now shift our focus to the two variable case of Theorem 3.5, motivated by the following example. Example 4.3. In [Har06, Corollary 3.9], Hara and Monsky independently described the possible values of fpt (f ) whenever f is homogeneous in two variables (under the standard grading) of degree 5 with an isolated singularity at the origin over an algebraically closed field (and hence, a product of five distinct linear forms), and p 6= 5; we recall their computation below (the description in terms of truncations is our own). 8 • If p ≡ 1 mod 5, then fpt (f ) = • If p ≡ 2 mod 5, then fpt (f ) = • If p ≡ 3 mod 5, then fpt (f ) = • If p ≡ 4 mod 5, then fpt (f ) = 2p−2 2 5 or 5p = 2p2 −3 = 25 2 5p2 2p−1 2 5p = 5 1 . 2p−3 2 5 or 5p = 2 5 1. 3 or 2p5p−1 3 2 5 1 or = 2p2 −2 5p2 2 5 3. = 2 5 2. The methods used in [Har06] rely on so-called “syzygy gap” techniques and the geometry of P1 , and hence differ greatly from ours. In this example, we observe the following: First, the F -pure threshold is always λ = 25 , or a truncation of 25 . Secondly, there seem to be fewer choices for the truncation point L than one might expect, given Theorem 3.5. In this subsection, we show that the two observations from Example 4.3 hold in general in the two variable setting. We now work in the context of Theorem 3.5 with n = 2, and relabel the variables so that f ∈ L[x, y]. Note that if deg f < deg xy, then fpt (f ) = 1, by Theorem 3.12 (an alternate justification: this inequality is satisfied if and only if, after possibly re-ordering the variables, f = x + y m for some m ≥ 1, in which case one can directly compute that νf (pe ) = pe − 1, and hence that fpt (f ) = 1). Thus, the interesting case here is when deg f ≥ deg xy. In this case, one obtains the following result. y]. Consider a homogeneous polynomial Theorem p 4.4 (cf. Theorem 3.5). Fix an N-grading on L[x, deg xy f with Jac (f ) = m and deg f ≥ deg xy. If fpt (f ) 6= deg f = ab , written in lowest terms, then q L y   ap % b deg xy deg xy − = fpt (f ) = deg f L deg f b · pL for some integer L satisfying the following properties: (1) If p ∤ b, then 1 ≤ L ≤ ord(p, b). e (2) If p > q b,Lthen ya < Jap % bK for all 1 ≤ e ≤ L − 1. (3) 1 ≤ ap % b ≤ b − a for all possible values of p. Proof. Assuming fpt (f ) 6= deg xy deg f , the bounds for E in Theorem 3.5 become ' &q y apL % b + a 0≤E ≤1− . b As the rounded term above is always either one or two, the inequality forces it to equal one, so xy that E = 0, which shows that fpt (f ) is a truncation of deg deg f . Moreover, the fact that the rounded q L y term above equals one also implies that ap % b + a ≤ b.  Remark 4.5. Though the first two points in Theorem 4.4 appear in Theorem 3.5, the third condition is special to the setting of two variables. Indeed, this extra condition will be key in eliminating potential candidate F -pure thresholds. For example, this extra condition allows us to recover the data in Example 4.3. Rather than justify this claim, we present two new examples. Example 4.6. Let f ∈ L[x, y] be as in Theorem 4.4, with • If p ≡ 1 mod 3, then fpt (f ) = • If p ≡ 2 mod 3, then fpt (f ) = 1 3 1 3 or or 1 3 1 1 3 1 = = 1 3 1 3 − − deg(xy) deg f 1 3p . 2 3p or = 13 . For p ≥ 5, the following hold: 1 3 2 = 1 3 − 1 . 3p2 In Example 4.6, the second and third points of Theorem 4.4 were uninteresting, as they did not “whittle away” any of the candidate F -pure thresholds identified by the first point of Theorem 4.4. The following example is more interesting, as we will see that both of the second and third points of Theorem 4.4, along with Proposition 3.3, will be used to eliminate potential candidates. 9 deg(xy) 2 deg f = 7 . For p ≥ 11, the following 2 . = 72 − 7p 1 2 4 2 2 1 7 − 7p or 7 2 = 7 − 7p2 . 2 4 2 5 2 1 2 2 7 − 7p2 or 7 3 = 7 − 7p3 or 7 4 = 7 − 7p4 . 2 1 7 − 7p . 2 3 2 2 1 7 − 7p or 7 2 = 7 − 7p2 . 5 or 72 2 = 72 − 7p22 . = 72 − 7p 1 Example 4.7. Let f ∈ L[x, y] be as in Theorem 4.4, with • • • • • • If p ≡ 1 mod 7, then fpt (f ) = If p ≡ 2 mod 7, then fpt (f ) = If p ≡ 3 mod 7, then fpt (f ) = If p ≡ 4 mod 7, then fpt (f ) = If p ≡ 5 mod 7, then fpt (f ) = If p ≡ 6 mod 7, then fpt (f ) = 2 7 2 7 or 2 7 2 7 2 7 2 7 1 2 1 1 or 2 7 = = = = 2 7 hold: For the sake of brevity, we only indicate how to deduce the lists for p ≡ 3 mod 7 and p ≡ 4 mod 7. Similar methods can be used for the remaining cases.
(1)
2 (5) = 2p−6 = p−3 (p ≡ 3 mod 7). In this case, it follows from Lemma 2.5 that 72 7 and 7 7 . In light of this, the second point of Proposition 3.3, which shows that the first digit of fpt (f ) must be the smallest digit, implies that fpt (f ) 6= 27 . Thus, the first point of Theorem 4.4 states that   2 for some 1 ≤ L ≤ ord(p, 7) = 6, as p ≡ 3 mod 7. fpt (f ) = 7 L q y However, as 2 6≤ 2p4 % 7 = 1, the second point of Theorem 4.4 eliminates the possibilities that L = 5 or 6. Moreover, as J2p % 7K = 6 6≤ 7 − 2 = 5, the third point of Theorem 4.4 eliminates the possibility that L = 1. Thus, the only remaining possibilities are fpt (f ) = 27 2 , 27 3 , and 27 4 .
(1)
(2) (p ≡ 4 mod 7). As before, we compute that 72 = 2p−1 = p−4 is greater than 72 7 7 , and 2 hence it again follows the second point of Proposition 3.3 that fpt (f ) 6= 7 . Consequently, the first point of Theorem 4.4 states that   2 for some 1 ≤ L ≤ ord(p, 7) = 3, as p ≡ 4 mod 7. fpt (f ) = 7 L q y However, we observe that 2 6≤ 2p2 % 7 = 1, and hence the second point of Theorem 4.4 eliminates the possibility that L = 2 or 3. Thus, the only remaining option is that fpt (f ) = 27 1 . Remark 4.8 (Minimal lists). In many cases, we are able to verify that the “whittled down” list obtained through the application of Theorems 4.4 and 3.5 and Proposition 3.3 is, in fact, minimal. For example, every candidate listed in Example 4.3 is of the form fpt (f ), where f varies among the polynomials x5 + y 5 , x5 + xy 4 , and x5 + xy 4 + 7x2 y 3 , and p various among the primes less than or equal to 29. An extreme example of the “minimality” of the lists of candidate thresholds appears below. Note that, in this example, the list of candidate thresholds is so small that it actually determines the precise value of fpt (f ) for p ≫ 0. Example 4.9 (F -pure thresholds are precisely determined). Let f ∈ L[x, y] be as in Theorem 4.4, = 53 ; for example, we may take f = x5 + x3 y + xy 2 , under the grading given by with deg(xy) deg f (deg x, deg y) = (1, 2). Using Theorem 4.4 and Proposition 3.3 in a manner analogous to that used in Example 4.7, we obtain the following complete description of fpt (f ) for p ≥ 7. • If p ≡ 1 mod 5, then fpt (f ) = 35 . • If p ≡ 2 mod 5, then fpt (f ) = 35 • If p ≡ 3 mod 5, then fpt (f ) = 35 • If p ≡ 4 mod 5, then fpt (f ) = 1 2 3 5 1 = = = 3 5 3 5 3 5 − − − 1 5p . 2 5p2 . 2p 5 . 10 We conclude this section with one final example illustrating “minimality”. In this instance, however, we focus on the higher dimensional case. Although the candidate list for F -pure thresholds produced by Theorem 3.5 is more complicated (due to the possibility of having a non-zero “E” term when n > 2), the following example shows that we can nonetheless obtain minimal lists in these cases using methods analogous to those used in this section’s previous examples. Example 4.10 (Minimal lists for n ≥ 3). Let f ∈ L[x, y, z] satisfy the hypotheses of Theorem 3.5, xyz 2 with deg deg f = 3 . Using the bounds for E and L therein, we obtain the following for p ≥ 5: • If p ≡ 1 mod 3, then fpt (f ) = • If p ≡ 2 mod 3, then fpt (f ) = 2 3 2 2 2 3 1 = 3 − 3p . 2 2 1 2 1 2 4 3 1 = 3 − 3p or 3 1 − p = 3 − 3p fact, if f = x9 + xy 4 + z 3 , homogeneous or We claim that this list is minimal. In under the grading determined by (deg x, deg y, deg z) = (1, 2, 3), we obtain each of these possibilities as p varies. 5. F - PURE THRESHOLDS OF HOMOGENEOUS POLYNOMIALS : D ETAILS Here, we prove the statements referred to in Section 3; we begin with some preliminary results. 5.1. Bounding the defining terms of the F -pure threshold. This subsection is dedicated to deriving bounds for νf (pe ). Our methods for deriving lower bounds are an extension of those employed by Bhatt and Singh in [BS]. Lemma 5.1. L[x1 , · ·k· , xn ] is homogeneous undernsome N-grading, then for every e ≥ 1, j If f ∈ P o P deg x deg x i νf (pe ) ≤ (pe − 1) · deg f i . In particular, fpt (f ) ≤ min deg f , 1 . e / Proof. By Definition 3.2, it suffices to establish the upper bound on νf (pe ). However, as f νf (p ) ∈ e e e m[p ] , there is a supporting monomial µ = xa11 · · · xann of f νf (p ) not in m[p ] , and comparing degrees shows that X X νf (pe ) · deg f = deg µ = ai · deg xi ≤ (pe − 1) · deg xi .  Corollary 5.2. n Let f ∈ L[x o1 , · · · , xn ] be a homogeneous polynomial under some N-grading, and P deg xi a e e write λ = min deg f , 1 = b in lowest terms. If fpt (f ) 6= λ, then ∆e := p hλie − p hfpt (f )ie defines a non-negative, non-decreasing, unbounded sequence. Moreover, if p ∤ b, then 1 ≤ min {e : ∆e 6= 0} ≤ ord(p, b). Proof. By Lemma 5.1, the assumption that fpt (f ) 6= λ implies that fpt (f ) < λ, the so the asserted properties of {∆e }e follow from Lemma 2.7. Setting s := ord(p, b), it follows from Lemma 2.5 that λ := . λ(1) : · · · : λ(s) (base p). By means of contradiction, suppose ∆s = 0, so that hλis = hfpt (f )is , i.e., so that (5.1.1) fpt (f ) = . λ(1) : · · · : λ(s) : fpt (f )(s+1) : fpt (f )(s+2) : · · · (base p). As fpt (f ) ≤ λ, comparing the tails of the expansions of fpt (f ) and λ shows that . fpt (f )(s+1) : · · · : fpt (f )(2s) (base p) ≤ . λ(1) : · · · : λ(s) (base p). On the other hand, comparing the first s digits appearing in the second point of Proposition 3.3, recalling the expansion (5.1.1), shows that . λ(1) : · · · : λ(s) (base p) ≤ . fpt (f )(s+1) : · · · fpt (f )(2s) (base p), 11 and thus we conclude that fpt (f )(s+e) = λ(s+e) for every 1 ≤ e ≤ s, i.e., that ∆2s = 0. Finally, a repeated application of this argument will show that ∆ms = 0 for every m ≥ 1, which implies that fpt (f ) = λ, a contradiction.  Notation 5.3. If R is any N-graded ring, and M is a graded R-module, [M ]d will denote the degree d component of M , and [M ]≤d and [M ]≥d the obvious [R]0 submodules of M . Furthermore, we P use HM (t) := d≥0 dim [M ]d · td to denote the Hilbert series of M . For the remainder of this subsection, we work in the following context. Setupp5.4. Fix an N-grading on R = L[x1 , · · · , xn ], and consider a homogeneous polynomial f ∈ m with Jac (f ) = m. In this context, ∂1 (f ), · · · , ∂n (f ) form a homogeneous system of parameters for R, and hence a regular sequence. Consequently, if we set Jk = (∂1 (f ), · · · , ∂k (f )), the sequences ∂k (f ) 0 → (R/Jk−1 ) (− deg f + deg xk ) −→ R/Jk−1 → R/Jk → 0 are exact for every 1 ≤ k ≤ n. Furthermore, using the fact that series is additive across Qn the Hilbert 1 short exact sequences, the well-known identities HR (t) = i=1 1−tdeg xi and HM (−s) (t) = ts HM (t) imply that HR/ Jac(f ) (t) = (5.1.2) n Y 1 − tdeg f −deg xi i=1 1 − tdeg xi , an identity that will play a key role in what follows.
e e Lemma 5.5. Under Setup 5.4, we have that m[p ] : Jac (f ) \ m[p ] ⊆ [R]≥(pe +1)·P deg xi −n·deg f . Proof. To simplify notation, set J = Jac (f ). By (5.1.2), the degree of HR/J (t) (a polynomial, as √ P J = m) is N := n deg f − 2 deg xi , and so [R/J]d = 0 whenever d ≥ N + 1. It follows that [R]≥N +1 ⊆ J, and to establish the claim, it suffices to show that   e e (5.1.3) m[p ] : [R]≥N +1 \ m[p ] ⊆ [R]≥(pe −1)·P deg xi −N = [R]≥(pe +1)·P deg xi −n·deg f . By means of contradiction, suppose (5.1.3) is false. Consequently, there exists a monomial   e e e µ = xp1 −1−s1 · · · xpn −1−sn ∈ m[p ] : [R]≥N +1 such that deg µ ≤ (pe − 1) · deg xi − (N + 1). This condition implies that the monomial µ◦ :=  e] s1 [p s x1 · · · xnn is in [R]≥N +1 , and as µ ∈ m : [R]≥N +1 , it follows that µµ◦ (which is apparently equal to (x1 · · · xn )p e −1 e ) is in m[p ] , a contradiction. l Proposition 5.6. In the setting of Setup 5.4, if p ∤ (νf (pe ) + 1), then νf (pe ) ≥ (pe + 1) · P deg xi deg f  m −n . e
e e e Proof. The Leibniz rule shows that ∂i m[p ] ⊆ m[p ] , and so differentiating f νf (p )+1 ∈ m[p ] shows e e that (νf (pe ) + 1) · f
νf (p ) · ∂i (f ) ∈ m[p ] for all i. Our assumption that p ∤ νf (pe ) + 1 then implies that e e e f νf (p ) ∈ m[p ] : J \ m[p ] ⊆ [R]≥(pe +1)·P deg xi −n·deg f , where the exclusion follows by definition, and the final containment by Lemma 5.5. Therefore, X deg f · νf (pe ) ≥ (pe + 1) · deg xi − n · deg f, and the claim follows.  12 Corollary 5.7. In the setting of Setup 5.4, write λ = min (e) fpt (f ) , the eth nP o deg xi deg f , 1 = a b, in lowest terms. If digit of fpt (f ), is not equal to p − 1, then   Jape % bK + a e e p hλie − p hfpt (f )ie ≤ n − . b Proof. By Proposition 3.3, νf (pe ) = pe hfpt (f )ie ≡ fpt (f )(e) mod p, and so the condition that fpt (f )(e) 6= p − 1 is equivalent to the condition that p ∤ P (νf (pe ) + 1). In light of this, we are free to apply Proposition 5.6. In what follows, we set δ := ( deg xi ) · (deg f )−1 .  First, suppose that min {δ, 1} = 1, so that a = b = 1. Then (Jape % bK + a) · b−1 = 2, and so it suffices to show that pe h1ie − pe hfpt (f )ie ≤ n − 2. However, the assumption that min {δ, 1} = 1 implies that δ ≥ 1, and Proposition 5.6 then shows that pe · hfpt (f )ie = νf (pe ) ≥ ⌈(pe + 1) · δ − n⌉ ≥ ⌈pe + 1 − n⌉ = pe − 1 + 2 − n = pe · h1ie + 2 − n. If instead min {δ, 1} = δ, Proposition 5.6 once again shows that pe hfpt (f )ie = νf (pe ) ≥ ⌈(pe + 1) · δ − n⌉ = ⌈pe · δ + δ − n⌉     Jape % bK + δ − n = pe · hδie + b · pe   e Jap % bK e + δ − n, = p · hδie + b the second to last equality following from Lemma 2.5.  Example 5.8 (Illustrating that Corollary 5.7 is not an equivalence). If p = 2 and f is any L∗ -linear (e) 15 combination of x15 6= 1, then ∆e := 2e 3−1 e − 1 , · · · , x5 , Corollary 5.7 states that if fpt (f ) e 2 hfpt (f )ie ≤ 4. We claim that the converse fails when e = 4. Indeed, a direct computation, made possible by [Hera], shows that fpt (f ) = 18 , and comparing the base 2 expansions of fpt (f ) = 18 and λ = 13 shows that ∆4 = 4, even though fpt (f )(4) = 1 = p − 1. 5.2. Proofs of the main results. In this subsection, we return to the statements in Section 3 whose proofs were postponed. For the benefit of the reader, we restate these results here. p Theorem 3.5. Fix annN-gradingoon R. Consider a homogeneous polynomial f with Jac (f ) = m, P deg xi a and write λ := min deg f , 1 = b in lowest terms. (1) If fpt (f ) 6= λ, then q ! y apL % b + b · E E fpt (f ) = λ − = hλiL − L L b·p p   J apL % b K+a 2 . for some (L, E) ∈ N with L ≥ 1 and 0 ≤ E ≤ n − 1 − b (2) If p > (n − 2) · b and p ∤ b, then 1 ≤ L ≤ ord(p, b); note that ord(p, 1) = 1. (3) If p > (n − 2) · b and p > b, then a < Jape % bK for all 1 ≤ e ≤ L − 1. (4) If p > (n − 1) · b, then there exists a unique pair (L, E) satisfying the conclusions of (1). Proof. We begin by establishing (1): The two descriptions of fpt (f ) are equivalent by Lemma 2.5, and so it suffices to establish the identity in terms of truncations. Setting ∆e := pe hλie − 13 pe hfpt (f )ie , Corollary 5.2, states that {∆e }e≥1 is a non-negative, non-decreasing, unbounded sequence; in particular, min {e : ∆e 6= 0} ≥ 1 is well-defined, and we claim that n o ℓ := min {e : ∆e 6= 0} ≤ L := max e : fpt (f )(e) 6= p − 1 , l e m the latter also being well-defined. Indeed, set µe := Jap %b bK+a . As 1 ≤ µe ≤ 2, the sequence {n − µe }e≥1 is bounded above by n − 1, and therefore ∆e > n − µe for e ≫ 0. For such e ≫ 0, Corollary 5.7 implies that fpt (f )(e) = p − 1, which demonstrates that L is well-defined. Note that, by definition, ∆ℓ = λ(ℓ) − fpt (f )(ℓ) ≥ 1, so that fpt (f )(ℓ) ≤ λ(ℓ) − 1 ≤ p − 2; by definition of L, it follows that ℓ ≤ L. As fpt (f )(e) = p − 1 for e ≥ L + 1, (5.2.1) fpt (f ) = hfpt (f )iL + 1 ∆L 1 E = hλiL − L + L = hλiL − L , L p p p p where E := ∆L − 1. In order to conclude this step of the proof, it suffices to note that (5.2.2) 1 ≤ ∆ℓ ≤ ∆L ≤ n − µL ≤ n − 1; indeed, the second bound in (5.2.2) follows from the fact that L ≥ ℓ, the third follows from Corollary 5.7, and the last from the bound 1 ≤ µe ≤ 2. For point (2), we continue to use the notation adopted above. We begin by showing that (5.2.3) ∆e = 0 for all 0 ≤ e ≤ L − 1 whenever p > (n − 2) · b. As the sequence ∆e is non-negative and non-decreasing, it suffices to show that ∆L−1 = 0. Therefore, by way of contradiction, we suppose that ∆L−1 ≥ 1. By definition 0 ≤ fpt (f )(L) ≤ p − 2, and hence ∆L = p · ∆L−1 + λ(L) − fpt (f )(L) ≥ λ(L) + 2. Comparing this with (5.2.2) shows that λ(L) + 2 ≤ ∆L ≤ n − 1, so that λ(L) ≤ n − 3. On the other hand, if p > (n − 2) · b, then it follows from the explicit formulas in Lemma 2.5 that q e−1 y ap % b · p − Jape % bK p−b (n − 2) · b − b (e) ≥ > = n − 3 for every e ≥ 1. (5.2.4) λ = b b b In particular, setting e = L in this identity shows that λ(L) > n−3, contradicting our earlier bound. Thus, we conclude that (5.2.3) holds, which when combined with (5.2.2) shows that L = min {e : ∆e 6= 0}. In summary, we have just shown that L = ℓ when p > (n − 2) · b. If we assume further that p ∤ b, the desired bound L = ℓ ≤ ord(p, b) then follows from Corollary 5.2. We now focus on point (3), and begin by observing that (5.2.5) fpt (f ) = λ(1) : · · · : λ(L−1) : λ(L) − ∆L : p − 1 (base p) whenever p > (n − 2) · b. Indeed, by (5.2.3), the first L − 1 digits of fpt (f ) and λ agree, while fpt (f )(e) = p − 1 for e ≥ L + 1, by definition of L. Finally, (5.2.3) shows that ∆L = λ(L) − fpt (f )(L) , so that fpt (f )(L) = λ(L) − ∆L . Recall that, by the second point of Proposition 3.3, the first digit of fpt (f ) is its smallest digit, and it follows from (5.2.5) that λ(1) ≤ λ(e) for all 1 ≤ e ≤ L, with this inequality being strict for e = L. However, it follows from the explicit formulas in Lemma 2.5 that whenever p > b, q y λ(1) ≤ λ(e) ⇐⇒ a · p − Jap % bK ≤ ape−1 % b · p − Jape % bK q y ⇐⇒ a ≤ ape−1 % b , 14 where q the second y equivalence relies on the fact that p > b. Summarizing, we have just shown that a ≤ ape−1 % b for all 1 ≤ e ≤ L whenever p > (n − 2) · b and p > b; relabeling our index, we see that a ≤ Jape % bK for all 0 ≤ e ≤ L − 1 whenever p > (n − 2) · b and p > b. It remains to show that this bound is strict for 1 ≤ e ≤ L − 1. By contradiction, assume that a = Jape % bK for some such e. In this case, a ≡ a · pe mod b, and as a and b are relatively prime, we conclude that pe ≡ 1 mod b, so that ord(p, b) | e. However, by definition 1 ≤ e ≤ L − 1 ≤ ord(p, b) − 1, where the last inequality follows point (2). Thus, we have arrived at a contradiction, and therefore conclude that our asserted upper bound is strict for 1 ≤ e ≤ L − 1. To conclude our proof, it remains to establish the uniqueness statement in point (4). To this end, let (L′ , E ′ ) denote any pair of integers satisfying the conclusions of point (1) of this Theorem; that is, ′ fpt (f ) = hλiL′ − E ′ · p−L with 1 ≤ E ′ ≤ n − 1 − µL′ ≤ n − 2. A modification of (5.2.4) shows that λ(e) > n − 2, and hence that λ(e) ≥ E ′ + 1, whenever p > (n − 1) · b, and it follows that ′ ′ ′ fpt (f ) = hλiL′ − E ′ · p−L = .λ(1) : · · · : λ(L −1) : λ(L ) − (E + 1) : p − 1 whenever p > (n − 1) · b. The uniqueness statement then follows from comparing this expansion with (5.2.5) and invoking the uniqueness of non-terminating base p expansions.  Corollary 3.8. In the setting of Theorem 3.5, if fpt (f ) 6= λ and p ∤ b, then pM · fpt (f ) ∈ N, where M := 2 · φ(b) + ⌈log2 (n − 1)⌉ , and φ denotes Euler’s phi function. Proof. We adopt the notation used in the 3.5. In particular, ℓ ≤ L and fpt (f ) ∈  proof of Theorem  p−L · N. Setting s = ord(p, b), and k = logp (n − 1) in Lemma 2.7 shows that (5.2.6) ∆ℓ+s+⌈log ⌈logp (n−1)⌉ + 1 ≥ n. ⌉≥p p (n−1) By definition of L, Corollary 5.7 states that ∆L ≤ n − 1, and as {∆e }e≥1 is non-decreasing, (5.2.6)   then shows that L is bounded above by ℓ + s + logp (n − 1) . To obtain a uniform bound, note that  ℓ ≤ s, by Corollary 5.2, while s ≤ φ(b), by definition, and logp (n − 1) ≤ log2 (n − 1), as p ≥ 2. P P deg xi Theorem 3.12. In the context of Theorem 3.5, suppose that deg xi > deg f , so that ρ := deg f , then fpt (f ) = 1. is greater than 1. If p > n−3 ρ−1 Proof. We begin with the following elementary manipulations, the first of which relies on the assumption that ρ − 1 is positive: Isolating n − 3 in our assumption that p > (n − 3) · (ρ − 1)−1 − 1 shows that (p + 1) · (ρ − 1) > n − 3, and adding p + 1 and subtracting n from both sides then shows that (p + 1) · ρ − n > p − 2; rounding up, we see that (5.2.7) ⌈(p + 1) · ρ − n⌉ ≥ p − 1. Assume, by means of contradiction, that fpt (f ) 6= 1. By hypothesis, 1 = min {ρ, 1}, and Corollary 5.2 then states that 1 = min {e : pe h1ie − pe hfpt (f )ie ≥ 1}; in particular, νf (p) = fpt (f )(1) = p · hfpt (f )i1 ≤ p h1i1 − 1 = p − 2. However, this bound allows us to apply Proposition 5.6, which when combined with (5.2), implies that νf (p) ≥ ⌈(p + 1) · ρ − n⌉ ≥ p − 1. Thus, we have arrived at a contradiction, which allows us to conclude that fpt (f ) = 1. 15  6. A PPLICATIONS TO LOG CANONICAL THRESHOLDS Given a polynomial fQ over Q, we will denote its log canonical threshold by lct (fQ ). In this article, we will not need to refer to the typical definition(s) of lct (fQ ) (e.g., via resolution of singularities), and will instead rely on the limit in (6.0.8) below as our definition. However, so that the reader unfamiliar with this topic may better appreciate (6.0.8), we present the following characterizations. In what follows, we fix fQ ∈ Q[x1 , · · · , xn ].   n n (1) If π : X → AQ is a log resolution of the pair AQ , V(fQ ) , then lct (fQ ) is the supremum over all λ > 0 such that the coefficients of the divisor Kπ − λ · π ∗ div(f ) are all greater than −1; here, Kπ denotes the relative canonical divisor of π. (2) For every λ > 0, consider the function Γλ (fQ ) : Cn → R given by (z1 , · · · , zn ) 7→ |f (z1 , · · · , zn )|−2λ , where | · | ∈ R denotes the norm of a complex number; note that Γλ (fQ ) has a pole at all (complex) zeros of fQ . In this setting, lct (fQ ) := sup {λ : Γλ (fQ ) is locally R-integrable} , where here, “locally R-integrable” means that we identify Cn = R2n , and require that this function be (Lebesque) integrable in a neighborhood of every point in its domain. (3) The roots of the Bernstein-Sato polynomial bfQ of fQ are all negative rational numbers, and −lct (fQ ) is the largest such root [Kol97]. For more information on these invariants, the reader is referred to the surveys [BL04, EM06]. We now recall the striking relationship between F -pure and log canonical thresholds: Though there are many results due to many authors relating characteristic zero and characteristic p > 0 invariants, the one most relevant to our discussion is the following theorem, which is due to Mustaţă and the fourth author. Theorem 6.1. [MZ, Corollary 3.5, 4.5] Given an polynomial fQ over Q, there exist constants C ∈ R>0 and N ∈ N (depending only on fQ ) with the following property: For p ≫ 0, either fpt (fp ) = lct (fQ ), or 1 C ≤ lct (fQ ) − fpt (fp ) ≤ . N p p Note that, as an immediate corollary of Theorem 6.1, (6.0.8) fpt (fp ) ≤ lct (fQ ) for all p ≫ 0 and lim fpt (fp ) = lct (fQ ) . p→∞ We point out that (6.0.8) (which follows from the work of Hara and Yoshida) appeared in the literature well before Theorem 6.1 (see, e.g., [MTW05, Theorem 3.3, 3.4]). 6.1. Regarding uniform bounds. Though the constants C ∈ R>0 and N ∈ N appearing in Theorem 6.1 are known to depend only on fQ , their determination is complicated (e.g., they depend on numerical invariants coming from resolution of singularities), and are therefore not explicitly described. In Theorem 6.2 below, we give an alternate proof of this result for homogeneous polynomials with an isolated singularity at the origin; in the process of doing so, we also identify explicit values for C and N . p Theorem 6.2. If fQ ∈nQ[x1 , · · · ,o xn ] is homogeneous under some N-grading, with Jac (fQ ) = m, P deg xi a then lct (fQ ) = min deg f , 1 , which we write as b in lowest terms. Moreover, if fpt (fp ) 6= lct (fQ ), then b−1 n − 1 − b−1 ≤ lct (f ) − fpt (f ) ≤ for p ≫ 0, Q p p pord(p,b) where ord(p, b) denotes the order of p mod b (which equals one when b = 1, by convention). 16 p Proof. As the reduction of ∂k (f ) mod p equals ∂k (fp ) for large values of p, the equality Jac (fQ ) = m reduces n P modop for p ≫ 0. Taking p → ∞, it follows from Theorem 3.5 and (6.0.8) that lct (fQ ) = min (6.1.1) deg xi deg f , 1 , and in light of this, Theorem 3.5 states that q L y ap % b E lct (fQ ) − fpt (fp ) = + L. L b·p p If lct (fQ ) 6= 1, then q L y ap % b 1 1 − b−1 ≤ ≤ . b · pL b · pL pL Furthermore, Theorem 3.5q implies ythat 1 ≤ L ≤ φ(b) and 0 ≤ E ≤ n − 2 for p ≫ 0. If instead lct (fQ ) = a = b = 1, then apL % b = φ(b) = 1, and hence q L y ap % b b−1 = . b · pL pL Moreover, in this case, Theorem 3.5 shows that L = 1 and 0 ≤ E ≤ n − 3 for p ≫ 0. Finally, it is left to the reader to verify that substituting these inequalities into (6.1.1) produces the desired bounds in each case.  Remark 6.3 (On uniform bounds). Of course, ord(p, b) ≤ φ(b), where φ denotes Euler’s phi function. By enlarging p, if necessary, it follows that the lower bound in Theorem 6.2 is itself bounded below by p−φ(b)−1 . In other words, in the language of Theorem 6.1, we may take N = φ(b) + 1 and C = n − 1 − b−1 . Remark 6.4 (Regarding sharpness). The bounds appearing in Theorem 6.2 are sharp: If d > 2 and fQ = xd1 + · · · + xdd , then lct (fQ ) = 1, and Theorem 6.2 states that (6.1.2) 1 d−2 ≤ lct (fQ ) − fpt (fp ) ≤ p p whenever fpt (fp ) 6= 1 and p ≫ 0. However, it is shown in [Hera, Corollary 3.5] that lct (fQ ) − fpt (fp ) = 1 − fpt (fp ) = Jp % dK − 1 p whenever p > d. If d is odd and p ≡ 2 mod d, then the lower bound in (6.1.2) is obtained, and similarly, if p ≡ d − 1 mod d, then the upper bound in (6.1.2) is obtained; in both these cases, Dirichlet’s theorem guarantees that there are infinitely many primes satisfying these congruence relations. 6.2. On the size of a set of bad primes. In this subsection, we record some simple observations regarding the set of primes for which the F -pure threshold does not coincide with the log canonical threshold, and we begin by recalling the pcase of elliptic curves: Let fQ ∈ Q[x, y, z] be a homogeneous polynomial of degree three with Jac (fQ ) = m, so that E := V(f ) defines an elliptic curve in P2Q . As shown in the proof of Theorem 6.2, the reductions fp ∈ Fp [x, y, z] satisfy these same conditions for p ≫ 0, and thus define elliptic curves Ep = V(fp ) ⊆ P2Fp for all p ≫ 0. Recall that the elliptic curve Ep is called supersingular if the natural Frobenius action on the local cohomology 2 module H(x,y,z) (Fp [x, y, z]/(fp )) is injective, or equivalently, if (fp )p−1 ∈ / (xp , y p , z p ) (see, e.g, [Sil09, Chapters V.3 and V.4] for these and other characterizations of supersingularity). Using these descriptions, one can show that Ep is supersingular if and only if fpt (fp ) = 1 [MTW05, Example 4.6]. In light of this, Elkies’ well-known theorem on the set of supersingular primes, which states that Ep is supersingular for infinitely many primes p, can be restated as follows. 17 Theorem 6.5. [Elk87] If fQ ∈ Q[x, y, z] is as above, the set of primes {p : fpt (fp ) 6= lct (fQ )} is infinite. Recall that given a set S of prime numbers, the density of S, δ(S), is defined as δ(S) = lim n→∞ # {p ∈ S : p ≤ n} . : p ≤ n} # {p In the context of elliptic curves over Q, the set of primes {p : fpt (fp ) 6= lct (fQ )}, which is infinite by Elkies’ result, may be quite large (i.e., have density 12 ), or may be quite small (i.e., have density zero); see [MTW05, Example 4.6] for more information. This discussion motivates the following question. Question 6.6. For which polynomials fQ is the set of primes {p : fpt (fp ) 6= lct (fQ )} infinite? In the case that this set is infinite, what is its density? As illustrated by the case of an elliptic curve, Question 6.6 is quite subtle, and one expects it to be quite difficult to address in general. However, as we see below, when the numerator of lct (fQ ) is not equal to 1, one is able to give a partial answer to this question using simple methods. Our main tool will be Proposition 3.3, which provides us with a simple criterion for disqualifying a rational number from being an F -pure threshold. We stress the fact that Proposition 6.7 is not applicable when lct (fQ ) = 1, and hence sheds no light on the elliptic curve case discussed above. Proposition 6.7. Let fQ denote any polynomial over Q, and write lct (fQ ) = ab in lowest terms. If a 6= 1, then the set of primes for which lct (fQ ) is not an F -pure threshold (of any polynomial) is infinite, and contains all primes p such that pe · a ≡ 1 mod b for some e ≥ 1. In particular, δ ({p : fpt (fp ) 6= lct (fQ )}) ≥ 1 . φ(b) Proof. As a and b are relatively prime, there exists c ∈ N such that a · c ≡ 1 mod b. We claim that {p : p ≡ c mod b} ⊆ {p : pe · a ≡ 1 mod b for some e ≥ 1} ⊆ {p : lct (fQ ) is not an F -pure threshold in characteristic p > 0} . 1 Once we establish this, the proposition will follow, as δ ({p : p ≡ c mod b}) = φ(b) by Dirichlet’s theorem. By definition of c, the first containment holds by setting e = 1, and so it suffices to establish the second containment. However, if pe · a ≡ 1 mod b for some e ≥ 1, then Lemma 2.5 shows that q y q y Jape % bK · p − ape+1 % b p − ape+1 % b (e+1) lct (fQ ) = = . b b On the other hand, Lemma 2.5 also shows that lct (fQ )(1) = a · p − Jap % bK , b and as a ≥ 2, by assumption, we see that lct (fQ )(1) > lct (fQ )(e) for all p ≫ 0. In light of this, the second point of Proposition 3.3, which shows that the first digit of an F -pure threshold must be its smallest, shows that lct (fQ ) could not be the F -pure threshold of any polynomial in characteristic p > 0.  We conclude this section with the following example, which follows immediately from Corollary 4.2, and which illustrates a rather large family of polynomials whose set of “bad” primes greatly exceeds the bound given by Proposition 6.7. 18 Example (under the standard grading) of degree d p 6.8. If fQ ∈ Q[x1 , · · · , xd−1 ] is homogeneous  with Jac (fQ ) = m, then {p : p 6≡ 1 mod d} ⊆ p : fpt (fp ) 6= lct (fQ ) = 1 − 1d . In particular, δ ({p : fpt (fp ) 6= lct (fQ )}) ≥ δ ({p : p 6≡ 1 mod d}) = 1 − 7. A SPECIAL CASE OF ACC AND LOCAL m- ADIC CONSTANCY FOR 1 . φ(d) F - PURE THRESHOLDS Motivated by the relationship between F -pure thresholds and log canonical thresholds, Blickle, Mustaţă, and Smith conjectured the following. Conjecture 7.1. [BMS09, Conjecture 4.4] Fix an integer n ≥ 1. (1) The set {fpt (f ) : f ∈ L[x1 , · · · , xn ]} satisfies the ascending chain condition (ACC); i.e., it contains no strictly increasing, infinite sequence. (2) For every f ∈ L[x1 , · · · , xn ], there exists an integer N (which may depend on f ) such that fpt (f ) ≥ fpt (f + g) for all g ∈ mN . As discussed in [BMS09, Remark 4.5], the first conjecture implies the second, which states that the F -pure threshold function f 7→ fpt (f ) is locally constant (in the m-adic topology). In this section, we confirm the first conjecture for a restricted set of F -pure thresholds (see we confirm the second in the case that f is homogeneous under Proposition 7.3). Additionally, p some N-grading with Jac (f ) = m (see Propositions 7.8 and 7.10). 7.1. A special case of ACC. Definition 7.2. For every ω ∈ Nn , let Wω denote the set of polynomials f ∈ L[x1 , · · · , xn ] satisfying the following conditions: p (1) Jac (f ) = m. (2) f is homogeneous under the grading determined by (deg x1 ,n· P · · , deg xno) = ω. deg xi (3) p ∤ deg f (and hence, does not divide the denominator of min deg f , 1 , in lowest terms). S Given N ∈ N, set W4N := ω Wω , where the union is taken over all ω = (ω1 , . . . , ωn ) ∈ Nn with ωi ≤ N for each 1 ≤ i ≤ n. Proposition 7.3. For every N ∈ N and µ ∈ (0, 1], the set {fpt (f ) : f ∈ W4N } ∩ (µ, 1] is finite. In particular, this set of F -pure thresholds satisfies ACC. Proof. Fix f ∈ W4N such that fpt (f ) > µ. By definition, there exists an N-grading on L[x1 , · · · , xn ] such that deg xi ≤ N for all 1 ≤ i ≤ N , and under which f is homogeneous. Moreover, by Lemma 5.1, Pn deg xi n·N µ < fpt (f ) ≤ i=1 ≤ . deg f deg f Consequently, deg f ≤ n·N µ , and it follows that     Pn a n·N i=1 deg xi , 1 ⊆ S := (0, 1] ∩ ∈Q:b≤ , λ := min deg f b µ a finite set. We will now show that fpt (f ) can take on only finitely many values: If fpt (f ) 6= λ, then by Corollary 3.8, there exists an integer Mλ , depending onlynon λ and n,o such that pMλ · fpt (f ) ∈ N. If M := max {Mλ : λ ∈ S}, it follows that fpt (f ) ∈ set. 19 a pM : a ∈ N ∩ (0, 1], a finite  7.2. A special case of local m-adic constancy of the F -pure threshold function. Throughout this subsection, we fix an N-grading on L[x1 , · · · , xn ]. Lemma 7.4. Consider f ∈ m such that pL · fpt (f ) ∈ N for some L ∈ N. If g ∈ m, then fpt (f + g) ≤ fpt (f ) ⇐⇒ (f + g)p L L s L ·fpt(f ) s L ∈ m[p ] . Proof. If (f + g)p ·fpt(f ) ∈ m[p ] , then (f + g)p ·fpt(f ) ∈ m[p ] for s ≥ L. Consequently, νf +g (ps ) < ps · fpt (f ) for s ≫ 0, and hence fpt (f + g) ≤ fpt (f ). We now focus on the remaining implication. L )−1 + p1L shows By the hypothesis, pL · fpt (f ) − 1 ∈ N, and hence the identity fpt (f ) = p ·fpt(f pL that pL · fpt (f ) − 1 hfpt (f )iL = . pL If fpt (f + g) ≤ fpt (f ), the preceding identity and Proposition 3.3 show that νf +g (pL ) = pL hfpt (f + g)iL ≤ pL hfpt (f )iL = pL fpt (f ) − 1, and consequently, this bound for νf +g (pL ) shows that (f + g)p L fpt(f ) L ∈ m[p ] . P e Lemma 7.5. If h is homogeneous and h ∈ / m[p ] , then deg h ≤ (pe − 1) · ni=1 deg xi . Proof. Every supporting monomial of h is of the form xp1 e −a 1  e · · · xpn −an , where each ai ≥ 1. Then n n X X e e deg xi . (p − ai ) deg xi ≤ (p − 1) deg h = i=1 i=1  Proposition 7.6. Fix f ∈ m homogeneous. If g ∈ [R]≥deg f +1 , then fpt (f ) ≤ fpt (f + g). Proof. If suffices to show that for every e ≥ 1, νf (pe ) ≤ νf +g (pe ); i.e., if N := νf (pe ), then (f + P e e N
N −k k g ∈ m[p ] ; g)N ∈ / m[p ] . Suppose, by way of contradiction, that (f + g)N = f N + N k=1 k f e note that, as f N ∈ / m[p ] by definition, each monomial summand of f N must cancel with one of PN N
N −k k g . However, for any monomial summand µ of any f N −k gk , k ≥ 1, k=1 k f deg µ ≥ (N − k) deg f + k(deg f + 1) > N deg f = deg(f N ), and such cancelation is impossible. Lemma 7.7. Fix f ∈ m homogeneous such that λ := e e e [R]≥deg f +1 , then (f + g)p hλie ≡ f p hλie mod m[p ] . P deg xi deg f  ≤ 1. If (pe − 1) · λ ∈ N and g ∈ Proof. We claim that (7.2.1) fp e hλi −k e e gk ∈ m[p ] for all 1 ≤ k ≤ pe hλie . Indeed, suppose that (7.2.1) is false. As g ∈ [R]≥deg f +1 and µ is a supporting monomial of we also have that e f p hλie −k gk , (7.2.2) deg µ ≥ deg f · (pe hλie − k) + (deg f + 1) · k = deg f · pe hλie + k. However, as (pe − 1) · λ ∈ N, it follows from Lemma 2.6 that pe hλie = (pe − 1) · λ. Substituting this into (7.2.2) shows that n X deg xi , deg µ ≥ deg f · (pe − 1) · λ + k = k + (pe − 1) i=1 which contradicts Lemma 7.5 as k ≥ 1. Thus, (7.2.1) holds, and it follows from the Binomial e e e  Theorem that (f + g)p hλie ≡ f p hλie mod m[p ] . 20 We are now able to prove our first result on the m-adic constancy of the F -pure threshold function, which does not require the isolated singularity hypothesis. P deg xi ≤ 1, and suppose that either Proposition 7.8. Fix f ∈ m homogeneous such that λ := deg f L fpt (f ) = λ, or fpt (f ) = hλiL and (p − 1) · λ ∈ N for some L ≥ 1. Then fpt (f + g) = fpt (f ) for each g ∈ [R]≥deg f +1 . Proof. By Proposition 7.6, it suffices to show that fpt (f ) ≥ fpt (f + g). First say that fpt (f ) = λ. e e It is enough to show that for all e ≥ 1, (f + g)νf (p )+1 ∈ m[p ] , so that νf (pe ) ≥ νf +g (pe ). By the e e Binomial Theorem, it suffices to show that for all 0 ≤ k ≤ νf (pf ) + 1, f νf (p )+1−k gk ∈ m[p ] . To this e end, take any monomial µ of such an f νf (p )+1−k gk . Then (7.2.3) deg µ ≥ (νf (pe )+ 1− k)·deg f + k ·(deg f + 1) = (νf (pe )+ 1)·deg f + k ≥ (νf (pe )+ 1)·deg f. By Lemma 3.3, νf (pe ) = pe hλie , and by definition, hαie ≥ α − 1 pe for all 0 < α ≤ 1. Then by (7.2.3), deg µ ≥ (pe hλie + 1) · deg f     1 e ≥ p λ − e + 1 · deg f p e = p · λ · deg f X = pe · deg xi . e We may now conclude that µ ∈ m[p ] Lemma 7.5. Now say that fpt (f ) = hλiL and (pL − 1) · λ ∈ N for some L ≤ 1. By Lemma 7.4, it suffices L L to show that (f + g)p ·fpt(f ) ∈ m[p ] . Indeed, pL · fpt (f ) > pL hfpt (f )iL = νf (pL ) (the equality L L L L L by Proposition 3.3), so that f p ·fpt(f ) ∈ m[p ] ; thus, (f + g)p ·fpt(f ) ≡ f p ·fpt(f ) ≡ 0 mod m[p ] by  Lemma 7.7. We see that the hypotheses of Proposition 7.8 are often satisfied in Example 7.9 below. We also see that the statement of the proposition is sharp in the sense that there exist f and g satisfying its hypotheses such that fpt (f ) = hλiL for some L ≥ 1, (pL − 1) · λ ∈ / N, and fpt (f + g) > fpt (f ). Example 7.9. Let f = x15 + xy 7 ∈ L[x, y], which is homogeneous with deg f = 15 under the grading determined by (deg x, deg y) = (1, 2), and has an isolated singularity at the origin when p ≥ 11. It follows from Theorem 4.4 that     1+2 1 fpt (f ) = = , 15 L 5 L where 1 ≤ L ≤ ord(p, 5) ≤ 4, or L = ∞ (i.e., fpt (f ) = 15 ). Furthermore, as f is a binomial, we can use the algorithm given in [Herb], and recently implemented by Sara Malec, Karl Schwede, and the third author in an upcoming Macaulay2 package, to compute the exact value fpt (f ), and hence, the exact value of L for a fixed p. We list some of these computations in Figure 7.9. We see that the hypotheses of Proposition 7.8 are often satisfied in this example, and it follows that fpt (f ) = fpt (f + g) for every g ∈ [R]≥16 whenever either “∞” appears in the second column 3 or “Yes” appears in the third. When p = 17, however, we have that fpt (f ) = 15 1 = 17 , and when  14 12 2 8 13 2 14 2 3 [17] g ∈ x y, x y , y , x y , x y ⊆ [R]≥16 , one may verify that (f + g) ∈ / m , so that it follows 3 from Lemma 7.4 that fpt (f  + g) > 17 . For another example of this behavior, it can be computed that when p = 47 and g ∈ x12 y 2 , x10 y 3 , x8 y 4 , x4 y 6 , x9 y 4 , x10 y 4 , then fpt (f + g) > fpt (f ). We now present the main result of this subsection. 21 p L (pL − 1) · 15 ∈ N? 11 1 Yes 13 1 No 17 1 No 19 2 Yes 23 4 Yes 29 ∞ – 31 1 Yes p L (pL − 1) · 15 ∈ N? 37 4 Yes 41 1 Yes 43 ∞ – 47 1 No 53 4 Yes 59 2 Yes 61 1 Yes p L (pL − 1) · 15 ∈ N? 67 1 No 71 ∞ – 73 3 No 79 2 Yes 83 2 No 97 1 No 101 1 Yes F IGURE 1. Some data on F -pure thresholds of f = x15 + xy 7 ∈ L[x, y]. Proposition 7.10. Suppose that f ∈ Fp [x1 , · · · , xn ] is homogeneous under some N-grading such p P that Jac (f ) = (x1 , . . . , xn ) and deg f ≥ deg xi . Then fpt (f + g) = fpt (f ) for each g ∈ P [R]≥n deg f − deg xi +1 . P deg xi Proof. Let λ = deg f . If fpt (f ) = λ, then Proposition 7.8 implies that fpt (f ) = fpt (f + g). For the remainder of this proof, we will assume that fpt (f ) 6= λ. By Proposition 7.6, it suffices to show that fpt (f ) ≥ fpt (f + g). Since fpt (f ) = hλiL − pEL for some integers E ≥ 0 and L ≥ 1 by L L Theorem 3.5, it is enough to show that (f + g)p fpt(f ) ∈ m[p ] by Lemma 7.4. To this end, note that 1 E n−1 E (7.2.4) fpt (f ) = hλiL − L ≥ λ − L − L ≥ λ − L , p p p p where the first inequality follows from Lemma 2.6, and the second from our bounds on E. SupL L / m[p ] . As pose, by way of contradiction, that (f + g)p ·fpt(f ) ∈ pL ·fpt(f ) (f + g) =f pL ·fpt(f ) + pL ·fpt(f )  L X p k=1 pL pL hfpt (f )i  · fpt (f ) pL ·fpt(f )−k k f g , k L L L p ·fpt(f ) ∈ m[p ] , and so there the inequality · fpt (f ) > L = νf (p ) implies that f L L must exist 1 ≤ k ≤ pL · fpt (f ) for which f p ·fpt(f )−k gk ∈ / m[p ] . We will now show, as in the proof of Lemma 7.7, that this is impossible for degree reasons. Indeed, for such a k, there exists a P L L supporting monomial µ of f p fpt(f )−k gk not contained in m[p ] , so that deg µ ≤ (pL − 1) · deg xi by Lemma 7.5. However, as g ∈ [R]≥n·deg f −P deg xi +1 ,   X (7.2.5) deg µ ≥ deg f · (pL · fpt (f ) − k) + k · n · deg f − deg xi + 1 . P The derivative with respect to k of the right-hand side of (7.2.5) P is (n − 1) deg f − deg xi + 1, which is always nonnegative by our assumption that deg f ≥ deg xi . Thus, the right hand side of (7.2.5) is increasing with respect to k, and as k ≥ 1,   X deg µ ≥ deg f · (pL · fpt (f ) − 1) + n · deg f − deg xi + 1   X ≥ deg f · (pL · λ − n) + n · deg f − deg xi + 1 X
X = pL · deg f · λ − deg xi + 1 = pL − 1 · deg xi + 1, where the second inequality above is a consequence of (7.2.4). Thus, we have arrived at a contraL L diction, and we conclude that (f + g)p ·fpt(f ) ∈ m[p ] , completing the proof.  22 R EFERENCES [BL04] [BMS06] [BMS08] [BMS09] [BS] [BST] [BSTZ09] [Elk87] [EM06] [Fed83] [Har98] [Har06] [Hera] [Herb] [Her12] [HW02] [HY03] [Kol97] [MTW05] [MZ] [Sch07] [Sch08] [Sil09] [Smi97a] [Smi97b] [Smi00] [STZ12] [Tak04] [TW04] Manuel Blickle and Robert Lazarsfeld. An informal introduction to multiplier ideals. In Trends in commutative algebra, volume 51 of Math. Sci. Res. Inst. Publ., pages 87–114. Cambridge Univ. Press, Cambridge, 2004. Nero Budur, Mircea Mustaţǎ, and Morihiko Saito. Roots of bernstein-sato polynomials for monomial ideals: a positive characteristic approach. Math. Res. Lett., 13(1):125–142, 2006. Manuel Blickle, Mircea Mustaţǎ, and Karen E. Smith. Discreteness and rationality of F -thresholds. Michigan Math. J., 57:43–61, 2008. Special volume in honor of Melvin Hochster. Manuel Blickle, Mircea Mustaţă, and Karen E. Smith. F -thresholds of hypersurfaces. Trans. Amer. Math. Soc., 361(12):6549–6565, 2009. Bhargav Bhatt and Anurag K. Singh. F -thresholds of Calabi-Yau hypersurfaces. arXiv:1307.1171. Manuel Blickle, Karl Schwede, and Kevin Tucker. F -singularities via alterations. Preprint, arXiv:1107.3807. Manuel Blickle, Karl Schwede, Shunsuke Takagi, and Wenliang Zhang. Discreteness and rationality of F jumping numbers on singular varieties. Mathematische Annalen, Volume 347, Number 4, 917-949, 2010, 06 2009. Noam D. Elkies. The existence of infinitely many supersingular primes for every elliptic curve over Q. Invent. Math., 89(3):561–567, 1987. Lawrence Ein and Mircea Mustaţă. Invariants of singularities of pairs. In International Congress of Mathematicians. Vol. II, pages 583–602. Eur. Math. Soc., Zürich, 2006. Richard Fedder. F -purity and rational singularity. Trans. Amer. Math. Soc., 278(2):461–480, 1983. Nobuo Hara. A characterization of rational singularities in terms of injectivity of Frobenius maps. Amer. J. Math., 120(5):981–996, 1998. Nobuo Hara. F-pure thresholds and F-jumping exponents in dimension two. Math. Res. Lett., 13(5-6):747–760, 2006. With an appendix by Paul Monsky. Daniel J. Hernández. F -invariants of diagonal hypersurfaces. To appear in Proceedings of the AMS. Daniel J. Hernández. F -pure thresholds of binomial hypersurfaces. To appear in Proceedings of the AMS. Daniel J. Hernández. F -purity of hypersurfaces. Math. Res. Lett., 19(02):1–13, 2012. Nobuo Hara and Kei-Ichi Watanabe. F-regular and F-pure rings vs. log terminal and log canonical singularities. J. Algebraic Geom., 11(2):363–392, 2002. Nobuo Hara and Ken-Ichi Yoshida. A generalization of tight closure and multiplier ideals. Trans. Amer. Math. Soc., 355(8):3143–3174 (electronic), 2003. János Kollár. Singularities of pairs. In Algebraic geometry—Santa Cruz 1995, volume 62 of Proc. Sympos. Pure Math., pages 221–287. Amer. Math. Soc., Providence, RI, 1997. Mircea Mustaţǎ, Shunsuke Takagi, and Kei-ichi Watanabe. F-thresholds and Bernstein-Sato polynomials. In European Congress of Mathematics, pages 341–364. Eur. Math. Soc., Zürich, 2005. Mircea Mustaţǎ and Wenliang Zhang. Estimates for F -jumping numbers and bounds for HartshorneSpeiser-Lyubeznik numbers. arXiv:1110.5687. Karl Schwede. A simple characterization of Du Bois singularities. Compos. Math., 143(4):813–828, 2007. Karl Schwede. Generalized test ideals, sharp F -purity, and sharp test elements. Math. Res. Lett., 15(6):1251– 1261, 2008. Joseph H. Silverman. The arithmetic of elliptic curves, volume 106 of Graduate Texts in Mathematics. Springer, Dordrecht, second edition, 2009. Karen E. Smith. F -rational rings have rational singularities. Amer. J. Math., 119(1):159–180, 1997. Karen E. Smith. Vanishing, singularities and effective bounds via prime characteristic local algebra. In Algebraic geometry—Santa Cruz 1995, volume 62 of Proc. Sympos. Pure Math., pages 289–325. Amer. Math. Soc., Providence, RI, 1997. Karen E. Smith. The multiplier ideal is a universal test ideal. Comm. Algebra, 28(12):5915–5929, 2000. Special issue in honor of Robin Hartshorne. Karl Schwede, Kevin Tucker, and Wenliang Zhang. Test ideals via a single alteration and discreteness and rationality of F -jumping numbers. Math. Res. Lett., 19(1):191–197, 2012. Shunsuke Takagi. F-singularities of pairs and inversion of adjunction of arbitrary codimension. Invent. Math., 157(1):123–146, 2004. Shunsuke Takagi and Kei-ichi Watanabe. On F-pure thresholds. J. Algebra, 282(1):278–297, 2004. D EPARTMENT OF M ATHEMATICS , U NIVERSITY Email address: dhernan@math.utah.edu OF U TAH , S ALT L AKE C ITY, UT 84112 D EPARTMENT OF M ATHEMATICS , U NIVERSITY Email address: lcn8m@virginia.edu OF V IRGINIA , C HARLOTTESVILLE , VA 22904 23 D EPARTMENT OF M ATHEMATICS , U NIVERSITY Email address: ewitt@umn.edu OF M INNESOTA , M INNEAPOLIS , MN 55455 D EPARTMENT OF M ATHEMATICS , U NIVERSITY Email address: wzhang15@unl.edu OF N EBRASKA , L INCOLN , NE 68588 24 

HOMOMORPHISMS INTO TOTALLY DISCONNECTED, LOCALLY COMPACT GROUPS WITH DENSE IMAGE arXiv:1509.00156v2 [math.GR] 3 Jan 2018 COLIN D. REID AND PHILLIP R. WESOLEK Abstract. Let φ : G → H be a group homomorphism such that H is a totally disconnected locally compact (t.d.l.c.) group and the image of φ is dense. We show that all such homomorphisms arise as completions of G with respect to uniformities of a particular kind. Moreover, H is determined up to a compact normal subgroup by the pair (G, φ−1 (L)), where L is a compact open subgroup of H. These results generalize the well-known properties of profinite completions to the locally compact setting. Contents 1. Introduction 2. Preliminaries 3. A general construction for completions 4. Classification and factorization of completions 5. Canonical completions of Hecke pairs 6. Invariant properties of completions 7. Completions compatible with homomorphisms References 1 3 5 8 11 14 18 20 1. Introduction For G a (topological) group, the profinite completion Ĝ of G is the inverse limit of the finite (continuous) quotients of G. One can obtain other profinite groups by forming the inverse limit of a suitable subset of the set of finite quotients of G. Such a profinite group H is always a quotient of Ĝ, and obviously the composition map G → Ĝ → H has dense image. On the other hand, given a (topological) group G, one can ask which profinite groups H admit a (continuous) homomorphism ψ : G → H with dense image. Letting ι : G → Ĝ be the canonical inclusion, it turns out that there is always a continuous quotient map ψe : Ĝ → H such that ψ = ψe ◦ ι; cf. [9, Lemma 3.2.1]. In this way one obtains a complete description of all profinite groups H and homomorphisms ψ : G → H such that the image of G is dense in H: all such groups and morphisms arise exactly by forming inverse limits of suitable sets of finite quotients of G. The first named author is an ARC DECRA fellow. Research supported in part by ARC Discovery Project DP120100996. The second author was supported by ERC grant #278469. 1 2 COLIN D. REID AND PHILLIP R. WESOLEK The aim of the present paper is to extend the well-known and well-established description of homomorphisms to profinite groups with dense image to homomorphisms into totally disconnected locally compact (t.d.l.c.) groups with dense image. That is to say, we will develop a theory of t.d.l.c. completions. Using the language of uniformities, we shall see that this theory generalizes the profinite case. Our approach generalizes previous work by Schlichting ([10]) and Belyaev ([1, §7]); Schlichting’s completion has also been studied in subsequent work, e.g. [11]. The novel contributions of this work are to present a unified theory of t.d.l.c. completions and to identify properties that hold for every completion. 1.1. Statement of results. We shall state our results in the setting of Hausdorff topological groups. If one prefers, the group G to be completed can be taken to be discrete. The topological group setting merely allows for finer control over the completions; for example, given a topological group G, there is often an interesting difference between completions of G as a topological group and completions of G as a discrete group. Definition 1.1. For a topological group G, a (t.d.l.c.) completion map is a continuous homomorphism ψ : G → H with dense image such that H is a t.d.l.c. group. We call H a t.d.l.c. completion of G. All t.d.l.c. completions arise as completions with respect to a certain family of uniformities. Definition 1.2. Let G be a group and let S be a set of open subgroups of G. We say that S is a G-stable local filter if the following conditions hold: (a) S is non-empty; (b) Any two elements of S are commensurate; T (c) Given a finite subset {V1 , . . . , Vn } of S, then ni=1 Vi ∈ S, and given V ≤ W ≤ G such that |W : V | is finite, then V ∈ S implies W ∈ S; (d) Given V ∈ S and g ∈ G, then gV g −1 ∈ S. Each G-stable local filter S is a basis at 1 for a (not necessarily Hausdorff) group topology on G, and thus, there is an associated right uniformity Φr (S) on G. The completion with respect to this uniformity, denoted by ĜS , turns out to be a t.d.l.c. group (Theorem 3.9); we denote by βG,S : G → ĜS the canonical inclusion homomorphism. All t.d.l.c. completions moreover arise in this way. Theorem 1.3 (see Theorem 4.3). If G is a topological group and φ : G → H is a t.d.l.c. completion map, then there is a G-stable local filter S and a unique topological group isomorphism ψ : ĜS → H such that φ = ψ ◦ β(G,S) . We next consider completion maps φ : G → H where a specified subgroup U of G is the preimage of a compact open subgroup of H. In this case, there are two canonical completions of G such that U is the preimage of a compact open subgroup of the completion: the Belyaev completion, denoted by ĜU , and the Schlichting completion, denoted by G//U . These completions are the ‘largest’ and ‘smallest’ completions in the following sense. We denote by βU : G → ĜU and βG/U : G → G//U the canonical inclusion homomorphisms. Theorem 1.4 (see Theorem 5.4). Suppose that G is a group and that φ : G → H is a t.d.l.c. completion map. Letting U ≤ G be the preimage of some compact open subgroup of H, then HOMOMORPHISMS INTO T.D.L.C. GROUPS 3 there are unique continuous quotient maps ψ1 : ĜU → H and ψ2 : H → G//U with compact kernels such that the following diagram commutes: G ✇ ❏❏❏❏ ✇✇ ❏❏βG/U ✇ ✇ ❏❏ φ ✇✇ ❏❏ ✇ $ {✇✇  /H / G//U. βU ĜU ψ1 ψ2 We conclude by identifying several properties that are “independent” of the completion. Doing so requires identifying a notion of size. Two subgroups U and V of a group G are commensurate if U ∩ V has finite index in both U and V . A subgroup W ≤ G is commensurated in G if W is commensurate with gW g −1 for all g ∈ G. We write [U ] for the set of closed subgroups commensurate with U . We call the collection [U ] a size if some (equivalently, any) W ∈ [U ] is commensurated; for [U ] a size, observe that [U ] = [gU g −1 ] for all g ∈ G. A compact open subgroup R of a t.d.l.c. group H is commensurated in H. Given a completion map ψ : G → H, the preimage U := ψ −1 (R) is an open subgroup of G that is commensurated in G. We thus obtain a size [U ], and the size [U ] does not depend on the choice of R. We say that [U ] is the size of ψ. Theorem 1.5 (see §6). Let G be a topological group. For each of the following properties, either every completion of G of size α has the property, or every completion of G of size α fails to have the property. (1) Being σ-compact. (2) Being compactly generated. (3) Being amenable. (4) Being uniscalar. (5) Having a quotient isomorphic to N where N is any specified t.d.l.c. group that has no non-trivial compact normal subgroups. Theorem 1.6 (See §6). Let G be a topological group. For each of the following properties, either every second countable completion of G of size α has the property, or every second countable completion of G of size α fails to have the property. (1) Being elementary. (2) Being elementary with transfinite rank β. Acknowledgments 1.7. The first named author would like to thank Aleksander Iwanow for pointing out the article [1] in response to an earlier preprint. 2. Preliminaries A quotient of a topological group must have closed kernel (such that the resulting quotient topology is Hausdorff). Topological group isomorphism is denoted by ≃. We use “t.d.”, “l.c.”, and “s.c.” for “totally disconnected”, “locally compact”, and “second countable”, respectively. For a topological group G, the set U (G) is defined to be the collection of compact open subgroups of G. Definition 2.1. A Hecke pair is a pair of groups (G, U ) where G is a topological group and U is an open subgroup of G that is commensurated. 4 COLIN D. REID AND PHILLIP R. WESOLEK 2.1. Uniformities. Our approach to completions is via uniform spaces; our discussion of uniform spaces follows [2]. Definition 2.2. Let X be a set. A uniformity is a set Φ of binary relations on X, called entourages, with the following properties: (a) Each A ∈ Φ is reflexive, that is, {(x, x) | x ∈ X} ⊆ A. (b) For all A, B ∈ Φ, there exists C ∈ Φ such that C ⊆ A ∩ B. (c) For all A ∈ Φ, there exists B ∈ Φ such that B ◦ B := {(x, z) | ∃y ∈ X : {(x, y), (y, z)} ⊆ B} is a subset of A. (d) For all A ∈ Φ, there exists B ∈ Φ such that the set {(y, x) | (x, y) ∈ B} is a subset of A. A set with a uniformity is called a uniform space. Two uniformities Φ, Φ′ on a set X are equivalent if every A ∈ Φ contains some A′ ∈ Φ′ and vice versa. Definition 2.3. Let (X, Φ) be a uniform space. A filter f of subsets of X is called a minimal Cauchy filter if f is a ⊆-least filter such that for all U ∈ Φ there is A ∈ f with A × A ⊆ U . A basis at 1 of a topological group gives rise to two canonical uniformities: Definition 2.4. Suppose that G is a topological group (not necessarily Hausdorff) and let B be a basis at 1. The left B-uniformity Φl (B) consists of entourages of the form Ul := {(x, y) | x−1 y ∈ U } (U ∈ B). The right B-uniformity Φr (B) consists of entourages of the form Ur := {(x, y) | xy −1 ∈ U } (U ∈ B). For both the left and right uniformities, the uniformity itself depends on the choice of basis, but the equivalence class of the uniformity does not. We will thus omit references to the basis where it is not significant. (For definiteness, one can take B to be the set of all open identity neighborhoods, but it is often convenient to take another basis.) We will also use the definite article when referring to the left or the right uniformity. Definition 2.5 ([2, II.3.7]). The completion of a uniform space (X, Φ) is defined to be X̂ := {f | f is a minimal Cauchy filter}. along with the uniformity Φ̂ given by entourages of the form Û := {(f, g) | ∃A ∈ f ∩ g with A × A ⊆ U } where U ∈ Φ. There is a canonical, continuous completion map β : X → X̂ which has dense image, defined by x 7→ fx where fx is the minimal Cauchy filter containing the neighborhoods of x. Note that as a topological space, X̂ is determined by the equivalence class of Φ. In particular, if G is a topological group equipped with the right uniformity, we let β : G → Ĝ be the completion map associated to this uniformity. Since β has dense image, there is at most one way to equip Ĝ with a continuous group multiplication and inverse such that β is a homomorphism; if these exist, we can say that Ĝ is a topological group in a canonical sense. The completion Ĝ admits such a group multiplication exactly when the left and right uniformities are equivalent; equivalently, the inverse function preserves the set of minimal Φr -Cauchy filters. HOMOMORPHISMS INTO T.D.L.C. GROUPS 5 Theorem 2.6 ([2, III.3.4 Theorem 1]). Suppose that G is a topological group and that Φr is the right uniformity. The completion Ĝ is a topological group if and only if the inverse map carries minimal Φr -Cauchy filters to minimal Φr -Cauchy filters. Theorem 2.7 ([2, III.3.4 Theorem 1]). Suppose that G is a topological group, Φr is the right uniformity, and the completion Ĝ is a topological group. Then the following hold: (1) The map β : G → Ĝ is a continuous homomorphism with dense image. (2) Multiplication on Ĝ is defined as follows: given f, f ′ ∈ Ĝ, then f f ′ is the minimal Cauchy filter of subsets of G generated by sets AB ⊂ G where A ∈ f and B ∈ f ′ . 3. A general construction for completions We describe a procedure for producing t.d.l.c. completions. The main idea in the construction is to form uniform completions of G with respect to a family of group topologies that are in general coarser than the natural topology of G. Definition 3.1. Let G be a group and let S be a set of open subgroups of G. We say that S is a G-stable local filter if the following conditions hold: (a) S is non-empty; (b) Any two elements of S are commensurate; (c) S is aTfilter in its commensurability class: that is, given a finite subset {V1 , . . . , Vn } of S, then ni=1 Vi ∈ S, and given V ≤ W ≤ G such that |W : V | is finite, then V ∈ S implies W ∈ S; (d) S is stable under conjugation in G: that is, given V ∈ S and g ∈ G, then gV g −1 ∈ S. We say S is a G-stable local filter of size [U ] if in addition S ⊆ [U ]. Remark 3.2. If S is a G-stable local filter, then S is a filter of [V ] for any V ∈ S. Furthermore, [V ] must be stable under the conjugation action of G. Lemma 3.3. Let G be a group and let S be a G-stable local filter. Then S is a basis at 1 for a (not necessarily Hausdorff ) group topology on G. Proof. Let T be the topology generated by all right translates of elements of S. Since S is invariant under conjugation in G, it is clear that every left coset of an element of S is a union of right cosets of elements of S and vice versa; hence T is invariant under inverses. We see that the multiplication map m : (g, h) 7→ gh is continuous with respect to T by observing that given U ∈ S and g, h ∈ G, then m−1 (U gh) contains the open neighborhood (U g) × (g −1 U gh) of (g, h).  We remark that the largest quotientTon which S induces a Hausdorff group topology is G/K where K is the normal subgroup U ∈S U . Equipping G with the group topology induced from S, we can form the left and right uniformities Φl (S) and Φr (S). Definition 3.4. Let S be a G-stable local filter. We define a right S-Cauchy filter f in G to be a minimal Cauchy filter with respect to the uniformity Φr (S). In other words, f is a filter of subsets of G with the following properties: (a) For every V ∈ S, there is exactly one right coset V g of V in G such that V g ∈ f ; (b) Every element of f contains a right coset of some element of S. 6 COLIN D. REID AND PHILLIP R. WESOLEK Left S-Cauchy filters are defined similarly with respect to the left uniformity. Notice that for each g ∈ G, there is a corresponding principal right S-Cauchy filter fg generated by {V g | V ∈ S}. Where the choice of S is clear, we will write ‘Cauchy’ to mean ‘S-Cauchy’. The next series of results will establish that the hypotheses of Theorem 2.6 are satisifed, so completing G with respect to Φr (S) produces a topological group. Lemma 3.5. Let G be a group, N be a commensurated subgroup of G, and g ∈ G. Then there are h1 , . . . , hn ∈ G such that for T all h ∈ G, the set N g ∩ hN is a (possibly empty) union of finitely many right cosets of N ∩ ni=1 hi N h−1 i . Proof. Suppose N g ∩hN 6= ∅ and put R := N ∩g−1 N g. For all h ∈ G, we have (N g ∩hN )R = N g ∩ hN , so N g ∩ hN is a union of left cosets of R in G. The left cosets of R in G that are subsets of N g are exactly those of the form gtR for t ∈ g−1 N g; indeed, xR ⊆ N g ⇔ g −1 xR ⊆ g−1 N g ⇔ g −1 x ∈ g−1 N g. Since R has finite index in g−1 N g, we deduce that only finitely many left cosets of the form gtR exist. It follows that the set {N g ∩ hN | h ∈ G and N g ∩ hN 6= ∅} is finite. Let h1 , . . . , hn ∈ G satisfy {N g ∩ hN | h ∈ G and N g ∩ hN 6= ∅} = {N g ∩ h1 N, . . . , N g ∩ hn N }. T Setting M := N ∩ ni=1 hi N h−1 i , we see that M (N g ∩ hN ) = N g ∩ hN for all h ∈ G with N g ∩ hN 6= ∅. Therefore, N g ∩ hN is a union of right cosets of M . That this union is finite follows as in the previous paragraph.  Lemma 3.6. Let G be a topological group, S be a G-stable local filter, and f be a set of subsets of G. Then f is a right S-Cauchy filter in G if and only if f is a left S-Cauchy filter in G. Proof. By symmetry, it suffices to assume f is a right Cauchy filter and prove that f is a left Cauchy filter. Fixing V ∈ S, the filter f contains some right coset V g of V . Applying Lemma 3.5 to V and g, we produce a finite intersection W of conjugates of V such that W ≤ V and such that for each h ∈ G, the set V g ∩ hV is a union of right cosets of W . Since S is closed under conjugation and finite intersection, we additionally have W ∈ S. Since f is right Cauchy, there is a unique right coset W k of W S contained in f , and since ∅ 6∈ f , it must be the case that W k ⊆ V g. Observing that V g = h∈G V g ∩ hV , there is some h ∈ G such that W k intersects V g ∩ hV . Lemma 3.5 then ensures that indeed W k ⊆ V g ∩ hV , hence hV ∈ f . We conclude that f contains a left coset of V for every V ∈ S. Since f is a filter and ∅ 6∈ f , in fact f contains exactly one left coset of V proving (a) of the definition of a left S-Cauchy filter. Given any element A ∈ f , then A contains V g for some V ∈ S and g ∈ G; in particular, A contains the left coset g(g−1 V g) of g−1 V g ∈ S, proving (b). We thus deduce that f is also a left S-Cauchy filter.  Corollary 3.7. For G a topological group and S a G-stable local filter, the set of right SCauchy filters in G is preserved by the map on subsets induced by taking the inverse. Proof. The map x 7→ x−1 induces a correspondence between the set of left Cauchy filters and the set of right Cauchy filters. By Lemma 3.6, these two sets coincide, so the set of right Cauchy filters is invariant under taking inverses.  HOMOMORPHISMS INTO T.D.L.C. GROUPS 7 We may now apply Theorem 2.6 to produce a completion of G with respect to Φr (S), denoted ĜS . Specifically, • The elements of ĜS are the (right) S-Cauchy filters in G. • The set ĜS is equipped with a uniformity with entourages of the form EU := {(f, f ′ ) | ∃g ∈ G : U g ∈ f ∩ f ′ } (U ∈ S) and topology generated by this uniformity. • The map given by β(G,S) : G → ĜS ; g 7→ fg is continuous, since the topology induced T by Φr (S) is coarser than the topology on G. The image is dense, and the kernel is S. • There is a unique continuous group multiplication on ĜS such that β(G,S) is a homomorphism. In fact, we can define multiplication on ĜS as follows: given f, f ′ ∈ ĜS , then f f ′ is the minimal Cauchy filter of subsets of G generated by sets AB ⊂ G where A ∈ f and B ∈ f ′ . Definition 3.8. For G a topological group and S a G-stable local filter, we call ĜS the completion of G with respect to S. We now establish a correspondence between t.d.l.c. completions and completions with respect to a G-stable local filter. Theorem 3.9. If G is a group and S is a G-stable local filter, then the following hold: (1) The topological group ĜS is a t.d.l.c. completion of G. (2) There is a one-to-one correspondence between U (ĜS ) and S given as follows: For a −1 (R) is the corresponding element of compact open subgroup R of ĜS , the subgroup β(G,S) S, and for V ∈ S, the subgroup β(G,S) (V ) is the corresponding compact open subgroup of ĜS . (3) For V ∈ S, the subgroup β(G,S) (V ) is naturally isomorphic as a topological group to the profinite completion of V with respect to the quotients V /N such that N ∈ S and N E V . Proof. For V ∈ S, set BV := {f ∈ ĜS | V ∈ f }. In view of the definition of the topology of ĜS , the collection {BV | V ∈ S} is a base of identity neighborhoods in ĜS . Moreover, each BV is closed under multiplication and inverse, so BV is a subgroup of ĜS . Therefore, ĜS has a base of identity neighborhoods consisting of open subgroups. Fix U ∈ S and write β := β(G,S) . The image β(U ) is a dense subgroup of BU , so in fact BU = β(U ). Define ) ( \ uV u−1 | U ≥ V ∈ S . NU := u∈U The set NU is precisely the set of elements of S that are normal in U , and these subgroups necessarily have finite index in U . Form ÛS , the profinite completion of U with Qrespect to the finite quotients {U/N | N ∈ NU }. Representing ÛS as a closed subgroup of N ∈NU U/N in the usual way, we define θ : BU → ÛS by setting θ(f ) := (N g)N ∈NU where N g is the unique coset of N in f . One verifies that θ is an isomorphism of topological groups, proving (3). The set {BV | V ∈ S} is thus a basis at 1 of compact open subgroups, so ĜS is a t.d.l.c. group. Since β has a dense image, ĜS is a t.d.l.c. completion of G, verifying (1). 8 COLIN D. REID AND PHILLIP R. WESOLEK Finally, we have a map η : S → U (ĜS ) given by η : V 7→ BV . We observe from the definition of BV that in fact V = β −1 (BV ), so η is injective, with inverse on U (ĜS ) ∩ η(S) given by R 7→ β −1 (R). Let R be a compact open subgroup of ĜS . Since {BV | V ∈ S} is a base of identity neighborhoods in ĜS , there is some V ∈ S such that BV ≤ R. The subgroup BV has finite index in R, since R is compact and BV is open, so β −1 (BV ) = V has finite index in β −1 (R). By the definition of an S-stable local filter, we therefore have β −1 (R) ∈ S. We conclude that η is a bijection of the form required for (2).  For V ∈ S, we will often abuse notation slightly and say that the profinite completion V̂S equals the compact open subgroup β(G,S) (V ) of ĜS . We will also write β in place of β(G,S) when G and S are clear from the context. 4. Classification and factorization of completions Theorem 3.9 gives a method for producing t.d.l.c. completions of a group G. In fact, just as in the profinite case, we see that all t.d.l.c. completions of G arise in this way. We first prove a criterion for whether a homomorphism from G to a t.d.l.c. group factors through ĜS . Proposition 4.1. Let φ : G → H be a continuous homomorphism such that H is a t.d.l.c. group and let S be a G-stable local filter. Then the following are equivalent: (1) There is a continuous homomorphism ψ : ĜS → H such that φ = ψ ◦ β(G,S) . (2) Every open subgroup of H contains φ(V ) for some V ∈ S. Moreover, if (1) holds, then ψ is unique. Proof. Suppose that (1) holds and let U be an open subgroup of H. The preimage ψ −1 (U ) is an open subgroup of ĜS , so R ≤ ψ −1 (U ) for some compact open subgroup R of ĜS . Theorem 3.9 ensures that V := β −1 (R) ∈ S, and V is such that φ(V ) = ψ(β(V )) ≤ U . We conclude (2). Additionally, since β(G) is dense in ĜS , the equation φ = ψ ◦ φ determines the restriction of ψ to β(G) and hence determines ψ uniquely as a continuous map. Conversely, suppose that every open subgroup of H contains φ(V ) for some V ∈ S. For f ∈ ĜS , define \ fˆ := {φ(V g) | g ∈ G, V ∈ S, and V g ∈ f }. Since H is a t.d.l.c. group, the open subgroups of H form a basis of identity neighborhoods. By (1), it follows that the set {φ(V ) | V ∈ S} contains arbitrarily small subgroups of H, so its intersection 1̂ is the trivial group. For general f ∈ ĜS , |fˆ| ≤ 1 since fˆ is an intersection of cosets of arbitrarily small subgroups of H. Fix R ∈ U (H) and Q ∈ S such that φ(Q) ≤ R, so in particular φ(Q) is compact. Fix h ∈ G such that Qh ∈ f . Letting g ∈ G and V ∈ S be such that V g ∈ f , there is some k ∈ G such that (V ∩Q)k ∈ f . The collection f is a proper filter, so we see that (V ∩Q)k ⊆ V g ∩Qh. T In particular, φ(V ∩ Q)k is contained in φ(V g). We can thus write fˆ as C∈C C where C = {φ(V g) | g ∈ G, V ∈ S, V g ⊆ Qh and V g ∈ f }. T For any Q1 g1 , . . . , Qn gn ∈ f such that Qi gi ⊆ Qh, the intersection ni=0 Qi gi is non-empty as it is an element of f . Thus C is a family of closed subsets of the compact set φ(Qg) with the finite intersection property. We conclude that fˆ is nonempty, so |fˆ| = 1. We now define a function ψ : ĜS → H by setting ψ(f ) to be the unique element of fˆ. One verifies that ψ is a homomorphism satisfying φ = ψ ◦ β. HOMOMORPHISMS INTO T.D.L.C. GROUPS 9 To see that ψ is continuous, fix (Vα )α∈I a basis at 1 of compact open subgroups for H and for each α ∈ I, choose Wα ∈ S such that φ(Wα ) ≤ Vα . As in the previous paragraph, T we observe that α∈I φ(Wα ) = {1}. Consider a convergent net fδ → f in ĜS . For each subgroup Wα , there is ηα ∈ I such that fγ fγ−1 contains Wα for all γ, γ ′ ≥ ηα . We conclude ′ ′ that ψ(fγ fγ−1 ′ ) ∈ φ(Wα ) ≤ Vα for all such γ, γ . In other words, ψ(fδ ) is a Cauchy net in H with respect to the right uniformity of H. Since H is locally compact, it is complete with respect to this uniformity, and ψ(fδ ) converges. It now follows that ψ(f ) = limδ∈I ψ(fδ ), so ψ is continuous as claimed, completing the proof that (2) implies (1).  As a corollary, we note the case of Proposition 4.1 when H is itself the completion with respect to a G-stable local filter. Corollary 4.2. Let G be a group with G-stable local filters S1 and S2 . Then the following are equivalent: (1) There is a continuous homomorphism ψ : ĜS1 → ĜS2 such that β(G,S2 ) = ψ ◦ β(G,S1 ) . (2) For all V2 ∈ S2 , there exists V1 ∈ S1 such that V1 ≤ V2 . Proof. Set β1 := β(G,S1 ) and β2 := β(G,S2 ) . Suppose that (1) holds and take V2 ∈ S2 . There exists L ∈ U (ĜS2 ) such that V2 = β2−1 (L). Applying Proposition 4.1 with the filter S1 gives V1 ∈ S1 such that β2 (V1 ) ≤ L. We deduce that V1 ≤ V2 , verifying (2). Conversely, suppose that (2) holds. For every open subgroup U of ĜS2 , we have β2 (V2 ) ≤ U for some V2 ∈ S2 by Theorem 3.9(2). By hypothesis, there exists V1 ∈ S1 such that V1 ≤ V2 , so β2 (V1 ) ≤ U . The conclusion (1) now follows by Proposition 4.1.  We will now demonstrate that all t.d.l.c. completions arise as completions with respect to G-stable local filters. Theorem 4.3. If G is a group and H is a t.d.l.c. completion of G via φ : G → H, then the set of all preimages of compact open subgroups of H is a G-stable local filter S, and moreover, there is a unique topological group isomorphism ψ : ĜS → H such that φ = ψ ◦ β(G,S) . Proof. Setting S := {φ−1 (M ) | M ∈ U (H)}, one verifies that S is a G-stable local filter. Define ψ : ĜS → H as in Proposition 4.1, so ψ is the unique continuous homomorphism such that φ = ψ ◦ β. It is easily verified that ψ is injective. The group ĜS has a base of identity neighborhoods of the form {V̂S | V ∈ S}. Given V ∈ S, say that φ−1 (P ) = V for P ∈ U (H). The image φ(V ) is dense in P , and the image β(V ) is dense in V̂S . Since φ = ψ ◦ β, we infer that ψ(V̂S ) is also a dense subgroup of P . The map ψ is continuous and V̂S is compact, so in fact ψ(V̂S ) = P . Hence, ψ is an open map. Since the image of ψ is dense, it follows that ψ is surjective and thereby bijective. The map ψ is an open continuous bijective homomorphism. We conclude that ψ is an isomorphism of topological groups, completing the proof.  Remark 4.4. Theorem 4.3 shows that, up to isomorphism, all t.d.l.c. completions of a group G have the form ĜS for S some G-stable local filter. This applies to the particular case where the t.d.l.c. completion is a profinite group. For instance, the profinite completion of a discrete group G is ĜS where S is the set of all subgroups of finite index in G. We conclude this section by characterizing properties of the kernel of the homomorphism ψ obtained in Corollary 4.2 in terms of the G-stable local filters. In light of Theorem 4.3, 10 COLIN D. REID AND PHILLIP R. WESOLEK these characterizations in fact apply to any continuous homomorphism ψ : H1 → H2 where H1 and H2 are completions of G, via taking Si to be the set of preimages of the compact open subgroups of Hi . Proposition 4.5. Let G be a group with G-stable local filters S1 and S2 such that for all V2 ∈ S2 there is V1 ∈ S1 with V1 ≤ V2 . Let ψ : ĜS1 → ĜS2 be the unique continuous homomorphism such that β(G,S2 ) = ψ ◦ β(G,S1 ) , as given by Corollary 4.2. Then the following holds: (1) The kernel of ψ is compact if and only if S2 ⊆ S1 . If S2 ⊆ S1 , then ψ is also a quotient map. (2) The kernel of ψ is discrete if and only if there exists V1 ∈ S1 such that for all W1 ∈ S1 there is W2 ∈ S2 with V1 ∩ W2 ≤ W1 . (3) The kernel of ψ is trivial if and only if for all V1 ∈ S1 , \ V1 = {V2 ∈ S2 | V1 ≤ V2 }. Proof. Set β1 := β(G,S1 ) and β2 := β(G,S2 ) . Proof of (1). Suppose that ker ψ is compact. Taking V ∈ S2 , the group ψ −1 (V̂S2 ) is a compact open subgroup of ĜS such that the preimage under β1 is V . Given the characterization of compact open subgroups of ĜS1 established in Theorem 3.9(2), it follows that V ∈ S1 . Hence, S2 ⊆ S1 . Conversely, suppose that S2 ⊆ S1 and let V ∈ S2 . Given the construction of ψ in the proof of Proposition 4.1, we see that ker ψ ≤ V̂S1 ; in particular, ker ψ is compact. We see from Theorem 3.9 that ψ(V̂S1 ) = β2 (V ) = V̂S2 , so the image of ψ is open. Since the image of ψ is also dense, it follows that ψ is surjective. We conclude that ψ is a quotient map with compact kernel, as required. Proof of (2). Suppose that ker ψ is discrete, so there is V1 ∈ S1 such that ker ψ ∩ (Vˆ1 )S1 = {1}. The map ψ restricts to a topological group isomorphism from (Vˆ1 )S1 to its image. Take W1 ∈ S1 and set Y := V1 ∩ W1 . The subgroup (Ŷ )S1 is open in (Vˆ1 )S1 , so ψ((Ŷ )S1 ) is open in ψ((Vˆ1 )S1 ). There thus exists W2 ∈ S2 such that ψ((Vˆ1 )S1 ) ∩ (Ŵ2 )S2 ≤ ψ((Ŷ )S1 ). Since ψ is injective on (Vˆ1 )S1 , it follows that (Vˆ1 )S1 ∩ ψ −1 ((Ŵ2 )S2 ) ≤ (Ŷ )S1 . We deduce that V1 ∩ W2 ≤ Y ≤ W1 since ψ ◦ β1 = β2 . Conversely, suppose that V1 ∈ S1 is such that for all W1 ∈ S1 there is W2 ∈ S2 with V1 ∩ W2 ≤ W1 . We claim that (Vˆ1 )S1 intersects ker ψ trivially, which will demonstrate that ker ψ is discrete. Suppose that f ∈ (Vˆ1 )S1 ∩ ker ψ and suppose for contradiction that f is non-trivial. There is then some W1 ∈ S1 such that f contains a nontrivial coset W1 g of W1 ; we may assume that W1 ≤ V1 . Since f ∈ (Vˆ1 )S1 , we additionally have g ∈ V1 . By the hypotheses, there is W2 ∈ S2 such that V1 ∩ W2 ≤ W1 , and there is Y ∈ S1 such that Y ≤ V1 ∩ W2 . Letting g′ ∈ G be such that Y g′ ∈ f , it must be the case that Y g′ ⊆ W2 since ψ(f ) is trivial. On the other hand, Y g ′ ⊆ W1 g, so g ′ ∈ V1 . We now see that g′ ∈ V1 ∩ W2 ≤ W1 and that W1 g = W1 g′ = W1 . Thus W1 g is the trivial coset of W1 , a contradiction. We conclude that (Vˆ1 )S1 ∩ ker ψ = {1} as claimed. HOMOMORPHISMS INTO T.D.L.C. GROUPS 11 Proof of (3). Suppose that ψ is injective. Let V1 ∈ S1 and set R := {V2 ∈ S2 | V1 ≤ V2 }. The subgroup (Vˆ1 )S1 is a compact open subgroup of ĜS1 , so K := ψ((Vˆ1 )S1 ) is a compact, hence profinite, subgroup of ĜS2 . It follows that K is the intersection of all compact open subgroups of ĜS2 that contain K. All compact open subgroups of ĜS2 are of the form ŴS2 for some W ∈ S2 . For ŴS2 ≥ K compact and open, we thus have β2−1 (K) ≤ W , and since β1 = ψ ◦ β2 , we deduce further that T V1 = β1−1 ((V̂T1 )S1 ) ≤ W . Hence, K = {ŴS2 | W ∈ R}. For g ∈ R, it is the case that β2 (g) ∈ β2 (W ) for all W ∈ R, so β2 (g) ∈ K. In other words, ψ ◦ β1 (g) ∈ ψ((Vˆ1 )S1 ). T As ψ is injective, β1 (g) ∈ (Vˆ1 )S1 , and thus, g ∈ β1−1 ((Vˆ1 )S1 ) = V1 ,Tshowing that R ⊆ V1 . On the other hand it is clear from the definition of R that V1 ⊆ R, so equality holds as required. T Conversely, suppose that for all V1 ∈ S1 , it is the case that V1 = {V2 ∈ S2 | V1 ≤ V2 }. Fix f ∈ ker ψ. For each V1 ∈ S1 , we may write f = β1 (g)u for g ∈ G and u ∈ (V̂1 )S1 , since β1 (G) is dense in ĜS1 , and it follows that β2 (g) ∈ ψ((V̂1 )S1 ). We now infer that g ∈ V2 for all V2 ∈ S2 such that V2 ≥ V1 ; by our hypothesis, it follows that g ∈ V1 . Recalling that f = β1 (g)u, it follows that f ∈ (V̂1 )S1 . The subgroups (V̂1 )S1 form a basis at 1, so indeed f = 1. We conclude that ψ is injective as required.  5. Canonical completions of Hecke pairs We now consider the t.d.l.c. completions H of G such that a specified subgroup U of G is the preimage of a compact open subgroup of the completion. The pair (G, U ) is then a Hecke pair. Let us make several definitions to organize our discussion of t.d.l.c. completions. Recall that all preimages of compact open subgroups of a given range group H are commensurate, so the following definition does not depend on the choice of U . Definition 5.1. Given a t.d.l.c. completion map φ : G → H, the size of φ is defined to be [U ], where U is the preimage of a compact open subgroup of H. We also say that [U ] is the size of H when the choice of completion map is not important, or can be inferred from the context. A completion map for a Hecke pair (G, U ) is a continuous homomorphism φ : G → H with dense image such that H is a t.d.l.c. group and U is the preimage of a compact open subgroup of H. We say that H is a completion of (G, U ). When H is also second countable, we call H a second countable completion. Given a Hecke pair (G, U ), there are two canonical G-stable local filters containing U , defined as follows: The Belyaev filter is [U ]. The Schlichting filter SG/U for (G, U ) is the filter of [U ] generated by the conjugacy class of U – that is, ) ( n \ gi U gi−1 ≤ V . SG/U := V ∈ [U ] | ∃g1 , . . . , gn ∈ G such that i=1 Definition 5.2. The Belyaev completion for (G, U ), denoted by ĜU , is defined to be Ĝ[U ] . The canonical inclusion map β(G,[U ]) is denoted by βU . The Schlichting completion for 12 COLIN D. REID AND PHILLIP R. WESOLEK (G, U ), denoted by G//U , is defined to be ĜSG/U . The canonical inclusion map β(G,SG/U ) is denoted by βG/U . Remark 5.3. We stress that the commensurability class [U ] does not determine the Schlichting filter SG/U ; indeed, the only situation when there is only one Schlichting filter of a given size is when the Schlichting filter is equal to the Belyaev filter of that size. Given any G-stable local filter S of size [U ] that contains U , we have SG/U ⊆ S ⊆ [U ]. Amongst G-stable local filters that contain U , the Schlichting filter is minimal whilst [U ] is maximal. The Belyaev and Schlichting completions are thus maximal and minimal completions of a Hecke pair in the following strong sense: Theorem 5.4. Suppose that G is a group and that φ : G → H is a completion map. Letting U ≤ G be the preimage of some compact open subgroup of H, then (G, U ) is a Hecke pair, H is a completion of (G, U ), and there are unique continuous quotient maps ψ1 : ĜU → H and ψ2 : H → G//U with compact kernels such that the following diagram commutes: G ✇ ❏❏❏❏ ✇✇ ❏❏βG/U ✇ ✇ ❏❏ φ ✇✇ ❏❏ ✇ $ {✇✇  /H / G//U. βU ĜU ψ1 ψ2 Proof. Fix L a compact open subgroup of H and set U := φ−1 (L). The subgroup U is an open commensurated subgroup of G, so (G, U ) is a Hecke pair. Since φ has dense image, we conclude that H is a completion of (G, U ). Set S := {φ−1 (V ) | V ∈ U (H)}. By Theorem 4.3, S is a G-stable local filter, and there is a unique topological group isomorphism ψ : ĜS → H such that φ = ψ ◦ β(G,S) . Observe that U ∈ S by Theorem 3.9(2). Since S contains U , Proposition 4.5 ensures there are unique continuous quotient maps π1 and π2 with compact kernels such that the following diagram commutes: G ❏❏❏ ✈✈ ❏❏❏βG/U ✈✈ ✈ ❏❏❏ β ✈ (G,S) ❏❏ ✈✈  ✈ % z ✈ / ĜS / G//U. βU ĜU π1 π2 It follows that ψ1 := ψ ◦ π1 and ψ2 := π2 ◦ ψ −1 make the desired diagram commute and both are continuous quotient maps with compact kernels. Uniqueness follows since ψ, π1 , and π2 are unique.  Theorem 5.4 shows all possible completions of a Hecke pair (G, U ) differ only by a compact normal subgroup. The locally compact, non-compact structure of a t.d.l.c. completion thus depends only on the Hecke pair; contrast this with the many different profinite completions a group can admit. We give precise statements illustrating this phenomenon in Section 6. We conclude this section by making two further observations. First, the Schlichting completion has a natural description. Suppose (G, U ) is a Hecke pair and let σ(G,U ) : G → Sym(G/U ) be the permutation representation given by left multiplication. We consider Sym(G/U ) to be equipped with the topology of pointwise convergence. Proposition 5.5. For (G, U ) a Hecke pair, there is a unique topological group isomorphism ψ : G//U → σ(G,U ) (G) such that σ(G,U ) = ψ ◦ βG/U . HOMOMORPHISMS INTO T.D.L.C. GROUPS 13 b is a Proof. For Y ⊆ G, set Yb := σ(G,U ) (Y ). The orbits of σ(U ) are finite on G/U , so U b b = Stab b (U ), hence it is open. It now follows that G profinite group. On the other hand, U G is a t.d.l.c. completion of G. b is given by stabilizers of finite sets of cosets. Such stabilizers A basis for the topology on G T b σ(g−1 ) with F ⊆ G finite. For every V ∈ U (G), b the are exactly of the form g∈F σ(g)U T −1 subgroup σ(G,U ) (V ) therefore contains g∈F gU g −1 for some F ⊆ G finite. The G-stable local b is thus exactly the Schlichting filter SG/U . Theorem 4.3 filter S := {σ −1 (V ) | V ∈ U (G)} (G,U ) now implies the proposition.  Via Theorem 5.4 and Proposition 5.5, we recover a result from the literature. Corollary 5.6 (Shalom–Willis [11, Lemma 3.5 and Corollary 3.7]). Suppose that G is a t.d.l.c. group and that φ : G → H is a completion map. If U is the preimage of a compact open subgroup of H, then there is a unique quotient map ψ : H → G//U and closed embedding ι : G//U → Sym(G/U ) such that σ(G,U ) = ι ◦ ψ ◦ φ. We next show the Belyaev completion satisfies a stronger universality property. Theorem 5.7. Let φ : G → H be a continuous homomorphism such that H is a t.d.l.c. group. Suppose that U is a commensurated open subgroup of G such that φ(U ) is profinite. Then there is a unique continuous homomorphism ψ : ĜU → H such that φ = ψ ◦ βU . If in addition φ has dense image and φ(U ) is open in H, then ψ is a quotient map. Proof. Let β := βU and set L := φ(U ). Let V be an open subgroup of H. Then V ∩ L is open and of finite index in L, since L is compact. In particular, we see that W = φ−1 (V ) ∩ U is an open subgroup of U of finite index. For all open subgroups V of H, there is thus W ∈ [U ] such that φ(W ) ≤ V . We now obtain the unique continuous homomorphism ψ : ĜU → H such that φ = ψ ◦ β via Proposition 4.1. Now suppose that φ(U ) is open in H and that φ has dense image. The group Û := β(U ) is a compact open subgroup of ĜU , so ψ(Û ) is compact in H; in particular, ψ(Û ) is closed in H. We see that ψ(Û ) ≥ φ(U ), so ψ(Û ) is indeed open in H. The image of ψ is therefore an open and dense subgroup of H, so ψ is surjective. Since H is a Baire space, it follows that ψ is a quotient map.  A standard universal property argument now shows that Theorem 5.7 characterizes the Belyaev completion up to topological isomorphism, so one can take Theorem 5.7 as the definition of the Belyaev completion. Remark 5.8. We see that the problem of classifying all continuous homomorphisms with dense image from a specified group G (possibly discrete) to an arbitrary t.d.l.c. group can, in principle, be broken into two steps: (1) Classify the possible sizes of completion; in other words, classify the commensurability classes α = [U ] of open subgroups of G that are invariant under conjugation. (This typically amounts to classifying commensurated open subgroups.) (2) For each such class α, form the Belyaev completion Ĝα and classify the quotients of Ĝα with compact kernel. A number of researchers have already considered (1). Shalom and Willis classify commensurated subgroups of many arithmetic groups in [11]. Other examples include classifications 14 COLIN D. REID AND PHILLIP R. WESOLEK of commensurated subgroups of almost automorphism groups ([5]) and Burger-Mozes simple universal groups ([6]). 6. Invariant properties of completions By Theorem 5.4, the t.d.l.c. completions of a group G of a given size differ only by a compact normal subgroup, so ought to have the same “non-compact” properties. We here make several precise statements showing that this intuition has substance. 6.1. First invariant properties. Proposition 6.1. Let G be a group and let H be a t.d.l.c. completion of G of size α. Then, (1) H is σ-compact if and only if |G : W | is countable for some (equivalently, any) W ∈ α; (2) H is compactly generated if and only if G is generated by finitely many left cosets of W for some (equivalently, any) W ∈ α. Proof. Let β : G → H be the completion map and V ∈ U (H). By hypothesis, U := β −1 (V ) is an element of α. For (1), if H is σ-compact, then V has only countably many left cosets in H. Since U = β −1 (V ), it follows that |G : U | is countable. Conversely, suppose that |G : W | is countable for some W ∈ α. It follows that |G : U | is countable. Since β(G) is dense in H, there are only countably many left cosets of β(U ) in H, so H is σ-compact. For (2), if H is compactly generated, then since β(G) is dense in H, there exists a finite symmetric A ⊆ β(G) such that H = hAiV ; see for instance [12, Proposition 2.4]. Say A = β(B) for a finite subset B of G. For every g ∈ G, there thus exists v ∈ V and g′ ∈ hBi such that β(g) = β(g ′ )v. Since β −1 (V ) = U , it follows further that v = β(u) for some u ∈ U . Thus, β(hB, U i) = β(G), and as ker β ≤ U , we infer that G = hB, U i. In particular, G is generated by finitely many left cosets of U . Conversely, suppose that G is generated by finitely many left cosets of some W ∈ α. It follows that G is generated by finitely many left S cosets b1 U, b2 U, . . . , bn U of U . The image β( ni=1 bi U ) generates a dense subgroup of H, and S hence the compact subset X := ni=1 β(bi )β(U ) generates a dense open subgroup of H and therefore generates H.  Corollary 6.2. For each of the following properties, either every completion of a group G of size α has the property, or every completion of G of size α fails to have the property. (1) Being σ-compact. (2) Being compactly generated. We next consider possible quotient groups and amenability. Proposition 6.3. For each of the following properties, either every completion of a group G of size α has the property, or every completion of G of size α fails to have the property. (1) Having a quotient isomorphic to N where N is any specified t.d.l.c. group that has no non-trivial compact normal subgroups. (2) Being amenable. Proof. For i ∈ {1, 2}, let φi : G → Hi be a completion map of size α and let Ui be the preimage under φi of a compact open subgroup of Hi . The subgroup Ui is a member of α, so by Theorem 5.4, there are quotient maps πi : ĜU → Hi with compact kernel. It therefore suffices to show that for any t.d.l.c. group H and compact normal subgroup K of H, the group H has the property if and only if H/K does. HOMOMORPHISMS INTO T.D.L.C. GROUPS 15 For (1), if π : H → N is a quotient map, then π(K) is a compact normal subgroup of N . Since N has no non-trivial compact normal subgroup, we deduce that K ≤ ker π, so N is a quotient of H if and only if N is a quotient of H/K. For (2), recall that every compact subgroup is amenable and that if L is a closed normal subgroup of the locally compact group H, then H is amenable if and only if both H/L and L are amenable. Since K is compact, we deduce that H is amenable if and only if H/K is amenable.  Let us now consider topological countability axioms. These are more delicate and depend on the choice of G-stable local filter, as opposed to only the size. Proposition 6.4. Suppose G is a topological group and S is a G-stable local filter. Then ĜS is first countable if and only if the set {V ∈ S | V ≤ U } is countable, for some (equivalently, any) U ∈ S. Proof. Fix U ∈ S. Let β : G → ĜS be the completion map and set SU := {V ∈ S | V ≤ U }. By Theorem 3.9, we have |SU | = |O| where O is the set of open subgroups of U . If SU is countable, then O is a countable base of identity neighborhoods, so ĜS is first countable. Conversely if ĜS is first countable, then there is a countable base B of identity neighborhoods consisting of open subsets of U . Since U is profinite, each B ∈ B contains a subgroup of U of finite index, so there are only finitely many open subgroups of U that contain B. Hence, O is countable, implying that SU is countable.  Proposition 6.5. Let G be a topological group and S be a G-stable local filter. Then ĜS is second countable if and only if {V ∈ S | V ≤ U } and |G : U | are countable, for some (equivalently, any) U ∈ S. Proof. Via [4, (5.3)], a locally compact group is second countable if and only if it is σ-compact and first countable. The proposition now follows from Propositions 6.4 and 6.1.  Corollary 6.6. If (G, U ) is a Hecke pair such that |G : U | is countable, then the Schlichting completion G//U is a t.d.l.c.s.c. group. 6.2. Elementary groups. Let us now consider a more complicated algebraic property, the property of being an elementary group. Definition 6.7. The class E of elementary groups is the smallest class of t.d.l.c.s.c. groups such that (i) E contains all second countable profinite groups and countable discrete groups. (ii) E is closed under taking closed subgroups. (iii) E is closed under taking Hausdorff quotients. (iv) E is closed under forming group extensions. S (v) If G is a t.d.l.c.s.c. group and G = i∈N Oi where (Oi )i∈N is an ⊆-increasing sequence of open subgroups of G with Oi ∈ E for each i, then G ∈ E . We say that E is closed under countable increasing unions. The operations (ii)-(v) are often called the elementary operations. It turns out operations (ii) and (iii) follow from the others, and (iv) can be weakened to (iv)′ : E is closed under extensions of profinite groups and discrete groups. These results are given by [12, Theorem 1.3]. 16 COLIN D. REID AND PHILLIP R. WESOLEK Remark 6.8. If G is a t.d.l.c.s.c. group that is non-discrete, compactly generated, and topologically simple, then G is not a member of E . The class E is thus strictly smaller than the class of all t.d.l.c.s.c. groups. The class of elementary groups comes with a canonical successor ordinal valued rank called the decomposition rank and denoted by ξ(G); see [12, Section 4]. The key property of the decomposition rank that we will exploit herein is that it is well-behaved under applying natural group building operations. T For a t.d.l.c. group G, recall that the discrete residual of G is defined to be Res(G) := {O E G | O open}. Proposition 6.9. For G a non-trivial elementary group, the following hold. (1) If H is a t.d.l.c.s.c. group, and ψ : H → G is a continuous, injective homomorphism, then H is elementary with ξ(H) ≤ ξ(G). ([12, Corollary 4.10]) (2) If L E G S is closed, then ξ(G/L) ≤ ξ(G). ([12, Theorem 4.19]) (3) If G = i∈N Oi with (Oi )i∈N an ⊆-increasing sequence of compactly generated open subgroups of G, then ξ(G) = supi∈N ξ(Res(Oi )) + 1. If G is compactly generated, then ξ(G) = ξ(Res(G)) + 1. ([12, Lemma 4.12]) (4) If G is residually discrete, then G is elementary with ξ(G) ≤ 2. ([12, Observation 4.11]) (5) If G is elementary and lies in a short exact sequence of topological groups {1} → N → G → Q → {1}, then ξ(G) ≤ (ξ(N ) − 1) + ξ(Q). (Here (ξ(N ) − 1) denotes the predecessor of ξ(N ), which exists as ξ(N ) is a successor ordinal.) ([8, Lemma 3.8]) Proposition 6.10. Let G be a group. (1) Either every second countable completion of G of size α is elementary, or all second countable completions of G of size α are non-elementary. (2) If every second countable completion of size α is elementary, then for any two second countable completions H and L of G with size α, we have ξ(L) ≤ 1 + ξ(H). (3) If some second countable completion of size α is elementary with transfinite rank β, then every second countable completion of size α is elementary with rank β. Proof. Let H and L be second countable completions of G of size α. By Theorem 4.3 we may assume H = ĜS1 and L = ĜS2 where S1 , S2 ⊆ α are G-stable local filters. Let S3 be the smallest G-stable local filter containing S1 ∪ S2 . Via Proposition 6.5, {V ∈ Si | V ≤ U } is countable for i ∈ {1, 2} and U ∈ Si . It follows that {V ∈ S3 | V ≤ U } is countable for U ∈ S1 , and hence ĜS3 is first countable. By Corollary 6.2(1), ĜS3 is also σ-compact and hence second countable. By Proposition 4.5, both ĜS1 and ĜS2 are quotients of ĜS3 with compact kernel. It follows that ĜSi is elementary for i ∈ {1, 2} if and only if ĜS3 is elementary; in particular, it is not possible for ĜS1 to be elementary and ĜS2 to be non-elementary. This proves (1). Moreover, if ĜS3 is elementary, then via Proposition 6.9 we obtain the inequalities ξ(ĜSi ) ≤ ξ(ĜS3 ) ≤ 1 + ξ(ĜSi ) (i ∈ {1, 2}). In particular, if ξ(ĜS3 ) is finite then ξ(ĜS1 ), ξ(ĜS2 ) ∈ {ξ(ĜS3 ) − 1, ξ(ĜS3 )}, and if ξ(ĜS3 ) is transfinite then it is equal to both ξ(ĜS1 ) and ξ(ĜS2 ). This proves (2) and (3).  HOMOMORPHISMS INTO T.D.L.C. GROUPS 17 It can be the case that there are no second countable completions of a given size. In light of Corollary 6.6, however, a group G has a second countable completion of size α if and only if |G : U | is countable for some U ∈ α. With this in mind, we make the following definition. Definition 6.11. A size α of a group G is called elementary if there is U ∈ α such that |G : U | is countable and some (all) second countable completions are elementary. A Hecke pair (G, U ) is called elementary if |G : U | is countable and some (all) s.c. completions are elementary. 6.3. The scale function and flat subgroups. We conclude this section by considering the scale function and flat subgroups in relation to completions; these concepts were introduced in [13] and [14] respectively, although the term “flat subgroup” is more recent. Definition 6.12. For G a t.d.l.c. group, the scale function s : G → Z is defined by s(g) := min{|gU g −1 : gU g −1 ∩ U | | U ∈ U (G)}. A compact open subgroup U of G is tidy for g ∈ G if it achieves s(g). We say g is uniscalar if s(g) = s(g−1 ) = 1. A subset X of G is flat if there exists a compact open subgroup U of G such that for all x ∈ X, the subgroup U is tidy for x; in this case, we say U is tidy for X. If X is a finitely generated flat subgroup, the rank of X is the least number of generators for the quotient group X/{x ∈ X | s(x) = 1}. The scale function and flatness are clearly locally compact non-compact phenomena. In relation to t.d.l.c. completions, they only depend on the size [U ]. Proposition 6.13. For φ : G → H a t.d.l.c. completion of size [U ], the following hold: (1) For ŝ and s the scale functions for ĜU and H, s ◦ φ = ŝ ◦ βU . (2) For X ⊆ G, the subset φ(X) is flat if and only if βU (X) is flat. (3) If K ≤ G is a finitely generated subgroup, then φ(K) is flat with rank k if and only if βU (K) is flat with rank k. Proof. By Theorem 5.4, we can factorize φ as φ = π ◦βU , where π : ĜU → H is a quotient map with compact kernel. The result [7, Lemma 4.9] ensures that s ◦ π = ŝ, hence s ◦ φ = ŝ ◦ βU , proving (1). Appealing again to [7, Lemma 4.9], if U is tidy for g in ĜU , then π(U ) is tidy for π(g) in H, and conversely if V is tidy for π(g) in H, then π −1 (V ) is tidy for g in ĜU . Therefore, if βU (X) has a common tidy subgroup, then so does φ(X). Conversely, if φ(X) has a common tidy subgroup V in H, then π −1 (V ) is a common tidy subgroup for βU (X). We conclude that φ(X) has a common tidy subgroup if and only if βU (X) does, verifying (2). Finally, if K is a subgroup of G, then φ(K) is a flat subgroup of H if and only if βU (K) is a flat subgroup of ĜU by (2). The rank of φ(K) is the number of generators of the factor φ(K)/LH where LH := {x ∈ φ(K) | s(x) = 1}. Letting LĜU be the analogous subgroup of ĜU , it follows from (1) that the map π induces an isomorphism π̃ : βU (K)/LĜU → φ(K)/LH . We conclude that βU (K) has rank k if and only if φ(K) has rank k, proving (3).  The next corollary is immediate from Proposition 6.13 and the fact the scale function is continuous (see [13]). Corollary 6.14. For G a group and U a commensurated open subgroup of G, either all t.d.l.c. completions of G of size α are uniscalar, or no completion of size α is uniscalar. 18 COLIN D. REID AND PHILLIP R. WESOLEK 7. Completions compatible with homomorphisms e of For an injective homomorphism θ : G → L, we may wish to find a t.d.l.c. completion G e G such that θ extends to an injective homomorphism from G to L. More precisely, we say the e is compatible with θ if there is a continuous injective t.d.l.c. completion map β : G → G e homomorphism ψ : G → L such that θ = ψ ◦ β; note that in this case ψ is necessarily unique. Here we do not insist that L be locally compact; indeed, in many interesting examples, L itself will not be locally compact (see Remark 7.3 below). We can characterize the t.d.l.c. completions compatible with θ in terms of commensurated subgroups. Theorem 7.1. Let θ : G → L be an injective continuous homomorphism of topological groups. (1) Suppose that H is an open commensurated subgroup of G such that the closure θ(H) of θ(H) in L is profinite and set H ∗ := θ −1 (θ(H)). Then H ∗ is open and commensurated in G, and there is a t.d.l.c. completion map β : G → ĜH,θ compatible with θ such that H ∗ is the preimage of a compact open subgroup of ĜH,θ . Moreover, β is unique up to isomorphisms of ĜH,θ and is determined by the pair ([H ∗ ], θ). e is a t.d.l.c. completion of G compatible with θ and ψ : G e→L (2) Suppose that β : G → G e is such that θ = ψ ◦ β, let U be a compact open subgroup of G, and set H := β −1 (U ). Then H is a commensurated subgroup of G that is the preimage of a profinite subgroup e ≃ ĜH,θ . θ(H) = ψ(U ) of L, and G Proof. For (1), H is an open commensurated subgroup of G such that the closure K = θ(H) of θ(H) in L is profinite. The image θ(G) additionally commensurates K; consider, for instance, [5, Lemma 2.7]. We thus conclude that H ∗ := θ −1 (K) is commensurated in G. Now let R be the set of closed subgroups of L that are commensurate with K and let S be the set of θ-preimages of elements of R. The collection S forms a G-stable local filter, and setting ĜH,θ := ĜS , we obtain a t.d.l.c. completion β : G → ĜH,θ . Let L′ be the group hθ(G), Ki, equipped with the unique group topology such that the inclusion of K into L′ is continuous and open. The map θ induces a continuous homomorphism θ ′ from G to L′ , and Theorem 4.3 provides a unique topological group isomorphism ψ ′ : ĜH,θ → L′ such that θ ′ = ψ ′ ◦ β. In particular, since the natural inclusion of L′ into L is continuous, we obtain a continuous injective homomorphism ψ : ĜH,θ → L such that θ = ψ ◦ β. Thus β is compatible with θ. It is also clear that given θ, the construction of β is determined by the commensurability class of K among closed subgroups of L, and hence by the commensurability class of H ∗ , since K = θ(H ∗ ) and the mapping · 7→ θ(·) preserves commensurability classes of subgroups. To see that β is unique up to isomorphisms of the range, Theorem 4.3 ensures that it is enough to show the following: given a t.d.l.c. completion β(G,T ) : G → ĜT that is compatible with θ, where T is a G-stable local filter and H ∗ ∈ T , then T = S. Suppose ψ2 is the injective continuous homomorphism from ĜT to L such that θ = ψ2 ◦ β(G,T ) . The collection T is the set of β(G,T ) -preimages of compact open subgroups of ĜT . The images of the compact open subgroups of ĜT give rise to a collection R′ of compact subgroups of L, so T is the set of θ-preimages of elements of R′ . We see that K ∈ R′ and that all elements of R′ are commensurate with K, so R′ ⊆ R and hence T ⊆ S. The argument that S ⊆ T is similar. e so H is commensurated in G. Let K := ψ(U ). Since For (2), U is commensurated in G, ψ is injective and continuous and U is compact, we see that K is closed in L and isomorphic HOMOMORPHISMS INTO T.D.L.C. GROUPS 19 to U as a topological group; in particular, K is profinite. The injectivity of ψ ensures that e we see that U = ψ −1 (K), so H = β −1 ψ −1 (K) = θ −1 (K). Since G has dense image in G, e ≃ ĜH,θ follows from the β(H) is dense in U and hence θ(H) is dense in K. The fact that G uniqueness result established in part (1).  To illustrate the theorem, let us spell out what it means for certain classes of action. Given a group G acting faithfully on a structure X, a t.d.l.c. completion of the action is a faithful e on the same structure such that G e contains a dense copy of G with action of a t.d.l.c. group G its original action. Say that a unitary representation X of a group H is locally finite if X is the closure of the union of an increasing family (Xi ) of finite-dimensional subrepresentations, such that the kernel of the action of H on Xi has finite index for each i. Corollary 7.2. For G a group with H a commensurated subgroup, suppose that one of the following hold: (1) G acts faithfully by permutations on a set X, and H has finite orbits on X. (2) G acts faithfully by homeomorphisms on a compact metrizable zero-dimensional space X, and the action of H on X is equicontinuous. (3) X is a faithful complex unitary representation of G, and X is locally finite as a representation of H. e y X of the action such that H has compact open Then there is a unique t.d.l.c. completion G e which is continuous in cases (1) and (2), and strongly continuous in case (3). closure in G, Moreover, all t.d.l.c. completions of the action of G on X with the given continuity property arise in this way. Proof. For (1), consider G and H as subgroups of Sym(X) with the permutation topology. As is well-known, a subgroup H of Sym(X) has compact closure if and only if it has finite orbits. Furthermore, any topological permutation group that acts continuously on X must map continuously into Sym(X), and conversely, Sym(X) itself acts continuously on X. For (2), consider G and H as subgroups of Homeo(X) with the compact-open topology. By the Arzelà–Ascoli theorem, the condition that H be equicontinuous on X is exactly the e is a group acting faithfully by condition that H has compact closure in Homeo(X). If G e to Homeo(X) is homeomorphisms on X, then the corresponding homomorphism from G e on X is continuous. continuous if and only if the action of G For (3), consider G and H as subgroups of the unitary group U(X) with the strong operator topology and let H denote the closure of H in U(H). Suppose that X is locally finite as a representation of H, with (Xi ) the corresponding increasing family of finite-dimensional subrepresentations. For each i, H acts on Xi with an open kernel of finite index, so H is totally disconnected. In addition, given a net (hj )j∈J in H, there is a subnet (hj(k) ) that is eventually constant on each Xi . It follows that (hj(k) ) converges pointwise on X to a unitary map; in other words, hj(k) converges in H. Thus H is compact and hence a profinite group. Conversely, suppose H is a subgroup of G with compact closure in U(X). By standard results (see for instance [3, Theorem 5.2]), X is an orthogonal sum of finite-dimensional irreducible representations Xj of H, and on each Xj , H acts as a compact Lie group. If H is profinite, then the Lie quotients of H are in fact finite, so H acts on Xj with a kernel of finite index. We conclude that X is a locally finite representation of H and hence of H. Therefore, H has profinite closure in U(X) if and only if X is locally finite as a representation of H. In all cases, the conclusions now follow by Theorem 7.1.  20 COLIN D. REID AND PHILLIP R. WESOLEK Remark 7.3. If X is a countably infinite set, the Cantor set, or the infinite-dimensional separable complex Hilbert space, then Sym(X), Homeo(X), and U(X) respectively are wellknown examples of Polish groups that are not locally compact. As such, simply taking the closure of the image of ψ : G → L, with L one of the aforementioned groups, will not always produce a locally compact group, and moreover, there are interesting examples of continuous actions of t.d.l.c. groups that do not arise from taking the closure in L. For example, Thompson’s group V acts faithfully by homemorphisms on the standard ternary Cantor set X ⊂ [0, 1] and has a commensurated subgroup H consisting of those elements of V that act by isometries of the visual metric. In particular, the action of H is equicontinuous. There is thus a unique t.d.l.c. completion Ve y X of the action on X such that the closure of H in Ve is a compact open subgroup. The group Ve is known as Neretin’s group N2,2 of piecewise homotheties of the ternary Cantor set, and it carries a strictly finer topology than the one induced by Homeo(X). References [1] V. V. Belyaev. Locally finite groups containing a finite inseparable subgroup. Siberian Math. J., 34(2):218– 232, 1993. [2] N. Bourbaki. General topology. Chapters 1–4. Elements of Mathematics (Berlin). Springer-Verlag, Berlin, 1989. Translated from the French, Reprint of the 1966 edition. [3] G. B. Folland. A course in abstract harmonic analysis. CRC Press, Boca Raton, FL, 2000. [4] A. S. Kechris. Classical descriptive set theory, volume 156 of Graduate Texts in Mathematics. SpringerVerlag, New York, 1995. [5] A. Le Boudec and P. Wesolek. Commensurated subgroups in tree almost automorphism groups. Groups Geom. Dyn. to appear. [6] F. Le Maı̂tre and P. Wesolek. On strongly just infinite profinite branch groups. J. Group Theory, 20(1):1– 32, 2017. [7] C. D. Reid. Dynamics of flat actions on totally disconnected, locally compact groups. New York J. Math., 22:115–190, 2016. [8] C. D. Reid and P. R. Wesolek. Dense normal subgroups and chief factors in locally compact groups. Proc. Lond. Math. Soc. to appear. [9] L. Ribes and P. Zalesskii. Profinite groups, volume 40 of Ergebnisse der Mathematik und ihrer Grenzgebiete. 3. Folge. A Series of Modern Surveys in Mathematics [Results in Mathematics and Related Areas. 3rd Series. A Series of Modern Surveys in Mathematics]. Springer-Verlag, Berlin, second edition, 2010. [10] G. Schlichting. Operationen mit periodischen Stabilisatoren. Arch. Math., 34(1):97–99, 1980. [11] Y. Shalom and G. A. Willis. Commensurated subgroups of arithmetic groups, totally disconnected groups and adelic rigidity. Geom. Funct. Anal., 23(5):1631–1683, 2013. [12] P. Wesolek. Elementary totally disconnected locally compact groups. Proc. Lond. Math. Soc. (3), 110(6):1387–1434, 2015. [13] G. Willis. The structure of totally disconnected, locally compact groups. Math. Ann., 300(2):341–363, 1994. [14] G. Willis. Tidy subgroups for commuting automorphisms of totally disconnected groups: An analogue of simultaneous triangularisation of matrices. New York J. Math., 10:1–35, 2004. University of Newcastle, School of Mathematical and Physical Sciences, University Drive, Callaghan NSW 2308, Australia E-mail address: colin@reidit.net Binghamton University, Department of Mathematical Sciences, PO Box 6000, Binghamton New York 13902 USA E-mail address: pwesolek@binghamton.edu 

Submitted to the Annals of Applied Probability ITERATIVE COLLABORATIVE FILTERING FOR SPARSE MATRIX ESTIMATION ∗ arXiv:1712.00710v1 [math.ST] 3 Dec 2017 By Christian Borgs‡ , Jennifer Chayes‡ Christina Lee‡ and Devavrat Shah§ Microsoft Research New England‡ and Massachusetts Institute of Technology§ The sparse matrix estimation problem consists of estimating the distribution of an n × n matrix Y , from a sparsely observed single instance of this matrix where the entries of Y are independent random variables. This captures a wide array of problems; special instances include matrix completion in the context of recommendation systems, graphon estimation, and community detection in (mixed membership) stochastic block models. Inspired by classical collaborative filtering for recommendation systems, we propose a novel iterative, collaborative filtering-style algorithm for matrix estimation in this generic setting. We show that the mean squared error (MSE) of our estimator converges to 0 at the rate of O(d2 (pn)−2/5 ) as long as ω(d5 n) random entries from a total of n2 entries of Y are observed (uniformly sampled), E[Y ] has rank d, and the entries of Y have bounded support. The maximum squared error across all entries converges to 0 with high probability as long as we observe a little more, Ω(d5 n ln2 (n)) entries. Our results are the best known sample complexity results in this generality. ∗ A preliminary version of this work has been accepted to appear in proceedings of the Neural Information Processing Systems Conference, 2017. The results have been improved, strengthened, and expanded with new extensions since the preliminary version. † This work is supported in parts by NSF under grants CMMI-1462158 and CMMI1634259, by DARPA under grant W911NF-16-1-0551, and additionally by a NSF Graduate Fellowship and Claude E. Shannon Research Assistantship. MSC 2010 subject classifications: Primary 62F12, 62G05; secondary 60B20, 68W40 Keywords and phrases: graphon estimation, matrix estimation, latent variable models, similarity based collaborative filtering, recommendation systems 1 2 BORGS-CHAYES-LEE-SHAH CONTENTS 1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 1.1 Problem Statement . . . . . . . . . . . . . . . . . . . 1.2 Related Work . . . . . . . . . . . . . . . . . . . . . . 1.3 Summary of Contributions . . . . . . . . . . . . . . . 2 Model, Assumptions and Variations . . . . . . . . . . . . 2.1 Notations, Assumptions . . . . . . . . . . . . . . . . 2.1.1 Mixed Membership Models . . . . . . . . . . 2.1.2 Finite Degree Polynomials . . . . . . . . . . . 2.2 Variation: Asymmetric . . . . . . . . . . . . . . . . . 2.3 Variation: Categorical Labels . . . . . . . . . . . . . 3 Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 Algorithm Intuition . . . . . . . . . . . . . . . . . . . 3.2 Algorithm Details . . . . . . . . . . . . . . . . . . . . 3.3 Reducing computation by subsampling vertices . . . 3.4 Choosing radius parameter r . . . . . . . . . . . . . 3.5 Computational Complexity . . . . . . . . . . . . . . 3.6 Belief Propagation and Non-Backtracking Operator . 3.7 Knowledge of the observation set E . . . . . . . . . . 4 Main Results . . . . . . . . . . . . . . . . . . . . . . . . . 4.1 Modified algorithm with subsampled Anchor vertices 4.2 Asymmetric Matrix . . . . . . . . . . . . . . . . . . . 4.3 Categorical Valued Data . . . . . . . . . . . . . . . . 4.4 Non-uniform sampling . . . . . . . . . . . . . . . . . 5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . 6 Proof Sketch . . . . . . . . . . . . . . . . . . . . . . . . . References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . Author’s addresses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 3 4 7 9 9 10 11 12 14 15 15 16 18 19 20 22 22 22 28 30 32 33 36 37 41 43 Appendices A Proof of Lemmas . . . . . . . . . . . . . . . . . . . . . . . . . . A.1 Sparsity of Local Neighborhood Grows Exponentially . . . A.2 Concentration of Paths within Local Neighborhood . . . . A.3 Showing that Majority of Local Neighborhoods are Good A.4 Concentration of Edges Between Local Neighborhoods . . A.5 Existence of Close Neighbors . . . . . . . . . . . . . . . . A.6 Error Bounds of Final Estimator . . . . . . . . . . . . . . B Proof of Main Results . . . . . . . . . . . . . . . . . . . . . . . B.1 Bounding the Max Error . . . . . . . . . . . . . . . . . . . B.2 Using Subsampled Anchor Vertices . . . . . . . . . . . . . . . . . . . . . . . . 44 . 44 . 49 . 58 . 64 . 82 . 94 . 109 . 109 . 110 ITERATIVE COLLABORATIVE FILTERING 3 1. Introduction. We consider the question of sparse matrix completion with noisy observations. As a prototype for such a problem, consider a noisy observation of a social network where observed interactions are signals of true underlying connections. We might want to predict the probability that two users would choose to connect if recommended by the platform, e.g. LinkedIn. As a second example, consider a recommendation system where we observe movie ratings provided by users, and we may want to predict the probability distribution over ratings for specific movie-user pairs. A popular collaborative filtering approach suggests using “similarities” between pairs of users to estimate the probability of a connection being formed or a movie being liked. Traditionally, the similarities between pair of users in a social network is computed by comparing the set of their friends or in the context of movie recommendation, by comparing commonly rated movies. In the sparse setting, however most pairs of users have no common friends, or most pairs of users have no commonly rated movies; thus there is insufficient data to compute the traditional similarity metrics. In this work, the primary interest is to provide a principled way to extend such a simple, intuitive approach to compute similarities between pair of users so to achieve sparse matrix completion. We propose to do so by incorporating information within a larger radius neighborhood rather than restricting only to immediate neighbors. In the process, we establish that it achieves best-known sample complexity which matches well known, conjectured lower bound for a special instance of the generic problem, the mixed membership stochastic block model. 1.1. Problem Statement. The question discussed above can be mathematically formulated as a matrix estimation problem. Let F be an n × n matrix which we would like to estimate, and let Z be a noisy signal of matrix F such that E[Z] = F . The available data is denoted by (E, M ), where E ⊂ [n] × [n] denotes the subset of indices for which data is observed, and M is the n × n symmetric data matrix where M (u, v) = Z(u, v) for (u, v) ∈ E, and M (u, v) = 0 for (u, v) ∈ / E. We can equivalently represent the data with an undirected weighted graph G with vertex set [n], edge set E, and edge weights given by M . We shall use graph and matrix notations in an interchangeable manner. Given the data (E, M ), we would like to estimate the original matrix F . We assume a uniform sampling model, where each entry is observed with probability p independently of all other entries. We shall assume that each u ∈ [n] is associated to a latent variable αu ∈ X1 , which is drawn independently across indices [n] as per distribution P1 over a bounded compact space X1 . We shall assume that the expected data 4 BORGS-CHAYES-LEE-SHAH matrix can be described by the latent function f , i.e. F (u, v) = f (αu , αv ), where f : X1 × X1 → R is a symmetric function. We note that such a structural assumption or the so-called Latent Variable Model is a canonical representation for exchangeable arrays as shown by Aldous and Hoover Aldous (1981); Hoover (1981); Austin (2012). For each observation, we assume that E[Z(u, v)] = F (u, v), Z(u, v) is bounded and {Z(u, v)}1≤u<v≤n are independent conditioned on the node latent variables. The goal is to find smallest p, as a function of n and structural properties of f , so that there exists an algorithm that can produce F̂ , an estimate of P matrix F , so that the Mean-Squared-Error (MSE) between F̂ and F , 1 2 u,v∈[n] (F̂ (u, v) − F (u, v)) , converges to 0 as n → ∞. n2 1.2. Related Work. The matrix estimation problem introduced above, as special cases, includes problems from different areas of literature: matrix completion popularized in the context of recommendation systems, graphon estimation arising from the asymptotic theory of graphs, and community detection using the stochastic block model or its generalization known as the mixed membership stochastic block model. The key representative results for each of these are mentioned in Table 1. We discuss the scaling of the sample complexity with respect to d (model complexity, usually rank) and n for polynomial time algorithms, including results for both mean squared error convergence, exact recovery in the noiseless setting, and convergence with high probability in the noisy setting. Now we provide a brief overview of prior works reported in the Tables 1. In the context of matrix completion, there has been much progress under the low-rank assumption and additive noise model. Most theoretically founded methods are based on spectral decompositions or minimizing a loss function with respect to spectral constraints Keshavan, Montanari and Oh (2010a,b); Candes and Recht (2009); Candès and Tao (2010); Recht (2011); Negahban and Wainwright (2011); Davenport et al. (2014); Chen and Wainwright (2015); Chatterjee (2015). A work that is closely related to ours is by Lee et al. (2016). It proves that a similarity based collaborative filteringstyle algorithm provides a consistent estimator for matrix completion under the generic model when the latent function is Lipschitz, not just low rank; however, it requires Õ(n3/2 ) samples. In a sense, our work can be viewed as an algorithmic generalization of Lee et al. (2016) that handles the sparse sampling regime and a generic noise model. Table 1 Sample Complexity of Related Literature Sample Complexity Data/Noise model Expected matrix Guarantee KMO09 KMO10 NW11 CW15 C14 LLSS16 CT09 KMO09 R11 ω(dn) Ω(dn max(log n, d)), ω(dn) ω(dn log n) Ω(n max(d, log2 n)) ω(dn log6 n) Ω(n3/2 ) Ω(dn log2 n max(d, log4 n)) Ω(dn max(d, log n)) Ω(dn log2 n) noiseless additive iid Gaussian additive iid Gaussian additive iid Gaussian independent bounded additive iid bounded noiseless noiseless noiseless rank d rank d rank d rank d rank d Lipschitz rank d rank d rank d MSE→ 0 MSE→ 0 MSE→ 0 MSE→ 0 MSE→ 0 MSE→ 0 exact recovery exact recovery exact recovery 1-bit matrix completion CW15 DPBW14 Ω(n max(d log n, log2 n, d2 )) Ω(n max(d, log n)), ω(dn) binary entries binary entries rank d rank d MSE→ 0 MSE→ 0 SBM AS15a; AS16 AS15a XML14 ω(n)∗ Ω(n log n)∗ Ω(n log n)∗ binary entries binary entries binary entries d blocks d blocks (SBM) rank d partial recovery exact recovery MSE→ 0 MMSBM AGHK13 SH17 Ω(d2 n polylog n) Ω(d2 n) binary entries binary entries rank d rank d whp error → 0 detection ACC13 ZLZ16 BCCG15 Ω(n2 ) Ω(n2 ) ω(n) binary entries binary entries binary entries monotone row sum piecewise Lipschitz monotone row sum MSE→ 0 MSE→ 0 MSE→ 0 This work This work ω(d5 n) Ω(d5 n log2 n) rank d, Lipschitz rank d, Lipschitz MSE→ 0 whp error → 0 matrix completion graphon matrix est. independent bounded independent bounded does not indicate dependence on d. *result ITERATIVE COLLABORATIVE FILTERING Paper 5 6 BORGS-CHAYES-LEE-SHAH Most of the results in matrix completion require additive noise models, which do not extend to setting when the observations are binary or quantized. The Universal Singular Value Thresholding (USVT) estimator Chatterjee (2015) is able to handle general bounded noise, although it requires a few log factors more in its sample complexity. Our work removes the extra log factors while still allowing for general bounded noise. There is also a significant amount of literature which looks at the estimation problem when the data matrix is binary, also known as 1-bit matrix completion, stochastic block model (SBM) parameter estimation, or graphon estimation. The latter two terms are found within the context of community detection and network analysis, as the binary data matrix can alternatively be interpreted as the adjacency matrix of a graph – which are symmetric, by definition. Under the SBM, each vertex is associated to one of d community types, and the probability of an edge is a function of the community types of both endpoints. Estimating the n×n parameter matrix becomes an instance of matrix estimation. In SBM, the expected matrix is at most rank d due to its block structure. Precise thresholds for cluster detection (better than random) and estimation have been established by Abbe and Sandon (2015a,b, 2016). Our work, both algorithmically and technically, draws insight from this sequence of works, extending the analysis to a broader class of generative models through the design of an iterative algorithm, and improving the technical results with precise MSE bounds. The mixed membership stochastic block model (MMSBM) allows each vertex to be associated to a length d vector, which represents its weighted membership in each of the d communities. The probability of an edge is a function of the weighted community memberships vectors of both endpoints, resulting in an expected matrix with rank at most d. Recent work by Steurer and Hopkins (2017) provides an algorithm for weak detection for MMSBM with sample complexity d2 n, when the community membership vectors are sparse and evenly weighted. They provide partial results to support a conjecture that d2 n is a computational lower bound, separated by a gap of d from the information theoretic lower bound of dn. This gap was first shown in the simpler context of the stochastic block model Decelle et al. (2011). Xu, Massoulié and Lelarge (2014) proposed a spectral clustering method for inferring the edge label distribution for a network sampled from a generalized stochastic block model. When the expected function has a finite spectrum decomposition, i.e. low rank, then they provide a consistent estimator for the sparse data regime, with Ω(n log n) samples. Graphon estimation extends SBM and MMSBM to the generic Latent Variable Model where the probability of an edge can be any measurable ITERATIVE COLLABORATIVE FILTERING 7 function f of real-valued types (or latent variables) associated to each endpoint. Graphons were first defined as the limiting object of a sequence of large dense graphs Borgs et al. (2008); Diaconis and Janson (2008); Lovász (2012), with recent work extending the theory to sparse graphs Borgs et al. (2014a,b, 2016); Veitch and Roy (2015). In the graphon estimation problem, we would like to estimate the function f given an instance of a graph generated from the graphon associated to f . Gao, Lu and Zhou (2015); Klopp, Tsybakov and Verzelen (2015) provide minimax optimal rates for graphon estimation; however a majority of the proposed estimators are not computable in polynomial time, since they require optimizing over an exponentially large space (e.g. least squares or maximum likelihood) Wolfe and Olhede (2013); Borgs et al. (2015); Borgs, Chayes and Smith (2015); Gao, Lu and Zhou (2015); Klopp, Tsybakov and Verzelen (2015). Borgs et al. (2015) provided a polynomial time method based on degree sorting in the special case when the expected degree function is monotonic. To our knowledge, existing positive results for sparse graphon estimation require either strong monotonicity assumptions Borgs et al. (2015), or rank constraints as assumed in the SBM, the 1-bit matrix completion, and in this work. We call special attention to the similarity based methods which are able to bypass the rank constraints, relying instead on smoothness properties of the latent function f (e.g. Lipschitz) Zhang, Levina and Zhu (2015); Lee et al. (2016). They hinge upon computing similarities between rows or columns by comparing commonly observed entries. Similarity based methods, also known in the literature as collaborative filtering, have been successfully employed across many large scale industry applications (Netflix, Amazon, Youtube) due to its simplicity and scalability Goldberg et al. (1992); Linden, Smith and York (2003); Koren and Bell (2011); Ning, Desrosiers and Karypis (2015); however the theoretical results have been relatively sparse. These recent results suggest that the practical success of these methods across a variety of applications may be due to its ability to capture local structure. A key limitation of this approach is that it requires a dense dataset with sufficient entries in order to compute similarity metrics, requiring that each pair of rows or columns has a growing number of overlapped observed entries, which does not hold when p = o(n−1/2 ). 1.3. Summary of Contributions. In this work, we present a novel algorithm for estimating F = [F (i, j)] by extending the notion of similarity for sparse regime in an intuitive and simple way: rather than only considering directly overlapped entries as done in Zhang, Levina and Zhu (2015); Lee et al. (2016), we consider longer “paths” of data associated to each row, 8 BORGS-CHAYES-LEE-SHAH expanding the set of associated datapoints until there is sufficient overlap. We show that this does not introduce bias and variance due to the sparse sampling. In fact, the MSE associated with the resulting estimate does converge to 0 as long as the latent function f when regarded as an integral operator has finite spectrum with rank d and p = ω(d5 n). More precisely, if f is piecewise Lipschitz with rank d and the latent variables are sampled uniformly over the unit interval, we prove that the mean squared error (MSE) of our estimates converges to zero at a rate of O(d2 (pn)−1/2+θ ) as long as the sparsity p = ω(dmax(1/2θ,4/(1−2θ)) n−1 ) for some θ ∈ (0, 41 ) (i.e. ω(d5 n) total observations for θ = 1/10). In addition, with high probability, the maximum squared error converges to zero at a rate of O(d2 (pn)−1/2+θ ) as long as the sparsity p = Ω(dmax(1/2θ,4/(1−2θ)) n−1 ln2 (n)). Our analysis applies to a generic noise setting as long as Z(i, j) has bounded support. Our work takes inspiration from Abbe and Sandon (2015a,b, 2016), which estimates clusters of the stochastic block model by computing distances from local neighborhoods around vertices. We improve upon their analysis to provide MSE bounds for the general latent variable model with finite spectrum, which includes a larger class of generative models such as mixed membership stochastic block models, while they consider the stochastic block model with non-overlapping communities. We show that our results hold even when the rank d increases with n, as long as d = o((pn)2θ,(1−2θ)/4 ). for θ ∈ (0, 41 ). As compared to spectral methods such as Keshavan, Montanari and Oh (2010b); Recht (2011); Davenport et al. (2014); Chen and Wainwright (2015); Chatterjee (2015), our analysis handles the general bounded noise model and holds for sparser regimes with respect to n for constant d, only requiring p = ω(n−1 ). In summary, as can be seen from Table 1, our result provides the best sample complexity with respect to n for the general matrix estimation problem with bounded entries noise model and constant rank d, as the other models either require extra log factors, or impose additional requirements on the noise model or the expected matrix. Similarly, ours is the best known sample complexity for high probability max-error convergence to 0 for the general rank d bounded entries setting, as other results either assume block constant or noiseless. Recently, Steurer and Hopkins (2017) showed a partial result that this computational lower bound holds for algorithms that rely on fitting low-degree polynomials to the observed data. Given the conjectured lower bound (with partial support in Steurer and Hopkins (2017)), it seems that our result is nearly optimal if not optimal in terms of its dependence on both n for MSE convergence as well as high probability (near) exact recovery. ITERATIVE COLLABORATIVE FILTERING 9 2. Model, Assumptions and Variations. 2.1. Notations, Assumptions. Recall that, of our interest is an n × n symmetric matrix F ; Z is a noisy signal of matrix F such that E[Z] = F . The available data is denoted by (E, M ), where E ⊂ [n]×[n] denotes the subset of indices for which data is observed, and M is the n×n symmetric data matrix where M (u, v) = Z(u, v) for (u, v) ∈ E, and M (u, v) = 0 for (u, v) ∈ / E. That is, our observations can be denoted as an undirected weighted graph G with vertex set [n], edge set E, and edge weights given by M . For all u, v ∈ [n], we shall assume that Z(u, v) are independent across indices with E[Z(u, v)] = F (u, v); and F (u, v), Z(u, v) ∈ [0, 1]. We assume that f has finite spectrum with rank d when regarded as an integral operator, i.e. for any αu , αv ∈ X1 , f (αu , αv ) = d X λk qk (αu )qk (αv ), k=1 where λk ∈ R for 1 ≤ k ≤ d, qk are orthonormal `2 functions for 1 ≤ k ≤ d such that Z Z 2 qk (y) dP1 (y) = 1 and qk (y)qh (y)dP1 (y) = 0 for k 6= h. X1 X1 We assume that there exists some Bq such that supy∈[0,1] |qk (y)| ≤ Bq for all k. Let us define (2.1) φ(ξ) := ess inf α0 ∈X1 P1 ! d o n X α ∈ X1 s.t. λ2k (qk (α) − qk (α0 ))2 < ξ 2 .. k=1 Let Λ denote the d × d diagonal matrix with {λk }k∈[d] as the diagonal entries, and let Q denote the d × n matrix where Q(k, u) = qk (αu ). Since Q is a random matrix depending on the sampled α, it is not guaranteed to be an orthonormal matrix (even though qk are orthonormal functions). By definition, it follows that F = QT ΛQ. Let d0 ≤ d be the number of distinct valued eigenvalues amongst λk , 1 ≤ k ≤ d. Let Λ̃ denote the d × d0 matrix where Λ̃(a, b) = λab−1 . For convenience, we shall define ρ ∈ Rd as Z (2.2) ρ = [ρk ]k∈[d] , where ρk = E[qk (α)] = qk (y)dP1 (y). X1 10 BORGS-CHAYES-LEE-SHAH We note that the finite spectrum assumption also implies that the model can be represented by latent variables in the d dimensional Euclidean space, where the latent variable for node i would be the vector (q1 (αi ), . . . qd (αi )), and the latent function would be linear, having the form X f (~q, ~q0 ) = λk qk qk0 = q T Λq 0 . k This condition also implies that the expected matrix F is low rank, which includes scenarios such as the mixed membership stochastic block model and finite degree polynomials. Although we assume observations are sampled independently with probability p, we will also discuss a solution for dealing with non-uniform sampling in Section 5. Since the finite spectrum condition imposes that the model can be described by a linear function, we will not need the additional Lipschitz condition, although we will need bounds on the underestimator φ which captures the local geometry in the d-dimensional representation. 2.1.1. Mixed Membership Models. We show that any mixed membership model can be represented with a finite spectrum latent variable model. In the mixed membership model, each node is associated to a vector π ∈ ∆d , sampled iid from a distribution P . For two nodes with respective types π and π 0 , the observed interaction is X f (π, π 0 ) = πi πj0 Bij = π T Bπ 0 , ij where B ∈ [0, 1]d×d and assumed to be symmetric. Since B is symmetric, there exists a diagonal decomposition B = U Λ̃U T with uk denoting the eigenvectors, such that the function f : ∆d × ∆d → [0, 1] corresponds to f (π, π 0 ) = d X λ̃k uTk πuTk π 0 . k=1 We can verify that Z ∆d Z f (π, π 0 )dP dP < ∞. ∆d Let W : L2 (∆d , [0, 1]) → L2 (∆d , [0, 1]) denote the Hilbert-Schmidt integral operator associated to the kernel f , such that for some function η ∈ ITERATIVE COLLABORATIVE FILTERING 11 L2 (∆d , [0, 1]), Z 0 (W η)(π ) = f (π, π 0 )η(π)dP ∆d Z = = = d X λ̃k uTk πuTk π 0 η(π)dP ∆d k=1 Z d X T 0 uTk πη(π)dP λ̃k uk π ∆d k=1 d X λ̃k gk (π 0 )hgk ηi, k=1 where the inner product between two functions η, η 0 is defined as Z 0 hη, η i = η(π)η 0 (π)dP, ∆d and gk (·) is a set of functions such that gk (π) = uTk π. Therefore, the complement of the kernel of W must be contained within the span of gk , such that W must have finite spectrum with rank at most d. 2.1.2. Finite Degree Polynomials. We demonstrate that finite degree polynomials lead to latent variable models with finite spectrum. Let f (x, y) be any finite degree symmetric polynomial, represented in the form f (x, y) = d X d X cij xi y j , i=0 j=0 were symmetry implies that for all ij, cij = cji . Let x = (1, x, x2 , . . . xd ) and y = (1, y, y 2 , . . . y d ), and let C denote the (d+1)×(d+1) matrix with entries [cij ]. It follows that f (x, y) = xT Cy. Then we can use the same argument as above to show that the Hilbert-Schmidt integral operator associated to the kernel f has finite spectrum. Since C is symmetric, there exists a diagonal decomposition B = U Λ̃U T with uk denoting the eigenvectors, such that the function f corresponds to f (x, y) = d X λ̃k uTk xuTk y. k=1 Let W : L2 (∆d , [0, 1]) → L2 (∆d , [0, 1]) denote the Hilbert-Schmidt integral operator associated to the kernel f , such that for some function η ∈ 12 BORGS-CHAYES-LEE-SHAH L2 (∆d , [0, 1]), Z f (x, y)η(x)dP (W η)(y) = X1 d X Z = = = X1 k=1 d X λ̃k uTk xuTk yη(x)dP λ̃k uTk y k=1 d X Z X1 uTk xη(x)dP λ̃k gk (y)hgk ηi, k=1 where gk (·) is a set of functions such that gk (y) = uTk y. Therefore, the complement of the kernel of W must be contained within the span of gk , such that W must have finite spectrum with rank at most d. 2.2. Variation: Asymmetric. If the model that we would like to learn is in fact asymmetric, we can actually transform it to an equivalent symmetric model. Consider an n × m matrix F which we would like to learn, where E is the set of observed indices generated via uniform sampling with density p, and we assume the independent bounded noise model where Z(u, v) ∈ [0, 1], and E[Z(u, v)] = Fuv . Row u ∈ [n] is associated with latent variable αu ∈ X1 drawn independently and as per distribution P1 , and column v ∈ [m] is associated with latent variable βv ∈ X2 drawn independently and as per distribution P2 . Let F (u, v) = f (αu , βv ) ∈ [0, 1], where f has finite spectrum according to d X λk q1k (αu )q2k (βv ), f (αu , βv ) = k=1 where q1k and q2k are orthonormal `2 functions with respect to the measures P1 and P2 respectively. We show how to translate this to an approximately equivalent symmetric model satisfying the assumptions in Section 2.1. We augment the latent space to be X10 = (X1 , 0) ∪ (0, X2 ) ⊂ X1 × X2 , where    n n+m  P1 if y ∈ (X1 , 0) P 0 (y) =  .  m P2 if y ∈ (0, X2 ) n+m 13 ITERATIVE COLLABORATIVE FILTERING We define the symmetric function f 0 to be   0 if θu , θv ∈ (X1 , 0)    0 if θu , θv ∈ (0, X2 ) f 0 (θu , θv ) = .  f (αu , βv ) if θu = (αu , 0), θv = (0, βv )    f (α , β ) if θ = (0, β ), θ = (α , 0) v u u u v v Then we can verify that f 0 has finite spectrum   √ √ d  d  X X λk nm 0 λk nm 0 qk (θu )qk (θv ) − qk (θu )qk0 (θv ), f (θu , θv ) = n+m n+m k=1 k=1 with orthogonal eigenfunctions (
n+m 1/2 qk (θu ) = and ( qk0 (θu ) = q1k (αu ) 2n
n+m 1/2 q2k (βu ) 2m n+m 1/2 q1k (αu ) 2n
n+m 1/2 − 2m q2k (βu )
if θu = (αu , 0) if θu = (0, βu ) if θu = (αu , 0) if θu = (0, βu ) . We can verify that this in fact equals f 0 as defined above, and that the qk , qk0 functions are orthonormal:  Z  Z Z 1 1 q1k (α)2 dP1 (α) + q2k (β)2 dP2 (β) qk (y)2 dP1 (y) = 0 2 2 X1 X2 X1     1 1 = + = 1. 2 2 The same holds for qk0 . For k 6= h, Z qk (y)qh (y)dP1 (y) X10  Z  Z 1 1 = q1k (α)q1h (α)dP1 (α) + q2k (β)q2h (β)dP2 (β) 2 2 X1 X2 = 0. And for qk0 and qk , Z qk (y)qk0 (y)dP1 (y) X10  Z  Z 1 1 q1k (α)q1k (α)dP1 (α) + q2k (β)(−q2k (β))dP2 (β) = 2 2 X1 X2     1 1 = − = 0. 2 2 14 BORGS-CHAYES-LEE-SHAH Therefore, the (n + m) × (n + m) matrix F 0 is effectively a permuted version of   0 F . FT 0 The only difference in these models is that in the original asymmetric model we fix n vertices with latent type in X1 , and m vertices with latent type in X2 , whereas in the symmetric model we sample n + m vertices total, where n each vertex is type X1 with probability n+m , and type X2 with probability m . These two models are asymptotically equivalent for large n, m, as long n+m as the ratio remains fixed. 2.3. Variation: Categorical Labels. If the edge labels are categorical instead of real-valued, then it doesn’t make sense to compute the expected value averaged over different edge label values, but rather the goal would be instead to estimate the distribution over the different categories or labels. This is particularly suitable for a setting in which there is no obvious metric between the categories such that an aggregate statistic would not be meaningful. Assume that there is a finite m category types, and the distribution over categories is a function of the latent variables. Futhermore, we assume that the model within each category satisfies the assumptions mentioned above in Section 2.1. Assume a uniform sampling model with density p. Each u ∈ [n] is associated to a latent variable αu ∈ X1 , drawn independently as per distribution P1 over a compacted bounded space X1 . F (u, v) = f (αu , αv ) ∈ ∆m , where f is a symmetric function, and ∆m denotes the m dimensional probability simplex. For each observed entry (u, v) ∈ E, the observed datapoint Z(u, v) ∈ [m] is drawn from the distribution f (αu , αv ), such that P(Z(u, v) = i) = fi (αu , αv ). Assume that each of the functions {fi }i∈[m] has finite spectrum with rank di when regarded as an integral operator, fi (αu , αv ) = di X λik qik (αu )qik (αv ), k=1 where qik are orthonormal `2 functions such that Z Z 2 qik (y) dP1 (y) = 1 and qik (y)qih (y)dP1 (y) = 0 for k 6= h. X1 X1 Our basic estimator can be modified to estimate a categorical distribution, where the error is measured in terms of total variation distance. Let M i be an i = I(M (u, v) = i), such that E[M i |(u, v) ∈ n × n binary matrix where Muv uv ITERATIVE COLLABORATIVE FILTERING 15 E] = P(Z(u, v) = i) = fi (αu , αv ). We can apply the algorithm to the data (E, M i ) to estimate fi for each i. Since the estimates should sum to 1 across all categories, we apply the algorithm to the data matrices associated to the first m − 1 categories, i.e. M i for i ∈ [m − 1], and then we let the estimate for the m-th category be equal to 1 minus the sum of the m − 1 earlier estimates. 3. Algorithm. The algorithm that we propose uses the concept of local approximation, first determining which datapoints are similar in value, and then computing neighborhood averages for the final estimate. All similaritybased collaborative filtering methods have the following basic format: 1. Compute distances between pairs of vertices, e.g., R dist(u, a) ≈ X1 (f (αu , t) − f (αa , t))2 dP1 (t). (3.1) 2. Form estimate by averaging over “nearby” datapoints, 1 P F̂ (u, v) = |Euv (3.2) (a,b)∈Euv M (a, b), | where Euv := {(a, b) ∈ E s.t. dist(u, a) < ξ(n), dist(v, b) < ξ(n)}. We will choose the threshold ξ(n) depending on dist in order to guarantee that it is small enough to drive the bias to zero, ensuring the included datapoints are close in value, yet large enough to reduce the variance, ensuring |Euv | diverges. 3.1. Algorithm Intuition. Various similarity-based algorithms differ in the distance computation (Step 1). For dense datasets, i.e. p = ω(n−1/2 ), previous works have proposed and analyzed algorithms which approximate the L2 distance of (3.1) by using variants of the finite sample approximation, P dist(u, a) = |O1ua | y∈Oua (F (u, y) − F (a, y))2 , (3.3) where y ∈ Oua iff (u, y) ∈ E and (a, y) ∈ E Airoldi, Costa and Chan (2013); Zhang, Levina and Zhu (2015); Lee et al. (2016). For sparse datasets, with high probability, Oua = ∅ for almost all pairs (u, a), such that this distance cannot be computed. In this paper we are interested in the sparse setting when p is significantly smaller than n−1/2 , down to the lowest threshold of p = ω(n−1 ). If we visualize the data via a graph with edge set E, then (3.3) corresponds to comparing common neighbors of vertices u and a. A natural extension when u and a have no common neighbors, is to instead compare the r-hop 16 BORGS-CHAYES-LEE-SHAH neighbors of u and a, i.e. vertices y which are at distance exactly r from both u and a. We compare the product of weights along edges in the path from u to y and a to y respectively, which in expectation approximates R Qr−2 Q i∈[r−1] P1 (ti ) X1r−1 f (αu , t1 )( s=1 f (ts , ts+1 ))f (tr−1 , αy )d P r = k λk qk (αu )qk (αy ) (3.4) = eTu QT Λr Qey . We choose a large enough r such that there are sufficiently many “common” vertices y which have paths to both u and a, guaranteeing that our distance can be computed from a sparse dataset. 3.2. Algorithm Details. We present and discuss details of each step of the algorithm, which primarily involves computing pairwise distances (or similarities) between vertices. The parameters of the algorithm are c1 , c2 , c3 , r, ξ1 (n) and ξ2 (n). Step 1: Sample Splitting. We partition the datapoints into disjoint sets, which are used in different steps of the computation to minimize correlation across steps for the analysis. Each edge in E is independently placed into E1 , E2 , or E3 , with probabilities c1 , c2 , and 1 − c1 − c2 respectively. Matrices M1 , M2 , and M3 contain information from the subset of the data in M associated to E1 , E2 , and E3 respectively. M1 is used to define local neighborhoods of each vertex, M2 is used to compute similarities of these neighborhoods, and M3 is used to average over datapoints for the final estimate in (3.2). Step 2: Expanding the Neighborhood. We first expand local neighborhoods of radius r around each vertex. Let Su,s denote the set of vertices which are at distance s from vertex u in the graph defined by edge set E1 . Specifically, i ∈ Su,s if the shortest path in G1 = ([n], E1 ) from u to i has a length of s. Let Bu,s denote the set of vertices which are at distance at most s from vertex u in the graph defined by E1 , i.e. Bu,s = ∪st=1 Su,t . Let Tu denote a breadth-first tree in G1 rooted at vertex u. The breadth-first property ensures that the length of the path from u to i within Tu is equal to the length of the shortest path from u to i in G1 . If there is more than one valid breadth-first tree rooted at u, choose one uniformly at random. Let Nu,r ∈ [0, 1]n denote the following vector with support on the boundary of the r-radius neighborhood of vertex u (we also call Nu,r the neighborhood ITERATIVE COLLABORATIVE FILTERING 17 boundary): (Q Nu,r (i) = (a,b)∈pathTu (u,i) M1 (a, b) if i ∈ Su,r , if i ∈ / Su,r , 0 where pathTu (u, i) denotes the set of edges along the path from u to i in the tree Tu . The sparsity of Nu,r (i) is equal to |Su,r |, and the value of the coordinate Nu,r (i) is equal to the product of weights along the path from u to i. Let Ñu,r denote the normalized neighborhood boundary such that ln(1/p) Ñu,r = Nu,r /|Su,r |. We will choose radius r = 86ln(c . 1 pn) Step 3: Computing the distances. We present two variants for estimating the distance. 1. For each pair (u, v), compute dist1 (u, v) according to 1−c1 p
c2 p Ñu,r − Ñv,r
T
M2 Ñu,r+1 − Ñv,r+1 . 2. For each pair (u, v), compute distance according to P dist2 (u, v) = i∈[d0 ] zi ∆uv (r, i), where ∆uv (r, i) is defined as 1−c1 p
c2 p Ñu,r − Ñv,r
T
M2 Ñu,r+i − Ñv,r+i , 0 and z ∈ Rd is a vector that satisfies Λ2r+2 Λ̃z = Λ2 1. z always exists and is unique because Λ̃ is a Vandermonde matrix (recall definitions from Section 2.1), and Λ−2r 1 lies within the span of its columns. Computing dist1 does not require knowledge of the spectrum of f . In our analysis we prove that the expected squared error of the estimate computed in (3.2) using dist1 converges to zero with n for p = ω(n−1+ ) for some  > 0, i.e. p must be polynomially larger than n−1 . Although computing dist2 requires full knowledge of the eigenvalues (λ1 . . . λd ) to compute the vector z, the expected squared error of the estimate computed in (3.2) using dist2 conveges to zero for p = ω(n−1 ), which includes the sparser settings when p is only larger than n−1 by polylogarithmic factors. It seems plausible that the technique employed by Abbe and Sandon (2015b) could be used to design a modified algorithm which does not need to have prior knowledge of the spectrium. They achieve this for the stochastic block model case by bootstrapping the algorithm with a method which estimates the spectrum first and then computes pairwise distances with the estimated eigenvalues. 18 BORGS-CHAYES-LEE-SHAH Step 4: Averaging datapoints to produce final estimate. The estimate F̂ (u, v) is computed by averaging over nearby points defined by the distance estimates dist1 (or dist2 ). Recall that Bq ≥ 1 was assumed in the model definition to upper bound supy∈[0,1] |qk (y)|. Let Euv1 denote the set of undirected edges (a, b) such that (a, b) ∈ E3 and 1 both dist1 (u, a) and dist1 (v, b) are less than ξ1 (n) = 33Bq d|λ1 |2r+1 (c1 pn)− 2 +θ . Here θ ∈ (0, 41 ) is a parameter whose choice may affect the performance of the algorithm. The final estimate F̂ (u, v) produced by using dist1 is computed by averaging over the undirected edge set Euv1 , X 1 (3.5) F̂ (u, v) = M3 (a, b). |Euv1 | (a,b)∈Euv1 Let Euv2 denote the set of undirected edges (a, b) such that (a, b) ∈ E3 , and 1 both dist2 (u, a) and dist2 (v, b) are less than ξ2 (n) = 33Bq d|λ1 |(c1 pn)− 2 +θ . Again, θ ∈ (0, 41 ) is a parameter whose choice may affect the performance of the algorithm. The final estimate F̂ (u, v) produced by using dist2 is computed by averaging over the undirected edge set Euv2 , X 1 F̂ (u, v) = (3.6) M3 (a, b). |Euv2 | (a,b)∈Euv2 3.3. Reducing computation by subsampling vertices. The most expensive part of the algorithm as stated above, is that we would need to compute n2 pairwise distances. If the network was generated from the stochastic block model where the function f is piecewise constant, there would be only k types of vertices to distinguish amongst in order to determine the local distances. Therefore, if we had a representative vertex from each of the k communities, it would be sufficient to compare a vertex with the k representative vertices, instead of all n other vertices (which all repeat these k types). This idea was used in Abbe and Sandon (2016) to obtain a nearly-linear time algorithm for clustering. In our setting however, we do not have k distinct communities, but rather the function f may be smooth. We can extend this idea by subsampling sufficiently many “anchor” vertices K ⊂ [n] that cover the space well, i.e. for any vertex u ∈ [n], there exists some anchor vertex i ∈ K which is “close” to u in the sense that kΛQ(eu − ei )k22 is small. Then for all n vertices, we compute the distances with each of the anchor vertices, and we let π : [n] → K be a mapping from each vertex to the anchor vertex that minimizes the estimated distance (as computed in Steps 2 and 3 above), π(u) = arg min dist(u, i). i∈K 19 ITERATIVE COLLABORATIVE FILTERING The final estimate then is given by F̂ (u, v) = F̂ (π(u), π(v)) = 1 |Eπ(u)π(v) | X M3 (a, b), (a,b)∈Eπ(u)π(v) where Eπ(u)π(v) denotes the set of undirected edges (a, b) such that (a, b) ∈ E3 and both dist(π(u), a) and dist(π(v), b) are less than some threshold ξ(n) (as defined in Step 4 above). We can compute Eπ(u)π(v) because for each anchor vertex, we have estimates of the distances to all other vertices. The pair-wise distance calculations required by the above algorithm scales as O(n|K|). 3.4. Choosing radius parameter r. The parameter r used to grow the local neighborhoods in Step 1 of the algorithm must be chosen very precisely. When r is either too small or too large, the size of the nodes at the neighorhood boundaries will be too small such that there is not sufficient overlap. The vectors Ñu,r and Ñv,r+1 will be too sparse, and the measureT M Ñ ments Ñu,r 2 v,r+1 will be noisy. Therefore, our analysis recommends that r is chosen to satisfy the following conditions   ln(1/c1 p) ln(1/c1 p) r + d0 ≤ 8 7ln(9c (3.7) = Θ ln(c1 pn) , 1 pn/8)   ln(1/p) (3.8) r + 12 ≥ 8 ln(7|λ6dln(1/p) = Θ 2 ln(c1 pn) . | c1 pn/8|λ1 |) In order to guarantee that there exists an integer value of r which satisfies (3.7) and (3.8), we need to impose additional restrictions on c1 , p, and |λd | (we assume |λ1 | is constant with respect to n). The assumption that |λd | = ω((c1 pn)−1/4 ) guarantees that the left hand side of the second inequality in (3.7) grows asymptotically with n. We need to ensure that the difference between the upper and lower bounds on r is at least 1, which is guaranteed by (when using dist2 ) 7 ln(1/c1 p) 6 ln(1/p) 1 − ≥ d0 − + 1. 2 8 ln(9c1 pn/8) 8 ln(7λd c1 pn/8|λ1 |) 2 Because |λ1 | is constant with respect to n, asymptotically this inequality reduces to 7 ln(1/c1 p) 6 ln(1/p) 1 − ≥ d0 + . 8 ln(c1 pn) 2 8 ln(λ2d c1 pn) If p = o(n−1+ ) for all  > 0, then for a constant c1 = Θ(1), the inequality is satisfied asymptotically if we guarantee that |λd | = ω((c1 pn)−1/15 ) and d = 20 BORGS-CHAYES-LEE-SHAH o(ln(n)/ ln(pn)). If p = n−1+ for some  > 0, then the inequality is satisfied asymptotically if we assume that |λd | = ω(n−γ ) ∀γ > 0, and c1 is chosen 0 0 such that c1 pn = ω(1) and c1 pn = o(n(6+1)/(8d +11) ) = o((p6 n7 )1/(8d +11) ). When the algorithm is using dist1 , the bound would be simplified with plugging in 1 instead of d0 . 3.5. Computational Complexity. Let’s discuss the precise computational complexity of the algorithm. Since the algorithm involves growing local neighborhoods and computing similarities between these neighborhoods, the cost of computation depends P 0 on the size of these neighborhoods, which is denoted by |Bu,r+d0 | = r+d s=1 |Su,s |. In Lemmas A.1 and A.2, we proved that with high probability the neighborhood sizes grow exponentially, |Su,s | < ln(1/p) ( 9c18pn )s , assuming that ( 9c18pn )s < (c1 p)−7/8 . The choice of r = 86ln(c 1 pn) guarantees that this assumption holds for all s ≤ r + d0 for a sufficiently large size matrix as long as d0 is not too large such as any constant. It follows that |Bu,r+d0 | is dominated by |Su,r+d0 | and is bounded above by (c1 p)−7/8 with high probability. First
we discuss the original algorithm which compares distances between all n2 pairs of vertices. Recall that we split the edges into sets E1 , E2 , E3 which takes O(|E|) = O(pn2 ) random coin tosses, which we assume to be O(1) amount of work. Then we build the r + d0 -radius neighborhood vectors Nu,r+d0 , which takes O(|Bu,r+d0 |) = O(p−7/8 ) time for a total of O(np−7/8 ) for all n vertices. The bound of p−7/8 comes from the analysis and is related to how the radius r is chosen to satisfy conditions in Lemmas A.1 and A.2. These lemmas guarantee that the neighborhood grows exponentially with the expansion factor of Θ(c1 pn). When r is too large, we “run out of vertices” such that the neighborhood growth rate will slow down. Next we compute the distances which involve at most d0 (≤ d) com
T
putations of the form Ñu,r − Ñv,r M2 Ñu,r+i − Ñv,r+i . The complexity of computing the inner product of two vectors is bounded above by the
sparsity of one of the vectors. By definition, the sparsity of Ñu,r − Ñv,r is bounded by (|Bu,r | + |Bv,r |) = O(p−7/8 ), and in expectation the sparsity of any row or column of M2 is c2 pn, such that the average sparsity of
T Ñu,r − Ñv,r M2 is bouned by O(p−7/8 c2 pn). Therefore the computation of dist1 (u, v) given the neighborhood
vectors is bounded by O(p−7/8 c2 pn), with a total of O(p−7/8 c2 pn3 ) for all n2 pairwise distances. Since dist2 (u, v) requires computing the above expression for i ∈ [d0 ], this would be further multiplied by d0 . In addition, computing dist2 requires the vector z which involves computing the pseudoinverse of a d0 × d matrix. This takes at most O(d3 ) time. Finally we have to compute the estimate for each value by ITERATIVE COLLABORATIVE FILTERING 21 averaging over data-points within the neighborhoods. The number of datapoints included in the weighted average is |Euv2 | = O(pn2 ), which is as large as O(pn4 ) computation if we separately average for each of the n(n − 1)/2 distinct locations to estimate. Thus, the total computational complexity can be bounded above as (3.9) O(pn2 + np−7/8 + d0 p−7/8 pn3 + d3 + pn4 ) = O(d0 p−7/8 pn3 + pn4 ), where d3 = O(pn4 ) for pn = ω(1), since d < n. Next we discuss the computational complexity of the modified algorithm which selects anchor vertices and computes distances only to the anchor vertices. First we select the |K| vertices at random, which takes O(|K| log(n)) time. Then we split the edges into sets E1 , E2 , E3 which takes O(|E|) = O(pn2 ) time. We build the r-radius neighborhood vectors Nur , which takes O(|Bu,r |) = O(( 9c18pn )r ) = O(p−7/8 ) time for a total of O(np−7/8 ) for all n vertices. Next we compute the distances which involve at most d0 compu
T
tations of the form Ñu,r − Ñv,r M2 Ñu,r+i − Ñv,r+i . The complexity of computing the inner product of two vectors is bounded above by the spar
T sity of one of the vectors. Therefore, since the sparsity of Ñu,r − Ñv,r M2 is bounded by (|Bu,r | + |Bv,r |)c2 pn, the computation is thus bounded by 1 1 O(p 8 c2 n) for each distance computation. This results in a total of O(p 8 c2 n2 |K|) for all n|K| pairwise distance between any vertex and an anchor. If we are computing dist2 , we have an additional O(d3 ) from calculating the pseudoinverse of Λ̃. As we compute the distances, we can keep track for each vertex which anchor is the closest, such that we know π(u) for every u ∈ [n] without additional computational cost. Finally we have to compute the estimate for each value by averaging over datapoints within the neighborhoods. Since F̂ (u, v) = F̂ (π(u), π(v)), we only need to compute estimates for |K|(|K|−1)/2 locations, and the rest of the estimates follow directly from the mapping π. The number of datapoints included in the weighted average is |Euv2 | = O(pn2 ), which results in a most O(pn2 |K|2 ) computation for all locations to estimate. This results in a total computational complexity of O(|K| log(n) + pn2 + np−7/8 + d0 p−7/8 pn2 |K| + pn2 |K|2 ) (3.10) = O(d0 p−7/8 pn2 |K| + pn2 |K|2 ). That is, subsampling the vertices can significantly reduce the computational complexity in (3.9) if |K|  n. We will show in our analysis, that it is 1 3θ sufficient to choose |K| to be on the order of Θ(d−3/2 (c1 pn) 4 + 2 ). If the data is very sparse, then |K| may be only logarithmic in n and still achieve the same error guarantees as if we computed all pairwise distances. 22 BORGS-CHAYES-LEE-SHAH 3.6. Belief Propagation and Non-Backtracking Operator. The idea of comparing vertices by looking at larger radius neighborhoods was introduced in Abbe and Sandon (2015a), and has connections to belief propagation Decelle et al. (2011); Abbe and Sandon (2016) and the non-backtracking operator Krzakala et al. (2013); Karrer, Newman and Zdeborová (2014); Mossel, Neeman and Sly (2017); Massoulié (2014); Bordenave, Lelarge and Massoulié (2015). The non-backtracking operator was introduced to overcome the issue of sparsity. For sparse graphs, vertices with high-degree dominate the spectrum, such that the informative components of the spectrum get hidden behind the high degree vertices. The non-backtracking operator avoids paths that immediately return to the previously visited vertex in a similar manner as belief propagation, and its spectrum has been shown to be more wellbehaved, perhaps adjusting for the high degree vertices, which get visited very often by paths in the graph. In our algorithm, the neighborhood paths are defined by first selecting a rooted tree at each vertex, thus enforcing that each vertex along a path in the tree is unique. This is important in our analysis, as it guarantees that the distribution of vertices at the boundary of each subsequent depth of the neighborhood is unbiased, since the sampled vertices are freshly visited. 3.7. Knowledge of the observation set E. In our algorithm, we assumed that we observed the edge set E. Specifically, this means that we are able to distinguish between entries of the matrix that have value zero because they are not observed, i.e. (i, j) ∈ / E, or if the entry was observed to be value zero, i.e. (i, j) ∈ E and M (i, j) = Z(i, j) = 0. This fits well for applications such as recommendations, where the system does know the information of which entries are observed or not. Some social network applications contain this information (e.g. facebook would know if they have recommended a link which was then ignored) but other network information may lack this information, e.g. we do not know if link does not exist because observations are sparse, or because observations are dense but the probability of an edge is small. 4. Main Results. We prove bounds on the estimation error of our algorithm in terms of the mean squared error (MSE), defined as h i P 1 2 . MSE ≡ E n(n−1) ( F̂ (u, v) − F (u, v)) u6=v It follows from the model that for any αu , αv ∈ X , R Pd 2 2 2 k=1 λk (qk (αu ) − qk (αv )) X1 (f (αu , y) − f (αv , y)) dP1 (y) = = kΛQ(eu − ev )k22 . 23 ITERATIVE COLLABORATIVE FILTERING The key part of the analysis is to show that the computed distances are in fact good estimates of kΛQ(eu − ev )k22 . The analysis essentially relies on showing that the neighborhood growth around a vertex behaves according to its expectation, according to some properly defined notion. The radius r must be small enough to guarantee that the growth of the size of the neighborhood boundary is exponential, increasing at a factor of approximately c1 pn. However, if the radius is too small, then the boundaries of the respective neighborhoods of the two chosen vertices would have a small intersection, so that estimating the similarities based on the small intersection of datapoints would result in high variance. Therefore, the choice of r is critical to the algorithm and analysis. We are able to prove bounds on the squared error when r is chosen to satisfy the following conditions:   ln(1/c1 p) ln(1/c1 p) r + d0 ≤ 8 7ln(9c = Θ ln(c1 pn) , 1 pn/8)   ln(1/p) r + 12 ≥ 8 ln(7|λ6 ln(1/p) = Θ ln(c1 pn) . |2 c1 pn/8|λ1 |) d Recall that d0 (≤ d) denotes the number of distinct valued eigenvalues in the spectrum of f , (λ1 . . . λd ), and determines the number of different radius “measurements” involved in computing dist2 (u, v). Computing dist1 (u, v) only involves a single measurement, thus the left hand side of (3.7) can be reduced to r + 1 instead of r + d0 . When p is above a threshold, we choose c1 to decrease with n to ensure (3.7) can be satisfied, sparsifying the edge set E1 used for expanding the neighborhood around a vertex . When the sample probability is polynomially larger than n−1 , i.e. p = n−1+ for some  > 0, these constraints imply that r is a constant with respect to n. However, if p = Õ(n−1 ), we will need r to grow with n according to a rate of 6 ln(1/p)/8 ln(c1 pn). Theorems 4.1 and 4.2 provide bound on the error of any given entry with high probability. These high probability error bounds naturally result in a bound on the mean squared error as well. We state result for dist1 first. Theorem 4.1. Let the following hold for some θ ∈ (0, 14 ),  > 0: 1. Conditions on sampling probability, p. (4.1) p = o(n−1+1/(5+8θ) ), p = ω(n−1+ ), and c1 pn = ω(1). 2. Conditions on neighborhood radius, r. We have r = Θ(1), i.e. not scaling with n, such that (4.2)  9c1 pn 8 r+1 −7/8 ≤ (c1 p) and   1 7λ2d c1 pn r+ 2 8|λ1 | ≥ p−6/8 24 BORGS-CHAYES-LEE-SHAH 3. Condition on spectrum of Λ. The smallest eigenvalue λd is such that   1 1 |λd | = ω (c1 pn)− min( 4 , 2 +θ) (4.3) 4. Condition on distribution of latent features. Define ξ1LB , ξ1U B as 2 ξ1LB = (4.4) Bq d|λ1 | 1 (c1 pn) 2 −θ 2 and ξ1U B =  |λ1 | |λd | 2r 65Bq d|λ1 | 1 (c1 pn) 2 −θ . They satisfy (4.5)     2θ −1 1 pn) φ(ξ1LB ) = ω max p, n−3/4 , ξ1LB . exp − (c8B 2d q For any u, v ∈ [n], with probability greater than       2 2 (c1 pn)2θ (n − 1)φ(ξ2LB ) c3 pn2 ξ1U B φ(ξ1LB ) 1 − O d exp − + exp − + exp − , 8Bq2 d 24 8 the error of the estimate produced by the algorithm when using dist1 and parameter r is bounded by 1/2 !  r  3 2 √ Bq d |λ1 | |λ1 | |F̂ (u, v) − f (αu , αv )| = O(Bq dξ1U B ) = O 1 −θ |λd | (c1 pn) 2 And subsequently 2 MSE = O(Bq2 dξ1U B )+        2 2 2θ c pn2 ξ1U (n−1)φ(ξ1LB ) 1 pn) B φ(ξ1LB ) O d exp − (c8B + exp − 3 + exp − . 2d 24 8 q The following result provides similar guarantee for dist2 . Theorem 4.2. Let the following hold for some θ ∈ (0, 41 ): 1. Conditions on sampling probability, p. p = o(n−1+1/(5+8θ) ) (4.6) and c1 pn = ω(1). 2. Conditions on neighborhood radius, r. (4.7)  9c1 pn 8 r+d0 ≤ (c1 p)−7/8 and   1 7λ2d c1 pn r+ 2 8|λ1 | ≥ p−6/8 25 ITERATIVE COLLABORATIVE FILTERING 3. Condition on spectrum of Λ. The smallest eigenvalue λd is such that   1 1 (4.8) |λd | = ω (c1 pn)− min( 4 , 2 +θ) . The number of distinct magnitude eigenvalues d0 satisfies (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ) r≥ , ln(2)   θ 1 ≥ (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ). ln 2(1 + 2θ) p (4.9) (4.10) 4. Condition on distribution of latent features. Define ξ1LB , ξ1U B as 2 ξ2LB = (4.11) Bq d|λ1 | 1 (c1 pn) 2 −θ 2 and ξ2U B = 65Bq d|λ1 | 1 (c1 pn) 2 −θ . They satisfy (4.12)     2θ −1 1 pn) exp − (c8B . φ(ξ2LB ) = ω max p, n−3/4 , ξ2LB 2d q For any u, v ∈ [n], with probability greater than     2    (c1 pn)2θ ξ2U B c3 pn2 φ(ξ2LB )2 (n − 1)φ(ξ2LB ) 1 − O d exp − + exp − (1 − o(1)) + exp − , 8Bq2 d 24 8 the error of the estimate produced by the algorithm when using dist2 and parameter r is bounded by  1/2 ! √ Bq3 d2 |λ1 | |F̂ (u, v) − f (αu , αv )| = O(Bq dξ2U B ) = O . 1 −θ (c1 pn) 2 And subsequently,   Bq3 d2 |λ1 | MSE = O + 1 (c1 pn) 2 −θ        2 2 (c1 pn)2θ c3 pn2 ξ2U (n−1)φ(ξ2LB ) B φ(ξ2LB ) + exp − . O d exp − + exp − 24 8 8Bq2 d In order to guarantee that the mean squared error converges to zero, we need to enforce that d = o((c1 pn)min(2θ,(1−2θ)/4) ). Thus, by choosing θ = 1/10, then this condition would imply p = ω(d5 n) for constant c1 . Furthermore, we can in fact obtain bounds on the maximum error over all entries with high probability as presented in Theorems 4.3 and 4.4. The results follow from using union bound to control the error among all entries. We first state result for dist1 . 26 BORGS-CHAYES-LEE-SHAH Theorem 4.3. Let (4.1)-(4.5) hold for some θ ∈ (0, 41 ),  > 0. Then, with probability at least        2 2θ ξ c pn2 φ(ξ )2 1 pn) 1LB ) + nd exp − (c4B , 1 − O n2 exp − 1U B 3 24 1LB (1 − o(1)) + n exp − (n−1)φ(ξ 2d 8 q the maximum error over all entries of the estimate produced by the algorithm when using dist1 and parameter r is bounded by √ kF̂ − F kmax ≡ max |F̂ (i, j) − F (i, j)| = O(Bq dξ1U B ) i,j 1/2 ! r  3 2  Bq d |λ1 | |λ1 | . =O 1 −θ |λd | (c1 pn) 2 The following result provides similar guarantee for dist2 . Theorem 4.4. Let (4.6)-(4.12) hold for some θ ∈ (0, 14 ). Then, with probability at least   2      2θ ξ c pn2 φ(ξ )2 1 pn) 2LB ) 1 − O n2 exp − 2U B 3 24 2LB (1 − o(1)) + n exp − (n−1)φ(ξ + nd exp − (c4B , 2d 8 q the maximum error over all entries of the estimate produced by the algorithm when using dist2 and parameter r is bounded by √ kF̂ − F kmax = O(Bq dξ2U B )  1/2 ! Bq3 d2 |λ1 | . =O 1 −θ (c1 pn) 2 The probability of error stated in Theorems 4.3 and 4.4 have n in the coef2θ ficient such that we additionally   need to enforce that (c1 pn) = Ω(d ln(dn)) 2θ 1 pn) to guarantee nd exp − (c4B converges to zero. 2d q Local geometry. The local geometry of the latent probability measure with respect to the latent function affects the error results through the function φ. As an example, consider the case where the latent variables are sampled from the uniform distribution on [0, 1] and the function f is piecewise L-Lipschitz with minimum piece size of ` ≤ 1. We can show that φ(ξ) ≥ min(`, ξ/2L), which we can then plug into the results in the above theorems. The above bounds can be reduced as the terms involving φ will be dominated by others. We can show that the mean squared error for the algorithm using dist1 is bounded by   2r 3 2   Bq d |λ1 | |λ1 | 2 2 2 − 21 +θ MSE = O(Bq dξ1U B ) = O = O d (c pn) . 1 1 |λd | −θ (c1 pn) 2 27 ITERATIVE COLLABORATIVE FILTERING Similarly, if φ(ξ) ≥ min(`, ξ/2L) the mean squared error bound for the algorithm using dist2 is bounded by     1 Bq3 d2 |λ1 | MSE = O = O d2 (c1 pn)− 2 +θ . 1 −θ (c1 pn) 2 For Theorems 4.3 and 4.4, if φ(ξ) ≥ min(`, ξ/2L)  the probability of error  (c1 pn)2θ expression would be dominated by the term nd exp − 4B 2 d . q Comparing results of dist1 and dist2 . We compare the simplified results between the mean squared error bounds for Theorems 4.1 and 4.2 in the setting where the latent variables are sampled from the uniform distribution on [0, 1] and the function f is piecewise L-Lipschitz with minimum piece size of ` ≤ 1. Since φ(ξ) ≥ min(`, ξ/2L), the MSE bound when using dist1 reduces to 1 2 2r 3 2 −( 2 −θ) O(Bq2 dξ1U ), B ) = O((|λ1 |/|λd |) Bq d |λ1 |(c1 pn) while the MSE bound when using dist2 reduces to   2 3 2 −( 21 −θ) O(Bq2 dξ2U ) = O B d |λ |(c pn) . 1 1 B q The only difference is the factor of (|λ1 |/|λd |)2r , which grows to be large when r grows asymptotically with respect to n. As observed from the conditions stated in (3.7), r is constant with respect to n when p = n−1+ for some  > 0. In fact, the reason why the error blows up with a factor of (|λ1 |/|λd |)2r when using dist1 is because we compute the distance by summing product of weights over paths of length 2r. From (3.4), we see that in expectation, when we take the product of edge weights over a path of length r from u to y, instead of computing f (αu , αy ) = eTu QΛQey , the expression concentrates around eTu QΛr Qey , which contains extra factors of Λr−1 . Therefore, by computing over neighborhoods of radius r, the calculation in dist1 will approximate kΛr+1 Q(eu − ev )k22 rather than our intended kΛQ(eu − ev )k22 , thus leading to an error factor of |λd |−2r . We chose ξ1 (n) such that we multiply the estimated distance by |λ1 |2r . Therefore, if |λd | = |λ1 |, then the error does not grow with r. On the other hand, dist2 is computed from a combination of d0 measurements at different path lengths using specific coefficients in order to adjust for this amplification. Each measurement approximates kΛr+i Q(eu − ev )k22 for a different value of i, which can also be written as a linear combination of 28 BORGS-CHAYES-LEE-SHAH the terms (ek Q(eu − ev ))2 . Essentially the different measurements allow us to separate between these terms for all k with distinct values of λk . This cor1 rection leads to a MSE bound of O(Bq3 d2 |λ1 |(c1 pn)− 2 +θ ), which converges even in the ultra sparse sampling regime of p = ω(n−1 dmax(1/(2θ),4/(1−2θ)) ). For a choice of θ = 1/10, this would reduce to p = ω(d5 n−1 ). As compared to information theoretic lower bounds, this sample complexity is optimal with respect to n although the exponent of d is suboptimal. 4.1. Modified algorithm with subsampled Anchor vertices. Recall that in Section 3.3, we discussed a modification of the algorithm to reduce computation by subsampling anchor
vertices, and comparing only to anchor vertices rather than computing all n2 pairwise distances. Let K denote the set of anchor vertices. In order to prove error bounds for the modified algorithm for some entry located at index (u, v), we need to ensure that with high probability there exist anchor vertices that are within “close” distances from both u and v. Then we need the distance estimates between the vertices u, v and the anchor vertices to be accurate enough such that the anchor vertices π(u) and π(v) which minimize dist(u, π(u)) and dist(v, π(v)) (for dist ∈ {dist1 , dist2 }) will also be close in terms of kΛQ(eu − eπ(u) )k2 and kΛQ(ev − eπ(v) )k2 . Finally, since the algorithm estimates F̂ (u, v) = F̂ (π(u), π(v)), it only remains to show that |F̂ (π(u), π(v)) − f (απ(u) , απ(v) )| is bounded, which follows from the proof which we showed before. We state the error bound and proof for the modified algorithm that uses dist2 to estimate pairwise distances, but an equivalent result holds when using dist1 . Theorem 4.5. Let (4.6)-(4.12) hold for some θ ∈ (0, 41 ). For some ξ > 0 and for any u, v ∈ [n], with probability at least    (c1 pn)2θ 1 − O exp (−|K|φ(ξ)) + |K|d exp − 4Bq2 d   2  ξ2U B c3 pn2 φ(ξ2LB )2 − O exp − (1 − o(1)) , 24 the error of the estimate produced by the modified algorithm presented in Section 3.3, using dist2 with parameter r and |K| anchor vertices, is upper bounded by  √  |F̂ (u, v) − f (αu , αv )| = O Bq d min(ξ2U B , ξ) . ITERATIVE COLLABORATIVE FILTERING 29 In the piecewise Lipschitz case where φ(ξ) ≥ min(`, ξ/2L), it follows from the theorem that if we want to match the error bounds of Theorem 4.2, it is sufficient to choose the number of anchor vertices 1 3θ 1 3θ L(c1 pn) 4 + 2 = O(d−3/2 (c1 pn) 4 + 2 ) |K| ≥ 1/2 5 3 2(Bq d |λ1 |) and ξ= Bq d|λ1 | 1 (c1 pn) 2 −θ !1/2 = O(ξ2U B ). Thus it follows that with probability at least    (c1 pn)2θ 1 − O |K|d exp − , 4Bq2 d the error of the estimate produced by the modified algorithm presented in Section 3.3 using dist2 with parameter r is upper bounded by  !1/2  Bq3 d2 |λ1 | . |F̂ (u, v) − f (αu , αv )| ≤ O  1 (c1 pn) 2 −θ We can also present bounds on the maximum error. If we want to bound the maximum error, we need to bound the event that all vertices are close to at least one anchor vertex. In order to prove this event holds with high probability, we will additionally show that we can bound the number of balls of diameter ξ needed to cover the space X1 with respect to the measure P1 . Theorem 4.6. Let (4.6)-(4.12) hold for some θ ∈ (0, 41 ). In addition, let |K|φ(ξ/4)2 ≥ 2. (4.13) For some ξ > 0, with probability at least      |K|φ(ξ/4) (c1 pn)2θ 1 − O nd exp − + exp − 4Bq2 d 8   2  ξ2U B c3 pn2 φ(ξ2LB )2 2 − O |K| exp − (1 − o(1)) , 24 the error of the estimate produced by the modified algorithm presented in Section 3.3, using dist2 with parameter r and |K| anchor vertices, is upper bounded by √ max |F̂ (u, v) − f (αu , αv )| = O(Bq d min(ξ2U B , ξ)). (u,v)∈[n]×[n] 30 BORGS-CHAYES-LEE-SHAH In the piecewise Lipschitz case where φ(ξ) ≥ min(`, ξ/2L), it follows from the theorem that if we want to match the max error bounds of Theorem 4.2, it is sufficient to choose the number of anchor vertices 1 |K| ≥ 3θ 1 3θ 16L(c1 pn) 4 + 2 = O(d−3/2 (c1 pn) 4 + 2 ) 1/2 5 3 (Bq d |λ1 |) and ξ= Bq d|λ1 | 1 (c1 pn) 2 −θ !1/2 = O(ξ2U B ). Thus it follows that with probability at least    (c1 pn)2θ , 1 − O nd exp − 4Bq2 d the error of the estimate produced by the modified algorithm presented in Section 3.3 using dist2 with parameter r is upper bounded by  !1/2  Bq3 d2 |λ1 | . max |F̂ (u, v) − f (αu , αv )| = O  1 (u,v)∈[n]×[n] (c1 pn) 2 −θ 4.2. Asymmetric Matrix. Consider an n × m matrix F which we would like to learn, where F (u, v) = f (αu , βv ) ∈ [0, 1], and f has finite spectrum according to d X f (αu , βv ) = λk q1k (αu )q2k (βv ), k=1 where q1k and q2k are orthonormal `2 functions. We use the transformation presented in 2.2 and simply translate the results from the symmetric matrix model. The new dimensions would be (n + m) × (n + m), the rank is 2d, √ the eigenvalues are each multiplied by nm/(n p + m), and the maximum eigenfunction magnitude Bq is multiplied by (n + m)/ min(n, m). Let φ denote the underestimator function for the probability measure after transforming to the symmetric latent model. We require that the ratio between the dimensions of the matrix m and n must be a constant, i.e. they must grow in proportion. We state the results for the algorithm when using dist2 , although equivalent results also carry over if we use dist1 . Theorem 4.7. Let the following hold for some θ ∈ (0, 41 ): 31 ITERATIVE COLLABORATIVE FILTERING 0. Scaling of n, m. n = Θ(1). m (4.14) 1. Conditions on sampling probability, p. (4.15) p = o((n + m)−1+1/(5+8θ) ) and c1 p(n + m) = ω(1). 2. Conditions on neighborhood radius, r. (4.16)  9c1 p(n+m) 8 r+d0 ≤ (c1 p) −7/8 and   1 √ 7λ2d c1 p nm r+ 2 8|λ1 | ≥ p−6/8 3. Condition on spectrum of Λ. The smallest eigenvalue λd is such that   1 1 (4.17) |λd | = ω (c1 p(n + m))− min( 4 , 2 +θ) The number of distinct magnitude eigenvalues d0 satisfies (4.18) (4.19) (2d0 − 1) ln(2|λ1 |/λgap ) + ln(2d0 ) r≥ , ln(2)   θ 1 ln ≥ (2d0 − 1) ln(2|λ1 |/λgap ) + ln(2d0 ). 2(1 + 2θ) p 4. Condition on distribution of latent features. Define ξ1LB , ξ1U B as q 2Bq d|λ1 | 2 ξ2LB = (n+m)nm min(n,m) (c p(n+m)) 12 −θ , 1 q 130Bq d|λ1 | 2 nm ξ2U B = (n+m) min(n,m) (4.20) 1 −θ (c1 p(n+m)) 2 They satisfy (4.21)     2θ −1 1 p(n+m)) . φ(ξ2LB ) = ω max p, (n + m)−3/4 , ξ2LB exp − min(n,m)(c 2 16(n+m)B d q For any u, v ∈ [m] × [n], with probability greater than     2  2θ ξ2U B c3 p(n+m)2 φ(ξ2LB )2 1 p(n+m)) (1 − o(1)) 1 − O d exp − min(n,m)(c + exp − 2 24 16(n+m)Bq d    (n+m−1)φ(ξ2LB ) − O exp − , 8 32 BORGS-CHAYES-LEE-SHAH for any u, v ∈ [n], the error of the estimate produced by the algorithm when using dist2 and parameter r is bounded by  1/2 ! √ Bq3 d2 |λ1 | |F̂ (u, v) − f (αu , αv )| = O(Bq dξ2U B ) = O . 1 −θ (c1 p(n+m)) 2 It follows that the mean squared error is bounded by    min(n, m)(c1 p(n + m))2θ Bq3 d2 |λ1 | MSE = O + d exp − 1 16(n + m)Bq2 d (c1 p(n+m)) 2 −θ      2 2 c3 p(n+m)2 ξ2U (n+m−1)φ(ξ2LB ) B φ(ξ2LB ) + O exp − + exp − . 24 8 4.3. Categorical Valued Data. If the edge labels take values within m category types, we use the reduction presented in 2.3, where the data is split into m different matrices, each containing the information for a separate category (or edge label). For each category or label i ∈ [m], the associated matrix Fi represents the probability that each datapoint is labeled with i, such that P(Z(u, v) = i) = Fi (u, v) = fi (αu , αv ), where f is a symmetric function having finite spectrum with rank di , fi (αu , αv ) = di X λik qik (αu )qik (αv ), k=1 where {qik }k∈[di ] are orthonormal L2 functions, and d0i denote the number of distinct valued eigenvalues. Let ri , Biq , φi denote the associated parameters for the model associated to each label or category. The algorithm then is applied to each matrix separately to estimate the probability of each category across the different entries. Since we need the estimates across different categories for the same entry to sum to 1, we can simply let the estimate for the m-th category one minus the sum of the estimates for the first m − 1 categories. To obtain an error bound, we can simply use union bound across the m−1 applications of the algorithm, which simply multiplies the error probability by m − 1. We state the results for the algorithm when using dist2 , although equivalent results also carry over if we use dist1 . Theorem 4.8. For some θ ∈ (0, 14 ), let (4.6) hold as well as (4.16)(4.12) for all i ∈ [m − 1] (with ri , di , φi , ξ2LB (i) and ξ2U B (i) in place of r, d, φ, ξ2LB and ξ2U B ). Then, for any (u, v) ∈ [n] × [n], with probability greater ITERATIVE COLLABORATIVE FILTERING 33 than      2  X  2 2 2θ ξ (i)c pn φ (ξ (i)) 1 pn) di exp − (c8B 1−O + exp − 2U B 3 24 i 2LB (1 − o(1))  2 d i∈[m−1] iq i    + O exp − (n−1)φi 8(ξ2LB (i)) , the total variation distance between the true label distribution and the estimate computed by combining m − 1 estimates for each label using the algorithm with dist2 is bounded by   X  B 3 d2 |λi1 | 1/2 1 X iq i . |F̂i (u, v) − fi (αu , αv )| = O  1 2 (c1 pn) 2 −θ i∈[m] i∈[m−1] 4.4. Non-uniform sampling. In the previous models, we always assumed a uniform sampling model, where each entry is observed independently with probability p. However, in reality the probability that entries are observed may not be uniform across all pairs (i, j). In this section we discuss an extension of our result that can handle variations in the sample probability as long as they are still independent across entries, and the sample probability is a function of the latent variables and scales in the same way with respect to n across all entries. The idea is simply to apply the algorithm twice, first to estimate the variations in the sample probability, and second to estimate the data entries normalized by these variations in sample probability. In order for our result to directly translate over, we assume that the data is sampled in two phases, where a subset of the entries are first sampled uniformly with probability p, and then this subset is further subsampled to obtain the final set of entries for which data is observed, allowing for variation across entries in the second sampling phase. The algorithm is assumed to have data on whether an entry is not observed because it was not sampled in the first round or second round. One application for which this type of data could be reasonable would be on an e-commerce platform, where users might be shown items at random (with density p), and each user chooses whether or not to buy and rate each product that s/he is shown, which would then be a function of the user and product features. In this case, the platform would have information about which user-product pairs were made available to the user (uniformly with probability p), and whether or not we observed a rating for each user-product pair. Given that a user decides to purchase and rate the product, the expected rating will be according to another latent function of the user and product features. 34 BORGS-CHAYES-LEE-SHAH Assume a model in which the probability of observing (i, j) is given by pg(αi , αj ), where p is the scaling factor (contains the dependence upon n and is fixed across all edges), and g allows for constant factor variations in the sample probability across entries as a function of the latent variables. Let matrix X indicate the presence of an observation or not, and let matrix M contain the data values for the subset of observed entries. We simply apply our algorithm twice, first on matrix X to estimate the function g, and then on data matrix M to estimate f times g. We can simply divide by the estimate for g to obtain the estimate for f . Assume that Xij distinguishes between an entry which is unobserved because of the uniform scaling parameter p, as opposed to unobserved because of the edge sampling variation g(·, ·). This restriction is due to the fact that our algorithm needs to distinguish between an entry which is observed to have value zero as opposed to an entry which is not observed at all, as discussed in Section 3.7. In fact, for estimating the function g, this requirement is reasonable, as the problem would be unidentifiable otherwise, since the same observation set could result from multiplying p by some constant and dividing the function g by the same constant. We directly apply the theorems from before to obtain error bounds. Let 2 r, Bq , d, d0 , λ1 , λd , ξ2U B , ξ2LB , φ refer to the parameters for estimating the ˜ d˜0 , λ̃1 , λ̃ ˜, ξ˜2 , ξ˜2LB , φ̃ refer to the parameters for function g, and let r̃, B̃q , d, d 1U B estimating the function f times g. We present the results for the algorithm using dist2 , however equivalent results follow when using dist1 . Theorem 4.9. Let the following hold for some θ ∈ (0, 14 ): 1. Conditions on sampling probability, p. p = o(n−1+1/(5+8θ) ) (4.22) and c1 pn = ω(1). 2. Conditions on neighborhood radius, r, r̃. (4.23)  9c1 pn 8  9c1 pn 8 r+d0 r̃+d˜0 ≤ (c1 p)−7/8 −7/8 ≤ (c1 p) and   1 7λ2d c1 pn r+ 2 8|λ1 |  and 7λ2˜c1 pn d 8|λ1 | r̃+ 1 2 ≥ p−6/8 , ≥ p−6/8 . 3. Condition on spectrum of Λ, Λ̃. The smallest eigenvalues λd , λ̃d˜ are such that   1 1 |λd |, |λ̃d | = ω (c1 pn)− min( 4 , 2 +θ) . (4.24) 35 ITERATIVE COLLABORATIVE FILTERING The number of distinct magnitude eigenvalues d0 and d˜0 satisfy (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ) r≥ , ln(2)   θ 1 ≥ (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ), ln 2(1 + 2θ) p ˜ ) + ln(d˜0 ) (d˜0 − 1) ln(2|λ̃1 |/λgap r̃ ≥ , ln(2)   θ 1 ˜ ) + ln(d˜0 ). ≥ (d˜0 − 1) ln(2|λ̃1 |/λgap ln 2(1 + 2θ) p (4.25) (4.26) (4.27) (4.28) 4. Condition on distribution of latent features. Define ξ2LB , ξ2U B , ξ˜2LB , ξ˜2U B : 2 ξ2LB = (4.29) 2 ξ˜2LB = Bq d|λ1 | 1 2 and ξ2U B = 1 2 and ξ˜2U B = (c1 pn) 2 −θ ˜ 1| B̃q d|λ (c1 pn) 2 −θ 65Bq d|λ1 | 1 , 1 . (c1 pn) 2 −θ ˜ 1| 65B̃q d|λ (c1 pn) 2 −θ They satisfy (4.30)     2θ −1 1 pn) exp − (c8B φ(ξ2LB ) = ω max p, n−3/4 , ξ2LB , 2d q     pn)2θ −1 exp − (c81B̃ φ̃(ξ˜2LB ) = ω max p, n−3/4 , ξ˜2LB . 2 d˜ q For any u, v ∈ [n], with probability greater than   2    ξ2U B c3 pn2 φ(ξ2LB )2 (c1 pn)2θ + exp − (1 − o(1)) 1 − O d exp − 8Bq2 d 24 !!    2θ (c pn) (n − 1)φ(ξ2LB ) 1 − O exp − − O d˜exp − 8 8B̃q2 d˜ ! !! 2 2 φ̃(ξ˜ 2 ˜ ξ˜2U c pn ) (n − 1) φ̃( ξ ) 2LB 2LB B 3 − O exp − (1 − o(1)) + exp − , 24 8 the error of the estimate produced by the algorithm when using dist2 is bounded by p √ |F̂ (u, v) − f (αu , αv )| = O((B̃q d˜ξ˜2U B + Bq dξ2U B )(g(αu , αv ) − Bq dξ2U B )−1 )   1/2 
1/2  3 2 3 2 − 1−2θ ˜ Bq d |λ1 | + B̃q d |λ̃1 | (c1 pn) 4 =O  × g(αu , αv ) − O Bq3 d3 |λ1 | 1 (c1 pn) 2 −θ 1/2 !!−1 . 36 BORGS-CHAYES-LEE-SHAH The error bounds show that if g(αi , αj ) is very small, then the error bounds for estimating f (αi , αj ) are larger. This is as expected, since a small value of g(αi , αj ) means that the probability of observing the data is small, such that there are fewer observations available to estimate f (αi , αj ). Since g(αi , αj ) is a constant with respect to n, as n goes to infinity, the estimator 1−2θ will still converge at the rate of (pn)− 4 . Proof. We bound the error as a function of the error of the estimate ĝ and the estimate fˆg. fˆg fˆuv − f (αu , αv ) = uv − f (αu , αv ) ĝuv fˆg − f (αu , αv )g(αu , αv ) f (αu , αv )g(αu , αv ) f (αu , αv )ĝuv = uv + − ĝuv ĝuv ĝuv fˆg − f (αu , αv )g(αu , αv ) + f (αu , αv )(g(αu , αv ) − ĝuv ) = uv . g(αu , αv ) − (g(αu , αv ) − ĝuv ) Therefore, using the fact that |f (αu , αv )| ≤ 1, |fˆuv − f (αu , αv )| ≤ |fˆg uv − f (αu , αv )g(αu , αv )| + |g(αu , αv ) − ĝuv | . g(αu , αv ) − |g(αu , αv ) − ĝuv | By application of Theorem 4.2 for bounding |fˆg uv − f (αu , αv )g(αu , αv )| and |g(αu , αv ) − ĝuv |, we obtain the desired result. 5. Discussion. In this paper, we presented a similarity based collaborative filtering algorithm which is provably consistent in sparse sampling regimes, as long as the sample probability p = ω(n−1 ). The algorithm computes similarity between two indices (rows, nodes or vertices) by comparing their local neighborhoods. Our model assumes that the data matrix is generated according to a latent variable model, in which the weight on an observed edge (u, v) is equal in expectation to a function f over associated latent variables αu and αv . We presented two variants for computing similarities (or distances) between vertices. Computing dist1 does not require knowledge of the spectrum of f , but the estimate requires p to be polynomially larger than n in order to guarantee the expected squared error converges to zero. Computing dist2 uses the knowledge of the spectrum of f , but it provides an estimate that is provably consistent for a significantly sparse regime, only requiring that p = ω(n−1 ). The mean squared error of both algorithms is 1 bounded by O((pn)− 2 +θ ) for any θ ∈ (0, 41 ). Since the computation is based on of comparing local neighborhoods within the graph, the algorithm can ITERATIVE COLLABORATIVE FILTERING 37 be easily implemented for large scale datasets where the data may be stored in a distributed fashion optimized for local graph computations. In practice, we do not know the model parameters, and we would use cross validation to tune the radius r and threshold ξ(n). If r is either too small or too large, then the vector Nu,r will be too sparse. The threshold ξ(n) trades off between bias and variance of the final estimate. Since we do not know the spectrum, dist1 may be easier to compute, and still enjoys good properties as long as r is not too large. When the sampled observations are not uniform across entries, the algorithm may require more modifications to properly normalize for high degree hub vertices, as the optimal choice of r may differ depending on the local sparsity. The key computational step of our algorithm involves comparing the expanded local neighborhoods of pairs of vertices to find the “nearest neighbors”. The local neighborhoods can be computed in parallel, as they are independent computations. Furthermore, the local neighborhood computations are suitable for systems in which the data is distributed across different machines in a way that optimizes local neighborhood queries. The most expensive part of our algorithm involves computing similarities for all pairs of vertices in order to determine the set of nearest neighbors. However, it would be possible to use approximate nearest neighbor techniques to greatly reduce the computation such that approximate nearest neighbor sets could be computed with significantly fewer than n2 pairwise comparisons. Additionally our modification of subsampling the vertices reduces the pairwise comparisons to n|K|. When the latent function governing the expected matrix behavior is piecewise Lipschitz, it is suffi1 3θ cient to choose |K| = Θ(d−3/2 (c1 pn) 4 + 2 ), which is significantly smaller than Θ(n). 6. Proof Sketch. The final estimate F̂ (u, v) is computed by averaging over datapoints, as specified in (3.2). In our analysis, we will show that dist1 (u, v) and dist2 (u, v) are close estimates for kΛQ(eu − ev )k22 , so that for a majority of the datapoints (a, b) ∈ Euv , it indeed holds that F (a, b) ≈ F (u, v), thus upper bounding the bias. We additionally show that the number of included datapoints, |Euv | is sufficiently large to reduce the variance. To ensure that F (u, v) is close to F (a, b) for (a, b) ∈ Euv , it would be sufficient to bound kΛQ(eu − ea )k2 and kΛQ(ev − eb )k2 , since |f (αu , αv ) − f (αa , αb )| = |eTu QT ΛQev − eTa QT ΛQeb | = |eTu QT ΛQ(ev − eb ) − (ea − eu )T QT ΛQeb | √ √ ≤ Bq dkΛQ(ev − eb )k2 + Bq dkΛQ(eu − ea )k2 , 38 BORGS-CHAYES-LEE-SHAH where the last step follows from assuming that |qk (α)| ≤ Bq for all k ∈ [d] and α ∈ X1 . Since Euv is defined by either the distances computed by dist1 or dist2 , by showing that dist1 (u, v) and dist2 (u, v) are good estimates of kΛQ(eu − ev )k22 , it will follow that kΛQ(ev − eb )k2 and kΛQ(eu − ea )k2 are small for a majority of (a, b) ∈ Euv . A key lemma of the proof is to show that with high probability, the difference between (6.1) 1−c1 p c2 p Ñu,r − Ñv,r
T M2 Ñu,r+i − Ñv,r+i and (eu − ev )T QT Λ2r+i+1 Q(eu − ev ) = kΛr+ i+1 2
Q(eu − ev )k22 is vanishing at a rate of O(Bq d|λ1 |2r+i (c1 pn)−1/2+θ ). This results in bounds on both dist1 and dist2 . There are a few steps that are involved in showing this bound. In Lemma 6.1, we show high probability concentration results on E1 , which is used to expand the neighborhoods, and determines Nu,v and Su,r . First we show that with high probability, the expansion factor |Su,s |/|Su,s−1 | is close to c1 pn. Second we show that for all k ∈ [d], eTk QÑu,s grows as λsk , according to its expected behavior. And we similarly show that kÑu,s k1 grows as O(|λ1 |s ). Lemma 6.1. For any u ∈ [n] and given θ ∈ (0, 14 ), if d = o((c1 pn)2θ ), with probability at least   (c1 pn)2θ 1 − 4(d + 2) exp − , 4Bq2 d the following statements hold: 1. For all s ∈ [r], |Su,s | 7c1 pn 9c1 pn ≤ . ≤ 8 |Su,s−1 | 8 2. For all s ∈ [r], k ∈ [d], t ∈ [r − s], |eTk QÑu,s+t − eTk Λt QÑu,s | ≤ 2|λk |t−1  1 (c1 pn) 2 −θ 3. For all s ∈ [r], kÑu,s k1 ≤ Bq2 d|λ1 |s (1 + o(1)). |λd |2 2|λ1 | s . ITERATIVE COLLABORATIVE FILTERING 39 Next, we show high probability concentration results on E2 , which is used to measure edges between the neighborhood boundaries. Specifically, we can show that (6.2) (6.3) 1−c1 p T 0 c2 p Ñu,r M2 Ñv,r T ≈ Ñu,r QΛQÑv,r0 0 ≈ eTu QΛr+r Qev . This allows us to proves bounds on kΛQ(eu −ev )k22 as a function of dist1 (u, v) and dist2 (u, v). Lemma 6.2. For any u, v ∈ [n] and given θ ∈ (0, 41 ), if d = o((c1 pn)2θ ), with probability at least ! 1/4

−1/4 2θ c (n − 1) 1 1 p)p 1 pn) −8d0 exp − (1−c −n exp − , 1−8(d+2) exp − (c4B 2 5|λ1 |Bq2 d qd 3 the following statements hold for large enough n: kΛQ(eu − ev )k22 ≥ |λ1 |−2r dist1 (u, v) − 32Bq d|λ1 |2r+1
, (c1 pn)−1/4 32Bq d|λ1 |2r+1
, (c1 pn)−1/4 32Bq d|λ1 | kΛQ(eu − ev )k22 ≤ |λd |−2r dist1 (u, v) + kΛQ(eu − ev )k22 ≥ dist2 (u, v) − kΛQ(eu − ev )k22 ≤ dist2 (u, v) + 1 (c1 pn)− 2 +θ 32Bq d|λ1 | 1 (c1 pn)− 2 +θ , ]. While the above lemmas provide concentration results for a single vertex or a pair of vertices, we need to show that the above properties hold for a majority of the vertices. Let Vu1 denote the set which consists of vertices a such that dist1 (u, a) < ξ1 (n), and similarly let Vu2 denote the set which consists of vertices a such that dist2 (u, a) < ξ2 (n). We can show that Vu1 and Vu2 consist mostly of vertices a such that kΛQ(eu − ea )k2 is small, and they do not contain many vertices for which kΛQ(eu − ea )k2 is large. If the latent variables are sampled uniformly from the unit interval and we assume f is Lipschitz, we will be able to prove lower bounds on the number of vertices for which kΛQ(eu − ea )k2 is small, i.e. vertices a such that f (αu , αb ) is close to f (αa , αb ) for all b. Then we use Lemmas 6.1 and 6.2 to show that a majority of vertices have “good distance estimates”, which leads to a lower bound on the fraction of truly close vertices within Vu1 and Vu2 . 40 BORGS-CHAYES-LEE-SHAH Let Wu1 denote the vertices within Vu1 whose true distance is also small, where the threshold is defined by the bounds presented in Lemma 6.2. A vertex v ∈ Wu1 if and only if v ∈ Vu1 and kΛQ(eu − ev )k22 <  |λ1 | |λd | 2r 65Bq d|λ1 | 1 (c1 pn) 2 −θ . Lemma 6.2 implies that for a vertex v, with high probability, v ∈ Vu1 implies that v ∈ Wu1 , which leads to a lower bound on the ratio |Wu1 |/|Vu1 |. Similarly, we let Wu2 denote the vertices within Vu2 whose true distance is also small, where the threshold is defined by the bounds presented in Lemma 6.2. A vertex v ∈ Wu2 if and only if v ∈ Vu2 and kΛQ(eu − ev )k22 < 65Bq d|λ1 | 1 (c1 pn) 2 −θ . Lemma 6.2 implies that for a vertex v, with high probability, v ∈ Vu2 implies that v ∈ Wu2 , which leads to a lower bound on the ratio |Wu2 |/|Vu2 |. Therefore, we can show in Lemma 6.3 that the majority of vertices within Vu1 and Vu2 are indeed truly close in function value. Lemma 6.3. Assume that the latent variables are sampled from the uniform distribution on [0, 1] and the function f is L-Lipschitz. For any u ∈ [n] and given θ ∈ (0, 41 ), if d = o((c1 pn)2θ ), with probability at least    (c1 pn)2θ 1 − O d exp − , 8Bq2 d the fraction of good neighbors is bounded below by  (c pn) 1−2θ  (c pn)2θ  4 |Wu1 | 1 1 =1−O exp − , |Vu1 | 8Bq2 d (Bq d|λ1 |)1/2  (c pn) 1−2θ  (c pn)2θ  4 |Wu2 | 1 1 =1−O exp − . 1/2 |Vu2 | 8Bq2 d (Bq d|λ1 |) Finally, we show high probability concentration results on E3 , which is used to compute the final estimate. Conditioned on the neighbors being chosen well, we show that E3 is evenly spread such that the majority of the entries of Euv are within Wu1 × Wv1 or Wu2 × Wv2 . This bounds the bias of the estimator. Then we show that the total number of these entries is ITERATIVE COLLABORATIVE FILTERING 41 sufficiently large to reduce the variance from entry-level noise. These high probability bounds on the squared error naturally lead to the stated bound on the MSE. The results can be extended beyond L-Lipschitz and α ∼ U [0, 1] as long as we can lower bound the probability that a randomly sampled vertex a is within a close neighborhood of a vertex u with respect to the quantity kΛQ(eu − ea )k22 . This includes piecewise Lipschitz functions and higher dimensional latent variable representations. References. Abbe, E. and Sandon, C. (2015a). Community detection in general stochastic block models: Fundamental limits and efficient algorithms for recovery. In Foundations of Computer Science (FOCS), 2015 IEEE 56th Annual Symposium on 670–688. IEEE. Abbe, E. and Sandon, C. (2015b). Recovering communities in the general stochastic block model without knowing the parameters. In Advances in neural information processing systems. Abbe, E. and Sandon, C. (2016). Detection in the stochastic block model with multiple clusters: proof of the achievability conjectures, acyclic BP, and the informationcomputation gap. Advances in neural information processing systems. Airoldi, E. M., Costa, T. B. and Chan, S. H. (2013). Stochastic blockmodel approximation of a graphon: Theory and consistent estimation. In Advances in Neural Information Processing Systems 692–700. Aldous, D. J. (1981). Representations for partially exchangeable arrays of random variables. J. Multivariate Anal. 11 581 - 598. Anandkumar, A., Ge, R., Hsu, D. and Kakade, S. (2013). A tensor spectral approach to learning mixed membership community models. In Conference on Learning Theory 867–881. Austin, T. (2012). Exchangeable random arrays. Technical Report, Notes for IAS workshop. Bordenave, C., Lelarge, M. and Massoulié, L. (2015). Non-backtracking spectrum of random graphs: community detection and non-regular ramanujan graphs. In Foundations of Computer Science (FOCS), 2015 IEEE 56th Annual Symposium on 1347–1357. IEEE. Borgs, C., Chayes, J. and Smith, A. (2015). Private graphon estimation for sparse graphs. In Advances in Neural Information Processing Systems 1369–1377. Borgs, C., Chayes, J. T., Lovász, L., Sós, V. T. and Vesztergombi, K. (2008). Convergent sequences of dense graphs I: Subgraph frequencies, metric properties and testing. Advances in Mathematics 219 1801–1851. Borgs, C., Chayes, J. T., Cohn, H. and Zhao, Y. (2014a). An Lp theory of sparse graph convergence I: limits, sparse random graph models, and power law distributions. arXiv preprint arXiv:1401.2906. Borgs, C., Chayes, J. T., Cohn, H. and Zhao, Y. (2014b). An Lp theory of sparse graph convergence II: LD convergence, quotients, and right convergence. arXiv preprint arXiv:1408.0744. Borgs, C., Chayes, J. T., Cohn, H. and Ganguly, S. (2015). Consistent nonparametric estimation for heavy-tailed sparse graphs. arXiv preprint arXiv:1508.06675. Borgs, C., Chayes, J. T., Cohn, H. and Holden, N. (2016). Sparse exchangeable graphs and their limits via graphon processes. arXiv preprint arXiv:1601.07134. 42 BORGS-CHAYES-LEE-SHAH Candes, E. and Recht, B. (2009). Exact matrix completion via convex optimization. Communications of the ACM 55 111–119. Candès, E. J. and Tao, T. (2010). The power of convex relaxation: Near-optimal matrix completion. IEEE Transactions on Information Theory 56 2053–2080. Chatterjee, S. (2015). Matrix estimation by universal singular value thresholding. The Annals of Statistics 43 177–214. Chen, Y. and Wainwright, M. J. (2015). Fast low-rank estimation by projected gradient descent: General statistical and algorithmic guarantees. arXiv preprint arXiv:1509.03025. Davenport, M. A., Plan, Y., van den Berg, E. and Wootters, M. (2014). 1-bit matrix completion. Information and Inference 3 189–223. Decelle, A., Krzakala, F., Moore, C. and Zdeborová, L. (2011). Asymptotic analysis of the stochastic block model for modular networks and its algorithmic applications. Phys. Rev. E 84 066106. Diaconis, P. and Janson, S. (2008). Graph limits and exchangeable random graphs. Rendiconti di Matematica VII 33-61. Gao, C., Lu, Y. and Zhou, H. H. (2015). Rate-optimal graphon estimation. The Annals of Statistics 43 2624–2652. Goldberg, D., Nichols, D., Oki, B. M. and Terry, D. (1992). Using Collaborative Filtering to Weave an Information Tapestry. Commun. ACM. Hoover, D. N. (1981). Row-column exchangeability and a generalized model for probability. In Exchangeability in Probability and Statistics (Rome, 1981) 281 - 291. Karrer, B., Newman, M. E. J. and Zdeborová, L. (2014). Percolation on Sparse Networks. Phys. Rev. Lett. 113 208702. Keshavan, R. H., Montanari, A. and Oh, S. (2010a). Matrix completion from a few entries. IEEE Transactions on Information Theory 56 2980–2998. Keshavan, R. H., Montanari, A. and Oh, S. (2010b). Matrix completion from noisy entries. Journal of Machine Learning Research 11 2057–2078. Klopp, O., Tsybakov, A. B. and Verzelen, N. (2015). Oracle inequalities for network models and sparse graphon estimation. To appear in Annals of Statistics. Koren, Y. and Bell, R. (2011). Advances in Collaborative Filtering. In Recommender Systems Handbook 145-186. Springer US. Krzakala, F., Moore, C., Mossel, E., Neeman, J., Sly, A., Zdeborov, L. and Zhang, P. (2013). Spectral redemption in clustering sparse networks. Proceedings of the National Academy of Sciences 110 20935-20940. Lee, C. E., Li, Y., Shah, D. and Song, D. (2016). Blind Regression: Nonparametric Regression for Latent Variable Models via Collaborative Filtering. In Advances in Neural Information Processing Systems 29 2155–2163. Linden, G., Smith, B. and York, J. (2003). Amazon.Com Recommendations: Item-toItem Collaborative Filtering. IEEE Internet Computing 7 76–80. Lovász, L. (2012). Large networks and graph limits 60. American Mathematical Society Providence. Massoulié, L. (2014). Community Detection Thresholds and the Weak Ramanujan Property. In Proceedings of the Forty-sixth Annual ACM Symposium on Theory of Computing. STOC ’14 694–703. ACM, New York, NY, USA. Mossel, E., Neeman, J. and Sly, A. (2017). A proof of the block model threshold conjecture. Combinatorica. Negahban, S. and Wainwright, M. J. (2011). Estimation of (near) low-rank matrices with noise and high-dimensional scaling. The Annals of Statistics 1069–1097. Ning, X., Desrosiers, C. and Karypis, G. (2015). Recommender Systems Handbook ITERATIVE COLLABORATIVE FILTERING 43 A Comprehensive Survey of Neighborhood-Based Recommendation Methods, 37-76. Springer US. Recht, B. (2011). A simpler approach to matrix completion. Journal of Machine Learning Research 12 3413–3430. Steurer, D. and Hopkins, S. (2017). Bayesian estimation from few samples: community detection and related problems. Veitch, V. and Roy, D. M. (2015). The class of random graphs arising from exchangeable random measures. arXiv preprint arXiv:1512.03099. Wolfe, P. J. and Olhede, S. C. (2013). Nonparametric graphon estimation. arXiv preprint arXiv:1309.5936. Xu, J., Massoulié, L. and Lelarge, M. (2014). Edge label inference in generalized stochastic block models: from spectral theory to impossibility results. In Conference on Learning Theory 903–920. Zhang, Y., Levina, E. and Zhu, J. (2015). Estimating network edge probabilities by neighborhood smoothing. arXiv preprint arXiv:1509.08588. One Memorial Drive, Cambridge, MA 02142 E-mail: borgs@microsoft.com E-mail: jchayes@microsoft.com E-mail: celee@mit.edu 77 Massachusetts Avenue, Cambridge, MA 02139 E-mail: devavrat@mit.edu 44 BORGS-CHAYES-LEE-SHAH APPENDIX A: PROOF OF LEMMAS In this section we provide detailed proofs of the lemmas outlined above, which we split further into a sequence of smaller lemmas. We order them according to the concentration results associated to different sets of random variables in the dataset, E1 , E2 , {θu }u∈[n] , and E3 , which also somewhat follows the steps of the algorithm. Lemma 6.1 follows directly from Lemmas A.2, A.4, A.8, and A.12. Lemma 6.2 follows directly from Lemmas A.12, A.19, A.17 and A.18.Lemma 6.3 follows directly from Lemmas A.12, A.13, A.20, A.21 and A.22. In figure 1, we show the dependencies amongst the Lemmas in Section 7. The main theorem results from the final Lemmas A.32, A.34, A.33, and A.35. The blue rectangles indicate Lemmas that prove probabilistic concentration statements, i.e. bounding the probabilities of bad events. The red ovals indicate Lemmas that show good properties result from conditioning on the good events, which leads to bounds on the error; these are not probabilistic statements. We introduce additional notation used within the proofs. For each vertex u ∈ [n], let Fu,0 ⊂ Fu,1 ⊂ Fu,2 . . . Fu,n be the filtration associated with revealing the latent vertex parameters and observed edge weights within local neighborhoods of vertex u with respect to the edge set E1 . Note that Su,s , Bu,s , {αv }v∈Su,s , Nu,s , and Ñu,s are all measurable with respect to Fu,s . Let FE1 be the sigma algebra associated with revealing all latent vertex parameters {αv }v∈[n] and edge weights of E1 . Let FE2 ,E1 be the sigma algebra associated with revealing all latent vertex parameters {αv }v∈[n] and edge weights of E1 and E2 . Note that dist1 (u, v), dist2 (u, v), Vu1 , Vu2 , Wu1 , and Wu2 are all measurable with respect to FE2 ,E1 . We will use I(·) to denote the indicator function, 1 to denote the vector of all ones, and ej to denote the standard basis vector with a one at coordinate j and zero elsewhere. A.1. Sparsity of Local Neighborhood Grows Exponentially. We first begin by showing concentration results pertaining to the sparsity of observed entries. Specifically, we prove in Lemma A.1 that |Su,s+1 | is approx
c | 1 − (1 − c p)|Su,s | conditioned on F . This will imately equal to |Bu,s 1 u,s allow us to show in Lemma A.2 that the size of the (s + 1)-radius boundary of vertex u is approximately c1 pn times larger than the size of the s-radius boundary of vertex u. Lemma A.1. For any s ∈ Z+ such that |Su,s | < (c1 p)−7/8 , and t ∈ 45 ITERATIVE COLLABORATIVE FILTERING Fig 1. Chart of Dependencies amongst Lemmas in Section 7. The blue rectangles indicate Lemmas that bound probabilities of bad events. The red ovals indicate Lemmas that show good properties result from conditioning on good events. L 7.4 L 7.2 L 7.1 L 7.3 L 7.5 L 7.7 L 7.6 L 7.9 L 7.8 L 7.10 L 7.11 L 7.12 L 7.15 L 7.14 L 7.13 L 7.18 L 7.17 L 7.16 L 7.19 L 7.22 L 7.20 L 7.21 L 7.24 L 7.25 L 7.23 L 7.31 L 7.30 L 7.26 L 7.32 L 7.33 L 7.27 L 7.28 L 7.34 L 7.29 L 7.35 46 BORGS-CHAYES-LEE-SHAH (0, 1 − c1 p|Su,s |),    c c P |Su,s+1 | − |Bu,s | 1 − (1 − c1 p)|Su,s | > |Bu,s ||Su,s |c1 pt ! c ||S |c p t2 |Bu,s u,s 1 ≤ 2 exp − . 3 Fu,s  Proof. Conditioned on Fu,s , the vertices which have a path to u which is c contains shorter than s+1 is denoted by Bu,s , such that the complement Bu,s all vertices which are eligible to be in Su,s+1 . A vertex is at distance s + 1 from vertex u if it is connected to a vertex which is at distance s from u, and it is not connected to any vertices at distance less than s from u. Therefore, in order to count the number of vertices that are at distance s + 1, we can c which have at least one edge to any count the number of vertices within Bu,s vertex in Su,s . By construction, an edge is in E1 with probability c1 p, such that for a c , vertex j ∈ Bu,s E[I(j ∈ Su,s+1 ) | Fu,s ] = P(∪b∈Su,s I((b, j) ∈ E1 ) / E1 ) = 1 − P(∩b∈Su,s I((b, j) ∈ = 1 − (1 − c1 p)|Su,s | . c such that i 6= j, I(i ∈ Su,s+1 ) is indepenFor a different vertex i ∈ Bu,s dent from I(j ∈ Su,s+1 ) since the involved edge sets are disjoint. Therefore, |Su,s+1 | conditioned on Fus is distributed as a Binomial random variable c | and 1 − (1 − c p)|Su,s | . Therefore, with parameters |Bu,s 1   c E [ |Su,s+1 || Fu,s ] = |Bu,s | 1 − (1 − c1 p)|Su,s | . By Chernoff’s bound, for some δ ∈ (0, 1),     c c P |Su,s+1 | − |Bu,s | 1 − (1 − c1 p)|Su,s | > δ|Bu,s |(1 − (1 − c1 p)|Su,s | ) (A.1) c |(1 − (1 − c p)|Su,s | ) δ 2 |Bu,s 1 < 2 exp − 3 ! . By the condition that |Su,s | < (c1 p)−7/8 = o((c1 p)−1 ), (A.2) 1 − (1 − c1 p)|Su,s | ≥ c1 p|Su,s |(1 − c1 p|Su,s |/2) = c1 p|Su,s |(1 − o(1)). 47 ITERATIVE COLLABORATIVE FILTERING Therefore, for any t ∈ (0, 1 − c1 p|Su,s |), we can verify that δ ∈ (0, 1) for the choice of t|Su,s |c1 p δ := . (1 − (1 − c1 p)|Su,s | ) By substituting this choice of δ into (A.1), it follows that    c c ||Su,s |c1 p | 1 − (1 − c1 p)|Su,s | > t|Bu,s P |Su,s+1 | − |Bu,s ! c ||S |2 c2 p2 t2 |Bu,s u,s 1 ≤ 2 exp − 3(1 − (1 − c1 p)|Su,s | ) ! c ||S |c p t2 |Bu,s u,s 1 ≤ 2 exp − , 3 Fu,s 
where the last inequality follows from 1 − (1 − c1 p)|Su,s | ≤ c1 p|Su,s |. Let us define (A.3) n
− 1 +θ o c c ||Su,h−1 |c1 p c1 pnBq4 2 |(1 − (1 − c1 p)|Su,h−1 | )| < |Bu,h−1 . A1u,h := ||Su,h | − |Bu,h−1 Lemma A.2. satisfies Assuming that c1 pn = ω(1), for any u ∈ [n] and s which  9c1 pn 8 s ≤ (c1 p)−7/8 , conditioned on ∩sh=1 A1u,h , for sufficiently large n, for all h ∈ [s], "  h   h # 9 7 |Su,h | ∈ c1 pn , c1 pn 8 8 Proof. Recall the definition of event A1u,h from (A.3). Using the fact c that (1 − c1 p)|Su,h−1 | ≥ 1 − c1 p|Su,h−1 | and |Bu,h−1 | ≤ n, conditioned on 1 event Au,h , 
− 1 +θ  c |Su,h | ≤ |Bu,h−1 | 1 − (1 − c1 p)|Su,h−1 | + |Su,h−1 |c1 p c1 pnBq4 2 
− 1 +θ  ≤ n c1 p|Su,h−1 | + |Su,h−1 |c1 p c1 pnBq4 2 = c1 pn|Su,h−1 | (1 + o(1)) , 48 BORGS-CHAYES-LEE-SHAH where the last inequality follows from the assumption that c1 pn = ω(1).
− 1 +θ Specifically for c1 pnBq4 ≥ 26/(1−2θ) , it follows by c1 pnBq4 2 ≤ 1/8 that   9 c1 pn|Su,h−1 |. |Su,h | ≤ 8 Therefore, conditioned on ∩sh=1 A1u,h , for all h ∈ [n], h   9 c1 pn . |Su,h | ≤ 8 (A.4) By assumption on the value of s, it follows that |Su,h | ≤ (c1 p)−7/8 = o((c1 p)−1 ) for all h ∈ [s]. Therefore, (A.5) 1 − (1 − c1 p)|Su,h | ≥ c1 p|Su,h |(1 − c1 p|Su,h |/2) = c1 p|Su,h |(1 − o(1)). Conditioned on event A1u,h , 
− 1 +θ  c |Su,h | ≥ |Bu,h−1 | 1 − (1 − c1 p)|Su,h−1 | − |Su,h−1 |c1 p c1 pnBq4 2 
− 1 +θ  c | c1 p|Su,h−1 |(1 − c1 p|Su,h−1 |/2) − |Su,h−1 |c1 p c1 pnBq4 2 ≥ |Bu,h−1 
− 1 +θ  c ≥ |Bu,h−1 |c1 p|Su,h−1 | 1 − (c1 p)1/8 /2 − c1 pnBq4 2 (A.6) c |c1 p|Su,h−1 | (1 − o(1)) , = |Bu,h−1 where the last step followed from the assumption that c1 pn = ω(1) and c |, it follows that c1 p = o(1). By plugging (A.4) into the definition of |Bu,h−1 c | |Bu,h−1 =n− h−1 X t=0 |Su,h | ≥ n − h−1   X 9 t=0 8 t c1 pn h 9 −1 8
c1 pn . 9 8 c1 pn − 1
=n−
By assumption on the value of s, (A.7) c |Bu,h−1 |≥n− (c1 p)−7/8 − 1
. 9 8 c1 pn − 1 Since c1 p = ω( n1 ), we have that n−1 (c1 p)−15/8 = o(n7/8 ). Further, (c1 p)−1 = ω(1). Therefore, it follows that (A.8) c |Bu,h−1 | = n − o(n7/8 ). ITERATIVE COLLABORATIVE FILTERING 49 Therefore, by plugging into (A.6), it follows that |Su,h | ≥ c1 pn|Su,h−1 |(1 − o(1)). (A.9) For sufficiently large n, |Su,h | ≥ (A.10) 7c1 pn |Su,h−1 |. 8 Therefore, conditioned on ∩sh=1 A1u,h , for all h ∈ [n],   h 7 |Su,h | ≥ c1 pn . 8 (A.11) A.2. Concentration of Paths within Local Neighborhood. We show concentration of the (s + 1)-radius boundary of node u in terms of eTk QNu,s+1 conditioned on Fu,s and |Su,s+1 |. Specifically, for every k ∈ [d], we prove that ek QÑu,s+1 is approximately equal to eTk ΛQÑu,s . For any k ∈ [d] and s ∈ Z+ , Lemma A.3.  P eTk QÑu,s+1 − eTk ΛQÑu,s >t Fu,s , Su,s+1  3|Su,s+1 |t2 ≤ 2 exp − 6kÑu,s k1 + 2Bq t Proof. By definition, X eTk QNu,s+1 = Nu,s+1 (i)qk (αi ) i∈Su,s+1 X = X I((b, i) ∈ Tu )Nu,s (b)M1 (b, i)qk (αi ). i∈Su,s+1 b∈Su,s Due to the construction of Tu as a BFS tree, for all i, X I((b, i) ∈ Tu ) = I(i ∈ Su,s+1 ) ∈ {0, 1}. b∈Su,s If there were more than one possible parent for i in the graph defined by edge set E1 , Tu is constructed by simply choosing one uniformly at random. Therefore, conditioned on I(i ∈ Su,s+1 ), the parent of i is equally likely to have been any of b ∈ Su,s , such that (A.12) E [I((b, i) ∈ Tu ) | I(i ∈ Su,s+1 ), Fu,s ] = 1 . |Su,s | ! 50 BORGS-CHAYES-LEE-SHAH For (b, i) ∈ Tu , by construction, M1 (b, i) = M (b, i) = Z(b, i). Recall that Z(b, i) ∈ [0, 1] is independent from the tree structure (the sampling of edges E), and its expectationPis a function of αb and αi . Let us define Xi = b∈Su,s I((b, i) ∈ Tu )Nu,s (b)Z(b, i)qk (αi ), such that eTk QNu,s+1 = X Xi . i∈Su,s+1 By the model assumptions that Nu,s (b) ∈ [0, 1], Z(b, i) ∈ [0, 1], and |qk (αi )| ∈ [0, Bq ], it follows that |Xi | ≤ Bq . If i ∈ / Su,s+1 , then Xi = 0 with probability 1. We next compute the expectation and variance of Xi and then proceed to apply Bernstein’s inequality to bound eTk QNu,s+1 . Recall that Nu,s and Su,s are fixed conditioned on Fu,s . Conditioned on Fu,s and Su,s+1 , Xi are identically distributed for i ∈ Su,s+1 . We can verify that for i 6= j, Xi is independent from Xj conditioned on Fus and Su,s+1 because αi is independent from αj , and the variables I((b, i) ∈ Tu ) are only correlated for the same i but not correlated across different potential children vertices, i.e. the tree constraint only enforces that each child has exactly one parent, but allows parents to have arbitrarily many children. Therefore eTk QNu,s+1 is a sum of iid random variables. We use the independence of sampled edge set E1 from the latent variables, along with (A.12) and the law of iterated expectations to show that E[Xi | i ∈ Su,s+1 , Fus ]  X = E I((b, i) ∈ Tu )Nu,s (b)Z(b, i)qk (αi )  i ∈ Su,s+1 , Fu,s  b∈Su,s X  1  = Nu,s (b)E[Z(b, i)qk (αi ) | Fu,s ] |Su,s | b∈Su,s X  1  Nu,s (b)E[E[Z(b, i) | αi , Fu,s ]qk (αi ) | Fu,s ]. = |Su,s | b∈Su,s For b ∈ Su,s , αb is fixed when conditioned on Fus , such that E[Z(b, i) | αi , Fus ] = ITERATIVE COLLABORATIVE FILTERING 51 f (αb , αi ). Additionally, by the spectral decomposition of f , E[Xi | i ∈ Su,s+1 , Fu,s ] X  1  Nu,s (b)E[f (αb , αi )qk (αi ) | Fu,s ] = |Su,s | b∈Su,s X  1  X λh qh (αb )qh (αi )qk (αi ) | Fu,s ] = Nu,s (b)E[ |Su,s | b∈Su,s h∈[d] X  1  = Nu,s (b)λk qk (αb ) |Su,s | b∈Su,s   1 = eT ΛQNu,s |Su,s | k (A.13) = eTk ΛQÑu,s , where we used the orthonormality property of the eigenfunctions qh (·). Using similar arguments, we can compute a bound on the variance of Xi , using P the fact that b∈Su,s I((b, i) ∈ Tu ) ∈ {0, 1}, such that I((b, i) ∈ Tu ) is only active for at most one value of b ∈ Su,s . Var[Xi | i ∈ Su,s+1 , Fu,s ] ≤ E[Xi2 | i ∈ Su,s+1 , Fu,s ]   X = E I((b, i) ∈ Tu )Nu,s (b)2 Z(b, i)2 qk (αi )2 | i ∈ Su,s+1 , Fu,s  b∈Su,s X  1  = Nu,s (b)2 E[Z(b, i)2 qk (αi )2 | Fu,s ]. |Su,s | b∈Su,s Because Z(b, j) ∈ [0, 1],R it follows that Z(b, j)2 ≤ 1. By assumption on f and qk (·), E[qk (αi )2 ] = X1 qk (y)2 dP1 (y) = 1. Therefore, Var[Xi | i ∈ Su,s+1 , Fu,s ] X  1  ≤ Nu,s (b)2 E[qk (αi )2 ] |Su,s | b∈Su,s X  1  Nu,s (b)2 . ≤ |Su,s | b∈Su,s By construction, Nu,s (b) ∈ [0, 1]. Therefore, Nu,s (b)2 ∈ [0, Nu,s (b)] and by 52 BORGS-CHAYES-LEE-SHAH definition Ñu,s = [Nu,s (b)/|Su,s |]. Subsequently X  1  (A.14) Nu,s (b) = kÑu,s k1 . Var[Xi | i ∈ Su,s+1 , Fu,s ] ≤ |Su,s | b∈Su,s P  By definition eTk QÑu,s+1 = X /|Su,s+1 |. Also we have that i i∈Su,s+1 |Xi | ≤ Bq . Therefore, by applying Bernstein’s inequality and using (A.13), (A.14), it follows that   P eTk QÑu,s+1 − eTk ΛQÑu,s > t Fu,s , Su,s+1   X =P | (Xi − E[Xi ])| > t|Su,s+1 | Fu,s , Su,s+1 ! 1 2 2 2 |Su,s+1 | t ≤ 2 exp − |Su,s+1 |kÑu,s k1 + 13 Bq t|Su,s+1 | ! 3|Su,s+1 |t2 . = 2 exp − 6kÑu,s k1 + 2Bq t Let us define ( (A.15) A2u,h,k := ek QÑu,h − eTk ΛQÑu,h−1 < (c1 pn) − 12 +θ  |λd |2 2|λ1 | h−1 ) . Lemma A.4. For any u ∈ [n], k ∈ [d], and r ∈ Z+ , conditioned on ∩rh=1 A2u,h,k , for any r0 ∈ Z+ and ∆ ∈ [r0 ] such that r0 + ∆ ≤ r, |eTk QÑu,r0 +∆ − eTk Λ∆ QÑu,r0 | ≤ 2|λk |∆−1 1 (c1 pn) 2 −θ  |λd |2 2|λ1 | r0 . Proof. Recall the definition of A2u,h,k from (A.15). We can write the telescoping sum eTk QÑu,r0 +∆ − eTk Λ∆ QÑur0 = = ∆  X h=1 ∆ X h=1 eTk Λ∆−h QÑu,r0 +h − eTk Λ∆−h+1 QÑu,r0 +h−1   T T λ∆−h e Q Ñ − e ΛQ Ñ . u,r +h u,r +h−1 k k 0 0 k  ITERATIVE COLLABORATIVE FILTERING 53 0 +∆ Conditioned on ∩rh=1 A2u,h,k , |eTk QÑu,r0 +∆ − eTk Λ∆ QÑu,r0 |  r +h−1 ∆ X |λd |2 0 |λk |∆−h ≤ 1 −θ 2|λ1 | 2 h=1 (c1 pn)  r ∆−1  h |λd |2 0 X |λk |∆−1 |λd |2 = 1 2|λ1 ||λk | (c1 pn) 2 −θ 2|λ1 | h=0 Recall that we ordered the eigenvalues such that |λd |2 ≤ |λ1 ||λk | which implies  r ∆−1 |λk |∆−1 |λd |2 0 X −h T T ∆ |ek QÑu,r0 +∆ − ek Λ QÑu,r0 | ≤ 2 (c1 pn)1/4 2|λ1 | h=0  r 2|λk |∆−1 |λd |2 0 ≤ . 1 (c1 pn) 2 −θ 2|λ1 | 1 Lemma A.5. Assume that |λk | = ω((c1 pn)− 2 +θ ). For any u, v ∈ [n], k ∈ [d], and r ∈ Z+ , conditioned on ∩rh=1 (A2u,h,k ∩ A2v,h,k ) for any r0 ∈ Z+ and ∆1 , ∆2 ∈ [r0 ] such that r0 + ∆1 ≤ r and r0 + ∆2 ≤ r, (eTk QÑu,r0 +∆1 )(eTk QÑv,r0 +∆2 ) − (eTk Λ∆1 QÑu,r0 )(eTk Λ∆2 QÑv,r0 ) !  r |λd |2 |λk | 0 4Bq |λk |∆1 +∆2 −1 = (1 + o(1)) 1 2|λ1 | (c1 pn) 2 −θ This bound also applies if u = v. Proof. Recall the definition of A2u,h,k from (A.15). Using the fact that 2(xy − ab) = (x − a)(y + b) + (x + a)(y − b), it follows that (eTk QÑu,r0 +∆1 )(eTk QÑv,r0 +∆2 ) − (eTk Λ∆1 QÑu,r0 )(eTk Λ∆2 QÑv,r0 ) 1 = ((eTk QÑu,r0 +∆1 ) − (eTk Λ∆1 QÑu,r0 ))((eTk QÑv,r0 +∆2 ) + (eTk Λ∆2 QÑv,r0 )) 2 (A.16) 1 + ((eTk QÑu,r0 +∆1 ) + (eTk Λ∆1 QÑu,r0 ))((eTk QÑv,r0 +∆2 ) − (eTk Λ∆2 QÑv,r0 )). 2 By rearranging, it follows that (eTk QÑv,r0 +∆2 ) + (eTk Λ∆2 QÑv,r0 ) = (eTk QÑv,r0 +∆2 ) − (eTk Λr0 +∆2 Qev ) + (eTk Λ∆2 QÑv,r0 ) (A.17) − (eTk Λr0 +∆2 Qev ) + 2(eTk Λr0 +∆2 Qev ). 54 BORGS-CHAYES-LEE-SHAH By substituting (A.17) into (A.16), and then applying Lemma A.4, it follows 0 +∆2 0 +∆1 A2v,h,k , A2u,h,k ∩rh=1 that conditioned on ∩rh=1 (eTk QÑu,r0 +∆1 )(eTk QÑv,r0 +∆2 ) − (eTk Λ∆1 QÑu,r0 )(eTk Λ∆2 QÑv,r0 ) ! ! r 1 2|λk |∆1 −1 |λd |2 0 4|λk |r0 +∆2 −1 T r0 +∆2 ≤ + 2 ek Λ Qev 1 2 (c1 pn) 12 −θ 2|λ1 | (c1 pn) 2 −θ ! ! r ∆2 −1 2|λ | |λd |2 0 1 4|λk |r0 +∆1 −1 k T r0 +∆1 Qeu . + 2 ek Λ + 1 1 2 2|λ1 | (c1 pn) 2 −θ (c1 pn) 2 −θ By the assumption that |qk (·)| ≤ Bq , eTk Λr0 +∆2 Qev = |λrk0 +∆2 qk (αv )| ≤ Bq |λk |r0 +∆2 . Therefore, it follows that (eTk QÑu,r0 +∆1 )(eTk QÑv,r0 +∆2 ) − (eTk Λ∆1 QÑu,r0 )(eTk Λ∆2 QÑv,r0 ) !  4|λ |∆1 +∆2 −2  |λ |2 |λ | r0  1 k d k ≤ 2(c1 pn)− 2 +θ + |λk |Bq 1 −θ 2|λ1 | (c1 pn) 2 !   r |λd |2 |λk | 0 4Bq |λk |∆1 +∆2 −1 = (1 + o(1)) , 1 2|λ1 | (c1 pn) 2 −θ 1 where the last step follows from |λk | = ω((c1 pn)− 2 +θ ). 1 Lemma A.6. Assume that |λd | = ω((c1 pn)− 2 +θ ). For any u, v ∈ [n] and r ∈ Z+ , conditioned on ∩dk=1 ∩rh=1 (A2u,h,k ∩ A2v,h,k ), for any r1 , r2 ∈ [r] and r0 ∈ Z+ such that r0 ≤ min(r1 , r2 ), T T |Ñu,r F Ñv,r2 − Ñu,r QT Λr1 +r2 −2r0 +1 QÑv,r0 | 1 0 !  r |λd |2 0 4Bq d|λ1 |r1 +r2 −2r0 ≤ (1 + o(1)) 1 2 (c1 pn) 2 −θ Proof. Recall the definition of A2u,h,k from (A.15). By definition, F = QT ΛQ, such that T T |Ñu,r F Ñv,r2 − Ñu,r QT Λr1 +r2 −2r0 +1 QÑv,r0 | 1 0 T T = |Ñu,r QT ΛQÑv,r2 − Ñu,r QT Λr1 +r2 −2r0 +1 QÑv,r0 | 1 0   X =| λk (eTk QÑu,r1 )(eTk QÑv,r2 ) − (eTk Λr1 −r0 QÑu,r0 )(eTk Λr2 −r0 QÑv,r0 ) |. k ITERATIVE COLLABORATIVE FILTERING 55 By Lemma A.5, conditioned on ∩dk=1 ∩rh=1 (A2u,h,k ∩ A2v,h,k ), if we substitute r1 − r0 for ∆1 , and r2 − r0 for ∆2 , and since |λk | ≤ |λ1 | for all k, it follows that T T QT Λr1 +r2 −2r0 +1 QÑv,r0 | |Ñu,r F Ñv,r2 − Ñur 0 1 !  r X |λd |2 |λk | 0 4Bq |λk |r1 +r2 −2r0 −1 ≤ |λk | (1 + o(1)) 1 2|λ1 | (c1 pn) 2 −θ k !  r |λd |2 0 4Bq d|λ1 |r1 +r2 −2r0 ≤ (1 + o(1)) . 1 2 (c1 pn) 2 −θ Next we prove concentration of kNu,s+1 k1 conditioned on Fus and Su,s+1 using similar arguments as the proof of Lemma A.3. Lemma A.7.  P For any s ∈ Z+ , kÑu,s+1 k1 − ρT ΛQÑu,s > t Fu,s , Su,s+1  3|Su,s+1 |t2 ≤ 2 exp − 6kÑu,s k1 + 2t ! where recall (cf. (2.2)) that R ρ = [ρk ]k∈[d] denotes the d dimensional vector such that ρk = E[qk (α)] = X1 qk (y)dP1 (y). Proof. Since Nu,s+1 is a nonnegative vector, X kNu,s+1 k1 = Nu,s+1 (i) i = X X I((b, i) ∈ Tu )Nu,s (b)M1 (b, i). i∈Su,s+1 b∈Su,s Due to the construction of Tu as a BFS tree, for all i, X I((b, i) ∈ Tu ) = I(i ∈ Su,s+1 ) ∈ {0, 1}. b∈Su,s If there were more than one possible parent for i in the graph defined by edge set E1 , Tu is constructed by simply choosing one uniformly at random. Therefore, conditioned on I(i ∈ Su,s+1 ), the parent of i is equally likely to have been any of b ∈ Su,s , such that (A.18) E [I((b, i) ∈ Tu ) | I(i ∈ Su,s+1 ), Fu,s ] = 1 . |Su,s | , 56 BORGS-CHAYES-LEE-SHAH For (b, i) ∈ Tu , by construction, M1 (b, i) = M (b, i) = Z(b, i). Recall that Z(b, i) ∈ [0, 1] is independent from the tree structure (the sampling of edges E), and its expectationPis a function of αb and αi . Let us define Xi = b∈Su,s I((b, i) ∈ Tu )Nu,s (b)Z(b, i), such that X kNu,s+1 k1 = Xi . i∈Su,s+1 By the model assumptions that Nu,s (b) ∈ [0, 1] and Z(b, i) ∈ [0, 1], it follows that |Xi | ≤ 1. If i ∈ / Su,s+1 , then Xi = 0 with probability 1. We next compute the expectation and variance of Xi and then proceed to apply Bernstein’s inequality to bound kNu,s+1 k1 . Recall that Nu,s and Su,s are fixed conditioned on Fu,s . Conditioned on Fu,s and Su,s+1 , Xi are identically distributed for i ∈ Su,s+1 . We can verify that for i 6= j, Xi is independent from Xj conditioned on Fus and Su,s+1 because αi is independent from αj , and the variables I((b, i) ∈ Tu ) are only correlated for the same i but not correlated across different potential children vertices, i.e. the tree constraint only enforces that each child has exactly one parent, but allows parents to have arbitrarily many children. Therefore kNu,s+1 k1 is a sum of iid random variables. We use the independence of sampled edge set E1 from the latent variables Θ, along with (A.18) and the law of iterated expectations to show that X  1  E[Xi | i ∈ Su,s+1 , Fu,s ] = Nu,s (b)E[Z(b, i) | Fu,s ]. |Su,s | b∈Su,s For b ∈ Su,s , αb is fixed when conditioned on Fu,s , such that E[Z(b, i) | αi , Fu,s ] = f (αb , αi ). Additionally, by the spectral decomposition of f , X  1  E[Xi | i ∈ Su,s+1 , Fu,s ] = Nu,s (b)E[f (αb , αi ) | Fu,s ] |Su,s | b∈Su,s X  1  X = Nu,s (b) λk qk (αb )E[qk (αi )] |Su,s | b∈Sus T k = ρ ΛQÑu,s . Using similar arguments, P we can also compute a bound on the variance of Xi , using the fact that b∈Su,s I((b, i) ∈ Tu ) ∈ {0, 1}, such that I((b, i) ∈ Tu ) 57 ITERATIVE COLLABORATIVE FILTERING is only active for at most one value of b ∈ Su,s . Var[Xi | i ∈ Su,s+1 , Fu,s ] ≤ E[Xi2 | i ∈ Su,s+1 , Fu,s ]  X I((b, i) ∈ Tu )Nu,s (b)2 Z(b, i)2 = E  i ∈ Su,s+1 , Fu,s  b∈Su,s X  1    = Nu,s (b)2 E Z(b, j)2 | Fu,s |Su,s | b∈Su,s Because Z(b, j) ∈ [0, 1] and Nu,s (b) ∈ [0, 1],   X   X 1 1 Var[Xi | i ∈ Su,s+1 , Fu,s ] ≤ Nu,s (b)2 ≤ Nu,s (b) = kÑu,s k1 . |Su,s | |Su,s | b∈Su,s b∈Su,s By an application of Bernstein’s inequality, !   2 3|S |t u,s+1 P kÑu,s+1 k1 − ρT ΛQÑu,s > t Fu,s , Su,s+1 ≤ 2 exp − 6kÑu,s k1 + 2t Let us define ( (A.19) A3u,h := − 12 +θ kÑu,h k1 − ρT ΛQÑu,h−1 < (c1 pn)  |λd |2 2|λ1 | h−1 ) . Lemma A.8. Assuming c1 pn = ω(1), for any u ∈ [n] and s ∈ Z+ , conditioned on ∩sh=1 (A3u,h ∩dk=1 A2u,h,k ), it holds that kÑu,s k1 ≤ Bq2 d|λ1 |s (1 + o(1)). Proof. Recall the definition of A2u,h,k from (A.15), and recall the definition of A3u,h from (A.19). We first rearrange the expression kÑu,s k1 into the sum of three expressions, kÑu,s k1 = kÑu,s k1 − ρT ΛQÑu,s−1 + ρT ΛQÑu,s−1 − ρT Λs Qeu + ρT Λs Qeu ≤ 1T Ñu,s − ρT ΛQÑu,s−1 + ρT ΛQÑu,s−1 − ρT Λs Qeu + ρT Λs Qeu = 1T Ñu,s − ρT ΛQÑu,s−1 + X k + X k ρk λsk qk (αu )   ρk λk eTk QÑu,s−1 − eTk Λs−1 Qeu 58 BORGS-CHAYES-LEE-SHAH The first expression is upper bounded conditioned on A3u,s . The second expression is upper bounded by Lemma A.4 conditioned on ∩sh=1 ∩dk=1 A2u,h,k , choosing r0 = 0 and ∆ = s − 1. The third expression is bounded by the assumption that |qk (·)| ≤ Bq , |ρk | = |E[qk (α)]| ≤ Bq , and |λk | ≤ |λ1 |. Therefore it follows that s−1  X 2|λk |s−1 |λd |2 − 21 +θ kÑu,s k1 ≤ (c1 pn) + Bq + Bq2 d|λ1 |s 1 −θ 2|λ1 | 2 (c pn) 1 k s−1    2 1 1 |λd | + Bq d|λ1 |s 2|λ1 |−1 (c1 pn)− 2 +θ + Bq ≤ (c1 pn)− 2 +θ 2|λ1 | = Bq2 d|λ1 |s (1 + o(1)). A.3. Showing that Majority of Local P Neighborhoods are Good. For each a ∈ [n], show concentration of b∈[n]\b I((a, b) ∈ E1 ). Let’s define event    X  (A.20) A4 := ∩a∈[n] I((a, b) ∈ E1 ) < c1 (p + (n − 1)−3/4 )(n − 1) .   b∈[n]\a Lemma A.9.
c1 (n − 1)1/4 P ¬A4 ≤ n exp − 3 ! . Proof. We can show this easily because this is just a sum of Bernoulli random variables. For a fixed a,   X E I((a, b) ∈ E1 ) = c1 p(n − 1) b∈[n]\a By Chernoff’s bound,   X P I((a, b) ∈ E1 ∪ E2 ) > (1 + p−1 (n − 1)−3/4 )c1 p(n − 1) b∈[n]\a c1 (n − 1)1/4 ≤ exp − 3 ! . We use union bound to get the final expression. ITERATIVE COLLABORATIVE FILTERING 59 Let us define A5u,s−1 := ∩sh=1 (A1u,h ∩ A3u,h ∩dk=1 A2u,h,k ). (A.21) We bound the probability of events A1u,s , A2u,s,k , and A3u,s conditioned on Fu,s−1 , A5u,s−1 . Lemma A.10. which satisfies Assuming that c1 pn = ω(1), and for any u ∈ [n] and s  9c1 pn 8 s ≤ (c1 p)−7/8 , for sufficiently large n, P(¬A1u,s |Fu,s−1 , A5u,s−1 )   (c1 pn)2θ ≤ 4 exp − 2−s . 4Bq2 Proof. Recall the definition of event A1u,h and A5u,s−1 from (A.3) and 1 (A.21). The event A5u,s−1 contains the event ∩s−1 h=1 Au,h , such that by Lemma 5 A.2 and the assumption on s, conditioned on Au,s−1 , we can verify that
− 1 +θ |Su,s | < (c1 p)−7/8 . Choosing t = c1 pnBq4 2 = o(1), we can verify that t < 1 − c1 p|Su,s | = 1 − o(1). By applying Lemma A.1 for this choice of t, it follows that   c |Bu,s−1 ||Su,s−1 |c1 p 1 5 (A.22) . P(¬Au,s |Fu,s−1 , Au,s−1 ) ≤ 2 exp − 3(c1 pn)1−2θ Bq2 c | ≥ n(1 − o(1)), such that for By (A.8) from the proof of Lemma A.2, |Bu,s−1 sufficiently large n,    3 7c1 pn s−1 c |Bu,s−1 ||Su,s−1 | ≥ n . 4 8 By plugging into (A.22) and using the inequality that for s ≥ 1, xs−1 ≥ 1 + (x − 1)(s − 1) for x ≥ 1, it follows that P(¬A1u,s |Fu,s−1 , A5u,s−1 )  !  (c1 pn)2θ 7c1 pn s−1 ≤ 2 exp − 4Bq2 8      (c1 pn)2θ 7c1 pn ≤ 2 exp − 1+ − 1 (s − 1) 4Bq2 8       (c1 pn)2θ (c1 pn)2θ 7c1 pn = 2 exp − exp − − 1 (s − 1) . 4Bq2 4Bq2 d 8 60 BORGS-CHAYES-LEE-SHAH In above, we used the fact that 7c1 pn/8 > 1 for n large enough since c1 pn = ω(1). For sufficiently large n,    (c1 pn)2θ 7c1 pn 1 exp − −1 ≤ , 2 4Bq 8 2 such that (A.23) P(¬A1u,s |Fu,s−1 , A5u,s−1 ) (A.24)   (c1 pn)2θ ≤ 2 exp − 2−(s−1) 4Bq2   (c1 pn)2θ ≤ 4 exp − 2−s . 4Bq2 1 Lemma A.11. Assuming that c1 pn = ω(1), |λd | = ω((c1 pn)− 4 ), for any u ∈ [n] and s ∈ Z+ which satisfies   9c1 pn s ≤ (c1 p)−7/8 , 8 conditioned on A5u,s−1 , for sufficiently large n,   (c1 pn)2θ P(¬A2u,s,k |Fu,s−1 , A5u,s−1 ) ≤ 4 exp − 2−s , 3Bq2 d and P(¬A3u,s |Fu,s−1 , A5u,s−1 )   (c1 pn)2θ ≤ 4 exp − 2−s . 3Bq2 d Proof. Recall the definitions of events A1u,h , A3u,h , A2u,h,k , and A5u,s−1  2 s−1 1 d| from (A.3), (A.19), (A.15), and (A.21). By plugging in t = (c1 pn)− 2 +θ |λ 2|λ1 | to Lemmas A.3 and A.7, it follows that ! 2 3|S |t u,s P(¬A2u,s,k |Fu,s−1 , A5u,s−1 ) ≤ 2 exp − 6kÑu,s−1 k1 + 2Bq t and P(¬A3u,s |Fu,s−1 , A5u,s−1 ) 3|Su,s |t2 ≤ 2 exp − 6kÑu,s−1 k1 + 2t ! 61 ITERATIVE COLLABORATIVE FILTERING We lower bound the expression in the exponent of the first bound using Lemmas A.2 and A.8. For sufficiently large n, 3 |t2  7c1 pn 8 s (c1 pn)−1+2θ  |λd |2 2|λ1 | 2(s−1) 3|Su,s ≥  2 s−1 1 6kÑu,s−1 k1 + 2Bq t d| 6Bq2 d|λ1 |s−1 (1 + o(1)) + 2Bq (c1 pn)− 2 +θ |λ 2|λ1 |     7c1 pn  s−1 −1+2θ 3 (c1 pn) 8   7c1 pn|λd |4 ≥ . s−1   1 32|λ1 |3 |λd |2 6Bq2 d(1 + o(1)) + 2Bq (c1 pn)− 2 +θ 2|λ 2 1| 1 Assuming that |λd | = ω((c1 pn)− 4 ), then Using the fact that it follows that x(s−1) 3|Su,s |t2 ≥ 6kÑu,s−1 k1 + 4Bq t  7c1 pn|λd |4 32|λ1 |3 > 1 for n large enough. ≥ 1 + (x − 1)(s − 1) for x, s ≥ 1, and |λd | ≤ |λ1 |, 7(c1 pn)2θ 16Bq2 d(1 + o(1))     7c1 pn|λd |4 1+ − 1 (s − 1) . 32|λ1 |3 Therefore, for sufficiently large n,       (c1 pn)2θ (c1 pn)2θ 7c1 pn|λd |4 P(¬A2u,s,k |Fu,s−1 , A5u,s−1 ) ≤ 2 exp − exp − − 1 (s − 1) . 3Bq2 d 3Bq2 d 32|λ1 |3 Further, because 7c1 pn|λd |4 32|λ1 |3 > 1, for n large enough    1 (c1 pn)2θ 7c1 pn|λd |4 −1 ≤ , exp − 2 3 3Bq d 32|λ1 | 2 such that P(¬A2u,s,k |Fu,s−1 , A5u,s−1 )   (c1 pn)2θ 2−(s−1) ≤ 2 exp − 3Bq2 d   (c1 pn)2θ = 4 exp − 2−s . 3Bq2 d Using similar arguments, for sufficiently large n, it follows that   (c1 pn)2θ 3 5 P(¬Au,s |Fu,s−1 , Au,s−1 ) ≤ 4 exp − 2−s . 3Bq2 d 62 BORGS-CHAYES-LEE-SHAH We consider a vertex u to have “good” neighborhood behavior if event A5u,r holds. Next we show a lower bound on the probability that the neighborhood of a vertex behaves well. 1 Lemma A.12. Assuming that c1 pn = ω(1) and |λd | = ω((c1 pn)− 4 ), for any r ∈ Z+ which satisfies   9c1 pn r ≤ (c1 p)−7/8 , 8 for sufficiently large n, P ¬A5u,r
  (c1 pn)2θ . ≤ 4(d + 2) exp − 4Bq2 d Proof. Recall the definition of event A5u,s−1 from (A.21).  
P ¬A5u,r = P ∪sh=1 ¬(A1u,h ∩ A3u,h ∩dk=1 A2u,h,k ) For a sequence of events {Zh }h∈[s] , P(∪h∈[s] ¬Zh ) = P({¬Zs ∩h∈[s−1] Zh } ∪ {∪h∈[s−1] ¬Zh }) = P(¬Zs | ∩h∈[s−1] Zh )P(∩h∈[s−1] Zh ) + P(∪h∈[s−1] ¬Zh ) ≤ P(¬Zs | ∩h∈[s−1] Zh ) + P(∪h∈[s−1] ¬Zh ). We can repeated apply this bound to show that X P(∪h∈[s] ¬Zh ) ≤ P(¬Zt | ∩h∈[t−1] Zh ). t∈[s] By using this inequality for our sequence of events as described in A5u,s−1 and applying union bound, it follows that
P ¬A5u,r r   X ≤ P ¬(A1u,h ∩ A3u,h ∩dk=1 A2u,h,k ) | A5u,s−1 ≤ s=1 r h X P(¬A1u,s |Fu,s−1 , A5u,s−1 ) + s=1 d X k=1 + P(¬A3u,s |Fu,s−1 , A5u,s−1 ) i .
P(¬A2u,s,k |Fu,s−1 , A5u,s−1 ) 63 ITERATIVE COLLABORATIVE FILTERING By Lemmas A.10 and A.11, for large enough n,
P ¬A5u,r ≤ r  X 4 exp  pn)2θ − (c14B 2 q  2 −s s=1 + d X   2θ 1 pn) 2−s 4 exp − (c3B 2d q k=1   2θ  1 pn) + 4 exp − (c3B 2−s 2 qd     (c1 pn)2θ (c1 pn)2θ + 4(d + 1) exp − ≤ 4 exp − 4Bq2 3Bq2 d   (c1 pn)2θ ≤ 4(d + 2) exp − . 4Bq2 d Next we show that with high probability a large fraction of the vertices have good local neighborhood behavior. Let us define       X 2θ (c1 pn) (A.25) A6u,v,r := . I(A5a,r ) ≥ (n − 2) 1 − exp −   8Bq2 d a∈[n]\u,v 1 Lemma A.13. Assuming that c1 pn = ω(1) and |λd | = ω((c1 pn)− 4 ), for any r ∈ Z+ which satisfies   9c1 pn r ≤ (c1 p)−7/8 , 8 for sufficiently large n, P ¬A6u,v,r
  (c1 pn)2θ . ≤ 4(d + 2) exp − 8Bq2 d Proof. Recall the definitions of events A5u,s−1 and A6u,v,r from (A.21) and (A.25). We use Lemma A.12 to upper bound the probability of event ¬A5u,s−1 and then plug it into Markov’s inequality to show that     2θ X (c pn) 1  P(¬A6u,v,r ) = P  I(¬A5a,r ) ≥ (n − 2) exp − 8Bq2 d a∈[n]\u,v P   5 (c1 pn)2θ a∈[n]\u,v E[I(¬Aa,r )] ≤ exp (n − 2) 8Bq2 d   (c1 pn)2θ ≤ 4(d + 2) exp − 8Bq2 d 64 BORGS-CHAYES-LEE-SHAH A.4. Concentration of Edges Between Local Neighborhoods. Suppose that we condition on all the latent variables {αi }i∈[n] and edge set E1 and associated weights in M1 such that vectors Nu,r1 and Nv,r2 are deterT M Ñ mined. The remaining randomness in computing Ñu,r 2 v,r2 comes from 1 the sampling of edges E2 and associated weights M2 , which are sampled independently with probability close to c2 p. Furthermore, we use the bounded assumption that Z(i, j) ∈ [0, 1] and f (αi , αj ) ∈ [0, 1] to apply concentration inequalities on the edge weights. Lemma A.14. For any u, v ∈ [n] and r1 , r2 ∈ Z+ , T T P Nu,r M2 Nv,r2 − EE2 [Nu,r M2 Nv,r2 | FE1 ] > t | FE1 1 1   3t2 (1 − c1 p) ≤ 2 exp − . T FN 12c2 pNur vr2 + 2(1 − c1 p)t 1
Proof. Recall that M2 is symmetric, and for some edge (i, j) ∈ E2 , M2 (i, j) = M (i, j) = Z(i, j), but for (i, j) ∈ / E2 , M2 (i, j) = 0. Therefore we T M N can rewrite the expression Nu,r as 2 v,r2 1 T Nu,r M2 Nv,r2 = 1 X Nu,r1 (i)Nv,r2 (j)M2 (i, j) i,j∈[n] = X Nu,r1 (i)Nv,r2 (j)I((i, j) ∈ E2 )Z(i, j) i,j∈[n] = X (Nu,r1 (i)Nv,r2 (j) + Nu,r1 (j)Nv,r2 (i))I((i, j) ∈ E2 )Z(i, j). i<j,(i,j)∈E / 1 The last step comes from the fact that our edges and associated weights are undirected, such that Z(i, j) = Z(j, i), and that we do not sample self-edges, such that (i, i) ∈ / E2 , leading to the strict inequality that i < j. Conditioned on FE1 , Nu,r1 and Nv,r2 are fixed. We have presented the expression as a sum of independent random variables conditioned on FE1 ,
n since Z(i, j) are independent for the distinct vertex pairs. The sum is 2
n over 2 − |E1 | terms, which are each bounded by 2, since Z(i, j) ∈ [0, 1], and the entries in vectors Nu,r1 and Nv,r2 are also in [0, 1]. By independence, it 65 ITERATIVE COLLABORATIVE FILTERING follows that   T Var Nu,r M2 Nv,r2 | FE1 1 X (Nu,r1 (i)Nv,r2 (j) + Nu,r1 (j)Nv,r2 (i))2 Var [I((i, j) ∈ E2 )Z(i, j) | FE1 ] = i<j,(i,j)∈E / 1 X ≤  (Nu,r1 (i)Nv,r2 (j) + Nu,r1 (j)Nv,r2 (i))2 E I((i, j) ∈ E2 )Z(i, j)2  F E1 . i<j,(i,j)∈E / 1 where the last line follows from the fact that Var[X] ≤ E[X 2 ]. We assumed that Z(i, j) is independent from I((i, j) ∈ E2 ). Conditioned on (i, j) ∈ / E1 , which occurs with probability 1 − c1 p, I((i, j) ∈ E2 ) is distributed according to a Bernoulli(c2 p/(1−c1 p)) random variable. By assumption, Z(i, j) ∈ [0, 1] such that E[Z(i, j)2 |αi , αj ] ≤ E[Z(i, j)|αi , αj ] = f (αi , αj ). Therefore,  T  Var Nu,r M2 Nv,r2 | FE1 1 X c2 p
≤ (Nu,r1 (i)Nv,r2 (j) + Nu,r1 (j)Nv,r2 (i))2 1−c f (αi , αj ). 1p i<j,(i,j)∈E / 1 By the fact that (a + b)2 ≤ 2a2 + 2b2 ,  T  Var Nu,r M2 Nv,r2 | FE1 1 X ≤ (2Nu,r1 (i)2 Nv,r2 (j)2 + 2Nu,r1 (j)2 Nv,r2 (i)2 ) c2 p
1−c1 p f (αi , αj ) i<j,(i,j)∈E / 1 = 2c2 p
1−c1 p X Nu,r1 (i)2 Nv,r2 (j)2 f (αi , αj ). i6=j,(i,j)∈E / 1 Since f (αi , αj ) ∈ [0, 1], and the entries of Nu,r1 and Nv,r2 are in [0, 1], it follows that X  T  2c2 p
Nu,r1 (i)Nv,r2 (j)f (αi , αj ) Var Nu,r M N | F ≤ 2 v,r E 2 1 1−c1 p 1 i6=j,(i,j)∈E / 1 ≤ 2c2 p 1−c1 p
X Nu,r1 (i)Nv,r2 (j)f (αi , αj ) (i,j)∈E / 1 = 2c2 p N T F Nv,r2 . 1 − c1 p u,r1 By Bernstein’s inequality,  T 
T M2 Nv,r2 − EE2 Nu,r M2 Nv,r2 FE1 > t | FE1 P Nu,r 1 1   3t2 (1 − c1 p) ≤ 2 exp − . T FN 12c2 pNu,r v,r2 + 2(1 − c1 p)t 1 66 BORGS-CHAYES-LEE-SHAH Let us define the following event n  T  T M2 Nv,r2 FE1 A7u,v,r1 ,r2 := Nu,r M2 Nv,r2 − EE2 Nu,r 1 1 1/2 o  (A.26) ≤ c2 p|λ1 |r1 +r2 |Su,r1 ||Sv,r2 |p−1/4 Lemma A.15. which satisfiy Assuming that c1 pn = ω(1), for any u, v ∈ [n] and r1 < r2  (A.27) and  7|λd |2 c1 pn 8|λ1 | 9c1 pn 8 r1  ≤ (r1 +r2 )/2 9c1 pn 8 r2 ≥ p−6/8 , ≤ (c1 p)−7/8 , conditioned on A4 ∩ A5u,r1 ∩ A5v,r2 ∩ A7u,v,r1 ,r2 , for sufficiently large n, 1−c1 p T c2 p Ñu,r1 M2 Ñv,r2 ≤ (1 − T − Ñu,r F Ñv,r2 1 −1/2 c1 p)c2 |λd |r1 +r2 p1/8 + (c1 pn + c1 n 1/4 )Bq2 d|λ1 |r1  8 7c1 pn r2 (1 + o(1)). Proof. Recall the definitions of events A4 , A5u,r1 , ∩A5v,r2 , and A7u,v,r1 ,r2 as given in (A.20), (A.21), and (A.26). Let us bound h i 1−c1 p T T FE1 − Nu,r F Nv,r2 . c2 p EE2 Ñu,r1 M2 Nv,r2 1 Recall that M2 is symmetric, and for some edge (i, j) ∈ E2 , M2 (i, j) = M (i, j) = Z(i, j), but for (i, j) ∈ / E2 , M2 (i, j) = 0. Since we do not sample T M N self-edges, (i, i) ∈ / E2 . Therefore we can rewrite the expression Nu,r 2 v,r2 1 as X T M N Nu,r = Nu,r1 (i)Nv,r2 (j)M2 (i, j) 2 v,r 2 1 i,j∈[n] = X Nu,r1 (i)Nv,r2 (j)I((i, j) ∈ E2 )Z(i, j). i6=j,(i,j)∈E / 1 Conditioned on (i, j) ∈ / E1 , which occurs with probability 1 − c1 p, I((i, j) ∈ E2 ) is distributed according to a Bernoulli(c2 p/(1 − c1 p)) random variable. Conditioned on FE1 , Nu,r1 and Nv,r2 are fixed, and since we assumed that 67 ITERATIVE COLLABORATIVE FILTERING Z(i, j) is independent from I((i, j) ∈ E2 ) and E[Z(i, j)|αi , αj ] = f (αi , αj ), it follows that   X c2 p T Nu,r1 (i)Nv,r2 (j)f (αi , αj ) EE2 [Nu,r1 M2 Nv,r2 | FE1 ] = (1 − c1 p) i6=j,(i,j)∈E / 1 T T M N Therefore the difference between EE2 [Nu,r 2 v,r2 | FE1 ] and Nu,r1 F Nv,r2 1 consists of the terms associated to edges in E1 and self edges,  T  T 1p Nu,r F Nv,r2 − 1−c FE1 c2 p EE2 Nu,r1 M2 Nv,r2 1 X X Nu,r1 (i)Nv,r2 (i)f (αi , αi ). Nu,r1 (i)Nv,r2 (j)f (αi , αj ) + = i∈[n] (i,j)∈E1 By the fact that f (αi , αj ) ∈ [0, 1], and the entries in vectors Nu,r1 and Nv,r2 are in [0, 1], it follows that  T  1−c1 p T Nu,r F N − E N M N F v,r 2 v,r E E u,r 2 2 2 1 c2 p 1 1 X X X ≤ Nu,r1 (i) I((i, j) ∈ E1 ) + Nu,r1 (i). i∈[n] j∈[n] Conditioned on A4 , for all i, (n − 1)−3/4 )(n − 1), such that i∈[n] P T Nu,r F Nv,r2 − 1 j∈[n] I((i, j) 1−c1 p c2 p EE2  ∈ E1 ) is bounded above by c1 (p + T Nu,r M2 Nv,r2 1 FE1  ≤ (c1 (p + (n − 1)−3/4 )(n − 1) + 1)kNu,r1 k1 . Conditioned on A5u,r1 , by Lemma A.8 we can upper bound kNu,r1 k1 = kÑu,r1 k1 |Su,r1 | such that  T  1−c1 p T E N M N F Nu,r F N − 2 v,r v,r E E u,r 2 2 2 1 c2 p 1 1 ≤ (c1 (p + (n − 1)−3/4 )(n − 1) + 1)|Su,r1 |Bq2 d|λ1 |r1 (1 + o(1)) = (c1 pn + c1 n1/4 )|Su,r1 |Bq2 d|λ1 |r1 (1 + o(1)). By combining the above bound with the bound given when conditioned on A7u,v,r1 ,r2 , and then dividing both sides by |Su,r1 ||Sv,r2 |, it follows that 1−c1 p T c2 p Ñu,r1 M2 Ñv,r2 = |Su,r1 |−1 |Sv,r2 |−1 ≤ (1 − c1 p) T − Ñu,r F Ñv,r2 1 1−c1 p T c2 p Nu,r1 M2 Nv,r2 |λ1 |r1 +r2 p−1/4 c2 p|Su,r1 ||Sv,r2 | T − Nu,r F Nv,r2 1 !1/2 + (c1 pn + c1 n 1/4 ) Bq2 d|λ1 |r1 |Sv,r2 | ! (1 + o(1)) 68 BORGS-CHAYES-LEE-SHAH Conditioned on A5u,r1 ∩ A5v,r2 , we can lower bound |Su,r1 | and |Sv,r2 | with Lemma A.2 to show that 1−c1 p T c2 p Ñu,r1 M2 Ñv,r2 ≤ (1 − c1 p)  |λ1 |r1 T − Ñu,r F Ñv,r2 1   +r −1/4 1/2 2p 8 7c1 pn c2 p (r1 +r2 )/2  r2 + (c1 pn + c1 n1/4 )Bq2 d|λ1 |r1 7c18pn (1 + o(1)) (r1 +r2 )/2  −1/2 1| = (1 − c1 p)c2 p−5/8 |λd |r1 +r2 7|λ8|λ 2 d | c1 pn  r2 (1 + o(1)). + (c1 pn + c1 n1/4 )Bq2 d|λ1 |r1 7c18pn By assumption on r1 and r2 in (A.27), 1−c1 p T T c2 p Ñu,r1 M2 Ñv,r2 − Ñu,r1 F Ñv,r2 −1/2 ≤ (1 − c1 p)c2 p−5/8 |λd |r1 +r2 p6/8 + (c1 pn + c1 n −1/2 = (1 − c1 p)c2 1/4 )Bq2 d|λ1 |r1  8 7c1 pn r2 8 7c1 pn r2 (1 + o(1)) |λd |r1 +r2 p1/8 + (c1 pn + c1 n 1/4 )Bq2 d|λ1 |r1  (1 + o(1)). We can bound the probability of event A7u,v,r1 ,r2 . Lemma A.16. satisfies 1 Assume that |λd | = ω((c1 pn)− 2 +θ ). For r1 ≤ r2 which  9c1 pn 8 and  7λ2d c1 pn |λ1 |8 r2 ≤ (c1 p)−7/8 , (r1 +r2 )/2 ≥ p−6/8 , conditioned on A5u,r1 ∩ A5v,r2 , for sufficiently large n, P ¬A7u,v,r1 ,r2 | FE1 , A5u,r1 , A5v,r2
(1 − c1 p)p−1/4 ≤ 2 exp − 5|λ1 |Bq2 d ! , ITERATIVE COLLABORATIVE FILTERING 69 Proof. Recall the definition of event A7u,v,r1 ,r2 from (A.26). For a choice
1/2 , by Lemma A.14 of t = c2 pλr11 +r2 |Sur1 ||Svr2 |p−1/4
P ¬A7u,v,r1 ,r2 | FE1 , A5u,r1 , A5v,r2   r1 +r2 −1/4 |Su,r1 ||Sv,r2 |p 3(1 − c1 p)c2 pλ1 ≤ 2 exp −
 r1 +r2 −1/4 1/2 T FN |p ||S |S + 2(1 − c p) c pλ 12c2 pNu,r v,r2 u,r1 1 2 v,r2 1 1 ! 3(1 − c1 p)c2 λr11 +r2 p−1/4 = 2 exp − . r1 +r2 −5/4
1/2 T F Ñ p (|Sur1 ||Svr2 |)−1/2 12c2 Ñur vr2 + 2(1 − c1 p) c2 λ1 1 T F Ñ Conditioned on A5u,r1 and A5v,r2 , we can bound Ñur vr2 by Lemma A.6, 1 choosing r0 = 0, ! r1 +r2 4B d|λ | q 1 T Ñur F Ñvr2 ≤ eTu QT Λr1 +r2 +1 Qev + (1 + o(1)) 1 1 (c1 pn) 2 −θ ! 4B d q ≤ |λ1 |r1 +r2 |λ1 |Bq2 d + (1 + o(1)) 1 (c1 pn) 2 −θ = |λ1 |r1 +r2 +1 Bq2 d(1 + o(1)). Conditioned on A5u,r1 and A5v,r2 , we can bound |Su,r1 ||Sv,r2 | by Lemma A.2, such that for sufficiently large n,  1/2 2(1 − c1 p) c2 |λ1 |r1 +r2 p−5/4 (|Su,r1 ||Sv,r2 |)−1/2 −(r1 +r2 )/2  1/2 (r1 +r2 )/2 −5/8 7c1 pn ≤ 2(1 − c1 p)(c2 ) |λ1 | p 8 (r1 +r2 )/2  8|λ1 | 1/2 r1 +r2 −5/8 = 2(1 − c1 p)(c2 ) |λd | . p 7|λd |2 c1 pn Then we use the condition assumed on r1 and r2 to show that  1/2 2(1 − c1 p) c2 |λ1 |r1 +r2 p−5/4 (|Su,r1 ||Sv,r2 |)−1/2 ≤ 2(1 − c1 p)(c2 )1/2 |λd |r1 +r2 p−5/8 p6/8 = 2(1 − c1 p)(c2 )1/2 |λd |r1 +r2 p1/8 = o(|λ1 |r1 +r2 +1 Bq2 d), where the last step follows from p1/8 = o(1), |λ1 | and Bq are constants, and d = Ω(1). 70 BORGS-CHAYES-LEE-SHAH Therefore, P ¬A7u,v,r1 ,r2 | FE1 , A5u,r2 , A5v,r2
3(1 − c1 p)c2 λr11 +r2 p−1/4 ≤ 2 exp − 12c2 |λ1 |r1 +r2 +1 Bq2 d(1 + o(1)) ! (1 − c1 p)p−1/4 = 2 exp − . 4|λ1 |Bq2 d(1 + o(1)) ! For sufficiently large n, P ¬A7u,v,r1 ,r2 | FE1 , A5u,r2 , A5v,r2
(1 − c1 p)p−1/4 ≤ 2 exp − 5|λ1 |Bq2 d ! . Recall that our evaluation of distance estimates dist1 and dist2 involves expressions of the form
T
1−c1 p
Ñ − Ñ M Ñ − Ñ . u,r v,r 2 u,r+i v,r+i c2 p By expanding this expression into its four terms, we can show by Lemma A.15 that this expression approximates (Ñu,r − Ñv,r )T F (Ñu,r+i − Ñv,r+i ) when conditioned on events A4 ∩ A5u,r+i ∩ A5v,r+i ∩ A7u,v,r,r+i ∩ A7v,u,r,r+i ∩ A7u,u,r,r+i ∩ A7v,v,r,r+i . This will allow us to prove bound on dist1 (u, v) and dist2 (u, v). Let us define the following event (A.28) A8u,v,r1 ,r2 := {A7u,v,r1 ,r2 ∩ A7v,u,r1 ,r2 ∩ A7u,u,r1 ,r2 ∩ A7v,v,r1 ,r2 } 1 Lemma A.17. Assume that θ ∈ (0, 14 ), |λd | = ω((c1 pn)− 2 +θ ), c1 pn = ω(1), and p = o(n−1+1/(5+8θ) ). For some u, v ∈ [n] and r which satisfies  7|λd |2 c1 pn 8|λ1 | and  9c1 pn 8 r+ 12 r+1 ≥ p−6/8 , ≤ (c1 p)−7/8 . 71 ITERATIVE COLLABORATIVE FILTERING Conditioned on A4 ∩ A5u,r+1 ∩ A5v,r+1 ∩ A8u,v,r,r+1 , for sufficiently large n, kΛQ(eu − ev )k22 is lower bounded by |λ1 |−2r dist1 (u, v) − 32Bq d|λ1 | 1 (c1 pn) 2 −θ , and kΛQ(eu − ev )k22 is upper bounded by −2r |λd | dist1 (u, v) + 32Bq d|λ1 | 1 (c1 pn) 2 −θ  |λ1 | |λd | 2r . Proof. Recall that dist1 (u, v) = 1−c1 p c2 p  Ñu,r − Ñv,r T   M2 Ñu,r+1 − Ñv,r+1 . By Lemma A.15, we upper bound the difference between dist1 (u, v) and (A.29)  Ñu,r − Ñv,r T   F Ñu,r+1 − Ñv,r+1 . By Lemma A.6, plugging in r0 = 0, r1 = r, and r2 = r + 1, we can upper bound the difference between (A.29) and (eu − ev )T QT Λ2r+2 Q(eu − ev ). 72 BORGS-CHAYES-LEE-SHAH Therefore, dist1 (u, v) − (eu − ev )T QT Λ2r+2 Q(eu − ev ) ≤ 4(1 − +4 −1/2 c1 p)c2 |λd |2r+1 p1/8 4Bq d|λ1 |2r+1 +4 ≤ 4(1 − +4 = 4(1 −  8 7c1 pn 2r+1 |λd |2r+1 p1/8 + 4(c1 pn + c1 n1/4 )Bq2 d |λd | 1 |λ1 | 2 ! 4Bq d|λ1 |2r+1 (1 + o(1))  8|λ1 | 7|λd |2 c1 pn r+ 1  2 −1/2 c1 p)c2 |λd |2r+1 p1/8 + 4((c1 pn) 1/2 + 1/2 c1 p−1/2 n−1/4 )Bq2 d  1 8 2 |λd |2r+1 1 (7|λ1 |) 2 −1/2 c1 p)c2 |λd |2r+1 p1/8 +  27 7|λ1 | 1 2 (1 + o(1))  6 p 8 (1 + o(1)) 1/2 1/2 (c1 p5/4 n1/2 + c1 p1/4 n−1/4 )Bq2 d|λd |2r+1 (1 + o(1)) ! 1 (c1 pn) 2 −θ (1 + o(1)) . Since θ ∈ (0, 41 ) and p = o(n−1+1/(5+8θ) ), it follows that p = o(n−4/5 ), which can be used to show that the first term dominates the second term in (A.30) because 1/2 2 (1 + o(1)) 1 (A.30) +4 1 ! (c1 pn) 2 −θ 4Bq d|λ1 |2r+1 8 7c1 pn (1 + o(1)) 1 (c1 pn) 2 −θ 4Bq d|λ1 |2r+1 r+1 (1 + o(1)) 1 −1/2 )Bq2 d|λ1 |r ! (c1 pn) 2 −θ = 4(1 − c1 p)c2 + 4(c1 pn + c1 n 1/4 1/2 1/2 1/2 c1 p5/4 n1/2 + c1 p1/4 n−1/4 = p1/8 (c1 p9/8 n1/2 + c1 p1/8 n−1/4 ) 1/2 = p1/8 c1 o(n−9/10 n1/2 + n−1/10 n−1/4 ) 1/2 = p1/8 c1 o(n−2/5 + n−7/20 ) = o(p1/8 ) By the assumption that p = o(n−1+1/(5+8θ) ) and θ ∈ (0, 14 ), it follows that 1 p1/8 = o((c1 pn)− 2 +θ ), such that the last term of (A.30) asymptotically dominates both the first and second terms. Therefore, for sufficiently large 73 ITERATIVE COLLABORATIVE FILTERING n, (A.31) dist1 (u, v) − (eu − ev )T QT Λ2r+2 Q(eu − ev ) ≤ 32Bq d|λ1 |2r+1 1 (c1 pn) 2 −θ ! . By definition, since |λ1 | ≥ · · · ≥ |λd |, (A.32) kΛQ(eu − ev )k22 ≥ |λ1 |−2r (eu − ev )T QT Λ2r+2 Q(eu − ev ), and (A.33) kΛQ(eu − ev )k22 ≤ |λd |−2r (eu − ev )T QT Λ2r+2 Q(eu − ev ). By combining (A.32) and (A.31),  kΛQ(eu − ev )k22 ≥ |λ1 |−2r dist1 (u, v) − 32Bq d|λ1 |2r+1 !  1 (c1 pn) 2 −θ 32Bq d|λ1 | = |λ1 |−2r dist1 (u, v) − . 1 (c1 pn) 2 −θ Similarly, by combining (A.33) and (A.31),  kΛQ(eu − ev )k22 ≤ |λd |−2r dist1 (u, v) + 32Bq d|λ1 |2r+1 !  1 (c1 pn) 2 −θ   32Bq d|λ1 | |λ1 | 2r −2r . = |λd | dist1 (u, v) + 1 (c1 pn) 2 −θ |λd | Lemma A.17 showed that dist1 approximates kΛQ(eu − ev )k22 well if r is constant and the condition number |λ1 |/|λd | is not too large. We can show that in fact dist2 approximates kΛQ(eu − ev )k22 well, adjusting for the distortion which was present in dist1 . 1 Lemma A.18. Assume that θ ∈ (0, 14 ), |λd | = ω((c1 pn)− 2 +θ ), c1 pn = ω(1), and p = o(n−1+1/(5+8θ) ). Let λgap = minij |λi − λj |. Assume that d0 74 BORGS-CHAYES-LEE-SHAH and r satisfy (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ) r≥ , ln(2)   1 θ ≥ (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ), ln 2(1 + 2θ) p  r+ 12 7|λd |2 c1 pn ≥ p−6/8 , 8|λ1 |   0 9c1 pn r+d ≤ (c1 p)−7/8 . 8 (A.34) (A.35) (A.36) (A.37) For any u, v ∈ [n], conditioned on A4 ∩ A5u,r+d0 ∩ A5v,r+d0 ∩i∈[d0 ] A8u,v,r,r+i , for sufficiently large n, 1 dist2 (u, v) − kΛQ(eu − ev )k22 ≤ 32Bq d|λ1 |(c1 pn)− 2 +θ . Proof. Recall that d0 is the number of distinct valued eigenvalues in Λ, thus it is upper bounded by d. The distance estimate dist2 is calculated according to dist2 (u, v) = X  T   1p z(i) 1−c Ñ − Ñ M Ñ − Ñ , u,r v,r 2 u,r+i v,r+i c2 p i∈[d0 ] 0 where z ∈ Rd is a vector that satisfies Λ2r+2 Λ̃z = Λ2 1. Such a vector z always exists and is unique because Λ̃ is a Vandermonde matrix, and Λ−2r 1 lies within the span of its columns. We will split the difference into a few terms dist2 (u, v) − kΛQ(eu − ev )k22  X
T 1p = z(i) 1−c c2 p (Ñu,r − Ñv,r ) M2 (Ñu,r+i − Ñv,r+i ) i∈[d0 ]  − (Ñu,r − Ñv,r )T QT Λi+1 Q(Ñu,r − Ñv,r )  T   X + z(i) Ñu,r − Ñv,r QT Λi+1 Q Ñu,r − Ñv,r − kΛQ(eu − ev )k22 . i∈[d0 ] 75 ITERATIVE COLLABORATIVE FILTERING First let’s consider the last line. By definition,    T X z(i) Ñu,r − Ñv,r QT Λi+1 Q Ñu,r − Ñv,r i∈[d0 ] X = z(i) X i∈[d0 ] 2   T λi+1 Ñ − Ñ e Q u,r v,r k k k By design, recall that z is chosen such that Λ2r+2 Λ̃z = Λ2 1, which implies P that i∈[d0 ] z(i)λi−1 = λ−2r for all k. Therefore, it follows that k k 2 X 2     X X −2r+2 T T z(i) λi+1 Ñ − Ñ = λ Ñ − Ñ , e Q e Q u,r v,r u,r v,r k k k k i∈[d0 ] k k such that  T   X z(i) Ñu,r − Ñv,r QT Λi+1 Q Ñu,r − Ñv,r − kΛQ(eu − ev )k22 i∈[d0 ] = X = X = X  2 X  T Q Ñ − Ñ e − λ2k (eTk Q(eu − ev ))2 λ−2r+2 u,r v,r k k k k    2   2 eTk Q Ñu,r − Ñv,r λ2k λ−2r k − (eTk Q(eu − ev ))2  k λ−2r+2 k eTk Q Ñu,r − Ñv,r − (eTk Λr Q(eu 2 − ev ))  . k By applying Lemma A.5 with a choice of r0 = 0, ∆1 = ∆2 = r, we can bound each term of the summation to show that  T   X z(i) Ñu,r − Ñv,r QT Λi+1 Q Ñu,r − Ñv,r − kΛQ(eu − ev )k22 i∈[d0 ] ≤ X 4Bq |λk |−1 λ2k 1 (c1 pn) 2 −θ k (A.38) ≤ ! 4Bq d|λ1 | 1 (c1 pn) 2 −θ (1 + o(1)) (1 + o(1)), where the last step comes from the assumption that |λk | ≤ |λ1 |. Now let’s consider bounding the difference  X
T 1p z(i) 1−c c2 p (Ñu,r − Ñv,r ) M2 (Ñu,r+i − Ñv,r+i ) i∈[d0 ]  − (Ñu,r − Ñv,r )T QT Λi+1 Q(Ñu,r − Ñv,r ) . 76 BORGS-CHAYES-LEE-SHAH By Lemma A.15, we upper bound the difference  1−c1 p c2 p  T   Ñu,r − Ñv,r M2 Ñu,r+i − Ñv,r+i  T   − Ñu,r − Ñv,r F Ñu,r+i − Ñv,r+i . By Lemma A.6, plugging in r0 = r, r1 = r, and r2 = r + i, we can upper bound the difference  T    T   Ñu,r − Ñv,r F Ñu,r+i − Ñv,r+i − Ñu,r − Ñv,r QΛi+1 Q Ñu,r − Ñv,r . Therefore, by plugging in the corresponding bounds in Lemmas A.15 and A.6,  1−c1 p c2 p  T   Ñu,r − Ñv,r M2 Ñu,r+i − Ñv,r+i  T   − Ñur − Ñvr QT Λi+1 Q Ñu,r − Ñv,r −1/2 ≤ 4(1 − c1 p)c2 |λd |2r+i p1/8 + 4(c1 pn + c1 n  +4 |λd |2 2 r 1/4 )Bq2 d|λ1 |r 4Bq d|λ1 |i 1 (c1 pn) 2 −θ  8 7c1 pn r+i (1 + o(1)) ! (1 + o(1)) . By definition of z, Λ2r+2 Λ̃z = Λ2 1, where we recall that Λ̃(a, b) = λb−1 a . b−1 Let us define a diagonal matrix D with Dbb = |λ1 | such that (Λ̃D)ab =  b−1 λa + . Let (Λ̃D) denote the pseudoinverse of (Λ̃D) if there are repeated |λ1 | eigenvalues, such that z = D−1 (Λ̃D)+ Λ−2r 1 and zi = eTi D−1 (Λ̃D)+ Λ−2r X eh h = X + −i+1 λ−2r . h (Λ̃D)ih |λ1 | h By a result from Gautschi in 1962 (On inverses of Vandermonde and confluent Vandermonde matrices), we can bound the sum of the entries of 77 ITERATIVE COLLABORATIVE FILTERING the pseudoinverse of a Vandermonde matrix with the following bound: X X X Y  |λ1 | + |λj |  + |(Λ̃D)ji | ≤ |λi − λj | 0 0 0 i∈[d ] j∈[d ] (A.39) i∈[d ] j6=i 0 ≤d  2|λ1 | mini,j |λi − λj | d0 −1 . We use λgap to denote the minimum gap between eigenvalues mini,j |λi −λj |. (Although they give the result for inverses, it extends to the pseudoinverse when there are repeated eigenvalues.) The above bound will become useful later. 78 BORGS-CHAYES-LEE-SHAH Therefore, X T 1p z(i) 1−c c2 p (Ñur − Ñvr ) M2 (Ñu,r+i − Ñv,r+i ) i∈[d0 ]
− (Ñur − Ñvr )T QT Λi+1 Q(Ñu,r − Ñv,r ) XX + T −i+1 1−c1 p = λ−2r h (Λ̃D)ih |λ1 | c2 p (Ñur − Ñvr ) M2 (Ñu,r+i − Ñv,r+i ) i∈[d0 ] h
− (Ñur − Ñvr )T QT Λi+1 Q(Ñu,r − Ñv,r ) h X X −1/2 + −i+1 ≤ λ−2r ( Λ̃D) |λ | 4(1 − c1 p)c2 |λd |2r+i p1/8 1 h ih h i∈[d0 ] +4  |λd |2 2 r   4Bq d|λ1 |i (1 + o(1)) 1 (c1 pn) 2 −θ  r+i i + 4(c1 pn + c1 n1/4 )Bq2 d|λ1 |r 7c18pn (1 + o(1)) X |λ | 2r X −1/2 |λd | i d ≤ 4|λ1 |(1 − c1 p)c2 p1/8 (Λ̃D)+ ih |λ1 | |λh | i∈[d0 ] h  + 4|λ1 |  4Bq d (1 + o(1)) 1 (c1 pn) 2 −θ X |λd |2 2|λh |2 r X (Λ̃D)+ ih i∈[d0 ] h + 4|λ1 |(c1 pn + c1 n1/4 )Bq2 d (1 + o(1)) X 8|λ1 | 7|λh |2 c1 pn r+ 1  2 |λh |2 |λ1 |2 −1/2 1/8 ≤ 4|λ1 |(1 − c1 p)c2 p (Λ̃D)+ ih i∈[d0 ] h XX 1/2 X (Λ̃D)+ ih h i∈[d0 ]  + 4|λ1 | 4Bq d  1 (c1 pn) 2 −θ + 4|λ1 |(c1 pn + c1 n 2−r (1 + o(1)) XX (Λ̃D)+ ih h i∈[d0 ] 1/4 )Bq2 d (1 + o(1)) X h p 6/8 X i∈[d0 ] (Λ̃D)+ ih  8 7|λ1 |c1 pn 1/2 .  8 7|λ1 |c1 pn i− 1 2 79 ITERATIVE COLLABORATIVE FILTERING We substitute in the bound from (A.39) to show that X T 1p z(i) 1−c c2 p (Ñur − Ñvr ) M2 (Ñu,r+i − Ñv,r+i ) i∈[d0 ]
− (Ñur − Ñvr )T QT Λi+1 Q(Ñu,r − Ñv,r ) 0  −1/2 1/8 0 2|λ1 | d −1 ≤ 4|λ1 |(1 − c1 p)c2 p d λgap    d0 −1 4Bq d −r 0 2|λ1 | + 4|λ1 | 2 (1 + o(1)) d 1 −θ λgap (c1 pn) 2 +  27 |λ1 | 7 1/2 (c1 pn + c1 n1/4 )Bq2 d (1 + o(1)) p6/8 (c1 pn)−1/2 d0  2|λ1 | λgap d0 −1 Since θ ∈ (0, 41 ) and p = o(n−1+1/(5+8θ) ), it follows that p = o(n−4/5 ), which can be used to show that 1/2 1/2 c1 p5/4 n1/2 + c1 p1/4 n−1/4 = o(p1/8 ). This implies that the first term asymptotically dominates the third term in the right hand side of the above inequality. Then we put together the previous bound with (A.38) to show that dist2 (u, v) − kΛQ(eu − ev )k22  d0 −1 (a) −1/2 1| ≤ 4|λ1 |(1 − c1 p)c2 p1/8 d0 2|λ (1 + o(1)) λgap    d0 −1 4Bq d −r 0 2|λ1 | + + 4|λ1 | 2 (1 + o(1)) d λgap 1 −θ (c1 pn) 2 (b) ≤ 4Bq d|λ1 | 1 (c1 pn) 2 −θ 4Bq d|λ1 | 1 (c1 pn) 2 −θ (1 + o(1)) (1 + o(1)). To justify inequality (b), we can verify that the second term on the right hand side of (a) is dominated by the third term due to the assumption stated in (A.34), which implies that 2−r d0  2|λ1 | λgap d0 −1 ≤ 1. By the assumption that p = o(n−1+1/(5+8θ) ), it follows that p(1−2θ)/8(1+2θ) = 1 o((c1 pn)− 2 +θ ). The assumption stated in (A.35) implies that p θ/2(1+2θ) 0 d  2|λ1 | λgap d0 −1 ≤ 1. 80 BORGS-CHAYES-LEE-SHAH We combine these two observations to show that the first term on the right hand side of (a) is dominated by the third term,   d0 −1 d0 −1 (1−2θ)/8(1+2θ) θ/2(1+2θ) 0 2|λ1 | 1| p1/8 d0 2|λ = p p d λgap λgap 1 = o((c1 pn)− 2 +θ ). Therefore, for sufficiently large n, dist2 (u, v) − kΛQ(eu − ev )k22 ≤ 32Bq d|λ1 | 1 (c1 pn) 2 −θ . We can bound the probability of event ∩i∈[d0 ] A8u,v,r,r+i . Lemma A.19. fies 1 Assume that |λd | = ω((c1 pn)− 2 +θ ). For r, d0 which satis 9c1 pn 8 r+d0 and  7λ2d c1 pn |λ1 |8 ≤ (c1 p)−7/8 , r+ 12 ≥ p−6/8 , conditioned on A5u,r+d0 ∩ A5v,r+d0 , for sufficiently large n, P ∪i∈[d0 ] ¬A8u,v,r,r+i | FE1 , A5u,r+d0 , A5v,r+d0
(1 − c1 p)p−1/4 ≤ 8d0 exp − 5|λ1 |Bq2 d ! , Proof. Recall the definition of A8u,v,r,r+i from (A.28). The result follows directly from applying union bound on the bound provided in Lemma A.16. Next we show that for any u, v ∈ [n], conditioned on A4 , A5u,r+d0 , A5v,r+d0 , and A6u,v,r+d0 , with high probability a large fraction of the vertices a ∈ [n] will satisfy ∩i∈[d0 ] A8u,a,r,r+i , and thus have good distance measurements dist1 (u, a) and dist2 (u, a). A9u,r,l := n X I(A5a,r+l ∩ A5u,r+l ∩li=1 A8u,a,r,r+i ) a∈[n]\u (A.40) ≥ (1 − c1 p)p−1/4 1 − exp − 10Bq2 d !! X a∈[n]\u I(A5a,r+l ) o 81 ITERATIVE COLLABORATIVE FILTERING Lemma A.20. 1 Assume that |λd | = ω((c1 pn)− 2 +θ ). For r, l which satisfies  9c1 pn 8 and  r+d0 7λ2d c1 pn 8|λ1 | ≤ (c1 p)−7/8 , r ≥ p−6/8 , for large enough n, P(¬A9u,r,l (1 − c1 p)p−1/4 | FΘ , A5u,r+l ) ≤ 8l exp − 10|λ1 |Bq2 d ! Proof. Recall the definition of A9u,r,l from (A.40). For readability, let δ and Vu∗ denote !! (1 − c1 p)p−1/4 , δ = 1 − exp − 10|λ1 |Bq2 d and Vu∗ = {a ∈ [n] \ u : A5a,r+l }. The event ¬A9u,r,l conditioned on FΘ is equivalent to ¬A9u,r,l =   X I(¬A5a,r+l ∪li=1 ¬A8u,a,r,r+i ) ≥ (n − 1) − δ|Vu∗ |  a∈[n]\u    X       X l 8 ∗ 5 = I(∪i=1 ¬Au,a,r,r+i ) ≥ (n − 1) − δ|Vu | I(¬Aa,r+l ) +   a∈[n]\u  a∈[n]\u:A5a,r+l   X  X = I(∪li=1 ¬A8u,a,r,r+i ) ≥ (n − 1) − δ|Vu∗ | − I(¬A5a,r+l )   ∗ a∈Vu a∈[n]\u       X l 8 ∗ ∗ = I(∪i=1 ¬Au,a,r,r+i ) ≥ (n − 1) − δ|Vu | − ((n − 1) − |Vu |)    a∈[n]\u:A5 a,r+l       X l 8 ∗ = I(∪i=1 ¬Au,a,r,r+i ) ≥ (1 − δ)|Vu | .   a∈[n]\u:A5  a,r+l 82 BORGS-CHAYES-LEE-SHAH Therefore, by Lemma A.19 we can bound the probability of I(∪li=1 ¬A8u,a,r,r+i ) and apply Markov’s inequality to show that h i P l ¬A8 5 5 E I(∪ ) | A , A i=1 u,a,r,r+i a∈Vu∗ u,r+l a,r+l P(¬A9u,r,l | FΘ , A5u,r+l ) ≤ ∗ (1 − δ)|Vu | ! (1 − c1 p)p−1/4 (1 − δ)−1 ≤ 8l exp − 5|λ1 |Bq2 d ! (1 − c1 p)p−1/4 = 8l exp − 10Bq2 d A.5. Existence of Close Neighbors. In Lemmas A.17 and A.18, we showed that the distance measurements dist1 and dist2 are good estimates of the true L2 functional differences. We would like to show that with high probability a majority of the entries (a, b) that are included in the final estimate of F̂ (u, v) are indeed points which are close in functional value, i.e. F (a, b) ≈ F (u, v). First we show that with high probability there is a large enough set of vertices a ∈ [n] such that kΛQ(eu − ea )k22 is small. Recall the definition of φ from (2.1), ! n o X φ(ξ) := ess inf α0 ∈X1 P1 α ∈ X1 s.t. λ2k (qk (α) − qk (α0 ))2 < ξ 2 . k By assumption on the form of function f (that it is finite spectrum with an orthonormal decomposition represented by {λk , qk }k∈[d] ), for any vertices u, v, Z X (f (αu , y)−f (αv , y))2 dP1 (y) = λ2k (qk (αu )−qk (αv ))2 = kΛQ(eu −ev )k22 . X1 k∈[d] In essence, for a measure one set of α0 ∈ X1 , φ(ξ) denotes the lower bound on the probability that the latent variable associated to a randomly chosen vertex v, denoted by αv , has a close L2 distance from α0 in terms of the function f (α0 , ·) compared to f (αv , ·). Because we assumed X1 is a compact metric space, and if f satisfies some continuity conditions, then we can show that the function φ(ξ) is well defined and positive for all ξ > 0. Depending if we use dist1 or dist2 , we need the size of the neighborhood around α0 to be chosen as a function of the different bounds in Lemmas A.17 and A.18. Recall that the algorithm chooses to average over points where ITERATIVE COLLABORATIVE FILTERING 83 dist1 or dist2 are less than ξ1 (n) or ξ2 (n) respectively. Let us then define the following thresholds which follow from plugging in ξ1 (n) or ξ2 (n) to the bounds presented in Lemmas A.17 and A.18. (A.41) 2 ξ1LB := (A.42) 2 ξ1U B (A.43) 2 ξ2LB := (A.44) 2 ξ2U B := Bq d|λ1 | , 1 (c1 pn) 2 −θ   |λ1 | 2r 65Bq d|λ1 | := , 1 |λd | (c1 pn) 2 −θ Bq d|λ1 | 1 , 1 . (c1 pn) 2 −θ 65Bq d|λ1 | (c1 pn) 2 −θ (A.45) For some ξ, let’s define event    X
(n − 1)φ(ξ)  2 2 := I kΛQ(e − e )k ≤ ξ ≥ A10 (A.46) . u a 2 u,ξ   2 a∈[n]\u We bound the probability of these events in Lemma A.21. Under the assumption that the latent variables are sampled from the uniform distribution on [0, 1] and the function f is piecewise L-Lipschitz with minimum piece size of ` ≤ 1, the function φ(ξ) is bounded by   ξ φ(ξ) ≥ min `, , 2L because Z (f (αu , y) − f (αv , y))2 dP1 (y) = X1 Z (f (αu , y) − f (αv , y))2 dy [0,1] = X λ2k (qk (αu ) − qk (αv ))2 k∈[d] = kΛQ(eu − ev )k22 . If f is piecewise constant, then L is 0, such that φ(ξ) ≥ `. Lemma A.21. For some function f (having spectral decomposition {λk , qk }k∈[d] ) and latent probability measure P1 such that ! X φ(ξ) := min P1 {α ∈ X1 s.t. λ2k (qk (α) − qk (α0 ))2 < ξ 2 } , α0 ∈X1 k 84 BORGS-CHAYES-LEE-SHAH it holds that  
(n − 1)φ(ξ) P ¬A10 ≤ exp − . u,ξ 8 Proof. Conditioned on αu , X I(kΛQ(eu − ea )k2 ≤ ξ) a∈[n]\u is a sum of (n − 1) independent Bernoulli random variables with probability parameter ! X P αa λ2k (qk (αa ) − qk (αu ))2 < ξ 2 ≥ φ(ξ). k Therefore, it stochastically dominates a Binomial((n−1), φ(ξ)) random variable, such that by stochastic dominance, for any t and αu ,   X P I(kΛQ(eu − ea )k2 ≤ ξ) ≤ t αu  ≤ P (Binomial(n − 1, φ(ξ)) ≤ t | αu ) . a∈[n]\u We choose t = 21 (n − 1)φ(ξ) and show that by Chernoff’s bound,   X 1 P I(kΛQ(eu − ea )k2 ≤ ξ) ≤ (n − 1)φ(ξ) 2 a∈[n]\u   Z 1 = P Binomial(n − 1, φ(ξ)) ≤ (n − 1)φ(ξ) αu dP1 (αu ) 2 X1   Z φ(ξ)(n − 1) ≤ exp − dP1 (αu ) 8 X1   (n − 1)φ(ξ) = exp − . 8 Define (A.47) (A.48) (A.49) (A.50) Vu1 := {a ∈ [n] : dist1 (u, a) < ξ1 (n)}, 2 Wu1 := {a ∈ [n] : dist1 (u, a) < ξ1 (n), kΛQ(eu − ea )k22 < ξ1U B }, Vu2 := {a ∈ [n] : dist2 (u, a) < ξ2 (n)}, 2 Wu2 := {a ∈ [n] : dist2 (u, a) < ξ2 (n), kΛQ(eu − ea )k22 < ξ2U B }. ITERATIVE COLLABORATIVE FILTERING 85 In the following lemma we will assume that φ satisfies    (c1 pn)2θ φ(ξ1LB ) = ω exp − . 8Bq2 d In fact when the latent variables are sampled from the uniform distribution on [0, 1] and the function f is piecewise L-Lipschitz with minimum piece size 1 2 of ` ≤ 1, then φ(ξ) ≥ min(`, ξ/2L). By design, ξ1LB = Bq d|λ1 |(c1 pn)− 2 +θ , such that any exponentially decaying term is dominated by φ(ξ1LB ), which is at least polynomial in (c1 pn). Lemma A.22. 1 Assume that |λd | = ω((c1 pn)− 2 +θ ) and    (c1 pn)2θ φ(ξ1LB ) = ω exp − . 8Bq2 d For p = o(n−1+1/(5+8θ) ) and r which satisfies   9c1 pn r+1 ≤ (c1 p)−7/8 , 8 and  7λ2d c1 pn |λ1 |8 r+ 12 ≥ p−6/8 , conditioned on 9 6 5 5 4 A10 u,ξ1LB ∩ Au,r,1 ∩ Au,v,r+1 ∩ Au,r+1 ∩ Av,r+1 ∩ A , it follows that nφ(ξ1LB ) (1 − o(1)), 2 nφ(ξ1LB ) |Wu1 | ≥ (1 − o(1)), 2   |Wu1 | 2 (c1 pn)2θ =1− exp − (1 + o(1)). |Vu1 | φ(ξ1LB ) 8Bq2 d    2θ 1 pn) and r, d0 satisfy Additionally assuming φ(ξ2LB ) = ω exp − (c8B 2d |Vu1 | ≥ q (A.51) (A.52) (A.53) (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ) r≥ , ln(2)   θ 1 ln ≥ (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ), 2(1 + 2θ) p   0 9c1 pn r+d ≤ (c1 p)−7/8 , 8 86 BORGS-CHAYES-LEE-SHAH for any u, v ∈ [n], conditioned on 9 6 5 5 4 A10 u,ξ2LB ∩ Au,r,d0 ∩ Au,v,r+d0 ∩ Au,r+d0 ∩ Av,r+d0 ∩ A , nφ(ξ2LB ) (1 − o(1)), 2 nφ(ξ2LB ) |Wu2 | ≥ (1 − o(1)), 2   2 (c1 pn)2θ |Wu2 | (1 + o(1)). =1− exp − |Vu2 | φ(ξ2LB ) 8Bq2 d |Vu2 | ≥ 10 9 6 5 Proof. Recall the definitions of A10 u,ξ1LB , Au,ξ2LB , Au,r,1 , Au,v,r+d0 , Au,r+d0 , A5v,r+d0 from (A.46), (A.40), (A.25), and (A.21). Recall the definitions of ξ1LB , ξ1U B , ξ2LB and ξ2U B from (A.41), (A.42), (A.43), and (A.44). Recall the definitions of Vu1 , Wu1 , Vu2 , and Wu2 from (A.47), (A.48), (A.49), and (A.50). Vu1 always contains the vertex u since dist1 (u, u) = 0. For all vertices 2 a such that A5a,r+d0 and A8u,a,r,r+d0 hold, if kΛQ(eu − ea )k22 < ξ1LB , then Lemma A.17 along with the definition of ξ1LB implies that dist1 (u, a) < 1 (c1 pn)− 2 +2θ , such that a ∈ Vu1 . Therefore, the set Vu1 is lower bounded by 2 minus the number the number of vertices such that kΛQ(eu − ea )k22 < ξ1LB 5 8 of vertices such that either Aa,r+1 or Aua,r,r+1 are violated (not satisfied). (A.54) |Vu1 | ≥ 1 + X a∈[n]\u 2 I(kΛQ(eu − ea )k22 < ξ1LB )− X a∈[n]\u I(¬A5a,r+1 ∪ ¬A8u,a,r,r+1 ) ITERATIVE COLLABORATIVE FILTERING 87 Conditioned on A9u,r,1 , A6u,v,r+1 , A5v,r+1 , and A5u,r+1 , by definition X I(¬A5a,r+1 ∪ ¬A5u,r+1 ∪ ¬A8u,a,r,r+1 ) a∈[n]\u ≤ (n − 1) − (1 − c1 p)p−1/4 1 − exp − 10|λ1 |Bq2 d !! X I(A5a,r+1 ) a∈[n]\u !!     (1 − c1 p)p−1/4 (c1 pn)2θ ≤ (n − 1) − 1 − exp − 1 + (n − 2) 1 − exp − 10|λ1 |Bq2 d 8Bq2 d ! !   (c1 pn)2θ (1 − c1 p)p−1/4 + exp − ≤ (n − 1) exp − 10|λ1 |Bq2 d 8Bq2 d (A.55) (c1 pn)2θ ≤ (n − 1) exp − 8Bq2 d   (1 + o(1)), where the last inequality follows from the condition that p = o(n−1+1/(5+8θ) ) such that p−1/4 = ω((c1 pn)2θ ). Conditioned on A10 u,ξ1LB , it follows from (A.55), (A.54), and the definition of A10 u,ξ1LB X X 2 |Vu1 | ≥ 1 + I(kΛQ(eu − ea )k22 < ξ1LB )− I(¬A5a,r+1 ∪ ¬A8u,a,r,r+1 ) a∈[n]\u a∈[n]\u    φ(ξ1LB ) (c1 pn)2θ ≥ 1 + (n − 1) − exp − (1 + o(1)) 2 8Bq2 d     φ(ξ1LB ) (c1 pn)2θ ≥n − exp − (1 − o(1)) 2 8Bq2 d  ≥ nφ(ξ1LB ) (1 − o(1)). 2    2θ 1 pn) where the last step followed from the assumption that φ(ξ1LB ) = ω exp − (c8B . 2d q With similar arguments, we can bound |Vu2 | using instead Lemma A.18 9 6 5 and ξ2LB instead of ξ1LB to show that conditioned on A10 u,ξ2LB , Au,r,d0 , Au,v,r+d0 , Av,r+d0 , and A5u,r+d0 ,     (c1 pn)2θ nφ(ξ2LB ) φ(ξ2LB ) |Vu2 | ≥ n − exp − (1 − o(1)) ≥ (1 − o(1)). 2 8Bq2 d 2 Wu1 always contains the vertex u since dist1 (u, u) = 0 and kΛQ(eu − eu )k22 = 0. For all vertices a such that A5a,r+1 and A8u,a,r,r+1 hold, if kΛQ(eu − 88 BORGS-CHAYES-LEE-SHAH 2 2 ea )k22 < ξ1LB < ξ1U B , then Lemma A.17 implies that dist1 (u, a) < ξ1 (n), and a ∈ Wu1 . Therefore, the set Wu1 is lower bounded by the number of 2 vertices such that kΛQ(eu − ea )k22 < ξ1LB minus the number of vertices such 5 8 that either Aa,r+1 or Aua,r,r+1 are violated (not satisfied). Using the same arguments as for proving the lower bound on |Vu1 |, we can obtain the same lower bound on |Wu1 |, X X 2 I(¬A5a,r+1 ∪ ¬A8u,a,r,r+1 ) I(kΛQ(eu − ea )k22 < ξ1LB )− |Wu1 | ≥ 1 + a∈[n]\u a∈[n]\u  ≥n ≥    φ(ξ1LB ) (c1 pn)2θ (1 + o(1)) − exp − 2 8Bq2 d nφ(ξ1LB ) (1 − o(1)). 2 We use similar arguments to lower bound |Wu2 |,     (c1 pn)2θ φ(ξ2LB ) − exp − (1 + o(1)) |Wu2 | ≥ n 2 8Bq2 d ≥ nφ(ξ2LB ) (1 − o(1)). 2 We also bound the ratio of |Wu1 | to |Vu1 | using (A.55) and the definition of A10 u,ξ1LB , |Wu1 | |Vu1 | ≥ 1+ ≥ 1+ ≥ ≥ 2 2 a∈[n]\u I(kΛQ(eu − ea )k2 < ξ1LB ) P 5 2 2 a∈[n]\u I(¬Aa,r+1 a∈[n]\u I(kΛQ(eu − ea )k2 < ξ1LB ) + 1LB ) 1 + (n−1)φ(ξ 2    ) (c1 pn)2θ (n − 1) φ(ξ1LB + exp − (1 + o(1)) 2 2 8Bq d 1+ P P nφ(ξ1LB )   2θ 1 pn) nφ(ξ1LB ) + 2(n − 1) exp − (c8B (1 + o(1)) 2 qd   2θ 1 pn) nφ(ξ1LB ) − 2(n − 1) exp − (c8B (1 + o(1)) 2d q nφ(ξ1LB )      n−1 2 (c1 pn)2θ =1− exp − (1 + o(1)) n φ(ξ1LB ) 8Bq2 d   2 (c1 pn)2θ =1− exp − (1 + o(1)). φ(ξ1LB ) 8Bq2 d ∪ ¬A8u,a,r,r+1 ) 89 ITERATIVE COLLABORATIVE FILTERING Using similar arguments we can bound the ratio of |Wu2 | to |Vu2 |, showing that   |Wu2 | 2 (c1 pn)2θ (1 + o(1)). ≥1− exp − |Vu2 | φ(ξ2LB ) 8Bq2 d For each a ∈ [n], we show concentration of Let’s define event P b∈[n]\b I((a, b) ∈ E1 ∪ E2 ). (A.56) A11 := ∩a∈[n]   X  b∈[n]\a   I((a, b) ∈ E1 ∪ E2 ) < (c1 + c2 )(p + (n − 1)−3/4 )(n − 1) .  Lemma A.23.
(c1 + c2 )(n − 1)1/4 P ¬A11 ≤ n exp − 3 ! . Proof. We can show this easily because this is just a sum of Bernoulli random variables. For a fixed a,   X E I((a, b) ∈ E1 ∪ E2 ) = (c1 + c2 )p(n − 1) b∈[n]\a By Chernoff’s bound,   X P I((a, b) ∈ E1 ∪ E2 ) > (1 + p−1 (n − 1)−3/4 )(c1 + c2 )p(n − 1) b∈[n]\a (c1 + c2 )(n − 1)1/4 ≤ exp − 3 ! . We use union bound to get the final expression. The estimate F̂ (u, v) is computed by averaging over nearby points defined by the distance estimates dist1 and dist2 . Let Euv1 denote the set of undirected edges (a, b) such that (a, b) ∈ E3 and both dist1 (u, a) and dist1 (v, b) are less than ξ1 (n). We remove duplicate entries, i.e. for some 90 BORGS-CHAYES-LEE-SHAH a, b ∈ Vu1 ∩ Vv1 , we only count the edge (a, b) or (b, a) once. The final estimate F̂ (u, v) produced by using dist1 is computed by averaging over the undirected edge set Euv1 , X 1 (A.57) M3 (a, b). F̂ (u, v) = |Euv1 | (a,b)∈Euv1 Let Euv2 denote the set of undirected edges (a, b) such that (a, b) ∈ E3 , and both dist2 (u, a) and dist2 (v, b) are less than ξ2 (n). The final estimate F̂ (u, v) produced by using dist2 is computed by averaging over the undirected edge set Euv2 , X 1 F̂ (u, v) = (A.58) M3 (a, b). |Euv2 | (a,b)∈Euv2 We would like to show that amongst the entries which are eligible to be in E3 , i.e. entries which are not in E1 or in E2 , the ratio of Wu1 × Wv1 ∗ , n∗ , and w ∗ and Vu1 × Vv1 is lower bounded. Let us define n∗uv1 , wuv1 uv2 uv2 according to (A.59) 1 n∗uv1 = |(Vu1 × Vv1 ) \ E1 , E2 | − |((Vu1 ∩ Vv1 ) × (Vu1 ∩ Vv1 )) \ E1 , E2 | 2 (A.60) 1 ∗ wuv1 = |(Wu1 × Wv1 ) \ E1 , E2 | − |((Wu1 ∩ Wv1 ) × (Wu1 ∩ Wv1 )) \ E1 , E2 | 2 (A.61) 1 n∗uv2 = |(Vu2 × Vv2 ) \ E1 , E2 | − |((Vu2 ∩ Vv2 ) × (Vu2 ∩ Vv2 )) \ E1 , E2 | 2 (A.62) 1 ∗ wuv2 = |(Wu2 × Wv2 ) \ E1 , E2 | − |((Wu2 ∩ Wv2 ) × (Wu2 ∩ Wv2 )) \ E1 , E2 |. 2 We will use Lemma A.23, to argue that A11 holds with high probability, such that there are not “too many” entries which are already part of edge sets E1 , E2 , leaving most entries still eligible to be sampled in E3 and used to compute the final estimate. We will additionally assume that     (c1 pn)2θ −3/4 . φ(ξ1LB ) = ω max p, n , exp − 8Bq2 d One example that would satisfy this condition would be if the latent variables are sampled from the uniform distribution on [0, 1] and the function f is 91 ITERATIVE COLLABORATIVE FILTERING piecewise L-Lipschitz with minimum piece size of ` ≤ 1, such that φ(ξ) ≥ 1 2 min(`, ξ/2L). By design, ξ1LB = Bq d|λ1 |(c1 pn)− 2 +θ . Using the assumption that p = o(n−1+1/(5+8θ) ) for θ < 41 , and r = Θ(1), it follows that φ(ξ1LB ) dominates p and n−3/4 .     2θ 1 pn) Lemma A.24. Assume that φ(ξ1LB ) = ω max p, n−3/4 , exp − (c8B 2d q 1 and |λd | = ω((c1 pn)− 2 +θ ). For p = o(n−1+1/(5+8θ) ) for some θ < ω(n−1+ ) for some  > 0, and r = Θ(1) which satisfies   9c1 pn r+1 ≤ (c1 p)−7/8 , 8 and  7λ2d c1 pn |λ1 |8 r+ 12 1 4, p = ≥ p−6/8 , 9 9 6 5 10 conditioned on A11 ∩ A10 u,ξ1LB ∩ Av,ξ1LB ∩ Au,r ∩ Av,r,1 ∩ Au,v,r+1 ∩ Au,r+1 ∩ A5v,r+1 ∩ A4 , it follows that n2 φ(ξ1LB )2 (1 − o(1)) 8 n2 φ(ξ1LB )2 ≥ (1 − o(1)) 8   4 (c1 pn)2θ ≥1− exp − (1 + o(1)). φ(ξ1LB ) 8Bq2 d n∗uv1 ≥ ∗ wuv1 ∗ wuv1 n∗uv1 Proof. Conditioned on A11 , by definition,   1 n∗uv1 ≥ |Vu1 | |Vv1 | − (c1 + c2 )(p + (n − 1)−3/4 )(n − 1) − |Vu1 ||Vv1 | 2   1 ≥ |Vu1 | |Vv1 | − (c1 + c2 )(pn + n1/4 ) , 2 and   1 ∗ wuv1 ≥ |Wu1 | |Wv1 | − (c1 + c2 )(p + (n − 1)−3/4 )(n − 1) − |Wu1 ||Wv1 | 2   1 ≥ |Wu1 | |Wv1 | − (c1 + c2 )(pn + n1/4 ) . 2 By substituting bounds from Lemma A.22, it follows that   nφ(ξ1LB ) nφ(ξ1LB ) (1 − o(1)) (1 − o(1)) − (c1 + c2 )(pn + n1/4 ) . n∗uv1 ≥ 2 4 92 BORGS-CHAYES-LEE-SHAH By assumption, φ(ξ1LB ) dominates p and n−3/4 , such that n∗uv1 ≥ n2 φ(ξ1LB )2 (1 − o(1)), 8 and ∗ wuv1   φ(ξ1LB ) nφ(ξ1LB ) 1/4 (1 − o(1)) (1 − o(1)) − (c1 + c2 )(pn + n ) ≥ 2 4 n2 φ(ξ1LB )2 = (1 − o(1)). 8 Similarly, we can bound the ratio by ∗ wuv1 n∗uv1
|Wu1 | 12 |Wv1 | − (c1 + c2 )(pn + n1/4 )
≥ |Wu1 | 12 |Wv1 | − (c1 + c2 )(pn + n1/4 ) + |Wu1 |(|Vv1 | − |Wv1 |) + (|Vu1 | − |Wu1 |)|Vv1 | |Wu1 |(|Vv1 | − |Wv1 |) + (|Vu1 | − |Wu1 |)|Vv1 |
=1− 1 |Wu1 | 2 |Wv1 | − (c1 + c2 )(pn + n1/4 ) + |Wu1 |(|Vv1 | − |Wv1 |) + (|Vu1 | − |Wu1 |)|Vv1 | |Vu1 |(|Vv1 | − |Wv1 |) + (|Vu1 | − |Wu1 |)|Vv1 |
=1− |Wu1 | 21 |Wv1 | − (c1 + c2 )(pn + n1/4 ) ! |Vu1 ||Vv1 | 1 − |Wv1 |/|Vv1 | =1− . |Wu1 ||Wv1 | 12 − (c1 + c2 )(pn + n1/4 )|Wv1 |−1 By substitutingbounds from  Lemma A.22, and by the assumption that φ(ξ) (c1 pn)2θ dominates exp − 8B 2 d , it follows that q   |Wu1 | 2 (c1 pn)2θ =1− exp − (1 + o(1)) = 1 − o(1), |Vu1 | φ(ξ1LB ) 8Bq2 d which then implies |Vu1 ||Vv1 | = (1 + o(1))2 . |Wu1 ||Wv1 | Again substituting bounds from Lemma A.22, and by the assumption that φ(ξ) dominates p and n−3/4 , it follows that |Wv1 | ≥ nφ(ξ1LB ) (1 − o(1)) = ω(pn + n1/4 ), 2 93 ITERATIVE COLLABORATIVE FILTERING such that   1 − |Wv1 |/|Vv1 | 1 1/4 )|W |−1 v1 2 − (c1 + c2 )(pn + n   2θ (c pn) 2 1 (1 + o(1)) exp − 2 φ(ξ1LB ) 8Bq d  = 1 2 1/4 ) − (c + c )(pn + n (1 + o(1)) 1 2 2 nφ(ξ1LB )    −1 1 (c1 pn)2θ 2 (1 + o(1)) exp − − o(1)(1 + o(1)) = φ(ξ1LB ) 8Bq2 d 2   4 (c1 pn)2θ = (1 + o(1)). exp − φ(ξ1LB ) 8Bq2 d Therefore we put it all together to show that   ∗ wuv1 4 (c1 pn)2θ 2 ≥ 1 − (1 + o(1)) exp − (1 + o(1)) n∗uv1 φ(ξ1LB ) 8Bq2 d   (c1 pn)2θ 4 exp − (1 + o(1)). =1− φ(ξ1LB ) 8Bq2 d Lemma A.25.     2θ 1 pn) Assume that φ(ξ2LB ) = ω max p, n−3/4 , exp − (c8B 2d q − 12 +θ and |λd | = ω((c1 pn) which satisfies (A.63) (A.64) (A.65) (A.66) ). For p = o(n−1+1/(5+8θ) ) for θ ∈ (1, 1 4 ), and r, d0 (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ) r≥ , ln(2)   1 θ ln ≥ (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ), 2(1 + 2θ) p  2 r+ 12 7λd c1 pn ≥ p−6/8 |λ1 |8   0 9c1 pn r+d ≤ (c1 p)−7/8 , 8 10 9 9 for any u, v ∈ [n], conditioned on A11 ∩ A10 u,ξ2LB ∩ Av,ξ2LB ∩ Au,r,d0 ∩ Av,r ∩ 94 BORGS-CHAYES-LEE-SHAH A6u,v,r+d0 ∩ A5u,r+d0 ∩ A5v,r+d0 ∩ A4 , it follows that n2 φ(ξ2LB )2 (1 − o(1)) 8 n2 φ(ξ2LB )2 ≥ (1 − o(1)) 8   4 (c1 pn)2θ (1 + o(1)). ≥1− exp − φ(ξ2LB ) 8Bq2 d n∗uv2 ≥ ∗ wuv2 ∗ wuv2 n∗uv2 Proof. This proof follows the same argument as the proof of Lemma A.24. Conditioned on A11 we use the definition of A11 along with substituting bounds from Lemma A.22 and using the assumption that p = o(n−1+1/(5+8θ) ) in order to complete the proof. A.6. Error Bounds of Final Estimator. Finally, conditioned on {αu }u∈[n] , E1 , E2 , we want to show concentration properties of E3 , specifically to prove bounds on the final estimate F̂ (u, v). The estimate F̂ (u, v) is computed by averaging over nearby points defined by the distance estimates dist1 and dist2 . Let Euv1 denote the set of undirected edges (a, b) such that (a, b) ∈ E3 and both dist1 (u, a) and dist1 (v, b) are less than 1 ξ1 (n) = 33Bq d|λ1 |2r+1 (c1 pn)− 2 +θ . We remove duplicate entries, i.e. for some a, b ∈ Vu1 ∩ Vv1 , we only count the edge (a, b) or (b, a) once. The final estimate F̂ (u, v) produced by using dist1 is computed by averaging over the undirected edge set Euv1 , (A.67) F̂ (u, v) = 1 |Euv1 | X M3 (a, b). (a,b)∈Euv1 Let Euv2 denote the set of undirected edges (a, b) such that (a, b) ∈ E3 , and 1 both dist2 (u, a) and dist2 (v, b) are less than ξ2 (n) = 33Bq d|λ1 |(c1 pn)− 2 +θ . The final estimate F̂ (u, v) produced by using dist2 is computed by averaging over the undirected edge set Euv2 , (A.68) F̂ (u, v) = 1 |Euv2 | X M3 (a, b). (a,b)∈Euv2 ∗ , n∗ , and w ∗ . We In Lemmas A.24 and A.25, we bounded n∗uv1 , wuv1 uv2 uv2 will use those results to bound |Euv1 |, |Euv2 |, |Euv1 ∩ (Wu1 × Wv1 )|, and ∗ , n∗ , and w ∗ |Euv2 ∩ (Wu2 × Wv2 )|. Recall the definitions of n∗uv1 , wuv1 uv2 uv2 from (A.59), (A.60), (A.61), and (A.62). 95 ITERATIVE COLLABORATIVE FILTERING Let us define the following events    n∗uv1 c3 p 12 (A.69) Au,v,1 := |Euv1 | ∈ (1 ± ξ1U B ) , 1 − c1 p − c2 p    n∗uv2 c3 p 12 (A.70) Au,v,2 := |Euv1 | ∈ (1 ± ξ2U B ) , 1 − c1 p − c2 p    ∗ c p wuv1 3 13 (A.71) Au,v,1 := |(Wu1 × Wv1 ) ∩ E3 | ≥ (1 − ξ1U B ) , 1 − c1 p − c2 p    ∗ c p wuv2 3 13 (A.72) Au,v,2 := |(Wu2 × Wv2 ) ∩ E3 | ≥ (1 − ξ2U B ) . 1 − c1 p − c2 p Observe that by definition, Vu1 , Vv1 , Wu1 , Wv1 , Vu2 , Vv2 , Wu2 , and Wv2 are measurable with respect to FE1 ,E2 . Lemma A.26.     2θ 1 pn) Assume that φ(ξ1LB ) = ω max p, n−3/4 , exp − (c8B 2d q − 12 +θ o(n−1+1/(5+8θ) ) and |λd | = ω((c1 pn) ). For p = for some θ < ω(n−1+ ) for some  > 0 and r = Θ(1) which satisfies  9c1 pn 8 r+1 and  7λ2d c1 pn |λ1 |8 1 4, p = ≤ (c1 p)−7/8 , r+ 12 ≥ p−6/8 , 9 9 6 5 10 conditioned on A11 ∩ A10 u,ξ1LB ∩ Av,ξ1LB ∩ Au,r,1 ∩ Av,r,1 ∩ Au,v,r+1 ∩ Au,r+1 ∩ A5v,r+1 ∩ A4 ,  2 2 2 ξ1U B c3 pn φ(ξ1LB ) (1 − o(1)) , P ≤ 2 exp − 24  2 
ξ1U B c3 pn2 φ(ξ1LB )2 13 P ¬Au,v,1 |FE1 ,E2 ≤ exp − (1 − o(1)) . 16 ¬A12 u,v,1 |FE1 ,E2
 13 Proof. We will first consider A12 u,v,1 , but the proof for Au,v,1 is identical. ∗ Conditioned on FE1 ,E2 , |Euv1 | is distributed as a sum of nuv1 Bernoulli random variables with parameter c3 p/(1−c1 p−c2 p), i.e. the probability that an edge is in E3 conditioned that it is not in E1 ∪ E2 . Therefore in expectation, E [|Euv1 | | FE1 ,E2 ] = c3 pn∗uv1 . 1 − c1 p − c2 p 96 BORGS-CHAYES-LEE-SHAH By Chernoff’s bound for binomials, ¬A12 u,v,1 |FE1 ,E2 P
  2 ∗ ξ1U B nuv1 c3 p ≤ 2 exp − 3(1 − c1 p − c2 p) We can lower bound n∗uv1 by Lemma A.24, such that   2
ξ1U n2 φ(ξ1LB )2 12 B c3 p (1 − o(1)) P ¬Au,v,1 |FE1 ,E2 ≤ 2 exp − 3(1 − c1 p − c2 p) 8  2  ξ1U B c3 pn2 φ(ξ1LB )2 ≤ 2 exp − (1 − o(1)) . 24 ∗ We can use equivalent proof steps with the corresponding bounds for wuv1 from Lemma A.24 to show that   2 ∗
ξ1U 13 B wuv1 c3 p P ¬Au,v,1 |FE1 ,E2 ≤ exp − 2(1 − c1 p − c2 p)  2  ξ c3 pn2 φ(ξ1LB )2 ≤ exp − 1U B (1 − o(1)) . 16 Lemma A.27.     2θ 1 pn) Assume that φ(ξ2LB ) = ω max p, n−3/4 , exp − (c8B 2d q 1 and |λd | = ω((c1 pn)− 2 +θ ). For p = o(n−1+1/(5+8θ) ) for some θ < ω(n−1+ ) for some  > 0, and r, d0 which satisfies (A.73) (A.74) (A.75) (A.76) 1 4, p = (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ) r≥ , ln(2)   θ 1 ln ≥ (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ), 2(1 + 2θ) p r+ 12  2 7λd c1 pn ≥ p−6/8 |λ1 |8   0 9c1 pn r+d ≤ (c1 p)−7/8 , 8 10 9 9 for any u, v ∈ [n], conditioned on A11 ∩ A10 u,ξ2LB ∩ Av,ξ2LB ∩ Au,r,d0 ∩ Av,r,d0 ∩ A6u,v,r+d0 ∩ A5u,r+d0 ∩ A5v,r+d0 ∩ A4 ,  2 
ξ2U B c3 pn2 φ(ξ2LB )2 12 P ¬Au,v,2 |FE1 ,E2 ≤ 2 exp − (1 − o(1)) 24  2 
ξ2U B c3 pn2 φ(ξ2LB )2 13 P ¬Au,v,2 |FE1 ,E2 ≤ exp − (1 − o(1)) . 16 97 ITERATIVE COLLABORATIVE FILTERING Proof. Using the same proof steps as the proof of Lemma A.26. Observe that |Euv2 | and |Euv2 ∩ (Wu2 × Wv2 )| conditioned on FE1 ,E2 are Binomial random variables. Therefore by applying Chernoff’s bound and using results in Lemma A.25   2 ∗
ξ2U 12 B nuv2 c3 p P ¬Au,v,2 |FE1 ,E2 ≤ 2 exp − , 3(1 − c1 p − c2 p)  2  ξ c3 pn2 φ(ξ2LB )2 ≤ 2 exp − 2U B (1 − o(1)) , 24 and P ¬A13 u,v,2
  2 ∗ ξ2U B wuv2 c3 p ≤ exp − 2(1 − c1 p − c2 p)  2  ξ2U B c3 pn2 φ(ξ2LB )2 ≤ exp − (1 − o(1)) . 16 Next, conditioned on FE1 ,E2 , we want to show concentration of 1 |Euv1 | and X (a,b)∈Eu,v1 1 |Euv2 | (f (αa , αb ) − M3 (a, b)) X (f (αa , αb ) − M3 (a, b)) (a,b)∈Eu,v2 Let’s define the following events    1  X (f (α , α ) − M (a, b)) (A.77) A14 := < ξ , a 3 1U B b u,v,1  |Euv1 |  (a,b)∈Euv1     1 X (A.78) < ξ . (f (α , α ) − M (a, b)) A14 := a 3 2U B b u,v,2   |Euv2 | (a,b)∈Euv2 Lemma A.28.     2θ 1 pn) Assume that φ(ξ1LB ) = ω max p, n−3/4 , exp − (c8B 2d q − 12 +θ o(n−1+1/(5+8θ) ) and |λd | = ω((c1 pn) ). For p = for some θ < ω(n−1+ ) for some  > 0, and r = Θ(1) which satisfies   9c1 pn r+1 ≤ (c1 p)−7/8 , 8 1 4, p = 98 BORGS-CHAYES-LEE-SHAH and  7λ2d c1 pn |λ1 |8 r+ 12 ≥ p−6/8 , 11 10 10 9 9 6 conditioned on A12 u,v,1 ∩ A ∩ Au,ξ1LB ∩ Av,ξ1LB ∩ Au,r,1 ∩ Av,r,1 ∩ Au,v,r+1 ∩ A5u,r+1 ∩ A5v,r+1 ∩ A4 , it holds that P ¬A14 u,v,1 FE2 ,E1 , Euv1
  2 2 c3 pn2 ξ1U B φ(ξ1LB ) (1 − o(1)) . ≤ 2 exp − 16 Proof. First we show that the expression has zero mean.   X E (f (αa , αb ) − M3 (a, b)) FE1 ,E2 , Euv1  (a,b)∈Euv1 X = (f (αa , αb ) − E [Z(a, b) | FE1 ,E2 ]) = 0 (a,b)∈Euv1 We can also compute the variance. Each undirected edge is distinct, thus conditioned on FE1 ,E2 and Euv1 , the individual terms of the summation are independent, and Z(a, b) ∈ [0, 1], therefore   X Var  (f (αa , αb ) − M3 (a, b)) FE1 ,E2 , Euv1  (a,b)∈Euv1 = X Var [Z(a, b) | FE1 ,E2 ] (a,b)∈Euv1 ≤ X 1 (a,b)∈Euv1 ≤ |Euv1 |. By Bernstein’s inequality, P ¬A14 u,v1   2
3|Euv1 |ξ1U B . FE2 ,E1 ,Θ , Euv ≤ 2 exp − 6 + 2ξ1U B 11 10 10 9 By Lemma A.24, conditioned on A12 u,v,1 ∩ A ∩ Au,ξ1LB ∩ Av,ξ1LB ∩ Au,r,1 ∩ 99 ITERATIVE COLLABORATIVE FILTERING A9v,r,1 ∩ A6u,v,r+1 ∩ A5u,r+1 ∩ A5v,r+1 ∩ A4 ,  c3 pn∗u,v,1 |Euv1 | ≥ (1 − ξ1U B ) 1 − c1 p − c2 p n2 φ(ξ1LB )2 ≥ (1 − ξ1U B )c3 p (1 − o(1)) 8 c3 pn2 φ(ξ1LB )2 (1 − o(1)). = 8  Therefore, P ¬A14 u,v,1 FE1 ,E2 , Euv1 Lemma A.29.
  2 3ξ1U c3 pn2 φ(ξ1LB )2 B ≤ 2 exp − (1 − o(1)) 6 + 2ξ1U B 8   2 2 c3 pn2 ξ1U B φ(ξ1LB ) (1 − o(1)) . ≤ 2 exp − 16     2θ 1 pn) Assume that φ(ξ2LB ) = ω max p, n−3/4 , exp − (c8B 2d q 1 and |λd | = ω((c1 pn)− 2 +θ ). For p = o(n−1+1/(5+8θ) ) for some θ < 14 , and r, d0 which satisfies (A.79) (A.80) (A.81) (A.82) (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ) r≥ , ln(2)   θ 1 ln ≥ (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ), 2(1 + 2θ) p  2 r+ 12 7λd c1 pn ≥ p−6/8 |λ1 |8   0 9c1 pn r+d ≤ (c1 p)−7/8 , 8 11 10 10 9 for any u, v ∈ [n], conditioned on A12 u,v,2 ∩ A ∩ Au,ξ2LB ∩ Av,ξ2LB ∩ Au,r,d0 ∩ 9 6 5 5 4 Av,r,d0 ∩ Au,v,r+d0 ∩ Au,r+d0 ∩ Av,r+d0 ∩ A , it holds that P ¬A14 u,v,2 FE1 ,E2 , Euv2
  2 2 c3 pn2 ξ2U B φ(ξ2LB ) ≤ 2 exp − (1 − o(1)) . 16 Proof. We use the same arguments from the proof of Lemma A.28 to prove the bound on A14 u,v,2 . We apply Bernstein’s inequality for sums of independent bounded random variables, and we use Lemma A.25, conditioned 100 BORGS-CHAYES-LEE-SHAH 11 10 9 6 5 5 4 on A12 u,v,2 , A , Au,ξ2LB ∩ Au,r,d0 ∩ Au,v,r+d0 ∩ Au,r+d0 ∩ Av,r+d0 ∩ A , to show that   2 2
c3 pn2 ξ2U 14 B φ(ξ2LB ) (1 − o(1)) . P ¬Au,v,2 FE1 ,E2 , Euv2 ≤ 2 exp − 16    2θ −1 1 pn) In the next Lemma, we assume that φ(ξ1LB ) = ω ξ1LB . exp − (c8B 2d q We again verify that this is satisfied for a piecewise Lipschitz latent function, as both φ(ξ1LB ) and ξ1LB decay polynomially in c1 pn.     2θ −1 1 pn) Lemma A.30. Assume that φ(ξ1LB ) = ω max p, n−3/4 , ξ1LB exp − (c8B 2d q − 12 +θ o(n−1+1/(5+8θ) ) and |λd | = ω((c1 pn) ). For p = for some θ < ω(n−1+ ) for some  > 0, and r = Θ(1) which satisfies   9c1 pn r+1 ≤ (c1 p)−7/8 , 8 and  7λ2d c1 pn |λ1 |8 (2r+1)/2 1 4, p = ≥ p−6/8 , 6 5 9 12 11 10 conditioned on A13 u,v,1 ∩ Au,v,1 ∩ A ∩ Au,ξ1LB ∩ Au,r,1 ∩ Au,v,r+1 ∩ Au,r+1 ∩ A5v,r+1 A4 , 1 X |Euv1 | (f (αa , αb ) − f (αu , αv )) = O (a,b)∈Euv1  |λ1 | |λd | r  Bq3 d2 |λ1 | 1/2 ! 1 (c1 pn) 2 −θ . Proof. We first decompose the sum into indices (a, b) ∈ (Wu1 × Wv1 ) and those indices in the complement. 1 X |Euv1 | = (a,b)∈Euv1 1 X |Euv1 | + (f (αa , αb ) − f (αu , αv )) (f (αa , αb ) − f (αu , αv ))I((a, b) ∈ Wu1 × Wv1 ) (a,b)∈Euv1 1 |Euv1 | X (a,b)∈Euv1 (f (αa , αb ) − f (αu , αv ))I((a, b) ∈ (Wu1 × Wv1 )c ) . ITERATIVE COLLABORATIVE FILTERING 101 Recall that Wu1 indicates vertices a such that kΛQ(ea − eu )k2 ≤ ξ1U B . Note that |(f (αa , αb ) − f (αu , αv ))| = |eTa QT ΛQeb − eTu QT ΛQev | √ ≤ Bq d (kΛQ(ea − eu )k2 + kΛQ(eb − ev )k2 ) , such that (a, b) ∈ Wu1 × Wv1 also implies that √ (f (αa , αb ) − f (αu , αv )) ≤ 2Bq dξ1U B . Therefore, the first term is bounded by 1 |Euv1 | X (a,b)∈Euv1 √ √ 2Bq dξ1U B |Euv1 ∩ (Wu1 × Wv1 )| 2Bq dξ1U B I((a, b) ∈ Wu1 × Wv1 ) = . |Euv1 | Because the function f takes value in [0, 1], the second term is trivially bounded by X 1 |Euv1 ∩ (Wu1 × Wv1 )| . I((a, b) ∈ (Wu1 × Wv1 )c ) = 1 − |Euv1 | |Euv1 | c (a,b)∈Euv1 ∩(Wu1 ×Wv1 ) 12 Conditioned on A13 u,v,1 ∩ Au,v,1 , 1 |Euv1 | X (f (αa , αb ) − f (αu , αv )) (a,b)∈Euv1 √ |Euv1 ∩ (Wu1 × Wv1 )| ≤ 1 − (1 − 2Bq dξ1U B ) |Euv1 |  ∗  wuv1 c3 p (1 − ξ ) √ 1U B 1−c1 p−c2 p  ∗  ≤ 1 − (1 − 2Bq dξ1U B ) n c3 p (1 + ξ1U B ) 1−cuv1 1 p−c2 p   ∗ √ 2ξ1U B wuv1 . ≤ 1 − (1 − 2Bq dξ1U B ) 1 − 1 + ξ1U B n∗uv1 Therefore by Lemma A.24, 1 |Euv1 | X (f (αa , αb ) − f (αu , αv )) (a,b)∈Euv1 ≤ 1 − (1 − 2Bq √     (c1 pn)2θ exp − (1 + o(1)) φ(ξ1LB ) 8Bq2 d   4 (c1 pn)2θ + exp − (1 + o(1)). φ(ξ1LB ) 8Bq2 d 2ξ1U B dξ1U B ) 1 − 1 + ξ1U B √ 2ξ1U B ≤ 2Bq dξ1U B + 1 + ξ1U B   1− 4 102 BORGS-CHAYES-LEE-SHAH    2θ −1 1 pn) , such that the first and By assumption, φ(ξ1LB ) = ω ξ1LB exp − (c8B 2 qd second term dominate and by the definition of ξ1U B , 1 √ (f (αa , αb ) − f (αu , αv )) ≤ O(Bq dξ1U B ) X |Euv1 | (a,b)∈Euv1 =O Lemma A.31.  |λ1 | |λd | r  Bq3 d2 |λ1 | 1 (c1 pn) 2 −θ 1/2 ! .     2θ −1 1 pn) Assume that φ(ξ2LB ) = ω max p, n−3/4 , ξ2LB exp − (c8B 2d q − 12 +θ and |λd | = ω((c1 pn) which satisfies ). For p = o(n−1+1/(5+8θ) ) for some θ < 1 4, and r, d0 (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ) r≥ , ln(2)   θ 1 ln ≥ (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ), 2(1 + 2θ) p  2 r+ 12 7λd c1 pn ≥ p−6/8 |λ1 |8  0  9c1 pn r+d ≤ (c1 p)−7/8 , 8 (A.83) (A.84) (A.85) (A.86) 10 12 11 10 for any u, v ∈ [n], conditioned on A13 u,v,2 ∩ Au,v,2 ∩ A ∩ Au,ξ2LB ∩ Av,ξ2LB ∩ A9u,r,d0 ∩ A9v,r,d0 ∩ A6u,v,r+d0 ∩ A5u,r+d0 ∩ A5v,r+d0 A4 , 1 |Euv2 | X (a,b)∈Euv2  (f (αa , αb ) − f (αu , αv )) = O Bq3 d2 |λ1 | 1 (c1 pn) 2 −θ 1/2 ! . Proof. We prove this using a similar argument as the proof of Lemma 103 ITERATIVE COLLABORATIVE FILTERING A.30. We first decompose the sum 1 X |Euv2 | = 1 X |Euv2 | + (f (αa , αb ) − f (αu , αv )) (a,b)∈Euv2 (f (αa , αb ) − f (αu , αv ))I((a, b) ∈ Wu2 × Wv2 ) (a,b)∈Euv2 1 X |Euv2 | (f (αa , αb ) − f (αu , αv ))I((a, b) ∈ (Wu2 × Wv2 )c ) . (a,b)∈Euv2 For (a, b) ∈ Wu2 × Wv2 , √ |(f (αa , αb ) − f (αu , αv ))| ≤ 2Bq dξ2U B , such that the first term is bounded by 1 |Euv2 | X (a,b)∈Euv2 √ √ 2Bq dξ2U B |Euv2 ∩ (Wu2 × Wv2 )| 2Bq dξ2U B I((a, b) ∈ Wu2 × Wv2 ) = . |Euv2 | Because the function f takes value in [0, 1], the second term is trivially bounded by 1 |Euv2 | X I((a, b) ∈ (Wu2 × Wv2 )c ) = 1 − (a,b)∈Euv2 ∩(Wu2 ×Wv2 )c |Euv2 ∩ (Wu2 × Wv2 )| . |Euv2 | 12 Conditioned on A13 u,v,2 ∩ Au,v,2 , 1 |Euv2 | X (f (αa , αb ) − f (αu , αv )) (a,b)∈Euv1 √ ≤ 1 − (1 − 2Bq dξ2U B ) 1 − 1 2(c1 pn)− 4 +θ 1 1 + (c1 pn)− 4 +θ Therefore by Lemma A.25, using the assumption that    (c1 pn)2θ −1 φ(ξ1LB ) = ω ξ2LB exp − , 8Bq2 d ! ∗ wuv2 . n∗uv2 104 BORGS-CHAYES-LEE-SHAH and the definition of ξ2U B , 1 X |Euv2 | (f (αa , αb ) − f (αu , αv )) (a,b)∈Euv2 √  = O(Bq dξ2U B ) = O Bq3 d2 |λ1 | 1 (c1 pn) 2 −θ 1/2 ! . Let us define the following events 14 13 12 11 10 10 9 9 A15 u,v,1 := Au,v,1 ∩ Au,v,1 ∩ Au,v,1 ∩ A ∩ Au,ξ1LB ∩ Av,ξ1LB ∩ Au,r,1 ∩ Av,r,1 (A.87) ∩ A6u,v,r+1 ∩ A5u,r+1 ∩ A5v,r+1 ∩ A4 , 9 9 10 14 13 12 11 10 A15 u,v,2 := Au,v,2 ∩ Au,v,2 ∩ Au,v,2 ∩ A ∩ Au,ξ2LB ∩ Av,ξ2LB ∩ Au,r,d0 ∩ Av,r,d0 (A.88) ∩ A6u,v,r+d0 ∩ A5u,r+d0 ∩ A5v,r+d0 ∩ A4 . Lemma A.32. Let     (c1 pn)2θ −3/4 −1 φ(ξ1LB ) = ω max p, n , ξ1LB exp − , 8Bq2 d 1 and |λd | = ω((c1 pn)− 2 +θ ). For p = o(n−1+1/(5+8θ) ) for some θ < ω(n−1+ ) for some  > 0, and r = Θ(1) which satisfies  9c1 pn 8 r+1 and  7λ2d c1 pn |λ1 |8 1 4, p = ≤ (c1 p)−7/8 , r+ 12 ≥ p−6/8 , for any (u, v) ∈ [n] × [n], conditioned on A15 u,v,1 , the error of the estimate output when using dist1 is bounded by 1/2 !  r  3 2 √ Bq d |λ1 | |λ1 | |F̂ (u, v) − f (αu , αv )| = O(Bq dξ1U B ) = O . 1 −θ |λd | (c1 pn) 2 105 ITERATIVE COLLABORATIVE FILTERING Proof. To obtain a final bound on the estimate, we separate out the noise terms with the bias terms. |F̂ (u, v) − f (αu , αv )| = X 1 M3 (a, b) − f (αu , αv ) |Euv1 | (a,b)∈Euv1 = X 1 |Euv1 | (M3 (a, b) − f (αa , αb )) (a,b)∈Euv1 + 1 |Euv1 | X (f (αa , αb ) − f (αu , αv )) . (a,b)∈Euv1 Conditioned on A14 u,v,1 , the first term is bounded by ξ1U B . By Lemma A.30, the second term is bounded by O(Bq dξ1U B ). Combining these two bounds leads to the desired result.     2θ −1 1 pn) exp − (c8B Lemma A.33. Assume that φ(ξ2LB ) = ω max p, n−3/4 , ξ2LB 2d q − 12 +θ and |λd | = ω((c1 pn) which satisfies (A.89) (A.90) (A.91) (A.92) ). For p = o(n−1+1/(5+8θ) ) for some θ < 1 4, and r, d0 (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ) r≥ , ln(2)   θ 1 ln ≥ (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ), 2(1 + 2θ) p  2 r+ 12 7λd c1 pn ≥ p−6/8 |λ1 |8   0 9c1 pn r+d ≤ (c1 p)−7/8 , 8 for any (u, v) ∈ [n] × [n], conditioned on A15 u,v,2 , the error of the estimate output when using dist2 is bounded by  1/2 ! √ Bq3 d2 |λ1 | |F̂ (u, v) − f (αu , αv )| = O(Bq dξ2U B ) = O . 1 −θ (c1 pn) 2 Proof. The proof is equivalent to the proof of Lemma A.32, and follows from combining the definition of A14 u,v,1 along with Lemma A.31. 106 BORGS-CHAYES-LEE-SHAH Lemma A.34.     2θ −1 1 pn) Assume that φ(ξ1LB ) = ω max p, n−3/4 , ξ1LB exp − (c8B 2d q 1 1 and |λd | = ω((c1 pn)− min( 4 , 2 −θ) ). For p = o(n−1+1/(5+8θ) ) for some θ < 14 , c1 pn = ω(1), p = ω(n−1+ ) for some  > 0, and r = Θ(1) which satisfies  9c1 pn 8 r+1 and  7λ2d c1 pn |λ1 |8 ≤ (c1 p)−7/8 , r+ 12 ≥ p−6/8 ,        2 2 c pn2 ξ1U (n−1)φ(ξ1LB ) (c1 pn)2θ B φ(ξ1LB ) + exp − 3 + exp − . P(¬A15 u,v,1 ) ≤ O d exp − 8B 2 d 24 8 q If φ(ξ) ≥ min(`, ξ/2L), then   
(c1 pn)2θ P ¬A15 = O exp − . u,v,1 8Bq2 d Proof. For events Z1 , Z2 , Z3 , we can use the inequality which results from P(¬Z1 ∪ ¬Z2 ∪ ¬Z3 ) = P(¬Z1 |Z2 ∩ Z3 )P(Z2 ∩ Z3 ) + P(¬Z2 ∪ ¬Z3 ) ≤ P(¬Z1 |Z2 ∩ Z3 ) + P(¬Z2 ∪ ¬Z3 ). Applying the above inequality multiple times and applying the union bound allows us to decompose the desired probability computation into each event ITERATIVE COLLABORATIVE FILTERING 107 for which we have previously defined bounds.
P ¬A15 u,v,1 12 11 10 10 9 9 6 5 5 4 ≤ P ¬A14 u,v,1 ∪ ¬Au,v,1 | A ∩ Au,ξ1LB ∩ Au,ξ1LB ∩ Au,r,1 ∩ Av,r,1 ∩ Au,v,r+1 ∩ Au,r+1 ∩ Av,r+1 ∩ A
11 10 10 9 9 6 5 5 4 + P ¬A13 u,v,1 | A ∩ Au,ξ1LB ∩ Au,ξ1LB ∩ Au,r,1 ∩ Av,r,1 ∩ Au,v,r+1 ∩ Au,r+1 ∩ Av,r+1 ∩ A
10 9 9 6 5 5 4 + P ¬A11 ∪ ¬A10 u,ξ1LB ∪ ¬Au,ξ1LB ∪ ¬Au,r,1 ∪ ¬Av,r,1 ∪ ¬Au,v,r+1 ∪ ¬Au,r+1 ∪ ¬Av,r+1 ∪ ¬A
12 11 10 10 9 9 6 5 5 4 ≤ P ¬A14 u,v,1 | Au,v,1 ∩ A ∩ Au,ξ1LB ∩ Au,ξ1LB ∩ Au,r,1 ∩ Av,r,1 ∩ Au,v,r+1 ∩ Au,r+1 ∩ Av,r+1 ∩ A
11 10 10 9 9 6 5 5 4 + P ¬A12 u,v,1 | A ∩ Au,ξ1LB ∩ Au,ξ1LB ∩ Au,r,1 ∩ Av,r,1 ∩ Au,v,r+1 ∩ Au,r+1 ∩ Av,r+1 ∩ A
11 10 10 9 9 6 5 5 4 + P ¬A13 u,v,1 | A ∩ Au,ξ1LB ∩ Au,ξ1LB ∩ Au,r,1 ∩ Av,r,1 ∩ Au,v,r+1 ∩ Au,r+1 ∩ Av,r+1 ∩ A
10 + P ¬A11 + P(¬A10 u,ξ1LB ) + P(¬Au,ξ1LB )
+ P(¬A9u,r,1 ∪ ¬A5u,r+1 ) + P(¬A9v,r,1 ∪ ¬A5v,r+1 ) + P(¬A6u,v,r+1 ) + P ¬A4
9 9 6 5 5 4 10 12 11 10 ≤ P ¬A14 u,v,1 | Au,v,1 ∩ A ∩ Au,ξ1LB ∩ Au,ξ1LB ∩ Au,r,1 ∩ Av,r,1 ∩ Au,v,r+1 ∩ Au,r+1 ∩ Av,r+1 ∩ A
9 9 6 5 5 4 10 11 10 + P ¬A12 u,v,1 | A ∩ Au,ξ1LB ∩ Au,ξ1LB ∩ Au,r,1 ∩ Av,r,1 ∩ Au,v,r+1 ∩ Au,r+1 ∩ Av,r+1 ∩ A
11 10 10 9 9 6 5 5 4 + P ¬A13 u,v,1 | A ∩ Au,ξ1LB ∩ Au,ξ1LB ∩ Au,r,1 ∩ Av,r,1 ∩ Au,v,r+1 ∩ Au,r+1 ∩ Av,r+1 ∩ A
9 5 10 + P ¬A11 + P(¬A10 u,ξ1LB ) + P(¬Au,ξ1LB ) + P(¬Au,r,1 |Au,r+1 )
+ P(¬A5u,r+1 ) + P(¬A9v,r,1 |A5v,r+1 ) + P(¬A5v,r+1 ) + P(¬A6u,v,r+1 ) + P ¬A4 . We then bound this expression by using Lemmas A.9, A.12, A.13, A.20, A.21, A.23, A.26, and A.28 to show that
P ¬A15 u,v,1    2  2 2 c3 pn2 ξ1U ξ1U B c3 pn2 φ(ξ1LB )2 B φ(ξ1LB ) ≤ 2 exp − (1 − o(1)) + exp − (1 − o(1)) 16 16 !  2  ξ1U B c3 pn2 φ(ξ1LB )2 (c1 + c2 )(n − 1)1/2 + 2 exp − (1 − o(1)) + n exp − 24 3 !   (1 − c1 p)p−1/4 (n − 1)φ(ξ1LB ) + 16 exp − + 2 exp − 8 10|λ1 |Bq2 d !     (c1 pn)2θ (c1 pn)2θ c1 (n − 1)1/4 + 4(d + 2) exp − + 8(d + 2) exp − + n exp − . 8Bq2 d 4Bq2 d 3 Using the assumption that p = o(n−1+1/(5+8θ) ) for some θ < 14 implies that n1/4 − ln(n) = ω((c1 pn)2θ ) and p−1/4 = ω((c1 pn)2θ ), such that        2 2
c3 pn2 ξ1U (c1 pn)2θ (n−1)φ(ξ1LB ) B φ(ξ1LB ) P ¬A15 = O d exp − + exp − + exp − . u,v,1 24 8 8B 2 d q
108 BORGS-CHAYES-LEE-SHAH Furthermore, φ(ξ) ≥ min(`, ξ/2L) and r = Θ(1) also implies that (n − 2 2 2θ 1)φ(ξ1LB ) = ω((c1 pn)2θ ) and c3 pn2 ξ1U B φ(ξ1LB ) = ω((c1 pn) ), such that P ¬A15 u,v,1
   (c1 pn)2θ . = O d exp − 8Bq2 d     2θ −1 1 pn) Assume that φ(ξ2LB ) = ω max p, n−3/4 , ξ2LB exp − (c8B 2d Lemma A.35. q 1 1 and |λd | = ω((c1 pn)− min( 4 , 2 −θ) ). For p = o(n−1+1/(5+8θ) ) for some θ < 14 , c1 pn = ω(1), and some r, d0 which satisfies (A.93) (A.94) (A.95) (A.96) (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ) , r≥ ln(2)   θ 1 ln ≥ (d0 − 1) ln(2|λ1 |/λgap ) + ln(d0 ), 2(1 + 2θ) p  2 r+ 12 7λd c1 pn ≥ p−6/8 |λ1 |8  0  9c1 pn r+d ≤ (c1 p)−7/8 , 8 for any u, v ∈ [n], P(¬A15 u,v,2 )     2    2θ ξ c pn2 φ(ξ )2 1 pn) 2LB ) = O d exp − (c8B + exp − 2U B 3 24 2LB (1 − o(1)) + exp − (n−1)φ(ξ . 2d 8 q If φ(ξ) ≥ min(`, ξ/2L), then   
(c1 pn)2θ P ¬A15 = O d exp − . u,v,2 8Bq2 d Proof. We use the same proof steps as the proof for Lemma A.34, and plug in Lemmas A.9, A.12, A.13, A.20, A.21, A.23, A.27, and A.29 to show ITERATIVE COLLABORATIVE FILTERING 109 that
P ¬A15 u,v,2    2  2 2 c3 pn2 ξ2U ξ2U B c3 pn2 φ(ξ2LB )2 B φ(ξ2LB ) ≤ 2 exp − (1 − o(1)) + exp − (1 − o(1)) 16 16 !  2  ξ2U B c3 pn2 φ(ξ2LB )2 (c1 + c2 )(n − 1)1/2 + 2 exp − (1 − o(1)) + n exp − 24 3 !   (1 − c1 p)p−1/4 (n − 1)φ(ξ2LB ) + 16d0 exp − + 2 exp − 8 10|λ1 |Bq2 d !     (c1 pn)2θ (c1 pn)2θ c1 (n − 1)1/4 + 4(d + 2) exp − + 8(d + 2) exp − + n exp − . 8Bq2 d 4Bq2 d 3 Using the assumption that p = o(n−1+1/(5+8θ) ) for some θ < 14 implies that n1/4 − ln(n) = ω((c1 pn)2θ ) and p−1/4 = ω((c1 pn)2θ ), such that   2   
ξ2U B c3 pn2 φ(ξ2LB )2 (c1 pn)2θ + exp − P ¬A15 = O d exp − (1 − o(1)) 2 u,v,2 24 8Bq d   2LB ) + exp − (n−1)φ(ξ . 8 Furthermore, φ(ξ) ≥ min(`, ξ/2L) also implies that (n−1)φ(ξ2LB ) = ω((c1 pn)2θ ) 2 2 2θ and c3 pn2 ξ2U B φ(ξ2LB ) = ω((c1 pn) ), such that P ¬A15 u,v,2
   (c1 pn)2θ . = O d exp − 8Bq2 d APPENDIX B: PROOF OF MAIN RESULTS In this section we combine the Lemmas to show the Theorems presented in Section 4. Theorem 4.1 follows from combining Lemmas A.34 and A.32 to obtain high probability error bounds for the estimate produced by dist1 . Theorem 4.2 follows from combining Lemmas A.35 and A.33 to obtain high probability error bounds for the estimate produced by dist2 . B.1. Bounding the Max Error. If we want to bound the max error as in Theorems 4.3 and 4.4, i.e. show that for all vertices the error is bounded, the events we would need to hold (when using dist1 ) would be 13 12 11 10 9 5 4 ∩(u,v)∈[n]×[n] (A14 u,v,1 ∩ Au,v,1 ∩ Au,v,1 ) ∩ A ∩u∈[n] (Au,ξ1LB ∩ Au,r,1 ∩ Au,r+1 ) ∩ A 110 BORGS-CHAYES-LEE-SHAH If these events hold, then for all (u, v) ∈ [n] × [n], √ |F̂ (u, v) − f (αu , αv )| = O(Bq dξ1U B ) = O 2 (F̂ (u, v) − f (αu , αv ))2 = O(Bq2 dξ1U B) = O   |λ1 | |λd | |λ1 | |λd | r  2r 1/2 ! Bq3 d2 |λ1 | 1 (c1 pn) 2 −θ Bq3 d2 |λ1 | 1 (c1 pn) 2 −θ  . By using union bound along with Lemmas A.9, A.12, A.20, A.21, A.23, A.26, and A.28, these events hold with probability at least   n(n + 1)  2  2 2 c pn2 ξ1U ξ1U B c3 pn2 φ(ξ1LB )2 B φ(ξ1LB ) 1 − n(n + 1) exp − 3 (1 − o(1)) − exp − (1 − o(1)) 16 16  2  2   ξ1U B c3 pn2 φ(ξ1LB )2 (c1 +c2 )(n−1)1/2 − 2n(n + 1) exp − + n exp − (1 − o(1)) 24 3 !   −1/4 (n − 1)φ(ξ1LB ) (1 − c1 p)p − n exp − − 8n exp − 8 10|λ1 |Bq2 d !   c1 (n − 1)1/4 (c1 pn)2θ − n exp − − 4n(d + 2) exp − 4Bq2 d 3    2     ξ c pn2 φ(ξ )2 (c1 pn)2θ 1LB ) . = 1 − O n2 exp − 1U B 3 24 1LB (1 − o(1)) + n exp − (n−1)φ(ξ + nd exp − 2 8 4B d q Note that since we have n and n2 in the coefficient, we need the exponential term to decay sufficiently fast to guarantee that this probability converges to zero as n goes to infinity. If φ(ξ) ≥ min(`, ξ/2L) and (c1 pn)2θ = ω(4Bq2 d log(nd)), then the second to last term dominates and the probability of error reduces to       (c1 pn)2θ (c1 pn)2θ O n(d + 2) exp − = O (d + 2) exp − . 4Bq2 d 5Bq2 d An equivalent result and proof holds for bounding the max error for estimates computed by dist2 . B.2. Using Subsampled Anchor Vertices. Recall that in Section 3.3, we discussed a modification of the algorithm to reduce computation by subsampling for anchor vertices, and comparing only to anchor vertices
rather than computing all n2 pairwise distances. In order to prove error bounds for the modified algorithm, we need to ensure that there exists an anchor vertex that is within close distance from the target vertex u. Let K denote the set of anchor vertices. For some location (u, v) we need to first prove that with high probability there are anchor vertices within a 111 ITERATIVE COLLABORATIVE FILTERING “close” distance to u and v. Then we need the distance estimates between the vertices u, v and the anchor vertices to be close such that the anchor vertices π(u) and π(v) which minimize dist2 (u, π(u)) and dist2 (v, π(v)) will also be close in terms of kΛQ(eu − eπ(u) )k2 and kΛQ(ev − eπ(v) )k2 . Finally, since the algorithm estimates F̂ (u, v) = F̂ (π(u), π(v)), it only remains to show that |F̂ (π(u), π(v)) − f (απ(u) , απ(v) )| is bounded, which follows from the proof which we showed before. Define event A0u,ξ := {min kΛQ(ei − eu )k22 ≤ ξ 2 } i∈K Lemma B.1. P(¬A0u,ξ ) ≤ exp(−|K|φ(ξ)). Proof. Because the set K is sampled at random amongst the vertices, the latent variables are sampled i.i.d. from P1 . Recall that by definition, φ(ξ) lower bounds the probability that a randomly sampled vertex satisfies kΛQ(ei − eu )k22 ≤ ξ 2 . Therefore, P(¬A0u,ξ ) = P(∩i∈K {kΛQ(ei − eu )k22 ≥ ξ 2 }) = P1 (kΛQ(ei − eu )k22 ≥ ξ 2 )|K| = (1 − P1 (kΛQ(ei − eu )k22 ≤ ξ 2 ))|K| ≤ (1 − φ(ξ))|K| ≤ exp(−|K|φ(ξ)). Proof of Theorem 4.5. Conditioned on the event that A4 ∩A5u,r+d0 ∩l∈K (A5l,r+d0 ∩i∈[d0 ] A8l,u,r,r+i ), we proved in Lemma A.18 that for all l ∈ K, 1 |dist2 (u, l) − kΛQ(eu − el )k22 | ≤ 32Bq d|λ1 |(c1 pn)− 2 +θ . Let π(u) denote arg mini∈K dist2 (i, u), and let π ∗ (u) denote arg mini∈K kΛQ(ei − eu )k22 . If we additionally condition on A0u,ξ , by definition kΛQ(eπ∗ (u) −eu )k22 ≤ ξ 2 . It follows that 1 kΛQ(eu − eπ(u) )k22 − 32Bq d|λ1 |(c1 pn)− 2 +θ ≤ dist2 (π(u), u) ≤ dist2 (π ∗ (u), u) 1 ≤ kΛQ(eu − eπ∗ (u) )k22 + 32Bq d|λ1 |(c1 pn)− 2 +θ 1 ≤ ξ 2 + 32Bq d|λ1 |(c1 pn)− 2 +θ . 112 BORGS-CHAYES-LEE-SHAH Therefore, 1 kΛQ(eu − eπ(u) )k22 ≤ ξ 2 + 64Bq d|λ1 |(c1 pn)− 2 +θ . Similarly, conditioned on A0v,ξ ∩ A4 ∩ A5v,r+d0 ∩l∈K (A5l,r+d0 ∩i∈[d0 ] A8l,v,r,r+i ), it follows that 1 kΛQ(ev − eπ(v) )k22 ≤ ξ 2 + 64Bq d|λ1 |(c1 pn)− 2 +θ . Therefore, conditioned on A0u,ξ ∩A0v,ξ ∩A4 ∩A5u,r+d0 ∩A5v,r+d0 ∩l∈K (A5l,r+d0 ∩i∈[d0 ] (A8l,u,r,r+i ∩ A8l,v,r,r+i )), |f (απ(u) , απ(v) ) − f (αu , αv )| = |eTπ(u) QT ΛQeπ(v) − eTu QT ΛQev | √
≤ Bq d kΛQ(eπ(u) − eu )k2 + kΛQ(eπ(v) − ev )k2 1/2 √  1 ≤ 2Bq d ξ 2 + 64Bq d|λ1 |(c1 pn)− 2 +θ . Conditioned on A15 π(u),π(v),2 , by Lemma A.33, the error of the estimated output when using dist2 is bounded by  1/2 ! √ Bq3 d2 |λ1 | |F̂ (π(u), π(v)) − f (απ(u) , απ(v) )| = O(Bq dξ2U B ) = O , 1 −θ (c1 pn) 2 2 (F̂ (π(u), π(v)) − f (απ(u) , απ(v) )) = 2 O(Bq2 dξ2U B)  =O Bq3 d2 |λ1 | 1 (c1 pn) 2 −θ  . Using the modification of the algorithm as specified in Section 3.3, we estimate F̂ (u, v) = F̂ (π(u), π(v)). Therefore, |F̂ (u, v) − f (αu , αv )| ≤ |F̂ (π(u), π(v)) − f (απ(u) , απ(v) )| + |f (απ(u) , απ(v) ) − f (αu , αv )| 1/2 √  √ 1 ≤ O(Bq dξ2U B ) + 2Bq d ξ 2 + 64Bq d|λ1 |(c1 pn)− 2 +θ √ = O(Bq d min(ξ2U B , ξ)). 0 0 5 5 The event that this holds is A15 π(u),π(v),2 ∩Au,ξ ∩Av,ξ ∩Au,r+d0 ∩Av,r+d0 ∩l∈K (A5l,r+d0 ∩i∈[d0 ] (A8l,u,r,r+i ∩ A8l,v,r,r+i )), which by Lemmas A.35, A.19, A.12, ITERATIVE COLLABORATIVE FILTERING 113 B.1, is bounded above by 0 0 5 5 5 8 8 P(¬A15 π(u),π(v),2 ∪ ¬Au,ξ ∪ ¬Av,ξ ∪ ¬Au,r+d0 ∪ ¬Av,r+d0 ∪l∈K (¬Al,r+d0 ∪i∈[d0 ] (¬Al,u,r,r+i ∪ ¬Al,v,r,r+i ))) 0 0 5 5 5 ≤ P(¬A15 π(u),π(v),2 ) + P(¬Au,ξ ) + P(¬Av,ξ ) + P(¬Au,r+d0 ∪ ¬Av,r+d0 ∪l∈K ¬Al,r+d0 ) + P(∪l∈K ∪i∈[d0 ] (¬A8l,u,r,r+i ∪ ¬A8l,v,r,r+i ) | A5u,r+d0 ∩ A5v,r+d0 ∩l∈K A5l,r+d0 ) X 0 0 5 5 ≤ P(¬A15 ) + P(¬A ) + P(¬A ) + P(¬A ) + P(¬A ) + P(¬A5l,r+d0 ) 0 0 u,ξ v,ξ u,r+d v,r+d π(u),π(v),2 l∈K + X + X P(∪i∈[d0 ] ¬A8l,u,r,r+i | A5u,r+d0 ∩ A5v,r+d0 ∩l∈K A5l,r+d0 ) l∈K P(∪i∈[d0 ] ¬A8l,v,r,r+i | A5u,r+d0 ∩ A5v,r+d0 ∩l∈K A5l,r+d0 ) l∈K ! (1 − c1 p)p−1/4 ≤ + 2 exp(−|K|φ(ξ)) + 16|K|d exp − 5|λ1 |Bq2 d   (c1 pn)2θ + 4(|K| + 2)(d + 2) exp − 4Bq2 d    2  2 2 c3 pn2 ξ2U ξ2U B c3 pn2 φ(ξ2LB )2 B φ(ξ2LB ) ≤ 2 exp − (1 − o(1)) + exp − (1 − o(1)) 16 16 !   2 ξ2U B c3 pn2 φ(ξ2LB )2 (c1 + c2 )(n − 1)1/2 (1 − o(1)) + n exp − + 2 exp − 24 3 !   (n − 1)φ(ξ2LB ) (1 − c1 p)p−1/4 + 2 exp − + 16d0 exp − 8 10|λ1 |Bq2 d !     (c1 pn)2θ (c1 pn)2θ c1 (n − 1)1/4 + 8(d + 2) exp − + n exp − + 4(d + 2) exp − 8Bq2 d 4Bq2 d 3 ! (1 − c1 p)p−1/4 2 exp(−|K|φ(ξ)) + 16|K|d0 exp − 5|λ1 |Bq2 d   (c1 pn)2θ + 4(|K| + 2)(d + 2) exp − 4Bq2 d     2  ξ2U B c3 pn2 φ(ξ2LB )2 (c1 pn)2θ + exp − (1 − o(1)) . = O exp(−|K|φ(ξ)) + |K|d exp − 4Bq2 d 24 P(¬A15 π(u),π(v),2 ) 0 If we want to bound the maximum error, we need to in addition bound the event that all vertices are close to at least one anchor vertex. In order to prove this event holds with high probability, we will show that we can 114 BORGS-CHAYES-LEE-SHAH bound the number of balls of diameter ξ needed to cover the space X1 with respect to the measure P1 . Lemma B.2. There exist a set of H disjoint subsets of the latent space X1 , denoted Y1 . . . YH , such that the measure of the union of the sets is 1, the number of subsets is bounded by φ(ξ/4)−1 , the measure of each subset is bounded below by φ(ξ/4), and the diameter of each subset is bounded above by ξ. In mathematical notation, 1. 2. 3. 4. P1 (∪i∈[L] Yi ) = 1, H ≤ φ(ξ/4)−1 , P1 (Yi ) ≥ φ(ξ/4) for all i ∈ [H], For all i ∈ [H], for any a, b ∈ Yi , kΛQ(ea − eb )k22 ≤ ξ 2 . Proof. We will prove the existence of this set by constructing it inductively. Let B(α0 , ξ/2) denote the ball of radius ξ/2 centered at point α0 . ( ) Z 1/2 ξ 2 B(α0 , ξ/2) := α ∈ X1 s.t. (f (α, y) − f (α0 , y)) dP1 (y) ≤ 2 X1 By definition, the diameter of B(α0 , ξ/2) is at most ξ. Let C ⊂ X1 denote a finite set of “centers” which we will use to inductively construct the subsets. First let C = ∅. As long as P1 (∪t∈C B(t, ξ/2)) < 1, there must exist some point α ∈ X1 \ (∪t∈C B(t, ξ/2)) such that P1 (B(α, ξ/4)) ≥ φ(ξ/4) by definition of the function φ. Additionally, by the triangle inequality, we can guarantee that for this choice of α, the ball B(α, ξ/4) is disjoint/nonoverlapping with the set ∪t∈C B(t, ξ/4), otherwise α would have to be within distance ξ/2 of one of the centers. We then add α to the set C and repeat the process. For each element that we add to C, P1 (∪t∈C B(t, ξ/2)) decreases by at least φ(ξ/4), such that after choosing at most φ(ξ/4)−1 center, we will have covered the space. Once this process finishes, we construct the disjoint subsets by defining Yi to be all points that are within distance ξ/2 of the i-th center in C for which the i-th center is the closest point within C. We can verify that by construction, the four desired properties are satisfied. Lemma B.3. For |K|φ(ξ/4)2 ≥ 2, P(∪u∈[n] ¬A0u,ξ )  |K|φ(ξ/4) ≤ exp − 8  Proof. Lemma B.2 proved that there exists a set of subsets Y1 . . . YH , each of diameter at most ξ, which cover measure 1 of the latent space. ITERATIVE COLLABORATIVE FILTERING 115 Therefore, if we can ensure that there is at least one anchor vertex in each of the H subsets, then it follows that P1 (∪i∈K B(i, ξ)) = 1, since B(t, ξ/2) ⊂ B(i, ξ) for i ∈ B(t, ξ/2). Next we bound the probability that amongst the |K| randomly chosen vertices, there is at least one in each of the H subsets. Consider iteratively drawing each random anchor vertex. If there is some subset Yt which does not yet contain an anchor vertex, the probability of randomly choosing a vertex in Yt is at least φ(ξ/4), since by construction P1 (Yt ) ≥ φ(ξ/4). Therefore, for each newly sampled anchor vertex, the probability that it is a member of an unrepresented subset is at least φ(ξ/4) as long as there still exists subsets to be represented. Thus the number of distinct subsets that the anchor vertices cover stochastically dominates a Binomial random variable with parameters |K| (the number of coupons drawn) and φ(ξ/4). The probability that all H subsets are represented after |K| randomly chosen vertices, is lower bounded by the probability that a Binomial(|K|, φ(ξ/4)) random variable is larger than or equal to H, which is bounded above by φ(ξ/4)−1 . By using Chernoff Hoeffding’s inequality for sums of Bernoulli random variables, it follows that P(∩t∈[L] {∃i ∈ Ks.t.i ∈ Yt }) ≥ P(Bin(|K|, φ(ξ/4)) ≥ H) ≥ P(Bin(|K|, φ(ξ/4)) ≥ φ(ξ/4)−1 )   (|K|φ(ξ/4) − φ(ξ/4)−1 )2 ≥ 1 − exp − 2φ(ξ/4)|K|   (|K|φ(ξ/4)2 − 1)2 = 1 − exp − . 2φ(ξ/4)3 |K| If we condition on choosing |K| such that |K|φ(ξ/4)2 ≥ 2, then it follows that   |K|2 φ(ξ/4)4 −1 P(Bin(|K|, φ(ξ/4)) ≥ φ(ξ/4) ) ≥ 1 − exp − 8φ(ξ/4)3 |K|   |K|φ(ξ/4) = 1 − exp − . 8 Therefore this probability is bounded by 1 − ε for     2 1 1 |K| ≥ max 4 ln , . φ(ξ/4) ε φ(ξ/4) 116 BORGS-CHAYES-LEE-SHAH Proof of Theorem 4.6. If we wanted the maximum error to be bounded, we need to show the following events hold 13 12 11 10 4 0 5 8 ∩(a,b)∈K×K (A14 a,b,2 ∩ Aa,b,2 ∩ Aa,b,2 ) ∩ A ∩a∈K Aa,ξ2LB ∩ A ∩u∈[n] (Au,ξ ∩ Au,r+d0 ∩l∈K,i∈[d0 ] Al,u,r,r+i ) If these events hold, as shown in Theorem 4.5, for all (u, v) ∈ [n] × [n], √ |F̂ (u, v) − f (αu , αv )| = O(Bq d min(ξ2U B , ξ)). By Lemmas B.3, A.9, A.12, A.19, A.21, A.23, A.27, and A.29, for |K|φ(ξ/4)2 ≥ 2, these events hold with probability at least    2  2 2 c pn2 ξ2U ξ2U B c3 pn2 φ(ξ2LB )2 |K|(|K|+1) B φ(ξ2LB ) 1 − |K|(|K| + 1) exp − 3 (1 − o(1)) − exp − (1 − o(1)) 16 2 16 !  2  2 2 ξ2U B c3 pn φ(ξ2LB ) (c1 + c2 )(n − 1)1/2 − 2|K|(|K| + 1) exp − (1 − o(1)) + n exp − 24 3 !   |K|φ(ξ/4) (1 − c1 p)p−1/4 (n − 1)φ(ξ2LB ) 0 − exp(− ) − 8n|K|d exp − − |K| exp − 8 8 5|λ1 |Bq2 d !   c1 (n − 1)1/4 (c1 pn)2θ − n exp − − 4n(d + 2) exp − 4Bq2 d 3       2  ξ2U B c3 pn2 φ(ξ2LB )2 (c1 pn)2θ |K|φ(ξ/4) 2 = 1 − O nd exp − + exp − + |K| exp − (1 − o(1)) . 4Bq2 d 8 24 

An Encoding for Order-Preserving Matching Travis Gagie1 , Giovanni Manzini2 , and Rossano Venturini3 1 2 arXiv:1610.02865v2 [cs.DS] 17 Feb 2017 3 School of Computer Science and Telecommunications, Diego Portales University and CEBIB, Santiago, Chile travis.gagie@mail.udp.cl Computer Science Institute, University of Eastern Piedmont, Alessandria, Italy and IIT-CNR, Pisa, Italy giovanni.manzini@uniupo.it Department of Computer Science, University of Pisa, Pisa, Italy and ISTI-CNR, Pisa, Italy rossano.venturini@unipi.it Abstract Encoding data structures store enough information to answer the queries they are meant to support but not enough to recover their underlying datasets. In this paper we give the first encoding data structure for the challenging problem of order-preserving pattern matching. This problem was introduced only a few years ago but has already attracted significant attention because of its applications in data analysis. Two strings are said to be an order-preserving match if the relative order of their characters is the same: e.g., 4, 1, 3, 2 and 10, 3, 7, 5 are an orderpreserving match. We show how, given a string S[1..n] over an arbitrary alphabet and a constant c ≥ 1, we can build an O(n log log n)-bit encoding such that later, given a pattern P [1..m] with m ≤ logc n, we can return the number of order-preserving occurrences of P in S in O(m) time. Within the same time bound we can also return the starting position of some order-preserving match for P in S (if such a match exists). We prove that our space bound is within a constant factor of optimal; our query time is optimal if log σ = Ω(log n). Our space bound contrasts with the Ω(n log n) bits needed in the worst case to store S itself, an index for order-preserving pattern matching with no restrictions on the pattern length, or an index for standard pattern matching even with restrictions on the pattern length. Moreover, we can build our encoding knowing only how each character compares to O(logc n) neighbouring characters. 1998 ACM Subject Classification E.1 Data Structures; F.2.2 Nonnumerical Algorithms and Problems; H.3 Information Storage and Retrieval. Keywords and phrases Compact data structures; encodings; order-preserving matching. Digital Object Identifier 10.4230/LIPIcs.CVIT.2016.23 1 Introduction As datasets have grown even faster than computer memories, researchers have designed increasingly space-efficient data structures. We can now store a sequence of n numbers from {1, . . . , σ} with σ ≤ n in about n words, and sometimes n log σ bits, and sometimes even nH bits, where H is the empirical entropy of the sequence, and still support many powerful queries quickly. If we are interested only in queries of the form “what is the position of the smallest number between the ith and jth?”, however, we can do even better: regardless of σ or H, we need store only 2n + o(n) bits to be able to answer in constant time [19]. Such a data structure, that stores enough information to answer the queries it is meant to support but not enough to recover the underlying dataset, is called an encoding [37]. As well © Travis Gagie, Giovanni Manzini and Rossano Venturini; licensed under Creative Commons License CC-BY 42nd Conference on Very Important Topics (CVIT 2016). Editors: John Q. Open and Joan R. Acces; Article No. 23; pp. 23:1–23:14 Leibniz International Proceedings in Informatics Schloss Dagstuhl – Leibniz-Zentrum für Informatik, Dagstuhl Publishing, Germany 23:2 An Encoding for Order-Preserving Matching as the variant of range-minimum queries mentioned above, there are now efficient encoding data structures for range top-k [12, 22, 25], range selection [33], range majority [34], range maximum-segment-sum [21] and range nearest-larger-value [18] on sequences of numbers, and range-minimum [24] and range nearest-larger-value [29, 30] on two-dimensional arrays of numbers; all of these queries return positions but not values from the sequence or array. Perhaps Orlandi and Venturini’s [35] results about sublinear-sized data structures for substring occurrence estimation are the closest to the ones we present in this paper, in that they are more related to pattern matching than range queries: they showed how we can store a sequence of n numbers from {1, . . . , σ} in significantly less than n log σ bits but such that we can estimate quickly and well how often any pattern occurs in the sequence. Encoding data structures can offer better space bounds than traditional data structures that store the underlying dataset somehow (even in succinct or compressed form), and possibly even security guarantees: if we can build an encoding data structure using only public information, then we need not worry about it being reverse-engineered to reveal private information. From the theoretical point of view, encoding data structures pose new interesting combinatorial problems and promise to be a challenging field for future research. In this paper we give the first encoding for order-preserving pattern matching, which asks us to search in a text for substrings whose characters have the same relative order as those in a pattern. For example, in 6, 3, 9, 2, 7, 5, 4, 8, 1, the order-preserving matches of 2, 1, 3 are 6, 3, 9 and 5, 4, 8. Kubica et al. [32] and Kim et al. [31] formally introduced this problem and gave efficient online algorithms for it. Other researchers have continued their investigation, and we briefly survey their results in Section 2. As well as its theoretical interest, this problem has practical applications in data analysis. For example, mining for correlations in large datasets is complicated by amplification or damping — e.g., the euro fluctuating against the dollar may cause the pound to fluctuate similarly a few days later, but to a greater or lesser extent — and if we search only for sequences of values that rise or fall by exactly the same amount at each step we are likely to miss many potentially interesting leads. In such settings, searching for sequences in which only the relative order of the values is constrained to be the same is certainly more robust. In Section 2 we review some previous work on order-preserving pattern matching. In Section 3 we review the algorithmic tools we use in the rest of the paper. In Section 4 we prove our first result showing how, given a string S[1..n] over an arbitrary alphabet [σ] and a constant c ≥ 1, we can store O(n log log n) bits — regardless of σ — such that later, given a pattern P [1..m] with m < logc n, in O(n logc n) time we can scan our encoding and report all the order-preserving matches of P in S. Our space bound contrasts with the Ω(n log n) bits needed in the worst case, when log σ = Ω(log n), to store S itself, an index for order-preserving pattern matching with no restriction on the pattern length, or an index for standard pattern matching even with restrictions on the pattern length. (If S is a permutation then we can recover it from an index for unrestricted order-preserving pattern matching, or from an index for standard matching of patterns of length 2, even when they do not report the positions of the matches. Notice this does not contradict Orlandi and Venturini’s result, mentioned above, about estimating substring frequency, since that permits additive error.) In fact, we build our representation of S knowing only how each character compares to 2 logc n neighbouring characters. We show in Section 5 how to adapt and build on this representation to obtain indexed
order-preserving pattern matching, instead of scan-based, allowing queries in O m log3 n time but now reporting the position of only one match. In Section 6 we give our main result showing how to speed up our index using weak prefix T. Gagie, G. Manzini and R. Venturini 23:3 search and other algorithmic improvements. The final index is able to count the number of occurrences and return the position of an order-preserving match (if one exists) in O(m) time. This query time is optimal if log σ = Ω(log n). Finally, in Section 7 we show that our space bound is optimal (up to constant factors) even for data structures that only return whether or not S contains any order-preserving matches. 2 Previous Work Although recently introduced, order-preserving pattern matching has received considerable attention and has been studied in different settings. For the online problem, where the pattern is given in advance, the first contributions were inspired by the classical KnuthMorris-Pratt and Boyer-Moore algorithms [3, 10, 31, 32]. The proposed algorithms have guaranteed linear time worst-case complexity or sublinear time average complexity. However, for the online problem the best results in practice are obtained by algorithms based on the concept of filtration, in which some sort of “order-preserving” fingerprint is applied to the text and the pattern [4, 5, 6, 8, 9, 16, 13]. This approach was successfully applied also to the harder problem of matching with errors [6, 23, 27]. There has also been work on indexed order-preserving pattern matching. Crochemore et al. [11] showed how, given a string S[1..n], in O(n log(n)/ log log n) time we can build an O(n log n)-bit index such that later, given a pattern P [1..m], we can return the starting positions of all the occ order-preserving matches of P in S in optimal O(m + occ) time. Their index is a kind of suffix tree, and other researchers [38] are trying to reduce the space bound to n log σ + o(n log σ) bits, where σ is the size of the alphabet of S, by using a kind of Burrow-Wheeler Transform instead (similar to [20]). Even if they succeed, however, when σ = nΩ(1) the resulting index will still take linear space — i.e., Ω(n) words or Ω(n log n) bits. In addition to Crochemore et al.’s result, other offline solutions have been proposed combining the idea of fingerprint and indexing. Chhabra et al. [7] showed how to speed up the search by building an FM-index [17] on the binary string expressing whether in the input text each element is smaller or larger than the next one. By expanding this approach, Decaroli et al. [13] show how to build a compressed file format supporting order-preserving matching without the need of full decompression. Experiments show that this compressed file format takes roughly the same space as gzip and that in most cases the search is orders of magnitude faster than the sequential scan of the text. We point out that these approaches, although interesting for the applications, do not have competitive worst case bounds on the search cost as we get from Crochemore et al.’s and in this paper. 3 Background In this section we collect a set of algorithmic tools that will be used in our solutions. In the following we report each result together with a brief description of the solved problem. More details can be obtained by consulting the corresponding references. All the results hold in the unit cost word-RAM model, where each memory word has size w = Ω(log n) bits, where n is the input size. In this model arithmetic and boolean operations between memory words require O(1) time. Rank queries on binary vector. In the next solutions we will need to support Rank queries on a binary vector B[1..n]. Given an index i, Rank(i) on B returns the number of 1s in the prefix B[1..i]. We report here a result in [28]. CVIT 2016 23:4 An Encoding for Order-Preserving Matching I Theorem 1. Given a binary vector B[1..n], we can support Rank queries in constant time by using n + o(n) bits of space. Elias-Fano representation. In the following we will need to encode an increasing sequence of values in almost optimal space. There are several solutions to this problem, we report here the result obtained with the, so-called, Elias-Fano representation [14, 15]. I Theorem 2. An increasing sequence of n values up to u can be represented by using
log nu + O(n) = n log nu + O(n) bits, so that we can access any value of the sequence in constant time. Minimal perfect hash functions. In our solution we will make use of Minimal perfect hash functions (Mphf) [26] and Monotone minimal perfect hash functions (Mmphf) [1]. Given a subset of S = {x1 , x2 , . . . , xn } ⊆ U of size n, a minimal perfect hash function has to injectively map keys in S to the integers in [n]. Hagerup and Tholey [26] show how to build a space/time optimal minimal perfect hash function as stated by the following theorem. I Theorem 3. Given a subset of S ⊆ U of size n, there is a minimal perfect hash function for S that can be evaluated in constant time and requires n log e + o(n) bits of space. A monotone minimal perfect hash function is a Mphf h() that preserves the lexicographic ordering, i.e., for any two strings x and y in the set, x ≤ y if and only if h(x) ≤ h(y). Results on Mmphfs focus their attention on dictionaries of binary strings [1]. The results can be easily generalized to dictionaries with strings over larger alphabets. The following theorem reports the obvious generalization of Theorem 3.1 in [1] and Theorem 2 in [2]. I Theorem 4. Given a dictionary of n strings drawn from the alphabet [σ], there is a monotone minimal perfect hash function h() that occupies O(n log(` log σ)) bits of space, where ` is the average length of the strings in the dictionary. Given a string P [1..m], h(P ) is computed in O(1 + m log σ/w) time. Weak prefix search. The Prefix Search Problem is a well-known problem in data-structure design for strings. It asks for the preprocessing of a given set of n strings in such a way that, given a query-pattern P , (the lexicographic range of) all the strings in the dictionary which have P as a prefix can be returned efficiently in time and space. Belazzougui et al. [2] introduced the weak variant of the problem that allows for a onesided error in the answer. Indeed, in the Weak Prefix Search Problem the answer to a query is required to be correct only in the case that P is a prefix of at least one string in dictionary; otherwise, the algorithm returns an arbitrary answer. Due to these relaxed requirements, the data structures solving the problem are allowed to use space sublinear in the total length of the indexed strings. Belazzougui et al. [2] focus their attention on dictionaries of binary strings, but their results can be easily generalized to dictionaries with strings over larger alphabets. The following theorem states the obvious generalization of Theorem 5 in [2]. I Theorem 5. Given a dictionary of n strings drawn from the alphabet [σ], there exists a data structure that weak prefix searches for a pattern P [1..m] in O(m log σ/w + log(m log σ)) time. The data structure uses O(n log(` log σ)) bits of space, where ` is the average length of the strings in the dictionary. T. Gagie, G. Manzini and R. Venturini 23:5 We remark that the space bound in [2] is better than the one reported above as it is stated in terms of the hollow trie size of the indexed dictionary. This measure is always within O(n log `) bits but it may be much better depending on the dictionary. However, the weaker space bound suffices for the aims of this paper. 4 An Encoding for Scan-Based Search As an introduction to our techniques, we show an O(n log log n) bit encoding supporting scan-based order-preserving matching. Given a sequence S[1..n] we define the rank encoding E(S)[1..n] as E(S)[i] =    0.5              j     if S[i] is lexicographically smaller than any character in {S[1], . . . , S[i − 1]},       j + 0.5              |{S[1], . . . , S[i − 1]}| + 0.5 if S[i] is larger than the lexicographically jth character in {S[1], . . . , S[i − 1]} but smaller than the lexicographically (j + 1)st, if S[i] is equal to the lexicographically jth character in {S[1], . . . , S[i − 1]}, if S[i] is lexicographically larger than any character in {S[1], . . . , S[i − 1]}. This is similar to the representations used in previous papers on order-preserving matching. We can build E(S) in O(n log n) time. However, we would ideally need E(S[i..n]) for i = 1, . . . , n, since P [1..m] has an order-preserving match in S[i..i + m − 1] if and only if E(P ) = E(S[i..i + m − 1]). Assuming P has polylogarithmic size, we can devise a more space efficient encoding. I Lemma 6. Given S[1..n] and a constant c ≥ 1 let ` = logc n. We can store O(n log log n) bits such that later, given i and m ≤ `, we can compute E(S[i..i + m − 1]) in O(m) time. Proof. For every position i in S which is multiple of ` = logc n, we store the ranks of the characters in the window S[i..i + 2`]. The ranks are values at most 2`, thus they are stored in O(log `) bits each. We concatenate the ranks of each window in a vector V , which has length O(n) and takes O(n log `) bits. Every range S[i..i + m − 1] of length m ≤ ` is fully contained in at least one window and in constant time we can convert i into i0 such that V [i0 ..i0 + m − 1] contains the ranks of S[i], . . . , S[i + m − 1] in that window. Computing E(S[i..i + m − 1]) naïvely from these ranks would take O(m log m) time. We can speed up this computation by exploiting the fact that S[i..i + m − 1] has polylogaritmic length. Indeed, a recent result [36] introduces a data structure to represent a small dynamic set S of O(wc ) integers of w bits each supporting, among the others, insertions and rank queries in O(1) time. Given an integer x, the rank of x is the number of integers in S that are smaller than or equal to x. All operations are supported in constant time for sets of size O(wc ). This result allows us to compute E(S[i..i + m − 1]) in O(m) time. Indeed, we can use the above data structure to insert S[i..i + m − 1]’s characters one after the other and compute their ranks in constant time. J It follows from Lemma 6 that given S and c, we can store an O(n log log n)-bit encoding of S such that later, given a pattern P [1..m] with m ≤ logc n, we can compute E(S[i..i+m−1]) CVIT 2016 23:6 An Encoding for Order-Preserving Matching for each position i in turn and compare it to E(P ), and thus find all the order-preserving matches of P in O(nm) time. (It is possible to speed this scan-based algorithm up by avoiding computing each E(S[i..i+m−1]) from scratch but, since this is only an intermediate result, we do not pursue it further here.) We note that we can construct the encoding in Lemma 6 knowing only how each character of S compares to O(logc n) neighbouring characters. I Corollary 7. Given S[1..n] and a constant c ≥ 1, we can store an encoding of S in O(n log log n) bits such that later, given a pattern P [1..m] with m ≤ logc n, we can find all the order-preserving matches of P in S in O(nm) time. We will not use Corollary 7 in the rest of this paper, but we state it as a baseline easily proven from Lemma 6. 5 Adding an Index to the Encoding Suppose we are given S[1..n] and a constant c ≥ 1. We build the O(n log log n)-bit encoding of Lemma 6 for ` = logc n + log n and call it S` . Using S` we can compute E(S 0 ) for any substring S 0 of S of length |S 0 | ≤ ` in O(|S 0 |) time. We now show how to complement S` with a kind of “sampled suffix array” using O(n log log n) more bits, such that we can search for a pattern P [1..m] with m ≤ logc n and return the starting position of
an order-preserving match for P in S, if there is one. Out first solution has O m log3 n query time; we will improve the query time to O(m) in the next section. We define the rank-encoded suffix array R[1..n] of S such that R[i] = j if E(S[i..n]) is the lexicographically jth string in {E(S[1..n]), E(S[2..n]), . . . , E(S[n])}. Note that E(S[i..n]) has length n − i + 1. Figure 1 shows an example. Our algorithm consists of a searching phase followed by a verification phase. The goal of the searching phase is to identify a range [l, r] in R which contains all the encodings prefixed by E(P ), if any, or an arbitrary interval if P does not occur. The verification phase has to check if there is at least an occurrence of P in this interval, and return one position at which P occurs. Searching phase. Similarly to how we can use a normal suffix array and S to support normal pattern matching, we could use R and S to find all order-preserving matches for a pattern P [1..m] in O(m log n) time via binary search, i.e., at each step we choose an index i, extract S[R[i]..R[i] + m − 1], compute its rank encoding and compare it to E(P ), all in O(m) time. If m ≤ ` we can compute E(S[R[i]..R[i] + m − 1]) using S` instead of S, still in O(m) time, but storing R still takes Ω(n log n) bits. Therefore, for our searching phase we sample and store only every sample-th element of R, by position, and every element of R equal 1 or n or a multiple of sample, where sample = blog n/ log log nc. This takes O(n log log n) bits. Notice we can still find in O(m log n) time via binary search in the sampled R an order-preserving match for any pattern P [1..m] that has at least sample order-preserving matches in S. If P has fewer than sample orderpreserving matches in S but we happen to have sampled a cell of R pointing to the starting position of one of those matches, then our binary search still finds it. Otherwise, we find an interval of length at most sample − 1 which contains pointers at least to all the orderpreserving matches for P in S; on this interval we perform the verification phase. Verification phase. The verification phase receives a range R[l, r] (although R is not stored completely) and has to check if that range contains the starting position of an order T. Gagie, G. Manzini and R. Venturini i 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 R[i] L[i] B[i] D[i] 30 29 22 13 2 23 8 14 20 3 16 24 11 9 15 28 7 19 12 1 21 10 27 6 18 26 17 5 25 4 2 2 1.5 0.5 4 2 2 3 0.5 3.5 1 3 3 3.5 1 2 3 2 0.5 3.5 1.5 3 3 1 3 3 1.5 1.5 4 5 2 2 4 2.5 1.5 1.5 1 2 2 4 4 2 1 0.5 2.5 3 1 3 2 2.5 4 23:7 E(S[R[i]..n]) 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 0.5 1.5 5 5.5 6.5 1 0.5 0.5 1 0.5 1.5 4 4.5 1 4 3.5 3.5 2 3 6 7 7.5 2 0.5 0.5 1.5 2.5 3.5 5.5 2.5 2 5 4 8 4 1 1 0.5 1.5 6 7 1 6 5 4 2 3 6 0.5 0.5 1.5 4.5 5.5 6.5 1 0.5 0.5 2.5 2.5 5.5 3 0.5 1 0.5 1.5 6 7 1 6 5 4 2 3 6 7 7.5 2 0.5 1 0.5 1.5 4 4.5 1 4 3.5 3.5 2 3 6 7 7.5 2 0.5 1.5 1.5 1.5 1.5 2.5 6 7 7.5 2 0.5 1.5 2.5 3.5 5.5 2.5 2 5 4 7.5 4 1 1 0.5 1.5 6 7 1 6 5 4 2 3 6 7 0.5 1.5 3.5 4.5 1 4 3.5 3.5 2 3 6 7 7.5 2 0.5 1.5 3.5 4.5 5.5 1 0.5 2.5 1 0.5 1 0.5 1.5 4 5 1 4 3.5 3.5 2 3 6 7 7.5 2 0.5 2.5 2.5 4.5 3 0.5 1 0.5 1.5 6 7 1 6 5 4 2 3 6 7 7.5 2 1 0.5 1.5 3.5 4.5 1 4 3.5 3.5 2 3 6 7 7.5 2 1.5 0.5 1.5 0.5 0.5 3 2.5 5.5 3 0.5 1 0.5 1.5 6 7 1 6 5 4 2 3 6 7 7.5 2 1.5 0.5 2 1.5 1.5 1.5 2.5 6 7 7.5 2 1.5 1 0.5 1 0.5 1.5 4 4.5 1 4 3.5 3.5 2 3 6 7 7.5 2 1.5 1.5 0.5 2 2.5 3.5 5.5 2.5 2 5 4 8 4 1 1 0.5 1.5 6 7 1 6 5 4 2 3 1.5 1.5 1.5 1.5 2.5 6 6.5 7.5 2 1.5 1.5 3.5 2 0.5 1 0.5 1.5 5 6 1 5 4.5 4 2 3 6 7 7.5 2 1.5 2.5 0.5 1.5 2.5 0.5 0.5 4 3 5.5 3 0.5 1 0.5 1.5 6 7 1 6 5 4 2 3 6 7 7.5 2 1.5 2.5 0.5 3 2.5 2.5 2 2.5 6 7 7.5 2 1.5 2.5 3.5 0.5 1.5 2.5 3.5 1 3 2.5 2.5 2 2.5 6 7 7.5 2 1.5 2.5 3.5 1.5 1 4 3 5.5 3 0.5 1 0.5 1.5 6 7 1 6 5 4 2 3 6 7 7.5 2 1.5 2.5 3.5 4.5 1 1.5 2.5 3.5 4.5 2.5 2 5 4 6.5 4 1 1 0.5 1.5 6 7 1 6 5 4 2 3 6 7 7.5 782 82 6782 2 Figure 1 The rank-encoded suffix array R[1..30] for S[1..30] = 3 9 7 2 3 5 6 8 4 3 6 5 9 5 2 2 0 1 5 6 0 5 4 3 1 2 5 6 7 1, with L[i], B[i] and D[i] computed for sample = 4. Stored values are shown in boldface. preserving match for P and, if so, return its position. This is done by adding auxiliary data structures to the sampled entries of R. Suppose that for each unsampled element R[i] = j we store the following data. the smallest number L[i] (if one exists) such that S[j − 1..j + L[i] − 1] has at most logc n order-preserving matches in S; the rank B[i] = E(S[j−1..j+L[i]−1]rev )[L[i]+1] ≤ L[i]+1/2 of S[j−1] in S[j..j+L[i]−1], where the superscript rev indicates that the string is reversed; the distance D[i] to the cell of R containing j − 1 from the last sampled element x such that E(S[x..x + L[i]]) is lexicographically smaller than E(S[j − 1..j + L[i] − 1]). Figure 1 shows the values in L, B and D for our example. Assume we are given P [1..m] and i and told that S[R[i]..R[i] + m − 1] is an orderpreserving match for P , but we are not told the value R[i] = j. If R[i] is sampled, of course, then we can return j immediately. If L[i] does not exist or is greater than m then P has at least logc n ≥ sample order-preserving matches in S, so we can find one in O(m) time: we consider the sampled values from R that precede and follow R[i] and check with Lemma 6 whether there are order-preserving matches starting at those sampled values. Otherwise, CVIT 2016 23:8 An Encoding for Order-Preserving Matching from L[i], B[i] and P , we can compute E(S[j − 1..j + L[i] − 1]) in O(m log m) time: we take the length-L[i] prefix of P ; if B[i] is an integer, we prepend to P [1..L[i]] a character equal to the lexicographically B[i]th character in that prefix; if B[i] is r + 0.5 for some integer r with 1 ≤ r < L[i], we prepend a character lexicographically between the lexicographically rth and (r + 1)st characters in the prefix; if B[i] = 0.5 or B[i] = L[i] + 0.5, we prepend a character lexicographically smaller or larger than any in the prefix, respectively. We can then find in O(m log n) time the position in R of x, the last sampled element such that E(S[x..x + L[i]]) is lexicographically smaller than E(S[j − 1..j + L[i] − 1]). Adding D[i] to this position gives us the position i0 of j − 1 in
R. Repeating this
procedure until we reach a sampled cell of R takes O m log2 n/ log log n = O m log2 n time, and we can then compute and return j. As the reader may have noticed, the procedure is very similar to how we use backward stepping to locate occurrences of a pattern with an FM-index [17], so we refer to it as a backward step at position i. Even if we do not really know whether S[R[i]..R[i] + m − 1] is an order-preserving match for P , we can still start at the cell R[i] and repeatedly apply this procedure: if we do not find a sampled cell after sample − 1 repetitions, then S[R[i]..R[i] + m − 1] is not an orderpreserving match for P ; if we do, then we add the number of times we have repeated the procedure to the contents of the sampled cell to obtain the contents of R[i] = j. Then, using S` we compute E(S[j..k + m − 1]) in O(m)
time, compare it to E(P ) and, if they are the same, return j. This still takes O m log2 n time. Therefore, after our searching phase, if we find an interval [l, r] of length at most sample − 1 which contains pointers to all the order-preserving matches for P in S (instead of an order-preserving match directly),
then we can check each cell in that interval with this procedure, in a total of O m log3 n time. If R[i] = j is the starting position of an order-preserving match for a pattern P [1..m] with m ≤ logc n that has at most sample order-preserving matches in S, then L[i] ≤ logc n. Moreover, if R[i0 ] = j − 1 then L[i0 ] ≤ logc n + 1 and, more generally, if R[i00 ] = j − t then L[i00 ] ≤ logc n + t. Therefore, we can repeat the stepping procedure described above and find j without ever reading a value in L larger than logc n + log n and, since each value in B is bounded in terms of the corresponding value in L, without ever reading a value in B larger than logc n + log n + 1/2. It follows that we can replace any values in L and B greater than logc n + log n + 1/2 by the flag −1, indicating that we can stop the procedure when we read it. With this modification, each value in L and B takes O(log log n) bits so, since each value in D is less than logc n + log n and also takes O(log log n) bits, L, B and D take a total of O(n log log n) bits. Since also the encoding S` from Lemma 6 with ` = logc n + log n takes O(n log log n) bits, the following intermediate theorem summarizes our results so far. I Theorem 8. Given S[1..n] and a constant c ≥ 1, we can store an encoding of S in
O(n log log n) bits such that later, given a pattern P [1..m] with m ≤ logc n, in O m log3 n time we can return the position of an order-preserving match of P in S (if one exists). A complete search example. Suppose we are searching for order-preserving matches for P = 2 3 1 2 in the string S[1..30] shown in Figure 1. Binary search on R tells us that pointers to all the matches are located in R strictly between R[16] = 28 and R[19] = 12, because E(S[28..30] = E(6 7 1) = 0.5 1.5 0.5 ≺ E(P ) = E(2 3 1 2) = 0.5 1.5 0.5 2 ≺ E(S[12..14]) = E(5 9 5) = 0.5 1.5 1 ; notice R[16] = 28 and R[19] = 12 are stored because 16, 28 and 12 are multiples of sample = 4. T. Gagie, G. Manzini and R. Venturini 23:9 We first check whether R[17] points to an order-preserving match for P . That is, we assume (incorrectly) that it does; we take the first L[17] = 3 characters of P ; and, because B[17] = 1.5, we prepend a character between the lexicographically first and second, say 1.5. This gives us 1.5 2 3 1, whose encoding is 0.5 1.5 2.5 0.5. Another binary search on R shows that R[20] = 1 is the last sampled element x such that E(S[x..x + 3]), in this case 0.5 1.5 1.5 0.5, is lexicographically smaller than 0.5 1.5 2.5 0.5. Adding D[17] = 4 to 20, we would conclude that R[24] = R[17] − 1 (which happens to be true in this case) and that 0.5 1.5 2.5 0.5 is a prefix of E(S[R[24]..n]) (which also happens to be true). Since R[24] = 6 is sampled, however, we compute E(S[7..10]) = 0.5 1.5 0.5 0.5 and, since it is not the same as P ’s encoding, we reject our initial assumption that R[17] points to an order-preserving match for P . We now check whether R[18] points to an order preserving match for P . That is, we assume (correctly this time) that it does; we take the first L[18] = 3 characters of P ; and, because B[18] = 1.5, we prepend a character between the lexicographically first and second, say 1.5. This again gives us 1.5 3 2 1, whose encoding is 0.5 1.5 2.5 0.5. As before, a binary search on R shows that R[20] = 1 is the last sampled element x such that E(S[x..x + 3]) is lexicographically smaller than 0.5 1.5 2.5 0.5. Adding D[18] = 5 to 20, we conclude (correctly) that R[25] = R[18] − 1 and that 0.5 1.5 2.5 0.5 is a prefix of E(S[R[25]..n]) Repeating this procedure with L[25] = 4, B[25] = 1 and D[25] = 3, we build a string with encoding 0.5 1.5 2.5 0.5, say 2 3 4 1, and prepend a character equal to the lexicographically first, 1. This gives us 1 2 3 4 1, whose encoding is 0.5 1.5 2.5 3.5 1. Another binary search shows that R[24] = 6 is the last sampled element x such that E(S[x..x + 4]) is lexicographically smaller than 0.5 1.5 2.5 3.5 1. We conclude (again correctly) that R[27] = R[18] − 2 and that 0.5 1.5 2.5 3.5 1 is a prefix of E(S[R[27]..n]). Finally, repeating this procedure with L[27] = 2, B[27] = 2.5 and D[27] = 3, we build a string with encoding 0.5 1.5, say 1 2, and prepend a character lexicographically greater than any currently in the string, say 3. This gives us 3 1 2, whose encoding is 0.5 0.5 1.5. A final binary search show that R[8] = 14 is the last sampled element x such that E(S[x..x + 2]) is lexicographically smaller than 0.5 0.5 1.5. We conclude (again correctly) that R[11] = R[18] − 3 and that 0.5 0.5 1.5 is a prefix of E(S[R[11]..n]). Since R[11] = 16 is sampled, we compute E(S[19..22]) = 0.5 1.5 0.5 2 and, since it matches P ’s encoding, we indeed report S[19..22] as an order-preserving match for P . 6 Achieving O(m) query time In this section we prove our main result: I Theorem 9. Given S[1..n] and a constant c ≥ 1, we can store an encoding of S in O(n log log n) bits such that later, given a pattern P [1..m] with m ≤ logc n, in O(m) time we can return the position of an order-preserving match of P in S (if one exists). In O(m) time we can also report the total number of order-preserving occurrences of P in S.
Compared to Theorem 8, we improve the query time from O m log3 n to O(m). This is achieved by speeding up several steps of the algorithm described in the previous section. Speeding up pattern’s encoding. Given a pattern P [1..m], the algorithm has to compute its encoding E(P [1..m]). Doing this naïvely as in the previous section would cost O(m log m) time, which is, by itself, larger than our target time complexity. However, since m is polylogarithmic in n, we can speed this up as we sped up the computation of the rankencoding of S[i..i + m − 1] in the proof of Lemma 6, and obtain E(P ) in O(m) time. Indeed, CVIT 2016 23:10 An Encoding for Order-Preserving Matching we can insert P ’s characters one after the other in the data structures of [36] and compute their ranks in constant time. Dealing with short patterns. The approach used by our solution cannot achieve a o(sample) query time. This is because we answer a query by performing Θ(sample) backward steps regardless of the pattern’s length. This means that for very short patterns, namely m = o(sample) = o(log n/ log log n), the solution cannot achieve O(m) query time. However, we can precompute and store the answers of all these short patterns in o(n) bits. Indeed, the encoding of a pattern of length at most m = o(log n/ log log n) is a binary string of √ length o(log n). Thus, there are o( n) possible encodings. For each of these encodings we explicitly store the number of its occurrence and the position of one of them in o(n) bits. From now on, thus, we can safely assume that m = Ω(log n/ log log n). Speeding up searching phase. The searching phase of the previous algorithm has two important drawbacks. First, it costs O(m log n) time and, thus, it is obviously too expensive for our target time complexity. Second, binary searching on the sampled entries in R gives too imprecise results. Indeed, it finds a range [l, r] of positions in R which may be potential matches for P . However, if the entire range is within two consecutive sampled positions, we are only guaranteed that all the occurrences of P are in the range but there may exist positions in the range which do not match P . This uncertainty forces us to explicitly check every single position in the range until a match for P is found, if any. This implies that we have to check r − l + 1 = O(sample) positions in the worst case. Since every check has a cost proportional to m, this gives ω(m) query time. We use the data structure for weak prefix search of Theorem 5 to index the encodings of all suffixes of the text truncated at length ` = logc + log n. This way, we can find the range [l, r] of suffixes prefixed by E(P [1..m]) in O(m log log n/w + log(m log log n)) = O(m log log n/w + log log n) time with a data structure of size O(n log log n) bits. This is because E(P [1..m]) is drawn from an alphabet of size O(logc n), and both m and ` are in O(logc n). Apart from its faster query time, this solution has stronger guarantees. Indeed, if the pattern P has at least one occurrence, the range [l, r] contains all and only the occurrences of P . Instead, if the pattern P does not occur, [l, r] is an arbitrary and meaningless range. In both cases, just a single check of any position in the range is enough to answer the order-preserving query. This property gives a O(log n/ log log n) factor improvement over the previous solution. Speeding up verification phase. It is clear by the discussion above that the verification phase has to check only one position in the range [l, r]. If the range contains at least one sampled entry of R, we are done. Otherwise, we have to perform at most sample backward steps as in the previous solution. We now improve the computation of every single backward step. Assume we have to compute a backward step at i, where R[i] = j. Before performing the backward step, we have to compute the encoding E(S[j − 1..j + L[i] − 1]), given B[i], L[i], and E(S[j..j + u]) for some u ≥ L[i]. This is done as follows. We first prepend 0.5 to E(S[j..j + u]) and take its prefix of length L[i]. Then, we increase by one every value in the prefix which is larger than B[i]. These operations can be done in O(1 + L[i] log log n/w) = O(1 + m log log n/w) time by exploiting word parallelism of the RAM model. Indeed, we can operate on O(w/ log log n) symbols of the encoding in parallel. Now the backward step at i is i0 = k + D[i], where k is the only sampled entry in R whose encoding is prefixed by E(S[j − 1..j + L[i] − 1]). Notice that there cannot be more than one otherwise S[j − 1..j + L[i] − 1] would occur more than sample times, which was T. Gagie, G. Manzini and R. Venturini 23:11 excluded in the construction. Thus, the problem is to compute k, given i and E(S[j − 1..j + L[i] − 1]). It is crucial to observe that E(S[j − 1..j + L[i] − 1]) depends only on S and L[i] and not on the pattern P we are searching for. Thus, there exists just one valid E(S[j − 1..j + L[i] − 1]) that could be used at query time for a backward step at i. Notice that, if the pattern P does not occur, the encoding that will be used at i may be different, but in this case it is not necessary to compute a correct backward step. Consider the set E all these, at most n, encodings. The goal is to map each encoding in E to the sampled entry in R that it prefixes. This can be done as follows. We build a monotone minimal perfect hash function h() on E to map each encoding to its lexicographic rank. Obviously, the encodings that prefix a certain sampled entry i in R form a consecutive range in the lexicographic ordering. Moreover, none of these ranges overlaps because each encoding prefixes exactly one sampled entry. Thus, we can use a binary vector B to mark each of these ranges, so that, given the lexicographic rank of an encoding, we can infer the sampled entry it prefixes. The binary vector is obtained by processing the sampled entries in R in lexicographic order and by writing the size of its range in unary. It is easy to see that the sampled entry prefixed by x = E(S[j − 1..j + L[i] − 1]) can be computed as Rank1 (h(x)) in constant time. The data structures that stores B and supports Rank requires O(n) bits (see Theorem 1). The evaluation of h() is the dominant cost, and, thus, a backward step is computed in O(1 + m log log n/w) time. The overall space usage of this solution is O(n log log n) bits, because B has at most 2n bits and h() requires O(n log log n) bits by Theorem 4. Since we perform at most sample backward steps, it follows that the overall query time is O(sample × (1 + m log log n/w) = O(m). The equality follows by observing that sample = O(log n/ log log n), m = Ω(log n/ log log n) and w = Ω(log n). We finally observe that we could use the weak prefix search data structure instead of h() to compute a backward step. However, this would introduce a term O(log n) in the query time, which would be dominant for short patterns, i.e., m = o(log n). Query algorithm. We report here the query algorithm for a pattern P [1..m], with m = Ω(log n/ log log n). Recall that for shorter patterns we store all possible answers. We first compute E(P [1..m]) in O(1 + m log log n/w) time. Then, we perform a weak prefix search to identify the range [l, r] of encodings that are prefixed by E(P [1..m]) in O(m log log n/w + log log n) time. If P has at least one occurrence, the search is guaranteed to find the correct range; otherwise, the range may be arbitrary but the subsequent check will identify the mistake and report zero occurrences. In the checking phase, there are only two possible cases. The first case occurs when [l, r] contains a sampled entry, say i, in R. Thus, we can use the encoding from Lemma 6 to compare E(S[R[i]..R[i] + m]) and E(P [1..m]) in O(m) time. If they are equal, we report R[i]; otherwise, we are guaranteed that there is no occurrence of P in S. The second case is when there is no sampled entry in [l, r]. We arbitrarily select an index i ∈ [l, r] and we perform a sequence of backward steps starting from i. If P has at least one occurrence, we are guaranteed to find a sampled entry e in at most sample backward steps. The overall time of these backward steps is O(sample × m log log n/w) = O(m). If e is not found, we conclude that P has no occurrence. Otherwise, we explicitly compare E(S[R[e]+b..R[e]+m+b]) and E(P [1..m]) in O(m) time, where b is the number of performed backward steps. We report R[e] + b only in case of a successful comparison. Note that if P occurs, then the number of its occurrences is r − l + 1. CVIT 2016 23:12 An Encoding for Order-Preserving Matching 7 Space Lower Bound In this section we prove that our solution is space optimal. This is done by showing a lower bound on the space that any data structure must use to solve the easier problem of just establishing if a given pattern P has at least one order-preserving occurrence in S. More precisely, in this section we prove the following theorem. I Theorem 10. Any encoding data structure that indexes any S[1..n] over the alphabet [σ] with log σ = Ω(log log n) which, given a pattern P [1..m] with m = log n, establishes if P has any order-preserving occurrence in S must use Ω(log log n) bits of space. By contradiction, we assume that there exists a data structure D that uses o(n log log n) bits. We prove that this implies that we can store any string S[1, n] in less than n log σ bits, which is clearly impossible. We start by splitting S into n/m blocks of size m = log n characters each. Let Bi denote the ith block in this partition. Observe that if we know both the list L(Bi ) of characters that occur in Bi together with their number of occurrences and E(Bi ), we can recover Bi . This is because E(Bi ) implicitly tells us how to permute the characters in L(Bi ) to obtain Bi . Obviously, if we are able to reconstruct each Bi , we can reconstruct S. Thus, our goal is to use D together with additional data structures to obtain E(Bi ) and L(Bi ), for any Bi . We first directly encode L(Bi ) for each i by encoding the sorted sequence of characters σ with Elias-Fano representation. By Theorem 2, we know that this requires m log m + O(m) σ bits. Summing up over all the blocks, the overall space used is n log m + O(n) bits. Now it remains to obtain the encodings of all the blocks. Consider the set E of the encodings of all the substrings of S of length m. We do not store E because it would require too much space. Instead, we use a minimal perfect hash function h() on E. This requires O(n) bits by Theorem 3. This way each distinct encoding is bijectively mapped to a value in [n]. For each block Bi , we store h(Bi ). This way, we are keeping track of those elements in E that are blocks and their positions in S. This requires O(n) bits, because there are n/ log n blocks and storing each value needs O(log n) bits. We are now ready to retrieve the encoding of all the blocks, which is the last step to be able to reconstruct S. This is done by searching in D for every possible encoding of exactly m characters. The data structure will be able to tell us the ones that occurs in S, i.e., we are retrieving the entire set E. For each encoding e ∈ E, we check if h(e) is the hash of any of the blocks. In this way we are able to associate the encodings in E to the original block. Thus, we are able to reconstruct S by using D and additional data structures which uses n log σ − n log log n + O(n) bits of space. This implies that D cannot use o(n log log n) bits. 8 Conclusion We have given an encoding data structure for order-preserving pattern matching: given a string S of length n over an arbitrary alphabet and a constant c ≥ 1, we can store O(n log log n) bits such that later, given a pattern P of length m ≤ logc n, in O(m) time we can return the position of an order-preserving match of P in S (if one exists) and report the number of such matches. Our space bound is within a constant factor of optimal, even for only detecting whether a match exists, and our time bound is optimal when the alphabet size is at least logarithmic in n. We can build our encoding knowing only how each character of S compares to O(logc n) neighbouring characters. We believe our results will help open up a new line of research, where space is saved by restricting the set of possible queries or by relaxing the acceptable answers, that will help us deal with the rapid growth of datasets. T. Gagie, G. Manzini and R. Venturini 23:13 References 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 Djamal Belazzougui, Paolo Boldi, Rasmus Pagh, and Sebastiano Vigna. Monotone minimal perfect hashing: searching a sorted table with o (1) accesses. In Proceedings of the twentieth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 785–794. SIAM, 2009. Djamal Belazzougui, Paolo Boldi, Rasmus Pagh, and Sebastiano Vigna. Fast prefix search in little space, with applications. In European Symposium on Algorithms, pages 427–438. Springer, 2010. Djamal Belazzougui, Adeline Pierrot, Mathieu Raffinot, and Stéphane Vialette. Single and multiple consecutive permutation motif search. In International Symposium on Algorithms and Computation, pages 66–77. Springer, 2013. Domenico Cantone, Simone Faro, and M Oguzhan Külekci. An efficient skip-search approach to the order-preserving pattern matching problem. In Stringology, pages 22–35, 2015. Tamanna Chhabra, Simone Faro, M Oğuzhan Külekci, and Jorma Tarhio. Engineering order-preserving pattern matching with simd parallelism. Software: Practice and Experience, 2016. Tamanna Chhabra, Emanuele Giaquinta, and Jorma Tarhio. Filtration algorithms for approximate order-preserving matching. In International Symposium on String Processing and Information Retrieval, pages 177–187. Springer, 2015. Tamanna Chhabra, M Oguzhan Külekci, and Jorma Tarhio. Alternative algorithms for order-preserving matching. In Stringology, pages 36–46, 2015. Tamanna Chhabra and Jorma Tarhio. Order-preserving matching with filtration. In International Symposium on Experimental Algorithms, pages 307–314. Springer, 2014. Tamanna Chhabra and Jorma Tarhio. A filtration method for order-preserving matching. Information Processing Letters, 116(2):71–74, 2016. Sukhyeun Cho, Joong Chae Na, Kunsoo Park, and Jeong Seop Sim. A fast algorithm for order-preserving pattern matching. Information Processing Letters, 115(2):397–402, 2015. Maxime Crochemore, Costas S Iliopoulos, Tomasz Kociumaka, Marcin Kubica, Alessio Langiu, Solon P Pissis, Jakub Radoszewski, Wojciech Rytter, and Tomasz Waleń. Orderpreserving indexing. Theoretical Computer Science, 638:122–135, 2016. Pooya Davoodi, Gonzalo Navarro, Rajeev Raman, and S Srinivasa Rao. Encoding range minima and range top-2 queries. Philosophical Transactions of the Royal Society of London A: Mathematical, Physical and Engineering Sciences, 372(2016):20130131, 2014. Gianni Decaroli, Travis Gagie, and Giovanni Manzini. A compact index for order-preserving pattern matching. In Data Compression Conference, 2017. To appear. Peter Elias. Efficient storage and retrieval by content and address of static files. Journal of the ACM (JACM), 21(2):246–260, 1974. Robert M. Fano. On the number of bits required to implement an associative memory. Technical Report Memorandum 61, Project MAC, Computer Structures Group, Massachusetts Institute of Technology, 1971. Simone Faro and M Oğuzhan Külekci. Efficient algorithms for the order preserving pattern matching problem. In International Conference on Algorithmic Applications in Management, pages 185–196. Springer, 2016. Paolo Ferragina and Giovanni Manzini. An experimental study of a compressed index. Information Sciences, 135(1):13–28, 2001. Johannes Fischer. Combined data structure for previous-and next-smaller-values. Theoretical Computer Science, 412(22):2451–2456, 2011. Johannes Fischer and Volker Heun. A new succinct representation of rmq-information and improvements in the enhanced suffix array. In Combinatorics, Algorithms, Probabilistic and Experimental Methodologies, pages 459–470. Springer, 2007. CVIT 2016 23:14 An Encoding for Order-Preserving Matching 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 Arnab Ganguly, Rahul Shah, and Sharma V Thankachan. pbwt: Achieving succinct data structures for parameterized pattern matching and related problems. In Proceedings of the Twenty-Eighth Annual ACM-SIAM Symposium on Discrete Algorithms, pages 397–407. SIAM, 2017. Paweł Gawrychowski and Patrick K Nicholson. Encodings of range maximum-sum segment queries and applications. In Annual Symposium on Combinatorial Pattern Matching, pages 196–206. Springer, 2015. Paweł Gawrychowski and Patrick K Nicholson. Optimal encodings for range top-k, selection, and min-max. In International Colloquium on Automata, Languages, and Programming, pages 593–604. Springer, 2015. Paweł Gawrychowski and Przemysław Uznański. Order-preserving pattern matching with k mismatches. Theoretical Computer Science, 638:136–144, 2016. Mordecai Golin, John Iacono, Danny Krizanc, Rajeev Raman, Srinivasa Rao Satti, and Sunil Shende. Encoding 2d range maximum queries. Theoretical Computer Science, 609:316–327, 2016. Roberto Grossi, John Iacono, Gonzalo Navarro, Rajeev Raman, and Satti Srinivasa Rao. Encodings for range selection and top-k queries. In European Symposium on Algorithms, pages 553–564. Springer, 2013. Torben Hagerup and Torsten Tholey. Efficient minimal perfect hashing in nearly minimal space. In Annual Symposium on Theoretical Aspects of Computer Science, pages 317–326. Springer, 2001. Tommi Hirvola and Jorma Tarhio. Approximate online matching of circular strings. In International Symposium on Experimental Algorithms, pages 315–325. Springer, 2014. Guy Jacobson. Space-efficient static trees and graphs. In Foundations of Computer Science, 1989., 30th Annual Symposium on, pages 549–554. IEEE, 1989. Varunkumar Jayapaul, Seungbum Jo, Rajeev Raman, Venkatesh Raman, and Srinivasa Rao Satti. Space efficient data structures for nearest larger neighbor. Journal of Discrete Algorithms, 36:63–75, 2016. Seungbum Jo, Rajeev Raman, and Srinivasa Rao Satti. Compact encodings and indexes for the nearest larger neighbor problem. In International Workshop on Algorithms and Computation, pages 53–64. Springer, 2015. Jinil Kim, Peter Eades, Rudolf Fleischer, Seok-Hee Hong, Costas S Iliopoulos, Kunsoo Park, Simon J Puglisi, and Takeshi Tokuyama. Order-preserving matching. Theoretical Computer Science, 525:68–79, 2014. Marcin Kubica, Tomasz Kulczyński, Jakub Radoszewski, Wojciech Rytter, and Tomasz Waleń. A linear time algorithm for consecutive permutation pattern matching. Information Processing Letters, 113(12):430–433, 2013. Gonzalo Navarro, Rajeev Raman, and Srinivasa Rao Satti. Asymptotically optimal encodings for range selection. In 34th International Conference on Foundation of Software Technology and Theoretical Computer Science, page 291, 2014. Gonzalo Navarro and Sharma V Thankachan. Encodings for range majority queries. In CPM, pages 262–272, 2014. Alessio Orlandi and Rossano Venturini. Space-efficient substring occurrence estimation. Algorithmica, 74(1):65–90, 2016. Mihai Patrascu and Mikkel Thorup. Dynamic integer sets with optimal rank, select, and predecessor search. In Foundations of Computer Science (FOCS), 2014 IEEE 55th Annual Symposium on, pages 166–175. IEEE, 2014. Rajeev Raman. Encoding data structures. In International Workshop on Algorithms and Computation, pages 1–7. Springer, 2015. Rahul Shah. Personal communication, 2016. 

On the role of synaptic stochasticity in training low-precision neural networks Carlo Baldassi,1, 2, 3 Federica Gerace,2, 4 Hilbert J. Kappen,5 Carlo Lucibello,2, 4 Luca Saglietti,2, 4 Enzo Tartaglione,2, 4 and Riccardo Zecchina1, 2, 6 arXiv:1710.09825v2 [cond-mat.dis-nn] 20 Mar 2018 1 Bocconi Institute for Data Science and Analytics, Bocconi University, Milano, Italy 2 Italian Institute for Genomic Medicine, Torino, Italy 3 Istituto Nazionale di Fisica Nucleare, Sezione di Torino, Italy 4 Dept. of Applied Science and Technology, Politecnico di Torino, Torino, Italy 5 Radboud University Nijmegen, Donders Institute for Brain, Cognition and Behaviour 6525 EZ Nijmegen, The Netherlands 6 International Centre for Theoretical Physics, Trieste, Italy Stochasticity and limited precision of synaptic weights in neural network models are key aspects of both biological and hardware modeling of learning processes. Here we show that a neural network model with stochastic binary weights naturally gives prominence to exponentially rare dense regions of solutions with a number of desirable properties such as robustness and good generalization performance, while typical solutions are isolated and hard to find. Binary solutions of the standard perceptron problem are obtained from a simple gradient descent procedure on a set of real values parametrizing a probability distribution over the binary synapses. Both analytical and numerical results are presented. An algorithmic extension aimed at training discrete deep neural networks is also investigated. Learning can be regarded as an optimization process over the connection weights of a neural network. In nature, synaptic weights are known to be plastic, low precision and unreliable, and it is an interesting issue to understand if this stochasticity can help learning or if it is an obstacle. The debate about this issue has a long history and is still unresolved (see [1] and references therein). Here, we provide quantitative evidence that the stochasticity associated with noisy low precision synapses can drive elementary supervised learning processes towards a particular type of solutions which, despite being rare, are robust to noise and generalize well — two crucial features for learning processes. In recent years, multi-layer (deep) neural networks have gained prominence as powerful tools for tackling a large number of cognitive tasks [2]. In a K-class classification task, neural network architectures are typically trained as follows. For any input x ∈ X (the input space X typically being a tensor space) and for a given set of parameters W , called synaptic weights, the network defines a probability density function P (y | x, W ) over the K possible outcomes. This is done through composition of affine transformations involving the synaptic weights W , element wise non-linear operators, and finally a softmax operator that turns the outcome of previous operations into a probability density function [3]. The weights W are adjusted, in a supervised learning scenario, using a training set D of M known input-output associations, M D = {(xµ , y µ )}µ=1 . The learning problem is reframed into the problem of maximizing a log-likelihood L̃ (W ) over the synaptic weights W : max L̃ (W ) := W X (x,y)∈D log P (y | x, W ) (1) The maximization problem is approximately solved using variants of the Stochastic Gradient Descent (SGD) procedure over the loss function −L̃ (W ) [4]. In a Bayesian approach instead one is interested in computing the posterior distribution P (W | D) ∝ P (D | W ) P (W ), where P (W ) is some prior over the weights W . In deep networks, unfortunately, the exact computation of P (W | D) is typically infeasible and various approximated approaches have been proposed [5–7]. Shallow neural network models, such as the perceptron model for binary classification, are amenable to analytic treatment while exposing a rich phenomenology. They have attracted great attention from the statistical physics community for many decades [8–16]. In the perceptron problem we have binary outputs y ∈ {−1, +1}, 2 while inputs x and weights W are N -components vectors. Under some statistical assumptions on the training set D and for large N , single variable marginal probabilities P (Wi | D) can be computed efficiently, using Belief Propagation [17–19]. The learning dynamics has also been analyzed, in particular in the online learning setting [11, 20]. In a slightly different perspective the perceptron problem is often framed as the task of minimizing the error-counting Hamiltonian min H (W ) := W X (x,y)∈D Θ −y N X i=1 Wi xi ! , (2) where Θ (x) is the Heaviside step function, Θ (x) = 1 if x > 0 and 0 otherwise. As a constraint satisfaction problem, it is said to be satisfiable (SAT) if zero energy (i.e. H (W ) = 0) configurations exists, unsatisfiable (UNSAT) otherwise. We call solutions such configurations. Statistical physics analysis, assuming random and uncorrelated D, shows a sharp threshold at a certain αc = M/N , when N grows large, separating a SAT phase from an UNSAT one. Moreover, restricting the synaptic space to binary values,Wi = ±1, leads to a more complex scenario: most solutions are essentially isolated and computationally hard to find [13, 21]. Some efficient algorithms do exist though [12, 22] and generally land in a region dense of solutions. This apparent inconsistency has been solved through a large deviation analysis which revealed the existence of sub-dominant and dense regions of solutions [14, 23]. This analysis introduced the concept of Local Entropy [14] which subsequently led to other algorithmic developments [24–26] (see also [27] for related analysis). In the generalization perspective, solutions within a dense region may be loosely considered as representative of the entire region itself, and therefore act as better pointwise predictors than isolated solutions, since the optimal Bayesian predictor is obtained averaging all solutions [14]. Here, we propose a method to solve the binary perceptron problem (2) through a relaxation to a distributional space. We introduce a perceptron problem with stochastic discrete weights, and show how the learning process is naturally driven towards dense regions of solutions, even in the regime in which they are exponentially rare compared to the isolated ones. In perspective, the same approach can be extended to the general learning problem (1), as we will show. Denote with Qθ (W ) a family of probability distributions over W parametrized by a set of variables θ. Consider the following problem: max L (θ) := θ X (x,y)∈D log EW ∼Qθ P (y | x, W ) (3) Here L (θ) is the log-likelihood of a model where for each training example (x, y) ∈ D the synaptic weights are independently sampled according to Qθ (W ). Within this scheme two class predictors can be devised for any input x: ŷ1 (x) = argmaxy P (y | x, Ŵ ), ´where Ŵ = argmaxW Qθ (W ), and ŷ2 (x) = argmaxy dW P (y | x, W ) Qθ (W ). In this paper we will analyze the quality of the training error given by the first predictor. Generally, dealing with Problem (3) is more difficult than dealing with Problem (1), since it retains some of the difficulties of the computation of P (W | D). Also notice that for any maximizer W ? of Problem (1) we have that δ (W − W ? ) is a maximizer of Problem (3) provided that it belongs to the parametric family, as can be shown using Jensen’s inequality. Problem (3) is a "distributional" relaxation of Problem (1). Optimizing L (θ) instead of L̃ (W ) may seem an unnecessary complication. In this paper we argue that there are two reasons for dealing with this kind of task. First, when the configuration space of each synapse is restricted to discrete values, the network cannot be trained with SGD procedures. The problem, while being very important for computational efficiency and memory gains, has been tackled only very recently [5, 28]. Since variables θ typically lie in a continuous manifold instead, standard continuous optimization tools can be applied to L (θ). Also, the learning dynamics on L (θ) enjoys some additional properties when compared to the dynamics on L̃ (W ). In the latter case additional regularizers, such as dropout and L2 norm, are commonly used to improve generalization properties [4]. The SGD in the θ-space instead already incorporates the kind of natural regularization intrinsic in the Bayesian approach and the robustness associated to high local entropy [14]. Here we make a case for these arguments by a numerical 3 1.0 0.8 E 0.6 q 0.4 0.2 100 200 300 400 5000.0 1.0 N = 1001 0.8 N = 10001 0.6 0.4 1.0 E (%) 0.2 0.5 0.0 0.56 0.64 0.72 0.0 0.56 0.60 0.64 0.68 0.72 q success probability E 0.5 0.4 0.3 0.2 0.1 0.00 training epoch Figure 1. (Left) The training error and the squared norm against the number of training epochs, for α = 0.55 and N = 10001, averaged over 100 samples. (Right) Success probability in the classification task as a function of the load α for networks of size N = 1001, 10001 averaging 1000 and 100 samples respectively. In the inset we show the average training error at the end of GD as a function of α. and analytical study of the proposed approach for the binary perceptron. We also present promising preliminary numerical results on deeper networks. Learning for the Stochastic Perceptron. Following the above discussion, we now introduce our binary stochastic perceptron model. For each input x presented, N synaptic weights W = (W1 , . . . , WN ), Wi ∈ {−1, +1}, are randomly extracted according to the distribution Qm (W ) = N  Y 1 + mi i=1 2 δWi ,+1 + 1 − mi δWi ,−1 2  (4) where δa,b is the Kronecker delta symbol. We will refer to the set m = (mi )i , where mi ∈ [−1, 1] ∀i, as the magnetizations or the control parameters. We choose the probability P (y | x, W ) on the class y ∈ {−1, +1} for a given input x as follows: P (y | x, W ) = Θ y N X i=1 Wi xi ! . (5) While other possibilities for P (y | x, W ) could be considered, this particular choice is directly related to the form of the Hamiltonian in Problem (2), which we ultimately aim to solve. Given a training set D = M {(xµ , y µ )}µ=1 , we can then compute the log-likelihood function of Eq. (3), with the additional assumption that N is large and the central limit theorem applies. It reads L (m) = X (x,y)∈D log H y − pP P i i mi xi (1 − m2i ) x2i ! , (6) √ ´∞ 2 where H (x) := x dz e−z /2 / 2π. Minimizing −L (m) instead of finding the solutions of Problem (2) allows us to use the simplest method for approximately solving continuous optimization problems, the Gradient Descent (GD) algorithm:  
t t mt+1 ← clip m + η ∂ L m . mi i i (7) We could have adopted the more efficient SGD approach, however in our case simple GD is already effective. In the last equation η is a suitable learning rate and clip (x) := max (−1, min (1, x)), applied element-wise. Parameters are
randomly initialized to small values, m0i ∼ N 0, N −1 . At any epoch t in the GD dynamics a binarized configuration Ŵit = sign (mti) can be used to compute the training error Ê t = 1 MH Ŵ t . We con- sider a training set D where each input component xµi is sampled uniformly and independently in {−1, 1} (with this choice we can set y µ = 1 ∀µ without loss of generality). The evolution of the network during GD is shown in Fig. 1. The training error goes progressively to zero while the P mean 2squared norm of the control variables q?t = N1 i (mti ) approaches one. Therefore the distribution Qm concentrates around a single configuration as the training is progressing. This natural flow is similar to the annealing of the coupling parameter manually performed in local entropy inspired algorithms [25, 26]. We also show in Fig. 1 the probability over the realizations of D of finding a solution of the binary problem as function of the load α = M/N . The algorithmic capacity of GD is approximately αGD ≈ 0.63. This value has to be compared to the theoretical capacity αc ≈ 0.83, above which there are almost surely no solutions [9], and state-of-the-art algorithms based on message passing heuristics for which we have a range of capacities αM P ∈ [0.6, 0.74] [12, 22, 29]. Therefore GD reaches loads only slightly worse than those reached by much more fine tuned algorithms, a surprising results for such a simple procedure. Also, for α slightly above αGD 4 Z= ˆ Y Ω dmi δ i X i N m2i − q? N ! eβL(m) (8) where Ω = [−1, 1] , β is an inverse temperature, and we constrained the squared norm to q? N in order to mimic the natural flow of q?t in the training process. The dependence on the training set D is implicit in last equation. We shall denote with ED the average over the training sets with i.i.d. input and output components uniform in {−1, 1}. We investigate the average properties of the system for large N and fixed load α = M/N using the replica method in the Replica Symmetric (RS) ansatz [35]. Unfortunately the RS solution becomes locally unstable for very large β. Therefore, instead of taking the infinite β limit to maximize the likelihood we will present the results obtained for β large but still in the RS region. The details of the free energy cal- 0.07 0.06 0.05 0.04 0.03 0.02 0.01 0.00 0.7 0.20 GD GD slow equilibrium training error 0.15 stochastic binary continuous 0.10 S(d) the training error remains comparably low, as shown in Fig. 1. In our experiments most variants of the GD procedure of Eq. (7) performed just as well: e.g. SGD ors GD computed on the fields hti = tanh−1 (mti ) rather than the magnetizations[30]. Other updates rules for the control parameters can be derived as multiple pass of on-line Bayesian learning [31, 32]. Variations of rule (7) towards biological plausibility are discussed in the SM. Deep Networks. We applied our framework to deep neural networks with binary stochastic weights and sign activation functions. Using an uncorrelated neuron approximation, as in Ref. [6], we trained the network using the standard SGD algorithm with backpropagation. We give the details in the SM. On the MNIST benchmark problem [33], using a network with three hidden layers we achieved ∼ 1.7% test error, a very good result for a network with binary weights and activations and with no convolutional layers [34]. No other existing approach to the binary perceptron problem has been extended yet to deeper settings. Statistical mechanics Analysis. We now proceed with the analytical investigation of the equilibrium properties of the stochastic perceptron, which partly motivates the good performance of the GD dynamics. The starting point of the analysis is the partition function 0.05 0.8 q 0.9 1.0 0.00 0.00 0.04 d 0.08 0.12 Figure 2. (Left) Energy of the Binarized Configuration versus the control variable q? . We show the equilibrium prediction of Eq. (9), and numerical results from the GD algorithm and a GD algorithm variant where after each update we rescale the norm of m to q? until convergence before moving to the next value of q? according to a certain schedule. The results are averaged over 20 random realizations of the training set with N = 10001. (Right) Entropy of binary solutions at fixed distance d from BCs of the spherical, binary and stochastic perceptron (q? = 0.7, 0.8 and 0.9 from bottom to top) at thermodynamic equilibrium. In both figures α = 0.55, also β = 20 for the stochastic perceptron and β = ∞ for the spherical and binary ones. culation and of the stability check can be found in the SM. Energy of the Binarized Configuration. We now analyze some properties of the mode of the distribution Qm (W ), namely Ŵi = sign (mi ), that we call Binarized Configuration (BC). The average training error per pattern is:  * !+ X X 1  ED  Θ −y sign (mi ) xi E = lim N →∞ αN i (x,y)∈D (9) where h•i is the thermal average over m according to the partition function (8), which implicitly depends on D, q? and β. The last equation can be computed analytically within the replica framework (see SM). In Fig. 2 (Left) we show that for large β the BC becomes a solution of the problem when q? approaches one. This is compared to the values of the training error obtained from GD dynamics at corresponding values of q? , and a modified GD dynamics where we let the system equilibrate at fixed q? . The latter case, although we are at finite N and we are considering a dynamical process that could suffer the presence of local minima, is in rea- 5 sonable agreement with the equilibrium result of Eq. (9). Geometrical structure of the solution space. Most solutions of the binary perceptron problem are isolated [13], while a subdominant but still exponentially large number belongs to a dense connected region [14]. Solutions in the dense region are the only ones that are algorithmically accessible. Here we show that BCs of the stochastic binary perceptron typically belong to the dense region, provided q? is high enough. To prove this we count the number of solutions at a fixed Hamming distance d from typical BC (this corresponds to fixing an overlap p = 1 − 2d). Following the approach of Franz and Parisi [36] we introduce the constrained partition function Z(d, m) = X Y Θ y W (x,y)∈D X W i xi i × δ N (1 − 2d) − X ! sign (mi ) Wi i ! , (10) N where the sum is over the {−1, +1} binary configurations. The Franz-Parisi entropy S (d) is then given by S(d) = lim N →∞ 1 ED hlog Z (d, m)i . N (11) We show how to compute S (d) in the SM. In Fig. 2 (Right) we compare S (d) for the stochastic perceptron with the analogous entropies obtained substituting the expectation h•i over m in Eq. (11) with a uniform sampling from the solution space of the spherical (the model of Ref. [8]) and the binary (as in Ref. [13]) perceptron. The distance gap between the BC and the nearest binary solutions (i.e., the value of the distance after which S(d) becomes positive) vanishes as q? is increased: in this regime the BC belongs to the dense cluster and we have an exponential number of solutions at any distance d > 0. Typical binary solutions and binarized solutions of the continuous perceptron are isolated instead (finite gap, corresponding to S(d) = 0 at small distances). In the SM we provide additional numerical results on the properties of the energetic landscape in the neighbor- hood of different types of solutions, showing that solutions in flatter basins achieve better generalization than those in sharp ones. Conclusions. Our analysis shows that stochasticity in the synaptic connections may play a fundamental role in learning processes, by effectively reweighting the error loss function, enhancing dense, robust regions, suppressing narrow local minima and improving generalization. The simple perceptron model allowed us to derive analytical results as well as to perform numerical tests. Moreover, as we show in the SM, when considering discretized priors, there exist a connection with the dropout procedure which is popular in modern deep learning practice. However, the most promising immediate application is in the deep learning scenario, where this framework can be extended adapting the tools developed in Refs. [6, 7], and where we already achieved state-of-the-art results in our preliminary investigations. Hopefully, the general mechanism shown here can also help to shed some light on biological learning processes, where the role of low precision and stochasticity is still an open question. Finally, we note that this procedure is not limited to neural network models; for instance, application to constraint satisfaction problems is straightforward. CB, HJB and RZ acknowledge ONR Grant N0001417-1-2569. [1] H Sebastian Seung. Learning in spiking neural networks by reinforcement of stochastic synaptic transmission. Neuron, 40(6):1063–1073, 2003. [2] Yann LeCun, Yoshua Bengio, and Geoffrey Hinton. Deep learning. Nature, 521(7553):436–444, 2015. [3] David JC MacKay. Information theory, inference and learning algorithms. Cambridge university press, 2003. [4] Léon Bottou, Frank E Curtis, and Jorge Nocedal. Optimization methods for large-scale machine learning. arXiv preprint arXiv:1606.04838, 2016. [5] Daniel Soudry, Itay Hubara, and R Meir. Expectation Backpropagation: parameter-free training of multilayer neural networks with real and discrete weights. Neural Information Processing Systems 2014, 2(1):1–9, 2014. [6] José Miguel Hernández-Lobato and Ryan P Adams. Probabilistic Backpropagation for Scalable Learning of 6 [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] Bayesian Neural Networks. Journal of Machine Learning Research, 37:1–6, feb 2015. Oran Shayar, Dan Levi, and Ethan Fetaya. Learning discrete weights using the local reparameterization trick. ArXiv e-print, 2017. Elizabeth Gardner. The space of interactions in neural network models, 1988. Werner Krauth and Marc Mézard. Storage capacity of memory networks with binary couplings, 1989. Manfred Opper and Ole Winther. A Mean Field Approach to Bayes Learning in Feed-Forward Neural Networks. Physical Review Letters, 76(11):1964–1967, mar 1996. Sara Solla and Ole Winther. Optimal perceptron learning: an online Bayesian approach. In David Saad, editor, On-Line Learning in Neural Networks, pages 1–20. Cambridge University Press, 1998. Alfredo Braunstein and Riccardo Zecchina. Learning by Message Passing in Networks of Discrete Synapses. Physical Review Letters, 96(3):030201, jan 2006. Haiping Huang and Yoshiyuki Kabashima. Origin of the computational hardness for learning with binary synapses. Physical Review E, 90(5):052813, nov 2014. Carlo Baldassi, Alessandro Ingrosso, Carlo Lucibello, Luca Saglietti, and Riccardo Zecchina. Subdominant Dense Clusters Allow for Simple Learning and High Computational Performance in Neural Networks with Discrete Synapses. Physical Review Letters, 115(12):128101, 2015. Silvio Franz and Giorgio Parisi. The simplest model of jamming. Journal of Physics A: Mathematical and Theoretical, 49(14):145001, apr 2016. Simona Cocco, Rémi Monasson, and Riccardo Zecchina. Analytical and numerical study of internal representations in multilayer neural networks with binary weights. Physical Review E, 54(1):717–736, jul 1996. Marc Mézard. The space of interactions in neural networks: Gardner’s computation with the cavity method. Journal of Physics A: Mathematical and General, 22(12):2181–2190, jun 1989. Giorgio Parisi, Marc Mézard, and Miguel Angel Virasoro. Spin glass theory and beyond. World Scientific Singapore, 1987. Andrea Montanari and Marc Mézard. Information, Physics and Computation. Oxford Univ. Press, 2009. David Saad. On-line learning in neural networks. Cambridge University Press, 1998. Lenka Zdeborová and Marc Mézard. Constraint satisfaction problems with isolated solutions are hard. Journal of Statistical Mechanics: Theory and Experiment, 2008(12):P12004, oct 2008. Carlo Baldassi, Alfredo Braunstein, Nicolas Brunel, and Riccardo Zecchina. Efficient supervised learning in net- [23] [24] [25] [26] [27] [28] [29] [30] [31] [32] [33] [34] [35] works with binary synapses. Proceedings of the National Academy of Sciences, 104(26):11079–11084, jun 2007. Carlo Baldassi, Federica Gerace, Carlo Lucibello, Luca Saglietti, and Riccardo Zecchina. Learning may need only a few bits of synaptic precision. Physical Review E, 93(5):052313, 2016. Carlo Baldassi, Alessandro Ingrosso, Carlo Lucibello, Luca Saglietti, and Riccardo Zecchina. Local entropy as a measure for sampling solutions in constraint satisfaction problems. Journal of Statistical Mechanics: Theory and Experiment, 2016(2):023301, 2016. Carlo Baldassi, Christian Borgs, Jennifer T. Chayes, Alessandro Ingrosso, Carlo Lucibello, Luca Saglietti, and Riccardo Zecchina. Unreasonable effectiveness of learning neural networks: From accessible states and robust ensembles to basic algorithmic schemes. Proceedings of the National Academy of Sciences, 113(48):E7655–E7662, nov 2016. Pratik Chaudhari, Anna Choromanska, Stefano Soatto, Yann LeCun, Carlo Baldassi, Christian Borgs, Jennifer T. Chayes, Levent Sagun, and Riccardo Zecchina. Entropy-SGD: Biasing Gradient Descent Into Wide Valleys. ArXiv e-prints, nov 2016. Alfredo Braunstein, Luca Dall’Asta, Guilhem Semerjian, and Lenka Zdeborová. The large deviations of the whitening process in random constraint satisfaction problems. Journal of Statistical Mechanics: Theory and Experiment, 2016(5):053401, may 2016. Matthieu Courbariaux, Itay Hubara, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1. ArXiv e-prints, page 9, feb 2016. Carlo Baldassi and Alfredo Braunstein. A max-sum algorithm for training discrete neural networks. Journal of Statistical Mechanics: Theory and Experiment, 2015(8):P08008, 2015. This has the advantage that it doesn’t require clipping. Manfred Opper and Ole Winther. A Bayesian approach to on-line learning. In David Saad, editor, On-line learning in neural networks, pages 363–378. 1998. Thomas P Minka. Expectation Propagation for Approximate Bayesian Inference F d. Uncertainty in Artificial Intelligence (UAI), 17(2):362–369, 2001. Yann LeCun, Corinna Cortes, and Christopher JC Burges. Mnist handwritten digit database. AT&T Labs [Online]. Available: http://yann. lecun. com/exdb/mnist, 2, 2010. Carlo Baldassi, Christian Borgs, Jennifer T. Chayes, Carlo Lucibello, Luca Saglietti, Enzo Tartaglione, and Riccardo Zecchina. In preparation. Marc Mézard and Andrea Montanari. Information, Physics, and Computation. Oxford University Press, 7 January 2009. [36] Silvio Franz and Giorgio Parisi. Recipes for metastable states in spin glasses. Journal de Physique I, 5(11):1401– 1415, 1995. On the role of synaptic stochasticity in training low-precision neural networks Supplementary Material Carlo Baldassi,1, 2, 3 Federica Gerace,2, 4 Hilbert J. Kappen,5 Carlo Lucibello,2, 4 Luca Saglietti,2, 4 Enzo Tartaglione,2, 4 and Riccardo Zecchina1, 2, 6 1 Bocconi Institute for Data Science and Analytics, Bocconi University, Milano, Italy 2 Italian Institute for Genomic Medicine, Torino, Italy 3 Istituto Nazionale di Fisica Nucleare, Sezione di Torino, Italy 4 Dept. of Applied Science and Technology, Politecnico di Torino, Torino, Italy 5 Radboud University Nijmegen, Donders Institute for Brain, Cognition and Behaviour 6525 EZ Nijmegen, The Netherlands 6 International Centre for Theoretical Physics, Trieste, Italy CONTENTS I. Replica Symmetric solution and stability analysis 1 II. Energy of the Binarized Configuration 4 III. Franz-Parisi entropy 6 IV. Binary control variables 8 V. Stochastic deep networks 9 VI. A weighted perceptron rule 10 VII. Average energy around algorithmic solutions 12 References 13 I. REPLICA SYMMETRIC SOLUTION AND STABILITY ANALYSIS In this Section we show how to compute the average the free entropy of the stochastic perceptron model discussed in the main text, using the replica trick and under Replica Symmetry (RS) assumptions. The limit of validity of the RS ansatz will also be discussed. The statistical physics model of a perceptron with N stochastic binary synapses is defined by the partition function: ZN = ˆ Y N Ω i=1 dmi δ N X i=1 m2i − q? N ! eβL(m) . (1) 2 M The partition function depends implicitly on the inverse temperature β, a training set D = {y µ , xµ }µ=1 , M = αN N for some α > 0, y µ ∈ {−1, 1} , xµ ∈ {−1, 1}N , and a norm parameter q? . The integration is in the box Ω = [−1, 1] . The log-likelihood is given, for large N , by M X L (m) = with H (x) = 1 2 erfc  x √ 2  ! PN y µ i=1 mi xµi − p , N (1 − q? ) log H µ=1 (2) . As usual in statistical physics, we are interested in the large N limit of the system, an assumption which has already been exploited in expressing L(m) as in Eq. (2), as already stated. Also, the average ED over random instances of the training set is considered: xµi are uniformly and independently distributed over {−1, 1} and without loss of generality we set y µ = 1 ∀µ. Notice that although L(m) is concave on the box Ω, the norm constraint decomposes Ω into disjoint domains, therefore the maximization problem (i.e. the large β limit) is non-trivial. We define the average asymptotic free entropy as φ = lim N →∞ 1 ED log ZN N (3) where the limit is taken at fixed α. In order to compute φ we shall resort to the standard machinery of the replica method [13, 20]. The replicated partition function of the model is given by n ED ZN = ED ˆ n Y N Y Ω⊗n a=1 i=1 dmai n Y a=1 δ N X 2 (mai ) i=1 − q? N ! n M Y Y H µ=1 a=1 After some manipulations, at the leading order in N , we obtain n ED ZN β PN µ a ! x m − p i=1 i i . N (1 − q? ) ˆ Y dq̂aa Y dq̂ab dqab N nφ[q̂,q] ∼ e , 2π 2π a (4) (5) a<b where the replicated free entropy is given by φ [q̂, q] = − 1 X q̂ab qab + GS [q̂] + αGE [q] , 2n (6) a,b where we defined qaa ≡ q? for convenience. In last equation we also defined GS [q̂] = GE [q] = 1 log n 1 log n ˆ Y [−1,1]n a 1 dma e 2 P ab q̂ab ma mb ,   ˆ Y dûa dua − 1 Pab qab ûa ûb +iûa ua Y β ua e 2 H −√ . 2π 1 − q? a a (7) (8) 3 Saddle point evaluation of the replicated partition function yields the following identities: q̂ab = −α hhûa ûb iiE qab = hhma mb iiS m2a qaa ≡ q? = S a > b, (9) a > b, (10) (11) . Here we denoted with hh•iiS and hh•iiE the expectations taken according to the single-body partition function in the logarithms of Eq. (7) and Eq. (8) respectively. Notice that last equation is an implicit equation for q̂aa . We perform the saddle point evaluation and analytic continuation of n ↓ 0 within the Replica Symmetric (RS) ansatz. Therefore we have qab = q0 , q̂ab = q̂0 for a 6= b and q̂aa = q̂1 ∀a. The RS prediction for the average free entropy is then given by 1 (q0 q̂0 − q? q̂1 ) + GS (q̂0 , q̂1 ) + αGE (q0 , q? ) 2 φRS = extr q0 ,q̂0 ,q̂1 (12) where ˆ ˆ 1 1 2 √ dm e 2 (q̂1 −q̂0 )m + q̂0 zm , (13) −1 √  √  ˆ ˆ q0 z + q? − q0 u √ GE (q0 ) = Dz log Du H β − . (14) 1 − q? ´ ´ z2 e− 2 . Saddle point conditions yield the set of equations In last equation we used the notation Dz = √dz 2π GS (q̂0 , q̂1 ) = q0 = −2 Dz log ∂GS ; ∂ q̂0 q̂0 = −2α ∂GE ; ∂q0 0=2 ∂GS − q? , ∂ q̂1 (15) that we solve iteratively. Last equation is an implicit equation for q̂1 . This derivation is the starting point for the more complicated calculations of the energy of the Binarized Configurations (BCs) in Section II and of the Franz-Parisi entropy in Section III. While we conjecture φRS to be exact at low β, in the region of high β that we need to explore in order to maximize the log-likelihood L(m) it may be necessary to use a replica symmetry breaking formalism. A necessary, but not sufficient, condition for the validity of the RS formalism is the local stability condition for the free energy functional of Eq. (6) at the RS stationary point. The stability criterion involving the eigenvalues of the Hessian can be rephrased, with a slight adaption of the derivation of Ref. [14], as αγE γS < 1. (16) Here γE and γS are the relevant eigenvalues of the Hessians of GE [q] and GS [q̂] respectively and for small n. They are given by γE = ˆ γS = ˆ h
2 i 2 Dz û2 (z) − û (z) , h i2 2 Dz m2 (z) − (m (z)) . (17) (18) 4 RS instability lines 40 35 βc 30 25 20 15 0.78 0.8 0.82 0.84 0.86 q* 0.88 0.9 0.92 Figure 1. Critical value βc for the stability of the RS solution for different loads α = 0.5, 0.55, 0.6 (from top to bottom) as a function of q? . Above βc the RS solution is unstable and a replica symmetry breaking ansatz should be considered to obtain the correct solution. Expectations in lasts equations are defined by ûk (z) ≡ mk (z) ≡ √ dûdu k − 12 (q? −q0 )û2 +iûu+iû q0 z β H (u) 2π û e ´ dûdu − 1 (q −q )û2 +iûu+iû√q z ? 0 0 H β (u) 2 2π e √ ´1 2 1 dm mk e 2 (q̂1 −q̂0 )m + q̂0 zm −1 √ ´1 1 2 dm e 2 (q̂1 −q̂0 )m + q̂0 zm −1 ´ (19) (20) In Fig. 1 we show the stability line βc (q? ), defined by the condition αγE γS = 1, for different values of α. Due to numerical problem arising in computing integrals at high β, we explore a small q? window. We note that βc (q? ) stays finite in the range of parameters we explored and that the β ↑ ∞ limit of the RS solution cannot be taken carelessly. Nonetheless the βc (q? ) is generally quite high, although decreasing with α. In the main text, where we presented the results for α = 0.55, we set the inverse temperature to β = 20, where the RS results are supposedly correct and quantitatively close to the β = +∞ limit. II. ENERGY OF THE BINARIZED CONFIGURATION We now show how to compute the average energy E (also called training error) associated to a typical Binarized Configuration (BC). In the thermodynamic limit it is written as E = lim ED N →∞ "* Θ −y 1 X i sign (mi ) x1i !+# , (21) 5 where the average h•i is over m sampled according to the partition function (1) and Θ(x) is the Heaviside step function. Along the lines of previous Section, we resort to the replica approach, although here the replica of index 1 is distinguished from the others: E = lim  ˆ lim ED  N →∞ n→0 Y Ω⊗n a,i dmai n Y δ a=1 N X 2 (mai ) i=1 − q? N ! Θ − X i sign m1i
x1i ! e β P  a a L(m )  . (22) P a b I and the conjugated Lagrangian multipliers q̂ab , In addition to the order parameters qab = N1 P i mi mi of Section
we also have to introduce the overlaps pa = N1 i sign m1i mai and the corresponding multipliers p̂a . We obtain the following expression for the mean energy E: E = lim lim N →∞ n→0 ˆ Y dqab a<b Y dq̂ab Y dpa dp̂a eN nφ̃[q,q̂,p,p̂] Ẽ [q, p] 2π a 2π (23) a≤b The free entropy functional φ [q, q̂, p, p̂] in this case reads 1X 1 X q̂ab qab − p̂a pa + GS [q̂, p̂] + αGE [q] 2n n a a,b   ˆ Y X X
1 1 q̂ab ma mb + sign m1 GS [q̂, p̂] = log dma exp  p̂a ma  n 2 n [−1,1] a a a,b     ˆ Y X X 1 dua dûa Y β ua 1 GE [q] = log H −√ exp i ua ûa − qab ûa ûb  . n 2π 2 1 − q? a a a φ̃ [q, q̂, p, p̂] = − (24) (25) (26) a,b and the other term appearing in the integrand is given by   ˆ Y ˆ P P P ˆY 1 ˆ2 dua dûa ua dũdũ ˆ ˆ 1 β Ẽ [q, p] = H −√ Θ (−ũ) ei a ua ûa − 2 ũ +iũũ− 2 a,b qab ûa ûb −ũ a pa ûa . (27) 2π 2π a 1 − q? a Saddle point evaluation of φ̃ with respect to pa readily gives p̂a = 0. On this submanifold, φ̃ reduces to the functional φ of previous Section, the matrix qab and q̂ab can be evaluated at saddle point in terms of q0 , q̂1 , q̂0 within the RS ansatz and analytic continuation to n = 0 is finally obtained. Saddle point conditions with respect to p̂a instead, i.e. ∂ φ̃/∂ p̂a = 0, fix the parameters pa ≡ p̃ ∀a > 1 and p1 ≡ p (here there is a little abuse of notation, the scalar value p has not to be confused to the n-dimensional vector of previous equations). In conclusion, and in the small n limit, after solving Eqs. (15) for q0 , q̂1 , q̂0 we compute the saddle point values of p and p̃ by p= ˆ p̃ = ˆ Dz ´1 Dz ´ 1 −1 2 √ dm sign (m) m e 2 (q̂∗ −q̂0 )m + √ ´1 1 2 dm e 2 (q̂∗ −q̂0 )m + q̂0 zm −1 1 −1 1 dm e 2 (q̂∗ −q̂0 )m 2 q̂0 zm ,  ´  2 √ 1 1 sign (m) dm m e 2 (q̂∗ −q̂0 )m + q̂0 zm −1 . h´ i2 √ 1 1 2 2 (q̂∗ −q̂0 )m + q̂0 zm dm e −1 (28) √ + q̂0 z0 m (29) 6 The value of E is then simply given by Ẽ evaluated on the saddle point. After some manipulation of the integrals appearing in Eq. (27) we finally arrive to E= ˆ Dz ´ Du H   √ p−p̃ u+ √p̃q z 0  rq? −q0  2 1− p̃q − 0 ´ (p−p̃)2 q? −q0   √ √ q −q0 u+ q0 z H β − ? √1−q ?  √  √ q −q0 u+ q0 z Du H β − ? √1−q ? (30) . This result is presented as the equilibrium curve in Figure 2 (Left) of the main text. III. FRANZ-PARISI ENTROPY In last Section we obtained some analytical proving that typical BCs of the stochastic perceptron can achieve essentially zero training error if β and q∗ are large enough, and if the load α is below some critical capacity. This BCs are therefore solution of the binary perceptron problem. While typical (most numerous) solutions of the binary solutions problem are known to be isolated [17], we will show here that typical BCs belong to the dense solution region uncovered in Ref. [5]. Notice that, while for q∗ = 1 the stochastic perceptron reduces to binary one, the limit q∗ → 1 of many observables will not be continuous due to this phenomena. Most noticeably, as shown in [5], the generalization error of solutions in the dense region is typically lower than the generalization error of isolated solutions. We are interested in counting the number of solutions of the binary perceptron problem at fixed Hamming distancePd from a typical BC of the stochastic perceptron. For notation convenience we work at fixed overlap p = N1 i Wi sign(mi ), which can be linked to d by the relation p = 1 − 2d. Following [12, 17] we define the Franz-Parisi entropy as 1 S (p) = lim ED N →∞ N    X    + X Y X log  Θ y Wi xi × δ pN − sign (mi ) Wi  , * i W (x,y)∈D (31) i where the expectation h•i P over m is defined as usual according to Gibbs measure of the stochastic perceptron given by Eq. (1). The sum W is over the binary configuration space {−1, +1}N . The computation of S(p) is lengthy but straightforward, and can be done along the lines of Refs. [12, 17] using the replica method within the RS ansatz. Here we have the additional complication of some extra order parameters, due to the presence of the signin the constraint. We will present here just the final result. First, the order parameters q0 , q̂0 and q̂1 can be independently fixed solving the saddle point equations (15). S(p) is then given by S (p) = extr Q0 ,Q̂0 ,s0 ,s1 ,ŝ0 ,ŝ1 ,p̂ 1 P FP − Q̂ (1 − Q) + ŝ0 s0 − ŝ1 s1 − p̂p + GF S (Q̂0 , ŝ0 , ŝ1 , p̂) + αGE (Q0 , s0 , s1 ), 2 (32) where the entropic contribution is given by P GF S (Q̂0 , ŝ0 , ŝ1 , p̂) = ˆ Dz ´1 −1 dm ´ 1 Dη e 2 (q̂1 −q̂0 )W̃ ´1 −1 dm e 2 √ + q̂0 zm √ 1 2 2 (q̂1 −q̂0 )m + A (m, η, z) q̂0 zm , (33) 7 with  A (m, η, z) = log 2 cosh (ŝ1 − ŝ0 ) m + p̂ sign (m) + and the energetic contribution by P GF E (Q0 , s0 , s1 ) = ˆ Dz0 ´ with a = q? − q0 − s  Q̂0 q̂0 − ŝ20 ŝ0  η+ √ z , q̂0 q̂0 (34)   √   √ √ s s −s bη+ √q0 z0 q0 z0 + az1 + 1√ 0 η b 0 √ √ log H − Dη Dz1 H β − 1−q? 1−Q0  √  √ , ´ q z + q −q z Dz1 H β − 0 0√1−q?? 0 1 (s1 − s0 ) Q0 − s0 2  1− s0 (q0 − s0 ) Q0 q0 − s20  ; b= Q0 q0 − s20 . q0 (35) (36) The extremization condition of Eq. (32) reads Q̂0 = −2α Q0 = 1 − 2 P ∂GF S ∂ Q̂0 P ∂GF ∂GF P ∂GF P E ; ŝ0 = −α E ; ŝ1 = α E ; ∂Q0 ∂s0 ∂s1 ; s0 = − (37) P P P ∂GF ∂GF ∂GF S S S ; s1 = ; 0= − p. ∂ŝ0 ∂ŝ1 ∂ p̂ (38) This system of coupled equations can be solved once again by iteration, with last equation being solved for p̂ at each step with Newton method. The solution can then be plugged into Eq. (32), thus obtaining the final expression for the Franz-Parisi entropy. In Figure 2 (Right) of the main text we show the results for S(p) at α = 0.55, β = 20 and different values of q? . Due to convergence problems in finding the fixed point of Eqs. (37,38), some of the curves could not be continued to large values of d = (1 − p)/2. It is not clear if the problem is purely numerical and caused by the many integrals appearing in the equations and by the large value of β, or if it is an hint of a replica symmetry breaking transition. Nonetheless the region we are interested in exploring is that of low d, where the curve at q∗ = 0.9 reaches the origin, meaning that typical BCs are in the dense region of binary solution at this point. In the same figure we also compare the S(p) with two similar Franz-Parisi entropies, which we denote here by Sbin (p) and Ssph (p), for the binary and the spherical perceptron respectively. These two entropies are defined as in Eq. (31), the only difference being that the expectation h•i over m is according to Z= with dν(m) = spherical one. Q i (δ(mi ˆ dν(m)  Y  X Θ yµ mi xµi , µ (39) i − 1) + δ(mi − 1))dmi in the binary case and dν(m) = Q i dmi δ P i
m2i − N in the 8 0.16 analytic BP N=1001 0.14 0.12 S 0.1 0.08 0.06 0.04 0.02 0 0 0.01 0.02 0.03 d 0.04 0.05 0.06 Figure 2. Franz-Parisi entropies for α = 0.55 and q? = 0.7, 0.8, 0.9 (from top to bottom). (Purple) Average case Franz-Parisi entropy S(d) as given by Eq. Eq. (32) for β = 20. (Green) Single sample Franz-Parisi entropies computed with Belief Propagation, averaged over 100 samples. We also derived and implemented a single sample version of the Franz-Parisi calculation, performed as follows. For a given realization of D we establish a slow schedule of q? values and perform a GD on mt where after each update we rescale the squared norm of mt to q? N until convergence, before moving to the next value of q? . At any point of the schedule, the configuration mt is binarized and given and the constrained entropy of the binary perceptron is computed using the Bethe approximation given by the Belief Propagation algorithm (see Ref. [6] for a detailed exposition). The result of the simulations are presented in Fig. 2. The slight deviation of the BP results from the analytic curves could be explained by several causes: 1) finite size effects; 2) the analytic prediction is for non-zero (although low) temperature; 3) the reference configuration mt is selected through the GD dynamics, while in the analytic computation m is sampled according to the thermal measure defined by partition function of Eq. (1). IV. BINARY CONTROL VARIABLES In order to make a straightforward connection with the large deviation analyses proposed in Ref. [5], we have √ also considered the case in which the control variables mi are discretized as well: mi = q? σi , with σi ∈ {−1, 1}. In this case the log-likelihood of the stochastic perceptron model reads: L(σ) = M X µ=1 log H − r ! P q? y µ i σi xµi pP µ . 2 1 − q? i (xi ) (40) The statistical physics analysis proposed in the main text can be easily adapted to this case. Fig. (3) shows the average energy E, associated to a typical configuration σ, as a function of q? . The analytic results are found to be in reasonable agreement with the estimation of the training error obtained through a Markov Chain Monte Carlo on the system with Hamiltonian given by −L(σ), with inverse temperature β = 15. Moreover, instead of assuming σ to be the parameters controlling Qm (W ), from which the stochastic binary synapses are sampled, it is instructive to take a different perspective: consider a model where the synapses are 9 0.016 MC analytic 0.014 training error 0.012 0.01 0.008 0.006 0.004 0.002 0 0.6 0.65 0.7 0.75 0.8 q* 0.85 0.9 0.95 1 Figure 3. Stochastic perceptron with binary control. Energy of the clipped center versus q? . Red curve, MC simulation at N = 1001, averaged over 100 samples. Green curve, analytic results determined through the replica approach. Storage load α = 0.55, inverse temperature β = 15. binary and deterministic, and where we introduce a dropout mask [21], randomly setting to zero a fraction p of the inputs. In this scenario, we can write the log-likelihood of obtaining a correct classification over the independent realizations of the dropout mask for each datapoint. For large N the resulting log-likelihood is exactly that given by Eq. (40), once we set q? = 1 − p. Moreover, in the case of a single layer network we are considering, the dropout mask on the inputs could be equivalently applied to the synapses, as in the drop-connect scheme [22]. We can thus see a clear analogy between the dropout/dropconnect schemes and the learning problem analyzed throughout this paper, even though in standard machine learning practice the synaptic weights σi are not constrained to binary values. V. STOCHASTIC DEEP NETWORKS The stochastic framework can be extended to train deep networks models with binary synapses and binary activations using standard deep learning techniques, once some approximations to the log-likelihood estimation are taken into account. Since this extension is beyond the scope of the present paper, here we only sketch the training algorithm and give some preliminary results on its performance, reserving a detailed overview and extensive testing to a future publication [3]. Consider a multi-layer perceptron with L hidden neuron layers, with synaptic weights Wij` , ` = 0, . . . . , L, and sign activations: τj`+1 = sign X j Wij` τj` + b`i  , ` = 0, . . . , L, (41) where τ 0 = x is the input of the network, and b` are continuous biases to be optimized. In our stochastic framework the weights Wij` are independent binary (±1) random variables with means m`ij to be optimized. For a fixed activation trajectory (τ ` )` and wide layers, expectations with respect to W can be taken. Also, adapting the 10 scheme of Ref. [15], the probabilistic iteration of the neuron activation distribution P` (τ ` ) across the layers can be performed within a factorized approximation, in terms of the neuron activation’s means a`i : a`+1 i = 2H −P P j j m`ij a`j + b`i 1 − (m`ij )2 (a`j )2 ! −1 Finally an approximated softmax layer can be defined on the last layer output aL+1 , and consequently a crossentropy loss function can be used in the training. We experimented this approach on the MNIST dataset, where we trained networks a 3 hidden layers of width 801. We approximately minimized the loss function using Adam optimizer with standard parameters and learning rate η = 10−2 . We used in our simulations the Julia deep learning library Knet [23], providing automatic differentiation, backpropagation and GPU acceleration. At the end of the training the resulting binarized configuration, Ŵij` = sign(m`ij ), achieved ∼ 1.7% test error. Our implementation of the current state of the art algorithm [10] on the same network, using batch normalization and with learning rate η = 10−3 , achieves approximately the same result. For the sake of comparison, we note that a standard neural network with the same structure but with ReLU activations and continuous weights we obtain ∼ 1.4% test error. Given the heavy constraints on the weights, the discontinuity of the sign activation and the peculiarity of the training procedures, it is quite astonishing to observe only a slight degradation in the performance of binary networks when compared to their continuous counterparts. Further improvements of our results, up to ∼ 1.3% test error, can be obtained applying dropout on the input and the hidden layers. VI. A WEIGHTED PERCEPTRON RULE In the main text we introduced a stochastic perceptron model where the stochastic synapses could be integrated out thanks to the central limit theorem. Therefore we could express the log-likelihood of the model L(m) as an easy  QN  1−mi i to compute function of the parameters m governing the distribution Qm (W ) = i=1 1+m 2 δWi ,+1 + 2 δWi ,−1 . We used deterministic gradient ascent as a procedure to optimize m. At convergence, the binarized configuration Wi = sign(mi ) is proposed as an approximate solution of the binary problem. This learning rule (i.e. Eq. (7) in the main text) can be rewritten as m0i M 1 X = clip mi + η K M µ=1 µ yµ h − µ σ ! µ y µ xµi (xµi )2 h + σµ (σ µ )3 !! , (42) pP P µ µ 2 2 where we defined h = i mi xµi , σ µ = i (1 − mi )(xi ) and K(x) = ∂x log H(x). We now proceed to modify the learning rule to test its adaptability to biological scenarios. As a premise, we note that the emergence of a discretized set of synaptic strengths, as encoded by our model, is an experimentally observed property of many neural systems [8, 19]. Inspection of (42) shows an Hebbian structure, where the synaptic µ strength is reinforced on the base of presynaptic and postsynaptic activity, with a modulating factor K(−y µ h /σ µ ) that can be interpreted as a reward signal [18]. The sum over the examples in the training set can be changed with the random extraction of a single index µ. In this way the algorithm can be naturally extended to an online learning scenario. We revert to the original µ stochastic variables, sampling Wi ∼ Qmi ∀i and we replace the average pre-activation value h with √ its realization P µ hµ = i Wi xi . Further butchering of (42) is obtained crudely replacing σ µ by the constant σ = 0.5N . The final stochastic rule reads 11 1.0 0.8 E 0.6 q 0.4 0.2 0.0 200 400 600 800 1000 1.0 0.8 0.6 0.4 0.2 0.00.3 q success probability E 0.5 0.4 0.3 0.2 0.1 0.00 training epoch N = 1001 N = 10001 5 4 E (%) 3 2 1 0 0.4 0.5 0.6 0.4 0.5 0.6 Figure 4. Performance of learning rule Eq. (43). Results on system size N = 1001 are averaged over 100 samples, learning rate η = 10−2 σ. Experiments at N = 10001 are averaged over 10 samples, learning rate η = 10−3 σ. (Left) The training error and the squared norm against the number of training epochs for α = 0.55 and N = 1001. (Right) Success probability within 2000 epochs in the classification task as a function of the load α = M/N . In the inset we show the average training error (in percentage) at the end of GD. m0i   y µ hµ = clip mi + ηK − σ  y µ xµi (xµ )2 hµ + i 3 σ σ  . (43) We measure the performance of rule (43), on randomly genererated training sets with uniform i.i.d. xµi = ±1 and y = ±1. We present the results in Fig. 4. We observe a degradation in performance with respect to rule (42) and longer convergence times. Nonetheless, the algorithm is still able to efficiently classify an extensive number of examples (for the considered system sizes) up to a load α = M/N ∼ 0.45. As with gradient descent, above the algorithmic threshold we observe a graceful increase of the training error of the binarized configurations returned by the algorithm. Learning rule (43) could be utterly simplified if we discarded the last term on the right hand side and we replaced K(x) with the Heaviside theta function Θ(x), which takes value 0 for negative arguments and 1 otherwise. The new rule would read µ m0i = clip (mi + ηΘ (−y µ hµ ) y µ xµi ) . (44) P We first observe that, if we choose hµ = i sign(mi )xµi , we obtain a variant of the clipped perceptron (CP) algorithm, analyzed in Refs. [1, 4]. The performances of this rule were shown to degrade rapidly with N . For example, rule (44) with deterministic hµ fails solution within 2000 epochs at N = 2001 and α = 0.3. P to find µ Instead, we find that if we consider hµ = i Wi xi , with Wi sampled according to mi , we obtain a stochastic 0 version  of the rule able to succeed in the same setting. We note also that the ordinary perceptron rule, mi = µ clip mi + ηΘ(−y µ h )y µ xµi , is not able to provide binary solutions, even at very low α. Although a proper assessment of the scaling behaviour of these algorithms with N is beyond the scope of this work, we report that rule (43) consistently outperforms both variants of rule (44). Moreover its performance can 12 16 teacher SA CP CP−S GD BP+R 14 12 Ê (%) 10 8 6 4 2 0 0 0.02 0.04 0.06 0.08 0.1 d Figure 5. The energy of Eq. 45 as a function of the Hamming distance dN from the teacher and from solutions found by different algorithms. N = 1001 and α = 0.4 in all simulations. Curves are averaged over 40 samples. be further improved using the actual value σ µ instead of σ. In the next section we show that stochastic variant of (44), which is a novel contribution of this paper and it is tied to the stochastic framework we have investigated, typically lands in a flatter region compared to the deterministic version. VII. AVERAGE ENERGY AROUND ALGORITHMIC SOLUTIONS In order to characterize the energy landscape around a reference configuration W , a standard tool is the constrained entropy S(W, d) (also called local entropy), counting the number of solutions (i.e. zero energy configurations) at distance d from W . The average properties of S(W, d) have been analyzed in the main text. If for any small d > 0 we have S(W, d) > 0, we say that the configuration W belongs to a dense cluster. Otherwise, we say that W is isolated [5, 17]. Along with S(W, d), we can consider a simpler observable that can help to build a picture of an heterogeneous energy landscape, made of wide and sharp basins. Following Ref. [7], we thus define the average misclassification error made by configurations at distance d from W , as Ê(W, d) = EW 0 |W,d where the expectation is defined by EW 0 |W,d • = M X 1 X Θ −y µ Wi0 xµi M µ=1 i ! , P P W W 0) W 0 • × δ (N (1 − 2d) − P P i i0 i W 0 δ (N (1 − 2d) − i Wi Wi ) (45) (46) Notice that configurations W 0 partecipating to EW 0 |W,d are not required to be solutions of the problem: we can easily sample them by choosing dN spins at random to be flipped in W . In our tests the expectation is approximated by 103 Monte Carlo samples. 13 Teacher SA CP CP-S GD BP+R 1 0.578(3) 0.628(3) 0.644(3) 0.642(3) 0.657(2) Table I. Generalization accuracy in the teacher-student scenario. N = 1001, α = 0.4, averages over 40 samples. We explored the behavior of Ê(W, d) as a function of d for different solutions W of the problem, obtained from different algorithms. We compare: the Gradient Descent (GD) algorithm investigated in (Eq. 7); Pthe main text µ the two variants of rule 44 (from previous section), the one with deterministic hµ = sign(m )x (CP) and i i i the one with stochastic hµ sampled according to m (CP-S); the Belief Propagation algorithm with reinforcement √ P P heuristic (BP+R) of Ref. [9]; Simulated Annealing (SA) on the Hamiltonian µ Θ1 (−y µ i Wi0 xµi / N ), where Θ1 (x) = x Θ(x). In order to compare the properties of the algorithmic solutions and the typical isolated solutions, it is useful to consider the so-called teacher-student scenario [11]. As before, we generate uniformly i.i.d. xµi = ±1, but we assign the labels according to a teacher configuration W T (we can choose WiT = 1 ∀i without loss of generality). In this scenario, the teacher has the same statistical properties of the typical solutions of the training set, therefore it is an isolated solution itself [5, 17]. Results are presented in Fig. 5. For the isolated teacher we see a rapid increase of the average energy around the origin. The same happens for solutions discovered by SA, which we can classify as isolated as well. Also, SA was the slowest algorithm to reach a solution (unsurprisingly, since it is known to scale badly with the system size [2, 16, 17]). Solutions from CP-S, GD and BP+R instead are surrounded by a much flatter average landscape and, remarkably, they all give similar results. These three algorithms are implicitly or explicitly devised to reach robust basins: GD and CP-S are derived within our stochastic framework, while the reinforcement term in BP+R has been shown in [2] to be linked to local entropy maximization. Solutions from CP algorithm, while not in basins as sharp as the ones found by SA, do not achieve the same robustness as ones from those three algorithms. We give some additional details on the simulation’s protocol. The setting chosen, N = 1001 and α = 0.4 (a low load regime in the teacher-student scenario), is such that each algorithm finds a solution on each instance within small computational time (a few minutes). As soon as a solution W is discovered the algorithm is arrested. Results could be slightly different for some algorithms if they were allowed to venture further within the basin of solutions. For CP and CP-S we set η = 2 ∗ 10−3 , while η = 0.1 for GD. The reinforcement parameter in BP+R is updated according to 1 − rt+1 = (1 − rt )(1 − 10−3 ) while the inverse temperature schedule in SA is β t+1 = β t (1 + 5 ∗ 10−3 ) . In Table I we report the probability of correctly classifying a new example generated by the teacher, usually called generalization accuracy, for each of the algorithms investigated. We note a clear correlation between the flatness of the basin as seen in Fig. 5 and the ability to classify correctly new examples, with SA having the worst performances and BP+R the best ones. [1] Carlo Baldassi. Generalization Learning in a Perceptron with Binary Synapses. Journal of Statistical Physics, 136(5):902– 916, sep 2009. [2] Carlo Baldassi, Christian Borgs, Jennifer T. Chayes, Alessandro Ingrosso, Carlo Lucibello, Luca Saglietti, and Riccardo Zecchina. Unreasonable effectiveness of learning neural networks: From accessible states and robust ensembles to basic algorithmic schemes. Proceedings of the National Academy of Sciences, 113(48):E7655–E7662, nov 2016. [3] Carlo Baldassi, Christian Borgs, Jennifer T. Chayes, Carlo Lucibello, Luca Saglietti, Enzo Tartaglione, and Riccardo Zecchina. In preparation. 14 [4] Carlo Baldassi, Alfredo Braunstein, Nicolas Brunel, and Riccardo Zecchina. Efficient supervised learning in networks with binary synapses. Proceedings of the National Academy of Sciences, 104(26):11079–11084, jun 2007. [5] Carlo Baldassi, Alessandro Ingrosso, Carlo Lucibello, Luca Saglietti, and Riccardo Zecchina. Subdominant Dense Clusters Allow for Simple Learning and High Computational Performance in Neural Networks with Discrete Synapses. Physical Review Letters, 115(12):128101, 2015. [6] Carlo Baldassi, Alessandro Ingrosso, Carlo Lucibello, Luca Saglietti, and Riccardo Zecchina. Local entropy as a measure for sampling solutions in constraint satisfaction problems. Journal of Statistical Mechanics: Theory and Experiment, 2016(2):023301, 2016. [7] Carlo Baldassi and Riccardo Zecchina. Efficiency of quantum vs. classical annealing in nonconvex learning problems. Proceedings of the National Academy of Sciences, 115(7):1457–1462, feb 2018. [8] Thomas M Bartol, Cailey Bromer, Justin P Kinney, Michael A Chirillo, Jennifer N Bourne, Kristen M Harris, and Terrence J Sejnowski. Hippocampal spine head sizes are highly precise. bioRxiv, page 016329, 2015. [9] Alfredo Braunstein and Riccardo Zecchina. Learning by Message Passing in Networks of Discrete Synapses. Physical Review Letters, 96(3):030201, jan 2006. [10] Matthieu Courbariaux, Itay Hubara, Daniel Soudry, Ran El-Yaniv, and Yoshua Bengio. Binarized Neural Networks: Training Deep Neural Networks with Weights and Activations Constrained to +1 or -1. ArXiv e-prints, page 9, feb 2016. [11] Andreas Engel. Statistical mechanics of learning. Cambridge University Press, 2001. [12] Silvio Franz and Giorgio Parisi. Recipes for metastable states in spin glasses. Journal de Physique I, 5(11):1401–1415, 1995. [13] Elizabeth Gardner. The space of interactions in neural network models, 1988. [14] Elizabeth Gardner and Bernard Derrida. Optimal storage properties of neural network models, 1988. [15] José Miguel Hernández-Lobato and Ryan P Adams. Probabilistic Backpropagation for Scalable Learning of Bayesian Neural Networks. Journal of Machine Learning Research, 37:1–6, feb 2015. [16] Heinz Horner. Dynamics of learning for the binary perceptron problem. Zeitschrift für Physik B Condensed Matter, 86(2):291–308, 1992. [17] Haiping Huang and Yoshiyuki Kabashima. Origin of the computational hardness for learning with binary synapses. Physical Review E, 90(5):052813, nov 2014. [18] Yonatan Loewenstein and H. Sebastian Seung. Operant matching is a generic outcome of synaptic plasticity based on the covariance between reward and neural activity. Proceedings of the National Academy of Sciences, 103(41):15224–15229, 2006. [19] Daniel H. O’Connor, Gayle M. Wittenberg, and Samuel S.-H. Wang. Graded bidirectional synaptic plasticity is composed of switch-like unitary events. Proceedings of the National Academy of Sciences, 102(27):9679–9684, 2005. [20] Giorgio Parisi, Marc Mézard, and Miguel Angel Virasoro. Spin glass theory and beyond. World Scientific Singapore, 1987. [21] Nitish Srivastava, Geoffrey Hinton, Alex Krizhevsky, Ilya Sutskever, and Ruslan Salakhutdinov. Dropout: A simple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 2014. [22] Li Wan, Matthew Zeiler, Sixin Zhang, Yann Lecun, and Rob Fergus. Regularization of neural networks using dropconnect. In Proceedings of the 30th International Conference on Machine Learning (ICML-13), 2013. [23] Deniz Yuret. Knet: beginning deep learning with 100 lines of julia. In Machine Learning Systems Workshop at NIPS 2016, 2016. 

COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM FOR SMOOTHING SPLINE By Zuofeng Shang† and Guang Cheng∗ Indiana University-Purdue University at Indianapolis and Purdue University arXiv:1512.09226v2 [math.ST] 21 Jul 2017 July 19, 2017 In this paper, we explore statistical versus computational tradeoff to address a basic question in the application of a distributed algorithm: what is the minimal computational cost in obtaining statistical optimality? In smoothing spline setup, we observe a phase transition phenomenon for the number of deployed machines that ends up being a simple proxy for computing cost. Specifically, a sharp upper bound for the number of machines is established: when the number is below this bound, statistical optimality (in terms of nonparametric estimation or testing) is achievable; otherwise, statistical optimality becomes impossible. These sharp bounds partly capture intrinsic computational limits of the distributed algorithm considered in this paper, and turn out to be fully determined by the smoothness of the regression function. As a side remark, we argue that sample splitting may be viewed as an alternative form of regularization, playing a similar role as smoothing parameter. 1. Introduction. In the parallel computing environment, divide-and-conquer (D&C) method distributes data to multiple machines, and then aggregates local estimates computed from each machine to produce a global one. Such a distributed algorithm often requires a growing number of machines in order to process an increasingly large dataset. A practically relevant question is “how many processors do we really need in this parallel computing?” or “shall we allocate all our computational resources in the data analysis?” Such questions are related to the minimal computational cost of this distributed method (which will be defined more precisely later). The major goal of this paper is to provide some “theoretical” insights for the above questions from a statistical perspective. Specifically, we consider a classical nonparametric regression setup: (1.1) yl = f (l/N ) + l , l = 0, 1, . . . , N − 1, where l ’s are iid random errors with E{l } = 0 and V ar(l ) = 1, in the following distributed ∗ † Assistant Professor. Corresponding Author. Professor. Research Sponsored by NSF (CAREER Award DMS-1151692, DMS-1418042), Simons Fellowship in Mathematics and Office of Naval Research (ONR N00014-15-1-2331). Guang Cheng gratefully acknowledges Statistical and Applied Mathematical Sciences Institute (SAMSI) for the hospitality and support during his visit in the 2013-Massive Data Program. 1 2 Z. SHANG AND G. CHENG algorithm: Entire Data (N ) Divide −−−−→ Subset 1 (n) Machine 1 Subset 2 (n) Machine 2 ··· Subset s (n) w w Superw  machine fb1 fb2 −→ −→ ··· ··· fbw s w Aggrew gate Machine s −→ Oracle Estimate denoted as fbN D&C Estimate P ¯ f = (1/s) sj=1 fbj We assume that the total sample size is N , the number of machines is s and the size of each subsample is n. Hence, N = s × n. Each machine produces an individual smoothing spline estimate fbj to be defined in (2.2) ([13]). A known property of the above D&C strategy is that it can preserve statistical efficiency for a wide-ranging choice of s (as demonstrated in Figure 1), say log s/ log N ∈ [0, 0.4], while largely reducing computational burden as log s/ log N increases (as demonstrated in Figure 2). An important observation from Figure 1 is that there is an obvious blowup for mean squared errors of f¯ when the above ratio is beyond some threshold, e.g, 0.8 for N = 10000. Hence, we are interested in knowing whether there exists a critical value of log s/ log N in theory, beyond which statistical optimality no longer exists. For example, mean squared errors will never achieve minimax optimal lower bound (at rate level) no matter how smoothing parameters are tuned. Such a sharpness result partly captures the computational limit of the particular D&C algorithm considered in this paper, also complementing the upper bound results in [10, 16, 17] N=500 N=1000 0.4 0.0 0.2 MSE 0.6 N=10000 0.0 0.2 0.4 0.6 0.8 log(s)/log(N) Fig 1. Mean-square errors (MSE) of f¯ based on 500 independent replications under different choices of N and s. The values of MSE stay at low levels for various choice of s with log s/ log N ∈ [0, 0.7]. True regression function is f0 (z) = 0.6b30,17 (z) + 0.4b3,11 (z) with ba1 ,a2 the density function for Beta(a1 , a2 ). COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM 3 2.0 1.0 0.0 Time (seconds) 3.0 N=10000 0.0 0.2 0.4 0.6 0.8 log(s)/log(N) Fig 2. Computing time of f¯ based on a single replication under different choices of s when N = 10, 000. The larger the s, the smaller the computing time. Our first contribution is to establish a sharp upper bound of s under which f¯ achieves the minimax optimal rate N m/(2m+1) , where m represents the smoothness of f0 . By “sharp” upper bound, we mean the largest possible upper bound for s to gain statistical optimality. This result is established by directly computing (non-asymptotic) upper and lower bounds of mean squared error of f¯. These two bounds hold uniformly as s diverges, and thus imply that the rate of mean squared error transits once s reaches the rate N 2m/(2m+1) , which we call as phase transition in divide-and-conquer estimation. In fact, the choice of smoothing parameter, denoted as λ, also plays a very subtle role in the above phase transition. For example, λ is not necessarily chosen at an optimal level when s attains the above bound as illustrated in Figure 3. Our second contribution is a sharp upper bound of s under which a simple Wald-type testing method based on f¯ is minimax optimal in the sense of [6]. It is not surprising that our testing method is consistent no matter s is fixed or diverges at any rate. Rather, this sharp bound is entirely determined by analyzing its (non-asymptotic) power. Specifically, we find that our testing method is minimax optimal if and only if s does not grow faster than N (4m−1)/(4m+1) . Again, we observe a subtle interplay between s and λ as depicted in Figure 3. One theoretical insight obtained in our setup is that a more smooth regression function can be optimally estimated or tested at a shorter time. In addition, the above Figure 3 implies that s and λ play an interchangeable role in obtaining statistical optimality. Therefore, we argue that it might be attempting to view sample splitting as an alternative form of regularization, complementing the use of penalization in smoothing spline. In practice, we propose to select λ via a distributed version of generalized cross validation (GCV); see [14]. In the end, we want to mention that our theoretical results are developed in one-dimensional models under fixed design. This setting allows us to develop proofs based on exact analysis of various Fourier series, coupled with properties of circulant Bernoulli polynomial kernel matrix. The major 4 Z. SHANG AND G. CHENG b b 4m 4 m+1 2m 2m+1 0 2m 2m+1 0 a 4 m−1 4 m+1 a Fig 3. Two lines indicate the choices of s  N a and λ  N −b , leading to minimax optimal estimation rate (left) and minimax optimal testing rate (right). Whereas (a, b)’s outside these two lines lead to suboptimal rates. Results are based on smoothing spline regression with regularity m ≥ 1. goal of this work is to provide some theoretical insights in a relatively simple setup, which are useful in extending our results to more general setup such as random or multi-dimensional design. Efforts toward this direction have been made by [8] who derived upper bounds of s for optimal estimation or testing in various nonparametric models when design is random and multi-dimensional. 2. Smoothing Spline Model. Suppose that we observe samples from model (1.1). The regression function f is smooth in the sense that it belongs to an m-order (m ≥ 1) periodic Sobolev space: (∞ X S m (I) = fν ϕν (·) : ν=1 where I := [0, 1] and for k = 1, 2, . . ., ϕ2k−1 (t) = √ ∞ X ν=1 ) fν2 γν < ∞ , 2 cos(2πkt), ϕ2k (t) = √ 2 sin(2πkt), γ2k−1 = γ2k = (2πk)2m . The entire dataset is distributed to each machine in a uniform manner as follows. For j = 1, . . . , s, the jth machine is assigned with samples (Yi,j , ti,j ), where Yi,j = yis−s+j−1 and ti,j = is − s + j − 1 N for i = 1, . . . , n. Obviously, t1,j , . . . , tn,j are evenly spaced points (with a gap 1/n) across I. At the jth machine, we have the following sub-model: (2.1) Yi,j = f (ti,j ) + i,j , i = 1, . . . , n, where i,j = is−s+j−1 , and obtain the jth sub-estimate as fbj = arg min `j,n,λ (f ). f ∈S m (I) COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM 5 Here, `j,n,λ represents a penalized square criterion function based on the jth subsample: n λ 1 X (2.2) (Yi,j − f (ti,j ))2 + J(f, f ), `j,n,λ (f ) = 2n 2 i=1 R with λ > 0 being a smoothing parameter and J(f, g) = I f (m) (t)g (m) (t)dt1 3. Minimax Optimal Estimation. In this section, we investigate the impact of the number of machines on the mean squared error of f¯. Specifically, Theorem 3.1 provides an (non-asymptotic) upper bound for this mean squared error, while Theorem 3.2 provides a (non-asymptotic) lower bound. Notably, both bounds hold uniformly as s diverges. From these bounds, we observe an interesting phase transition phenomenon that f¯ is minimax optimal if s does not grow faster than N 2m/(2m+1) and an optimal λ  N −2m/(2m+1) is chosen, but the minimax optimality breaks down if s grows even slightly faster (no matter how λ is chosen). Hence, the upper bound of s is sharp. Moreover, λ does not need to be optimal when this bound is attained. In some sense, a proper sample splitting can compensate a sub-optimal choice of λ. In this section, we assume that l ’s are iid zero-mean random variables with unit variance. Denote mean squared error as where kf k2 = qR Theorem 3.1. MSEf0 (f ) := Ef0 {kf − f0 k22 }, 2 I f (t) dt. For simplicity, we write Ef0 as E later. Define h = λ1/(2m) . (Upper Bounds of Variance and Squared Bias) Suppose h > 0, and N is divisible by n. Then there exist absolute positive constants bm , cm ≥ 1 (depending on m only) such that   Z πnh 1 2 −1 −1 ¯ ¯ (3.1) E{kf − E{f }k2 } ≤ bm N + (N h) dx , (1 + x2m )2 0 p (3.2) kE{f¯} − f0 k2 ≤ cm J(f0 )(λ + n−2m + N −1 ) for any fixed 1 ≤ s ≤ N . From (6.2) and (6.3) in Appendix, we can tell that f¯ − E{f¯} is irrelevant to f0 . So is the upper bound for the (integrated) variance in (3.1). However, this is not the case for the (integrated) bias kE{f¯} − f0 k2 , whose upper bound depends on f0 through its norm J(f0 ). In particular, the (integrated) bias becomes zero if f0 is in the null space, i.e., J(f0 ) = 0, according to (3.2). Since MSEf0 (f¯) = E{kf¯ − E{f¯}k22 } + kE{f¯} − f0 k22 , (3.3) Theorem 3.1 says that (3.4)  Z −1 −1 ¯ MSEf0 (f ) ≤ bm N + (N h) 0 1 For simplicity, we denote J(f, f ) = J(f ) later. πnh  1 dx + c2m J(f0 )(λ + n−2m + N −1 ). (1 + x2m )2 6 Z. SHANG AND G. CHENG When we choose h  N −1/(2m+1) and n−2m = O(λ), it can be seen from (3.4) that f¯ is minimax optimal, i.e., kf¯ − f0 k2 = OP (N −m/(2m+1) ). Obviously, the above two conditions hold if λ  N −2m/(2m+1) and s = O(N 2m/(2m+1) ). (3.5) From now on, we define the optimal choice of λ as N −2m/(2m+1) , denoted as λ∗ ; according to [16]. Alternatively, the minimax optimality can be achieved if s  N 2m/(2m+1) and nh = o(1), i.e., λ = o(λ∗ ). In other words, a sub-optimal choice of λ can be compensated by a proper sampling splitting strategy. See Figure 3 for the subtle relation between s and λ. It should be mentioned that λ∗ depends on N (rather than n) for achieving optimal estimation rate. In practice, we propose to select λ via a distributed version of GCV; see [14]. Remark 3.1. Under random design and uniformly bounded eigenfunctions, Corollary 4 in [16] showed that the above rate optimality is achieved under the following upper bound on s (and λ = λ∗ ) s = O(N (2m−1)/(2m+1) / log N ). For example, when m = 2, their upper bound is N 0.6 / log N (versus N 0.8 in our case). We improve their upper bound by applying a more direct proof strategy. To understand whether our upper bound can be further improved, we prove a lower bound result in a “worst case” scenario. Specifically, Theorem 3.2 implies that once s is beyond the above upper bound, the rate optimality will break down for at least one true f0 . Theorem 3.2. (Lower Bound of Squared Bias) Suppose h > 0, and N is divisible by n. Then for any constant C > 0, it holds that sup f0 ∈S m (I) J(f0 )≤C kE{f¯} − f0 k22 ≥ C(am n−2m − 8N −1 ), where am ∈ (0, 1) is an absolute constant depending on m only, for any fixed 1 < s < N . It follows by (3.3) that (3.6) sup f0 ∈S m (I) J(f0 )≤C MSEf0 (f¯) ≥ sup f0 ∈S m (I) J(f0 )≤C kE{f¯} − f0 k22 ≥ C(am n−2m − 8N −1 ). It is easy to check that the above lower bound is strictly slower than the optimal rate N −2m/(2m+1) if s grows faster than N 2m/(2m+1) no matter how λ is chosen. Therefore, we claim that N 2m/(2m+1) is a sharp upper bound of s for obtaining an averaged smoothing spline estimate. 7 COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM In the end, we provide a graphical interpretation for our sharp bound result. Let s = N a for 0 ≤ a ≤ 1 and λ = N −b for 0 < b < 2m. Define ρ1 (a), ρ2 (a) and ρ3 (a) as Upper bound of squared bias: N −ρ1 (a)  λ + n−2m + N −1 , Lower bound of squared bias: N −ρ2 (a)  max{n−2m − N −1 , 0}, Z πnh 1 dx, Upper bound of variance: N −ρ3 (a)  N −1 + (N h)−1 (1 + x2m )2 0 based on Theorems 3.1 and 3.2. A direct examination reveals that ρ1 (a) = min{2m(1 − a), 1, b} ( 2m(1 − a), a > (2m − 1)/(2m) ρ2 (a) = ∞, a ≤ (2m − 1)/(2m) ρ3 (a) = max{a, (2m − b)/(2m)} ρ1 ρ2 2m 2m ρ3 1 ρ2 (a) 2m 2 m+1 2m 2 m+1 2m 2 m+1 ρ3 (a) ρ1 (a) 0 2m 2 m+1 1 a 0 2 m−1 2m 2m 2 m+1 1 a 0 2m 2 m+1 1 a Fig 4. Plots of ρ1 (a), ρ2 (a), ρ3 (a) versus a, indicated by thick solid lines, under λ = N −2m/(2m+1) . ρ1 (a), ρ2 (a) and ρ3 (a) indicate upper bound of squared bias, lower bound of squared bias and upper bound of variance, respectively. ρ2 (a) is plotted only for (2m − 1)/(2m) < a ≤ 1; when 0 ≤ a ≤ (2m − 1)/(2m), ρ2 (a) = ∞, which is omitted. Figure 4 displays ρ1 , ρ2 , ρ3 for λ = N −2m/(2m+1) . It can be seen that when a ∈ [0, 2m/(2m + 1)], upper bounds of squared bias and variance maintain at the same optimal rate N −2m/(2m+1) , while the exact bound of squared bias increases above N −2m/(2m+1) when a ∈ (2m/(2m + 1), 1). This explains why transition occurs at the critical point a = 2m/(2m + 1) (even when the upper bound of variance decreases below N −2m/(2m+1) when a ∈ (2m/(2m + 1), 1)). It should be mentioned that when λ 6= N −2m/(2m+1) , i.e., b 6= 2m/(2m+1), suboptimal estimation almost always occurs. More explicitly, b < 2m/(2m + 1) yields ρ1 (a) < 2m/(2m + 1) for any 0 ≤ a ≤ 1. While b > 2m/(2m + 1) yields ρ2 (a) < 2m/(2m + 1) for any 2m/(2m + 1) < a ≤ 1; yields ρ3 (a) < 2m/(2m + 1) for any 0 ≤ a < 2m/(2m + 1). The only exception is a = 2m/(2m + 1) which yields ρ1 = ρ2 = ρ3 = 2m/(2m + 1) for any b > 2m/(2m + 1). Remark 3.2. As a side remark, we notice that each machine is assigned with n  N 1/(2m+1) samples when s attains its upper bound in the estimation regime. This is very similar as the local 8 Z. SHANG AND G. CHENG polynomial estimation where approximately N 1/(2m+1) local points are used for obtaining optimal estimation (although we realize that our data is distributed in a global manner). Remark 3.3. Under repeated curves with a common design, [2] observed a similar phase tran- sition phenomenon for the minimax rate of a two-stage estimate, where the rate transits when the number of sample curves is nearly N 2m/(2m+1) . This coincides with our observation for s. However, the common design assumption, upon which their results crucially rely, clearly does not apply to our divide-and-conquer setup, and our proof techniques are significantly different. Rather,Theorems 3.1 and 3.2 imply that the results in [2] may still hold for a non-common design. 4. Minimax Optimal Testing. In this section, we consider nonparametric testing: (4.1) H0 : f = 0 v.s. H1 : f ∈ S m (I)\{0}. In general, testing f = f0 (for a known f0 ) is equivalent to testing f∗ ≡ f − f0 = 0. So (4.1) has no loss of generality. Inspired by the classical Wald test ([11]), we propose a simple test statistic based on the f¯ as TN,λ := kf¯k22 . We find that testing consistency essentially requires no condition on the number of machines no matter it is fixed or diverges at any rate. However, our power analysis, which is non-asymptotically valid, depends on the number of machines in a nontrivial way. Specifically, we discover that our test method is minimax optimal in the sense of Ingster ([6]) when s does not grow faster than N (4m−1)/(4m+1) and λ is chosen optimally (different from λ∗ , though), but it is no longer optimal once s is beyond the above threshold (no matter how λ is chosen). This is a similar phase transition phenomenon as we observe in the estimation regime. Again, we notice an optimal choice of λ may not be necessary if the above upper bound of s is achieved. In this section, we assume that the model errors i,j ’s are iid standard normal for technical convenience. In fact, our results can be generalized to likelihood ratio test without assuming Gaussian errors. This extension is possible (technically tedious, though) since likelihood ratio statistic can be approximated by TN,λ through quadratic expansion; see [9]. Theorem 4.1 implies the consistency of our proposed test method with the following testing rule: φN,λ = I(|TN,λ − µN,λ | ≥ z1−α/2 σN,λ ), 2 where µN,λ := EH0 {TN,λ }, σN,λ := VarH0 {TN,λ } and z1−α/2 is the (1 − α/2) × 100 percentile of N (0, 1). The conditions required in Theorem 4.1 are so mild that our proposed testing is consistent no matter the number of machines is fixed or diverges at any rate. COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM Theorem 4.1. 9 (Testing Consistency) Suppose that h → 0, n → ∞ when N → ∞, and limN →∞ nh exists (which could be infinity). Then, we have under H0 , TN,λ − µN,λ d −→ N (0, 1), as N → ∞. σN,λ Our next theorem analyzes the non-asymptotic power of TN,λ , in which we pay particular attention to the impact of s on the separation rate of testing, defined as q dN,λ = λ + n−2m + σN,λ . Let B = {f ∈ S m (I) : J(f ) ≤ C} for a positive constant C. Theorem 4.2. (Upper Bound) Suppose that h → 0, n → ∞ when N → ∞, and limN →∞ nh exists (which could be infinity). Then for any ε > 0, there exist Cε , Nε > 0 s.t. for any N ≥ Nε , (4.2) inf f ∈B kf k2 ≥Cε dN,λ Pf (φN,λ = 1) ≥ 1 − ε. Under assumptions of Theorem 4.1, it can be shown that (see (6.40) in Appendix) ( n , if limN →0 nh = 0, 2 N2 σN,λ  (4.3) 1 , if limN →∞ nh > 0. N 2h p Given a range of λ leading to limN →∞ nh > 0, we have by (4.3) that dN,λ = λ + (N h1/2 )−1 . An optimal choice of λ (satisfying the above requirement) is λ∗∗ := N −4m/(4m+1) since it leads to the optimal separating rate d∗N,λ := N −2m/(4m+1) ; see [6]. Meanwhile, the constraint limN →∞ nh > 0 (together with the choice of λ∗∗ ) implies that s = O(N (4m−1)/(4m+1) ). (4.4) The above discussions illustrate that we can always choose λ∗∗ to obtain a minimax optimal testing (just as in the single dataset [9]) as long as s does not grow faster than N (4m−1)/(4m+1) . In the case that limN →∞ nh = 0, the minimax optimality can be maintained if s  N (4m−1)/(4m+1) , h = o(1) and nh = o(1). Such a selection of s gives us a lot of freedom in choosing λ that needs to satisfy λ = o(λ∗∗ ). A complete picture in depicting the relation between s and λ is given in Figure 3. We further discover in Theorem 4.3 that the upper bound (4.4) turns out to be sharp. Theorem 4.3. (Lower Bound) Suppose that s  N (4m−1)/(4m+1) , h → 0, n → ∞ when N → ∞, and limN →∞ nh exists (which could be infinity). Then there exists a positive sequence βN,λ with limN →∞ βN,λ = ∞ s.t. (4.5) lim sup N →∞ inf f ∈B kf k2 ≥βN,λ d∗N,λ Pf (φN,λ = 1) ≤ α. Recall that 1 − α is the pre-specified significance level. 10 Z. SHANG AND G. CHENG Theorem 4.3 says that when s  N (4m−1)/(4m+1) , the test φN,λ is no longer powerful even when kf k2  d∗N,λ . In other words, our test method fails to be optimal. Therefore, we claim that N (4m−1)/(4m+1) is a sharp upper bound of s to ensure our testing to be minimax optimal. Remark 4.1. As a side remark, the existence of limN →∞ nh can be replaced by the following weaker condition under which the results in Theorems 4.1, 4.2 and 4.3 still hold: Condition (R) : either lim nh = 0 or inf nh > 0. N →∞ N ≥1 Condition (R) aims to exclude irregularly behaved s such as in the following case where s vibrates too much along with N : (4.6) s= ( N b1 , N is odd, N b2 , N is even, where h  N −c for some c > 0, b1 , b2 ∈ [0, 1] satisfy b1 + c ≥ 1 and b2 + c < 1. Clearly, Condition (R) fails under (4.6). 5. Discussions. This paper offers “theoretical” suggestions on the allocation of data. In a relatively simple distributed algorithm, i.e., in m-order periodic splines with evenly spaced design, our recommendation proceeds as follows: • Distribute to s  N 2m/(2m+1) machines for obtaining an optimal estimate; • Distribute to s  N (4m−1)/(4m+1) machines for performing an optimal test. However, data-dependent formulae are still needed in picking a right number of machines in practice. This might be possible in light of Figure 3 indicating that sample splitting could be an alternative form of tuning. As for the choice of λ, we prove that it should be chosen in the order of N even when each subsample has size n. Hence, a distributed version of the generalized cross validation method is applied to each sub-sample; see [14]. Another theoretically interesting direction is how much adaptive estimation (where m is unknown) can affect the computational limits. Acknowledgments We thank PhD student Meimei Liu at Purdue for the simulation study. 6. Appendix. Proofs of our results are included in this section. COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM 11 6.1. Proofs in Section 3. Proof of Theorem 3.1. We do a bit preliminary analysis before proving (3.1) and (3.2). It follows from [13] that (S m (I), J) is a reproducing kernel Hilbert space with reproducing kernel function K(x, y) = ∞ X ϕν (x)ϕν (y) γν ν=1 =2 ∞ X cos(2πk(x − y)) (2πk)2m k=1 , x, y ∈ I. For convenience, define Kx (·) = K(x, ·) for any x ∈ I. It follows from the representer theorem ([13]) that the optimization to problem (2.2) has a solution fbj = (6.1) n X i=1 b ci,j Kti,j , j = 1, 2, . . . , s, where b cj = (b c1,j , . . . , b cn,j )T = n−1 (Σj + λIn )−1 Yj , Yj = (Y1,j , . . . , Yn,j )T , In is n × n identity matrix, and Σj = [K(ti,j , ti0 ,j )/n]1≤i,i0 ≤n . It is easy to see that Σ1 = Σ2 = · · · = Σs . For convenience, denote Σ = Σ1 . Similarly, define K 0 (x, y) = ∞ X ϕν (x)ϕν (y) γν2 ν=1 =2 ∞ X cos(2πk(x − y)) (2πk)4m k=1 , x, y ∈ I. For 1 ≤ j ≤ s, let Ωj = [K 0 (ti,j , ti0 ,j )/n]1≤i,i0 ≤n . It is easy to see that Ω1 = Ω2 = · · · = Ωs . For convenience, denote Ω = Ω1 , and let Φν,j = (ϕν (t1,j ), . . . , ϕν (tn,j )). It is easy to examine that f¯ = (6.2) ∞ X (6.3) j=1 Φν,j (Σ ν=1 ∞ X Ps E{f¯} = ∞ X = ν=1 and Ps N γν j=1 Φν,j (Σ ν=1 + λIn )−1 Yj ϕν + λIn )−1 (f0,j + j ) N γν Ps j=1 Φν,j (Σ + λIn )−1 f0,j N γν ϕν , ϕν , where f0,j = (f0 (t1,j ), . . . , f0 (tn,j ))T and j = (1,j , . . . , n,j )T . We now look at Σ and Ω. For 0 ≤ l ≤ n − 1, let cl = dl = ∞ 2 X cos(2πkl/n) , n (2πk)2m k=1 ∞ 2 X cos(2πkl/n) . n (2πk)4m k=1 Since cl = cn−l and dl = dn−l for l = 1, 2, . . . , n − 1, Σ and Ω are both symmetric circulant of order √ n. Let ε = exp(2π −1/n). Ω and Σ share the same normalized eigenvectors as 1 xr = √ (1, εr , ε2r , . . . , ε(n−1)r )T , r = 0, 1, . . . , n − 1. n 12 Z. SHANG AND G. CHENG Let M = (x0 , x1 , . . . , xn−1 ). Denote M ∗ as the conjugate transpose of M . Clearly, M M ∗ = In and Σ, Ω admits the following decomposition Σ = M Λc M ∗ , Ω = M Λd M ∗ , (6.4) where Λc = diag(λc,0 , λc,1 , . . . , λc,n−1 ) and Λd = diag(λd,0 , λd,1 , . . . , λd,n−1 ) with λc,l = c0 + c1 εl + . . . + cn−1 ε(n−1)l and λd,l = d0 + d1 εl + . . . + dn−1 ε(n−1)l . Direct calculations show that   (6.5) λc,l = P∞  (6.6) λd,l = P∞ 1 k=1 (2πkn)2m , P∞ 1 1 k=0 [2π(kn+l)]2m , k=1 [2π(kn−l)]2m +   2 P∞ 1 k=1 (2πkn)4m , P∞ P∞ 1 1 k=1 [2π(kn−l)]4m + k=0 [2π(kn+l)]4m , 2  It is easy to examine that l = 0, 1 ≤ l ≤ n − 1. l = 0, 1 ≤ l ≤ n − 1. λc,0 = 2c̄m (2πn)−2m , λd,0 = 2d¯m (2πn)−4m , (6.7) and for 1 ≤ l ≤ n − 1, λc,l = 1 1 + 2m [2π(n − l)] (2πl)2m ∞ ∞ X X 1 1 + , + 2m [2π(kn − l)] [2π(kn + l)]2m k=1 k=2 λd,l (6.8) 1 1 = + [2π(n − l)]4m (2πl)4m ∞ ∞ X X 1 1 + + , 4m [2π(kn − l)] [2π(kn + l)]4m k=2 and for c̄m := P∞ k=1 k −2m , cm := P∞ k=2 k cm (2πn)−2m ≤ cm (2πn)−2m ≤ dm (2πn)−4m ≤ dm (2πn)−4m ≤ k=1 −2m , d¯m := P∞ k=1 k −4m , dm := P∞ k=2 k ∞ X 1 ≤ c̄m (2πn)−2m , [2π(kn − l)]2m k=1 ∞ X 1 ≤ d¯m (2πn)−4m , [2π(kn − l)]4m k=2 ∞ X k=2 ∞ X k=1 −4m , 1 ≤ c̄m (2πn)−2m , [2π(kn + l)]2m 1 ≤ d¯m (2πn)−4m . [2π(kn + l)]4m For simplicity, we denote I = E{kf¯−E{f¯}k22 } and II = kE{f¯}−f0 k22 . Hence, MSEf0 (f¯) = I +II. COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM Proof of (3.1) Using (6.4) – (6.8), we get that ∞ Ps −1 2 X j=1 E{|Φν,j (Σ + λIn ) j | } I = N 2 γν2 ν=1 P s ∞ −1 T −1 X j=1 trace((Σ + λIn ) Φν,j Φν,j (Σ + λIn ) ) = N 2 γν2 ν=1 = = = = s ∞ X ΦTν,j Φν,j /n n X −1 trace (Σ + λI ) (Σ + λIn )−1 n N2 γν2 n N2 j=1 s X ν=1 trace (Σ + λIn )−1 Ω(Σ + λIn )−1 j=1 !

1 trace M (Λc + λIn )−1 Λd (Λc + λIn )−1 M ∗ N n−1 λd,l 1 X N (λ + λc,l )2 l=0 ≤ 2d¯m N (2c̄m + (2πn)2m λ)2 n−1 X (2π(n − l))−4m + (2πl)−4m +(1 + d¯m )N −1 (λ + (2π(n − l))−2m + (2πl)−2m )2 l=1 2d¯m ≤ N (2c̄m + (2πn)2m λ)2 X +2(1 + d¯m )N −1 1≤l≤n/2 (2πl)−4m + (2π(n − l))−4m (λ + (2πl)−2m + (2π(n − l))−2m )2 X 2d¯m (2πl)−4m ¯m )N −1 ≤ + 4(1 + d N (2c̄m + (2πn)2m λ)2 (λ + (2πl)−2m )2 1≤l≤n/2 Z 2(1 + d¯m ) πnh 1 2d¯m ≤ + dx N (2c̄m + (2πn)2m λ)2 πN h (1 + x2m )2 0   Z πnh 1 1 1 ≤ bm + dx , N Nh 0 (1 + x2m )2 where bm ≥ 1 is an absolute constant depending on m only. This proves (3.1). Proof of (3.2) √ Throughout, let η = exp(2π −1/N ). For 1 ≤ j, l ≤ s, define Σj,l = σj,l,r = ∞ 1 X ΦTν,j Φν,l , n γν ν=1    ∞ cos 2πk r − j−l X n N 2 , r = 0, 1, . . . , n − 1. n (2πk)2m k=1 13 14 Z. SHANG AND G. CHENG It can be shown that Σj,l is a circulant matrix with elements σj,l,0 , σj,l,1 , . . . , σj,l,n−1 , therefore, by [1] we get that Σj,l = M Λj,l M ∗ , (6.9) where M is the same as in (6.4), and Λj,l = diag(λj,l,0 , λj,l,1 , . . . , λj,l,n−1 ), with λj,l,r , for r = 1, . . . , n − 1, given by the following n−1 X λj,l,r = σj,l,t εrt t=0 2 n n−1 ∞ XX   cos 2πk nt − j−l N  εrt (2πk)2m t=0 k=1 Pn−1 (r−k)t Pn−1 (k+r)t ∞ X η −k(j−l) t=0 ε ε + η k(j−l) t=0 1 n (2πk)2m = = k=1 ∞ ∞ X X η −(qn−r)(j−l) η (qn+r)(j−l) = + , [2π(qn − r)]2m [2π(qn + r)]2m (6.10) q=1 q=0 and for r = 0, given by λj,l,0 = = (6.11) = n−1 X σj,l,t t=0 ∞ P kt k(j−l) 1 X n−1 t=0 ε η n Pn−1 −kt −k(j−l) ε η + t=0 2m (2πk) k=1 ∞ X η qn(j−l) + η −qn(j−l) q=1 (2πqn)2m . For p ≥ 0, 1 ≤ v ≤ n, 0 ≤ r ≤ n − 1 and 1 ≤ j ≤ s, define Ap,v,r,j = s s l=1 l=1 1X 1X λj,l,r x∗r ΦT2(pn+v)−1,l , Bp,v,r,j = λj,l,r x∗r ΦT2(pn+v),l . s s By direct calculation, we have for 1 ≤ v ≤ n − 1,   p Φ2(pn+v)−1,l xr = n/2 η (pn+v)(l−1) I(r + v = n) + η −(pn+v)(l−1) I(v = r) ,   p Φ2(pn+v),l xr = −n/2 η (pn+v)(l−1) I(r + v = n) − η −(pn+v)(l−1) I(v = r) , (6.12) and   p n/2I(r = 0) η (p+1)n(l−1) + η −(p+1)n(l−1) ,   p = −n/2I(r = 0) η (p+1)n(l−1) − η −(p+1)n(l−1) . Φ2(pn+n)−1,l xr = (6.13) Φ2(pn+n),l xr 15 COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM Let I(·) be an indicator function. Then we have for p ≥ 0, 1 ≤ j ≤ s and 1 ≤ v, r ≤ n − 1, s Bp,v,r,j 1X λj,l,r x∗r ΦT2(pn+v),l s l=1   s ∞ ∞ −(qn−r)(j−l) (qn+r)(j−l) X X X p 1 η η   = − −n/2 + s [2π(qn − r)]2m [2π(qn + r)]2m q=1 q=0 l=1   × η −(pn+v)(l−1) I(r + v = n) − η (pn+v)(l−1) I(r = v)  X p η −(pn+v)(j−1) = − −n/2  I(r + v = n) [2π(uN + pn + v)]2m = u≥−p/s − X u≥(p+1)/s η (pn+v)(j−1) I(r = v) [2π(uN − pn − v)]2m X η −(pn+v)(j−1) I(r + v = n) [2π(uN − pn − v)]2m u≥(p+1)/s  X η (pn+v)(j−1) − I(r = v) = ap,v x∗r ΦT2(pn+v),j , [2π(uN + pn + v)]2m + (6.14) u≥−p/s where ap,v = P 1 u≥−p/s [2π(uN +pn+v)]2m + P 1 u≥(p+1)/s [2π(uN −pn−v)]2m , For v = n, similar calculations give that  X p Bp,n,r,j = − −n/2I(r = 0)  u≥−p/s − + X u≥(p+2)/s X u≥(p+2)/s = (6.15) where ap,n = P for p ≥ 0, 1 ≤ v ≤ n − 1. η −(pn+n)(j−1) [2π(uN + pn + n)]2m η (pn+n)(j−1) [2π(uN − pn − n)]2m η −(pn+n)(j−1) [2π(uN − pn − n)]2m ap,n x∗r ΦT2(pn+n),j , 1 u≥−p/s [2π(uN +pn+n)]2m + P − X u≥−p/s η (pn+n)(j−1) [2π(uN + pn + n)]2m 1 u≥(p+2)/s [2π(uN −pn−n)]2m , for p ≥ 0.   16 Z. SHANG AND G. CHENG Similarly, we have p ≥ 0, 1 ≤ j ≤ s and 1 ≤ v, r ≤ n − 1,  X p η −(pn+v)(j−1) Ap,v,r,j = n/2  I(r + v = n) [2π(uN + pn + v)]2m u≥−p/s X + u≥(p+1)/s η (pn+v)(j−1) I(r = v) [2π(uN − pn − v)]2m X η −(pn+v)(j−1) I(r + v = n) [2π(uN − pn − v)]2m u≥(p+1)/s  X η (pn+v)(j−1) + I(r = v) [2π(uN + pn + v)]2m + u≥−p/s = ap,v x∗r ΦT2(pn+v)−1,j , (6.16) and for v = n, Ap,n,r,j  X p = n/2I(r = 0)  u≥−p/s X + η (pn+n)(j−1) [2π(uN − pn − n)]2m u≥(p+2)/s X η −(pn+n)(j−1) [2π(uN − pn − n)]2m u≥(p+2)/s  X η (pn+n)(j−1)  = ap,n x∗r ΦT + 2(pn+n)−1,j . [2π(uN + pn + n)]2m + (6.17) η −(pn+n)(j−1) [2π(uN + pn + n)]2m u≥−p/s It is easy to check that both (6.14) and (6.16) hold for r = 0. Summarizing (6.14)–(6.17), we have that for p ≥ 0, 1 ≤ j ≤ s, 1 ≤ v ≤ n and 0 ≤ r ≤ n − 1, (6.18) Ap,v,r,j = ap,v x∗r ΦT2(pn+v)−1,j , Bp,v,r,j = ap,v x∗r ΦT2(pn+v),j . To show (3.2), let f¯j = (E{f¯(t1,j )}, . . . , E{f¯(tn,j )})T , for 1 ≤ j ≤ s. It follows by (6.3) that ∞ Ps −1 X l=1 Φν,l (Σ + λIn ) f0,l T ¯ fj = Φν,j N γν ν=1 ! s ∞ 1 X 1 X ΦTν,j Φν,l = (Σ + λIn )−1 f0,l s n γν l=1 s X = 1 s = s 1X s ν=1 Σj,l (Σ + λIn )−1 f0,l l=1 l=1 M Λj,l (Λc + λIn )−1 M ∗ f0,l , 17 COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM together with (6.18), leading to that s M ∗ f¯j = 1X Λj,l (Λc + λIn )−1 M ∗ f0,l s l=1  1 Ps ∗ T = =   fµ0   µ=1 ∞ X l=1 λj,l,n−1 x∗n−1 ΦT µ,l λ+λc,n−1  1 Ps ∞ X n X +  0  f2(pn+v)−1  p=0 v=1 ∞ X n X p=0 v=1 +   0  f2(pn+v)    0  f2(pn+v)−1  ∞ X n X p=0 v=1 1 s   ∞ X n X l=1 s   0 f2(pn+v)   p=0 v=1 ∞ X n X p=0 v=1 = Ps .. .   0 f2(pn+v)−1   p=0 v=1 ∞ X n X + = 1 s  λj,l,0 x0 Φµ,l λ+λc,0 l=1 s   0  f2(pn+v)  Ps l=1 1 s 1 s Ps l=1  λj,l,0 x∗0 ΦT 2(pn+v)−1,l λ+λc,0 .. . λj,l,n−1 x∗n−1 ΦT 2(pn+v)−1,l λ+λc,n−1 l=1 Ps      λj,l,0 x∗0 ΦT 2(pn+v),l λ+λc,0 .. . λj,l,n−1 x∗n−1 ΦT 2(pn+v),l λ+λc,n−1 Ap,v,0,j λ+λc,0 .. . Ap,v,n−1,j λ+λc,n−1 Bp,v,0,j λ+λc,0 .. . Bp,v,n−1,j λ+λc,n−1              ap,v ∗ T λ+λc,0 x0 Φ2(pn+v)−1,j .. . ap,v ∗ T λ+λc,n−1 xn−1 Φ2(pn+v)−1,j ap,v ∗ T λ+λc,0 x0 Φ2(pn+v),j .. . ap,v ∗ T λ+λc,n−1 xn−1 Φ2(pn+v),j      .      18 Z. SHANG AND G. CHENG On the other hand, M ∗ f0,j = = ∞ X fµ0 M ∗ ΦTµ,j µ=1 ∞ X n X 0 M ∗ ΦT2(pn+v)−1,j f2(pn+v)−1 p=0 v=1 + ∞ X n X 0 M ∗ ΦT2(pn+v),j f2(pn+v) p=0 v=1   ∗ ΦT x 0 2(pn+v)−1,j ∞ X n X   .. 0   = f2(pn+v)−1 .   p=0 v=1 ∗ T xn−1 Φ2(pn+v)−1,j   ∗ ΦT x 0 2(pn+v),j ∞ X n X   .. 0 .  + f2(pn+v) .   p=0 v=1 T ∗ xn−1 Φ2(pn+v),j Therefore, M ∗ (f¯j − f0,j ) = ∞ X n X p=0 v=1 + (6.19)   0  f2(pn+v)−1  ∞ X n X p=0 v=1 where bp,v,r = ap,v λ+λc,r   0  f2(pn+v)  bp,v,0 x∗0 ΦT2(pn+v)−1,j .. .     bp,v,n−1 x∗n−1 ΦT2(pn+v)−1,j  bp,v,0 x∗0 ΦT2(pn+v),j  .. , .  ∗ T bp,v,n−1 xn−1 Φ2(pn+v),j − 1, for p ≥ 0, 1 ≤ v ≤ n and 0 ≤ r ≤ n − 1. It holds the trivial observation bks+g,v,r = bg,v,r for k ≥ 0, 0 ≤ g ≤ s − 1, 1 ≤ v ≤ n √ P∞ 0 0 and 0 ≤ r ≤ n − 1. Define Cg,r = −1f2(kN k=0 (f2(kN +gn+n−r)−1 − +gn+n−r) ) and Dg,r = √ P∞ 0 0 k=0 (f2(kN +gn+r)−1 + −1f2(kN +gn+r) ), for 0 ≤ g ≤ s − 1 and 0 ≤ r ≤ n − 1. Also denote Cg,r and Dg,r as their conjugate. By (6.12) and (6.13), and direct calculations we get that, for 1 ≤ j ≤ s and 1 ≤ r ≤ n − 1, δj,r ≡ = (6.20) ∞ n X X p=0 r n 2 0 f2(pn+v)−1 bp,v,r x∗r ΦT2(pn+v)−1,j v=1 ∞ h X p=0 + n X 0 bp,v,r x∗r ΦT2(pn+v),j f2(pn+v) v=1 0 (f2(pn+n−r)−1 − 0 +(f2(pn+r)−1 + √ √ 0 −1f2(pn+n−r) )bp,n−r,r η −(pn+n−r)(j−1) i 0 −1f2(pn+r) )bp,r,r η (pn+r)(j−1) , ! 19 COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM leading to that s X j=1 2 |δj,r | = s ∞ √ n X Xh 0 0 )bp,n−r,r η −(pn+n−r)(j−1) (f2(pn+n−r)−1 − −1f2(pn+n−r) 2 p=0 j=1 0 +(f2(pn+r)−1 + = 0 −1f2(pn+r) )bp,r,r η (pn+r)(j−1) i 2 s s−1  n X X Cg,r bg,n−r,r η −(gn+n−r)(j−1) + Dg,r bg,r,r η (gn+r)(j−1) 2 g=0 j=1 = √ n 2 s−1 X s X (Cg,r bg,n−r,r η −(gn+n−r)(j−1) + Dg,r bg,r,r η (gn+r)(j−1) ) g,g 0 =0 j=1 0 0 ×(Cg0 ,r bg0 ,n−r,r η (g n+n−r)(j−1) + Dg0 ,r bg0 ,r,r η −(g n+r)(j−1) ) = s−1 NX (|Cg,r |2 b2g,n−r,r + Cg,r Ds−1−g,r bg,n−r,r bs−1−g,r,r 2 g=0 +Dg,r Cs−1−g,r bg,r,r bs−1−g,n−r,r + |Dg,r |2 b2g,r,r ) (6.21) = s−1 NX |Cg,r bg,n−r,r + Ds−1−g,r bs−1−g,r,r |2 2 g=0 ≤ N = N s−1 X g=0 s−1 X g=0 (|Cg,r |2 b2g,n−r,r + |Ds−1−g,r |2 b2s−1−g,r,r ) (|Cg,r |2 b2g,n−r,r + |Dg,r |2 b2g,r,r ). It is easy to see that for 0 ≤ g ≤ s − 1 and 1 ≤ r ≤ n − 1, |Cg,r | 2 = ( ∞ X +( k=0 ≤ ∞ X ≤ k=0 × (6.22) and 2 |Dg,r | 0 2 f2(kN +gn+n−r) ) 0 2 0 2 2m (|f2(kN +gn+n−r)−1 | + |f2(kN +gn+n−r) | )(kN + gn + n − r) k=0 ∞ X ∞ X k=0 k=0 ∞ X × (6.23) 0 2 f2(kN +gn+n−r)−1 ) (kN + gn + n − r)−2m 0 2 0 2 2m (|f2(kN +gn+n−r)−1 | + |f2(kN +gn+n−r) | )(kN + gn + n − r) 2m (gn + n − r)−2m , 2m − 1 ≤ ∞ X k=0 × 0 2 0 2 2m (|f2(kN +gn+r) | + |f2(kN +gn+r)−1 | )(kN + gn + r) 2m (gn + r)−2m . 2m − 1 2 20 Z. SHANG AND G. CHENG For 1 ≤ g ≤ s − 1, we have ag,n−r ≤ λc,r , which further leads to |bg,n−r,r | ≤ 2. Meanwhile, by (6.5), we have 0 ≤ λc,r − a0,r ≤ (2π(n − r))−2m + 2c̄m (2πn)−2m ≤ (1 + 2c̄m )(2π(n − r))−2m . Then we have λ + λc,r − a0,r λ + λc,r λ + (1 + 2c̄m )(2π(n − r))−2m ≤ λ + (2πr)−2m + (2π(n − r))−2m λ + (2π(n − r))−2m ≤ (1 + 2c̄m ) , λ + (2πr)−2m + (2π(n − r))−2m |b0,r,r | = leading to 2 λ + (2π(n − r))−2m ≤ r (1 + 2c̄m ) λ + (2π(n − r))−2m + (2πr)−2m   λ + (2π(n − r))−2m −2m 2 ≤ r (1 + 2c̄m ) λ + (2π(n − r))−2m + (2πr)−2m r−2m b20,r,r −2m 2  ≤ (2π)2m (1 + 2c̄m )2 (λ + (πn)−2m ). (6.24) The last inequality can be proved in two different cases: 2r ≤ n and 2r > n. Similarly, it can be shown that (n − r)−2m b20,n−r,r ≤ (2π)2m (1 + 2c̄m )2 (λ + (πn)−2m ). Then we have by (6.22)–(6.24) that n−1 s−1 XX r=1 g=0 |Dg,r |2 b2g,r,r ≤ n−1 s−1 X ∞ XX r=1 g=1 k=0 0 2 0 2 2m (|f2(kN +gn+r) | + |f2(kN +gn+r)−1 | )(kN + gn + r) n−1 ∞ XX 2m 0 2 0 2 2m × (gn + r)−2m 22 + (|f2(kN +r) | + |f2(kN +r)−1 | )(kN + r) 2m − 1 r=1 k=0 2m −2m 2 × r b0,r,r 2m − 1 n−1 s−1 X ∞ XX 0 2 0 2 ≤ c0m (λ + n−2m ) (|f2(kN +gn+r) | + |f2(kN +gn+r)−1 | ) r=1 g=0 k=0 2m ×(2π(kN + gn + r)) (6.25) , 2m 8m , (1 + 2c̄m )2 2m−1 }. Similarly, one can show that where c0m = max{(2π)−2m 2m−1 n−1 s−1 XX r=1 g=0 (6.26) |Cg,r |2 b2g,n−r,r ≤ c0m (λ +n −2m ) n−1 s−1 X ∞ XX r=1 g=0 k=0 2m ×(2π(kN + gn + r)) . 0 2 0 2 (|f2(kN +gn+r) | + |f2(kN +gn+r)−1 | ) 21 COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM Combining (6.25) and (6.26) we get that n−1 s XX r=1 j=1 (6.27) |δj,r |2 ≤ 2c0m (λ + n−2m )N n−1 s−1 X ∞ XX r=1 g=0 k=0 2m ×(2π(kN + gn + r)) 2 2 0 0 (|f2(kN +gn+r) | + |f2(kN +gn+r)−1 | ) . To the end of proof of (3.2), by (6.19) we have for 1 ≤ j ≤ s, δj,0 ≡ = ∞ n X X 0 f2(pn+v)−1 bp,v,0 x∗0 ΦT2(pn+v)−1,j + n X 0 f2(pn+v) bp,v,0 x∗0 ΦT2(pn+v),j p=0 v=1 ∞  X 0 0 f2(pn+n)−1 bp,n,0 x∗0 ΦT2(pn+n)−1,j + f2(pn+n) bp,n,0 x∗0 ΦT2(pn+n),j v=1 p=0 = r = r !  √ n Xh 0 0 )bp,n,0 η −(p+1)n(j−1) (f2(pn+n)−1 − −1f2(pn+n) 2 p=0 i √ 0 0 +(f2(pn+n)−1 + −1f2(pn+n) )bp,n,0 η (p+1)n(j−1) r s−1 " ∞ √ nX X 0 0 −(gn+n)(j−1) = (f2(kN +gn+n)−1 − −1f2(kN +gn+n) )bg,n,0 η 2 g=0 k=0 # ∞ X √ 0 0 (gn+n)(j−1) + (f2(kN +gn+n)−1 + −1f2(kN +gn+n) )bg,n,0 η (6.28) ∞ k=0 s−1 h X n 2 g=0 i Cg,0 bg,n,0 η −(gn+n)(j−1) + Dg,n bg,n,0 η (gn+n)(j−1) , which, together with Cauchy-Schwartz inequality, (6.22)–(6.23), and the trivial fact |bg,n,0 | ≤ 2 for 0 ≤ g ≤ s − 1, leads to s X j=1 |δj,0 | 2 ≤ n s s−1 X X j=1 Cg,0 bg,n,0 η +n g=0 s s−1 X X j=1   s−1 s−1 X X = N |Cg,0 |2 b2g,n,0 + |Dg,n |2 b2g,n,0  g=0 (6.29) −(gn+n)(j−1) 2 ≤ 2c0m n−2m N Dg,n bg,n,0 η (gn+n)(j−1) 2 g=0 g=0 s−1 X ∞ X g=0 k=0 0 2 0 2 2m (|f2(kN . +gn+n)−1 | + |f2(kN +gn+n) | ) × (2π(kN + gn + n)) Combining (6.27) and (6.29) we get that s X n X j=1 i=1 (E{f¯}(ti,j ) − f0 (ti,j ))2 = n−1 s−1 XX r=0 g=0 |δj,r |2 ≤ 2c0m (λ + n−2m )N (6.30) ×(2π(kN + gn + ∞ n X s−1 X X 0 2 0 2 (|f2(kN +gn+i) | + |f2(kN +gn+i)−1 | ) i=1 g=0 k=0 i))2m = 2c0m (λ + n−2m )N J(f0 ). 22 Z. SHANG AND G. CHENG Next we will apply (6.30) to show (3.2). Since fbj is the minimizer of `j,n,λ (f ), it satisfies for 1 ≤ j ≤ s, n 1X − (Yi,j − fbj (ti,j ))Kti,j + λfbj = 0. n i=1 Taking expectations, we get that n 1X (E{fbj }(ti,j ) − f0 (ti,j ))Kti,j + λE{fbj }, n i=1 therefore, E{fbj } is the minimizer to the following functional n `0j (f ) = 1 X λ (f (ti,j ) − f0 (ti,j ))2 + J(f ). 2n 2 i=1 Define gj = E{fbj }. Since `0j (gj ) ≤ `0j (f0 ), we get n λ 1 X λ (gj (ti,j ) − f0 (ti,j ))2 + J(gj ) ≤ J(f0 ). 2n 2 2 i=1 This means that J(gj ) ≤ J(f0 ), leading to (6.31) Note that E{f¯} = 1 s Ps and m ≥ 1 we get that s s j=1 j=1 p 1 X (m) 1 X (m) k gj k2 ≤ kgj k2 ≤ J(f0 ). s s j=1 gj . Define g(t) = (E{f¯}(t) − f0 (t))2 . By [4, Lemma (2.24), pp. 58], (6.31) Z 1 N −1 1 X g(l/N ) − g(t)dt N 0 l=0 (6.32) ≤ 2 N Z 0 ≤ 2 1 k N s ≤ 2 1 k N s 1 s s 1X 0 1X gj (t) − f0 (t) × gj (t) − f00 (t) dt s s j=1 s X j=1 s X j=1 j=1 gj − f0 k2 × k (m) gj 1 s (m) 2 k2 − f0 s X j=1 ≤ gj0 − f00 k2 8J(f0 ) . N Combining (6.30) and (6.32) we get that kE{f¯} − f0 k22 ≤ c2m J(f0 )(λ + n−2m + N −1 ), where c2m = max{8, 2c0m }. This completes the proof of (3.2). P 0 0 Proof of Theorem 3.2. Suppose f0 = ∞ ν=1 fν ϕν with fν satisfying ( Cn−1 (2π(n + r))−2m , ν = 2(n + r) − 1, 1 ≤ r ≤ n/2, 0 2 (6.33) |fν | = 0, otherwise. COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM It is easy to see that J(f0 ) = P 0 2 1≤r≤n/2 |f2(n+r)−1 | (2π(n 23 + r))2m ≤ C. Consider the decomposition (6.19) and let δj,r be defined as in (6.20) and (6.28). It can be easily checked that Cg,r = 0 for 1 ≤ r ≤ n/2 and 0 ≤ g ≤ s − 1. Furthermore, for 1 ≤ r ≤ n/2, λc,r − a1,r = ∞ X (2π(un + r))−2m + u=0 ∞ X − u=1 ∞ ∞ X X (2π(un − r))−2m − (2π(uN + n + r))−2m u=1 u=0 (2π(uN − n − r))−2m ≥ (2πr)−2m . Therefore, b21,r,r = ≥ (6.34)   λ + λc,r − a1,r λ + λc,r 2 λ + (2πr)−2m λ + 2(1 + c̄m )(2πr)−2m 2 ≥ 1 . 4(1 + c̄m )2 Using (6.21) and (6.34), we have s X j=1 (f¯j − f0,j )T (f¯j − f0,j ) = ≥ = s n−1 X X j=1 r=0 X X X 1≤r≤n/2 = X 1≤r≤n/2 ≥ where am = 1 16(3π)2m (1+c̄m )2 X 1≤r≤n/2 |δj,r |2 s−1 NX |Cg,r bg,n−r,r + Ds−1−g,r bs−1−g,r,r |2 2 g=0 s−1 NX |Ds−1−g,r |2 b2s−1−g,r,r 2 g=0 s−1 NX |Dg,r |2 b2g,r,r 2 g=0 N |D1,r |2 b21,r,r 2 ≥ N 8(1 + c̄m )2 ≥ 16(3π)2m (1 X 1≤r≤n/2 NC + c̄m )2 0 |f2(n+r)−1 |2 n−2m ≡ am N Cn−2m , < 1 is an absolute constant depending on m only. Then the conclusion follows by (6.32). Proof is completed. 6.2. Proofs in Section 4. s X 1≤r≤n/2 j=1 1≤r≤n/2 = |δj,r |2 24 Z. SHANG AND G. CHENG Proof of Theorem 4.1. For 1 ≤ j, l ≤ s, define Ωj,l = σ ej,l,r = ∞ 1 X ΦTν,j Φν,l , n γν2 ν=1    ∞ cos 2πk r − j−l X n N 2 , r = 0, 1, . . . , n − 1. 4m n (2πk) k=1 Clearly Ωj,l is a circulant matrix with elements σ ej,l,0 , σ ej,l,1 , . . . , σ ej,l,n−1 . Furthermore, by arguments (6.9)–(6.11) we get that Ωj,l = M Γj,l M ∗ , (6.35) where M is the same as in (6.4), and Γj,l = diag(δj,l,0 , δj,l,1 , . . . , δj,l,n−1 ), with δj,l,r , for r = 1, . . . , n− 1, given by the following (6.36) δj,l,r = ∞ ∞ X X η −(qn−r)(j−l) η (qn+r)(j−l) + , [2π(qn − r)]4m [2π(qn + r)]4m q=1 q=0 and for r = 0, given by (6.37) δj,l,0 = ∞ X η qn(j−l) + η −qn(j−l) q=1 (2πqn)4m Define A = diag((Σ + λIn )−1 , . . . , (Σ + λIn )−1 ) and {z } | . s  Ω1,1 Ω1,2 · · · Ω1,s   Ω2,1 Ω2,2 · · · B=  ··· ···  ··· Ωs,1 Ωs,2 · · ·   Ω2,s  .  ···  Ωs,s Note that B is N × N symmetric. Under H0 , it can be shown that kf¯k22 = ∞  Ps X ν=1 = + λIn )−1 l N γν2 l=1 Φν,l (Σ s 1 X T j (Σ + λIn )−1 Ns j,l=1 s X = 1 Ns = 1 T 1 T  ABA =  ∆, Ns Ns 2 ∞ 1 X ΦTν,j Φν,l n γν2 ν=1 ! Tj (Σ + λIn )−1 Ωj,l (Σ + λIn )−1 l j,l=1 where  = (T1 , . . . , Ts )T and ∆ ≡ ABA. (Σ + λIn )−1 l 25 COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM 2 This implies that TN,λ = T ∆/(N s) with µN,λ = trace(∆)/(N s) and σN,λ = 2trace(∆2 )/(N s)2 . Define U = (TN,λ − µN,λ )/σN,λ . Then for any t ∈ (−1/2, 1/2), log E{exp(tU )} = log E{exp(tT ∆/(N sσN,λ ))} − tµN,λ /σN,λ 1 = − log det(IN − 2t∆/(N sσN,λ )) − tµN,λ /σN,λ 2 2 = ttrace(∆)/(N sσN,λ ) + t2 trace(∆2 )/((N s)2 σN,λ ) 3 +O(t3 trace(∆3 )/((N s)3 σN,λ )) − tµN,λ /σN,λ 3 )). = t2 /2 + O(t3 trace(∆3 )/((N s)3 σN,λ 3 ) = o(1) in order to conclude the proof. It remains to show that trace(∆3 )/((N s)3 σN,λ 2 ) and trace(∆3 ). We start from the In other words, we need to study trace(∆2 ) (used in σN,λ former. By direct calculations, we get trace(∆2 ) = trace(A2 BA2 B)   s s X X trace  M (Λc + λIn )−2 Γl,j (Λc + λIn )−2 Γj,l M ∗  = j=1 l=1 = s X j,l=1 s n−1 X
X trace (Λc + λIn )−2 Γl,j (Λc + λIn )−2 Γj,l = j,l=1 r=0 For 1 ≤ g ≤ s and 0 ≤ r ≤ n − 1, define Ag,r = ∞ X p=0 1 . [2π(pN + gn − r)]4m Using (6.36) and (6.37), it can be shown that for r = 1, 2, . . . , n − 1, s X j,l=1 2 |δj,l,r | = s s X X Ag,r η −gn(j−l) j,l=1 g=1 = s X j,l=1 +   + s X Ag,n−r η s X 0 Ag,r Ag0 ,r η −(g−g )n(j−l) + = s2 (6.38) ≥ s2 g=1 s X Ag,r Ag0 ,n−r η A2g,r + s2 −(g+g 0 −1)n(j−l) + (6.39) g=1 A2g,r = s X Ag,n−r Ag0 ,r η (g+g 0 −1)n(j−l) s X g=1 s X A2g,n−r + 2s2 s X Ag,r As+1−g,n−r g=1 A2g,n−r . g=1 Since s X 0 Ag,n−r Ag0 ,n−r η (g−g )n(j−l) g,g 0 =1 A2g,r + s2 g=1 s X g,g 0 =1 g,g 0 =1 s X 2 (g−1)n(j−l) g=1 g,g 0 =1 s X |δj,l,r |2 . (λ + λc,r )4 s X g=1  ∞ X  p=0 2 1 1  ≥ , 4m [2π(pN + gn − r)] [2π(n − r)]8m   26 Z. SHANG AND G. CHENG we get that n−1 X 2 trace(∆ ) ≥ P P s2 ( sg=1 A2g,r + sg=1 A2g,n−r ) (λ + λc,r )4 r=1 1 n−1 X [2π(n−r)] 8m 2 ≥ s (λ + λc,r r=1 ≥ 2s2 (2 + 2c̄m )4 = s2 8(1 + c̄m )4 X 1≤r≤n/2 X 1≤r≤n/2 s2 h−1 8(1 + c̄m )4 ≥ + Z 1 (2πr)8m )4 (λ 1 (2πr)8m 1 4 + (2πr) 2m ) 1 (1 + (2πrh)2m )4 nh/2 h 1 dx. (1 + (2πx)2m )4 Meanwhile, (6.38) indicates that for 1 ≤ r ≤ n − 1, s X j,l=1 |δj,l,r |2 ≤ 2s2 s X A2g,r + 2s2 g=1 s X A2g,n−r . g=1 From (6.39) we get that for 1 ≤ r ≤ n − 1, s X g=1 A2g,r ≤ cm , (2π(n − r))8m where cm > 0 is a constant depending on m only. Similar analysis to (6.38) shows that s X j,l=1 2 |δj,l,0 | s s X X = Ag,0 (η gn(j−l) + η −gn(j−l) ) 2 j,l=1 g=1 = 2s 2 ≤ 4s2 s X g=1 s X g=1 A2g,0 + 2s 2 s−1 X Ag,0 As−g,0 + 2s2 A2s,0 g=1 A2g,0 ≤ cm s2 (2πn)−8m . Therefore, 2 trace(∆ ) ≤ 4s2 Ps (λ 2 g=1 Ag,0 + λc,0 )4 n X 2 + 2s n−1 X r=1 Ps 2 g=1 Ag,r (λ + 1 (1 + (2πrh)2m )4 r=1 Z nh 1 ≤ 4cm s2 h−1 dx. 2m )4 (1 + (2πx) 0 ≤ 4cm s 2 + Ps 2 g=1 Ag,n−r λc,r )4 27 COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM By the above statements, we get that 2 σN,λ = 2trace(∆2 )/(N s)2  (6.40) ( n , N2 1 , N 2h if nh → 0, if limN nh > 0. To the end, we look at the trace of ∆3 . By direct examinations, we have trace(∆3 ) = trace(ABA2 BA2 BA) s s X X = trace [ M (Λc + λIn )−2 Γj,l (Λc + λIn )−2 Γl,k M ∗ ] j,k=1 = l=1
λIn )−2 Γk,j M ∗ ×M (Λc + s X
trace (Λc + λIn )−2 Γj,l (Λc + λIn )−2 Γl,k (Λc + λIn )−2 Γk,j j,k,l=1 = s n−1 X X δj,l,r δl,k,r δk,j,r . (λ + λc,r )6 j,k,l=1 r=0 For r = 1, 2, . . . , n − 1, it can be shown that   ∞ ∞ qn(j−l) −qn(j−l) X X η η  + δj,l,r δl,k,r δk,j,r =  (2π(qn − r))4m (2π(qn + r))4m q=0 q=1   ∞ ∞ −qn(l−k) qn(l−k) X X η η  × + 4m (2π(qn − r)) (2π(qn + r))4m q=1 q=0   ∞ ∞ −qn(k−j) qn(k−j) X X η η . × (6.41) + 4m (2π(qn − r)) (2π(qn + r))4m q=1 q=0 We next proceed to show that for 1 ≤ r ≤ n − 1, (6.42) s X l,j,k=1 δj,l,r δl,k,r δk,j,r 96m ≤ 12m − 1  4m 4m − 1 3 3 s  1 1 + 12m (2π(n − r)) (2πr)12m  . 28 Z. SHANG AND G. CHENG Using the trivial fact that Ag,r ≤ s ∞ X X = = = = ≤ ≤ j,l,k=1 q1 =1 s s X X 4m 4m−1 1 , (2π(gn−r))4m the first term in (6.41) satisfies ∞ ∞ X X η −q1 n(j−l) η −q2 n(j−l) η −q3 n(j−l) (2π(q1 n − r))4m (2π(q2 n − r))4m (2π(q3 n − r))4m Ag1 ,r η −g1 n(j−l) j,l,k=1 g1 =1 s X q2 =1 s X Ag2 ,r η −g2 n(l−k) g2 =1 Ag1 ,r Ag2 ,r Ag3 ,r g1 ,g2 ,g3 =1 s X Ag1 ,r Ag2 ,r Ag3 ,r  3 g1 ,g2 ,g3 =1 s X s3 A3g,r g=1 4m 4m − 1 × s X q3 =1 s X Ag3 ,r η −g3 n(k−j) g3 =1 η −g1 n(j−l) η −g2 n(l−k) η −g3 n(k−j) j,l,k=1 s s s X X X (g3 −g1 )n(j−1) (g1 −g2 )n(l−1) (g2 −g3 )n(k−1) η η j=1 η l=1 k=1 s X 1 (2π(gn − r))12m g=1  3 12m 4m 1 s3 . 12m − 1 4m − 1 (2π(n − r))12m s3 Similarly, one can show that all other terms in (6.41) are upper bounded by  3   12m 4m 1 1 3 s + . 12m − 1 4m − 1 (2π(n − r))12m (2πr)12m Therefore, (6.42) holds. It can also be shown by (6.37) and similar analysis that s X (6.43) j,l,k=1 δj,l,0 δl,k,0 δk,j,0 ≤ s3 (2πn)−12m . Using (6.42) and (6.43), one can get that s n−1 X X δj,l,r δl,k,r δk,j,r (λ + λc,r )6 3 trace(∆ ) = l,j,k=1 r=0 . s 3 r=1 n X 1 (2π(n−r))12m + 1 (2πr)12m (λ + λc,r ) +s 3 (λ 1 (2πn)12m + λc,0 )12m 1 (1 + (2πrh)2m )6 r=1 ( Z nh s3 n, if nh → 0, 1 3 −1 . s h dx  2m 6 (1 + (2πx) ) s3 h−1 , if limN nh > 0. 0 . s3 (6.44) n−1 X Combining (6.40) and (6.44), and using the assumptions n → ∞, h → 0, we get that ( n−1/2 , if nh → 0, 3 3 3 trace(∆ )/((N s) σN,λ ) . = o(1). h1/2 , if limN nh > 0. Proof is completed. COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM 29 Proof of Theorem 4.2. Throughout the proof, we assume that data Y1 , . . . , YN are generated from the sequence of alternative hypotheses: f ∈ B and kf k2 ≥ Cε dN,λ . Define fj = (f (t1,j ), . . . , f (tn,j ))T for 1 ≤ j ≤ s. Then it can be shown that N sTN,λ = N s = = = ∞ X f¯ν2 ν=1 s X YjT (Σ + λIn )−1 Ωj,l (Σ + λIn )−1 Yl j,l=1 s X YjT M (Λc + λIn )−1 Γj,l (Λc + λIn )−1 M ∗ Yl j,l=1 s X fTj M (Λc + λIn )−1 Γj,l (Λc + λIn )−1 M ∗ fl j,l=1 s X + fTj M (Λc + λIn )−1 Γj,l (Λc + λIn )−1 M ∗ l j,l=1 s X Tj M (Λc + λIn )−1 Γj,l (Λc + λIn )−1 M ∗ fl + j,l=1 s X + Tj M (Λc + λIn )−1 Γj,l (Λc + λIn )−1 M ∗ l j,l=1 ≡ T1 + T2 + T3 + T4 . (6.45) Next we will analyze all the four terms in the above. Let f = 1 ≤ l ≤ s, define dl,r = dl,r = x∗r fl . ∞ X n X Then it holds that f2(pn+v)−1 x∗r ΦT2(pn+v)−1,l + p=0 v=1 Using (6.12) and (6.13), we get that for 1 ≤ r ≤ n − 1, dl,r (6.46) ∞ X n X P∞ ν=1 fν ϕν . For 0 ≤ r ≤ n − 1 and f2(pn+v) x∗r ΦT2(pn+v),l . p=0 v=1 r    n η −(pn+v)(l−1) I(r + v = n) + η (pn+v)(l−1) I(r = v) = f2(pn+v)−1 2 p=0 v=1  r  ∞ n−1  X X n  −(pn+v)(l−1) + f2(pn+v) − − η I(r + v = n) − η (pn+v)(l−1) I(r = v) 2 p=0 v=1 r ∞ h √ nX = (f2(pn+n−r)−1 − −1f2(pn+n−r) )η −(pn+n−r)(l−1) 2 p=0 i √ +(f2(pn+r)−1 + −1f2(pn+r) )η (pn+r)(l−1) , ∞ n−1 X X 30 Z. SHANG AND G. CHENG and for r = 0, dl,0 = ∞ X f2(pn+n)−1 x∗0 ΦT2(pn+n)−1,l + = (6.47) f2(pn+n) x∗0 ΦT2(pn+n),l p=0 p=0 r ∞ X n 2 ∞ h X p=0 (f2(pn+n)−1 − +(f2(pn+n)−1 + √ √ −1f2(pn+n) )η −(pn+n)(l−1) i −1f2(pn+n) )η (pn+n)(l−1) . We first look at T1 . It can be examined directly that s X T1 = (dj,0 , . . . , dj,n−1 )diag j,l=1 (6.48) = n−1 X r=0 Ps  δj,l,n−1 δj,l,0 ,..., 2 (λ + λc,0 ) (λ + λc,n−1 )2  × (dl,0 , . . . , dl,n−1 )T j,l=1 δj,l,r dj,r dl,r . (λ + λc,r )2 Using similar arguments as (6.14)–(6.18), one can show that for p ≥ 0, 1 ≤ v ≤ n, 0 ≤ r ≤ n − 1 and 1 ≤ j ≤ s, s 1X δj,l,r x∗r ΦT2(pn+v)−1,l = bp,v x∗r ΦT2(pn+v)−1,j , s l=1 s 1X δj,l,r x∗r ΦT2(pn+v),l = bp,v x∗r ΦT2(pn+v),j , s (6.49) l=1 where bp,v =  P   P 1 1 u≥−p/s (2π(uN +pn+v))4m + u≥(p+1)/s (2π(uN −pn−v))4m , for 1 ≤ v ≤ n − P P 1 1 u≥(p+2)/s (2π(uN −pn−n))4m , for v = n. u≥−p/s (2π(uN +pn+n))4m + 1, COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM 31 By (6.49), we have s X δj,l,r dj,r dl,r = s X dj,r j=1 j,l=1 = s X l=1 dj,r j=1 + s X s X l=1 ∞ X n X p=0 v=1 = s X j=1 +  dj,r  ∞ X n X δj,l,r dl,r  ∞ X n X  δj,l,r f2(pn+v)−1 x∗r Φ2(pn+v)−1,l p=0 v=1 f2(pn+v) x∗r ΦT2(pn+v),l  ∞ X n X f2(pn+v)−1 p=0 v=1 f2(pn+v) p=0 v=1 = s s X j=1 +  dj,r  ∞ X n X p=0 v=1  p=0 v=1 δj,l,r x∗r ΦT2(pn+v)−1,l l=1 s X l=1 ∞ X n X s X  δj,l,r x∗r ΦT2(pn+v),l  f2(pn+v)−1 bp,v x∗r ΦT2(pn+v)−1,j  f2(pn+v) bp,v x∗r ΦT2(pn+v),j  . It then follows from (6.46) and (6.47), trivial facts bs−1−g,r = bg,n−r and Cg,n−r = Dg,r (both Cg,r and Dg,r are defined similarly as those in the proof of Theorem 3.1, but with f0 therein replaced 32 Z. SHANG AND G. CHENG by f ), and direct calculations that for 1 ≤ r ≤ n − 1 s X √ sn X X h (f2(pn+n−r)−1 + −1f2(pn+n−r) )η (pn+n−r)(j−1) 2 j=1 p=0 i √ +(f2(pn+r)−1 − −1f2(pn+r) )η −(pn+r)(j−1) ∞ s δj,l,r dj,r dl,r = j,l=1 ∞ h X √ × (f2(pn+n−r)−1 − −1f2(pn+n−r) )bp,n−r η −(pn+n−r)(j−1) p=0 +(f2(pn+r)−1 + √ −1f2(pn+r) )bp,r η (pn+r)(j−1) i s s−1 ∞ √ N XXXh (f2(kN +gn+n−r)−1 + −1f2(kN +gn+n−r) )η (gn+n−r)(j−1) 2 j=1 g=0 k=0 i √ +(f2(kN +gn+r)−1 − −1f2(kN +gn+r) )η −(gn+r)(j−1) = × s−1 X ∞ h X g=0 k=0 (f2(kN +gn+n−r)−1 − √ −1f2(kN +gn+n−r) )bks+g,n−r η −(gn+n−r)(j−1) i √ +(f2(kN +gn+r)−1 + −1f2(kN +gn+r) )bks+g,r η (gn+r)(j−1)   s s−1 s−1 X X X N  = Cg,r η (gn+n)(j−1) + Dg,r η −gn(j−1)  2 g=0 j=1 g=0   s−1 s−1 X X × bg,n−r Cg,r η −(gn+n)(j−1) + bg,r Dg,r η gn(j−1)  g=0 =  g=0 s−1 s−1 X N s X bg,n−r |Cg,r |2 + bs−1−g,r Cg,r Ds−1−g,r 2 g=0 g=0  s−1 s−1 X X + bs−1−g,n−r Dg,r Cs−1−g,r + bg,r |Dg,r |2  , g=0 which leads to (6.50) n−1 X r=1 g=0 Ps j,l=1 δj,l,r dj,r dl,r (λ + λc,r )2 Since J(f ) ≤ C, equivalently, (6.51) P∞ X n−1 Ns X = 2 2 ν=1 (f2ν−1 r=1 Ps−1 g=0 bg,r |Cs−1−g,r + (λ + λc,r )2 Dg,r |2 . 2 )(2πν)2m ≤ C, we get that + f2ν 2 2 (f2r−1 + f2r ) ≥ kf k22 − C(2πn)−2m . 1≤r≤n/2 Meanwhile, for 1 ≤ r ≤ n/2, using similar arguments as (6.25) and (6.26) one can show that there 33 COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM exists a constant c0m relying on C and m s.t. |Cs−1,r + D0,r |2 = f2r−1 + ∞ X k=0 + f2r + ∞ X k=1 and 2 |Cs−1,r + D0,r | (2πr) 2m " f2(kN +r) − ∞ X k=0 f2(kN +r)−1 k=1 f2(kN +N −r) !2 !2 1 2 2 (f + f2r−1 ) − c0m N −2m , 2 2r−1 ≥ (6.52) f2(kN +N −r)−1 + ∞ X ∞ ∞ X X 2 ≤ 4 ( f2(kN +N −r)−1 ) + ( f2(kN +N −r) )2 k=0 ∞ X +( ≤ k=0 2 f2(kN +r)−1 ) + ( k=0 ∞ X 4 k=0 +4 ∞ X k=0 +4 +4 ∞ X k=0 ∞ X ∞ X f2(kN +r) ) k=0 2 f2(kN +N −r)−1 (2π(kN # (2πr)2m 2m + N − r)) ∞ X k=0 (2π(kN + N − r))−2m ∞ X 2 2m f2(kN +N −r) (2π(kN + N − r)) (2π(kN + N − r))−2m k=0 ∞ X 2 2m f2(kN (2π(kN + r)) (2π(kN + r))−2m +r)−1 2 f2(kN +r) (2π(kN 8m 2m − 1 8m + 2m − 1 2m + r)) k=0 ≤ 2 k=0 ∞ X −2m (2π(kN + r)) k=0 ∞ X ! × (2πr)2m 2 2 −2m (f2(kN +N −r)−1 + f2(kN +N −r) )γkN +N −r (2π(N − r)) k=0 ∞ X 2 2 −2m (f2(kN +r)−1 + f2(kN +r) )γkN +r (2πr) k=0 ! × (2πr)2m , which, together with the fact N ≥ 2r for 1 ≤ r ≤ n/2, leads to that X (6.53) |Cs−1,r + D0,r |2 (2πr)2m ≤ c0m . 1≤r≤n/2 Furthermore, it can be verified that for 1 ≤ r ≤ n/2, λ2c,r − b0,r (2πr)−2m ≤ (λ + λc,r ) ((2πr)−2m + (2π(n − r))−2m + c̄m (2πn)−2m )2 − (2πr)−4m (2πr)−2m ((2πr)−2m + (2π(n − r))−2m )2 ≤ c0m n−2m , (6.54) which leads to that (λ + λc,r )2 − b0,r (2πr)−2m = (λ + λc,r )2 (6.55) λ2c,r − b0,r λ2 + 2λλc,r −2m (2πr) + (2πr)−2m (λ + λc,r )2 (λ + λc,r )2 ≤ 2λ + c0m n−2m . 34 Z. SHANG AND G. CHENG Then, using (6.48)–(6.50) and (6.51)–(6.55) one gets that T1 ≥ = (6.56) ≥ Ns 2 X 1≤r≤n/2  b0,r |Cs−1,r + D0,r |2 (λ + λc,r )2 X X )2  (λ + λc,r − b0,r Ns  |Cs−1,r + D0,r |2 − |Cs−1,r + D0,r |2  2 (λ + λc,r )2 1≤r≤n/2 1≤r≤n/2   Ns 1 2 0 −2m 0 −2m 0 0 −2m kf k2 − cm n − cm N − cm (2λ + cm n ) ≥ C 0 N sσN,λ , 2 2 where the last inequality follows by kf k22 ≥ 4C 0 (λ + n−2m + σN,λ ) for a large constant C 0 satisfying 2C 0 > 2c0m + (c0m )2 . To achieve the desired power, we need to enlarge C 0 further. This will be described later. Combining (6.56) with (6.40) and (6.56) we get that (6.57) T1  s uniformly for f ∈ B with kf k22 ≥ 4C 0 d2N,λ . Terms T2 and T3 can be handled similarly. To handle T2 , note that T2 = fT ∆, where f = (fT1 , . . . , fTs )T ,  = (T1 , . . . , Ts )T , and ∆ is defined in the proof of Theorem 4.1. We need to establish ∆ ≤ sIN . Define an arbitrary a = (aT1 , . . . , aTs )T ∈ RN , where each aj is an (real) n-vector. Let ξj = M ∗ aj and ξ = (ξ1T , . . . , ξsT )T . For simplicity, put ξj = (ξj,0 , . . . , ξj,n−1 )T for 1 ≤ j ≤ s. Then based on (6.37) and (6.36), we have aT ∆a = ξ ∗ [(Λc + λIn )−1 Γj,l (Λc + λIn )−1 ]1≤j,l≤s ξ = n−1 X s X n−1 X s ξj,r ξl,r r=0 j,l=1 ≤ r=1 + ≤ s 2s  δj,l,r (λ + λc,r )2 1 (2π(n−r))4m P∞ (λ Ps 1 q=1 (2πqn)4m n−1 s XX r=0 j=1 P 1 (2πr)4m + λc,r )2 + (λ + λc,0 )2 s 2 j=1 |ξj,r | 2 j=1 |ξj,0 | |ξj,r |2 = sξ ∗ ξ = saT a, therefore, ∆ ≤ sIN . This leads to that, uniformly for f ∈ B with kf k22 ≥ 4C 0 d2N,λ , Ef {T22 } = fT ∆2 f ≤ sT1 . Together with (6.57), we get that   1/2 sup (6.58) Pf |T2 | ≥ ε−1/2 T1 s1/2 ≤ ε. f ∈B √ kf k2 ≥2 C 0 dN,λ Note that (6.58) also applies to T3 . By Theorem 4.1, (T4 /(N s) − µN,λ )/σN,λ is OP (1) uniformly for f . Therefore, we can choose Cε0 > 0 s.t. Pf (|T4 /(N s) − µN,λ |/σN,λ ≥ Cε0 ) ≤ ε as N → ∞. COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM 35 It then follows by (6.56), (6.57) and (6.58) that for suitable large C 0 (e.g., C 0 ≥ 2(Cε0 + z1−α/2 )), √ uniformly for f ∈ B with kf k2 ≥ 2 C 0 dN,λ ,
Pf |TN,λ − µN,λ |/σN,λ ≥ z1−α/2 ≤ 3ε, as N → ∞. Proof is completed. Proof of Theorem 4.3. Define BN = bN 2/(4m+1) c, the integer part of N 2/(4m+1) . We prove the theorem in two cases: limN nh > 0 and nh = o(1). Case I: limN nh > 0. In this case, it can be shown by s  N (4m−1)/(4m+1) (equivalently n  BN , leading to BN h  nh hence BN h → ∞) that n−6m h−4m+1/2 N  (BN /n)6m . Choose g to be an integer satisfying n−6m h−4m+1/2 N  g 6m  (BN /n)6m . (6.59) Construct an f = (6.60) P∞ fν2 = ν=1 fν ϕν ( with C n−1 (2π(gn + r))−2m , ν = 2(gn + r) − 1, r = 1, 2, . . . , n − 1, 0, otherwise. It can be seen that (6.61) J(f ) = s−1 X 2 f2(gn+r)−1 (2π(gn + r))2m = C, r=1 and n−1 X kf k22 = 2 f2(gn+r)−1 r=1 n−1 C X (2π(gn + r))−2m n−1 = r=1 (6.62) 2 ≥ C(2π(gn + n))−2m = βN,λ N −4m/(4m+1) , 2 where βN,λ = C[BN /(2π(gn + n))]2m . Due to (6.59) and n  BN , we have gn + n  2BN , which further implies βN,λ → ∞ as N → ∞. Using the trivial fact bs−2−g,n = bg,n for 0 ≤ g ≤ s − 2, one can show that  s−2 s s−1 X X Ns  X 2 δj,l,0 dj,0 dl,0 = 2 |Cg0 ,0 |2 bg0 ,n + Cs−2−g0 ,0 Cg0 ,0 bg0 ,n + Cs−1,0 bs−1,n 2 j,l=1 g 0 =0 g 0 =0  s−2 X 2 Ds−2−g0 ,n Dg0 ,n bg0 ,n + Ds−1,n bs−1,n  + g 0 =0 (6.63) ≤ 2N s s−1 X g 0 =0 |Cg0 ,0 |2 bg0 ,n = 0, 36 Z. SHANG AND G. CHENG where the last equality follows by a trivial observation Cg0 ,0 = 0. It follows by (6.63), (6.48) and (6.50) that n−1 T1 = ≤ = (6.64) = Ns X 2 Ps−1 g 0 =0 bg 0 ,r |Cs−1−g 0 ,r (λ + λc,r )2 + Dg0 ,r |2 r=1 P P n−1 n−1 2 2 X s−1 X s−1 g 0 =0 bg 0 ,r |Cs−1−g 0 ,r | g 0 =0 bg 0 ,r |Dg 0 ,r | Ns + N s (λ + λc,r )2 (λ + λc,r )2 r=1 r=1 P P n−1 n−1 2 2 X s−1 X s−1 g 0 =0 bs−1−g 0 ,n−r |Cg 0 ,n−r | g 0 =0 bg 0 ,r |Dg 0 ,r | Ns + N s (λ + λc,r )2 (λ + λc,r )2 r=1 r=1 P 2 n−1 n−1 2 X s−1 X bg,r f2(gn+r)−1 g 0 =0 bg 0 ,r |Dg 0 ,r | 2N s = 2N s , (λ + λc,r )2 (λ + λc,r )2 r=1 r=1 where the last equality follows from the design of f , i.e., (6.60). Now it follows from (6.64) and the fact bg,r ≤ c0m (2π(gn + r))−4m , for some constant c0m depending on m only, that T1 ≤ 2N s = ≤ (6.65) n−1 X c0m (2π(gn r=0 C + r))−4m n−1 (2π(gn + r))−2m (λ + λc,r )2 n−1 2N sc0m C X (2π(gn + r))−6m n−1 (λ + λc,r )2 r=1 0 2cm C(2π)−6m N s(gn)−6m h−4m  sh−1/2  N sσN,λ , where the last “” follows from (4.3). By (6.58) we have that 1/2 |T2 + T3 | = T1 s1/2 OPf (1) = oPf (sh−1/4 ) = oPf (N sσN,λ ). Hence, by (6.45) and Theorem 4.1 we have TN,λ − µN,λ σN,λ = = T1 + T2 + T3 T4 /(N s) − µN,λ + N sσN,λ σN,λ T4 /(N s) − µN,λ d + oPf (1) −→ N (0, 1). σN,λ Consequently, as N → ∞ inf f ? ∈B kf ? k2 ≥βN,λ N −2m/(4m+1) Pf ? (φN,λ = 1) ≤ Pf (φN,λ = 1) → α. This shows the desired result in Case I. Case II: nh = o(1). The proof is similar to Case I although a bit technical difference needs to be emphasized. Since n  BN , it can be shown that N n−2m−1/2  (BN /n)6m . Choose g to be an integer satisfying (6.66) N n−2m−1/2  g 6m  (BN /n)6m . 37 COMPUTATIONAL LIMITS OF A DISTRIBUTED ALGORITHM Let f = P∞ J(f ) = C ν=1 fν ϕν and kf k22 with fν satisfying (6.60). Similar to (6.61) and (6.62) one can show that 2 N −4m/(4m+1) , where β 2 2m . It is clear that ≥ βN,λ N,λ = C[BN /(2π(gn + n))] βN,λ → ∞ as N → ∞. Then similar to (6.63), (6.48), (6.50) and (6.65) one can show that T1 n−1 X 2 bg,r f2(gn+r)−1 ≤ 2N s ≤ n−1 2N sc0m C X (2π(gn + r))−6m n−1 (λ + λc,r )2 ≤ r=1 (λ + λc,r )2 r=1 0 2cm C(2π)−2m N sg −6m n−2m  sn1/2  N sσN,λ , where the last line follows by (6.66) and (4.3). Then the desired result follows by arguments in the rest of Case I. Proof is completed. REFERENCES [1] Brockwell, P. J. and Davis, R. A. (1987). Time Series: Theory and Methods. Springer, New York. [2] Cai, T. T. and Yuan, M. (2011). Optimal Estimation of the Mean Function Based on Discretely Sampled Functional Data: Phase Transition. Annals of Statistics, 39, 2330–2355. [3] de Jong, P. (1987). A central limit theorem for generalized quadratic forms. Probability Theory & Related Fields, 75, 261–277. [4] Eggermont, P. P. B. and LaRiccia, V. N. (2009). Maximum Penalized Likelihood Estimation: Volume II. Springer Series in Statistics. [5] Hastie, T., Tibshirani, R. and Friedman, J. (2001). The Elements of Statistical Learning. Springer. [6] Ingster, Yu I. (1993). Asymptotically minimax hypothesis testing for nonparametric alternatives I–III. Mathematical Methods of Statistics, 2, 85–114; 3, 171–189; 4, 249–268. [7] Kosorok, M. R. (2008). Introduction to Empirical Processes and Semiparametric Inference. Springer: New York. [8] Liu, M., Shang, Z. and Cheng, G. (2017). How Many Machines Can We Use in Parallel Computing? Preprint. [9] Shang, Z. and Cheng, G. (2013). Local and Global Asymptotic Inference in Smoothing Spline Models. Annals of Statistics, 41, 2608–2638. [10] Shang, Z. and Cheng, G. (2015). A Bayesian Splitotic Theory for Nonparametric Models. Arxiv:1508.04175. [11] Shao, J. (2003). Mathematical Statistics, 2nd Ed. Springer Texts in Statistics. Springer, New York. [12] Sollich, P. and Williams, C. K. (2005). Understanding Gaussian process regression using the equivalent kernel. In Deterministic and statistical methods in machine learning. Springer, 211-228. [13] Wahba, G. (1990). Spline Models for Observational Data. SIAM, Philidelphia. [14] Xu, G., Shang, Z. and Cheng, G. (2017). Optimal tuning for divide-and-conquer kernel ridge regression with massive data. arXiv preprint arXiv:1612.05907 [15] Yang, Y., Shang, Z. and Cheng, G. (2017). Non-asymptotic theory for nonparametric testing. https://arxiv.org/abs/1702.01330. [16] Zhang, Y., Duchi, J. C., Wainwright, M. J. (2015). Divide and Conquer Kernel Ridge Regression: A Distributed Algorithm with Minimax Optimal Rates. Journal of Machine Learning Research, 16, 3299–3340. 38 Z. SHANG AND G. CHENG [17] Zhao, T., Cheng, G. and Liu, H. (2016) A Partially Linear Framework for Massive Heterogeneous Data. Annals of Statistics, 44, 1400-1437. [18] Zhou, D.-X. (2002). The Covering Number in Learning Theory. Journal of Complexity, 18, 739-767. Department of Mathematical Sciences Department of Statistics Indiana University-Purdue University at Indianapolis Purdue University 420 University Blvd 250 N. University Street Indianapolis, IN 46202 West Lafayette, IN 47907 Email: zuofengshang@gmail.com Email: chengg@purdue.edu 

A Homological Theory of Functions Greg Yang arXiv:1701.02302v3 [math.AC] 6 Apr 2017 Harvard University gyang@college.harvard.edu April 7, 2017 Abstract In computational complexity, a complexity class is given by a set of problems or functions, and a basic challenge is to show separations of complexity classes A 6= B especially when A is known to be a subset of B. In this paper we introduce a homological theory of functions that can be used to establish complexity separations, while also providing other interesting consequences. We propose to associate a topological space SA to each class of functions A, such that, to separate complexity classes A ⊆ B0 , it suffices to observe a change in “the number of holes”, i.e. homology, in SA as a subclass B ⊆ B0 is added to A. In other words, if the homologies of SA and SA∪B are different, then A 6= B0 . We develop the underlying theory of functions based on combinatorial and homological commutative algebra and Stanley-Reisner theory, and recover Minsky and Papert’s result [12] that parity cannot be computed by nonmaximal degree polynomial threshold functions. In the process, we derive a “maximal principle” for polynomial threshold functions that is used to extend this result further to arbitrary symmetric functions. A surprising coincidence is demonstrated, where the maximal dimension of “holes” in SA upper bounds the VC dimension of A, with equality for common computational cases such as the class of polynomial threshold functions or the class of linear functionals in F2 , or common algebraic cases such as when the Stanley-Reisner ring of SA is Cohen-Macaulay. As another interesting application of our theory, we prove a result that a priori has nothing to do with complexity separation: it characterizes when a vector subspace intersects the positive cone, in terms of homological conditions. By analogy to Farkas’ result doing the same with linear conditions, we call our theorem the Homological Farkas Lemma. 1 1.1 Introduction Intuition Let A ⊆ B0 be classes of functions. To show that B0 6= A, it suffices to find some B ⊆ B0 such that A ∪ B 6= A. In other words, we want to add something to A and watch it change. Let’s take a step back Consider a more general setting, where A and B are “nice” subspaces of a larger topological space C. We can produce a certificate of A ∪ B 6= A by observing a difference in the number of “holes” of A ∪ B and A. Figure 1 shows two examples of such certificates. 1 1.1 Intuition B A A B (a) A and B are both contractible (do not have holes), (b) A has a hole in its center, but B covers it, so that but their union A ∪ B has a hole. A ∪ B is now contractible. Figure 1: Certifying A ∪ B 6= A by noting that the numbers of 1-dimensional holes are different between A ∪ B and A. Sometimes, however, there could be no difference between the number of holes in A ∪ B and A. For example, if B in Figure 1a is slightly larger, then A ∪ B no longer has a hole in the center (see Figure 2). But if we take a slice of A ∪ B, we observe a change in the number of connected components (zeroth dimensional holes) from A to A ∪ B. L L∩A L ∩ (A ∪ B) A B Figure 2: A ∪ B and A are both contractible, but if we look at a section L of A ∪ B, we see that L ∩ A has 2 connected components, but L ∩ (A ∪ B) has only 1. From this intuition, one might daydream of attacking complexity separation problems this way: 1. For each class A, associate a unique topological space (specifically, a simplicial complex) SA . 2. Compute the number of holes in SA and SA∪B of each dimension, and correspondingly for each section by an affine subspace. 3. Attempt to find a difference between these quantities (a “homological” certificate). It turns out this daydream is not so dreamy after all! This work is devoted to developing such a homological theory of functions for complexity separation, which incidentally turns out to have intricate connection to other areas of computer science and combinatorics. Our main results can be summarized as follows: 1) Through our homological framework, we recover Marvin Minsky and Seymour Papert’s classical result that polynomial threshold functions do not compute parity unless degree is maximal [12], and in fact we discover multiple proofs, each “coresponding to a different hole”; the consideration of lower dimension holes yields a maximal principle for polynomial threshold functions that is used to extend Minsky and Papert’s result to arbitrary symmetric functions [3]. 2) We show that an algebraic/homological 2 1 INTRODUCTION quantity arising in our framework, the homological dimension dimh A of a class A, upper bounds the VC dimension dimVC A of A. Informally, this translates to the following remarkable statement: “The highest dimension of any holes in SA or its sections upper bounds the number of samples needed to learn an unknown function from A, up to multiplicative constants.” We furthermore show that equality holds in many common cases in computation (for classes like polynomial thresholds, F2 linear functionals, etc) or in algebra (when the Stanley-Reisner ring of SA is Cohen-Macaulay). 3) We formulate the Homological Farkas Lemma, which characterizes by homological conditions when a linear subspace intersects the interior of the positive cone, and obtain a proof for free from our homological theory of functions. While the innards of our theory relies on homological algebra and algebraic topology, we give an extended introduction in the remainder of this section to the flavor of our ideas in what follows, assuming only comfort with combinatorics, knowledge of basic topology, and a geometric intuition for “holes.” A brief note about notation: [n] denotes the set {0, . . . , n − 1}, and [n → m] denotes the set of functions from domain [n] to codomain [m]. The notation f :⊆ A → B specifies a partial function from domain A to codomain B. † represents the partial function with empty domain. 1.2 An Embarassingly Simple Example Let linfund ∼ = (Fd2 )∗ be the class of linear functionals of a d-dimensional vector space V over F2 . If d ≥ 2, then linfund does not compute the indicator function I1 of the singleton set {1 := 11 · · · 1}. This is obviously true, but let’s try to reason via a “homological way.” This will provide intuition for the general technique and set the stage for similar analysis in more complex settings. Let g : 0 → 0, 1 → 1. Observe that for every partial linear functional h ⊃ g strictly extending g, I1 intersects h nontrivially. (Because I1 is zero outside of g, and every such h must send at least one element to zero outside of g). I claim this completes the proof. Why? Combinatorially, this is because if I1 were a linear functional, then for any 2-dimensional subspace W of V containing {0, 1}, the partial function h :⊆ V → F2 , dom h = W , ( g(u) if u ∈ dom g h(u) = 1 − I1 (u) if u ∈ dom h \ dom g is a linear functional, and by construction, does not intersect I1 on W \ {0, 1}. Homologically, we are really showing the following The space associated to linfund , in its section by an affine subspace corresponding to g, “has a hole” that is “filled up” when I1 is added to linfund . “Wait, what? I’m confused. I don’t see anything in the proof resembling a hole?” 1.3 The Canonical Suboplex OK. No problem. Let’s see where the holes come from. 3 1.3   0 7→ 0 1   0 7→ 1 1   0 7→ 0 1   1 7→ 0 0   0 7→ 0 1   0 7→ 1 1   1 7→ 0 0 The Canonical Suboplex   1 7→ 1 0 [1 0]   1 7→ 1 1   1 7→ 1 0   1 7→ 1 0 [0 0] [0 1]   1 7→ 0 0   1 7→ 0 1 [0 0]   1 7→ 1 1 [0 1]   1 7→ 1 1 [1 0]   1 7→ 0 1 [1 1] (a) Step 1 and Step 2 for linfun02 . Step 1: Each simplex is labeled with a function f ∈ linfun02 , represented as a row vector. Step 2: Each vertex of each simplex is labeled by an input/output pair, here presented in the form of a column vector to a scalar. The collection of input/output pairs in a simplex Ff recovers the graph of f . Each face of Ff has an induced partial function label, given by the collection of input/output pairs on its vertices (not explicitly shown).   1 7→ 0 1 [1 1]   0 7→ 1 1 (b) Step 3 for linfun02 . The simplices Ff are glued together according to their labels. For example, F[0 0] and F[0 1] are glued together by their vertices with the common label [1 0]T 7→ 0, and not anywhere else because no other faces share a common label. Figure 3 Let’s first define the construction of the simplicial complex SC associated to any function class C, called the canonical suboplex. In parallel, we give the explicit construction in the case of C = linfun0d := linfun2  {0 7→ 0}. This is the same class as linfun2 , except we delete 0 from the domain of every function. It gives rise to essentially the same complex as linfun2 , and we will recover Slinfun2 explicitly at the end. Pick a domain, say [n] = {0, . . . , n − 1}. Let C ⊆ [n → 2] be a class of boolean functions on [n]. We construct a simplicial complex SC as follows: 1. To each f ∈ C we associate an (n − 1)-dimensional simplex Ff ∼ = 4n−1 , which will be a facet of SC . 2. Each of the n vertices of Ff is labeled by an input/output pair i 7→ f (i) for some i ∈ [n], and each face G of Ff is labeled by a partial function f ⊆ f , whose graph is specified by the labels of the vertices of G. See Figure 3a for the construction in Step 1 and Step 2 for linfun02 . 3. For each pair f, g ∈ C, Ff is glued together with Fg along the subsimplex G (in both facets) with partial function label f ∩ g. See Figure 3b for the construction for linfun02 . This is the simplicial complex associated to the class C, called the canonical suboplex SC of C. Notice that in the case of linfun0d , the structure of “holes” is not trivial at all: Slinfun0d has 3 holes in dimension 1 but no holes in any other dimension. An easy way to visualize this it to pick one of the triangular holes; If you put your hands around the edge, pull the hole wide, and flatten the entire complex onto a flat plane, then you get Figure 4a. It is easy to construct the canonical suboplex of linfund from that of linfun0d : Slinfund is just a cone over Slinfun0d , where the cone vertex has the label [0 0]T 7→ 0 (Figure 4b). This is because every function in linfund shares this input/output pair. Note that a cone over any base has no hole in any dimension, because any hole can be contracted to a point in the vertex of the cone. This is a fact we will use very soon. Let’s give another important example, the class of all functions. If C = [n → 2], then one can see that SC is isomorphic to the 1-norm unit sphere (also known as orthoplex) S1n−1 := {kxk1 = 4 1 INTRODUCTION (a) The shape obtained by stretching Slinfun0d along one of its triangular holes and then flatten everything onto a flat plane. This deformation preserves all homological information, and from this picture we see easily that Slinfun0d has 3 holes, each of dimension 1. (b) The canonical suboplex of linfund is just a cone over that of linfun0d . Here we show the case d = 2. Figure 4 0 7→ 1 a 7→ b 2 7→ 1 SC(a7→b) 1 7→ 0 1 7→ 1 SC 2 7→ 0 0 7→ 0 (a) The canonical suboplex of [3 → 2]. (b) SC(a7→b) is an affine section of SC . (c) we may recover Slinfun0d as a linear cut through the “torso” of Slinfund . Figure 5 1 : x ∈ Rn } (Figure 5a). For general C, SC can be realized as a subcomplex of S1n−1 . Indeed, for C = linfun02 ⊆ [3 → 2], it is easily seen that SC is a subcomplex of the boundary of an octahedron, which is isomorphic to S12 . Let C ⊆ [n → 2], and let f :⊆ [n] → [2] be a partial function. Define the filtered class C  f to be {g \ f : g ∈ C, g ⊇ f} ⊆ [[n] \ dom f → [2]] Unwinding the definition: C  f is obtained by taking all functions of C that extend f and ignoring the inputs falling in the domain of f. The canonical suboplex SCf can be shown to be isomorphic to an affine section of SC , when the latter is embedded as part of the L1 unit sphere S1n−1 . Figure 5b shows an example when f has a singleton domain. Indeed, recall linfun0d is defined as linfund  {0 7→ 0}, and we may recover Slinfun0d as a linear cut through the “torso” of Slinfund (Figure 5c). “OK. I see the holes. But how does this have anything to do with our proof of I1 6∈ linfund ?” 5 1.4 Nerve Lemma Figure 6: A continuous deformation of Slinfun02 into a complete graph with 4 vertices (where we ignore the sharp bends of the “outer” edges). Hold on tight! We are almost there. First let me introduce a “duality principle” in algebraic topology called the Nerve Lemma. Readers familiar with it can skip ahead to the next section. 1.4 Nerve Lemma Note that the canonical suboplex of linfun02 can be continuously deformed as shown in Figure 6 into a 1-dimensional complex (a graph), so that all of the holes are still preserved. Such a deformation produces a complex • whose vertices correspond exactly to the facets of the original complex, and • whose edges correspond exactly to intersections of pairs of facets, all the while preserving the holes of the original complex, and producing no new ones. Such an intuition of deformation is vastly generalized by the Nerve Lemma: Lemma 1.1 (Nerve Lemma (Informal)). Let U = {Ui }i be a “nice” cover (to be explained below) of a topological space X. The nerve NU of U is defined as the simplicial complex with vertices T {Vi : Ui ∈ U}, and with simplices {Vi }i∈S for each index set S such that {Ui : i ∈ S} is nonempty. Then, for each dimension d, the set of d-dimensional holes in X is bijective with the set of d-dimensional holes in NU . What kind of covers are nice? Open covers in general spaces, or subcomplex covers in simplicial (or CW) complexes, are considered “nice”, if in addition they satisfy the following requirements (acyclicity). P • Each set of the cover must have no holes. • Each nontrivial intersection of a collection of sets must have no holes. The example we saw in Figure 7 is an application of the Nerve Lemma for the cover by facets. Another example is the star cover: For vertex V in a complex, the open star St V of V is defined as the union of all open simplices whose closure meets V (see Figure 7 for an example). If the cover U consists of the open stars of every vertex in a simplicial complex X, then NU is isomorphic to X as complexes. Figure 7: The open star St P of vertex P OK! We are finally ready to make the connection to complexity! 6 1 INTRODUCTION   1 7→ 1 1   1 7→ 1 1 I1 (a) The canonical suboplex of linfun2  {[0 0]T 7→ 0, [1 1]T 7→ 1} is isomorphic to the affine section as shown, and it has two disconnected components, and thus “a single zeroth dimensional hole.” (b) When we add I1 to linfund to obtain D := linfund ∪ {I1 }, SDg now does not have any hole! Figure 8 1.5 The Connection It turns out that Slinfun0d = Slinfund (07→0) (a complex of dimension 2d − 2) has holes in dimension d − 1. The proof is omitted here but will be given in Section 2.3.6. This can be clearly seen in our example when d = 2 (Figure 4a), which has 3 holes in dimension d − 1 = 1. Furthermore, for every partial linear functional h (a linear functional defined on a linear subspace), Slinfund h also has holes, in dimension d − 1 − dim(dom h). Figure 8a show an example for d = 2 and h = [1 1]T 7→ 1. But when we add I1 to linfund to obtain D := linfund ∪ {I1 }, SDg now does not have any hole! Figure 8b clearly demonstrates the case d = 2. For general d, note that Slinfun0d has a “nice” cover by the open stars C := {St V : V has label u 7→ r for some u ∈ Fd2 \ {0} and r ∈ F2 }. When we added I1 to form D, the collection C 0 := C ∪ 4I1 obtained by adding the simplex of I1 to C is a “nice” cover of SD . Thus the nerve NC 0 has the same holes as SD , by the Nerve Lemma. But observe that NC 0 is a cone! . . . which is what our “combinatorial proof” of I1 6∈ linfund really showed. More precisely, a collection of stars S := {St V : V ∈ V} has nontrivial intersection iff there is a partial linear functional extending the labels of each V ∈ V. We showed I1 intersects every partial linear functional strictly extending g : 0 7→ 0, 1 7→ T 1. Therefore, a 0 I1 collection of stars S in C intersects nontrivially iff (S∪{4I1 }) 6= ∅. 0 In other words, in the nerve of C , 4I1 forms the vertex of a cone over all other St V ∈ C. In our example of linfun2 , this is demonstrated in Figure 9. Thus, to summarize, • NC 0 , being a cone, has no holes. • By the Nerve Lemma, SDg has no holes either. • Since Slinfund g has Figure 9: The nerve N 0 overC holes, we know D 6= linfund , i.e. I1 6∈ linfund , as desired. layed on D = linfun2 ∪ {I1 }. While this introduction took some length to explain the logic of Note that NC 0 is a cone over its our approach, much of this is automated in the theory we develop base of 2 points. in this paper, which leverages existing works on Stanley-Reisner theory and cellular resolutions. *** 7 1.6 Dimension theory In our proof, we roughly did the following • (Local) Examined the intersection of I1 with fragments of functions in linfund . • (Global) Pieced together the fragments with nontrivial intersections with I1 to draw conclusions about the “holes” I1 creates or destroys. This is the local-global philosophy of this homological approach to complexity, inherited from algebraic topology. This is markedly different from conventional wisdom in computer science, which seeks to show that a function, such as f = 3sat, has some property that no function in a class, say C = P, has. In that method, there is no global step that argues that some global property of C changes after adding f into it. Using our homological technique, we show, in Section 3, a proof of Minsky and Papert’s classical result that the class polythrkd of polynomial thresholds of degree k in d variables does not contain the parity function parityd unless k = d (Theorem 3.40). Homologically, there are many
Pk reasons. By considering high dimensions, we deduce that Spolythrk has a hole in dimension i=0 di that d is filled in by parityd . By considering low dimensions, we obtain a maximal principle for polynomial threshold functions from which we obtain not only Minsky and Papert’s result but also extensions to arbitrary symmetric functions. This maximal principle Theorem 3.51 says Theorem 1.2 (Maximal Principle for Polynomial Threshold). Let C := polythrkd , and let f : {−1, 1}d → {−1, 1} be a function. We want to know whether f ∈ C. Suppose there exists a function g ∈ C (a “local maximum” for approximating g) such that • for each h ∈ C that differs from g on exactly one input u, we have g(u) = f (u) = ¬h(u). If g 6= f , then f 6∈ C. (In other words, if f ∈ C, then the “local maximum” g must be a “global maximum”). Notice that the maximal principle very much follows the local-global philosophy. The “local maximum” condition is saying that when one looks at the intersection with f of g and its “neighbors” (local), these intersections together form a hole that f creates when added to C (global). The homological intuition, in more precise terms, is that a local maximum g 6= f ∈ C implies that the filtered class C  (f ∩ g) consists of a single point with label g, so that when f is added to C, a zero-dimensional hole is created. We also obtain an interesting characterization of when a function can be weakly represented by a degree bounded polynomial threshold function. A real function ϕ : U → R on a finite set U is said to weakly represent a function f : U → {−1, 1} if ϕ(u) > 0 ⇐⇒ f (u) = 1 and ϕ(u) < 0 ⇐⇒ f (u) = −1, but we don’t care what happens when ϕ(u) = 0. Our homological theory of function essentially says that f ∈ polythrkd (“f is strongly representable by a polynomial of degree k”) iff Spolythrk ∪{f }g has the same number of holes as Spolythrk g in each dimension and d d for each g. But, intriguingly, f is weakly representable by a polynomial of degree k iff Spolythrk ∪{f } d has the same number of holes as Spolythrk in each dimension (Corollary 3.46) — in other words, d we only care about filtering by g = † but no other partial functions. 1.6 Dimension theory Let C ⊆ [n → 2]. The VC Dimension dimVC C of C is the size of the largest set U ⊆ [n] such that C  U = {0, 1}U . Consider the following setting of a learning problem: You have an oracle, called the sample oracle, such that every time you call upon it, it will emit a sample (u, h(u)) from an unknown 8 1 INTRODUCTION distribution P over u ∈ [n], for a fixed h ∈ C. This sample is independent of all previous and all future samples. Your task is to learn the identity of h with high probability, and with small error (weighted by P ). A central result of statistical learning theory says roughly that Theorem 1.3 ([10]). In this learning setting, one only needs O(dimVC C) samples to learn h ∈ C with high probability and small error. It is perhaps surprising, then, that the following falls out of our homological approach. Theorem 1.4 (Colloquial version of Theorem 3.11). Let C ⊆ [n → 2]. Then dimVC C is upper bounded by one plus the highest dimension, over any partial function g, of any hole in SCg . This quantity is known as the homological dimension dimh C of C. In fact, equality holds for common classes in the theory of computation like linfund and polythrkd , and also when certain algebraic conditions hold. More precisely — for readers with algebraic background — Theorem 1.5 (Colloquial version of Corollary 3.34). dimVC C = dimh C if the Stanley-Reisner ring of SC is Cohen-Macaulay. These results suggest that our homological theory captures something essential about computation, that it’s not a coincidence that we can use “holes” to prove complexity separation. 1.7 Homological Farkas Farkas’ Lemma is a simple result from linear algebra, but it is an integral tool for proving weak and strong dualities in linear programming, matroid theory, and game theory, among many other things. Lemma 1.6 (Farkas’ Lemma). Let L ⊆ Rn be a linear subspace not contained in any coordinate hyperplanes, and let P = {x ∈ Rn : x > 0} be the positive cone. Then either • L intersects P , or • L is contained in the kernel of a nonzero linear functional whose coefficients are all nonnegative. but not both. Farkas’ Lemma is a characterization of when a linear subspace intersects the positive cone in terms of linear conditions. An alternate view important in computer science is that Farkas’ Lemma provides a linear certificate for when this intersection does not occur. Analogously, our Homological Farkas’ Lemma will characterize such an intersection in terms of homological conditions, and simultaneously provide a homological certificate for when this intersection does not occur. Before stating the Homological Farkas’ Lemma, we first introduce some terminology. For g : [n] → {1, −1}, let Pg ⊆ Rn denote the open cone whose points have signs given by g. Consider the intersection 4g of Pg with the unit sphere S n−1 and its interior 4̊g . 4̊g is homeomorphic to an open simplex. For g 6= ¬1, define Λ(g) to be the union of the facets F of 4g such that 4̊g and 4̊1 sit on opposite sides of the affine hull of F . Intuitively, Λ(g) is the part of ∂4g that can be seen from an observer in 4̊1 (illustrated by Figure 10a). The following homological version of Farkas’ Lemma naturally follows from our homological technique of analyzing the complexity of threshold functions. 9 1.7 Λ(g) Λ(g) 4¬1 4g Homological Farkas 4¬1 41 (a) An example of a Λ(g). Intuitively, Λ(g) is the part of ∂4g that can be seen from an observer in 41 . 4g 41 (b) An illustration of Homological Farkas’ Lemma. The horizontal dash-dotted plane intersects the interior of 41 , but its intersection with any of the Λ(f ), f 6= 1, ¬1 has no holes. The vertical dash-dotted plane misses the interior of 41 , and we see that its intersection with Λ(g) as shown has two disconnected components. Figure 10 Theorem 1.7 (Homological Farkas’ Lemma Theorem 3.43). Let L ⊆ Rn be a linear subspace. Then either • L intersects the positive cone P = P1 , or • L ∩ Λ(g) for some g 6= 1, ¬1 is nonempty and has holes. but not both. Figure 10b illustrates an example application of this result. One direction of the Homological Farkas’ Lemma has the following intuition. As mentioned before, Λ(g) is essentially the part of ∂4g visible to an observer Tom in 4̊1 . Since the simplex is convex, the image Tom sees is also convex. Suppose Tom sits right on L (or imagine L to be a subspace of Tom’s visual field). If L indeed intersects 4̊1 , then for L ∩ Λ(g) he sees some affine space intersecting a convex body, and hence a convex body in itself. Since Tom sees everything (i.e. his vision is homeomorphic with the actual points), L ∩ Λ(g) has no holes, just as Tom observes. In other words, if Tom is inside 4̊1 , then he cannot tell Λ(g) is nonconvex by his vision alone, for any g. Conversely, the Homological Farkas’ Lemma says that if Tom is outside of 4̊1 and if he looks away from 4̊1 , he will always see a nonconvex shape in some Λ(g). As a corollary to Theorem 1.7, we can also characterize when a linear subspace intersects a region in a linear hyperplane arrangement (Corollary 3.55), and when an affine subspace intersects a region in an affine hyperplane arrangement (Corollary 3.56), both in terms of homological conditions. A particular simple consequence, when the affine subspace either intersects the interior or does not intersect the closure at all, is illustrated in Figure 11. The rest of this paper is organized as follows. Section 2 builds the theory underlying our complexity separation technique. Section 2.1 explains some of the conventions we adopt in this work and more importantly reviews basic facts from combinatorial commutative algebra and collects important lemmas for later use. Section 2.2 defines the central objects of study in our theory, the Stanley-Reisner ideal and the canonical ideal of each function class. The section ends by giving a characterization of when an ideal is the Stanley-Reisner ideal of a class. Section 2.3 discusses how to extract homological data of a class from its ideals via cellular resolutions. We construct cellular 10 2 THEORY 3 2 f g Λ(g) Λ(f ) 1 Figure 11: Example application of Corollary 3.57. Let the hyperplanes (thin lines) be oriented such that the square S at the center is on the positive side of each hyperplane. The bold segments indicate the Λ of each region. Line 1 intersects S, and we can check that its intersection with any bold component has no holes. Line 2 does not intersect the closure S, and we see that its intersection with Λ(f ) is two points, so has a “zeroth dimension” hole. Line 3 does not intersect S either, and its intersection with Λ(g) consists of a point in the finite plane and another point on the circle at infinity. resolutions for the canonical ideals of many classes prevalent in learning theory, such as conjunctions, linear thresholds, and linear functionals over finite fields. Section 2.4 briefly generalizes definitions and results to partial function classes, which are then used in Section 2.5. This section explains, when combining old classes to form new classes, how to also combine the cellular resolutions of the old classes into cellular resolutions of the new classes. Section 3 reaps the seeds we have sowed so far. Section 3.1 looks at notions of dimension, the Stanley-Reisner dimension and the homological dimension, that naturally appear in our theory and relates them to VC dimension, a very important quantity in learning theory. We observe that in most examples discussed in this work, the homological dimension of a class is almost the same as its VC dimension, and prove that the former is always at least the latter. Section 3.2 characterizes when a class has Stanley-Reisner ideal and canonical ideal that induce Cohen-Macaulay rings, a very well studied type of rings in commutative algebra. We define Cohen-Macaulay classes and show that their homological dimensions are always equal to their VC dimensions. Section 3.3 discusses separation of computational classes in detail, and gives simple examples of this strategy in action. Here a consequence of our framework is the Homological Farkas Lemma. Section 3.4 formulates and proves the maximal principle for threshold functions, and derives an extension of Minsky and Papert’s result for general symmetric functions. Section 3.5 further extends Homological Farkas Lemma to general linear or affine hyperplane arrangements. Section 3.6 examines a probabilistic interpretation of the Hilbert function of the canonical ideal, and shows its relation to hardness of approximation. Finally, Section 5 considers major questions of our theory yet to be answered and future directions of research. 2 2.1 Theory Background and Notation In this work, we fix k to be an arbitrary field. We write N = {0, 1, . . . , } for the natural numbers. Let n, m ∈ N and A, B be sets. The notation f :⊆ A → B specifies a partial function f whose domain dom f is a subset of A, and whose codomain is B. The words “partial function” will often be abbreviated “PF.” We will use Sans Serif font for partial (possibly total) functions, ex. f, g, h, 11 2.1 Background and Notation but will use normal font if we know a priori a function is total, ex. f, g, h. We denote the empty function, the function with empty domain, by †. We write [n] for the set {0, 1, . . . , n − 1}. We write [A → B] for the set of total functions from A to B and [⊆ A → B] for the set of partial functions from A to B. By a slight abuse of notation, [n → m] (resp. [⊆ n → m] is taken to be a shorthand for [[n] → [m]] (resp. [⊆ [n] → [m]]). The set [2d ] is identified with [2]d via binary expansion (ex: 5 ∈ [24 ] is identified with (0, 1, 0, 1) ∈ [2]4 ). A subset of [A → B] (resp. [⊆ A → B]) is referred to as a class (resp. partial class), and we use C, D (resp. C, D), and so on to denote it. Often, a bit vector v ∈ [2d ] will be identified with the subset of [d] of which it is the indicator function. For A ⊆ B, relative set complement is written B \ A; when B is clearly the universal set from context, we also write Ac for the complement of A inside B. If {a, b} is any two-element set, we write ¬a = b and ¬b = a. P Denote the n-dimensional simplex {v ∈ Rn : i vi = 1} by 4n . Let X, Y be topological spaces (resp. simplicial complexes, polyhedral complexes). The join of X and Y as a topological space (resp. simplicial complex, polyhedral complex) is denoted by X ? Y . We abbreviate the quotient X/∂X to X/∂. We will use some terminologies and ideas from matroid theory in Section 2.3.5 and Section 3.3. Readers needing more background can consult the excellently written chapter 6 of [22]. 2.1.1 Combinatorial Commutative Algebra Here we review the basic concepts of combinatorial commutative algebra. We follow [11] closely. Readers familiar with this background are recommended to skip this section and come back as necessary; the only difference in presentation from [11] is that we say a labeled complex is a cellular resolution when in more conventional language it supports a cellular resolution. Let k be a field and S = k[x] be the polynomial ring over k in n indeterminates x = x0 , . . . , xn−1 . n−1 Definition 2.1. A monomial in k[x] is a product xa = xa00 · · · xn−1 for a vector a = (a0 , . . . , an−1 ) ∈ n a N of nonnegative integers. Its support supp x is the set of i where ai 6= 0. We say xa is squarefree if every coordinate of a is 0 or 1. We often use symbols σ, τ , etc for squarefree exponents, and identify them with the corresponding subset of [n]. An ideal I ⊆ k[x] is called a monomial ideal if it is generated by monomials, and is called a squarefree monomial ideal if it is generated by squarefree monomials. a Let ∆ be a simplicial complex. Definition 2.2. The Stanley-Reisner ideal of ∆ is defined as the squarefree monomial ideal I∆ = hxτ : τ 6∈ ∆i generated by the monomials corresponding the nonfaces τ of ∆. The Stanley-Reisner ring of ∆ is the quotient ring S/I∆ . Definition 2.3. The squarefree Alexander dual of squarefree monomial ideal I = hxσ1 , . . . , xσr i is defined as I ? = mσ1 ∩ · · · ∩ mσr . If ∆ is a simplicial complex and I = I∆ its Stanley-Reisner ideal, then the simplicial complex ∆? ?. Alexander dual to ∆ is defined by I∆? = I∆ Proposition 2.4 (Prop 1.37 of [11]). The Alexander dual of a Stanley-Reisner ideal I∆ can in fact be described as the ideal hxτ : τ c ∈ ∆i, with minimal generators xτ where τ c is a facet of ∆. 12 2 THEORY Definition 2.5. The link of σ inside the simplicial complex ∆ is linkσ ∆ = {τ ∈ ∆ : τ ∪ σ ∈ ∆ & τ ∩ σ = ∅}, the set of faces that are disjoint from σ but whose unions with σ lie in ∆. Definition 2.6. The restriction of ∆ to σ is defined as ∆  σ = {τ ∈ ∆ : τ ⊆ σ}. Definition 2.7. A sequence φ1 φ l Fl ← 0 F• : 0 ← F0 ←− F1 ← · · · ← Fl−1 ←− of maps of free S-modules is called a complex if φi ◦ φi+1 = 0 for all i. The complex is exact in homological degree i if ker φi = im φi+1 . When the free modules Fi are Nn -graded, we require that each homomorphism φi to be degree-preserving. Let M be a finitely generated Nn -graded module M . We say F• is a free resolution of M over S if F• is exact everywhere except in homological degree 0, where M = F0 / im φ1 . The image in Fi of the homomorphism φi+1 is the ith syzygy module of M . The length of F• is the greatest homological degree of a nonzero module in the resolution, which is l here if Fl 6= 0. The following lemma says that if every minimal generator of an ideal J is divisible by x0 , then its resolutions are in bijection with the resolutions of J/x0 , the ideal obtained by forgetting variable x0 . Lemma 2.8. Let I ⊆ S = k[x0 , . . . , xn−1 ] be a monomial ideal generated by monomials not divisible by x0 . A complex F• : 0 ← F0 ← F1 ← · · · ← Fl−1 ← Fl ← 0 resolves x0 I iff for S/x0 = k[x1 , . . . , xn−1 ], F• ⊗S S/x0 : 0 ← F0 /x0 ← F1 /x0 ← · · · ← Fl−1 /x0 ← Fl /x0 ← 0 resolves I ⊗S S/x0 . Definition 2.9. Let M be a finitely generated Nn -graded module M and F• : 0 ← F0 ← F1 ← · · · ← Fl−1 ← Fl ← 0 L be a minimal graded free resolution of M . If Fi = a∈Nn S(−a)βi,a , then the ith Betti number of M in degree a is the invariant βi,a = βi,a (M ). Proposition 2.10 (Lemma 1.32 of [11]). βi,a (M ) = dimk TorSi (k, M )a . Proposition 2.11 (Hochster’s formula, dual version). All nonzero Betti numbers of I∆ and S/I∆ lie in squarefree degrees σ, where e i−1 (linkσc ∆∗ ; k). βi,σ (I∆ ) = βi+1,σ (S/I∆ ) = dimk H Proposition 2.12 (Hochster’s formula). All nonzero Betti numbers of I∆ and S/I∆ lie in squarefree degrees σ, where e |σ|−i−1 (∆  σ; k). βi−1,σ (I∆ ) = βi,σ (S/I∆ ) = dimk H 13 2.1 Background and Notation Note that since we are working over a field k, the reduced cohomology can be replaced by reduced homology, since these two have the same dimension. Instead of algebraically constructing a resolution of an ideal I, one can sometimes find a labeled simplicial complex whose simplicial chain is a free resolution of I. Here we consider a more general class of complexes, polyhedral cell complexes, which can have arbitrary polytopes as faces instead of just simplices. Definition 2.13. A polyhedral cell complex X is a finite collection of convex polytopes, called faces or cells of X, satisfying two properties: • If P is a polytope in X and F is a face of P, then F is in X. • If P and Q are in X, then P ∩ Q is a face of both P and Q. In particular, if X contains any point, then it contains the empty cell ∅, which is the unique cell of dimension −1. Each closed polytope P in this collection is called a closed cell of X; the interior of such a polytope, written P̊, is called an open cell of X. By definition, the interior of any point polytope is the empty cell. The complex with only the empty cell is called the irrelevant complex. The complex with no cell at all is called the void complex. The void complex is defined to have dimension −∞; any other complex X is defined to have dimension dim(X) equal to the maximum dimension of all of its faces. Examples include any polytope or the boundary of any polytope. Each polyhedral cell complex X has a natural reduced chain complex, which specializes to the usual reduced chain complex for simplicial complexes X. Definition 2.14. Suppose X is a labeled cell complex, by which we mean that its r vertices have labels that are vectors a1 , . . . , ar in Nr . The label aF on an arbitrary face F of X is defined as the coordinatewise maximum maxi∈F ai over the vertices in F . The monomial label of the face F is xaF . In particular, the empty face ∅ is labeled with the exponent label 0 (equivalently, the monomial label 1 ∈ S). When necessary, we will refer explicitly to the labeling function λ, defined by λ(F ) = aF , and express each labeled cell complex as a pair (X, λ). Definition 2.15. Let X be a labeled cell complex. The cellular monomial matrix supported on X uses the reduced chain complex of X for scalar entries, with the empty cell in homological degree 0. Row and column labels are those on the corresponding faces of X. The cellular free chain complex FX supported on X is the chain complex of Nn -graded free S-modules (with basis) represented by the cellular monomial matrix supported on X. The free complex FX is a cellular resolution if it has homology only in degree 0. We sometimes abuse notation and say X itself is a cellular resolution if FX is. Proposition 2.16. Let (X, λ) be a labeled complex. If FX is a cellular resolution, then it resolves S/I where I = hxaV : V ∈ X is a vertex}. FX is in addition minimal iff for each cell F of X, λ(F ) 6= λ(G) for each face G of F . Proposition 2.17. If X is a minimal cellular resolution of S/I, then βi,a (I) is the number of i-dimensional cells in X with label a. 14 2 THEORY Given two vectors a, b ∈ Nn , we write a  b and say a precedes b, b − a ∈ Nn . Similarly, we write a ≺ b if a  b but a 6= b. Define Xa = {F ∈ X : aF  a} and X≺a = {F ∈ X : aF ≺ a}. Let us say a cell complex is acyclic if it is either irrelevant or has zero reduced homology. In the irrelevant case, its only nontrivial reduced homology lies in degree −1. Lemma 2.18 (Prop 4.5 of [11]). X is a cellular resolution iff Xb is acyclic over k for all b ∈ Nn . For X with squarefree monomial labels, this is true iff Xb is acyclic over k for all b ∈ [2]n . When FX is acyclic, it is a free resolution of the monomial quotient S/I where I = hxav : v ∈ X is a vertexi generated by the monomial labels on vertices. It turns out that even if we only have a nonminimal cellular resolution, it can still be used to compute the Betti numbers. Proposition 2.19 (Thm 4.7 of [11]). If X is a cellular resolution of the monomial quotient S/I, then the Betti numbers of I can be calculated as e i−1 (X≺b : k) βi,b (I) = dimk H as long as i ≥ 1. Lemma 2.18 and Proposition 2.19 will be used repeatedly in the sequel. We will also have use for the dual concept of cellular resolutions, cocellular resolutions, based on the cochain complex of a polyhedral cell complex. Definition 2.20. Let X 0 ⊆ X be two polyhedral cell complexes. The cochain complex C • (X, X 0 ; k) of the cellular pair (X, X 0 ) is defined by the exact sequence 0 → C • (X, X 0 ; k) → C • (X; k) → C • (X 0 ; k) → 0. The ith relative cohomology of the pair is H i (X, X 0 ; k) = H i C • (X, X 0 ; k). Definition 2.21. Let Y be a cell complex or a cellular pair. Then Y is called weakly colabeled if the labels on faces G ⊆ F satisfy aG  aF . In particular, if Y has an empty cell, then it W must be labeled as well. Y is called colabeled if, in addition, every face label aG equals the join aF of all the labels on facets F ⊇ G. Again, when necessary, we will specifically mention the labeling function λ(F ) = aF and write the cell complex (or pair) as (Y, λ). We have the following well known lemma from the theory of CW complexes. Lemma 2.22. Let X be a cell complex. A collection R of open cells in X is a subcomplex of X iff S R is closed in X. If Y = (X, X 0 ) is a cellular pair, then we treat Y as the collection of (open) cells in X \ X 0 , for the reason that C i (X, X 0 , k) has as a basis the set of open cells of dimension i in X \ X 0 . As Y being a complex is equivalent to Y being the pair (Y, {}) (where {} is the void subcomplex), in the sense that the reduced cochain complex of Y is isomorphic to the cochain complex of the pair (Y, {}), we will only speak of cellular pairs from here on when talking about colabeling. Definition 2.23. Let Y = (X, A) be a cellular pair and U a subcollection of open cells of Y . We say U is realized by a subpair (X 0 , A0 ) ⊆ (X, A) (i.e. X 0 ⊆ X, A0 ⊆ A) if U is the collection of open cells in X 0 \ A0 . Definition 2.24. Define Yb (resp. Y≺b and Yb ) as the collection of open cells with label  b (resp. ≺ b and b). 15 2.1 Background and Notation We often consider Yb , Y≺b , and Yb as subspaces of Y , the unions of their open cells. Proposition 2.25. Let Y be a cellular pair and U = Yb (resp. Y≺b and Yb ). Then U is realized by the pair (U, ∂U), where the first of the pair is the closure of U as a subspace in Y , and the second is the partial boundary ∂U := U \ U. Proof. See Appendix A. Note that if X 0 is the irrelevant complex, then H i (X, X 0 ; k) = H i (X; k), the unreduced cohoe i (X; k), the reduced cohomology of mology of X. If X 0 is the void complex, then H i (X, X 0 ; k) = H e i (X/X 0 ; k). X. Otherwise X 0 contains a nonempty cell, and it is well known that H i (X, X 0 ; k) ∼ =H 0 i 0 i ∼ e (•; k) = 0. In particular, when X = X, H (X, X ; k) = H Definition 2.26. Let Y be a cellular pair (X, X 0 ), (weakly) colabeled. The (weakly) cocellular monomial matrix supported on Y has the cochain complex C • (Y ; k) for scalar entries, with top dimensional cells in homological degree 0. Its row and column labels are the face labels on Y . The (weakly) cocellular free complex FY supported on Y is the complex of Nn -graded free S-modules (with basis) represented by the cocellular monomial matrix supported on Y . If FY is acyclic, so that its homology lies only in degree 0, then FY is a (weakly) cocellular resolution. We sometimes abuse notation and say Y is a (weakly) cocellular resolution if FY is. Proposition 2.27. Let (Y, λ) be a (weakly) colabeled complex or pair. If FY is a (weakly) cocellular resolution, then FY resolves I = hxaF : F is a top dimensional cell of Y i. It is in addition minimal iff for each cell F of Y , λ(F ) 6= λ(G) for each cell G strictly containing F . We say a cellular pair (X, X 0 ) is of dimension d if d is the maximal dimension of all (open) cells in X \ X 0 . If Y is a cell complex or cellular pair of dimension d, then a cell F of dimension k with label aF corresponds to a copy of S at homological dimension d − k with degree xaF . Therefore, Proposition 2.28. If Y is a d-dimension minimal (weakly) cocellular resolution of ideal I, then βi,a (I) is the number of (d − i)-dimensional cells in Y with label a. We have an acyclicity lemma for cocellular resolutions similar to Lemma 2.18 Lemma 2.29. Let Y = (X, A) be a weakly colabeled pair of dimension d. For any U ⊆ X, write U for the closure of U inside X. Y is a cocellular resolution iff for any exponent sequence a, K := Ya satisfies one of the following: 1) The partial boundary ∂K := K \ K contains a nonempty cell, and H i (K, ∂K) is 0 for all i 6= d and is either 0 or k when i = d, or e i (K) 2) The partial boundary ∂K is void (in particular does not contain the empty cell), and H is 0 for all i 6= d and is either 0 or k when i = d, or 3) K is void. Proof. See Appendix A. Lemma 2.30. Suppose Y = (X, A) is a weakly colabeled pair of dimension d. If Y supports a cocellular resolution of the monomial ideal I, then the Betti numbers of I can be calculated for all i as βi,b (I) = dimk H d−i (Y b , ∂Yb ; k). Proof. See Appendix A. Like with boundaries, we abbreviate the quotient K/∂K to K/∂, so in particular, the equation above can be written as e d−i (Yb /∂; k). βi,b (I) = dimk H 16 2 THEORY 00 01 (a) linfun1 suboplex. Dashed lines indicate facets of the complete suboplex not in Slinfun1 . Label 00 is the identically zero function; label 01 is the identity function. (b) Slinfun2 is a cone of what is shown, which is a subcomplex of the boundary complex of an octahedron. The cone’s vertex has label ((0, 0), 0), so that every top dimensional simplex meets it, because every linear functional sends (0, 0) ∈ (F2 )2 to 0. Figure 12: linfun1 and linfun2 suboplexes. 2.2 The Canonical Ideal of a Function Class Definition 2.31. An n-dimensional orthoplex (or n-orthoplex for short) is defined as any polytope combinatorially equivalent to {x ∈ Rn : kxk1 ≤ 1}, the unit disk under the 1-norm in Rn . Its boundary is a simplicial complex and has 2n facets. A fleshy (n − 1)-dimensional suboplex, or suboflex is the simplicial complex formed by any subset of these 2n facets. The complete (n − 1)-dimensional suboplex is defined as the suboplex containing all 2n facets. In general, a suboplex is any subcomplex of the boundary of an orthoplex. For example, a 2-dimensional orthoplex is equivalent to a square; a 3-dimensional orthoplex is equivalent to an octahedron. Let C ⊆ [n → 2] be a class of finite functions. There is a natural fleshy (n − 1)-dimensional suboplex SC associated to C. To each f ∈ C we associate an (n−1)-dimensional simplex Ff ∼ = 4n−1 , which will be a facet of SC . Each of the n vertices of Ff is labeled by a pair (i, f (i)) for some i ∈ [n], and each face G of Ff is labeled by a partial function f ⊆ f , whose graph is specified by the labels of the vertices of G. For each pair f, g ∈ C, Ff is glued together with Fg along the subsimplex G (in both facets) with partial function label f ∩ g. This produces SC , which we call the canonical suboplex of C. Example 2.32. Let [n → 2] be the set of boolean functions with n inputs. Then S[n→2] is the complete (n − 1)-dimensional suboplex. Each cell of S[n→2] is association with a unique partial function f :⊆ [n] → [2], so we write Ff for such a cell. Example 2.33. Let f ∈ [n → 2] be a single boolean function with domain [n]. Then S{f } is a single (n − 1)-dimensional simplex. Example 2.34. Let linfun2d ⊆ [2d → 2] be the class (F2 )d∗ of linear functionals mod 2. Figure 12 shows Slinfun2 for d = 1 and d = 2. d The above gluing construction actually make sense for any C ⊆ [n → m] (with general codomain [m]), even though the resulting simplicial complex will no longer be a subcomplex of S[n→2] . However, we will still call this complex the canonical suboplex of C and denote it SC as well. We name any such complex an m-suboplex. The (n − 1)-dimensional m-suboplex S[n→m] is called the complete (n − 1)-dimensional m-suboplex. 17 2.2 The Canonical Ideal of a Function Class The canonical suboplex of C ⊆ [n → m] can be viewed as the object generated by looking at the metric space Cp on C induced by a probability distribution p on [n], and varying p over all distributions in 4n−1 . This construction seems to be related to certain topics in computer science like derandomization and involves some category theoretic techniques. It is however not essential to the homological perspective expounded upon in this work, and thus its details are relegated to the appendix (See Appendix B). Definition 2.35. Let C ⊆ [n → m]. Write S for the polynomial ring k[x] with variables xi,j for i ∈ [n], j ∈ [m]. We call S the canonical base ring of C. The Stanley-Reisner ideal IC of C is defined as the Stanley-Reisner ideal of SC with respect to S, such that xi,j is associated to the “vertex” (i, j) of SC (which might not actually be a vertex of SC if no function f in C computes f (i) = j). The canonical ideal IC? of C is defined as the Alexander dual of its Stanley-Reisner ideal. By Proposition 2.4, the minimal generators of IC? are monomials xσ where σ c is the graph of a function in C. Let us define Γf to be the complement of graph f in [n] × [m] for any partial function f :⊆ [n] → [m]. Therefore, IC? is minimally generated by the monomials {xΓf : f ∈ C}. When the codomain [m] = [2], Γf = graph(¬f ), the graph of the negation of f , so we can also write IC? = hxgraph ¬f : f ∈ Ci. Example 2.36. Let [n → 2] be the set of boolean functions with domain [n]. Then I[n→2] is the ? ideal hxi,0 xi,1 : i ∈ [n]i, and I[n→2] is the ideal hxΓf : f ∈ [n → 2]i = hxgraph g : g ∈ [n → 2]i. Example 2.37. Let f ∈ [n → 2]. The singleton class {f } has Stanley-Reisner ideal hxi,¬f (i) : i ∈ [n]i and canonical ideal hxΓf i. The Stanley-Reisner ideal IC of a class C has a very concrete combinatorial interpretation. Proposition 2.38. Let C ⊆ [n → m]. IC is generated by all monomials of the following forms: 1. xu,i xu,j for some u ∈ [n], i 6= j ∈ [m], or 2. xgraph f for some partial function f :⊆ [n] → [m] such that f has no extension in C, but every proper restriction of f does. It can be helpful to think of case 1 as encoding the fact that C is a class of functions, and so for every function f , f sends u to at most one of i and j. For this reason, let us refer to monomials of the form xu,i xu,j , i 6= j as functional monomials with respect to S and write FMS , or FM when S is clear from context, for the set of all functional monomials. Let us also refer to a PF f of the form appearing in case 2 as an extenture of C, and denote by ex C the set of extentures of C. In this terminology, Proposition 2.38 says that IC is minimally generated by all the functional monomials and xgraph f for all extentures f ∈ ex C. Proof. The minimal generators of IC are monomials xa ∈ IC such that xa /xu,i 6∈ IC for any (u, i) ∈ a. By the definition of IC , a is a nonface, but each subset of a is a face of the canonical suboplex SC of C. Certainly pairs of the form {(u, i), (u, j)} for u ∈ [n], i 6= j ∈ [m] are not faces of SC , but each strict subset of it is a face unless (u, i) 6∈ SC or (u, j) 6∈ SC . In either case x(u,i) or x(u,j) or fall into case 2. If a minimal generator ω is not a pair of such form, then its exponent b cannot contain such {(u, i), (u, j)} either, or else r is divisible by xu,i xu,j . Therefore b is the graph of a partial function f :⊆ [n → m]. In particular, there is no f ∈ C extending f, or else graph f is a face of SC . But every proper restriction of f must have an extension in C. Thus ω is of the form stated in the proposition. One can also quickly see that xgraph f for any such f is a minimal generator of IC . 18 2 THEORY Taking the minimal elements of the above set, we get the following Proposition 2.39. The minimal generators of C ⊆ [n → m] are {xΓf : f ∈ ex C} ∪ {xu,i xu,j ∈ FM : (u 7→ i) 6∈ ex C, (u 7→ j) 6∈ ex C}. Are all ideals with minimal generators of the above form a Stanley-Reisner ideal of a function class? It turns out the answer is no. If we make suitable definitions, the above proof remains valid if we replace C with a class of partial functions (see Proposition 2.85). But there is the following characterization of the Stanley-Reisner ideal of a (total) function class. Proposition 2.40. Let I ⊆ S be an ideal minimally generated by {xgraph f : f ∈ F} ∪ {xu,i xu,j ∈ FM : (u 7→ i) 6∈ F, (u 7→ j) 6∈ F } for a set of partial functions F. Then I is the Stanley-Reisner ideal of a class of total functions C precisely when For any subset F ⊆ F, if F (u) defined as {f(u) : f ∈ F, u ∈ dom f} is equal to [m] for some u ∈ [n], then either |F (v)| > 1 for some v 6= u in [n], or [ ∨u F := f  (dom f \ {u}) (?) f∈F ,u∈dom f is a partial function extending some h ∈ F. Lemma 2.41. For I minimally generated as above, I = IC for some C iff for any partial f :⊆ [n] → [m], xgraph f 6∈ I implies xgraph f 6∈ I for some total f extending f. Proof of Lemma 2.41. Let ∆I be the Stanley-Reisner complex of I. Then each face of ∆I is the graph of a partial function, as I has all functional monomials as generators. A set of vertices σ is a face iff xσ 6∈ I. I = IC for some I iff ∆I is a generalized suboflex, iff the maximal cells of ∆I are all (n − 1)-dimensional simplices, iff every cell is contained in such a maximal cell, iff xgraph f 6∈ I implies xgraph f 6∈ I for some total f extending f. Proof of Proposition 2.40. (⇒). We show the contrapositive. Suppose for some F ⊆ F and u ∈ [n], F (u) = [m] but |F (v)| ≤ 1 for all v 6= u and g := ∨u F does not extend any f ∈ F. Then xgraph g 6∈ I, and every total f ⊇ g must contain one of f ∈ F, and so xgraph f ∈ I. Therefore I 6= IC for any C. (⇐). Suppose (?) is true. We show that for any nontotal function f :⊆ [n] → [m] such that graph f 6∈ I, there is a PF h that extends f by one point, such that xgraph h 6∈ I. By simple induction, x this would show that I = IC for some C. Choose u 6∈ dom f. Construct F := {g ∈ F : u ∈ dom g, f ⊇ g  (dom g \ {u})}. If F (u) 6= [m], then we can pick some i 6∈ F (u), and set h(u) = i and h(v) = f(v), ∀v 6= u. If h ⊇ k for some k ∈ F, then k ∈ F , but then k(u) 6= h(u) by assumption. Therefore h does not extend any PF in F, and xgraph h 6∈ I. If F (u) = [m], then by (?), either |F (v)| > 1 for some v 6= u or ∨u F extends some h ∈ F. The former case is impossible, as f ⊇ g  (dom g \ {u}) for all g ∈ F . The latter case is also impossible, as it implies that xgraph f ∈ I. 19 2.3 2.3 Resolutions Resolutions Sometimes we can find the minimal resolution of the Stanley-Reisner ideal of a class. For example, consider the complete class [n → 2]. Its Stanley-Reisner ideal is hxi,0 xi,1 : i ∈ [n]i as explained in Example 2.36. Theorem 2.42. Let X be an (n − 1)-simplex, whose vertex i is labeled by monomial xi,0 xi,1 . Then X is a minimal cellular resolution of S/I[n→2] . Proof. The vertex labels of X generate I[n→2] , and each face label is distinct from other face labels, so if X is a cellular resolution, then it resolves S/I[n→2] and is minimal. Therefore it suffices to show that FX is exact. By Lemma 2.18, we need to show that Xb is acyclic over k for all b ⊆ [n] × [2]. Xb can be described as the subcomplex generated by the vertices {i : (i, 0), (i, 1) ∈ b}, and hence is a simplex itself and therefore contractible. This completes the proof. Corollary 2.43. The Betti numbers of I[n→2] are nonzero only at degrees of the form σ= Y xi,0 xi,1 i∈U for subset U ⊆ [n]. In such cases, βi,σ (I[n→2] ) = I(i = |U | − 1). Similar reasoning also gives the minimal resolution of any singleton class. Theorem 2.44. Suppose f ∈ [n → 2]. Let X be an (n − 1)-simplex, whose vertex i is labeled by variable xi,¬f (i) . Then X is a minimal cellular resolution of S/I{f } . Corollary 2.45. The Betti numbers of I{f } are nonzero only at degrees of the form σ= Y xi,¬f (i) i∈U for subset U ⊆ [n]. In such cases, βi,σ (I{f } ) = I(i = |U | − 1). However, in general, minimally resolving the Stanley-Reisner ideal of a class seems difficult. Instead, we turn to the canonical ideal, which appears to more readily yield cellular resolutions, and as we will see, whose projective dimension corresponds to the VC dimension of the class under ? an algebraic condition. For example, a single point with label xΓf minimally resolves S/I{f } for any f ∈ [n → 2]. We say (X, λ) is a cellular resolution of a class C if (X, λ) is a cellular resolution of S/IC? . In the following, we construct the cellular resolutions of many classes that are studied in Computational Learning Theory. As a warmup, we continue our discussion of [n → 2] by constructing a cellular resolution of its canonical ideal. Theorem Q 2.46. Let P be the n-dimensional cube [0, 1]n , where vertex v ∈ [2]n is labeled with the monomial ni=1 xi,vi . Then P minimally resolves [n → 2]. 20 2 THEORY Proof. We first show that this labeled cell complex on the cube is a cellular resolution. Let σ ⊆ [n] × [2]. We need to show that Pσ is acyclic. If for some i, (i, 0) 6∈ σ & (i, 1) 6∈ σ, then Pσ is empty and thus acyclic. Otherwise, σ c defines a partial function f :⊆ [n] → [2]. Then Pσ is the “subcube” {v ∈ [0, 1]n : vi = ¬f (i), ∀i ∈ dom f }, and is therefore acyclic. This shows that P is a resolution. It is easy to see that all faces of P have unique labels in this form, and hence the resolution is minimal as well. ? P resolves S/I[n→2] by Example 2.36 The above proof readily yields the following description of [n → 2]’s Betti numbers. ? Corollary 2.47. The Betti numbers for I[n→2] are nonzero only at degrees of the form Γf for partial functions f :⊆ [n] → [2]. More precisely, ? ) = I(| dom f | = n − i) βi,Γf (I[n→2] We made a key observation in the proof of Theorem 2.46, that when neither (i, 0) nor (i, 1) is in σ for some i, then Pσ is empty and thus acyclic. A generalization to arbitrary finite codomains is true for all complexes X we are concerned with: Lemma 2.48. Let (X, λ) be a labeled complex in which each vertex i is labeledTwith Γf
i for partial function fi : [n] → [m]. Then the face label λ(F ) for a general face F is Γ i∈F fi . A fortiori Xσ is empty whenever σ is not of the form Γg for some partial function g :⊆ [n] → [m]. Proof. Treating the exponent labels, which are squarefree, as sets, we have !c ! [ [ \ \ c λ(F ) = Γfi = (graph fi ) = graph fi =Γ fi i∈F i∈F i∈F i∈F If σ is not of the form Γg, then for some a ∈ [n] and b 6= b0 ∈ [m], (a, b), (a, b0 ) 6∈ σ. But every exponent label is all but at most one of the pairs (a, ∗). So Xσ is empty. If we call a complex as described in the lemma partial-function-labeled, or PF-labeled for short, then any PF-labeled complex has a set of partial function labels, or PF labels for short, along with its monomial/exponent labels. If fF denotes the partial function label of face F and aF denotes the exponent label of face F , then they can be interconverted via fF = acF aF = ΓfF where on the right we identify a T partial function with its graph. Lemma 2.48 therefore says that F ⊆ G implies fF ⊇ fG , and fF = i∈F fi , for faces F and G. When we wish to be explicit about the PF labeling function, we use the symbol µ, such that µ(F ) = fF , and refer to labeled complexes as pairs (X, µ) or triples (X, λ, µ). We can furthermore reword Lemma 2.18 for the case of PF-labeled complexes. Write X⊇f (resp. X⊃f ) for the subcomplex with partial function labels weakly (resp. strictly) extending f. Lemma 2.49. A PF-labeled complex X is a cellular resolution iff X⊇f is acyclic over partial functions f. k for all A PF-colabeled complex or pair is defined similarly. The same interconversion equations hold. We can likewise reword Lemma 2.29. 21 2.3 Resolutions Lemma 2.50. Let (X, A) be a weakly PF-colabeled complex or pair of dimension d. (X, A) is a cocellular resolution if for any partial function f, (X, A)⊇f is either 1) representable as a cellular pair (Y, B) – that is, X \A as a collection of open cells is isomorphic to Y \ B as a collection of open cells, such that H i (Y, B) is 0 for all i 6= d, or 2) a complex Y (in particular it must contain a colabeled empty cell) whose reduced cohomology vanishes at all dimensions except d. Because any cellular resolution of a class C only has cells with degree Γf for some PF f, the Betti numbers βi,σ (IC? ) can be nonzero only when σ = Γg for some PF g. We define the Betti numbers of a class C as the Betti numbers of its canonical ideal IC? , and we denote βi,f (C) := βi,Γf (IC? ). Finally we note a trivial but useful proposition and its corollary. Proposition 2.51. Let C ⊆ [n → m], and let f :⊆ [n] → [m]. The subset of functions extending f, {h T ∈ C : f ⊆ h}, is the intersection of the collection of sets which extend the point restrictions of f , i∈dom f {h ∈ C : (i, f(i)) ⊆ h}. S If partial functions g1 , . . . , gk ∈ [n → m] satisfy t gt = f, then we also have {h ∈ C : f ⊆ h} = k \ {h ∈ C : gt ⊆ h}. t=1 Corollary 2.52. Let f :⊆ [n] S → [m]. Suppose X is a PF-labeled complex. If partial functions g1 , . . . , gk ∈ [n → m] satisfy t gt = f, then X⊇f = k \ X⊇gt . t=1 With these tools in hand, we are ready to construct cellular resolutions of more interesting function classes. 2.3.1 Delta Functions Let deltan ⊆ [n → 2] be the class of delta functions δi (j) = I(i = j). Form the abstract simplex X with vertices [n]. Label each vertex i with δi and induce PF labels on all higher dimensional faces in the natural way. One can easily check the following lemma. Lemma 2.53. For any face F ⊆ [n] with |F | > 1, its PF label fF is the function defined on [n] \ F , sending everything to 0. Conversely, for every partial f :⊆ [n] → [2] with im f ⊆ {0}, there is a unique face F with fF = f as long as n − | dom f | ≥ 2. Theorem 2.54. X is a (n − 1)-dimensional complex that minimally resolves deltan . Proof. We apply Lemma 2.49: We show for any f :⊆ [n] → [2], X⊇f is acyclic. If f sends two distinct elements to 1, then X⊇f is empty. If f sends exactly one element i to 1, then X⊇f is the single point i. If f is the empty function, then X⊇f is the whole simplex and thus acyclic. Otherwise, im f = {0}. If n − | dom f | = 1, then there is exactly one delta function extending f, so X⊇f is again a point. If n − | dom f | ≥ 2, then by Lemma 2.53, X⊇f is exactly one face F with fF = f, and therefore acyclic. X is furthermore minimal because all PF labels are distinct. 22 2 THEORY Tabulating the faces by their labels, we obtain Corollary 2.55. For i > 0, βi,f (deltan ) is nonzero only when im f ⊆ {0} and n − | dom f| ≥ 2, and i = n − | dom f| − 1. In that case, βi,f (deltan ) = 1. In particular, the top dimensional Betti number is βn−1,† (deltan ) = 1. 2.3.2 Weight-k Functions Write o := 0 ∈ [n → 2], the function that sends all inputs to 0. Let wt(f, k)n ⊆ [n → 2] be the class consisting of all functions g such that there are exactly k inputs u ∈ [n] such that g(u) 6= f (u). This is a generalization of delta, as wt(o, 1)n = delta. P WLOG, we consider the case f = o in this section. Consider the hyperplane Hk := {v ∈ Rn : i vi = k} and the polytope given by Pnk := [0, 1]n ∩ Hk . We inductively define its labeling function µkn and show that (Pnk , µkn ) is a minimal cellular resolution of wt(f, k)n . For n = 1, P10 and P11 are both a single point. Set µ01 (P10 ) = (0 7→ 0) and µ11 (P11 ) = (0 7→ 1). Then trivially, (P10 , µ01 ) is the minimal resolution of wt(o, 0) = {0 7→ 0} and (P11 , µ11 ) is the minimal resolution of wt(o, 1) = {0 7→ 1}. k , µk ) is a minimal cellular resolution of wt(o, k) Suppose that µkm is defined and that (Pm m for all 0 ≤ k ≤ m. Consider n = m + 1 and fix k. Write, for each u ∈ [n], b ∈ [2], Fu,b := [0, 1]u × {b} × [0, 1]n−u−1 for the corresponding facet of [0, 1]n . Then Pnk has boundary given by [ Fu,b ∩ Hk . u∈[n],b∈[2] k−1 ∼ P k and Fu,1 ∩ Hk ∼ (here ∼ But we have Fu,0 ∩ Hk = = Pn−1 = means affinely isomorphic). Thus, if n−1 G is a face of Fu,b ∩ Hk , we define the labeling functions µkn (G) : [n] → [2] i 7→ µk−b n−1 (G)(i) if i < u i 7→ b if i = u i 7→ µk−b n−1 (G)(i − 1) if i > u. If we represent functions as a string of {0, 1, .} (where . signifies “undefined”), then essentially µkn (G) is obtained by inserting b ∈ {0, 1} at the uth position in µk−b n−1 (G). It is easy to see that, when G is both a face of Fu,b ∩ Hk and a face of Fu0 ,b0 ∩ Hk , the above definitions of µkn (G) coincide. Finally, we set µkn (Pnk ) = †. This finishes the definition of µkn . In order to show that (Pnk , µkn ) is a minimal cellular resolution, we note that by induction k−b hypothesis, it suffices to show that (Pnk )⊇† = Pnk is acyclic, since (Pnk )⊇(u7→b)∪f ∼ )⊇f is = (Pn−1 k acyclic. But of course this is trivial given that Pn is a polytope. By an easy induction, the vertex labels of Pnk are exactly the functions of wt(o, k)n . Thus Theorem 2.56. (Pnk , µkn ) as defined above is a minimal resolution of wt(o, k)n . Corollary 2.57. For k 6= 0, n, C := wt(o, k)n has a Betti number βi,† (C) = I(i = n − 1). Furthermore, for each PF f, βi,f (wt(o, k)n ) is nonzero for at most one i, where it is 1. 23 2.3 2.3.3 Resolutions Monotone Conjunction Let L = {l1 , . . . , ld } be a set of literals. The class of monotone conjunctions monconjd over L is defined as the set of functions that can be represented as a conjunction of a subset of L. We represent each h ∈ monconjd as the set of literals L(h) in its conjunctive form, and for each subset (or indicator function thereof) T of literals, let Λ(T ) denote the corresponding function. For example, Λ{l1 , l3 } is the function that takes v ∈ [2]d to 1 iff v1 = v3 = 1. Theorem 2.58. Let X be the d-cube in which each vertex V ∈ [2]d has partial function label (that is in fact a total function) fV = Λ(V ), where on the RHS V is considered an indicator function for a subset of literals. Then X resolves monconjd minimally. We first show that the induced face labels of X are unique, and hence if X is a resolution, it is minimal. This will follow from the following three lemmas. Lemma 2.59. Let w be a partial function w :⊆ [d] → [2]. Let Σw be the set of monotone conjunctions {h T : li ∈ L(h) if w(i) = 1 and li 6∈ L(h) if w(i) = 0}. Then the intersection of functions (not literals) Σw is the partial function Λ(w) := f :⊆ [2d ] → [2],   0 f(v) = 1   undefined if vi = 0 for some i with w(i) = 1 if vi = 1 for all i with w(i) = 1 and for all i where w(i) is undefined otherwise. When w is a total function considered as a bit vector, Λ(w) coincides with the previous definition of Λ. If F is the face of the cube resolution X with the vertices {V : V ⊇ w} (here treating V ∈ [2]d ∼ = [d → 2] as a function), then the partial function label of F is f. T Proof. f is certainly contained in Σw . To see that the inclusion is an equality, we show that for any v not of the two cases above, there are two functions h, h0 that disagree on v. Such a v satisfies vi = 1 for all w(i) = 1 but vi = 0 for some w(i) being undefined. There is some h ∈ Σw with L(h) containing the literal li and there is another h0 ∈ Σw with li 6∈ L(h0 ). These two functions disagree on v. The second statement can be checked readily. The third statement follows from Lemma 2.48. Lemma 2.60. For any partial function f of the form in Lemma 2.59, there is a unique partial function w :⊆ d → 2 with f = Λ(w), and hence there is a unique cell of X with PF label f. Proof. The set A := w−1 1 ∪ (dom w)c is the set {i ∈ d : vi = 1, ∀v ∈ f −1 1}, by the second case in f’s definition. The set B := w−1 1 is the set of i ∈ d such that the bit vector v with vi = 0 and vj = 1 for all j 6= i is in f −1 0, by the first case in f’s definition. Then dom w = (A \ B)c , and w−1 0 = (dom w) \ (w−1 1). Lemma 2.61. The face labels of X are all unique. Proof. Follows from Lemma 2.59 and Lemma 2.60. Proof of Theorem 2.58. We show that X is a resolution (minimal by the above) by applying Lemma 2.49. Let f :⊆ [2d ] → [2] be a partial function and g0 , g1 be respectively defined by gt = f  f −1 t for t = 0, 1, so that f = g0 ∪ g1 . By Corollary 2.52, X⊇f = X⊇g0 ∩ X⊇g1 . We first show that X⊇g1 is a face of X, and thus is itself a cube. If h ∈ monconjd is a conjunction, then 24 2 THEORY T it can be seen that h extends g1 iff L(h) ⊆ L1 := v∈dom g1 {li : vi = 1} (check this!). Thus X⊇g1 is the subcomplex generated by the vertices V whose coordinates Vi satisfy Vi = 0, ∀i 6∈ L1 . This subcomplex is precisely a face of X. Now we claim that each cell of X⊇g0 ∩ X⊇g1 is a face of a larger cell which contains the vertex W with W1 = 1, ∀i ∈ L1 and Wi = 0, ∀i 6∈ L1 . This would imply that X⊇g0 ∩ X⊇g1 is contractible via the straight line homotopy to W . We note that if h, h0 ∈ monconjd and L(h) ⊆ L(h0 ), then h extends g0 only if h0 also extends g0 . (Indeed, h extends g0 iff ∀v ∈ dom g0 , vk = 0 while lk ∈ L(h) for some k. This still holds for h0 if h0 contains all literals appearing in h). This means that, if F is a face of X⊇g0 , then the face F 0 generated by {V 0 : ∃V ∈ F, V ⊆ V 0 } (where V and V 0 are identified with the subset of literals they correspond to) is also contained in X⊇g0 ; F 0 can alternatively be described geometrically as the intersection [0, 1]d ∩ (F + [0, 1]d ). If furthermore F is a face of X⊇g1 , then F 0 ∩ X⊇g1 contains W as a vertex, because W is inclusion-maximal among vertices in X⊇g1 (when identified with sets for which they are indicator functions for). This proves our claim, and demonstrates that X⊇f is ? contractible. Therefore, X is a (minimal) resolution, of Imonconj by construction. d Corollary 2.62. βi,f (monconjd ) is nonzero iff f = Λ(w) for some PF w :⊆ [d] → [2] and i = d − | dom w|, and in that case it is 1. In particular, the top dimensional nonzero Betti number is βd,17→1 (monconjd ) = 1. Proof. This follows from Lemma 2.59 and Lemma 2.60. We will refer to X as the cube resolution of monconjd . 2.3.4 Conjunction S Define L0 := di=1 {li , ¬li }. The class of conjunctions conjd is defined as the set of functions that can be represented as a conjunction of a subset of L0 . In particular, L0 contains the null function ⊥ : v 7→ 0, ∀v, which can be written as the conjunction l1 ∧ ¬l1 . We now describe the polyhedral cellular resolution of conjd , which we call the cone-over-cubes resolution, denoted COCd . Each nonnull function h has a unique representation as a conjunction of e be the inverse function taking a set literals in L0 . We define L(h) to be the set of such literals and Λ of consistent literals to the conjunction function. We assign a vertex Vh ∈ {−1, 0, 1}d × {0} ∈ Rd+1 to each nonnull h by   if li ∈ L(h) 1 (Vh )i = −1 if ¬li ∈ L(h)   0 otherwise for all 1 ≤ i ≤ d (and of course (Vh )d+1 = 0), so that the PF label fVh = h. We put in COCd all d z }| { faces of the (2, 2, . . . , 2) pile-of-cubes: these are the collection of 2d d-dimensional unit cubes with vertices among {−1, 0, 1}d × {0}. This describes all faces over nonnull functions. Finally, we assign the coordinate V⊥ = (0, . . . , 0, 1) ∈ Rd+1 , and put in COCd the (d + 1)dimensional polytope C which has vertices Vh for all h ∈ conjd , and which is a cone over the pile of cubes, with vertex V⊥ . (Note that this is an improper polyhedron since the 2d facets of C residing on the base, the pile of cubes, all sit on the same hyperplane.) Figure 13 shows the cone-over-cubes resolution for d = 2. Theorem 2.63. COCd is a (d + 1)-dimensional complex that minimally resolves conjd . 25 2.3 Resolutions ⊥ ¬l1 ∧ l2 > ¬l1 ¬l1 ∧ ¬l2 l1 ∧ l2 l2 l1 ¬l2 l1 ∧ ¬l2 Figure 13: Cone-over-cube resolution of conj2 . Labels are PF labels. Proof. Let X = COCd . We first shot that X is a resolution of conjd . We wish to prove that for any f :⊆ [2d ] → [2], the subcomplex of X⊇f is acyclic. First suppose that im f = {0}. Then X⊇f is a subcomplex that is a cone with V⊥ as the vertex, and hence contractible. Otherwise f sends some point u ∈ [2d ] ∼ = [2]d to 1. All h ∈ conjd extending f must have L(h) ui ui be a subset of {li : i ∈ [d]}, where li is the literal li if ui = 1 and ¬li if li = 0. The subcomplex of X consisting of these h is a single d-cube of the pile, given by the opposite pair of vertices 0 and 2u − 1 in Rd considered as the hyperplane containing the pile. But then this case reduces to the reasoning involved in the proof that the cube resolution resolves monconjd . Hence we conclude that X⊇f is acyclic for all f, and therefore X resolves conjd . We prove the uniqueness of PF labels and therefore the minimality of X through the following series of propositions. Each face of COCd containing the vertex V⊥ is a cone over some subpile-of-subcubes, which has vertices Pw = {V ∈ {−1, 0, 1}d × {0} : Vi = w(i), ∀i ∈ dom w} for some PF w :⊆ [d] → {−1, 1}. We shall write Cw for a face associated with such a w. Obviously dim Cw = d + 1 − | dom w|. Proposition 2.64. Let W ∈ {−1, 0, 1}d × {0} be defined by Wi = w(i), ∀i ∈ dom w, and Wi = 0 otherwise. Thus W is the “center” of the subpiles-of-subcubes mentioned above. Its PF label is a total function f = fW ∈ conjd . Then the face Cw has a unique PF label Λ0 (w) := f  f −1 0 = f ∩ ⊥ as a partial function :⊆ [2d ] → [2]. Proof. By Lemma 2.48, the PF label of Cw is the intersection of the PF labels of its vertices. Since Λ0 (w) = f ∩ ⊥, Λ0 (w) ⊇ fCw . Because L(f ) ⊆ L(fV ) for all V ∈ Pw , f (u) = 0 implies fV (u) = 0, ∀V ∈ Pw . Thus Λ0 (w) ⊆ fV , ∀V ∈ Cw =⇒ Λ0 (w) = fCw as desired. Uniqueness follows from the uniqueness of preimage of 0. S The rest of the faces in COCd reside in the base, and for each face F , {L(fV ) : V ∈ F } contains at most one of each pair {¬li , li }. Define the partial order C on {−1, 0, 1}d as the product order of partial order 0 C0 +1, −1. It is easy to check that V E W implies L(fV ) ⊆ L(fW ), which further implies fV −1 (0) ⊆ fW −1 (0) and fV −1 (1) ⊇ fW −1 (1). Each face F can be described uniquely by the least and the greatest vertices in F under this order, which we denote resp. as min F and max F . 26 2 THEORY Then the vertices in F are precisely those who fall in the interval [min F, max F ] under partial order C. Proposition 2.65. Let F be a face residing in the base of COCd and write V := min F and W := max F . Then F has a unique PF label fF = Λ(V, W ) :⊆ [2d ] → {−1, 0, 1},   if ui = (1 − Vi )/2 for some i with Vi 6= 0 0 u 7→ 1 if ui = (1 + Wi )/2 for all i with Wi 6= 0   undefined otherwise. T Proof. By the observation above, we see that fF = U ∈F fU has fF−1 (0) = fV−1 (0) and fF−1 (1) = −1 fW (1). Both sets are exactly of the form described above. It remains to check that the map F 7→ fF is injective. Let f = fF for some face F . We have (max F )i = 1 iff ∀u ∈ f −1 (1), ui = 1 and (max F )0 = 0 iff ∀u ∈ f −1 (1), ui = −1. Thus f determines max F . Let v be the bit vector defined by vj = (1 + (max F )j )/2 if (max F )j 6= 0 and vj = 0 otherwise. Let v i denote v with the ith bit flipped. Then (min F )i 6= 0 iff f(v i ) = 0. For all other i, we have (min F )i = (max F )i . This proves the uniqueness of the label fF . Proposition 2.66. Every face of COCd has a unique PF label. Proof. The only thing remaining to check after the two propositions above is that faces incident on the vertex V⊥ have different PF labels from all other faces. But it is obvious that functions of the form in the previous proposition have nonempty preimage of 1, so cannot equal Λ0 (w) for any w. Summarizing our results, we have the following Theorem 2.67. βi,f (conjd ) is nonzero iff f = Λ0 (w) for some w :⊆ [d] → {−1, 1} and i = d + 1 − | dom w| or f = Λ(V, W ) for some V, W ∈ {−1, 0, 1}d , V E W . In either case, the Betti number is 1. In particular, the top dimensional nonzero Betti number is βd+1,† (conjd ) = 1. 2.3.5 Threshold Functions Let U ⊆ Rd be a finite set of points. We are interested in the class of linear threshold functions linthrU on U , defined as the set of functions of the form ( 1 if c · u > r u 7→ 0 if c · u ≤ r for some c ∈ Rd , r ∈ R. We shall assume U affinely spans Rd ; otherwise, we replace Rd with the affine span of U , which does not change the class linthrU . When U = {−1, 1}d , this is the class of linear threshold functions on d bits, and we write linthrd for linthrU in this case. Define ms (u0 , . . . , ud−1 ) = (u0 · · · us−2 us−1 , u0 · · · us−2 us , · · · , ud−s · · · ud−2 ud−1 ) as the function that outputs degree s monomials of its input. For U = Mk , the image of {−1, 1}d under the map m≤k : u 7→ (m1 u, · · · , mk u) 27 2.3 Resolutions linthrU becomes polythrkd , the class of polynomial threshold functions on d bits with degree bound k. We will construct a minimal cocellular resolution for linthrU , which will turn out to be homeomorphic as a topological space to the d-sphere S d . 1 We first vectorize the set U by mapping each point to a vector, u 7→ ~u, (u1 , . . . , ud ) 7→ ~ . Each oriented affine (u1 , . . . , ud , 1). We refer to the image of U under this vectorization as U hyperplane H in the original affine space Rd (including the hyperplane at infinity, i.e. all points get labeled positive or all points get labeled negative) corresponds naturally and bijectively to a ~ in Rd+1 which can be identified by the normal vector ν(H) ~ on the unit sphere vector hyperplane H d d+1 ~ S ⊆R perpendicular to H and oriented the same way. ~ that contains ~u is exactly For each vector ~u ∈ Rd+1 , the set of oriented vector hyperplanes H ~ residing on the equator E~u := ν(H) ~ ⊥ ∩ Sd the set of those which have their normal vectors ν(H) d d of S . This equator divides S \ E~u into two open sets: v · ~u > 0 for all v in one T (let’s call this set R~u+ ) and v · ~u < 0 for all v in the other (let’s call this set R~u− ). Note that {E~u : u ∈ U } ~ (vector) is empty, since we have assumed at the beginning that U affinely spans Rd , and thus U d+1 d spans R . The set of all such equators for all ~u divides S into distinct open subsets, which form the top-dimensional (open) cells of a cell complex. More explicitly, each cell T F (not necessarily top-dimensional and possibly empty) of this complex has a presentation as {A~u : u ∈ U } where each A~u is one of {E~u , R~u+ , R~u− }. If the cell is nonempty, then this presentation is unique and we assign the PF label fF :⊆ U → [2] defined by   if A~u = R~u+ 1 fF (u) = 0 if A~u = R~u−   undefined otherwise. ~ for some oriented affine hyperplane H such that It is easily seen that any point in F is ν(H) −1 −1 U \ dom fF lies on H, fF (1) lies on the positive side of H, and fH (0) lies on the negative side of H. If F = ∅ is the empty cell, then we assign the empty function f∅ = † as its PF label. Figure 14 illustrates this construction. We claim this labeling gives a minimal cocellular resolution X of linthrU . We show this via Lemma 2.50. Suppose f is the empty function. Then X⊇f = X, which is a complex with nontrivial reduced cohomology only at dimension d, where the is 1 (case 2 of Lemma 2.50). T rank of+itsTcohomology − Now suppose f is nonempty. Then X⊇f = u:f(u)=1 R~u ∩ u:f(u)=0 R~u is an intersection of open halfspheres. It is either empty (case 3 of Lemma 2.50) or is homeomorphic, along with its boundary in S d , to the open d-disk and its boundary (Dd , ∂Dd ), which has cohomology only at degree d because Dd /∂ ∼ = S d , where its rank is 1 (case 1 of Lemma 2.50). Thus our claim is verified. X is in fact minimal, as each cell has a unique monomial label. We have proved the following. Theorem 2.68. The colabeled complex X constructed as above is a minimal cocellular resolution of linthrU . Definition 2.69. The colabeled complex X is called the coBall resolution of linthrU , written coBallU . 1 For readers familiar with hyperplane arrangements: The cocellular resolution is essentially S d intersecting the fan of the hyperplane arrangement associated with the matroid on U . The partial function labels on the resolution are induced from the covector labelings of the fan. 28 2 THEORY u2 = (1, −1) u3 = (1, 1) 11.. .1.1 u0 = (−1, −1) u1 = (−1, 1) .10. 1..0 c · u3 = 0 0.0. ..00 Figure 14: Cocellular resolution of linthrU , where U = {u0 , u1 , u2 , u3 } as labeled in the figure. Each equator is the orthogonal space to the vector ~ui ; the case for i = 3 is demonstrated in the figure. The text in monofont are the PF labels of vertices visible in this projection. For example, 1..0 represents the PF that sends u0 to 1 and u3 to 0, and undefined elsewhere. X can be made a polytope in an intuitive way, by taking the convex hull of all vertices on X 2 . In addition, we can obtain a minimal polyhedral cellular resolution Y by taking the polar of this polytope and preserving the labels across polars. Then the empty cell of X becomes the unique dimension d + 1 cell of Y . We call this cellular resolution Y the ball resolution, written BALLU or BALL when U is implicitly understood, of linthrU . For any partial function f :⊆ U → [2], define σf to be the function   + if f(u) = 1 σf(u) = − if f(u) = 0   0 if f(u) is undefined. Let L(U ) := {sgn(ψ  U ) : ψ is a affine linear map} be the poset of covectors of U , under the pointwise order 0 < +, −, with smallest element 0. Therefore the cocircuits (minimal covectors) are the atoms of L(U ). Recall that L(U ) has a rank function defined as ( rank(a) = 0 if a is a cocircuit; rank(b) = 1 + rank(a) if b covers a. and rank(0) = −1. From the construction of coBallU , it should be apparent that each PF label is really a covector (identified by σ). There is an isomorphism between L(U ) and the face poset of coBallU : F ⊆ G ⇐⇒ fF ⊆ fG ⇐⇒ σfF ≤ σfG . Noting that rank(σfF ) = dim F (and in particular, rank(σf∅ ) = −1 = dim ∅), this observation yields the following via Proposition 2.28 2 As remarked in the previous footnote, X is the intersection of a polyhedral fan with the unit sphere. Instead of intersecting the fan with a sphere, we can just truncate the fans to get a polytope. 29 2.3 Resolutions Theorem 2.70. The Betti number βi,f (linthrU ) is nonzero only when σf is a covector of U . In this case, βi,f (linthrU ) = 1 if i = d−rank(σf), and 0 otherwise. In particular, the top dimensional Betti number of linthrU is βd+1,† (linthrU ) = 1. Via Hochster’s dual formula, this means that the canonical suboplex of linthrU is a homological (d + 1)-sphere. Let’s look at the example of U = Md , so that linthrU = polythrdd = [{−1, 1}d → 2]. In ~ is an orthgonal basis for R2d , and thus the equators of coBallU are cut out by a set of this case, U pairwise orthogonal hyperplanes. In other words, under a change of coordinates, coBallU is just the d sphere S 2 −1 cut out by the coordinate hyperplanes, and therefore is combinatorially equivalent to the complete suboplex of dimension 2d − 1, with the PF labels given by the 12 (sgn +1) function. Its polar, BALLU , just recovers the cube resolution of [2d → 2] as discussed in the beginning of Section 2.3. When linthrU = polythrkd , notice a very natural embedding of the cocellular resolution coBallMk ,→ coBallMk+1 as the section of coBallMk+1 cut out by the orthogonal complement of {w ~ γ : γ is a monomial of degree k + 1}, where wγ is all 0s except at the position where γ appears in m≤k+1 , and ~· is the vectorization function as above, appending a 1 at the end. This corresponds to the fact that a polynomial threshold of degree k is just a polynomial threshold of degree k + 1 whose coefficents for degree k + 1 monomials are all zero. This is in fact a specific case of a much more general phenomenon. Let’s call a subset P ⊆ Rn openly convex if P is convex and ∀u, v ∈ P, ∃τ > 1 : τ u + (1 − τ )v ∈ P. Examples include any open convex set in Rn , any affine subspace of Rn , and the intersections of any of the former and any of the latter. Indeed, if P and Q are both openly convex, then, P ∩ Q is convex: for any u, v ∈ P ∩ Q, if the definition of openly convex for P yields τ = ρ > 1 and that for Q yields τ = ρ0 > 1, then we may take τ = min(ρ, ρ0 ) for P ∩ Q, which works because P ∩ Q is convex. An openly convex set is exactly one which is convex and, within its affine span, is equal to the interior of its closure. Our proof that coBallU is a minimal cocellular resolution can be refined to show the following Theorem 2.71. Let U ⊆ Rn be a point set that affinely spans Rn . Let L be an openly convex cone of the vector space Rn+1 . Define Y to be the intersection of X = coBallU with L, such that each nonempty open cell Y ∩ F̊ of Y gets the same exponent label λY (Y ∩ F̊ ) = λX (F̊ ) as the open cell F̊ of X, and Y has the empty cell ∅ with monomial label xλY (∅) = 1 ∈ S iff L is vector subspace. Then Y is a minimal cocellular resolution of hxλY (F̊ ) : F̊ is a top dimensional cell in Y i. We will need a technical lemma, distinguishing the case when L is a vector subspace and when it is not. Lemma 2.72. Let L be an openly convex cone in Rq . Then either L equals its vector span, or there is an open coordinate halfspace (i.e. {v ∈ Rq : vj > 0} or {v ∈ Rq : vj < 0}) that contains L. 30 2 THEORY Proof. See Appendix A. Proof of Theorem 2.71. It suffices to show that Y⊇f for any PF f satisfies one of the three conditions of Lemma 2.29, and the minimality would follow from the uniqueness of labels. If L is a vector subspace, Y⊇† = Y is a sphere (condition 2). Otherwise, Y⊇† = Y is contained in an open halfspace H, and thus by projection from the origin onto an affine subspace ∂H 0 parallel to ∂H, Y is homeomorphic to ∂H 0 ∩ L, an openly convex set of dimension dim Y (condition 1). Whether L is a vector space, for any nonempty PF f, Y⊇f is the intersection of the unit sphere (the underlying space of coBallMd ), L, and a number of open halfspaces, and thus the intersection of openly convex sets contained in an open halfspace. This is again homeomorphic to an openly convex set of dimension dim Y via projection to an affine subspace, if it is not empty. (condition 1/condition 3). Linear functionals on Md are bijective with real functions on the boolean d-cube {−1, 1}d . Therefore the cone L represents a cone of real functions when U = {−1, 1}d , and Y is a minimal cellular resolution of the threshold functions of L. In other words, we have the following corollary Corollary 2.73. Let C ⊆ [{−1, 1}d → 2] be the class obtained by strongly thresholding an openly convex cone L of real functions {−1, 1}d → R, i.e. C = { 21 (sgn(f )+1) : f ∈ L, ∀u ∈ {−1, 1}d [f (u) 6= 0]}. Then C has a minimal cocellular resolution of dimension equal to the dimension of the affine hull of L. This corollary specializes to the case when L is any vector subspace of boolean functions. The examples explored in the beginning of this section took L as degree bounded polynomials. We make the following formal definitions. Definition 2.74. Let L be a cone of real functions on {−1, 1}d . Suppose C = { 12 (sgn(f ) + 1) : f ∈ L, f (u) 6= 0, ∀u ∈ {−1, 1}d }. We say C is the strongly thresholded class of L, written C = thr L. We call C thresholded convex if L is openly convex. We call C thresholded linear if L is linear. While this corollary produces minimal cocellular resolutions for a large class of functions, it does not apply to all classes. For example, the corollary shows that the Betti numbers of thresholded convex classes are either 0 or 1, but as we show in the next section, the linear functionals over finite fields have very large Betti numbers, so cannot be a thresholded convex class. 2.3.6 Linear Functionals over Finite Fields d Let p be a prime power. Define linfunpd ∼ = Fd∗ p ⊆ [p → p] to be the class of linear functionals p over the d-dimensional vector space [p]d ∼ = Fdp . We will refer to elements of linfund as covectors. Denote the affine span of a set of elements g1 , . . . , gk by Lg1 , . . . , gk M. In this section we construct the minimal resolution of linfundp . Fix a linear order C on Fd∗ p . We construct as follows a DAG Td of depth d + 1 (with levels 1, ..., d + 1), whose nodes are of the form (f, V ) where V is an affine subspace of the dual space Fd∗ p and f is the C-least element of V . (Therefore if any affine subspace appears in a node, then it appears only in that node — indeed, every affine subspace appears in exactly one node.) There is only one node at level 1, which we call the root. This is the C-least element along with V = Fd∗ p . For any node (f, V ) where dim V > 1, we add as its children the nodes (g, W ) where W is a codimension-1 affine subspace of V not containing f , and g is the C-least element of W . By simple induction, one sees that all affine subspaces appearing on level i of Td has dimension d − i. In 31 2.3 Resolutions particular, the nodes at level d + 1, the leaf nodes, are all of the form (f, {f }). This completes the construction of Td . For each path (f1 , V1 = Fd∗ p ), (f2 , V2 ), . . . , (fd+1 , Vd+1 ) from the root to a leaf node, we have by construction f1 C f2 C · · · C fd C fd+1 . Therefore, every such path is unique. Lemma 2.75. Any node (f, V ) at level i of Td has exactly pd−i − 1 children. Proof. The children of (f, V ) are in bijection with the set of codimension-1 affine subspaces of V not containing f . Each nonzero covector in V ∗ defines a vector hyperplane in V , whose cosets determine p parallel affine hyperplanes. Exactly one of these affine hyperplanes contain f . Covectors f and g in V ∗ determine the same hyperplane if f = cg for some constant c ∈ Fdp , c 6= 0. As remarked above, d−i V has dimension d − i, and so has cardinality pd−i . Therefore there are p p−1−1 (p − 1) = pd−i − 1 affine hyperplanes of V not containing f . Q d−i − 1) maximal paths in the DAG T . (When d = 0, Lemma 2.76. There are Up (d) := d−1 d i=0 (p Up (d) := 1.) Proof. Immediately follows from the previous lemma. For example, suppose p = 2 and C is the right-to-left lexicographic order on the covectors: 0 · · · 00 C 0 · · · 01 C 0 · · · 10 C · · · C 1 · · · 10 C 1 · · · 11, where a covector (x1 , . . . , xd ) 7→ a1 x1 + · · · ad xd is abbreviated as the bitstring a1 a2 · · · ad . When d = 3, the root is (000, F3∗ 2 ). There are then seven dimension 3 − 1 = 2 affine planes in F3p not containing 000, so seven nodes at level 1: • Covector 001 for all affine planes containing 001, which are {001, 111, 101, 011}, {001, 111, 100, 010}, {001, 101, 110, 010}, {001, 100, 110, 011}. • There are 3 other affine planes, which correspond to the following nodes 1. (100, {111, 110, 100, 101}) 2. (010, {111, 011, 010, 110}) 3. (010, {010, 011, 101, 100}) Or, suppose we choose to order covectors by the number of 1s and then lexicographically, 0 · · · 000 ≺ 0 · · · 001 ≺ 0 · · · 010 ≺ 0 · · · 100 ≺ · · · 10 · · · 000 ≺ 0 · · · 011 ≺ 0 · · · 101 ≺ 0 · · · 110 ≺ · · · ≺ 1 · · · 11. Then the DAG will be exactly the same as above. Once we have built such a DAG Td , we can construct the corresponding cellular resolution X ? 3 The cellular resolution will be simplicial and pure of dimension d. Its vertex set is of Ilinfun p. d linthrp ∼ = Fd∗ ; each vertex has itself as the PF label. For each maximal path d p (f1 , V1 = Fd∗ p ), (f2 , V2 ), . . . , (fd+1 , Vd+1 ), we add a top simplex (of dimension d) with the vertex set {f1 , f2 , . . . , fd+1 }. As usual, the PF label of a face F ⊆ linthrpd is just the intersection of the PF labels of its vertices. Lemma 2.77. For an k-dimensional face F of X, its PF label is a linear functional on a vector subspace of Fdp of dimension d − k. 3 If we treat Td as a poset, then the cellular resolution as a complex is a quotient of the order complex of Td by identifying (f, V ) with (g, W ) iff f = g. 32 2 THEORY Proof. F has kT+ 1 vertices, f0 , . . . , fk . Their intersection is the partial function defined on the subspace W = ki=1 ker(fi − fi−1 ), and equals the restriction of fi to W for any i. The affine independence of {f0 , . . . , fk } implies the vector independence of {(f1 − f0 ), . . . , (fk − fk−1 )}. Therefore W has codimension k, as desired. Now, to show that X is a minimal resolution, we will require the following lemma. Lemma 2.78. Fix any linear order C on Fd∗ p . Suppose (g1 , g2 , . . . , gk ) is a sequence of covectors such that gi is the C-least element of the affine space generated by (gi , gi+1 , . . . , gk ). Then there is a maximal path in Td containing (g1 , . . . , gk ) as a subsequence. Proof. We proceed by induction on k. When k = 0, the claim is vacuously true. Assume k ≥ 1. We will show that there is a path ℘ from the root to a node (g1 , V ) with V containing W = Lg1 , . . . , gk M, the affine subspace generated by g1 , . . . , gk . Then we apply the induction hypothesis with Fd∗ p replaced by W and (g1 , g2 , . . . , gk ) replaced by (g2 , g3 , . . . , gk ) to obtain a path from (g1 , V ) to a leaf node, which would give us the desired result. The first node of ℘ is of course the root. We maintain the invariant that each node (f, W 0 ) added to ℘ so far satisfies W 0 ⊇ W . If we have added the nodes (f1 , V1 ), (f2 , V2 ), . . . , (fj , Vj ) in p, then either fj = g1 , in which case we are done, or Vj is strictly larger than W . In the latter case, there exists an affine subspace Vj+1 of Vj with W ⊆ Vj+1 ⊂ Vj and dim Vj+1 = dim Vj − 1, and we add (fj+1 , Vj+1 ) to the path ℘, with fj+1 being the C-least element of Vj+1 . This process must terminate because the dimension of Vj decreases with j, and when it does, we must have fj = g1 , and the path ℘ constructed will satisfy our condition. Theorem 2.79. X is a d-dimensional complex that minimally resolves linfunpd . Proof. To prove that X is a cellular resolution, it suffices to show that X⊇f for any partial function f :⊆ [pd → p] is acyclic. The set of f ∈ Fd∗ p extending f is an affine subspace W . Our strategy is to prove that if {g1 , . . . , gk } generates an affine subspace of W and is a face of X, then {g, g1 , . . . , gk } is also a face of X, where g is the C-least element of W . This would show that X⊇f is contractible and thus acyclic. But this is precisely the content of Lemma 2.78: Any such {g1 , . . . , gk } can be assumed to be in the order induced by being a subsequence of a maximal path of Td . This means in particular that gi is the least element of Lgi , . . . , gk M. A fortiori, {g, g1 , g2 , . . . , gk } must also satisfy the same condition because g is the least element of W . Therefore Lemma 2.78 applies, implying that {g, g1 , g2 , . . . , gk } is a face of X, and X is a cellular resolution as desired. The resolution is minimal since the PF label of any face P is a covector defined on a strictly larger subspace than those of its subfaces. Definition 2.80. The resolution X is called the flag resolution, FLAGpd , of linfunpd with respect to C. Theorem 2.81. The Betti number βi,g (linfunpd ) is nonzero only when g is a linear functional defined on a subspace of Fdp , and i = d − dim dom g. In this case, it is equal to Up (i) (as defined in Lemma 2.76). Proof. All the cells in the resolution X have exponent labels of the form Γg as stated in the theorem, and by Lemma 2.77, such cells must have dimension i = d − dim(dom g). It remains to verify that the number B of cells with PF label g is Up (i). The subset of Fd∗ p that extends g is an affine subspace W of dimension d − dim dom g = i. The number B is the number of sequences (g0 , . . . , gi ) ∈ W i+1 such that gj is the C-least element of Lgj , . . . , gi M for each j, and such that Lg0 , . . . , gi M = W . If we treat W ∼ = Fi∗ p and construct Ti on W , then B is exactly the number of maximal paths of Ti , which is Up (i) by Lemma 2.76. 33 2.3 Resolutions 1011 ..11 1..1 ...1 1101 .1.1 110. 1100 0111 Figure 15: An example of a nonpure minimal resolution of a boolean function class ⊆ [4 → 2]. The labels are PF labels. For example, ..11 represents a partial function sending 2 and 3 to 1, and undefined elsewhere. 1000 0010 .0.0 .00. ..00 1... 00.. ..1. 1111 0001 ...1 0..0 0.0. 0100 .1.. Figure 16: Another example of nonpure minimal resolution of a boolean function class ⊆ [4 → 2]. Only the vertices and edges are labeled (with PF labels). Note that the maximal cells are the three triangles incident on the vertex 1111 and the tetrahedron not incident on 1111. They have the same PF label, the empty function †. Therefore it is possible for a boolean function class to have nonzero Betti numbers in different dimensions for the same degree. As discussed in Section 2.3.5, we have the following corollary because the Betti numbers of linfun2d can be greater than 1. Corollary 2.82. linfun2d is not a thresholded convex class. 2.3.7 Abnormal Resolutions All of the classes exhibited above have pure minimal resolutions, but this need not be the case in general. Figure 15 gives an example of a nonpure minimal resolution of a class ⊆ [4 → 2]. It consists of a segment connected to a (solid) triangle. This example can be generalized as follows. Let C ⊆ [n + 1 → 2] be {¬δi = Ind(u 6= i) : i ∈ [n]} ∪ {g := I(u 6∈ {n − 1, n})}. Let X be the simplicial complex on vertex set C, consisting of an (n − 1)-dimensional simplex on {¬δi : i ∈ [n]}, and a segment attaching ¬δn−1 to g. With the natural PF labels, X minimally resolves C and is nonpure. Definition 2.83. We say a class C has pure Betti numbers if for every PF f, βi,f (C) 6= 0 for at most one i. All of the classes above discussed in the previous sections have pure Betti numbers. But this is not true in general. Figure 16 shows a minimal resolution of a class D ⊆ [4 → 2] that has three 34 2 THEORY triangles and one tetrahedron as its top cells, and they all have the empty function as the PF label. Thus β2,† (D) = β3,† (D) = 1. This example can be generalized as follows. Let D ⊆ [n → 2] be {δi : i ∈ [n]} ∪ {1}. Let X be the simplicial complex on vertex set D, consisting of an (n − 1)dimensional simplex on {δi : i ∈ [n]} and triangles on each triple {δi , δj , 1} for each i 6= j. With the natural PF labels, X is a minimal cellular resolution of D, and the cells with PF label † are exactly the (n − 1)-dimensional simplex and each of the triangles incident on 1. Thus the gap between the highest nontrivial Betti number and the lowest nontrivial Betti number for the same partial function can be linear in the size of the input space. 2.4 Partial Function Classes Most of the definitions we made actually apply almost verbatim to partial function classes C ⊆ [⊆ n → 2]. Here we list the corresponding definitions for PF classes and the propositions that hold PF classes as well as for function classes. We omit the proofs as they are similar to the ones given before. Definition 2.84. Let C ⊆ [⊆ n → m]. The canonical suboplex SC of C is the subcomplex of the complete (n − 1)-dimensional m-suboplex consisting of all cells Ff where f has an extension in C. The canonical base ring S of C is the same as the canonical base ring of [n → m]. The Stanley-Reisner ideal IC of C is defined as the Stanley-Reisner ideal of SC with respect to S. The canonical ideal of C is the dual ideal IC? of its Stanley-Reisner ideal. It is generated by {xΓf : f ∈ C}, and generated minimally by {xΓf : f ∈ C is maximal}. A Betti number βi,b (IC? ) is nonzero only if b = Γf for some partial function f with extension in C. Thus we define βi,f (C) = βi,Γf (IC? ). Proposition 2.85 (Counterpart of Proposition 2.38). Let C ⊆ [⊆ n → m]. Each minimal generator of IC is either 1) xu,i xu,j for some u ∈ [n], i 6= j ∈ [m], or 2) xgraph f for some partial function f :⊆ [n] → [m] such that f has no extension in C, but every proper restriction of f does. In addition, the set of all such monomials is exactly the set of minimal generators of IC . Definition 2.86. Let C ⊆ [⊆ n → m]. A labeled complex (X, λ) is a (co)cellular resolution of partial class C if (X, λ) is a (co)cellular resolution of S/IC? . Proposition 2.87 (Counterpart of Lemma 2.48). If (X, λ) is a cellular resolution T of a partial class C ⊆ [⊆ n → m], then it is PF-labeled as well. The PF label λ(F ) of a face F is V ∈F fV . Lemma 2.49 and Lemma 2.50 give conditions on when a PF-(co)labeled complex is a resolution, and they apply verbatim to resolutions of partial classes as well. Proposition 2.51 and Corollary 2.52 hold as well when C is replaced by a partial class C, but we will not use them in the sequel. 2.5 Combining Classes We first give a few propositions on obtaining resolutions of a combination of two classes C and D from resolutions of C and D. Proposition 2.88. Let I and J be two ideals of the same polynomial ring S. If (XI , λI ) is a polyhedral cellular resolution of S/I, and (XJ , λJ ) is a cellular resolution of S/J, then the join (XI ? XJ , λI ?λJ ) is a cellular resolution of S/(I +J), where we define λI ?λJ (F ?G) := lcm(λI (F ), λJ (G)). Proof. Let a be an exponent sequence. (XI ? XJ )a is precisely (XI )a ? (XJ )a , which is acyclic when both (XI )a and (XJ )a are acyclic. So XI ? XJ is a resolution. The 0-cells of XI ? XJ are just the 0-cells of XI union the 0-cells of XJ , with the same labels, so XI ? XJ resolves I + J. 35 2.5 Combining Classes Note however that in general XI ? XJ is not minimal even when XI and XJ both are. Proposition 2.89. Let C and D be classes ⊆ [m → n]. If (XC , λC ) is a cellular resolution of C, and (XD , λD ) is a cellular resolution of D, then the join (XC ? XD , λC ? λD ) is a cellular resolution of C ∪ D. If µC is the PF labeling function of XC and µD is the PF labeling function of XD , then the PF labeling function of XC ? XD is given by µC ? µD (F ? G) := µC (F ) ∩ µD (G). Proof. By the above proposition, (XC ?XD , λC ?λD ) resolves IC? +ID? , which has minimal generators {xΓf : f ∈ C ∪ D}. The characterization of µC ? µD follows from the the definition of λC ? λD . We will need to examine the “difference” between the Betti numbers of I + J and those of I and J. The following lemma gives a topological characterization of this difference. Lemma 2.90. Let I and J be two monomial ideals of the same polynomial ring S. Suppose (XI , λI ) is a polyhedral cellular resolution of S/I, and (XJ , λJ ) is a cellular resolution of S/J. Label XI ×XJ by the function λI × λJ : F × G 7→ lcm(λI (F ), λJ (G)) for nonempty cells F and G; the empty cell has exponent label 0. If σ is an exponent sequence, then there is a long exact sequence e i ((XI × XJ )≺σ ) → H e i ((XI )≺σ ) ⊕ H e i ((XJ )≺σ ) → H e i ((XI ? XJ )≺σ ) → · · · ··· → H where i decreases toward the right. Proof. One can check that (XI ? XJ )≺σ is the homotopy pushout of (XI )≺σ ← (XI × XJ )≺σ → (XJ )≺σ . The lemma then follows from the homotopy pushout exact sequence. We also have an algebraic version. Lemma 2.91. Let I and J be two monomial ideals of the same polynomial ring S. For each exponent sequence a, there is a long exact sequence · · · → kβi,a (I∩J) → kβi,a (I) ⊕ kβi,a (J) → kβi,a (I+J) → kβi−1,a (I∩J) → · · · Proof. We have a short exact sequence 0 → I ∩ J → I ⊕ J → I + J → 0. By Proposition 2.10, we can apply Tor(−, k) to obtain the long exact sequence as stated. The ideal I ∩ J is generated by {lcm(mi , mj ) : mi ∈ mingen(I), mj ∈ mingen(J)}. When I = IC? and J = ID? , I ∩ J = hxΓ(f ∩g) : f ∈ C, g ∈ D}. Define the Cartesian Intersection C  D of C and D ? . to be {f ∩ g : f ∈ C, g ∈ D}. This is a class of partial functions, and we can check IC? ∩ ID? = ICD So the above lemma can be restated as follows Lemma 2.92. Let C, D ⊆ [n → m]. For each PF f :⊆ [n] → [m], there is a long exact sequence · · · → kβi,f (CD) → kβi,f (C) ⊕ kβi,f (D) → kβi,f (C∪D) → kβi−1,f (CD) → · · · Next, we seek to produce from cellular resolutions of C and D a cellular resolution of the Cartesian Union C q D of two classes C ⊆ [U → V ], D ⊆ [U 0 → V 0 ], defined as the class with elements f q g : U t U 0 → V t V 0 for f ∈ C, g ∈ D, defined by ( f (u) if u ∈ U f q g(u) = g(u) else. We start with the general version for ideals, and specialize to function classes. 36 2 THEORY Proposition 2.93. Let I be an ideal of polynomial ring S and let J be an ideal of polynomial ring T such that S and T share no variables. If (XI , λI ) resolves S/I and (XJ , λJ ) resolves S/J, then (XI × XJ , λI q λJ ) resolves the ideal S/(I ⊗ J) with I ⊗ J := (I ⊗ T )(S ⊗ J) in the ring S ⊗ T , where λI q λJ (F × G) = λI (F )λJ (G) for any cells F ∈ XI and G ∈ XJ . (Here, tensor ⊗ is over base ring k). Furthermore, if XI and XJ are both minimal then (XI × XJ , λI q λJ ) is minimal as well. Proof. Let ω 0 , . . . , ω p−1 be minimal monomial generators of I and let γ 0 , . . . , γ q−1 be minimal monomial generators of J. The ideal I ⊗ J is generated by {ω i γ j : (i, j) ∈ [p] × [q]}, which are furthermore minimal because {ω i }i and {γ j }j are respectively minimal, and S and T share no variables. The complex XI × XJ has vertices Vi × Vj0 for vertices Vi ∈ XI and Vj ∈ XJ . If Vi has label ω i and Vj0 has label γ j , then Vi × Vj0 has label ω i γ j via λI q λJ . Thus XI × XJ resolves S/(I ⊗ J), if it is a resolution. And in fact, it is, because for any exponent sequence a wrt S and exponent sequence b wrt T , (XI × XJ )aqb = (XI )a × (XJ )b , which is acyclic (Here a q b is the exponent sequence whose values on variables in S come from a and whose values on variables in T come from b). The faces of a cell F × G ∈ XI × XJ are {F × G0 : G0 ⊆ ∂G, dim G0 = dim G − 1} ∪ {F 0 × G : F 0 ⊆ ∂F, dim F 0 = dim F − 1}. If λI (F ) 6= λI (F 0 ) for any F 0 ⊂ F and λJ (G) 6= λJ (G0 ) for any G0 ⊂ G, then λI q λJ (F × G) = λI (F )λJ (G) is not equal to any of λI (F 0 )λJ (G) or λI (F )λJ (G0 ) for any of the above F 0 or G0 . Therefore (XI × XJ , λI q λJ ) is minimal if XI and XJ are. Proposition 2.94. Let C ⊆ [U → V ] and D ⊆ [U 0 → V 0 ]. If (XC , λC ) is a cellular resolution of C, and (XD , λD ) is a cellular resolution of D, then the product (XC × XD , λC q λD ) is a cellular resolution of C q D. Furthermore, if XC and XD are both minimal then (XC × XD , λC q λD ) is minimal as well. Finally, we want to construct cellular resolutions of restrictions of a function class to a subset of its input space. Definition 2.95. Let C ⊆ [U → V ] and U 0 ⊆ U . Then the restriction class C  U 0 ⊆ [U 0 → V ] is defined as C  U 0 = {f  U 0 : f ∈ C}. Again we start with a general algebraic version and then specialize to restriction classes. Proposition 2.96. Let X := {xi : i ∈ [n]} and Y := {yj : j ∈ [m]} be disjoint sets of variables. Let I be an ideal of polynomial ring S = k[X t Y]. Suppose (X, λ) resolves I. Then (X, λ Y ) resolves the ideal I/hxi − 1 : xi ∈ Xi in the ring k[Y], where λ Y (F ) := λ(F )/hxi − 1 : xi ∈ Xi. Essentially, if we just ignore all the variables in X then we still get a resolution, though most of the time the resulting resolution is nonminimal even if the original resolution is. Proof. The subcomplex (X, λ Y )ya for a monomial ya in k[Y] is exactly the subcomplex (X, λ)x1 ya , and hence acyclic. One can easily see that the Stanley-Reisner ideal of C  U 0 is IC /hxu,v − 1 : u 6∈ U 0 , v ∈ V i and similarly the canonical ideal of C  U 0 is IC? /hxu,v − 1 : u 6∈ U 0 , v ∈ V i (both ideals are of the polynomial ring S[xu,v : u ∈ U 0 , v ∈ V ]). Then the following corollary is immediate. 37 2.5 Combining Classes A B zn−1 S Figure 17: S as the union A ∪ B. Proposition 2.97. Let C ⊆ [U → V ] and U 0 ⊆ U . If (X, λ) is a cellular resolution of C, then (X, λ  U 0 × V ) resolves C  U 0 , where λ  U 0 × V := λ  {xu,v : u ∈ U 0 , v ∈ V }. Similarly, if C is an algebraic free resolution of IC? , then C  U 0 := C/hxu,v − 1 : u 6∈ U 0 , v ∈ V i is an algebraic free ? resolution of ICU 0. Finally we show that there is a series of exact sequences relating the Betti numbers of C ⊆ [n → 2] to the Betti numbers of C  U ⊆ [n]. All of the below homology are with respect to k. Definition 2.98. Let C ⊆ [n → m] and f :⊆ [n] → [m]. The class C filtered by f, C  f, is {f \ f : f ⊆ f ∈ C}. For any U ⊆ [n] × [m] that forms the graph of a partial function f, we also write C  U = C  f. It should be immediate that SCU = linkU SC , so that by Hochster’s dual formula, e i−1 (SCf ). e i−1 (linkgraph f SC ) = dimk H βi,f (C) = dimk H Consider the standard embedding of the complete (n − 1)-dimensional suboplex S1n−1 ∼ = {z ∈ : kzk1 = 1}. Then SC ⊆ S1n−1 is the union of two open sets: A := SC ∩ {z ∈ Rn : |zn−1 | < 2/3} and B := SC ∩ {z ∈ Rn : |zn−1 | > 1/3} (see Figure 17). If all functions in C sends n − 1 to the same output, then B is homotopy equivalent to a single point; otherwise B contracts to 2 points. A deformation retracts onto SC[n−1] . The intersection A ∩ B deformation retracts to the disjoint union of two spaces, respectively homeomorphic to the links of SC with respect to the vertices (n − 1, 0), (n − 1, 1) ∈ [n] × [2]. We therefore have the following long exact sequence due to Mayer-Vietoris Rn e i+1 (SC ) → H e i (SC(n−1,0) ) ⊕ H e i (SC(n−1,1) ) → H e i (SC[n−1] ) ⊕ H e i (B) → H e i (SC ) → · · · ··· → H If every function f ∈ C has f (n − 1) = 1, then C  (n − 1, 1) = C  [n − 1]; a similar thing happens if all f (n − 1) = 0. So suppose C  {n − 1} = [2]. Then B ' ••, and neither C  (n − 1, 0) nor C  (n − 1, 1) are empty. Therefore the long exact sequence simplifies down to e i+1 (SC ) → H e i (SC(n−1,0) ) ⊕ H e i (SC(n−1,1) ) → H e i (SC[n−1] ) ⊕ ZI(i=0) → H e i (SC ) → · · · ··· → H Note that for any simplicial complex ∆, the link and restriction operations commute: linkτ (∆  σ) = (linkτ ∆)  σ. 38 2 THEORY Correspondingly, for function class C, filtering and restricting commute: C  U  V = C  V  U. Let U := graph f for some f :⊆ [n−1] → [2] and denote U0 := U ∪{(n−1, 0)}, U1 := U ∪{(n−1, 1)}. The above long exact sequence generalizes to the following, by replacing C with C  U and applying the commutativity above: e i+1 (SCU ) → H e i (SCU ) ⊕ H e i (SCU ) → H e i (SC[n−1]U ) ⊕ ZI(i=0) → H e i (SCU ) → · · · ··· → H 0 1 This yields via Hochster’s formulas the following sequence relating the Betti numbers of C and C  [n − 1]. Theorem 2.99. Let C ⊆ [n → 2], f :⊆ [n − 1] → [2], and f0 := f ∪ (n − 1 7→ 0), f1 := f ∪ (n − 1 7→ 1). We have an exact sequence · · · → kβi+1,f (C) → kβi,f0 (C)+βi,f1 (C) → kβi,f (C[n−1])+I(i=−1) → kβi,f (C) → · · · Using Theorem 2.99 we can recapitulate the following fact about deletion in oriented matroids. Below we write V \ u for V \ {u} in the interest of clarity. Corollary 2.100. Let V be a point configuration with affine span Rd and u ∈ V . Suppose V \ u has affine span Rd−e , where e is either 0 or 1. Then τ ∈ {−, 0, +}V \u is a covector of rank r of V \ u iff one of the following is true: 1. τ− := τ ∪ (u 7→ −) is a covector of rank r + e of V . 2. τ+ := τ ∪ (u 7→ +) is a covector of rank r + e of V . 3. τ0 := τ ∪ (u 7→ 0) is a covector of rank r + e of V , but τ− and τ+ are not covectors of V . Proof. Let C = linthrV and D = linthrV \u = C  (V \ u). Write f := σ −1 τ, f0 := σ −1 τ0 , f+ := σ −1 τ+ , f− := σ −1 τ− . βi,f (C) = 1 iff σf is a covector of V of rank d − i by Theorem 2.70. If Item 1 is true, but not Item 2, then τ0 cannot be a covector of V (or else subtracting a small multiple of τ− from τ0 yields τ+ ). As C and D both have pure Betti numbers, we have an exact sequence 0 → kβj,f− (C) → kβj,f (D) → 0 where j = d − rank τ− . This yields that τ is a covector of rank d − e − j = rank τ− − e. The case that Item 2 is true but not Item 1 is similar. If Item 1 and Item 2 are both true, then τ0 must also be a covector. Furthermore, it must be the case that rank τ− = rank τ+ = rank τ0 + 1. Again as C and D have pure Betti numbers, we have an exact sequence 0 → kβj+1,f0 (C) → kβj,f− (C)+βj,f+ (C) → kβj,f (D) → 0 where j = d − rank τ− . Thus τ is a covector of rank d − e − j = rank τ− − e. Finally, if Item 3 is true, we immediately have an exact sequence 0 → kβj,f (D) → kβj,f0 (C) → 0 with j = d − rank τ0 , so τ is a covector of rank d − e − j = rank τ0 − e. In general, if C ⊆ [n → 2] and C  [n − 1] are known to have pure Betti numbers, then Theorem 2.99 can be used to deduce the Betti numbers of C  [n − 1] directly from those of C. This strategy is employed in the proof of Corollary 3.32 in a later section. It is an open problem to characterize when a class has pure Betti numbers. 39 3 Applications 3.1 Dimension Theory In this section we investigate the relationships between VC dimension and other algebraic quantities derived from the Stanley-Reisner ideal and the canonical ideal. Definition 3.1. Suppose C ⊆ [n → 2]. We say C shatters a subset U ⊆ [n] if C  U = [U → 2]. The VC dimension of C, dimVC C, is defined as the largest k such that there is a subset U ⊆ [n] of size k that is shattered by C. The VC radius of C, radVC C, is defined as the largest k such that all subsets of [n] of size k are shattered by C. The VC dimension is a very important quantity in statistical and computational learning theory. For example, suppose we can obtain data points (u, f (u)) by sampling from some unknown distribution u ∼ P, where f is an unknown function known to be a member of a class C. Then the number of samples required to learn the identity of f approximately with high probability is O(dimVC C) [10]. Simultaneous ideas also popped up in model theory [17]. In this learning theory perspective, an extenture f of C is what is called a minimal nonrealizable sample: there is no function in C that realizes the input/output pairs of f, but there is such functions for each proper subsamples (i.e. restrictions) of f. Note that C shatters U iff ICU = IC ⊗S S/JU equals hxu,0 xu,1 : u ∈ U i as an ideal of S/JU , where JU = hxū,v − 1 : ū 6∈ U, v ∈ V i. In other words, every nonfunctional minimal monomial generator of IC gets killed when modding out by JU ; so C shatters U iff every extenture of C is defined on a point outside U . Therefore if we choose U to be any set with |U | < min{| dom f| : f ∈ ex C}, then C shatters U . Since dom f is not shattered by C if f is any extenture, this means that Theorem 3.2. For any C ⊂ [n → 2] not equal to the whole class [n → 2], radVC C = min{| dom f| : f ∈ ex C} − 1. Define the collapsing map π : k[xu,0 , xu,1 : u ∈ [n]] → k[xu : u ∈ [n]] by π(xu,i ) = xu . If U ⊆ [n] is shattered by C, then certainly all subsets of U are also shattered by C. Thus the collection of shattered sets form an abstract simplicial complex, called the shatter complex SHC of C. Theorem 3.3. Let I be the the Stanley-Reisner ideal of the shatter complex SHC in the ring S 0 = k[xu : u ∈ [n]]. Then π∗ IC = I + hx2u : u ∈ [n]i. Equivalently, U ∈ SHC iff xU 6∈ π∗ IC . Proof. U is shattered by C iff for every f : U → [2], f has an extension in C, iff xgraph f 6∈ IC , ∀f : U → [2], iff xU 6∈ π∗ IC . We immediately have the following consequence. Theorem 3.4. dimVC C = max{|U | : xU 6∈ π∗ IC }. Recall the definition of projective dimension [11]. Definition 3.5. The length of a minimal resolution of a module M is the called the projective dimension, projdim M , of M . We make the following definitions in the setting of function classes. Definition 3.6. For any C ⊆ [n → 2], the homological dimension dimh C is defined as the projective dimension of IC? , the length of the minimal resolution of IC? . The Stanley-Reisner dimension dimSR C is defined as the projective dimension of the Stanley-Reisner ring S/IC . 40 3 APPLICATIONS [n → 2] {f } deltan monconjd conjd linthrd polythrkd linfun2d dimh n 0 n−1 d d+1 d+1 Σk0 d dimSR n n n+1 2d+1 − d − 1 2d+1 − d − 1 2d+1 − d − 1 2d+1 − Σk0 = 2d + Σdk+1 2d+1 − d − 1 dimVC n 0 1 d [14] d [14] d + 1 [1] Σk0 [1] d Table 1: Various notions of dimensions for boolean function classes investigated in this work. Σkj :=
Pk d i=j i . The VC dimensions without citation can be checked readily. One can quickly verify the following lemma. Lemma 3.7. If S/IC has a minimal cellular resolution X, then dimSR C = dim X + 1. If C has a minimal cellular resolution X, then dimh C = dim X. The same is true for cocellular resolutions Y if we replace dim X with the difference between the dimension of a top cell in Y and that of a bottom cell in Y . Recall the definition of regularity [11]. Definition 3.8. The regularity of a Nn -graded module M is reg M = max{|b| − i : βi,b (M ) 6= 0}, where |b| = Pn j=1 bi . There is a well known duality between regularity and projective dimension. Proposition 3.9. [11, thm 5.59] Let I be a squarefree ideal. Then projdim(S/I) = reg(I ? ). This implies that the Stanley-Reisner dimension of C is equal to the regularity of IC? . For each minimal resolutions we have constructed, it should be apparent that max{|Γf| − i : βi,f (C) 6= 0} occurs when i is maximal, and thus for such an f with smallest domain it can be computed as #variables − | dom f| − dimh C. Altogether, by the results of Section 2.3, we can tabulate the different dimensions for each class we looked at in this work in Table 1. For all classes other than deltan , we see that dimh is very close to dimVC . We can in fact show the former is always at least thte latter. Proposition 3.10. Let C ⊆ [U → V ] and U 0 ⊆ U . Then dimh C ≥ dimh C  U 0 . Proof. Follows from Proposition 2.97. Theorem 3.11. For any C ⊆ [n → 2], dimh C ≥ dimVC C. Proof. Let U ⊆ [n] be the largest set shattered by C. We have by the above proposition that dimh C ≥ dimh C  U . But C  U is the complete function class on U , which has the cube minimal resolution of dimension |U |. Therefore dimh C ≥ |U | = dimVC C. As a consequence, we have a bound on the number of minimal generators of an ideal I expressable as a canonical ideal of a class, courtesy of the Sauer-Shelah lemma [10]. 41 3.1 Dimension Theory Corollary 3.12. Suppose ideal I equals IC? for some C ⊆ [n → 2]. Then I is minimally generated by a set no larger than O(nd ), where d is the projective dimension of I. However, in contrast to VC dimension, note that homological dimension is not monotonic: delta2d ⊆ conjd but the former has homological dimension 2d while the latter has homological dimension d + 1. But if we know a class C ⊆ [n → 2] has dimh C = dimVC C, then C ⊆ D implies dimh C ≤ dimh D by the monotonicity of VC dimension. We write this down as a corollary. Corollary 3.13. Suppose C, D ⊆ [n → 2]. If dimh C = dimVC C, then C ⊆ D only if dimh C ≤ dimh D. The method of restriction shows something more about the Betti numbers of C. Theorem 3.14. C shatters U ⊆ [n] iff for every partial function f :⊆ U → [2], there is some g :⊆ [n] → [2] extending f such that β|U |−| dom f|,g (C) ≥ 1. Proof. The backward direction is clear when we consider all total function f : U → [2]. ? From any (algebraic) resolution F of IC? , we get an (algebraic) resolution F  U of ICU by ignoring the variables {xu,v : u 6∈ U, v ∈ [2]}. If for some f :⊆ U → [2], for all g :⊆ [n] → [2] extending f, β|U |−| dom f|,g C = 0, then there is the (|U | − | dom f|)th module of F  U has no summand of degree Γg, which violates the minimality of the cube resolution of C  U . There is also a characterization of shattering based on the Stanley-Reisner ideal of a class. We first prove a trivial but important lemma. e n (∆) 6= 0 iff ∆ is complete.4 Lemma 3.15. Suppose ∆ is an n-dimensional suboplex. Then H Proof. The backward direction is clear. Write S1n for the complete n-dimensional suboplex. Suppose ∆ 6= S1n . Choose an n-dimensional simplex F not contained in ∆. Let ∇ be the complex formed by the n-dimensional simplices not contained in ∆ or equal to F . By Mayer-Vietoris for simplicial complexes, we have a long exact sequence e n (∇) ⊕ H e n (∆) → H e n (∇ ∪ ∆) → H e n−1 (∇ ∩ ∆) → · · · e n (∇ ∩ ∆) → H ··· → H Now ∇ ∪ ∆ is just S1n \ int F , which is homeomorphic to an n-dimensional disk, and hence cone m (∇ ∪ ∆) = 0, ∀m > 0, and therefore H e m (∇ ∩ ∆) ∼ e m (∇) ⊕ H e m (∆), ∀m > 0. tractible. Hence H =H e e e n (∆) = 0, as But ∇ ∩ ∆ has dimension at most n − 1, so Hn (∇ ∩ ∆) = 0, implying Hn (∇) = H desired. Theorem 3.16. Let C ⊆ [n → 2]. Suppose U ⊆ [n] and let τ = U × [2]. Then C shatters U iff β|U |−1,τ (IC ) 6= 0. Proof. C shatters U iff C  U = [U → 2]. The canonical suboplex of C  U is SCU = SC  τ . By the e |U |−1 (SC  τ ) 6= 0 iff H e |U |−1 (SC  τ ; k) 6= 0. By Hochster’s above lemma, SC  τ is complete iff H formula (Proposition 2.12), the dimension of this reduced cohomology is exactly β|U |−1,τ (IC ). The above yields another proof of the dominance of homological dimension over projective dimension. 4 The proof given actually works as is when ∆ is any pure top dimensional subcomplex of a simplicial sphere. 42 3 APPLICATIONS Second proof of Theorem 3.11. By Proposition 3.9, dimh C + 1 = projdim(S/IC? ) = reg(IC ). By Theorem 3.16, the largest shattered set U must satisfy β|U |−1,U ×[2] (IC ) 6= 0, so by the definition of regularity, dimVC C = |U | = |U × [2]| − (|U | − 1) − 1 ≤ reg(IC ) − 1 = dimh C. From the same regularity argument, we obtain a relation between homological dimension and the maximal size of any minimal nonrealizable samples. Theorem 3.17. For any minimal nonrealizable sample f of C, we have |f| ≤ dimh C + 1. Proof. Again, dimh C + 1 = reg(IC ). For each extenture (i.e. minimal nonrealizable sample) f, xgraph f is a minimal generator of IC , so we have β0,graph f (IC ) = 1. Therefore, |f| ≤ reg(IC ) = dimh C + 1. It is easy to check that equality holds for C = monconj, linfun, polythr. Combining Theorem 3.3, Theorem 3.14, and Theorem 3.16, we have the equivalence of three algebraic conditions Corollary 3.18. Let C ⊆ [n → 2] and U ⊆ [n]. The following are equivalent 1. C shatters U . 2. xU 6∈ π∗ IC . 3. ∀f :⊆ U → [2], there is some g :⊆ [n] → [2] extending f such that β|U |−| dom f |,g (C) ≥ 1. 4. β|U |−1,U ×[2] (IC ) 6= 0. The above result together with Corollary 3.13 implies several algebraic conditions on situations in which projective dimension of an ideal is monotone. Here we write down one of them. Corollary 3.19. Let S = k[xu,i : u ∈ [n], i ∈ [2]]. Suppose ideals I and J of S are generated by monomials of the form xΓf , f ∈ [n → 2]. If max{|U | : xU 6∈ π∗ I} = projdim I, then I ⊆ J implies projdim I ≤ projdim J. 3.2 Cohen-Macaulayness We can determine the Betti numbers of dimension 1 of any class of boolean functions. Let C ⊆ [n → 2]. Write C⊇f := {h ∈ C : h ⊇ f}. Then we have the following theorem. Theorem 3.20. The 1-dimensional Betti numbers satisfy ( 1 if |C⊇f | = 2 β1,f (C) = 0 otherwise. More precisely, let {f : f ∈ C} be a set of basis, each with degree Γf , and define M φ: Sf  IC? , φ(f ) = xΓf . f ∈C 43 3.2 Cohen-Macaulayness Let ω f,g = xΓf /xΓ(f ∩g) and ζf,g := ω f,g g − ω g,f f . Then ker φ has minimal generators {ζf,g : C⊇f = {f, g}, f ≺ g}, where ≺ is lexicographic ordering (or any linear order for that matter). We will use the following lemma from [6]. Lemma 3.21 ([6] Lemma 15.1 bis). ker φ is generated by {ζh,h0 : h, h0 ∈ C}. Proof of Theorem 3.20. It’s clear that the latter claim implies the former claim about Betti numbers. We first show that G = {ζf,g : C⊇f = {f, g}, f ≺ g} is a set of generators as claimed. By the lemma above, it suffices to show that ζh,h0 for any two functions h ≺ h0 ∈ C can be expressed as a linear combinations of G. Denote by kf − gk1 the L1 distance n − | dom(f ∩ g)|. We induct on the size of the disagreement p = kh − h0 k1 . When p = 1, ζf,g ∈ G, so there’s nothing to prove. Suppose the induction hypothesis is satisfied for p ≤ q and set p = q +1. Let f = h∩h0 . If C⊇f has size 2 then we are done. So assume |C⊇f | ≥ 3 and let h00 be a function in C⊇f distinct from h or h00 . There must be some u, u0 ∈ [n]\dom f such that h(u) = h00 (u) = ¬h0 (u) and h0 (u0 ) = h00 (u0 ) = ¬h(u0 ). Indeed, if such a u does not exist, then h00 (v) = h0 (v) for all v ∈ [n] \ dom f, and thus h00 = h0 , a contradiction; similarly, if u0 does not exist, we also derive a contradiction. Therefore kh − h00 k1 , kh0 − h00 k ≤ q, and by induction hypothesis, ζh,h00 and ζh0 ,h00 are both expressible as linear combination of G, and thus ζh,h0 = ζh,h00 − ζh00 ,h0 is also expressible this way. This proves that G is a set of generators. For any partial f, if C⊇f = {f, g}, then the degree xΓf strand of φ is the map of vector spaces kωf,g g ⊕ kωg,f f → kxΓf , (ω, ω 0 ) 7→ ω + ω 0 whose kernel is obviously kζf,g . Therefore, G must be a minimal set of generators. Definition 3.22. Let C ⊆ [n → 2] and f, g ∈ C. If Cf ∩g = {f, g}, then we say f and g are neighbors in C, and write f ∼C g, or f ∼ g when C is clear from context. Next, we discuss the conditions under which S/IC and S/IC? could be Cohen-Macaulay. Recall the definition of Cohen-Macaulayness. Definition 3.23 ([11]). A monomial quotient S/I is Cohen-Macaulay if its projective dimension is equal to its codimension codim S/I := min{supp ω : ω ∈ mingen(I ? )}. Cohen-Macaulay rings form a well-studied class of rings in commutative algebra that yields to a rich theory at the intersection of algebraic geometry and combinatorics. The mathematician Melvin Hochster famously wrote “Life is really worth living” in a Cohen-Macaulay ring [9]. By [5, Prop 1.2.13], we have that S/I is Cohen-Macaulay for I squarefree only if every minimal generator of I ? has the same support size. Then the following theorem shows that requiring S/IC? to be Cohen-Macaulay filters out most interesting function classes, including every class considered above except for singleton classes. We first make a definition to be used in the following proof and in later sections. Definition 3.24. Let D ⊆ [n → m]. We say S D is full if for every pair (u, v) ∈ [n] × [m], there is some function h ∈ D with h(u) = v — i.e. {graph h : h ∈ D} = [n] × [m]. Theorem 3.25. Let C ⊆ [n → 2]. The following are equivalent 1. S/IC? is Cohen-Macaulay. 44 3 APPLICATIONS 2. Under the binary relation ∼, C⊇f forms a tree for every PF f :⊆ [n] → [2]. 3. dimh C ≤ 1. Proof. We will show the equivalence of the first two items; the equivalence of the second and third items falls out during the course of the proof. First suppose that C is not full. Then IC has a minimal generator xu,b for some u ∈ [n], b ∈ [2]. If S/IC? is Cohen-Macaulay, then all minimal generators of IC must have the same support size, so for each functional monomial xv,0 xv,1 , either xv,0 or xv,1 is a minimal generator of IC . This means that ? C is a singleton class, and thus is a tree under ∼ trivially. Conversely, S/I{f } is Cohen-Macaulay ? for any f ∈ [n → 2] because the projective dimension of S/I{f } is dimh {f } + 1 = 1 which is the common support size of I{f } (Theorem 2.44). Now assume C is full. Then mingen(IC ) ⊇ FM and min{| supp ω| : ω ∈ mingen(IC )} = 2. Hence S/IC? is Cohen-Macaulay iff the projective dimension of S/IC? is 2 iff the homological dimension of C is 1. This is equivalent to saying that the 1-dimensional cell complex X with vertices f ∈ C and edges f ∼ g minimally resolves IC? with the obvious labeling, which is the same as the condition specified in the theorem. Corollary 3.26. Let C ⊆ [n → 2]. If S/IC? is Cohen-Macaulay, then C has a minimal cellular resolution and has pure Betti numbers which are 0 or 1. Example 3.27. Let o : [n] → [2] be the identically zero function. The class C := deltan ∪ {o} satisfies S/IC? being Cohen-Macaulay. Indeed, f ∼C g iff {f, g} = {δi , o} for some i, so ∼C forms a star graph with o at its center. For each nonempty f :⊆ [n] → [2], if im f = {0}, then C⊇f contains o and thus is again a star graph. If f(i) = 1 for a unique i, then C⊇f = δi , which is a tree trivially. Otherwise, C⊇f = ∅, which is a tree vacuously. It seems unlikely that any class C with Cohen-Macaulay S/IC? is interesting computationally, as Theorem 3.25 and Theorem 3.11 imply the VC dimension of C is at most 1. By the Sauer-Shelah lemma [10], any such class C ⊆ [n → 2] has size at most n + 1. In contrast, the classes C ⊆ [n → 2] with Cohen-Macaulay S/IC form a larger collection, and they all have cellular resolutions. For this reason, we say C is Cohen-Macaulay if S/IC is CohenMacaulay. Definition 3.28. Let n be the n-dimensional cube with vertices [2]n . A cublex (pronounced Q-blex) is a subcomplex of n . n has a natural PF labeling η = η  that labels each vertex V ∈ [2]n with the corresponding function η(V ) : [n] → [2] with η(V )(i) = Vi , and the rest of the PF labels are induced via intersection as in Lemma 2.48. Specifically, each face Fw is associated to a unique PF w :⊆ [n] → [2], such that Fw consists of all vertices V with η(V ) ⊇ w; we label such a Fw with η(Fw ) = w. A cublex X naturally inherits η  , which we call the canonical PF label function of X. Rephrasing Reisner’s Criterion [11, thm 5.53], we obtain the following characterization. Proposition 3.29 (Reisner’s Criterion). C ⊆ [n → 2] is Cohen-Macaulay iff e i−1 (SCf ; k) = 0 for all i 6= n − | dom f|. βi,f (C) = dimk H Theorem 3.30. Let C ⊆ [n → 2]. The following are equivalent. 1. C is Cohen-Macaulay. 45 3.2 Cohen-Macaulayness 2. dimSR C = n. 3. C = {η  (V ) : V ∈ X} for some cublex X such that X⊇f is acyclic for all f :⊆ [n] → [2]. Proof. (1 ⇐⇒ 2). This is immediate after noting that codim S/IC = n. (3 =⇒ 2). X is obviously a minimal cellular resolution of C, and for each f, the face Ff with PF label f, if it exists, has dimension n − | dom f|, so Reisner’s Criterion is satisfied. (2 =⇒ 3). Let X be the cubplex containing all faces Ff such that βi,f (C) 6= 0 for i = n−| dom f|. e i−1 (SCf ; k) 6= 0 iff SCf is the complete (i − 1)-dimensional suboplex This is indeed a complex: H by Lemma 3.15; hence for any g ⊇ f, SCg is the complete (j − 1)-dimensional suboplex, where j = n − | dom g|, implying that βj,g (C) = 1. We prove by induction on poset structure of f :⊆ [n] → [2] under containment that X⊇f is acyclic for all f. The base case of f being total is clear. Suppose our claim is true for all g ⊃ f. If X⊇f is an (n − | dom f|)-dimensional cube, then we are done. Otherwise, X⊇f = [ X⊇g . g⊃f | dom g|=| dom f|+1 By induction hypothesis, each of X⊇g is acyclic, so the homology of X⊇f is isomorphic to the homology of the nerve N of {X⊇g }. We have for any collection F of such g, \ X⊇g 6= ∅ ⇐⇒ ∃f ∈ C ∀g ∈ F[f ⊇ g]. g∈F e • (SCf ; k) = 0 (since Xf is empty), Therefore N is isomorphic to SCf as simplicial complexes. As H X⊇f is acyclic as well. X is obviously minimal since it has unique PF labels, and its vertex labels are exactly C. The minimal cublex cellular resolution of Cohen-Macaulay C constructed in the proof above is called the canonical cublex resolution of C. Corollary 3.31. If C ⊆ [n → 2] is Cohen-Macaulay, then C has a minimal cellular resolution and has pure Betti numbers which are 0 or 1. It should be easy to see that if C is Cohen-Macaulay, then so is the filtered class C  f for any PF f :⊆ [n] → [2]. It turns out this is also true for restrictions of C. Corollary 3.32 (Cohen-Macaulayness is preserved under restriction). If C ⊆ [n → 2] is CohenMacaulay, then so is C  U for any U ⊆ [n]. Its canonical cublex resolution is the projection of the canonical cublex resolution of C onto the subcube Fw of n , where w :⊆ [n] → [2] takes everything outside U to 0. Consequently, β•,f (C  U ) = 0 iff β•,f 0 (C) = 0 for all f 0 ⊇ f extending f to all of [n] \ U . Proof. It suffices to consider the case U = [n − 1] and then apply induction. Fix f :⊆ [n − 1] → [2], and let f0 := f ∪ (n − 1 7→ 0), f1 = f ∪ (n − 1 7→ 1). We wish to show βi,f (C  U ) = 0 for all i 6= n − 1 − | dom f|. We have three cases to consider. 1. β•,f0 (C) = β•,f1 (C) = 0. Certainly, β•,f (C) would also have to be 0 (the existence of the subcube Ff would imply the existence of Ff0 and Ff1 in the canonical cublex resolution of C). By Theorem 2.99, this implies β•,f (C  U ) = 0 as well. 46 3 APPLICATIONS 2. WLOG βi,f0 (C) = I(i = n − | dom f| − 1) and β•,f1 (C) = 0. Again, β•,f (C) = 0 for the same reason. So Theorem 2.99 implies βi,f (C  U ) = I(i = n − | dom f| − 1). 3. βi,f0 (C) = βi,f1 (C) = I(i = n − | dom f| − 1). Then C = [n → 2] and therefore βi,f (C) = I(i = n − | dom f|). Theorem 2.99 yields an exact sequence 0 → kβj+1,f (CU ) → kβj+1,f (C) → kβj,f0 (C)+βj,f1 (C) → kβj,f (CU ) → 0, where j = n − | dom f| − 1. Because C has pure Betti numbers by Corollary 3.31, the only solution to the above sequence is βi,f (C  U ) = I(i = n − | dom f| − 1). This shows by Proposition 3.29 that C  U is Cohen-Macaulay. The second and third statements then follow immediately. Lemma 3.33. If C ⊆ [n → 2] is Cohen-Macaulay, then βi,f (C) = I(i = n − | dom f|) iff f ∈ C for all total f extending f. Proof. βi,f (C) = I(i = n − | dom f|) iff linkf (SC ) is the complete suboplex iff f ∈ C for all total f extending f. Corollary 3.34. If C ⊆ [n → 2] is Cohen-Macaulay, then dimh C = dimVC C. Proof. dimh C is the dimension of the largest cube in the canonical cublex resolution of C, which by the above lemma implies C shatters a set of size dimh C. Therefore dimh C ≤ dimVC C. Equality then follows from Theorem 3.11. Example 3.35. The singleton class {f }, delta ∪ {o} as defined in Example 3.27, and the complete class [n → 2] are all Cohen-Macaulay. However, inspecting Table 1 shows that, for d ≥ 1, none of delta, monconj, conj, linthr, or linfun on d-bit inputs are Cohen-Macaulay, as their StanleyReisner dimensions are strictly greater than 2d . Likewise, polythrkd is not Cohen-Macaulay unless k = d. Consequently, the converse of Corollary 3.34 cannot be true. Example 3.36. We can generalize delta ∪ {o} as follows. Let nb(f )kn be the class of functions on [n] that differs from f ∈ [n → 2] on at most k inputs. Then nb(f )kn is Cohen-Macaulay; its canonical cublex resolution is the cublex with top cells all the k-dimensional cubes incident on f . For example, delta ∪ {o} = nb(o)1n . Finally, we briefly mention the concept of sequential Cohen-Macaulayness, a generalization of Cohen-Macaulayness. Definition 3.37 ([18]). A module M is sequential Cohen-Macaulay if there exists a finite filtration 0 = M0 ⊆ M1 ⊆ · · · ⊆ M r = M of M be graded submodules Mi such that 1. Each quotient Mi /MI−1 is Cohen-Macaulay, and 2. dim(M1 /M0 ) < dim(M2 /M1 ) < · · · < dim(Mr /Mr−1 ), where dim denotes Krull dimension. Sequentially Cohen-Macaulay rings S/I satisfy projdim S/I = max{| supp a| : xa ∈ mingen(I ? )} by a result of [7]. If S/IC is sequentially Cohen-Macaulay, this means it is actually Cohen-Macaulay, since all minimal generators of IC? have the same total degree. Thus what can be called “sequentially Cohen-Macaulay” classes coincide with Cohen-Macaulay classes. 47 3.3 Separation of Classes P link{P } S S Figure 18: The link of suboplex S with respect to vertex P is homeomorphic to the intersection of S with a hyperplane. 3.3 Separation of Classes In this section, unless specificed otherwise, all homologies and cohomologies are taken against k. ? Suppose C, D ⊆ [n → m]. If C ⊆ D, then C ∪ D = D, and IC? + ID? = IC∪D = ID? . In particular, it must be the case that for every i and σ, ? βi,σ (IC? + ID? ) = βi,σ (IC∪D ) = βi,σ (ID? ). Thus C ⊂ D if for some i and f, βi,f (C) 6= βi,f (C ∪ D). The converse is true too, just by virtue of β0,− encoding the elements of each class. By Theorem 3.20, C ⊂ D already implies that β1,− must differ between the two classes. However, we may not expect higher dimensional Betti numbers to certify strict inclusion in general, as the examples in Section 2.3.7 show. This algebraic perspective ties into the topological perspective discussed in the introduction as follows. Consider C ⊆ [2d → {−1, 1}] and a PF f :⊆ [2d ] → {−1, 1}. By Hochster’s dual formula e i−1 (SCf ; k) = dimk H e i−1 (linkgraph f SC ; k). When f = †, this (Proposition 2.11), βi,f (C) = dimk H quantity is the “number of holes of dimension i−1” in the canonical suboplex of C. When graph f = {(u, f(u))} has a singleton domain, linkgraph f SC is the section of SC by a hyperplane. More precisely, d d if we consider SC as embedded the natural way in S12 −1 = {z ∈ R2 : kzk1 = 1} (identifying each coordinate with a v ∈ [2d ] ∼ = [2]d ), linkgraph f SC is homeomorphic to SC ∩{z : zu = f(u)/2}. Figure 18 illustrates this. For general f, we have the homeomorphism linkgraph f SC ∼ = SC ∩ {z : zu = f(u)/2, ∀u ∈ dom f}. Thus comparing the Betti numbers of D and C ∪ D is the same as comparing “the number of holes” of SD and SC∪D and their corresponding sections. If PF-labeled complex (XC , µC ) resolves C and PF-labeled complex (XD , µD ) resolves D, then the join (XC ? XD , µC ? µD ) resolves C ∪ D by Proposition 2.89. The Betti numbers can then be computed by e i−1 ((XC ? XD )⊃f ; k) βi,f (C ∪ D) = dimk H via Proposition 2.19. Here are some simple examples illustrating this strategy. Theorem 3.38. Let d ≥ 2. Let I1 ∈ [2d → 2] be the indicator function u 7→ I(u = 1 = 1 · · · 1 ∈ [2]d ). Consider the partial linear functional g : 0 → 0, 1 → 1. Then βi,g (linfun2d ∪ {I1 }) = 0 for all i. 48 3 APPLICATIONS The proof below is in essence the same as the proof of I1 6∈ linfun2d given in the introduction, but uses the theory we have developed so far. The application of the Nerve Lemma there is here absorbed into the Stanley-Reisner and cellular resolution machineries. Proof. Let (X, µ) be the flag resolution of linfun2d and • be the one point resolution of {I1 }. Then X ? • is the cone over X, with labels µ0 (F ? •) = µ(F ) ∩ I1 and µ0 (F ) = µ(F ) for cells F in X. Consider Z := (X ? •)⊃g . Every cell F of X in Z has PF label a linear functional on a linear subspace of Fd2 strictly containing V := {0, 1}. As such, µ(F ) ∩ I1 strictly extends g, because µ(F ) sends something to 0 outside of V. This means Z is a cone over X⊃g , and thus is acyclic. Therefore βi,g (linfun2d ∪ {I1 }) = 0 for all i. But βd−1,g (linfun2d ) is nonzero, so we obtain the following corollary. Corollary 3.39. I1 6∈ linfun2d for d ≥ 2. Theorem 3.38 says the following geometrically: the canonical suboplex of linfun2d  g (a complex d of dimension 22 − 2) has holes in dimension d − 1, but these holes are simultaneously covered up when we add I1 to linfun2d . Theorem 3.40. Let parityd be the parity function on d bits. Then βi,† (polythrkd ∪{parityd }) = 0 for all i if k < d. ∼ {1, −1}, a 7→ (−1)a , Let us work over {−1, 1} instead of {0, 1}, under the bijection {0, 1} = so that parityd (u0 , . . . , ud−1 ) = u0 · · · ud−1 for u ∈ {−1, 1}d and polythrkd consists of sgn(p) for polynomials p with degree at most k not taking 0 on any point in {−1, 1}d . Proof. Fix k < d. Let (X, µ) denote the ball resolution of polythrkd and • be the one point resolution of {f }. Then X ? • is the cone over X, with labels µ0 (F ? •) = µ(F ) ∩ f and µ0 (F ) = µ(F ) for cells F in X. Consider Z := (X ? •)⊃† . Every PF label f :⊆ {−1, 1}d → {−1, 1} of X intersects parityd nontrivially if f 6= †. Otherwise, suppose p is a polynomial function such that p(u) > 0 ⇐⇒ f(u) = 1, p(u) < 0 ⇐⇒ f(u) = −1, and p(u) = 0 ⇐⇒ uQ6∈ dom f. Then by discrete Fourier transform 5 , the coeffient of p for the monomial parityd (u) = d−1 i=0 ui is X p(a)parityd (a) < 0 a∈{−1,1}d because whenever p(a) is nonzero, its sign is the opposite of parityd (a). This contradicts k < d. Thus in particular, the PF label of every cell of X except for the top cell (with PF label †) intersects parityd nontrivially. Therefore Z is a cone and thus β•,† (polythrkd ∪ {parityd }) = 0.
P But βe,† (polythrkd ) = 1, where e = kj=0 dj is the homological dimension of polythrkd . So we recover the following result by Minsky and Papert. Corollary 3.41 ([12]). parityd 6∈ polythrkd unless k = d. From the analysis below, we will see in fact that adding parityd to polythrkd causes changes to Betti numbers in every dimension up to dimVC polythrkd = dimh polythrkd , so in some sense parityd is maximally homologically separated from polythrkd . This “maximality” turns out to be equivalent to the lack of weak representation Corollary 3.46. 5 See the opening chapter of [15] for a good introduction to the concepts of Fourier analysis of boolean functions. 49 3.3 Separation of Classes By Lemma 2.90, the “differences” between the Betti numbers of C ∪ D and those of C and of D are given by the homologies of (XC × XD , µC × µD )⊃f . Suppose C consists of a single function f . Then XC is a single point with exponent label Γf . (XC × XD , µC × µD ) is thus isomorphic to XD as complexes, but the exponent label of each nonempty cell F ∈ XC × XD isomorphic to cell F 0 ∈ XD is now lcm(λD (F 0 ), Γf ), and the PF label of F is µD (F 0 ) ∩ f ; the empty cell ∅ ∈ XC × XD has the exponent label 0. We denote this labeled complex by (XD )f . Notice that (XD )f is a (generally nonminimal) cellular resolution of the PF class Df := D{f }, because (XD )f ⊇f = (XD )⊇f whenever f ⊆ f and empty otherwise, and therefore acyclic. So the (dimensions of) homologies of (XC × XD )⊃f are just the Betti numbers of Df . This is confirmed by Lemma 2.92. Another perspective is that SCD is the intersection SC ∩ SD , so by Mayer-Vietoris, ? ICD gives the “difference” in Betti numbers between β•,− (C) + β•,− (D) and β•,− (C ∪ D). ID?f determines the membership of f through several equivalent algebraic conditions. Lemma 3.42. Let D ⊆ [n → m] be a full class (see Definition 3.24). Then the following are equivalent: 1. f ∈ D 2. ID?f is principally generated by xΓf 3. ID?f is principal 4. βi,f (Df ) = 1 for exactly one partial f when i = 0 and equals 0 for all other i. 5. βi,f (Df ) = 0 for all f and all i ≥ 1. 6. βi,f (Df ) = 0 for all f 6= f and all i ≥ 1. Proof. (1 =⇒ 2 =⇒ 3) If f ∈ D, then ID?f is principally generated by xΓf . (3 =⇒ 2 =⇒ 1) If ID?f is principal, then it’s generated by xΓg for some partial function g. This implies that h∩f ⊆ g =⇒ graph h ⊆ Γf ∪graph g, ∀h ∈ D. But taking the union over all h ∈ D contradicts our assumption on D unless g = f . Thus there is some h ∈ D with h ∩ f = f =⇒ h = f . (3) ⇐⇒ 4) This should be obvious. (4 ⇐⇒ 5 ⇐⇒ 6) The forward directions are obvious. Conversely, if ID?f has more than one minimal generator, then its first syzygy is nonzero and has degrees  Γf , implying the negation of Item 5 and Item 6. Thus ID?f by itself already determines membership of f ∈ D. It also yields information on the Betti numbers of C ∪ D via Lemma 2.92. Thus in what follows, we study ID?f in order to gain insight into both of the membership question and the Betti number question. Let us consider the specific case of D = linthrU , with minimal cocellular resolution coBallU = (Y, µ). Then linthrU f has minimal cocellular resolution (Y, µf ), where we relabel cells F of Y by µf (F ) = µ(F ) ∩ f , so that, for example, the empty cell still has PF label the empty function. ~ forms a set of orthogonal basis for Choose U to be a set of n points such that the vectorization U Rn . Then linthrU = [U → 2], and Y is homeomorphic to the unit sphere S n−1 as a topological space and is isomorphic to the complete (n − 1)-dimensional suboplex as a simplicial complex. It has 2n top cells 4g , one forTeach function g ∈ [U → 2]; in general, it has a cell 4f for each PF f :⊆ U → 2, satisfying 4f = f ⊇f 4f . Γf Let us verify that βi,f (linthrf U ) equals βi,Γf (hx i) = I(i = 0 & f = f ) by Lemma 2.30. For any f ⊆ f , define f ♦ f to be the total function f ♦ f : u 7→ f (u) ∀u ∈ dom f, 50 u 7→ ¬f (u) ∀u 6∈ dom f. 3 APPLICATIONS ∂ f 4f ♦ f 4f 4f ♦ f 4¬f Figure 19: The bold segments form the partial boundary ∂ f 4f ♦ f . In particular, this partial boundary contains three vertices. It is exactly the part of 4f ♦ f visible to a spider in the interior of 4¬f , if light travels along the sphere. Define the (f, f)-star F(f, f) to be the collection of open cells 4̊g with PF label f ⊆ g ⊆ f ♦ f. This is exactly the collection of open cells realized by the cellular pair (4f ♦ f , ∂ f 4f ♦ f ), where ∂ f 4f ♦ f denotes the partial boundary of 4f ♦ f that is the union of the closed cells with PF labels (f ♦ f) \ (i 7→ f(i)) for each i ∈ dom f. In particular, F(f, f ) is realized by (4f , ∂4f ), and F(f, †) is realized by (4¬f , {}) (where {} is the void complex). In the following we suppress the subscript to write (4, ∂ f 4) for the sake of clarity. When f 6= †, f , ∂ f 4 is the union of faces intersecting 4¬f ; intuitively, they form the subcomplex of faces directly visible from an observer in the interior of 4¬f . This is illustrated in Figure 19. Then the part of (YU , µf U ) with PF label f is exactly the (f, f)-star. If f 6= f , the closed top cells in ∂ f 4 all intersect at the closed cell with PF label f ♦ f \ f = ¬(f \ f), and thus their union ∂ f 4 is contractible. This implies via the relative cohomology sequence e j (∂ f 4) ← H e j (4) ← H j (4, ∂ f 4) ← H e j−1 (∂ f 4) ← · · · ··· ← H e j (4) = dimk H j (4, ∂ f 4) = βn−1−j,f (linthrf ). If f = f , then ∂ f 4 = ∂4, so that 0 = dimk H U f e j (4/∂) ∼ H j (4, ∂ f 4) ∼ =H = kI(j=n−1) . This yields βk,f (linthrU ) = I(k = 0). The analysis of the Betti numbers of any thresholded linear class thr L is now much easier given the above. As discussed in Section 2.3.5, the cocellular resolution (Z, µZ ) of thr L is just the intersection of coBallU = (Y, µ) with L, with the label of an intersection equal to the label of the original cell, i.e. Z = Y ∩ L, µZ (F ∩ L) = µ(F ). Similarly, the cocellular resolution of (thr L)f f f is just (Z, µf Z ) with Z = Y ∩ L, µZ (F ∩ L) = µ (F ). If L is not contained in any coordinate hyperplane of Y , then thr L is full. By Lemma 3.42, f ∈ thr L iff βi,f (thr Lf ) = 0 for all i ≥ 1. This is equivalent by Lemma 2.30 to the statement that for all f, the degree Γf part of (Z, µf Z ), FL (f, f) := F(f, f) ∩ L, has the homological constraint H dim Z−i (FL (f, f), ∂FL (f, f)) = H dim Z−i (4f ♦ f ∩ L, ∂ f 4f ♦ f ∩ L) = 0, ∀i ≥ 0. But of course, f ∈ thr L iff L ∩ 4̊f 6= ∅. We therefore have discovered half of a remarkable theorem. Theorem 3.43 (Homological Farkas). Let L be a vector subspace of dimension l ≥ 2 of Rn not contained in any coordinate hyperplane, let P denote the positive cone {v ∈ Rn : v > 0}, and let 1 : [n] → {−1, 1}, j 7→ 1. For any g : [n] → {−1, 1}, define Ξ(g) to be the topological space represented by the complex ∂FL (1, 1 ∩ g). Then the following are equivalent:6 6 Our proof will work for all fields k of any characteristic, so the cohomologies can actually be taken against Z. 51 3.3 Separation of Classes 1. L intersects P. e • (Ξ(g); k) = 0 as long as 4g ∩ L 6= ∅. 2. For all g 6= 1, ¬1, H This theorem gives homological certificates for the non-intersection of a vector subspace with the positive cone, similar to how Farkas’ lemma [22] gives linear certificates for the same thing. Let’s give some intuition for why it should be true. As mentioned before, ∂F(1, 1 ∩ g) is essentially the part of 41 ♦(1∩g) = 4g visible to an observer Tom in 4̊¬1 , if we make light travel along the surface of the sphere, or say we project everything into an affine hyperplane. Since the simplex is convex, the image Tom sees is also convex. If L indeed intersects 4̊1 (equivalently 4̊¬1 ), then for Ξ(g) he sees some affine space intersecting a convex body, and hence a convex body in itself. As Tom stands in the interior, he sees everything (i.e. his vision is bijective with the actual points), and the obvious contraction he sees will indeed contract Ξ(g) to a point, and Ξ(g) has trivial cohomology. Conversely, this theorem says that if Tom is outside of 4̊1 (equivalently 4̊¬1 ), then he will be able to see the nonconvexity of ∂F(1, 1 ∩ g) for some g, such that its intersection with an affine space is no longer contractible to a single point. Proof of 1 =⇒ 2. Note that Ξ(g) is a complex of dimension at most l − 2, so it suffices to prove the following equivalent statement: e l−2−i (Ξ(g); k) = 0 for all i ≥ 0 as long as 4g ∩ L 6= ∅. For all g 6= 1, ¬1, H L intersects P iff L intersects 4̊1 iff 1 ∈ thr L. By Lemma 3.42, this implies βi,f (thr Lf ) = 0 for all f 6= 1, † and i ≥ 0. As we observed above, this means H l−1−i (FL (1, f), ∂FL (1, f)) = 0, ∀i ≥ 0. Write A = FL (1, f) and B = ∂FL (1, f) for the sake of brevity. Suppose A = 41 ♦ f ∩L is nonempty. Then for f 6= † as we have assumed, both A and B contain the empty cell, and therefore we have a relative cohomology long exact sequence with reduced cohomologies, e l−2−i (B) ← H e l−2−i (A) ← H l−2−i (A, B) ← · · · · · · ← H l−1−i (A, B) ← H e • (A) = 0, we have Because H e l−2−i (B), ∀i. H l−1−i (A, B) ∼ =H This yields the desired result after observing that 1 ♦ f 6= 1, ¬1 iff f 6= 1, †. Note that we cannot replace L with any general openly convex cone, because we have used Lemma 3.42 crucially, which requires thr L to be full, which can happen only if L is a vector subspace, by Lemma 2.72. The reverse direction is actually quite similar, using the equivalences of Lemma 3.42. But straightforwardly applying the lemma would yield a condition on when g = ¬1 as well which boils down to L ∩ 4¬1 6= ∅, that significantly weakens the strength of the theorem.7 To get rid of this condition, we need to dig deeper into the structures of Betti numbers of thr L. Theorem 3.44. Suppose L is linear of dimension l, thr L is a full class, and f 6∈ thr L. Let g be such that σg is the unique covector of L of the largest support with g ⊆ f (where we let g = 0 if no such covector exists). We say g is the projection of f to thr L, and write g = Π(f, L). Then the following hold: 7 Note that the condition says L intersects the closed cell 4¬1 , not necessarily the interior, so it doesn’t completely trivialize it. 52 3 APPLICATIONS 1. βi,g (thr Lf ) = I(i = l − 1 − rank σg). (Here rank denotes the rank wrt matroid of L as defined in Section 2.3.5) 2. For any h 6⊇ g, β•,h (thr Lf ) = 0. 3. For any PF r with domain disjoint from dom g, βi,r∪g (thr Lf ) = βi,r (thr Lf  ([n] \ dom g)). Note that such a σg would indeed be unique, since any two covectors with this property are consistent and thus their union gives a covector with weakly bigger support. Proof. (Item 1) The assumption on g is exactly that L intersects 4f at 4̊g ⊆ 4f and g is the S maximal such PF. Then F(f, g) ∩ L = f ♦ g⊇h⊇g 4̊h ∩ L = 4̊g ∩ L. Therefore βi,g (thr Lf ) = e l−1−i ((4g ∩ L)/∂) = I(l − 1 − i = dim(4g ∩ L)) (note that when 4g ∩ L is a point (resp. the H empty cell), the boundary is the empty cell (resp. the empty space), so that this equality still holds in those cases). But dim(4g ∩ L) is rank σg. So the Betti number is I(i = l − 1 − rank σg). ? (Item 2) We show that Ithr is generated by monomials of the form xΓf for f ⊇ g. It suffices Lf to demonstrate that for any function h ∈ thr L, the function h o g defined by ( g(u) if u ∈ dom g h o g(u) := h(u) otherwise. is also in thr L, as f ∩ (h o g) ⊇ f ∩ h. Let ϕ ∈ L be a function ϕ : U → R such that sgn(ϕ) = σg. If ψ ∈ L is any function, then for sufficiently small  > 0, sgn(ψ + ϕ) = sgn(ψ) o sgn(ϕ) = sgn(ψ) o g. Since L is linear, ψ + ϕ ∈ L, and we have the desired result. ? (Item 3) As shown above, the minimal generators of Ithr are all divisible by xΓg . The result Lf then follows from Lemma 2.8. Corollary 3.45. Suppose L is linear of dimension l ≥ 2 and thr L is a full class. Then f ∈ thr L iff βi,f (thr Lf ) = 0 for all f 6= f, † and all i ≥ 1. If l ≥ 1, then we also have f ∈ thr L iff βi,f (thr Lf ) = 0 for all f 6= f, † and all i ≥ 0. Proof. We show the first statement. The second statement is similar. The forward direction follows from Lemma 3.42. If β•,† (thr Lf ) = 0, then the same lemma also proves the backward direction. So assume otherwise, and in particular, f 6∈ thr L. By Theorem 3.44, it has to be the case that βi,† (thr Lf ) = I(i = l − 1 − (−1)) = I(i = l) since rank 0 = −1. Consequently, βl−1,f (thr Lf ) 6= 0 for some f ⊃ †. If l ≥ 2, then this contradicts the right side of the equivalence, as desired. We can now finish the proof of Theorem 3.43. Proof of 2 =⇒ 1 in Theorem 3.43. Assume l ≥ 2. (2) says exactly that βj,f (thr Lf ) = 0 for all f 6= 1, † and all j ≥ 1. So by Corollary 3.45, 1 ∈ thr L and therefore thr L intersects P. From the literature of threshold functions in theoretical computer science, we say a real function ϕ : U → R on a finite set U weakly represents a function f : U → {−1, 1} if ϕ(u) > 0 ⇐⇒ f (u) = 1 and ϕ(u) < 0 ⇐⇒ f (u) = −1, but we don’t care what happens when ϕ(u) = 0. In these terms, we have another immediate corollary of Theorem 3.44. f Corollary 3.46. A function f is weakly representable by polythrkd iff βi,† ((polythrkd ) 53 ) = 0, ∀i. 3.3 Separation of Classes This is confirmed by Theorem 3.40. By Lemma 2.92, this result means that f is weakly representable by polythrkd iff adding f to polythrkd did not change the homology of Spolythrk . d Remark 3.47. Item 3 of Theorem 3.44 reduces the characterization of Betti numbers of thr Lf to the case when f is not “weakly representable” by thr L. The following theorem says that as we perturb a function f 6∈ thr L by a single input u to obtain f u , a nonzero Betti number βi,f of “codimension 1” of thr Lf remains a nonzero Betti number of u “codimension 1” of thr Lf if we truncate f. Theorem 3.48 (Codimension 1 Stability). Suppose L is linear of dimension l ≥ 2 and thr L is a full class. Let f be a function not in thr L and write g = Π(f, L). Assume rank σg = s (so that βl−s−1,g (thr Lf ) = 1) and let f :⊆ [n] → [2] be such that βl−s−2,f (thr Lf ) 6= 0. Then βl−s−2,f (thr Lf ) = 1. Furthermore, if | dom(f \ g)| > 1 and u ∈ dom(f \ g), set f 0 := f \ (u 7→ f (u)). Then we have u βl−s−2,f 0 (thr Lf ) = 1 where ( f (v) f u (v) := ¬f (v) if v 6= u if v = u. Proof. By Theorem 3.44, it suffices to show this for g = †; then s = −1. Recall FL (f, f) = F(f, f)∩L. By Lemma 2.30, βl−1,f (thr Lf ) = dimk H 0 (FL (f, f), ∂FL (f, f)). If ∂FL (f, f) contains more than the empty cell, then the RHS is the zeroth reduced cohomology of the connected space FL (f, f)/∂, which is 0, a contradiction. Therefore ∂FL (f, f) = {∅}, i.e. as geometric realizations, L does not intersect ∂F(f, f). Consequently, βl−1,f (thr Lf ) = dimk H 0 (4f ♦ f , {∅}) = dimk H 0 (4f ♦ f ) = 1. Now f u ♦ f 0 = f ♦ f, and ∂ f 0 4f ♦ f ⊂ ∂ f 4f ♦ f since f 0 ⊂ f. Therefore ∂FL (f u , f 0 ) ⊆ ∂FL (f, f) u also does not intersect L. So βl−1,f 0 (thr Lf ) = dimk H 0 (4f ♦ f , {∅}) = dimk H 0 (4f ♦ f ) = 1. Below we give some examples of the computation of Betti numbers of thr Lf . Theorem 3.49. Let f := parityd ∈ [{−1, 1}d → {−1, 1}] and C := linthrd ⊆ [{−1, 1}d → {−1, 1}]. Then βi,f ∩1 (Cf ) = (2d−1 − 1)I(i = 1). Proof. Let (Y, µ) be the coball resolution of linthrd . Consider the cell F1 of Y with PF label 1. It has 2d facets since ( r 7→ −1 if r = s Is := r 7→ 1 if r 6= s is a linear threshold function (it “cuts” out a corner of the d-cube), so that each facet of F1 is the cell Gs := F1∩Is with PF label ( r 7→ 1 if r = s 1 ∩ Is = undefined otherwise. If for s, s0 ∈ {−1, 1}d , f (s) = f (s0 ), then Gs and Gs0 do not share a codimension 2 face (do not share a facet of their own). (If they do, then ( r 7→ + if r = 6 s, s0 r 7→ 0 else 54 3 APPLICATIONS is a covector of the d-cube. But this means that (s, s0 ) is an edge of the d-cube, implying that f (s) 6= f (s0 )). S Now note that ∂ f ∩1 F1 = {Gs : f (s) = 1}. For each Gs 6⊆ ∂ f ∩1 F1 , we have ∂Gs ⊆ ∂ f ∩1 F1 by Fn/2 the above reasoning. Let α := d + 1 = dimh C and n = 2d . Therefore, ∂ f ∩1 R1 ∼ = S α−2 \ i=1 Dn−2 , the (α − 2)-sphere with n/2 holes. So ( n/2−1 if m = α − 3 m e (∂ f ∩1 F1 ) = Z H 0 otherwise Hence e α−1−i (F1 , ∂ f ∩1 F1 ; k) βi,f ∩1 (C) = dimk H e α−2−i (∂ f ∩1 F1 ) = rank H = (n/2 − 1)I(α − 2 − i = α − 3) = (2d−1 − 1)I(i = 1). ⊆ [{−1, 1}d → Theorem 3.50. Let f := parityd ∈ [{−1, 1}d → {−1, 1}] and C := polythrd−1 d f d−1 {−1, 1}]. Then βi,f ∩1 (C ) = (2 − 1)I(i = 1). Proof. Let (Y, µ) be the coball resolution of linthrd . Consider the cell F1 of Y with PF label 1. It has 2d facets since ( r 7→ −1 if r = s Is := r 7→ 1 if r 6= s is a linear threshold function (it “cuts” out a corner of the d-cube), so that each facet of F1 is the cell Gs := F1∩Is with PF label ( r 7→ 1 if r = s 1 ∩ Is = undefined otherwise. Note that a function g : {−1, 1}d → {−1, 0, 1} is the sign function of a polynomial p with degree d − 1 iff im(gf ) ⊇ {−1, 1} (i.e. g hits both 1 and −1). Indeed, by Fourier Transform, the degree constraint on p is equivalent to hp, f i = X p(u)f (u) = 0. u∈{−1,1}d If im gf = im sgn(p)f does not contain −1, then this quantity is positive as long as p 6= 0, a contradiction. So suppose im gf ⊇ {−1, 1}. Set mp := #{u : g(u)f (u) = 1} and mn := #{u : g(u) g(u)f (u) = −1}. Define the polynomial p by p(u) = g(u) mp if g(u)f (u) = 1 and p(u) = mn if g(u)f (u) = −1. Then hp, f i = 0 and sgn p = g by construction, as desired. As ∂ f ∩1 F1 is the complex with the facets F := {Gs : f (s) = 1}, to find its homology it suffices to consider the nerve of the facet cover. For a function g : {−1, 1}d → {−1, 0, 1}, write ḡ for the partial function g  g −1 ({−1, 1}) (essentially, we are marking as undefined all inputs that g send to 0). But by the above, any proper subset G ⊂ F must have nontrivial intersection (which is a cell 55 3.4 The Maximal Principle for Threshold Functions T with PF label ḡ for some g with im g ⊇ {−1, 1}), while F must have PF label a subfunction g of h̄ for ( 1 if f (u) = −1 h(u) = 0 otherwise. T Again, by the last paragraph, this implies that F = ∅. Therefore, in summary, the nerve NF is e j (∂ f ∩1 F1 ) ∼ e j (NF ) = the boundary of a (n/2 − 1)-dimensional simplex, where n = 2d , so that H =H d−1 I(j = n/2 − 2) · Z. Let α = dimh polythrd = 2d − 1. Then e α−1−i (F1 , ∂ 1∩f F1 ; k) βi,1∩f (C) = dimk H e α−2−i (∂ 1∩f F1 ) = rank H = I(α − 2 − i = n/2 − 2) = I(i = 2d−1 − 1) as desired. 3.4 The Maximal Principle for Threshold Functions By looking at the 0th Betti numbers of thr Lf , we can obtain a “maximal principle” for thr L. Theorem 3.51. Suppose there exists a function g ∈ thr L such that • g 6= f and, • for each h ∈ thr L that differs from g on exactly one input u, we have g(u) = f (u) = ¬h(u). Then βi,f ∩g (thr Lf ) = I(i = 0) and f 6∈ thr L. Conversely, any function g ∈ thr L satisfying βi,f ∩g (thr Lf ) = I(i = 0) also satisfies condition (3.51). Informally, Theorem 3.51 says that if we look at the partial order on thr L induced by the mapping from thr L to the class of partial functions, sending g to g ∩ f , then, assuming f is in thr L, any function g that is a “local maximum” in thr L under this partial order must also be a global maximum and equal to f . We shall formally call any function g ∈ thr L satisfying condition 3.51 a local maximum with respect to f . Proof. Let coBall = (Y, µ) be the minimal cocellular resolution of thr L. Let Fg denote the face of Y with label g. Each facet of Fg has the label g ∩ h for some h differing from g on exactly one input. Condition (3.51) thus says that ∂ f ∩g Fg = ∂Fg . Therefore, if l = dim L, e l−1−i (Fg /∂ f ∩g ; k) βi,f ∩g (thr Lf ) = dimk H e l−1−i (Fg /∂; k) = dimk H = I(l − 1 − i = dim Fg ) = I(i = 0) This shows that f 6∈ thr L as desired. For the converse statement, we only need to note that the Betti number condition implies ∂ f ∩g Fg = ∂Fg , by reversing the above argument. For any f : {−1, 1}d → {−1, 1}, define thrdeg f to be the minimal degree of any polynomial P with 0 6∈ P ({−1, 1}d ) and sgn(P ) = f . The maximal principle enables us to compute thrdeg f for any symmetric f (a result that appeared in [3]). 56 3 APPLICATIONS Theorem 3.52. Suppose f : {−1, 1}d → {−1, 1} is symmetric, i.e. f (u) = f (π · u) for any permutation π. Let r be the number of times f changes signs. Then thrdeg f = r. P Proof. To show thrdeg f ≤ r: Let s(u) := i (1 − ui )/2. Because f is symmetric, it is a function of s, say f¯(s(u)) = f (u) 8 . Suppose WLOG f¯(0) >Q 0 and f¯ changes signs between s and s + 1 for s = t1 , . . . , tr . Then define the polynomial Q(s) := ri=1 (ti + 12 − s). One can immediately see that sgn Q(s) = f¯(s) = f (u). Therefore thrdeg f ≤ r. Q 1 To show thrdeg f ≥ r: Let k = r − 1 and consider the polynomial Q0 (s) = r−1 i=1 (ti + 2 − s) and 0 k its sign function ḡ(s) = sgn Q (s) ∈ polythrd . We show that g(u) = ḡ(s(u)) a local maximum. Since ḡ(s) = f¯(s) for all s ∈ [0, tr ], it suffices to show that for any v with s(v) > tr , the function ( g(u) if u 6= v g v (u) := ¬g(u) if u = v. is not in polythrkd . WLOG, assume v = (−1, . . . , −1, 1, . . . , 1) with σ := s(v) −1’s in front. For the sake of contradiction, suppose there exists degree k P polynomial P with sgn P = g v . Obtain through symmetrization the polynomial R(z1 , . . . , zσ ) :=P π∈Sσ P (π · z, 1, . . . , 1). R is a symmetric polynomial, so expressable as a univariate R0 (q) in q := j (1 − zj )/2 ∈ [0, σ] on the Boolean cube. Furthermore, sgn R0 (q) = ḡ(q) for all q 6= σ, and sgn R0 (σ) = −ḡ(σ). Thus R0 changes sign k + 1 times on [0, σ] but has degree at most k, a contradiction. This yields the desired result. The proof above can be extended to give information on the zeroth Betti numbers of polythrkd  {f }. Suppose f is again symmetric, and as in the proof above, r := thrdeg f X s(u) := (1 − ui )/2 i f¯(s(u)) = f (u) Q(s) := f¯(0) r Y 1 (ti + − s) 2 i=1 where f¯ changes signs between s and s + 1 for s = t1 , . . . , tr . Q Theorem 3.53. Let k < r and a < b ∈ [r] be such that b − a = k − 1. Set Q0 (s) = f¯(ta ) bi=a (ti + 1 0 k 2 − s) and g(u) := ḡ(s(u)) := sgn Q (s(u)). Then βi,f ∩g (polythrd  {f }) = I(i = 0). Proof. We prove the equivalent statement (by Theorem 3.51) that g is a local maximum. Since ḡ(s) = f¯(s) for s ∈ [ta , tb + 1], we just need to show that for any v with s(v) 6∈ [ta , tb + 1], the function ( g(u) if u 6= v g v (u) := ¬g(u) if u = v. is not in polythrkd . If s(v) > tb + 1, then WLOG assume v = (−1, . . . , −1, 1, . . . , 1) with σ := s(v) −1’s in front. For the sake of contradiction, suppose there exists degree k polynomial P with sgn P = g v . Obtain 8 f can be expressed as a polynomial in {−1, 1}d , and by the fundamental theorem of symmetric polynomials, f d is a polynomial in the elementary symmetric Ppolynomials. But with respect to the Boolean cube {−1, 1} , all higher symmetric polynomials are polynomials in i ui , so in fact f is a univariate polynomial in s. 57 3.5 Homological Farkas P through symmetrization the polynomial R(z1 , . . . , zσ ) :=P π∈Sσ P (π · z, 1, . . . , 1). R is a symmetric polynomial, so expressable as a univariate R0 (q) in q := j (1 − zj )/2 ∈ [0, σ] on the Boolean cube. Furthermore, sgn R0 (q) = ḡ(q) on q ∈ [0, σ − 1], and sgn R0 (σ) = −ḡ(σ). Thus R0 changes sign k + 1 times on [0, σ] but has degree at most k, a contradiction. If s(v) < ta , then WLOG assume v = (1, . . . , 1, −1, . . . , −1) with σ := s(v) −1’s in the back. For the sake of contradiction, suppose there exists degree kPpolynomial P with sgn P = g v . Obtain through symmetrization the polynomial R(z1 , . . . , zσ ) := π∈Sσ P P(π · z, −1, . . . , −1). R is a symmetric polynomial, so expressable as a univariate R0 (q) in q := j (1 − zj )/2 ∈ [0, d − σ] on the Boolean cube. Furthermore, sgn R0 (q) = ḡ(q + σ) on q ∈ [1, σ − 1], and sgn R0 (0) = −ḡ(σ). Thus R0 changes sign k + 1 times on [0, d − σ] but has degree at most k, a contradiction. 3.5 Homological Farkas Theorem 3.43 essentially recovers Theorem 1.7, after we define Λ(g) to be ∂FL (¬1, ¬1 ∩ g), and utilize the symmetry 1 ∈ thr L ⇐⇒ ¬1 ∈ thr L. Then Λ(g) indeed coincides with the union of facets of 4g whose linear spans separates 4g and 41 . We can generalize the homological Farkas’ lemma to arbitrary linear hyperplane arrangements. Let H = {Hi }ni=1 be a collection of hyperplanes in Rk , and {wi }i be a collection of row matrices such that Hi = {x ∈ Rk : wi x = 0}. Set W to be the matrix with rows wi . For b ∈ {−, +}n , define Rb := {x ∈ Rk : sgn(W x) = b}. Thus R+ = {x ∈ Rk : W x > 0}. Suppose W has full rank (i.e. the normals wi to Hi span the whole space Rk ), so that W is an embedding Rk  Rn . Each region Rb is the preimage of Pb , the cone in Rn with sign b. Therefore, Rb is linearly isomorphic to im W ∩ Pb , via W . Let L ⊆ Rk be a linear subspace of dimension l. Then L ∩ R+ 6= ∅ ⇐⇒ W (L) ∩ P+ 6= ∅ e l−2−i (W (L) ∩ Λ(b)) = 0∀i ≥ 0] ⇐⇒ ∀b 6= +, −, [W (L) ∩ Λ(b) 6= ∅ =⇒ H e l−2−i (L ∩ W −1 Λ(b)) = 0∀i ≥ 0] ⇐⇒ ∀b 6= +, −, [L ∩ W −1 Λ(b) 6= ∅ =⇒ H This inspires the following definition. Definition 3.54. Let H = {Hi }ni=1 and W be as above, with W having full rank. Suppose b ∈ {−, +}n . Then ΛH (b) is defined as the union of the facets of Rb ∩ S k−1 whose linear spans separate Rb and R+ . One can immediately see that ΛH (b) = W −1 Λ(b). In this terminology, we have shown the following Corollary 3.55. Let H = {Hi }ni=1 be a collection of linear hyperplanes in Rk whose normals span Rk (This is also called an essential hyperplane arrangement.). Suppose L ⊆ Rk is a linear subspace of dimension l. Then either • L ∩ R+ 6= ∅, or e l−2−i (L ∩ ΛH (b)) 6= 0 for some i ≥ 0, • there is some b 6= +, −, such that L ∩ ΛH (b) 6= ∅ and H but not both. 58 3 APPLICATIONS Rb ΛA (b) A sphere at infinity ¬A Λ¬A (¬b) ¬R¬b Figure 20: Illustration of the symbols introduced so far. This corollary can be adapted to the affine case as follows. Let A = {Ai }ni=1 be an essential oriented affine hyperplane arrangement in Rk−1 . The hyperplanes A divide Rk−1 into open, signed regions Rb , b ∈ {−, +}n such that Rb lies on the bi side of Ai . We can define ΛA (b) as above, as the union of facets F of Rb such that Rb falls on the negative side of the affine hull of F , along with their closures in the “sphere at infinity.” ~ = {A ~ i }n to be the Let Hb := {(x, b) : x ∈ Rk−1 }. Treating Rk−1 ,→ Rk as H1 , define A i=1 oriented linear hyperplanes vectorizing Ai in Rk . Vectorization produces from each Rb two cones ~ b, R ~ ¬b ⊆ Rk , defined by R ~ b := {v ∈ Rk : ∃c > 0, cv ∈ Rb } R ~ ¬b := {v ∈ Rk : ∃c < 0, cv ∈ Rb }. R Define ¬A as the hyperplane arrangement with the same hyperplanes as A but with orientation reversed. Let ¬Rb denote the region in ¬A with sign b. Set Λ¬A (b) analogously for ¬A, as the union of facets F of ¬Rb such that ¬Rb falls on the negative side of the affine hull of F , along their closures in the “sphere at infinity.” Thus the natural linear identity between ¬A and A identifies ¬R¬b with Rb , and Λ¬A (¬b) with the union of facets not in ΛA (b). See Figure 20. ~ i ∩H1 = Ai as oriented hyperplanes, and by symmetry, A ~ i ∩H−1 = Note that, by construction, A ¬Ai . By projection with respect to the origin, A and ¬A can be glued along the “sphere at infinity” ~ i ∩ S n−1 }i . Similarly, Rb and ¬Rb can be glued together along a subspace of the “sphere to form {A ~ b , and ΛA (b) and Λ¬A (b) can be glued together likewise to obtain Λ ~ (b). at infinity” to obtain R A ~ b = Rb t∞ ¬Rb We denote this “gluing at infinity” construction by − t∞ −, so that we write R and ΛA~ (b) = ΛA (b) t∞ Λ¬A (b). ~ be its vectorization in Rk . Let N be an affine subspace of Rk−1 of dimension l − 1, and let N Then ~ ∩R ~ + 6= ∅ ⇐⇒ N ~ ∩R ~ − 6= ∅ N ∩ R+ 6= ∅ ⇐⇒ N ~ ∩ Λ ~ (b) 6= ∅ =⇒ H e l−2−i (N ~ ∩ Λ ~ (b)) = 0∀i ≥ 0] ⇐⇒ ∀b 6= +, −, [N A A ~ ∩ Λ ~ (b) = (N ∩ ΛA (b)) t∞ (N ∩ Λ¬A (b)), so we get the following But N A Corollary 3.56. N does not intersect R+ iff there is some b 6= +, − such that (N ∩ ΛA (b)) t∞ (N ∩ Λ¬A (b)) is nonempty and is not nulhomotopic. 59 3.5 3 Homological Farkas 2 f g Λ(g) Λ(f ) 1 Figure 21: Example application of Corollary 3.57. Let the hyperplanes (thin lines) be oriented such that the square at the center is R+ . The bold segments indicate the Λ of each region. Line 1 intersects R+ , and we can check that its intersection with any bold component is nulhomotopic. Line 2 does not intersect R+ , and we see that its intersection with Λ(f ) is two points, so has nontrivial zeroth reduced cohomology. Line 3 does not intersect R+ either, and its intersection with Λ(g) consists of a point in the finite plane and another point on the circle at infinity. When N does not intersect the closure R+ , we can just look at ΛA (b) and the component at infinity for a homological certificate. Corollary 3.57. Let A = {Ai }ni=1 be an affine hyperplane arrangement in Rk−1 whose normals affinely span Rk−1 . Suppose R+ is bounded and let N be an affine subspace of dimension l − 1. Then the following hold: e • (N ∩ ΛA (b)) = 0. 1. If R+ ∩ N 6= ∅, then for all b 6= +, −, N ∩ ΛA (b) 6= ∅ =⇒ H 2. If R+ ∩ N = ∅, then for each j ∈ [0, l − 2], there exists b 6= +, − such that N ∩ ΛA (b) 6= ∅ e j (N ∩ ΛA (b)) = 0 for some j. and H ~ := A ~ ∪ {H0 }, where H0 is the linear hyperplane of Rk with last Proof. (Item 1) Consider B ~ 0 for the region with sign c with respect to coordinate 0, oriented toward positive side. Write R c ~ + does not intersect H0 other than at the B (where c ∈ {−, +}n+1 ). Because R+ is bounded, R 0 ~ ~ ∩ Λ ~ (c) is nulhomotopic if nonempty. origin. Then N ∩ R+ 6= ∅ ⇐⇒ N ∩ R+ 6= ∅ ⇐⇒ ∀c, N B ~ ∩ Λ ~ (c) ∼ Note that for any b ∈ {−, +}n , we have ΛB~ (bb+) ∼ = ΛA (b), and N = N ∩ ΛA (b). (Here B bb+ means + appended to b). Substituing c = bb+ into the above yields the result. (Item 2) The most natural proof here adopts the algebraic approach. ~ ) ⊆ [[n + 1] → ~ Consider C := thr W ~ (N Let WB~ : Rk → Rn+1 be the embedding matrix for B. B ~ as {−, +}]. This is the class of functions corresponding to all the sign vectors achievable by N + ~ Define C := C . Since N ∩ R+ = ∅, βi,† (C) = I(i = l) it traverses through the regions of B. by Theorem 3.44. By the minimality of Betti numbers, for all j ≥ 0, βl−1−j,f (C) 6= 0 for some e j (Ξ ~ (+ ♦ f) ∩ N ~ ) 6= 0 f :⊆ [n + 1] → {−, +}, f 6= †, + with n + 1 6∈ dom f. But this means that H B ~ ∼ by the proof of Theorem 3.44. Of course, (+ ♦ f)(n + 1) = −, meaning that ΞB~ (+ ♦ f) ∩ N = ΛA (¬(+ ♦ f)) ∩ N . For the desired result, we just set b = ¬(+ ♦ f). Figure 21 gives an example application of Corollary 3.57. 60 3 APPLICATIONS 3.6 Probabilistic Interpretation of Hilbert Function In this section we exhibit a probabilistic interpretation of the Hilbert function of a canonical ideal. For a graded module M over S = k[x0 , . . . , xn−1 ], the graded Hilbert function HF(M ; a) takes an exponent sequence a to the dimension over k of the component of M with degree a. Its generating function X HF(M ; a)xa HS(M ; x) = a is called the graded Hilbert series of M . It is known [11] that K(M ; a) HF(M ; a) = Qn−1 i=0 (1 − xi ) for some polynomial K in n variables. This polynomial is called the K-polynomial of M . If one performs fractional decomposition on this rational function, then one can deduce that the Hibert function coincides with a polynomial when a has large total degree. (This is briefly demonstrated below for the N-graded version). This polynomial is called the Hilbert polynomial and is written HP(M ; a). Let χM (x) denote the graded Euler characteristic of a module M : XX χM (x) = (−1)i βi,a (M )xa . i≥0 a0 For example, for M = IC? , we write χC (x) := χIC? (x), and it takes the form χC (x) = X f ∈C xΓf − X xΓ(f ∩g) + · · · . f ∼C g It is shown in [11, Thm 4.11, 5.14] that χI ? (1 − x) = K(I ? ; 1 − x) = K(S/I; x) = χS/I (x) for any squarefree monomial ideal I. Now let C ⊆ [n → 2], and let S be its canonical base ring. For any f 6∈ C, the minimal generators of IC?f are xΓ(f ∩g) for g ∈ C “closest” to f . In particular, for every function h ∈ C, | dom f ∩ h| ≤ | dom f| = 2d+1 − totdeg xΓf for some xΓf ∈ mingen(IC?f ). Definition 3.58. Define the hardness of approximating f with C as ℵ(f, C) = min{2d − | dom f ∩ h| : h ∈ C} Then ℵ(f, C) is the smallest total degree of any monomial appearing in χCf minus 2d . Therefore, log χCf (ζ, . . . , ζ) − 2d ζ→0+ log ζ log K(S/ICf ; ϑ, . . . , ϑ) = lim − 2d ϑ→1− log 1 − ϑ log HS(S/ICf ; ϑ, . . . , ϑ) = lim + 2d ϑ→1− log 1 − ϑ ℵ(f, C) = lim 61 3.6 Probabilistic Interpretation of Hilbert Function where the last equality follows from HS(S/I; t, . . . t) = K(S/I; t, . . . , t)/(1 − t)2 d+1 . The N-graded Hilbert series expands into a Laurent polynomial in (1 − t), HS(S/I; t, . . . , t) = a−1 a−r−1 + ··· + + a0 + · · · as ts r+1 (1 − t) (1 − t) such that the N-graded Hilbert polynomial HP(S/I; t, . . . , t) has degree r. Thus ℵ(f, C) = 2d − (r + 1) = 2d − deg HP(S/ICf ; t, . . . , t) − 1 = 2d − totdeg HP(S/ICf ) − 1 d+1 −1
d+1 Note that total number of monomials in degree k is k+2 = Θ(k 2 −1 ). Therefore, if we 2d+1 −1 define ℘(k; f, C) to be the probability monomial ω of degree k has supp ω ⊆ f ∩ h   that a random for some h ∈ C, then ℘(k; f, C) = Θ HP(S/ICf ;k) k2d+1 −1 , and log ℘(k; f, C) − 2d . k→∞ log k ℵ(f, C) = − lim Now, ℘(k) is really the probability that a PF f has extension in C and is extended by f , where f is chosen from the distribution Qk that assigns a probability to f proportional to the number of monomials of total degree k whose support is f. More precisely,
k−1 Qk (f) = | dom f|−1
k+2d+1 −1 2d+1 −1 Note that there is a nonzero probability of choosing an invalid partial function, i.e. a monomial that is divisible by xu,0 xu,1 for some u ∈ [2d ]. Under this distribution, a particular PF of size d + 1 is Θ(k) times as likely as any particular PF of size d. As k → ∞, Qk concentrates more and more probability on the PFs of large size. By a similar line of reasoning using C instead of Cf , we see that deg HP(S/IC ) + 1 = 2d , so we define ℵ(C) = 0. Therefore the probability that a PF f has extension in C when f is drawn from Qk is d ℘(k; C) ∼ k −2 . We deduce that Theorem 3.59. The probability that a PF f drawn from Qk is extended by f when it is known to have extension in C is Θ(k −ℵ(f,C) ). Note that we are assuming f and C are fixed, and in particular when we are interested in a parametrized family (fd , Cd ), there might be dependence on d that is not written here. The main point we want to make, however, is that the Betti numbers of C and Cf affect the behavior of these classes under certain kinds of probability distributions. By considering higher Betti numbers and their dependence on the parameter d, we may compute the dependence of ℘ on d as well. Conversely, carrying over results from subjects like statistical learning theory could yield bounds on Betti numbers this way. 62 4 DISCUSSION 4 Discussion We have presented a new technique for complexity separation based on algebraic topology and Stanley-Reisner theory, which was used to give another proof of Minsky and Papert’s lower bound on the degree of polynomial threshold function required to compute parity. We also explored the connection between the algebraic/topological quantity dimh C and learning theoretical quantity dimVC C, and surprisingly found that the former dominates the latter, with equality in common computational classes. The theory created in this paper seems to have consequences even in areas outside of computation, as illustrated the Homological Farkas Lemma. Finally, we exhibited a probabilistic interpretation of the Hilbert function that could provide a seed for future developments in hardness of approximation. 4.1 Geometric Complexity Theory For readers familiar with Mulmuley’s Geometric Complexity program [13], a natural question is perhaps in what ways is our theory different? There is a superficial similarity in that both works associate mathematical objects to complexity classes and focus on finding obstructions to equality of complexity classes. In the case of geometric complexity, each class is associated to a variety, and the obstructions sought are of representation-theoretic nature. In our case, each class is associated to a labeled simplicial complex, and the obstructions sought are of homological nature. But beyond this similarity, the inner workings of the two techniques are quite distinct. Whereas geometric complexity focuses on using algebraic geometry and representation theory to shed light on primarily the determinant vs permanent question, our approach uses combinatorial algebraic topology and has a framework general enough to reason about any class of functions, not just determinant and permanent. This generality allowed, for example, the unexpected connection to VC dimension. It remains to be seen, however, whether these two algebraic approaches are related to each other in some way. 4.2 Natural Proofs So this homological theory is quite different from geometric complexity theory. Can it still reveal new insights on the P = NP problem? Based on the methods presented in this paper, one might ? try to show P/poly 6= NP by showing that the ideal ISIZE(d c ){3SAT } is not principal, for any c and d large enough d. Could Natural Proofs [16] present an obstruction? A predicate P : [2d → 2] → [2] is called natural if it satisfies • (Constructiveness) It is polynomial time in its input size: there is an 2O(d) -time algorithm that on input the graph of a function f ∈ [2d → 2], outputs P(f ). • (Largeness) A random function f ∈ [2d → 2] satisfies P(f ) = 1 with probability at least 1 n. Razborov and Rudich’s celebrated result says that Theorem 4.1. [16] Suppose there is no subexponentially strong one-way functions. Then there exists a constant c such that no natural predicate P maps SIZE(dc ) ⊆ [2d → 2] to 0. This result implicates that common proof methods used for proving complexity separation of lower complexity classes, like Hastad’s switching lemma used in the proof of parity 6∈ AC0 [2], cannot be used toward P vs NP. ? In our case, since SIZE(dc ) has 2poly(d) functions, naively computing the ideal ISIZE(d c ){3SAT } is d already superpolynomial time in 2d , which violates the “constructiveness” of natural proofs. Even 63 4.3 Homotopy Type Theory ? if the ideal ISIZE(d c )3SATd is given to us for free, computing the syzygies of a general ideal is NP- hard in the number of generators Ω(2d ) [4]. Thus a priori this homological technique is not natural (barring the possibility that in the future, advances in the structure of SSIZE(dc ) yield poly(2d )-time ? algorithms for the resolution of ISIZE(d c ){3SAT } ). d 4.3 Homotopy Type Theory A recent breakthrough in the connection between algebraic topology and computer science is the emergence of Homotopy Type Theory (HoTT) [19]. This theory concerns itself with rebuilding the foundation of mathematics via a homotopic interpretation of type theoretic semantics. Some of the key observations were that dependent types in type theory correspond to fibrations in homotopy theory, and equality types correspond to homotopies. One major contribution of this subfield is the construction of a new (programming) language which “simplifies” the semantics of equality type, by proving, internally in this language, that isomorphism of types “is equivalent” to equality of types. It also promises to bring automated proof assistants into more mainstream mathematical use. As such, HoTT ties algebraic topology to the B side (logic and semantics) of theoretical computer science. Of course, this is quite different from what is presented in this paper, which applies algebraic topology to complexity and learning theory (the A side of TCS). However, early phases of our homological theory were inspired by the “fibration” philosophy of HoTT. In fact, the canonical suboplex was first constructed as a sort of “fibration” (which turned out to be a cosheaf, and not a fibration) as explained in Appendix B. It remains to be seen if other aspects of HoTT could be illuminating in future research. 5 Future Work In this work, we have initiated the investigation of function classes through the point of view of homological and combinatorial commutative algebra. We have built a basic picture of this mathematical world but left many questions unanswered. Here we discuss some of the more important ones. Characterize when dimVC = dimh , or just approximately. We saw that all of the interesting computational classes discussed in this work, for example, linthr and linfun, have homological dimensions equal to their VC dimensions. We also showed that Cohen-Macaulay classes also satisfy this property. On the other hand, there are classes like delta whose homological dimensions are very far apart from their VC dimensions. A useful criterion for when this equality can occur, or when dimh = O(dimVC ), will contribute to a better picture when the homological properties of a class reflect its statistical/computational properties. Note that adding the all 0 function to delta drops its homological dimension back to its VC dimension. So perhaps there is a notion of “completion” that involves adding a small number of functions to a class to round out the erratic homological behaviors? Characterize the Betti numbers of thr Lf . We showed that the Betti numbers of thr Lf has nontrivial structure, and that some Betti numbers correspond to known concepts like weak representation of f . However, we only discovered a corner of this structure. In particular, what do the “middle dimension” Betti numbers look like? We make the following conjecture. 64 5 FUTURE WORK Conjecture 5.1. Let C = polythrkd and f 6∈ C. For every PF f in Cf , there is some i for which βi,f \f (C) is nonzero. It can be shown that this is not true for general thr L classes, but computational experiments suggest this seems to be true for polynomial thresholds. How do Betti numbers of thr Lf change under perturbation of f ? We proved a stability theorem for the “codimension 1” Betti numbers. In general, is there a pattern to how the Betti numbers respond to perturbation, other than remaining stable? Does every boolean function class have a minimal cellular or cocellular resolution? It is shown in [20] that there exist ideals whose minimal resolutions are not (CW) cellular. A natural question to ask here is whether this negative results still holds when we restrict to canonical ideals of boolean, or more generally finite, function classes. If so, we may be able to apply techniques from algebraic topology more broadly. When does a class C have pure Betti numbers? If we can guarantee that restriction preserves purity of Betti numbers, then Theorem 2.99 can be used directly to determine the Betti numbers of restriction of classes. Is this guarantee always valid? How do we obtain classes with pure Betti numbers? Under what circumstances can we expect separation of classes using high dimensional Betti numbers? Betti numbers at dimension 0 just encode the members of a class, and Betti numbers at dimension 1 encode the “closeness” relations on pairs of functions from the class. On the other hand, the maximal dimension Betti number of thr Lf encodes information about weak representation of f . So it seems that low dimension Betti numbers reflect more raw data while higher dimension Betti numbers reflect more “processed” data about the class, which are probably more likely to yield insights different from conventional means. Therefore, the power of our method in this view seems to depend on the dimension at which differences in Betti number emerges (as we go from high dimension to low dimension). Extend the probabilistic interpretation of Hilbert function. One may be able to manipulate the distribution Qk in Section 3.6 to arbitrary shapes when restricted to total functions, by modifying the canonical ideal. This may yield concrete connections between probabilistic computation and commutative algebra. Prove new complexity separation results using this framework We have given some examples of applying the homological perspective to prove some simple, old separation results, but hope to find proofs for nontrivial separations in the future. 65 Appendices A Omitted Proofs Proof of Proposition 2.25. The set of open cells in U \ ∂U is obviously U. So we need to show that U and ∂U are both subcomplex of Y . The first is trivial by Lemma 2.22. Suppose U = Yb . An open cell F̊ is in ∂U only if its label aF 6 b. But then any cell in its boundary ∂F must fall inside ∂U as well, because its exponent label majorizes aF . Thus the closed cell satsifies F ∈ ∂U. This shows ∂U is closed and thus a subcomplex by Lemma 2.22. The case of U = Y≺b has the same proof. For U = Yb , the only difference is the proof of ∂U being closed. We note that an open cell F̊ is in ∂U iff F̊ ∈ U and its label aF  b. Thus any open cell G̊ in its boundary ∂F falls inside ∂U as well, because its exponent label aG  aF  b. So F ∈ ∂U, and ∂U is closed, as desired. Proof of Lemma 2.29. Let E be the chain complex obtained from cochain complex F(X,A) by placing cohomological degree d at homological degree 0. For each a, we show the degree xa part Ea of E has rank 0 or 1 homology at homological degree 0 and trivial homology elsewhere iff one of the three conditions are satisfied. As a homological chain L complex, Ea consists of free modules Eai at each homological degree i isomorphic to a direct sum F ∈∆d−i ((X,A)a ) S, where ∆i (X, A) denotes the pure i-skeleton of the pair (X, A) (i.e. the collection of open cells of dimension i in X \ A). Writing SF for the copy of the base ring S corresponding to the cell F , the differential is given componentwise by X d : Eai → SG ∈ Eai−1 , a 7→ sign(F, G)aF . facetsF ⊂G If K is void, this chain is identically zero. Otherwise if ∂K is empty, then Ea just reproduces the reduced simplicial cochain complex of K — reduced because the empty cell is in K and thus has a corresponding copy of S at the highest e d−i (K) is nonzero only possible at i = 0, as desired, homological degree in Ea . Then Hi (Ea ) = H and at this i, the rank of the homology is 0 or 1 by assumption. Finally, if ∂K contains a nonempty cell, then Ea recovers the relative cochain complex for (K, ∂K). Then Hi (Ea ) = H̃ d−i (K, ∂K) is nonzero only possible at i = 0, where the rank of the homology is again 0 or 1. This proves the reverse direction (⇐). For the forward direction (⇒), suppose ∂K only contains an empty cell (i.e. does not satisfy conditions 1 and 2). Then Ea is the nonreduced cohomology chain complex of K, and therefore it must be the case that H i (K) = Hd−i (Ea ) = 0 at all i 6= d. But H 0 (K) = 0 implies K is empty, yielding condition 3. Otherwise, if ∂K is void, this implies condition 2 by the reasoning in the proof of the backward direction. Similarly, if ∂K is nonempty, this implies condition 1. Proof of Lemma 2.30. Let E be the chain complex obtained from cochain complex FY by placing cohomological degree d at homological degree 0. Then βi,b (I) = dimk Hi (E⊗k)b = dimk H d−i (FY ⊗ k)b . But the degree b part of FY ⊗ k is exactly the cochain complex of the collection of open cells Yb . By Proposition 2.25, Yb is realized by (Y b , ∂Yb ), so H d−i (FY ⊗ k)b = H d−i (Y b , ∂Yb ), which yields the desired result. 66 B COSHEAF CONSTRUCTION OF THE CANONICAL SUBOPLEX 1 2 id 1 ¬id ⊥ 1 2 1 > (0, 1) 41 ( 12 , 12 ) (1, 0) Figure B.1: “Gluing” together the metric spaces [2 → 2]p for all p ∈ 41 . The distances shown are L1 distances of functions within each “fiber.” ⊥ is the identically 0 function; > is the identically 1 function; id is the identity; and ¬id is the negation function. If we ignore the metric and “untangle” the upper space, we get the complete 1-dimensional suboplex. Proof of Lemma 2.72. WLOG, we can replace Rq with the span of L, so we assume L spans Rq . We show by induction on q that if L 6⊆ H for every open halfspace, then 0 ∈ L. This would imply our result: As L is open in Rq , there is a ball contained in L centered at the origin. Since L is a cone, this means L = Rq . Note that L 6⊆ H for every open coordinate halfspace H is equivalent to that L ∩ H 6= ∅ for every open coordinate halfspace H. Indeed, if L ∩ H 0 = ∅, then Rq \ H 0 contains the open set L, and thus the interior int(Rq \ H) is an open coordinate halfspace that contains L. If L intersects every open coordinate halfspace, then certainly it cannot be contained in any single H, or else int(Rq \ H) does not intersect L. We now begin the induction. The base case of q = 1: L has both a positive point and negative point, and thus contains 0 because it is convex. Suppose the induction hypothesis holds for q = p, and let q = p + 1. Then for any halfspace H, L ∩ H and L ∩ int(Rq \ H) are both nonempty, and thus L intersects the hyperplane ∂H by convexity. Certainly L ∩ ∂H intersects every open coordinate halfspace of ∂H because the latter are intersections of open coordinate halfspaces of Rq with ∂H. So by the induction hypothesis, L ∩ ∂H contains 0, and therefore 0 ∈ L as desired. B Cosheaf Construction of the Canonical Suboplex Let C ⊆ [n → m] and p be a probability distribution on [n]. p induces an L1 metric space Cp by d(f, g) = n1 kf − gk1 . If we vary p over 4n−1 , then Cp traces out some kind of shape that “lies over” 4n−1 . For C = [2 → 2], this is illustrated in Figure B.1. In this setting, Impagliazzo’s Hardcore Lemma [2] would say something roughly like the following: Let C ⊆ [n → 2] and C be the closure of C under taking majority over “small” subsets of C. For any f ∈ [n → 2], either in the fiber [n → 2]U over the uniform distribution U, f is “close” to C, or in the fiber [n → 2]q for some q “close” to U, f is at least distance 1/2 +  from C. Thus this view of “fibered metric spaces” may be natural for discussion of hardness of approximation or learning theory. 67 B.1 Cosheaves and Display Space If we ignore the metrics and “untangle” the space, we get the canonical suboplex of [2 → 2], the complete 1-dimensional suboplex. In general, the canonical suboplex of a class C ⊆ [n → m] can be obtained by “gluing” together the metric spaces Cp for all p ∈ 4n−1 , so that there is a map Υ : SC → 4n−1 whose fibers are Cp (treated as a set). But how do we formalize this “gluing” process? In algebraic topology, one usually first tries to fit this picture into the framework of fibrations or the framework of sheaves. But fibration is the wrong concept, as our “fibers” over the “base space” 4n−1 are not necessarily homotopy equivalent, as seen in Figure B.1. So SC cannot be the total space of a fibration over base space 4n−1 . Nor is it the étalé space of a sheaf, as one can attest to after some contemplation. It turns out the theory of cosheaves provide the right setting for this construction. B.1 Cosheaves and Display Space Definition B.1. A precosheaf is a covariant functor F : O(X) → Set from the poset of open sets of a topological space X to the category of sets. For each inclusion ı : U ,→ V , the set map Fı : F(U ) → F(V ) is called the inclusion map from F(U ) to F(V ). A precosheaf F is further called a cosheaf if it satisfies the following cosheaf condition: For S every open covering {Ui }i of an open set U ⊆ X with {Ui }i = U , a a a F(Uk ) ← F(Uk ∩ Ul ) → F(Ul ) k k6=l l has pushout F(U ). Here each arrow is the coproduct of inclusion maps. There is a concept of costalk dual to the concept of stalks in sheaves. Definition B.2. Let F : O(X) → Set be a cosheaf and let p ∈ X. Then the costalk Fp is defined as the cofiltered limit Fp := lim F(U ), U ∈p of F(U ) over all open U containing p. Analogous to the étalé space of a sheaf, cosheaves have something called a display space [8] that compresses all of its information in a topological space. We first discuss the natural cosheaf associated to a continuous map. Let ψ : Y → X be a continuous map between locally path-connected spaces Y and X. We have a cosheaf F ψ : O(X) → Set induced as follows: For each U ∈ O(X), F ψ (U ) = π0 (ψ −1 U ), where π0 denotes the set of connected components. For an inclusion ı : U ,→ V , F ψ (ı) maps each component in Y of ψ −1 U into the component of ψ −1 V that it belongs to. For open cover {Ui }i with union U , a a a F ψ (Uk ) ← F ψ (Uk ∩ Ul ) → F ψ (Ul ) k k6=l l has pushout F ψ (U ). Indeed, this is just the standard gluing construction of pushouts in Set for each component of ψ −1 U . (An alternative view of F ψ is that it is the direct image cosheaf of F id , where id : Y → Y is the identity). Now we reverse the construction. Let X be a topological space, and F : O(X) → Set be a cosheaf. We construct the display space Y and a map ψ : Y → X such that F ∼ = F ψ . For the points of Y , we will take the disjoint union of all costalks, G |Y | := Fp . p∈X 68 B COSHEAF CONSTRUCTION OF THE CANONICAL SUBOPLEX Then the set-map |ψ| underlying the desired ψ will be Fp 3 y 7→ p ∈ X. Now we topologize Y by exhibiting a basis. For any U ∈ O(X), there is a canonical map G G gU := mp,U : Fp → F(U ) p∈U p∈U formed by the coproduct of the limit maps mp,U : Fp → F(U ). Then each fiber of gU is taken as an open set in Y : For each s ∈ F(U ), we define [s, U ] := gU−1 (s) as an open set. Note that [s, U ]∩[t, U ] = ∅ if s, t ∈ F(U ) but s 6= t. We claim that for s ∈ F(U ), t ∈ F(V ), G (1) [s, U ] ∩ [t, V ] = {[r, U ∩ V ] : F(iU )(r) = s & F(iV )(r) = t} where iU : U ∩ V → U and iV : U ∩ V → V are the inclusions. The inclusion of the RHS into the LHS should be clear. For the opposite direction, suppose p ∈ U ∩ V and y ∈ Fp with gU (y) = mp,U (y) = s and gV (y) = mp,V (y) = t. Since Fp is the cofiltered limit of {F(W ) : p ∈ W }, we have the following commutative diagram Fp mp,U ∩V mp,U mp,V F(U ∩ V ) F (k) F (j) F(U ) F(V ) Therefore there is an r ∈ F(U ∩ V ) such that mp,U ∩V (y) = r and F(j)(r) = s and F(k)(r) = t. Then y ∈ [r, U ∩ V ] ⊆ RHS of 1. Our claim is proved, and {[s, U ] : s ∈ F(U )} generates a topological basis for Y . Finally, to complete the verification that F ∼ = F ψ , we show that F(U ) ∼ = π0 (ψ −1 U ), natural over all U ∈ O(X). It suffices to prove that for each U ∈ O(X) and s ∈ F(U ), [s, U ] is connected; then F(U ) 3 s 7→ [s, U ] is a natural isomorphism. S partition i∈A {[si , Ui ]} t S Suppose for some s ∈ F(U ) this is not true: S there exists a nontrivial S {[sj , Uj ]} = j∈BS S [s, U ] of [s, U ] by open sets i∈A {[si , Ui ]} and j∈B {[sj , Uj ]}. We assume WLOG that i∈A Ui ∪ j∈B Uj = U (in case that for some x ∈ U , Fp = ∅, we extend each Ui and Uj to cover x). Then by the cosheaf condition, the pushout of the following a a a F(Uk ) ← F(Uk ∩ Ul ) → F(Ul ) k∈A∪B k6=l l∈A∪B is F(U ). By assumption, F(Ui  U )(si ) = s for all i ∈ A ∪ B. So there must be some i ∈ A, j ∈ B and t ∈ F(Ui ∩ Uj ) such that F(Ui ∩ Uj  Ui )(t) = si and F(Ui ∩ Uj  Uj )(t) = sj . This implies that [t, Ui ∩ Uj ] ⊆ [si , Ui ] ∩ [sj , Uj ]. If X is first countable and locally compact Hausdorff, or if X S S is metrizable, then by Lemma B.3, [t, Ui ∩ Uj ] is nonempty, and therefore 6 ∅, a contradiction, as desired. j∈B {[sj , Uj ]} = i∈A {[si , Ui ]} ∩ 69 B.1 Cosheaves and Display Space Lemma B.3. If X is first countable and locally compact Hausdorff, or if X is metrizable, then [s, U ] is nonempty for every U ∈ O(X) and s ∈ F(U ). Proof. We give the proof for the case when X is first countable and locally compact Hausdorff. The case of metrizable X is similar. For each x ∈ X, fix a countable local basis x ⊆ · · · ⊆ Bxn ⊆ Bxn−1 ⊆ · · · ⊆ Bx2 ⊆ Bx1 , with the property that Bxn ⊆ Bxn−1 and is compact. Fix such a U and s ∈ F(U ). Let U0 := U and s0 := s. We will form a sequence hUi , si i as follows. Given Ui−1 and si−1 , for each point x ∈ Ui−1 , choose a kx > i such that Bxkx is contained in Ui−1 . These sets {Bxkx }x form an open covering of Ui−1 , and by the sheaf condition, for some x, im F(Bxkx  Ui−1 ) contains si−1 . Then set Ui := Bxkx and choose any element of F(Bxkx  Ui−1 )−1 (si−1 ) to be si . Hence by construction si ∈ F(Ui ). Following this procedure for all i ∈ N, we obtain a sequence hUi , si ii≥0 T T withTthe property that U0 ⊇ U1 ⊇ U1 ⊇ U2 ⊇ U2 · · · . As each of Ui is compact, Ui , and hence Ui = Ui , is nonempty. i Let T z be one of its elements. Then Ui ⊆ Bz for all i ≥ 1. Therefore z must be the unique element of Ui , and the sequence hUi ii is a local basis of z. Furthermore, hsi ii is an element of the costalk at z, as it can easily be seen to be an element of the inverse limit limi→∞ F(Ui ) = lim{F(V ) : z ∈ V }. This shows that [s, U ] is nonempty. Note that without assumptions on X, Lemma B.3 cannot hold. In fact, something quite extreme can happen. Proposition B.4. There exists a cosheaf F : O(X) → Set whose costalks are all empty. Proof. This proof is based on Waterhouse’s construction [21]. Let X be an uncountable set with the cofinite topology. Define F(U ) to be the set of injective functions from the finite set X \ U to the integers. The map F(U  V ) just restricts a function g : X \U → Z to g  (X \V ) : X \V → Z. One can easily check that the cosheaf sequence is a pushout. Thus F is a cosheaf. For any x ∈ X, each point of the inverse limit of {F(U ) : x ∈ U } has the following description: a sequence of injective functions hfA : A  ZiA indexed by finite sets A ⊆ X, such that if A S ⊆ B are both finite sets, then fA ⊆ fB . Such a sequence would determine an injective function A fA : X → Z, but that is impossible as X was assumed to be uncountable. Back to our case of canonical suboplex. For any S = S[n→m] , there is a canonical embedding Ξ : S  4mn−1 ⊆ Rmn , defined by taking vertex Vu,i , (u, i) ∈ [n] ×P[m] to eu,i , the basis vector of Rmn corresponding to (u, i), and taking each convex combination n−1 u=0 p(u)Vu,f (i) in the simplex Pn−1 associated to f : [n] → [m] to u=0 p(u)eu,f (i) . The map Υ : SC → 4n−1 we sketched in the beginning of this section can then be formally described as Υ = Π ◦ Ξ  SC , where Π is the linear projection defined by eu,i 7→ eu ∈ 4n−1 . As we have shown, Υ induces a cosheaf F Υ : O(4n−1 ) → Set, sending each open U ⊆ 4n−1 to π0 (Υ−1 U ). For example, if U is in the interior of 4n−1 , then F Υ (U ) has size equal to the size of C. If U is a small ball around the vertex eu , then F Υ (U ) is bijective with the set of values C takes on u ∈ [n]. It is easy to check that the costalk FpΥ at each point p ∈ 4n−1 is just π0 (Υ−1 p) = |Cp |, the set underlying the metric space Cp , so we have successfully “glued” together the pieces into a topological space encoding the separation information in C. One may naturally wonder whether the cosheaf homology of such a cosheaf matches the homology of the display space. One can show that this is indeed the case for our canonical suboplex, via identification of the cosheaf homology with Cech homology and an application of the acyclic cover lemma. What is disappointing about this construction is of course that it ignores metric information in all of the costalks Cp . Directly replacing Set with the category Met of metric spaces with metric 70 REFERENCES maps (maps that do not increase distance) does not work, because it does not have coproducts. It remains an open problem whether one can find a suitable category to replace Set such that SC can still be expressed as the display space of a cosheaf on 4n−1 , while preserving metric information in each costalk, and perhaps more importantly, allows the expression of results like Impagliazzo’s Hardcore Lemma in a natural categorical setting. Perhaps a good starting point is to notice that the embedding Ξ actually preserves the L1 metric within each fiber Cp . References [1] Martin Anthony. Discrete Mathematics of Neural Networks. 2001. [2] Sanjeev Arora and Boaz Barak. Computational complexity: a modern approach. 2009. OCLC: 443221176. [3] J. Aspnes, R. Beigel, M. Furst, and S. Rudich. The expressive power of voting polynomials. Combinatorica, 14(2):135–148, June 1994. [4] Dave Bayer and Mike Stillman. Computation of Hilbert functions. Journal of Symbolic Computation, 14(1):31–50, 1992. [5] Winfried Bruns and Jurgen Herzog. Cohen-Macaulay Rings. Cambridge University Press, 1998. [6] David Eisenbud. Commutative algebra: with a view toward algebraic geometry. 1994. OCLC: 891662214. [7] Sara Faridi. The projective dimension of sequentially Cohen-Macaulay monomial ideals. arXiv:1310.5598 [math], October 2013. arXiv: 1310.5598. [8] J. Funk. The display locale of a cosheaf. Cahiers de Topologie et Gomtrie Diffrentielle Catgoriques, 36(1):53–93, 1995. [9] Melvin Hochster. The canonical module of a ring of invariants. Contemp. Math, 88:43–83, 1989. [10] Michael Kearns and Umesh Vazirani. An Introduction to Computational Learning Theory. January 1994. [11] Ezra Miller and Bernd Sturmfels. Combinatorial commutative algebra. Number 227 in Graduate texts in mathematics. Springer, New York, 2005. OCLC: ocm55765389. [12] Marvin Minsky and Seymour Papert. Perceptrons: An Introduction to Computational Geometry. MIT Press, 1969. [13] K. Mulmuley and M. Sohoni. Geometric Complexity Theory I: An Approach to the P vs. NP and Related Problems. SIAM Journal on Computing, 31(2):496–526, January 2001. [14] Thomas Natschlger and Michael Schmitt. Exact VC-dimension of Boolean monomials. Information Processing Letters, 59(1):19–20, July 1996. [15] Ryan O’Donnell. Analysis of boolean functions. Cambridge University Press, New York, NY, 2014. 71 REFERENCES [16] Alexander A Razborov and Steven Rudich. Natural Proofs. Journal of Computer and System Sciences, 55(1):24–35, August 1997. [17] Saharon Shelah. A combinatorial problem; stability and order for models and theories in infinitary languages. Pacific Journal of Mathematics, 41(1):247–261, 1972. [18] Richard P. Stanley. Combinatorics and commutative algebra. Number v. 41 in Progress in mathematics. Birkhuser, Boston, 2nd ed edition, 1996. [19] The Univalent Foundations Program. Homotopy Type Theory: Univalent Foundations of Mathematics. arXiv preprint arXiv:1308.0729, 2013. [20] Mauricio Velasco. Minimal free resolutions that are not supported by a CW-complex. Journal of Algebra, 319(1):102–114, January 2008. [21] William C. Waterhouse. An empty inverse limit. Proceedings of the American Mathematical Society, 36(2):618, 1972. [22] Gnter M. Ziegler. Lectures on Polytopes, volume 152 of Graduate Texts in Mathematics. Springer New York, New York, NY, 1995. 72 

Approximate Clustering with Same-Cluster Queries Nir Ailon1 , Anup Bhattacharya2 , Ragesh Jaiswal2 , and Amit Kumar2 arXiv:1704.01862v3 [cs.DS] 4 Oct 2017 2 1 Technion, Haifa, Israel.⋆ Department of Computer Science and Engineering, Indian Institute of Technology Delhi.⋆⋆ Abstract. Ashtiani et al. proposed a Semi-Supervised Active Clustering framework (SSAC), where the learner is allowed to make adaptive queries to a domain expert. The queries are of the kind “do two given points belong to the same optimal cluster?”, and the answers to these queries are assumed to be consistent with a unique optimal solution. There are many clustering contexts where such same cluster queries are feasible. Ashtiani et al. exhibited the power of such queries by showing that any instance of the k-means clustering problem, with additional margin assumption, can be solved efficiently if one is allowed O(k2 log k + k log n) same-cluster queries. This is interesting since the k-means problem, even with the margin assumption, is NP-hard. In this paper, we extend the work of Ashtiani et al. to the approximation setting showing that a few of such same-cluster queries enables one to get a polynomial-time (1 + ε)-approximation algorithm for the k-means problem without any margin assumption on the input dataset. Again, this is interesting since the k-means problem is NP-hard to approximate within a factor (1 + c) for a fixed constant 0 < c < 1. The number of same-cluster queries used is poly(k/ε) which is independent of the size n of the dataset. Our algorithm is based on the D2 -sampling technique, also known as the k-means++ seeding algorithm. We also give a conditional lower bound on the number of same-cluster queries showing that if the Exponential Time Hypothesis (ETH) holds, then any such efficient query algorithm  needs to make Ω polyklog k same-cluster queries. Our algorithm can be extended for the case when the oracle is faulty, that is, it gives wrong answers to queries with some bounded probability. Another result we show with respect to the k-means++ seeding algorithm is that a small modification to the kmeans++ seeding algorithm within the SSAC framework converts it to a constant factor approximation algorithm instead of the well known O(log k)-approximation algorithm. 1 Introduction Clustering is extensively used in data mining and is typically the first task performed when trying to understand large data. Clustering basically involves partitioning given data into groups or clusters such that data points within the same cluster are similar as per some similarity measure. Clustering is usually performed in an unsupervised setting where data points do not have any labels associated with them. The partitioning is done using some measure of similarity/dissimilarity between data elements. In this work, we extend the work of Ashtiani et al. [AKBD16] who explored the possibility of performing clustering in a semi-supervised active learning setting for center based clustering problems such as k-median/means. In this setting, which they call Semi-Supervised Active Clustering framework or SSAC in short, the clustering algorithm is allowed to make adaptive queries of the form: do two points from the dataset belong to the same optimal cluster?. where query answers are assumed to be consistent with a unique optimal solution. Ashtiani et al. [AKBD16] started the study of understanding the strength of this model. Do hard clustering problems become easy in this model? They explore such questions in the context of center-based clustering problems. Center based clustering problems such as k-means are extensively used to analyze large data clustering problems. Let us define the k-means problem in the Euclidean setting. ⋆ ⋆⋆ Email address: nailon@cs.technion.ac.il Email addresses: {anupb, rjaiswal, amitk}@cse.iitd.ac.in Definition 1 (k-means problem). Given a dataset X ⊆ Rd containing n points, and a positive integer k, find a set of k points C ⊆ Rd (called centers) such that the following cost function is minimized: X Φ(C, X) = min D(x, c). x∈X c∈C D(x, c) denotes the squared Euclidean distance between c and x. That is, D(x, c) = ||x − c||2 . Note that the k optimal centers c1 , ..., ck of the k-means problem define k clusters of points in a natural manner. All points for which the closest center is ci belong to the ith cluster. This is also known as the Voronoi partitioning and the clusters obtained in this manner using the optimal k centers are called the optimal clusters. Note that the optimal center for the 1-means problem P def. for any dataset X ⊆ Rd is the centroid of the dataset denoted by µ(X) = x∈X |X| x . This means that if X1 , ...., Xk are the optimal clusters for the k-means problem on any dataset X ⊆ Rd and c1 , ..., ck are the corresponding optimal centers, then ∀i, ci = µ(Xi ). The k-means problem has been widely studied and various facts are known about this problem. The problem is tractable when either the number k of clusters or the dimension d equal to 1. However, when k > 1 or d > 1, then the problem is known to be NP-hard [Das08, Vat09, MNV12]. There has been a number of works of beyond the worst-case flavour for k-means problem in which it is typically assumed that the dataset satisfies some separation condition, and then the question is whether this assumption can be exploited to design algorithms providing better guarantees for the problem. Such questions have led to different definitions of separation and also some interesting results for datasets that satisfy these separation conditions (e.g., [ORSS13, BBG09, ABS12]). Ashtiani et al. [AKBD16] explored one such separation notion that they call the γ-margin property. Definition 2 (γ-margin property). Let γ > 1 be a real number. Let X ⊆ Rd be any dataset and k be any positive integer. Let PX = {X1 , ..., Xk } denote k optimal clusters for the k-means problem. Then this optimal partition of the dataset PX is said to satisfy the γ-margin property iff for all i 6= j ∈ {1, ..., k} and x ∈ Xi and y ∈ Xj , we have: γ · ||x − µ(Xi )||< ||y − µ(Xi )||. Qualitatively, what this means is that every point within some cluster is closer to its own cluster center than a point that does not belong to this cluster. This seems to be a very strong separation property. Ashtiani et al. [AKBD16] showed that the k-means clustering problem is NP-hard even √ when restricted to instances that satisfy the γ-margin property for γ = 3.4 ≈ 1.84. Here is the formal statement of their hardness result. Theorem 1 (Theorem 10 in [AKBD16]). Finding an optimal solution to k-means objective function is NP-hard when k = Θ(nε ) for any √ ε ∈ (0, 1), even when there is an optimal clustering that satisfies the γ-margin property for γ = 3.4. In the context of the k-means problem, the same-cluster queries within the SSAC framework are decision questions of the form: Are points x, y such that x 6= y belong to the same optimal cluster? 3 Following is the main question explored by Ashitiani et al. [AKBD16]: √ Under the γ-margin assumption, for a fixed γ ∈ (1, 3.4], how many queries must be made in the SSAC framework for k-means to become tractable? 3 In case where the optimal solution is not unique, the same-cluster query answers are consistent with respect to any fixed optimal clustering. Ashtiani et al. [AKBD16] addressed the above question and gave a query algorithm. Their algorithm, in fact, works for a more general setting where the clusters are not necessarily optimal. In the more general setting, there is a target clustering X̄ = X̄1 , ..., X̄k of the given dataset X ⊆ Rd (not necessarily optimal clusters) such that these clusters satisfy the γ-margin property (i.e., for all i, x ∈ X̄i , and y ∈ / X̄i , γ · ||x − µ(X̄i )||< ||y − µ(X̄i )||) and the goal of the query algorithm is to output the clustering X̄. Here is the main result of Ashtiani et al. Theorem 2 (Theorems 7 and 8 in [AKBD16]). Let δ ∈ (0, 1) and γ > 1. Let X ⊆ Rd be any dataset containing n points, k be a positive integer, and X1 , ..., Xk be any target clustering  of X that satisfies the γ-margin property. Then there is a query algorithm A that makes k+log 1/δ O k log n + k2 log(γ−1) same-cluster queries and with probability at least (1 − δ) outputs the 4   k+log 1/δ clustering X1 , ..., Xk . The running time of algorithm A is O kn log n + k2 log(γ−1) . 4 The above result is a witness to the power of the SSAC framework. We extend this line of work by examining the power of same-cluster queries in the approximation algorithms domain. Our results do not assume any separation condition on the dataset (such as γ-margin as in [AKBD16]) and they hold for any dataset. Since the k-means problem is NP-hard, an important line of research work has been to obtain approximation algorithms for the problem. There are various efficient approximation algorithms for the k-means problem, the current best approximation guarantee being 6.357 by Ahmadian et al. [ANFSW16]. A simple approximation algorithm that gives an O(log k) approximation guarantee is the k-means++ seeding algorithm (also known as D 2 -sampling algorithm) by Arthur and Vassilvitskii [AV07]. This algorithm is commonly used in solving the k-means problem in practice. As far as approximation schemes or in other words (1 + ε)-approximation algorithms (for arbitrary small ε < 1) are concerned, the following is known: It was shown by Awasthi et al. [ACKS15] that there is some fixed constant 0 < c < 1 such that there cannot exist an efficient (1 + c) factor approximation unless P = NP. This result was improved by Lee et al. [LSW17] where it was shown that it is NP-hard to approximate the k-means problem within a factor of 1.0013. However, when either k or d is a fixed constant, then there are Polynomial Time Approximation Schemes (PTAS) for the k-means problem.4 Addad et al. [CAKM16] and Friggstad et al. [FRS16] gave a PTAS for the k-means problem in constant dimension. For fixed constant k, various PTASs are known [KSS10, FMS07, JKS14, JKY15]. Following is the main question that we explore in this work: For arbitrary small ε > 0, how many same-cluster queries must be made in an efficient (1 + ε)-approximation algorithm for k-means in the SACC framework? The running time should be polynomial in all input parameters such as n, k, d and also in 1/ε. Note that this is a natural extension of the main question explored by Ashtiani et al. [AKBD16]. Moreover, we have removed the separation assumption on the data. We provide an algorithm that makes poly(k/ε) same-cluster queries and runs in time O(nd · poly(k/ε)). More specifically, here is the formal statement of our main result: Theorem 3 (Main result: query algorithm). Let 0 < ε ≤ 1/2, k be any positive integer, and X ⊆ Rd . Then there is a query algorithm A that runs in time O(ndk 9 /ε4 ) and with probability at least 0.99 outputs a center set C such that Φ(C, X) ≤ (1 + ε) · ∆k (X). Moreover, the number of 4 This basically means an algorithm that runs in time polynomial in the input parameters but may run in time exponential in 1/ε. same-cluster queries used by A is O(k9 /ε4 ). Here ∆k (X) denotes the optimal value of the k-means objective function. Note that unlike Theorem 2, our bound on the number of same-cluster queries is independent of the size of the dataset. We find this interesting and the next natural question we ask is whether this bound on the number of same-cluster queries is tight in some sense. In other words, does there exist a query algorithm in the SSAC setting that gives (1+ε)-approximation in time polynomial in n, k, d and makes significantly fewer queries than the one given in the theorem above? We answer this question in the negative by establishing a conditional lower bound on the number of same-cluster queries under the assumption that ETH (Exponential Time Hypothesis) [IP01, IPZ01] holds. The formal statement of our result is given below. Theorem 4 (Main result: query lower bound). If the Exponential Time Hypothesis (ETH) holds, then there exists a constant c > 1 such that any c-approximation query algorithm for the k queries. k-means problem that runs in time poly(n, d, k) makes at least poly log k Faulty query setting The existence of a same-cluster oracle that answers such queries perfectly may be too strong an assumption. A more reasonable assumption is the existence of a faulty oracle that can answer incorrectly but only with bounded probability. Our query approximation algorithm can be extended to the setting where answers to the same-cluster queries are faulty. More specifically, we can get wrong answers to queries independently but with probability at most some constant q < 1/2. Also note that in our model the answer for a same-cluster query does not change with repetition. This means that one cannot ask the same query multiple times and amplify the probability of correctness. We obtain (1 + ε)-approximation guarantee for the k-means clustering problem in this setting. The main result is given as follows. Theorem 5. Let 0 < ε ≤ 1/2, k be any positive integer, and X ⊆ Rd . Consider a faulty SSAC setting where the response to every same-cluster query is incorrect with probability at most some constant q < 1/2. In such a setting, there is a query algorithm AE that with probability at least 0.99, outputs a center set C such that Φ(C, X) ≤ (1 + ε) · ∆k (X). Moreover, the number of same-cluster queries used by AE is O(k15 /ε8 ). The previous theorems summarise the main results of this work which basically explores the power of same-cluster queries in designing fast (1 + ε)-approximation algorithms for the k-means problem. We will give the proofs of the above theorems in Sections 3, 4, and 5. There are some other simple and useful contexts, where the SSAC framework gives extremely nice results. One such context is the popular k-means++ seeding algorithm. This is an extremely simple sampling based algorithm for the k-means problem that samples k centers in a sequence of k iterations. We show that within the SSAC framework, a small modification of this sampling algorithm converts it to one that gives constant factor approximation instead of O(log k)-approximation [AV07] that is known. This is another witness to the power of same-cluster queries. We begin the technical part of this work by discussing this result in Section 2. Some of the basic techniques involved in proving our main results will be introduced while discussing this simpler context. Other related work Clustering problems have been studied in different semi-supervised settings. Basu et al. [BBM04] explored must-link and cannot-link constraints in their semi-supervised clustering formulation. In their framework, must-link and cannot-link constraints were provided explicitly as part of the input along with the cost of violating these constraints. They gave an active learning formulation for clustering in which an oracle answers whether two query points belong to the same cluster or not, and gave a clustering algorithm using these queries. However, they work with a different objective function and there is no discussion on theoretical bounds on the number of queries. In contrast, in our work we consider the k-means objective function and provide bounds on approximation guarantee, required number of adaptive queries, and the running time. Balcan and Blum [BB08] proposed an interactive framework for clustering with split/merge queries. Given a clustering C = {C1 , . . .}, a user provides feedback by specifying that some cluster Cl should be split, or clusters Ci and Cj should be merged. Awasthi et al. [ABV14] studied a local interactive algorithm for clustering with split and merge feedbacks. Voevodski et al. [VBR+ 14] considered one versus all queries where query answer for a point s ∈ X returns distances between s to all points in X. For a k-median instance satisfying (c, ε)-approximation stability property [BBG09], the authors found a clustering close to true clustering using only O(k) one versus all queries. Vikram and Dasgupta [VD16] designed an interactive bayesian hierarchical clustering algorithm. Given dataset X, the algorithm starts with a candidate hierarchy T , and an initially empty set C of constraints. The algorithm queries user with a subtree T |S of hierarchy T restricted to constant sized set S ⊂ X of leaves. User either accepts T |S or provides a counterexample triplet ({a, b}, c) which the algorithm adds to its set of constraints C, and updates T . They consider both random and adaptive ways to select S to query T |S . Our Techniques We now give a brief outline of the new ideas needed for our results. Many algorithms for the k-means problem proceed by iteratively finding approximations to the optimal centers. One such popular algorithm is the k-means++ seeding algorithm [AV07]. In this algorithm, one builds a set of potential centers iteratively. We start with a set C initialized to the empty set. At each step, we sample a point with probability proportional to the square of the distance from C, and add it to C. Arthur and Vassilvitskii [AV07] showed that if we continue this process till |C| reaches k, then the corresponding k-means solution has expected cost O(log k) times the optimal k-means cost. Aggarwal et al. [ADK09] showed that if we continue this process till |C| reaches βk, for some constant β > 1, then the corresponding k-means solution (where we actually open all the centers in C) has cost which is within constant factor of the optimal k-means cost with high probability. Ideally, one would like to stop when size of C reaches k and obtain a constant factor approximation guarantee. We know from previous works [AV07, BR13, BJA16] that this is not possible in the classical (unsupervised) setting. In this work, we show that one can get such a result in the SSAC framework. A high-level way of analysing the k-means++ seeding algorithm is as follows. We first observe that if we randomly sample a point from a cluster, then the expected cost of assigning all points of this cluster to the sampled point is within a constant factor of the cost of assigning all the points to the mean of this cluster. Therefore, it suffices to select a point chosen uniformly at random from each of the clusters. Suppose the set C contains such samples for the first i clusters (of an optimal solution). If the other clusters are far from these i clusters, then it is likely that the next point added to C belongs to a new cluster (and perhaps is close to a uniform sample). However to make this more probable, one needs to add several points to C. Further, the number of samples that needs to be added to C starts getting worse as the value of i increases. Therefore, the algorithm needs to build C till its size becomes O(k log k). In the SSAC framework, we can tell if the next point added in C belongs to a new cluster or not. Therefore, we can always ensure that |C| does not exceed k. To make this idea work, we need to extend the induction argument of Arthur and Vassilvitskii [AV07] – details are given in Section 2. We now explain the ideas for the PTAS for k-means. We consider the special case of k = 2. Let X1 and X2 denote the optimal clusters with X1 being the larger cluster. Inaba et al. [IKI94] showed that if we randomly sample about O(1/ε) points from a cluster, and let µ′ denote the mean of this subset of sampled points, then the cost of assigning all points in the cluster to µ′ is within (1 + ε) of the cost of assigning all these points to their actual mean (whp). Therefore, it is enough to get uniform samples of size about O(1/ε) from each of the clusters. Jaiswal et al. [JKS14] had the following approach for obtaining a (1 + ε)-approximation algorithm for k-means (with running time being nd · f (k, ε), where f is an exponential function of k/ε) – suppose we sample about O(1/ε2 ) points from the input, call this sample S. It is likely to contain at least O(1/ε) from X1 , but we do not know which points in S are from X1 . Jaiswal et al. addressed this problem by cycling over all subsets of S. In the SSAC model, we can directly partition S into S ∩ X1 and S ∩ X2 using |S| same-cluster queries. Having obtained such a sample S, we can get a close approximation to the mean of X1 . So assume for sake of simplicity that we know µ1 , the mean of X1 . Now we are faced with the problem of obtaining a uniform sample from X2 . The next idea of Jaiswal et al. is to sample points with probability proportional to square of distance from µ1 . This is known as D 2 -sampling. Suppose we again sample about O(1/ε2 ) such points, call this sample S ′ . Assuming that the two clusters are far enough (otherwise the problem only gets easier), they show that S ′ will contain about O(1/ε2 ) points from X2 (with good probability). Again, in the SSAC model, we can find this subset by |S ′ | queries – call this set S ′′ . However, the problem is that S ′′ may not represent a uniform sample from X2 . For any point e ∈ X2 , let pe denote the conditional probability of sampling e given that a point from X2 is sampled when sampled using D 2 -sampling. They showed ε , where m denotes the size of X2 . In order for the sampling lemma of Inaba et al. pe is at least m [IKI94] to work, we cannot work with approximately uniform sampling. The final trick of Jaiswal et al. was to show that one can in fact get a uniform sample of size about O(ε|S ′′ |) = O(1/ε) from S ′′ . The idea is as follows – for every element e ∈ S ′′ , we retain it with probability peεm (which is at most 1), otherwise we remove it from S ′′ . It is not difficult to see that this gives a uniform sample from X2 . The issue is that we do not know m. Jaiswal et al. again cycle over all subsets of S ′ – we know that there is a (large enough) subset of S ′ which will behave like a uniform sample from X2 . In the SSAC framework, we first identify the subset of S ′ which belongs to X2 , call this S ′′ (as above). Now we prune some points from S ′′ such that the remaining points behave like a uniform sample. This step is non-trivial because as indicated above, we do not know the value m. Instead, 2 ε and m for most of the points of X2 . Therefore, S ′′ is likely to we first show that pe lies between m εpe contain such a nice point, call it v. Now, for every point e ∈ S ′′ , we retain it with probability 2p v (which we know is at most 1). This gives a uniform sample of sufficiently large size from X2 . For k larger than 2, we generalize the above ideas using a non-trivial induction argument. 2 k-means++ within SSAC framework The k-means++ seeding algorithm, also known as the D 2 -sampling algorithm, is a simple sampling procedure that samples k centers in k iterations. The description of this algorithm is given below. The algorithm picks the first center randomly from the set X of points and after having picked the first (i − 1) centers denoted by Ci−1 , it picks a point x ∈ X to be the ith center with probability proportional to minc∈Ci−1 ||x − c||2 . The running time of k-means++ seeding algorithm is clearly O(nkd). Arthur and Vassilvitskii [AV07] showed that this simple sampling procedure gives an O(log k) approximation in expectation for any dataset. Within the SSAC framework where the algorithm is allowed to make same-cluster queries, we can make a tiny addition to the k-means++ seeding algorithm to obtain a query algorithm that gives constant factor approximation guarantee and makes only O(k2 log k) same-cluster queries. The description of the query algorithm is given in Table 1 (see right). In iteration i > 1, instead of simply accepting the sampled point x as the ith center (as done in k-means++ seeding algorithm), the sampled point x is accepted only if it belongs to a cluster other than those to which centers in Ci−1 belong (if this does not happen, k-means++(X,k) Query-k-means++(X, k) - Randomly sample a point x ∈ X - Randomly sample a point x ∈ X - C ← {x} - C ← {x} - for i = 2 to k - for i = 2 to k - Sample x ∈ X using distribution D - for j = 1 to ⌈log k⌉ defined as D(x) = Φ(C,{x}) - Sample x ∈ X using distribution D Φ(C,X) - C ← C ∪ {x} defined as D(x) = Φ(C,{x}) Φ(C,X) - return(C) - if(NewCluster(C, x)){C ← C ∪ {x}; break} - return(C) NewCluster(C, x) - If(∃c ∈ C s.t. SameCluster(c, x)) return(false) - else return(true) Table 1. k-means++ seeding algorithm (left) and its adaptation in the SSAC setting (right) the sampling is repeated for at most ⌈log k⌉ times). Here is the main result that we show for the query-k-means++ algorithm. Theorem 6. Let X ⊆ Rd be any dataset containing n points and k > 1 be a positive integer. Let C denote the output of the algorithm Query-k-means++(X, k). Then E[Φ(C, X)] ≤ 24 · ∆k (X), where ∆k (X) denotes the optimal cost for this k-means instance. Furthermore, the algorithm makes O(k2 log k) same-cluster queries and the running time of the algorithm is O(nkd + k log k log n + k2 log k). The bound on the number of same-cluster queries is trivial from the algorithm description. For the running time, it takes O(nd) time to update the distribution D which is updated k times. This accounts for the O(nkd) term in the running time. Sampling an element from a distribution D takes O(log n) time (if we maintain the cumulative distribution etc.) and at most O(k log k) points are sampled. Moreover, determining whether a sampled point belongs to an uncovered cluster takes O(k) time. So, the overall running time of the algorithm is O(nkd+k log k log n+k 2 log k). We prove the approximation guarantee in the remaining discussion. We will use the following terminology. Let the optimal k clusters for dataset X are given as X1 , ..., Xk . For any i, ∆i (X) Pdenotes the optimal cost of the i-means problem on dataset X. Given this, note that ∆k (X) = ki=1 ∆1 (Xi ). For any non-empty center set C, we say that a point x is sampled from dataset X using D 2 -sampling w.r.t. center set C if the sampling probability of x ∈ X is given by D(x) = Φ(C,{x}) Φ(C,X) . The proof of Theorem 6 will mostly follow O(log k)-approximation guarantee proof of k-means++ seeding by Arthur and Vassilvitskii [AV07]. The next two lemmas from [AV07] are crucially used in the proof of approximation guarantee. Lemma 1. Let A be any optimal cluster and let c denote a point sampled uniformly at random from A. Then E[Φ({c}, A)] ≤ 2 · ∆1 (A). Lemma 2. Let C be any arbitrary set of centers and let A be any optimal cluster. Let c be a point sampled with D 2 -sampling with respect to the center set C. Then E[Φ(C ∪{c}, A)|c ∈ A] ≤ 8·∆1 (A). The first lemma says that a randomly sampled center from X provides a good approximation (in expectation) to the cost of the cluster to which it belongs. The second lemma says that for any center set C, given that a center c that is D 2 -sampled from X w.r.t. C belong to an optimal cluster A, the conditional expectation of the cost of the cluster A with respect to center set C ∪ {c} is at most 8 times the optimal cost of cluster A. Using the above two lemmas, let us try to qualitatively see why the k-means++ seeding algorithm behave well. The first center belongs to some optimal cluster A and from Lemma 1 we know that this center is good for this cluster. At the time the ith center is D 2 -sampled, there may be some optimal clusters which are still costly with respect to the center set Ci−1 . But then we can argue that it is likely that the ith sampled center c will belong to one of these costly clusters, and conditioned on the center being from one such cluster A, the cost of this cluster after adding c to the current center set is bounded using Lemma 2. The formal proof of O(log k) approximation guarantee in [AV07] involves setting up a clever induction argument. We give a similar induction based argument to prove Theorem 6. We prove the following lemma (similar to Lemma 3.3 in [AV07]). We will need the following definitions: For any center set C, an optimal cluster A is said to be “covered” if at least one point from A is in C, otherwise A is said to be “uncovered”. Let T be a union of a subset of the optimal clusters, then we will use the def. P notation ΦOP T (T ) = Xi ⊆T ∆1 (Xi ). Lemma 3. Let C ⊆ X be any set of centers such that the number of uncovered clusters w.r.t. C is u > 0. Let Xu denote the set of points of the uncovered clusters and Xc denote set of the points of the covered clusters. Let us run t iterations of the outer for-loop in Query-k-means++ algorithm such that t ≤ u ≤ k. Let C ′ denote the resulting set of centers after running t iterations of the outer for-loop. Then the following holds:   u−t t ′ · Φ(C, Xu ). (1) + E[Φ(C , X)] ≤ (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + k u Proof. Let us begin by analysing what happens when starting with C, one iteration of the outer forloop in query-k-means++ is executed. The following two observations will be used in the induction argument: Φ(C,Xc ) c) Observation 1: If Φ(C,X Φ(C,X) ≥ 1/2, then we have Φ(C,Xc )+Φ(C,Xu ) ≥ 1/2 which implies that Φ(C, Xu ) ≤ Φ(C, Xc ), and also Φ(C, X) ≤ 2 · Φ(C, Xc ). c) Observation 2: If Φ(C,X Φ(C,X) < 1/2, then the probability that no point will be added after one iteration  ⌈log k⌉
log k c) = k1 . is given by Φ(C,X < 12 Φ(C,X) We will now proceed by induction. We show that if the statement holds for (t − 1, u) and (t − 1, u − 1), then the statement holds for (t, u). In the basis step, we will show that the statement holds for t = 0 and u > 0 and u = t = 1. Basis step: Let us first prove the simple case of t = 0 and u > 0. In this case, C ′ = C. So, we have E[Φ(C ′ , X)] = Φ(C, X) which is at most the RHS of (1). Consider the case when u = t = 1. This means that there is one uncovered cluster and one iteration of the outer for-loop is executed. If a center from the uncovered cluster is added, then E[Φ(C ′ , X)] ≤ Φ(C, Xc ) + 8 · ΦOP T (Xu ) and if no center is picked, then Φ(C ′ , X) = Φ(C, X). The probability of adding a center from the log k  c) . So, we get E[Φ(C ′ , X)] ≤ p · (Φ(C, Xc ) + 8 · uncovered cluster is given by p = 1 − Φ(C,X Φ(C,X) ΦOP T (Xu )) + (1 − p) · Φ(C, X). Note that this is upper bounded by the RHS of (1) by observing c) that 1 − p ≤ Φ(C,X Φ(C,X) . Inductive step: As stated earlier, we will assume that the statement holds for (t − 1, u) and def. c) 1 (t − 1, u − 1) and we will show that the statement holds for (t, u). Suppose p = Φ(C,X Φ(C,X) ≥ 2 , then Φ(C, X) ≤ 2 · Φ(C, Xc ) and so Φ(C ′ , X) ≤ Φ(C, X) ≤ 2 · Φ(C, Xc ) which is upper bounded by the RHS of (1). So, the statement holds for (t, u) (without even using the induction assumption). So, for the rest of the discussion, we will assume that p < 1/2. Let us break the remaining analysis into two cases – (i) no center is added in the next iteration of the outer for-loop, and (ii) a center is added. In case (i), u does not change, t decreases by 1, and the covered and uncovered clusters remain the same after the iteration. So the contribution of this case to E[Φ(C ′ , X)] is at most     t−1 u−t+1 ⌈log k⌉ p · (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + · Φ(C, Xu ) (2) + k u Now, consider case (ii). Let A be any uncovered cluster w.r.t. center set C. For any point a ∈ A, let pa denote the conditional probability of sampling a conditioned on sampling a point from A. Also, let φa denote the cost of A given a is added as a center. That is, φa = Φ(C ∪ {a}, A). The contribution of A to the expectation E[Φ(C ′ , X)] using the induction hypothesis is: (1 − p⌈log k⌉ ) · Φ(C, A) X pa Φ(C, Xu ) a ∈A     t−1 u−t · (Φ(C, Xc ) + φa + 8 · ΦOP T (Xu ) − 8 · ∆1 (A)) · 2 + · (Φ(C, Xu ) − Φ(C, A)) + k u−1 This is at most   t−1 u−t · (Φ(C, Xu ) − Φ(C, A)) (1 − p (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + + k u−1 P The previous inequality follows from the fact that a∈A pa φa ≤ 8 · ∆1 (A) from Lemma 2. Summing over all uncovered clusters, the overall contribution in case (ii) is at most:      Φ(C, Xu ) t−1 u−t · Φ(C, Xu ) − (1 − p⌈log k⌉ ) · (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + + k u−1 u ⌈log k⌉ Φ(C, A) )· Φ(C, Xu )   P The above bound is obtained using the fact that A is uncovered Φ(C, A)2 ≥ u1 Φ(C, Xu )2 . So the contribution is at most     t−1 u−t ⌈log k⌉ (1 − p ) · (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + · Φ(C, Xu ) (3) + k u Combining inequalities (2) and (3), we get the following:     t−1 Φ(C, Xu ) u−t ′ E[Φ(C , X)] ≤ (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + · Φ(C, Xu ) + p⌈log k⌉ · + k u u     t−1 u−t = (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + · Φ(C, Xu ) + + k u   Φ(C, Xc ) ⌈log k⌉ Φ(C, Xu ) · Φ(C, X) u     Φ(C, Xc ) u−t t−1 · Φ(C, Xu ) + + ≤ (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + k u ku (using the Observation 2, that is p⌈log k⌉ ≤ 1/k)   u−t t · Φ(C, Xu ) + ≤ (Φ(C, Xc ) + 8 · ΦOP T (Xu )) · 2 + k u This completes the inductive argument and the proof. ⊔ ⊓ Let us now conclude the proof of Theorem 6 using the above lemma. Consider the center set C before entering the outer for-loop. This contains a single center c chosen randomly from the dataset X. Let c belong to some optimal cluster A. Let C ′ denote the center set after the execution of the outer for-loop completes. Applying the above lemma with u = t = k − 1, we get:   k−1 ′ E[Φ(C , X)] ≤ (Φ(C, A) + 8 · ∆k (X) − 8 · ∆1 (A)) · 2 + k ≤ 3 · (2 · ∆1 (A) + 8 · ∆k (X) − 8 · ∆1 (A)) (using Lemma 1) ≤ 24 · ∆k (X) 3 Query Approximation Algorithm (proof of Theorem 3) As mentioned in the introduction, our query algorithm is based on the D 2 -sampling based algorithm of Jaiswal et al. [JKS14, JKY15]. The algorithm in these works give a (1 + ε)-factor approximation for arbitrary small ε > 0. The running time of these algorithms are of the form nd · f (k, ε), where f is an exponential function of k/ε. We now show that it is possible to get a running time which is polynomial in n, k, d, 1/ε in the SSAC model. The main ingredient in the design and analysis of the sampling algorithm is the following lemma by Inaba et al. [IKI94]. Lemma 4 ([IKI94]). Let S be a set of points obtained by independently sampling M points uniformly at random with replacement from a point set X ⊂ Rd . Then for any δ > 0,     1 Pr Φ({µ(S)}, X) ≤ 1 + · ∆1 (X) ≥ (1 − δ). δM Here µ(S) denotes the geometric centroid of the set S. That is µ(S) = P s∈S s |S| Our algorithm Query-k-means is described in Table 2. It maintains a set C of potential centers of the clusters. In each iteration of step (3), it adds one more candidate center to the set C (whp), and so, the algorithm stops when |C| reaches k. For sake of explanation, assume that optimal clusters are X1 , X2 , . . . , Xk with means µ1 , . . . , µk respectively. Consider the ith iteration of step (3). At this time |C|= i − 1, and it has good approximations to means of i − 1 clusters among X1 , . . . .Xk . Let us call these clusters covered. In Step (3.1), it samples N points, each with probability proportional to square of distance from C (D 2 -sampling). Now, it partitions this set, S, into S ∩ X1 , . . . , S ∩ Xk in the procedure UncoveredClusters, and then picks the partition with the largest size such that the corresponding optimal cluster Xj is not one of the (i − 1) covered clusters. Now, we would like to get a uniform sample from Xj – recall that S ∩Xj does not represent a uniform sample. However, as mentioned in the introduction, we need to find an element s of Xj for which the probability of being in sampled is small enough. Therefore, we pick the element in S ∩ Xj for which this probability is smallest (and we will show that it has the desired properties). The procedure UncoveredCluster returns this element s. Finally, we choose a subset T of S ∩ Xj in the procedure UniformSample. This procedure is designed such that each element of Xj has the same probability of being in T . In step (3.4), we check whether the multi-set T is of a desired minimum size. We will argue that the probability of T not containing sufficient number of points is very small. If we have T of the desired size, we take its mean and add it to C in Step (3.6). We now formally prove the approximation guarantee of the Query-k-means algorithm. Theorem 7. Let 0 < ε ≤ 1/2, k be any positive integer, and X ⊆ Rd . There exists an algorithm that runs in time O(ndk9 /ε4 ) and with probability at least 14 outputs a center set C such that Φ(C, X) ≤ (1 + ε) · ∆k (X). Moreover, the number of same-cluster queries used by the algorithm is O(k9 /ε4 ). 12 3 23 2 Constants: N = (2 ε2)k , M = 64k , L = (2 ε4)k ε Query-k-means(X, k, ε) UncoveredCluster(C, S, R) (1) R ← ∅ - For all i ∈ {1, ..., k}: Si ← ∅ (2) C ← ∅ -i←1 (3) for i = 1 to k - For all y ∈ R: {Si ← y; i++} (3.1) D2 -sample a multi-set S of N points - For all s ∈ S: from X with respect to center set C - If (∃j, y s.t. y ∈ Sj & SameCluster(s, y)) (3.2) s ← UncoveredCluster(C, S, R) - Sj ← Sj ∪ {s} (3.3) T ← UniformSample(X, C, s) - else (3.4) If (|T |< M ) continue - Let i be any index s.t. Si is empty (3.5) R ← R ∪ {s} - Si ← {s} (3.6) C ← C ∪ µ(T ) - Let Si be the largest set such that i > |R| (4) return(C) - Let s ∈ Si be the element with smallest UniformSample(X, C, s) value of Φ(C, {s}) in Si -T ←∅ - return(s) - For i = 1 to L: - D2 -sample a point x from X with respect to center set C - If (SameCluster(s, x))   - With probability ε 128 · Φ(C,{s}) Φ(C,{x}) add x in multi-set T - return(T ) Table 2. Approximation algorithm for k-means(top-left frame). Note that µ(T ) denotes the centroid of T and D2 sampling w.r.t. empty center set C means just uniform sampling. The algorithm UniformSample(X, C, s) (bottom-left) returns a uniform sample of size Ω(1/ε) (w.h.p.) from the optimal cluster to which point s belongs. Note that the success probability of the algorithm may be boosted by repeating it a constant number of times. This will also prove our main theorem (that is, Theorem 3). We will assume that the dataset X satisfies (k, ε)-irreducibility property defined next. We will later drop this assumption using a simple argument and show that the result holds for all datasets. This property was also used in some earlier works [KSS10, JKS14]. Definition 3 (Irreducibility). Let k be a positive integer and 0 < ε ≤ 1. A given dataset X ⊆ Rd is said to be (k, ε)-irreducible iff ∆k−1 (X) ≥ (1 + ε) · ∆k (X). Qualitatively, what the irreducibility assumption implies is that the optimal solution for the (k − 1)-means problem does not give a (1+ ε)-approximation to the k-means problem. The following useful lemmas are well known facts. Lemma 5. For any dataset X ⊆ Rd and a point c ∈ Rd , we have: Φ({c}, X) = Φ(µ(X), X) + |X|·||c − µ(X)||2 . Lemma 6 (Approximate Triangle Inequality). For any three points p, q, r ∈ Rd , we have ||p − q||2 ≤ 2(||p − r||2 +||r − q||2 ) Let X1 , ..., Xk be optimal clusters of the dataset X for the k-means objective. Let µ1 , ..., µk denote the corresponding optimal k centers. That is, ∀i, µi = µ(Xi ). For all i, let mi = |Xi |. Also, P for all i, let ri = 5 x∈Xi ||x−µi || mi 2 . The following useful lemma holds due to irreducibility.5 This is Lemma 4 from [JKS14]. We give the proof here for self-containment. Lemma 7. For all 1 ≤ i < j ≤ k, ||µi − µj ||2 ≥ ε · (ri + rj ). Proof. For the sake of contradiction, assume that ||µi − µj ||2 < ε · (ri + rj ). WLOG assume that mi > mj . We have: Φ({µi }, Xi ∪ Xj ) = mi ri + mj rj + mj ||µi − µj ||2 (using Lemma 5) ≤ (1 + ε) · mi ri + (1 + ε) · mj rj (since mi > mj ) ≤ mi ri + mj rj + mj · ε · (ri + rj ) ≤ (1 + ε) · Φ({µi , µj }, Xi ∪ Xj ) This implies that the center set {µ1 , ..., µk } \ {µj } gives a (1 + ε)-approximation to the k-means objective which contradicts with the (k, ε)-irreducibility assumption on the data. ⊔ ⊓ Consider the algorithm Query-k-means in Figure 2. Let Ci = {c1 , ..., ci } denote the set of centers at the end of the ith iteration of the for loop. That is, Ci is the same as variable C at the end of iteration i. We will prove Theorem 7 by inductively arguing that for every i, there are i distinct clusters for which centers in Ci are good in some sense. Consider the following invariant: P(i): There exists a set of i distinct clusters Xj1 , Xj2 , ..., Xji such that  ε · ∆1 (Xjr ). ∀r ∈ {1, ..., i}, Φ({cr }, Xjr ) ≤ 1 + 16 P ε Note that a trivial consequence of P (i) is Φ(Ci , Xj1 ∪ ... ∪ Xji ) ≤ (1 + 16 ) · ir=1 ∆1 (Xjr ). We will show that for all i, P (i) holds with probability at least (1 − 1/k)i . Note that the theorem follows if P (k) holds with probability at least (1 − 1/k)k . We will proceed using induction. The base case P (0) holds since C0 is the empty set. For the inductive step, assuming that P (i) holds with probability at least (1 − 1/k)i for some arbitrary i ≥ 0, we will show that P (i + 1) holds with probability at least (1 − 1/k)i+1 . We condition on the event P (i) (that is true with probability at least (1 − 1/k)i ). Let Ci and Xj1 , ..., Xji be as guaranteed by the invariant P (i). For ease of notation and without loss of generality, let us assume that the index jr is r. So, Ci gives a good approximation w.r.t. points in the set X1 ∪ X2 ∪ .... ∪ Xi and these clusters may be thought of as “covered” clusters (in the approximation sense). Suppose we D 2 -sample a point p w.r.t. center set Ci . The probability that p belongs to some “uncovered cluster” Xr where r ∈ [i + 1, k] is given i ,Xr ) 2 as Φ(C Φ(Ci ,X) . If this quantity is small, then the points sampled using D sampling in subsequent iterations may not be good representatives for the uncovered clusters. This may cause the analysis to break down. However, we argue that since our data is (k, ε)-irreducible, this does not occur. 6 Lemma 8. Φ(Ci ,Xi+1 ∪...∪Xk ) Φ(Ci ,X) ≥ 4ε . Proof. For the sake of contradiction, assume that the above statement does not hold. Then we have: Φ(Ci , X) = Φ(Ci , X1 ∪ ... ∪ Xi ) + Φ(Ci , Xi+1 ∪ ... ∪ Xk ) (ε/4) ≤ Φ(Ci , X1 ∪ ... ∪ Xi ) + · Φ(Ci , X1 ∪ ... ∪ Xi ) (using our assumption) 1 − (ε/4) 1 · Φ(Ci , X1 ∪ ... ∪ Xi ) = 1 − ε/4 6 This is Lemma 5 in [JKS14]. We give the proof for self-containment. ≤ i 1 + ε/16 X ∆1 (Xj ) (using invariant) · 1 − ε/4 j=1 ≤ (1 + ε) · k X ∆1 (Xj ) j=1 ⊔ ⊓ This contradicts with the (k, ε)-irreducibility of X. The following simple corollary of the above lemma will be used in the analysis later. Corollary 1. There exists an index j ∈ {i + 1, ..., k} such that Φ(Ci ,Xj ) Φ(Ci ,X) ≥ ε 4k . The above corollary says that there is an uncovered cluster which will have a non-negligible representation in the set S that is sampled in iteration (i + 1) of the algorithm Query-k-means. The next lemma shows that conditioned on sampling from an uncovered cluster l ∈ {i + 1, ..., k}, ε times its sampling probability if it were the probability of sampling a point x from Xl is at least 64 ε · m1l ). 7 sampled uniformly from Xl (i.e., with probability at least 64 Lemma 9. For any l ∈ {i + 1, ..., k} and x ∈ Xl , Φ(Ci ,{x}) Φ(Ci ,Xl ) ≥ ε 64 · 1 ml . Proof. Let t ∈ {1, ..., i} be the index such that x is closest to ct among all centers in Ci . We have: Φ(Ci , Xl ) = ml · rl + ml · ||µl − ct ||2 (using Lemma 5) 2 ≤ ml · rl + 2ml · (||µl − µt || +||µt − ct ||2 ) (using Lemma 6) ε ≤ ml · rl + 2ml · (||µl − µt ||2 + · rt ) (using invariant and Lemma 5) 16 Also, we have: ||x − µt ||2 − ||µt − ct ||2 (using Lemma 6) 2 ||µl − µt ||2 ≥ − ||µt − ct ||2 (since ||x − µt ||≥ ||µl − µt ||/2) 8 ε ||µl − µt ||2 − · rt (using invariant and Lemma 5) ≥ 8 16 ||µl − µt ||2 ≥ (using Lemma 7) 16 Φ(Ci , {x}) = ||x − ct ||2 ≥ Combining the inequalities obtained above, we get the following: ||µl − µt ||2 Φ(Ci , {x})
≥ Φ(Ci , Xl ) 16 · ml · rl + 2||µl − µt ||2 + 8ε · rt 1 1 ε 1 ≥ · ≥ · 16 · ml (1/ε) + 2 + (1/8) 64 ml This completes the proof of the lemma. ⊔ ⊓ With the above lemmas in place, let us now get back to the inductive step of the proof. Let J ⊆ {i + 1, ..., k} denote the subset of indices (from the uncovered cluster indices) such that 7 This is Lemma 6 from [JKS14]. We give the proof for self-containment. Φ(Ci ,Xj ) ε Φ(Ci ,X) ≥ 8k . For any 2 i ,{y}) Yj , Φ(C Φ(Ci ,Xj ) ≤ mj . That ∀j ∈ J, index j ∈ J, let Yj ⊆ Xj denote the subset of points in Xj such that is, Yj consists of all the points such that the conditional probability ∀y ∈ of sampling any point y in Yj , given that a point is sampled from Xj , is upper bounded by 2/mj . Note that from Lemma 9, the conditional probability of sampling a point x from Xj , given that a ε · m1j . This gives the following simple and useful point is sampled from Xj , is lower bounded by 64 lemma. Lemma 10. For all j ∈ {i + 1, ..., k} the following two inequalities hold: Φ(C ,Y ) Φ(C ,X ) j j ε 1. Φ(Cii ,X) · Φ(Cii ,X) ≥ 128 , and 2. For any y ∈ Yj and any x ∈ Xj , ε 128 · Φ(Ci , {y}) ≤ Φ(Ci , {x}). Φ(Ci ,{y}) ε Φ(Ci ,Xj ) ≥ 64 1 ε i ,{x}) Xj , Φ(C Φ(Ci ,Xj ) ≥ 64 · mj Proof. Inequality (1) follows from the fact that |Yj |≥ mj /2, and Xj . Inequality (2) follows from the fact that for all x ∈ i ,{y}) Yj , Φ(C Φ(Ci ,Xj ) ≤ 2 mj . · 1 mj for all y ∈ and for all y ∈ ⊔ ⊓ Let us see the outline of our plan before continuing with our formal analysis. What we hope to get in line (3.2) of the algorithm is a point s that belongs to one of the uncovered clusters with index in the set J. That is, s belongs to an uncovered cluster that is likely to have a good representation in the D 2 -sampled set S obtained in line (3.1). In addition to s belonging to Xj for some j ∈ J, we would like s to belong to Yj . This is crucial for the uniform sampling in line (3.3) to succeed. We will now show that the probability of s returned in line (3.2) satisfies the above conditions is large. Lemma 11. Let S denote the D 2 -sample obtained w.r.t. center set Ci in line (3.1) of the algorithm. Pr[∃j ∈ J such that S does not contain any point from Yj ] ≤ 1 . 4k Proof. We will first get bound on the probability for a fixed j ∈ J and then use the union bound. Φ(C ,Yj ) ε ε2 ε From property (1) of Lemma 10, we have that for any j ∈ J, Φ(Cii ,X) · 8k = (210 ≥ 128 . Since the )k 12 3 number of sampled points is N = (2 ε2)k , we get that the probability that S has no points from Yj is at most 4k12 . Finally, using the union bound, we get the statement of the lemma. ⊔ ⊓ Lemma 12. Let S denote the D 2 -sample obtained w.r.t. center set Ci in line (3.1) of the algorithm and let Sj denote the representatives of Xj in S. Let max = arg maxj∈{i+1,...,k}|Sj |. Then Pr[max ∈ / 1 J] ≤ 4k . Φ(C ,X ) j ε Proof. From Corollary 1, we know that there is an index j ∈ {i + 1, ..., k} such that Φ(Cii ,X) . ≥ 4k ε Let α = N · 4k . The expected number of representatives from Xj in S is at least α. So, from Chernoff bounds, we have: Pr[|Sj |≤ 3α/4] ≤ e−α/32 On the other hand, for any r ∈ {i + 1, ..., k} \ J, the expected number of points in S from Xr is at ε · N = α/2. So, from Chernoff bounds, we have: most 8k Pr[|Sr |> 3α/4] ≤ e−α/24 So, the probability that there exists such an r is at most k · e−α/24 by union bound. Finally, the probability that max ∈ / J is at most Pr[|Sj |≤ 3α/4] + Pr[∃r ∈ {i + 1, ..., k} \ J||Sr |> 3α/4] which 12 3 1 ⊔ ⊓ is at most 4k due to our choice of N = (2 ε2)k . 1 From the previous two lemmas, we get that with probability at least (1 − 2k ), the s returned in line (3.2) belongs to Yj for some j ∈ J. Finally, we will need the following claim to argue that the set T returned in line (3.3) is a uniform sample from one of the sets Xj for j ∈ {i + 1, ..., k}. Lemma 13. Let S denote the D 2 -sample obtained w.r.t. center set Ci in line (3.1) and s be the point returned in line (3.2) of the algorithm. Let j denote the index of the cluster to which s belongs. 1 ), T returned in line (3.3) is a uniform If j ∈ J and s ∈ Yj , then with probability at least (1 − 4k 64k sample from Xj with size at least ε . Proof. Consider the call to sub-routine UniformSample(X, Ci, s) with s as given in the statement of the lemma. If j is the index of the cluster to which s belongs, then j ∈ J and s ∈ Yj . Let us define L random variables Z1 , ..., ZL one for every iteration of the sub-routine. These random variables are defined as follows: for any r ∈ [1, L], if the sampled point x belongs to the same cluster as s and it is picked to be included in multi-set S, then Zr = x, otherwise Zr = ⊥. We first note that for any r and any x, y ∈ Xj , Pr[Zr = x] = Pr[Zr = y]. This is because for any x ∈ Xj , we have ε ·Φ(C ,{s}) ε ·Φ(C ,{s}) i i ε i ,{s}) i ,{x}) 128 128 · Φ(C Pr[Zr = x] = Φ(C = 128 ≤1 Φ(Ci ,X) · Φ(Ci ,{x}) Φ(Ci ,X) . It is important to note that Φ(Ci ,{x}) from property (2) of Lemma 10 and hence valid in the probability calculations above. Let us now obtain a bound on the size of T . Let Tr = I(Zr ) be the indicator variable that is 1 if Zr 6= ⊥ and 0 otherwise. Using the fact that j ∈ J, we get that for any r: P ε ε ε ε ε3 x∈Xj Φ(Ci , {s}) E[Tr ] = Pr[Tr = 1] = · ≥ · · = 16 . 128 Φ(Ci , X) 128 8k 64 (2 )k Given that L = 223 k 2 , ε4 applying Chernoff bounds, we get the following:      64k 1 64k = 1 − Pr |T |≤ ≥ 1− Pr |T |≥ ε ε 4k  This completes the proof of the lemma. ⊔ ⊓ Since a suitable s (as required by the above lemma) is obtained in line (3.2) with probability at 1 least (1− 2k ), the probability that T obtained in line (3.3) is a uniform sample from some uncovered 1 1 )·(1− 4k ). Finally, the probability that the centroid µ(T ) of the multi-set cluster Xj is at least (1− 2k 1 ) using Inaba’s lemma. Combining T that is obtained is a good center for Xj is at least (1 − 4k 1 everything, we get that with probability at least (1 − k ) an uncovered cluster will be covered in the ith iteration. This completes the inductive step and hence the approximation guarantee of (1 + ε) holds for any dataset that satisfies the (k, ε)-irreducibility assumption. For the number of queries and running time, note that every time sub-routine UncoveredCluster is called, it uses at most kN same cluster queries. For the subroutine UniformSample, the number of same-cluster queries made is L. So, the total number of queries is O(k(kN + L)) = O(k5 /ε4 ). More specifically, we have proved the following theorem. Theorem 8. Let 0 < ε ≤ 1/2, k be any positive integer, and X ⊆ Rd such that X is (k, ε)irrducible. Then Query-k-means(X, k, ε) runs in time O(ndk5 /ε4 ) and with probability at least (1/4) outputs a center set C such that Φ(C, X) ≤ (1 + ε) · ∆k (X). Moreover, the number of samecluster queries used by Query-k-means(X, k, ε) is O(k5 /ε4 ). To complete the proof of Theorem 7, we must remove the irreducibility assumption in the above theorem. We do this by considering the following two cases: ε 1. Dataset X is (k, (4+ε/2)k )-irreducible. ε 2. Dataset X is not (k, (4+ε/2)k )-irreducible. In the former case, we can apply Theorem 8 to obtain Theorem 7. Now, consider the latter ε case. Let 1 < i ≤ k denote the largest index such that X is (i, (1+ε/2)k )-irreducible, otherwise i = 1. Then we have: k−i   ε ε · ∆k (X). · ∆k (X) ≤ 1 + ∆i (X) ≤ 1 + (4 + ε/2)k 4 This means that a (1 + ε/4)-approximation for the i-means problem on the dataset X gives the desired approximation for the k-means problem. Note that our approximation analysis works for the i-means problem with respect to the algorithm being run only for i steps in line (3) (instead of k). That is, the centers sampled in the first i iterations of the algorithm give a (1+ε/16)-approximation for the i-means problem for any fixed i. This simple observation is sufficient for Theorem 7. Note since Theorem 8 is used with value of the error parameter as O(ε/k), the bounds on the query and running time get multiplied by a factor of k4 . 4 Query Lower Bound (proof of Theorem 4) In this section, we will obtain a conditional lower bound on the number of same-cluster queries assuming the Exponential Time Hypothesis (ETH). This hypothesis has been used to obtain lower bounds in various different contexts (see [Man16] for reference). We start by stating the Exponential Time Hypothesis (ETH). Hypothesis 1 (Exponential Time Hypothesis (ETH)[IP01, IPZ01]): There does not exist an algorithm that can decide whether any 3-SAT formula with m clauses is satisfiable with running time 2o(m) . Since we would like to obtain lower bounds in the approximation domain, we will need a gap version of the above ETH hypothesis. The following version of the PCP theorem will be very useful in obtaining a gap version of ETH. Theorem 9 (Dinur’s PCP Theorem [Din07]). For some constants ε, d > 0, there exists a polynomial time reduction that takes a 3-SAT formula ψ with m clauses and produces another 3-SAT formula φ with m′ = O(m polylog m) clauses such that: – If ψ is satisfiable, then φ is satisfiable, – if ψ is unsatisfiable, then val(φ) ≤ 1 − ε, and – each variable of φ appears in at most d clauses. Here val(φ) denotes the maximum fraction of clauses of φ that are satisfiable by any assignment. The following new hypothesis follows from ETH and will be useful in our analysis. Hypothesis 2: There exists constants ε, d > 0 such that the following holds: There does not exist an algorithm that, given a 3-SAT formula ψ with m clauses with each variable appearing in atmost d clauses, distinguishes whether ψ is satisfiable or val(ψ) ≤ (1 − ε), Ω runs in time 2 m poly log m . The following simple lemma follows from Dinur’s PCP theorem given above. Lemma 14. If Hypothesis 1 holds, then so does Hypothesis 2. We now see a reduction from the gap version of 3-SAT to the gap version of the Vertex Cover problem that will be used to argue the hardness of the k-means problem. The next result is a standard reduction and can be found in a survey by Luca Trevisan [Tre04]. Lemma 15. Let ε, d > 0 be some constants. There is a polynomial time computable function mapping 3-SAT instances ψ with m variables and where each variable appears in at most d clauses, into graphs Gψ with 3m vertices and maximum degree O(d) such that if ψ is satisfiable, then Gψ has a vertex cover of size at most 2m and if val(ψ) ≤ (1 − ε), then every vertex cover of Gψ has size at least 2m(1 + ε/2). We formulate the following new hypothesis that holds given that hypothesis 2 holds. Eventually, we will chain all these hypothesis together. Hypothesis 3: There exists constants ε, d > 0 such that the following holds: There does not exist an algorithm that, given a graph G with n vertices and maximum degree d, distinguishes between the case when G has a vertex cover of size at most 2n/3 and the case when G has  a vertex cover of size at least 2n 3 Ω · (1 + ε), runs in time 2 n poly log n . The following lemma is a simple implication of Lemma 15 Lemma 16. If Hypothesis 2 holds, then so does Hypothesis 3. We are getting closer to the k-means problem that has a reduction from the vertex cover problem on triangle-free graphs [ACKS15]. So, we will need reductions from vertex cover problem to vertex cover problem on triangle-free graphs and then to the k-means problem. These two reductions are given by Awasthi et al. [ACKS15]. Lemma 17 (Follows from Theorem 21 [ACKS15]). Let ε, d > 0 be some constants. There is a polynomial-time computable function mapping any graph G = (V, E) with maximum degree d to a triangle-free graph Ĝ = (V̂ , Ê) such that the following holds: 3 2 –  |V̂ |= poly(d,1/ε)· |V | and maximum degree   of verticesin Ĝ is O(d /ε ), and Ĝ)| ≤ 1 − |V C( . – 1 − |V C(G)| ≤ (1 + ε) · 1 − |V C(G)| |V | |V | |V̂ | Here V C(G) denote the size of the minimum vertex cover of G. We can formulate the following hypothesis that will follow from Hypothesis 3 using the above lemma. Hypothesis 4: There exists constants ε, d > 0 such that the following holds: There does not exist an algorithm that, given a triangle-free graph G with n vertices and maximum 2n degree d, distinguishes between the case when G has a vertex cover of size at most 3 and  the case when G has a vertex cover of size at least 2n 3 Ω · (1 + ε), runs in time 2 The next claim is a simple application of Lemma 17. Lemma 18. If Hypothesis 3 holds, then so does Hypothesis 4. n poly log n . Finally, we use the reduction from the vertex cover problem in triangle-free graphs to the kmeans problem to obtain the hardness result for the k-means problem. We will use the following reduction from Awasthi et al. [ACKS15]. Lemma 19 (Theorem 3 [ACKS15]). There is an efficient reduction from instances of Vertex Cover (in triangle free graphs) to those of k-means that satisfies the following properties: – if the Vertex Cover instance has value k, then the k-means instance has cost at most (m − k) – if the Vertex Cover instance has value at least k(1 + ε), then the optimal k-means cost is at least m − (1 − Ω(ε))k. Here ε is some fixed constant > 0. Here m denotes the number of edges in the vertex cover instance. The next hypothesis follows from Hypothesis 4 due to the above lemma. Hypothesis 5: There exists constant c > 1 such that the following holds: There does not exist an algorithm that gives guarantee of c for the k-means problem that  an approximation  Ω runs in time poly(n, d) · 2 k poly log k . Claim. If Hypothesis 4 holds, then so does Hypothesis 5. Now using Lemmas 14, 16, 18, and 4, get the following result. Lemma 20. If the Exponential Time Hypothesis (ETH) holds then there exists a constant c > 1 such that any c-approximation algorithm for the k-means problem cannot have running time better   Ω than poly(n, d) · 2 k poly log k . This proves Theorem 4 since if there is a query algorithm that runs in time poly(n, d, k) and makes polyklog k same-cluster queries, then we can convert it to a non-query algorithm that runs in k time poly(n, d) · 2 poly log k in a brute-force manner by trying out all possible answers for the queries and then picking the best k-means solution. 5 Query Approximation Algorithm with Faulty Oracle In this section, we describe how to extend our approximation algorithm for k-means clustering in the SSAC framework when the oracle is faulty. That is, the answer to the same-cluster query may be incorrect. Let us denote the faulty same-cluster oracle as OE . We consider the following error model: for a query with points u and v, the query answer OE (u, v) is wrong independently with probability at most q that is strictly less than 1/2. More specifically, if u and v belong to the same optimal cluster, then OE (u, v) = 1 with probability at least (1 − q) and OE (u, v) = 0 with probability at most q. Similarly, if u and v belong to different optimal clusters, then OE (u, v) = 1 with probability at most q and OE (u, v) = 0 with probability at least (1 − q). The modified algorithm giving (1 + ε)-approximation for k-means with faulty oracle OE is given in Figure 3. Let X1 , . . . , Xk denote the k optimal clusters for the dataset X. Let C = {c1 , . . . , ci } denote the set of i centers chosen by the algorithm at the end of iteration i. Let S denote the sample obtained using D 2 -sampling in the (i + 1)st iteration. The key idea for an efficient (1 + ε)approximation algorithm for k-means in the SSAC framework with a perfect oracle was the following. Given a sample S, we could compute using at most k|S| same-cluster queries the partition S1 , . . . , Sk of S among the k optimal clusters such that Sj = S ∩ Xj for all j. In the following, we discuss how we achieve this partitioning of S among k optimal clusters when the oracle OE is faulty. We reduce the problem of finding the partitions of S among the optimal clusters to the problem of recovering dense (graph) clusters in a stochastic block model (SBM). An instance of an SBM is created as follows. Given any arbitrary partition V1 , . . . , Vk of a set V of vertices, an edge is added between two vertices belonging to the same partition with probability at least (1 − q) and between two vertices in different partitions with probability at most q. We first construct an instance I of an SBM using the sample S. By querying the oracle OE with all pairs of points u, v in S, we obtain a graph I on |S| vertices, where vertices in I correspond to the points in S, and an edge exists in I between vertices u and v if OE (u, v) = 1. Since oracle OE errs with probability at most q, for any u, v ∈ Sj for some j ∈ [k], there is an edge between u and v with probability at least (1 − q). Similarly, there is an edge (u, v) ∈ I for any two points u ∈ Sy and v ∈ Sz , y 6= z belonging to different optimal clusters with probability at most q. Note that the instance I created in this manner would be an instance of an SBM. Since q < 1/2, this procedure, with high probability, creates more edges between vertices belonging to the same partition than the number of edges between vertices in different partitions. Intuitively, the partitions of S would correspond to the dense (graph) clusters in SBM instance I, and if there were no errors, then each partition would correspond to a clique in I. One way to figure out the partitions S1 , . . . , Sk would be to retrieve the dense (graph) clusters from the instance I. Ailon et al. [ACX15] gave a randomized algorithm to retrieve all large clusters of any SBM instance. Their algorithm on a graph of n vertices retrieves √ all clusters of size at least n with high probability. Their main result in our context is given as follows. 13 3 23 2 Constants: N = (2 ε2)k , M = 64k , L = (2 ε4)k ε Faulty-Query-k-means(X, k, ε) UncoveredCluster(C, S, J) (1) J ← ∅ - For all i ∈ {1, ..., k}: Si ← ∅ (2) C ← ∅ -i←1 (3) for i = 1 to k - For all y ∈ J: {Si ← y; i++} (3.1) D2 -sample a multi-set S of N points - T1 , . . . , Tl = PartitionSample(S) for l < k from X with respect to center set C - for j = 1, . . . , l (3.2) s ← UncoveredCluster(C, S, J) - if IsCovered(C, Tj ) is FALSE (3.3) T ← UniformSample(X, C, s) - if ∃t such that St = ∅, then St = Tj (3.4) If (|T |< M ) continue - Let Si be the largest set such that i > |J| (3.5) J ← J ∪ {s} - Let s ∈ Si be the element with smallest (3.6) C ← C ∪ µ(T ) value of Φ(C, {s}) in Si (4) return(C) - return(s) UniformSample(X, C, s) -S←∅ - For i = 1 to L: PartitionSample(S) - D2 -sample point x ∈ X with respect to center set C - Construct SBM instance I by querying OE (s, t) ∀s, t ∈ S - U = U ∪ {x} - Run cluster recovery algorithm of Ailon et al. [ACX15] on I - T1 , . . . , Tl = PartitionSample(U ) for l < k - Return T1 , . . . , Tl for l < k - for j = 1, . . . , l - If (IsCovered(s, Tj ) is TRUE)   - ∀x ∈ Tj , with probability in multi-set S - return (S) ε 128 · Φ(C,{s}) Φ(C,{x}) add x IsCovered(C,U) - for c ∈ C – if for majority of u ∈ U , OE (c, u) = 1, Return TRUE - Return FALSE Table 3. Approximation algorithm for k-means (top-left frame) using faulty oracle. Note that µ(T ) denotes the centroid of T and D2 -sampling w.r.t. empty center set C means just uniform sampling. The algorithm UniformSample(X, C, s) (bottom-left) returns a uniform sample of size Ω(1/ε) (w.h.p.) from the optimal cluster in which point s belongs. Lemma 21 ([ACX15]). There exists a polynomial time algorithm that, given an instance of a √ stochastic block model on n vertices, retrieves all clusters of size at least Ω( n) with high probability, provided q < 1/2. We use Lemma 21 to retrieve the large clusters from our SBM instance I. We also need to make sure that the sample S is such that its overlap with at least one uncovered optimal cluster is large, where an optimal cluster Sj for some j is uncovered if C ∩ Sj = ∅. More formally, we would require p the following: ∃j ∈ [k] such that |Sj |≥ Ω( |S|), and Xj is uncovered by C. From Corollary 1, given a set of centers C with |C|< k, there exists an uncovered cluster such that any point sampled ε using D 2 -sampling would belong to that uncovered cluster with probability at least 4k . Therefore, ε 2 in expectation, D -sample S would contain at least 4k |S| points from one such uncovered optimal p 2 cluster. In order to ensure that this quantity is at least as large as |S|, we need |S|= Ω( 16k ε2 ). Our bounds for N and L, in the algorithm, for the size of D 2 -sample S satisfy this requirement with high probability. This follows from a simple application of Chernoff bounds. 12 2 Lemma 22. For D 2 -sample S of size at least 2 ε2k , there is at least one partition Sj = S ∩ Xj among the partitions returned by the sub-routine PartitionSample corresponding to an uncovered 1 cluster Xj with probability at least (1 − 16k ). Proof. From Corollary 1, for any point p sampled using D 2 -sampling, the probability that point p ε . In expectation, the number of points sampled belongs to some uncovered cluster Xj is at least 4k 10 ε|S| 2 k from uncovered cluster Xj is E[|Sj |] = 4k = ε . Exact recovery using Lemma 21 requires |Sj | to 6 1 be at least 2 εk . Using Chernoff bounds, the probability of this event is at least (1 − 16k ). ⊔ ⊓ Following Lemma 22, we condition on the event that there is at least one partition corresponding to an uncovered cluster among the partitions returned by the sub-routine PartitionSample. Next, we figure out using the sub-routine IsCovered which of the partitions returned by PartitionSample are covered and which are uncovered. Let T1 , . . . , Tl be the partitions returned by PartitionSample where l < k. Sub-routine IsCovered determines whether a cluster is covered or uncovered in the following manner. For each j ∈ [l], we check whether Tj is covered by some c ∈ C. We query oracle OE with pairs (v, c) for v ∈ Tj and c ∈ C. If majority of the query answers for some c ∈ C is 1, we say cluster Tj is covered by C. If for all c ∈ C and some Tj , the majority of the query answers is 0, then we say Tj is uncovered by C. Using Chernoff bounds, we show that with high probability uncovered clusters would be detected. Lemma 23. With probability at least (1 − correctly by the sub-routine IsCovered. 1 16k ), all covered and uncovered clusters are detected Proof. First, we figure out the probability that any partition Tj for j ∈ [l] is detected correctly as covered or uncovered. Then we use union bound to bound the probability that all clusters are detected correctly. Recall that each partition returned by PartitionSample has size at least 6 |Tj | ≥ 2 εk for j ∈ [l]. We first compute for one such partition Tj and some center c ∈ C, the probability that majority of the queries OE (v, c) where v ∈ Tj are wrong. Since each query answer is wrong independently with probability q < 1/2, in expectation the number of wrong query answers would be q|Tj |. Using Chernoff bound, the probability that majority of the queries is wrong is at most e− 1 2 26 k (1− 2q ) 3ε . The probability that the majority of the queries is wrong for at least one center 6 1 2 − 23εk (1− 2q ) c ∈ C is at most ke probability at least (1 − 6 . Again using union bound all clusters are detected correctly with 1 2 2 k k2 e− 3ε (1− 2q ) ) ≥ (1 − 1 16k ). ⊔ ⊓ 1 With probability at least (1 − 8k ), given a D 2 -sample S, we can figure out the largest uncovered optimal cluster using the sub-routines PartitionSample and IsCovered. The analysis of the Algorithm 3 follows the analysis of Algorithm 2. For completeness, we compute the probability of success, and the query complexity of the algorithm. Note that s in line (3.2) of the Algorithm 3 1 1 is chosen correctly with probability (1 − 4k )(1 − 8k ). The uniform sample in line (3.3) is chosen 1 1 properly with probability (1 − 4k )(1 − 8k ). Since, given the uniform sample, success probability 1 ), overall the probability of success becomes (1 − k1 ). For using Inaba’s lemma is at least (1 − 4k 6 query complexity, we observe that PartitionSample makes O( kε8 ) same-cluster queries to oracle 4 OE , query complexity of IsCovered is O( kε4 ). Since PartitionSample is called at most k times, 7 total query complexity would be O( kε8 ). Note that these are bounds for dataset that satisfies (k, ε)irreducibility condition. For general dataset, we will use O(ε/k) as the error parameter. This causes the number of same-cluster queries to be O(k15 /ε8 ). 6 Conclusion and Open Problems This work explored the power of the SSAC framework defined by Ashtiani et al. [AKBD16] in the approximation algorithms domain. We showed how same-cluster queries allowed us to convert the popular k-means++ seeding algorithm into an algorithm that gives constant approximation for the k-means problem instead of the O(log k) approximation guarantee of k-means++ in the absence of such queries. Furthermore, we obtained an efficient (1+ε)-approximation algorithm for the k-means problem within the SSAC framework. This is interesting because it is known that such an efficient algorithm is not possible in the classical model unless P = NP. Our results encourages us to formulate similar query models for other hard problems. If the query model is reasonable (as is the SSAC framework for center-based clustering), then it may be worthwhile exploring its powers and limitations as it may be another way of circumventing the hardness of the problem. For instance, the problem closest to center-based clustering problems such as k-means is the correlation clustering problem. The query model for this problem may be similar to the SSAC framework. It will be interesting to see if same-cluster queries allows us to design efficient algorithms and approximation algorithms for the correlation clustering problem for which hardness results similar to that of k-means is known. References ABS12. ABV14. ACKS15. ACX15. ADK09. AKBD16. Pranjal Awasthi, Avrim Blum, and Or Sheffet. Center-based clustering under perturbation stability. Information Processing Letters, 112(1):49 – 54, 2012. Pranjal Awasthi, Maria-Florina Balcan, and Konstantin Voevodski. Local algorithms for interactive clustering. In ICML, pages 550–558, 2014. Pranjal Awasthi, Moses Charikar, Ravishankar Krishnaswamy, and Ali Kemal Sinop. The Hardness of Approximation of Euclidean k-Means. In Lars Arge and János Pach, editors, 31st International Symposium on Computational Geometry (SoCG 2015), volume 34 of Leibniz International Proceedings in Informatics (LIPIcs), pages 754–767, Dagstuhl, Germany, 2015. Schloss Dagstuhl–Leibniz-Zentrum fuer Informatik. Nir Ailon, Yudong Chen, and Huan Xu. Iterative and active graph clustering using trace norm minimization without cluster size constraints. Journal of Machine Learning Research, 16:455–490, 2015. Ankit Aggarwal, Amit Deshpande, and Ravi Kannan. Adaptive sampling for k-means clustering. In Irit Dinur, Klaus Jansen, Joseph Naor, and Jos Rolim, editors, Approximation, Randomization, and Combinatorial Optimization. Algorithms and Techniques, volume 5687 of Lecture Notes in Computer Science, pages 15–28. Springer Berlin Heidelberg, 2009. Hassan Ashtiani, Shrinu Kushagra, and Shai Ben-David. Clustering with same-cluster queries. In NIPS. 2016. ANFSW16. Sara Ahmadian, Ashkan Norouzi-Fard, Ola Svensson, and Justin Ward. Better guarantees for k-means and euclidean k-median by primal-dual algorithms. arXiv preprint arXiv:1612.07925, 2016. AV07. David Arthur and Sergei Vassilvitskii. k-means++: the advantages of careful seeding. In Proceedings of the eighteenth annual ACM-SIAM symposium on Discrete algorithms, SODA ’07, pages 1027–1035, Philadelphia, PA, USA, 2007. Society for Industrial and Applied Mathematics. BB08. Maria-Florina Balcan and Avrim Blum. Clustering with interactive feedback. In International Conference on Algorithmic Learning Theory, pages 316–328. Springer Berlin Heidelberg, 2008. BBG09. Maria-Florina Balcan, Avrim Blum, and Anupam Gupta. Approximate clustering without the approximation. In Proc. ACM-SIAM Symposium on Discrete Algorithms, pages 1068–1077, 2009. BBM04. Sugato Basu, Arindam Banerjee, and Raymond J. Mooney. Active semi-supervision for pairwise constrained clustering. In Proceedings of the Fourth SIAM International Conference on Data Mining, Lake Buena Vista, Florida, USA, April 22-24, 2004, pages 333–344, 2004. BJA16. Anup Bhattacharya, Ragesh Jaiswal, and Nir Ailon. Tight lower bound instances for k-means++ in two dimensions. Theor. Comput. Sci., 634(C):55–66, June 2016. BR13. Tobias Brunsch and Heiko Rglin. A bad instance for k-means++. Theoretical Computer Science, 505:19 – 26, 2013. CAKM16. Vincent Cohen-Addad, Philip N. Klein, and Claire Mathieu. Local search yields approximation schemes for k-means and k-median in euclidean and minor-free metrics. 2016 IEEE 57th Annual Symposium on Foundations of Computer Science (FOCS), 00:353–364, 2016. Das08. Sanjoy Dasgupta. The hardness of k-means clustering. Technical Report CS2008-0916, Department of Computer Science and Engineering, University of California San Diego, 2008. Din07. Irit Dinur. The pcp theorem by gap amplification. J. ACM, 54(3), June 2007. FMS07. Dan Feldman, Morteza Monemizadeh, and Christian Sohler. A PTAS for k-means clustering based on weak coresets. In Proceedings of the twenty-third annual symposium on Computational geometry, SCG ’07, pages 11–18, New York, NY, USA, 2007. ACM. FRS16. Zachary Friggstad, Mohsen Rezapour, and Mohammad R. Salavatipour. Local search yields a PTAS for k-means in doubling metrics. 2016 IEEE 57th Annual Symposium on Foundations of Computer Science (FOCS), 00:365–374, 2016. IKI94. Mary Inaba, Naoki Katoh, and Hiroshi Imai. Applications of weighted Voronoi diagrams and randomization to variance-based k-clustering: (extended abstract). In Proceedings of the tenth annual symposium on Computational geometry, SCG ’94, pages 332–339, New York, NY, USA, 1994. ACM. IP01. Russell Impagliazzo and Ramamohan Paturi. On the complexity of k-sat. Journal of Computer and System Sciences, 62(2):367 – 375, 2001. IPZ01. Russell Impagliazzo, Ramamohan Paturi, and Francis Zane. Which problems have strongly exponential complexity? Journal of Computer and System Sciences, 63(4):512 – 530, 2001. JKS14. Ragesh Jaiswal, Amit Kumar, and Sandeep Sen. A simple D2 -sampling based PTAS for k-means and other clustering problems. Algorithmica, 70(1):22–46, 2014. JKY15. Ragesh Jaiswal, Mehul Kumar, and Pulkit Yadav. Improved analysis of D2 -sampling based PTAS for k-means and other clustering problems. Information Processing Letters, 115(2):100 – 103, 2015. KSS10. Amit Kumar, Yogish Sabharwal, and Sandeep Sen. Linear-time approximation schemes for clustering problems in any dimensions. J. ACM, 57(2):5:1–5:32, February 2010. LSW17. Euiwoong Lee, Melanie Schmidt, and John Wright. Improved and simplified inapproximability for kmeans. Information Processing Letters, 120:40 – 43, 2017. Man16. Pasin Manurangsi. Almost-polynomial ratio ETH-hardness of approximating densest k-subgraph. CoRR, abs/1611.05991, 2016. MNV12. Meena Mahajan, Prajakta Nimbhorkar, and Kasturi Varadarajan. The planar k-means problem is NPhard. Theor. Comput. Sci., 442:13–21, July 2012. ORSS13. Rafail Ostrovsky, Yuval Rabani, Leonard J. Schulman, and Chaitanya Swamy. The effectiveness of lloyd-type methods for the k-means problem. J. ACM, 59(6):28:1–28:22, January 2013. Tre04. Luca Trevisan. Inapproximability of combinatorial optimization problems. CoRR, cs.CC/0409043, 2004. Vat09. Andrea Vattani. The hardness of k-means clustering in the plane. Technical report, Department of Computer Science and Engineering, University of California San Diego, 2009. VBR+ 14. Konstantin Voevodski, Maria-Florina Balcan, Heiko Röglin, Shang-Hua Teng, and Yu Xia. Efficient clustering with limited distance information. CoRR, abs/1408.2045, 2014. VD16. Sharad Vikram and Sanjoy Dasgupta. Interactive bayesian hierarchical clustering. In Proceedings of the 33nd International Conference on Machine Learning, ICML 2016, New York City, NY, USA, June 19-24, 2016, pages 2081–2090, 2016. 

New avenue to the Parton Distribution Functions: Self-Organizing Maps arXiv:0810.2598v2 [hep-ph] 2 Nov 2008 H. Honkanena,b,∗ , S. Liutia,† a Department of Physics, University of Virginia, P.O. Box 400714, Charlottesville, VA 22904-4714, USA b Department of Physics and Astronomy, Iowa State University, Ames, IA 50011, USA J. Carnahanc , Y. Loitierec , P. R. Reynoldsc c Department of Computer Science, School of Engineering, University of Virginia, P.O. Box 400740 Charlottesville, VA 22904-4740, USA Abstract Neural network algorithms have been recently applied to construct Parton Distribution Function (PDF) parametrizations which provide an alternative to standard global fitting procedures. We propose a technique based on an interactive neural network algorithm using Self-Organizing Maps (SOMs). SOMs are a class of clustering algorithms based on competitive learning among spatially-ordered neurons. Our SOMs are trained on selections of stochastically generated PDF samples. The selection criterion for every optimization iteration is based on the features of the clustered PDFs. Our main goal is to provide a fitting procedure that, at variance with the standard neural network approaches, allows for an increased control of the systematic bias by enabling user interaction in the various stages of the process. PACS numbers: 13.60.Hb, 12.38.Bx, 84.35.+i ∗ † heli@iastate.edu sl4y@virginia.edu 1 Introduction Modelling experimental data always introduces bias, in the form of either a theoretical or systematical bias. The former is introduced by researchers with the precise structure of the model they use, which invariably constrains the form of the solutions. The latter form of bias is introduced by algorithms, such as optimization algorithms, which may favour some results in ways which are not justified by their objective functions, but rather depend on the internal operation of the algorithm. In this paper we concentrate on high energy hadronic interactions, which are believed to be described by Quantum Chromodynamics (QCD). Because of the properties of factorization and asymptotic freedom of the theory, the cross sections for a wide number of hadronic reactions can be computed using perturbation theory, as convolutions of perturbatively calculable hard scattering coefficients, with non perturbative Parton Distribution Functions (PDFs) that parametrize the large distance hadronic structure. The extraction of the PDFs from experiment is inherently affected by a bias, which ultimately dictates the accuracy with which the theoretical predictions can be compared to the high precision measurements of experimental observables. In particular, the form of bias introduced by PDFs will necessarily impact the upcoming searches of physics beyond the Standard Model at the Large Hadron Collider (LHC). This situation has in fact motivated an impressive body of work, and continuous, ongoing efforts to both estimate and control PDFs uncertainties. Currently, the established method to obtain the PDFs is the global analysis, a fitting procedure, where initial scale Q0 ∼ 1GeV ≤ Qmin dat ansatze, as a function of the momentum fraction x, for each parton flavour i in hadron h are evolved to higher scales according to the perturbative QCD renormalization group equations. All the available observables e.g. the proton structure function, F2p (x, Q2 ), are composed of the candidate PDFs and comparison with the data is made with the help of some statistical estimator such as the global χ2 , 2 χ = Ne XX (Datai − Theori ) Vij−1 (Dataj − Theorj ) , expt. i,j=1 1 (1) where the error matrix Vij consists of the statistical and uncorrelated systematic errors, as well as of the correlated systematic errors when available. The parameters in the ansatze are then adjusted and the whole process repeated until a global minimum has been found. The modern PDF collaborations (CTEQ [1] and references within , MRST [2– 4], Alekhin [5, 6], Zeus [7] and H1 [8]) also provide error estimates for the PDF sets. They all rely on some kind of variant of the Hessian method (see e.g. [9] for details), which is based on a Taylor expansion of the global χ2 around it’s minimum. When only the leading terms are kept, the displacement of χ2 can be written in terms of Hessian matrix Hij , which consists of second derivatives of χ2 with respect to the parameter displacements, evaluated at the minimum. The error estimate for the parameters themselves, or for any quantity that depends on those parameters, can then be obtained in terms of the inverse of the Hessian matrix, (∆X)2 = ∆χ2 X ∂X ∂yi i,j H −1
∂X . ij ∂y j (2) For details of PDF uncertainty studies see e.g. Refs. [10, 11]. The global analysis combined with Hessian error estimation is a powerful method, allowing for both extrapolation outside the kinematical range of the data and extension to multivariable cases, such as nuclear PDFs (nPDFs) [12–15]. In principle, when more data become available, the method could also be applied to Generalized Parton Distributions (GPDs), for which only model-dependent [16] or semi model-dependent [17, 18] solutions presently exist. However, there are uncertainties related to the method itself, that are difficult to quantify, but may turn out to have a large effect. Choosing global χ2 as a statistical estimator may not be adequate since the minimum of the global fit may not correspond to a minima of the individual data sets, and as a result the definition of ∆χ2 may be ambiguous. Estimates for the current major global analyses are that ∆χ2 = 50 − 100 is needed to obtain a ∼ 90% confidence interval [1, 2]. In principle this problem could be avoided by using the Lagrange multiplier method (see e.g.[19]), which does not assume quadratic behaviour for the errors around the minimum, instead of the Hessian method, 2 but this is computationally more expensive solution. Introducing a functional form at the initial scale necessarily introduces a parametrization dependence bias and theoretical assumptions behind the fits, such as s, s̄, c quark content, details of the scale evolution (e.g. higher order perturbative corrections, large/small x resummation), higher twists etc. as well as the data selection and treatment, e.g. kinematical cuts, all reflect into the final result of the analysis. Also, there may be systematical bias introduced by the optimization algorithm. The differences between the current global PDF sets tend to be larger than the estimated uncertainties [20], and these differences again translate to the predictions for the LHC observables, such as Higgs [21] or W ± and Z production cross sections [1]. A new, fresh approach to the PDF fitting has recently been proposed by NNPDF collaboration [22, 23] who have replaced a typical functional form ansatz with a more complex standard neural network (NN) solution and the Hessian method with Monte Carlo sampling of the data (see the references within [23] for the nonsinglet PDF fit and the details of the Monte Carlo sampling). Figure 1: Schematic diagram of a feed-forward neural network, from [24]. Neural network can be described as a computing solution that consists of interconnected processing elements, neurons, that work together to produce an output function. In a typical feed-forward NN (see Fig. 1) the output is given by the neurons in the last layer, as a non-linear function of the output of all neurons in the previous layer, which in turn is a function of the output of all neurons in the previous layer, and so on, starting from the first layer, which receives the input. For a NN with L layers and nl neurons in each PL−1 layer, the total number of the parameters is l=1 (nl nl+1 + nl+1 ). 3 In the beginning of the NNPDF fitting procedure a Monte Carlo sample of replicas of the experimental data is generated by jittering the central values of the data withing their errorbars using univariate Gaussian (or some other distribution if desired) random numbers for each independent error source. The number of the replicas is made so large that the Monte Carlo set of replicas models faithfully the probability distribution of the original data. For each replica a Genetic Algorithm (GA) fit is performed by first setting the NN parameters for each parton flavour to be fitted randomly, then making clones of the set of parameters, and mutating each of them randomly (multiple mutations). After scale evolution the comparison with the data is performed for all the clones, and the best clones are selected for a source of new clones, and the process repeated until the minimum for the χ2 has been found. Overfitting of the data is prevented by using only part of the data in the minimizing procedure, and using the other part to monitor the behaviour of the χ2 . When fitting PDFs one thus ends up with Nrep PDF sets, each initial scale parton distribution parametrized by a different NN. The quality of the global fit is then given by the χ2 computed from the averages over the sample of trained neural networks. The mean value of the parton distribution at the starting scale for a given value of x is found by averaging over the replicas, and the uncertainty on this value is the variance of the values given by the replicas. The NNPDF method circumvents the problem of choosing a suitable ∆χ2 , and it relies on GA which works on a population of solutions for each MC replica, thus having a lowered possibility of getting trapped in local minima. NN parametrizations are also highly complex, with large number of parameters, and thus unbiased compared to the ansatze used in global fits. The estimated uncertainties for NNPDF fits are larger than those of global fits, possibly indicating that the global fit uncertainties may have been underestimated. It should, however, be pointed out that the MC sampling of the data is not not tied to use of NNs, and it thus remains undetermined whether the large uncertainties would persist if the MC sampling was used with a fixed functional form. The complexity of NN results may also induce problems, especially when used in a purely automated fitting procedure. Since the effect of modifying individual NN parameters is unknown, the result may exhibit 4 strange or unwanted behaviour in the extrapolation region, or in between the data points if the data is sparse. In such a case, and in a case of incompatible data, the overfitting method is also unsafe to use. Implementation of information not given directly by the data, such as nonperturbative models, lattice calculations or knowledge from prior work in general, is also difficult in this approach. A possible method of estimating the PDF uncertainties could also be provided by Bayesian statistical analysis, as preliminarily studied in [25, 26] and explained in [27], in which the errors for the PDF parameters, or for an observable constructed from the PDFs, are first encapsulated in prior probabilities for an enlarged set of model parameters, and posterior distributions are obtained using computational tools such as Markov Chain Monte Carlo. Similar to NNPDF approach, this method allows for an inclusion of nonGaussian systematic errors for the data. In this introductory paper we propose a new method which relies on the use of Self-Organizing Maps (SOMs), a subtype of neural network. The idea of our method is to create means for introducing “Researcher Insight” instead of “Theoretical bias”. In other words, we want to give up fully automated fitting procedure and eventually develop an interactive fitting program which would allow us to “take the best of both worlds”, to combine the best features of both the standard functional form approach and the neural network approach. In this first step, we solely concentrate on single variable functions, free proton PDFs, but discuss the extension of the model to multivariable cases. In Section 2 we describe the general features of the SOMs, in Sections 3 and 4 we present two PDF fitting algorithms relying on the use of SOMs and finally in Section 5 we envision the possibilities the SOM method has to offer. 2 Self-Organizing Maps The SOM is a visualization algorithm which attempts to represent all the available observations with optimal accuracy using a restricted set of models. The SOM was developed by T. Kohonen in the early 1980’s ([28], see also [29]) to model biological brain functions, but has since then developed 5 into a powerful computational approach on it’s own right. Many fields of science, such as statistics, signal processing, control theory, financial analyses, experimental physics, chemistry and medicine, have adopted the SOM as a standard analytical tool. SOMs have been applied to texture discrimination, classification/pattern recognition, motion detection, genetic activity mapping, drug discovery, cloud classification, and speech recognition, among others. Also, a new application area is organization of very large document collections. However, applications in particle physics have been scarce so far, and mostly directed to improving the algorithms for background event rejection [30–32]. SOM consists of nodes, map cells, which are all assigned spatial coordinates, and the topology of the map is determined by a chosen distance metric Mmap . Each cell i contains a map vector Vi , that is isomorphic to the data samples used for training of the neural network. In the following we will concentrate on a 2-dimensional rectangular lattice for simplicity. A natural choice for the P topology is then L1 (x, y) = 2i=1 |xi − yi |, which also has been proved [33] to be an ideal choice for high-dimensional data, such as PDFs in our case. The implementation of SOMs proceeds in three stages: 1) initialization of the SOM, 2) training of the SOM and 3) associating the data samples with a trained map, i.e. clustering. During the initialization the map vectors are chosen such that each cell is set to contain an arbitrarily selected sample of either the actual data to be clustered, or anything isomorphic to them (see Fig. 2 for an example). The actual training data samples, which may be e.g. subset or the whole set of the actual data, are then associated with map vectors by minimizing a similarity metric Mdata . We choose Mdata = L1 . The map vector each data sample becomes matched against, is then the most similar one to the data sample among all the other map vectors. It may happen that some map vectors do no not have any samples associated with them, and some may actually have many. During the training the map vectors are updated by averaging them with the data samples that fell into the cells within a given decreasing neighbourhood, see Fig. 3. This type of training which is based on rewarding the winning node to become more like data, is called competitive learning. The initial value of a map vector Vi at SOM cell i then changes during the course of 6 Figure 2: A 2D grid SOM which cells get randomly associated with the type of data samples we would like to study, such as nonsinglet PDFs or observables. At this stage each cell gets associated with only one curve, the map vector. training as Vi (t + 1) = Vi (t) (1 − w(t) Nj,i (t)) + Sj (t) w(t) Nj,i(t) (3) where now Vi (t + 1) is the contents of the SOM cell i after the data sample Sj has been presented on the map. The neighbourhood, the radius, within which the map vectors are updated is given by the function Nj,i (t), centered on the winner cell j. Thus even the map vectors in those cells that didn’t find a matching data sample are adjusted, rewarded, to become more like 2 data. Typically Nj,i (t) = e−Mmap (j,i) /r(t) , where r(t) is a monotonously decreasing radius sequence. In the beginning of the training the neighbourhood may contain the whole map and in the end it just consists of the cell itself. Moreover, the updating is also controlled by w(t), which is a monotonously decreasing weight sequence in the range [0, 1]. As the training proceeds the neighbourhood function eventually causes the data samples to be placed on a certain region of the map, where the neighbouring map vectors are becoming increasingly similar to each other, and the weight sequence w(t) furthermore finetunes their position. In the end on a properly trained SOM, cells that are topologically close to each other will have map vectors which are similar to each other. In the final phase the actual data is matched against the map vectors of the trained map, and thus get distributed on the map according to the feature that was used 7 Figure 3: (Colour online) Each data sample Si is associated with the one map vector Vi it is most similar to. As a reward for the match, the winning map vector, as well as its neighbouring map vectors, get averaged with the data associated with the winning cell. as Mdata . Clusters with similar data now emerge as a result of unsupervised learning. For example, a map containing RGB colour triplets would initially have colours randomly scattered around it, but during the course of training it would evolve into patches of colour which smoothly blend with each other, see Fig. 4. This local similarity property is the feature that makes SOM suitable for visualization purposes, thus facilitating user interaction with the data. Since each map vector now represent a class of similar objects, the SOM is an ideal tool to visualize high-dimensional data, by projecting it onto a low-dimensional map clustered according to some desired similar feature. SOMs, however, also have disadvantages. Each SOM, though identical in size and shape and containing same type of data, is different. The clusters may also split in such a way that similar type of data can be found in several different places on the map. We are not aware of any mathematical or computational means of detecting if and when the map is fully trained, and whether there occurs splitting or not, other than actually computing the 8 Figure 4: (Colour online) SOM containing RGB colour triplets getting trained. Adapted from Ref. [34] similarities between the neighbouring cells and studying them. In this work we use the so-called batch-version of the training, in which all the training data samples are matched against the map vectors before the training begins. The map vectors are then averaged with all the training samples within the neighbourhood radius simultaneously. The procedure is repeated Nstep (free parameter to choose) times such that in every training step the same set of training data samples is associated with the evolving map and in Eq.(3) t now counts training steps. When the map is trained, the actual data is finally matched against the map vectors. In our study our training data are always going to be the whole data we want to cluster, and the last training step is thus the clustering stage. The benefit of the batch training compared to the incremental training, described earlier, is that the training is independent of the order in which the training samples are introduced on the map. 3 MIXPDF algorithm In this approach our aim is to both i) to be able to study the properties of the PDFs in a model independent way and yet ii) to be able to implement knowledge from the prior works on PDFs, and ultimately iii) to be able to guide the fitting procedure interactively with the help of SOM properties. At this stage it is important to distinguish between the experimental data 9 and the training data of the SOM. When we are referring to measured data used in the PDF fitting, such as F2 data, we always call it experimental data. The SOM training data in this study is going to be a collection of candidate PDF sets, produced by us, or something composed of them. A PDF set in the following will always mean a set of 8 curves, one for each independent ¯ s = s̄, c = c̄ and b = b̄ in this simplified parton flavour f = (g, uv , dv , ū, d, introductory study), that are properly normalized such that XZ 1 dxxff /p (x, Q2 ) = 1, (4) f 0 and conserve baryon number and charge Z 1 Z 1 2 dxfuv /p (x, Q ) = 2, dxfdv /p (x, Q2 ) = 1. 0 (5) 0 In order to proceed we have to decide how to create our candidate PDF sets, decide the details of the SOMs, details of the actual fitting algorithm, experimental data selection and details of the scale evolution. In this introductory paper our aim is not to provide a finalised SOMPDF set, but rather to explore the possibilities and restrictions of the method we are proposing. Therefore we refrain from using “all the possible experimental data” as used in global analyses, but concentrate on DIS structure function data from H1 [35], BCDMS [36, 37] and Zeus [38], which we use without additional kinematical cuts or normalization factors (except rejecting the data points below our initial scale). The parameters for the DGLAP scale evolution were chosen to be those of CTEQ6 (CTEQ6L1 for lowest order (LO)) [39], the initial scale being Q0 = 1.3 GeV. In next-to-leading order (NLO) case the evolution code was taken from [40] (QCDNUM17 beta release). We will start now with a simple pedagogical example, which we call MIXPDF algorithm, where we use some of the existing PDF sets as material for new candidate PDFs. At first, we will choose CTEQ6 [39], CTEQ5 [41], MRST02 [2, 42], Alekhin [5] and GRV98 [43] sets and construct new PDF sets from them such that at the initial scale each parton flavour in the range x = [10−5 , 1] is randomly selected from one of these five sets (we set the heavy flavours to be zero below their mass thresholds). The sumrules on this new 10 set are then imposed such that the original normalization of uv and dv are preserved, but the rest of the flavours are scaled together so that Eq.(4) is fulfilled. In this study we accept the <few% normalization error which results from the fact that our x-range is not x = [0, 1]. From now on we call these type of PDF sets database PDFs. We randomly initialize a small 5 × 5 map with these candidate database PDFs, such that each map vector Vi consists of the PDF set itself, and of the observables F2p (x, Q20 ) derived from it. Next we train the map with Nstep = 5 batch-training steps with training data that consists of 100 database PDFs plus 5 original “mother” PDF sets, which we will call init PDFs from now on. We choose the similarity criterion to be the similarity of observables F2p (x, Q2 ) with Mdata = L1 . The similarity is tested at a number of x-values (equidistant in logarithmic scale up to x ∼ 0.2, and equidistant in linear scale above that) both at the initial scale and at all the evolved scales where experimental data exist. On every training, after the matching, all the observables (PDFs) of the map vectors get averaged with the observables (PDFs, flavor by flavor) matched within the neighbourhood according to Eq. (3). The resulting new averaged map vector PDFs are rescaled again (such that uv and dv are scaled first) to obey the sumrules. From now on we will call these type of PDF sets map PDFs. The map PDFs are evolved and the observables at every experimental data scale are computed and compared for similarity with the observables from the training PDFs. After the training we have a map with 25 map PDFs and the same 105 PDF sets we used to train the map. The resulting LO SOM is shown in Fig. 5, with just F2p (x, Q20 )’s of each cell shown for clarity. The red curves in the figure are the map F2’s constructed from the map PDFs, black curves are the database F2’s and green curves are the init F2’s constructed from the init PDFs (CTEQ6, CTEQ5, MRST02, Alekhin and GRV98 parametrizations). It is obviously difficult to judge visually just by looking at the curves whether the map is good and fully trained. One hint about possible ordering may be provided by the fact that the shape of the F2p curve must correlate with the χ2 /N against the experimental data. The distribution of the χ2 /N values (no correlated systematic errors are taken into account for simplicity) of the map PDFs, shown in each cell , does indeed seem somewhat organized. 11 1 2 3 4 2.46 2.20 4.36 4.98 3.28 2.49 2.38 1.54 2.59 4.04 3.22 2.51 1.45 2.75 6.55 5.51 4.27 3.26 2.00 9.49 7.24 4.72 4.58 2.32 4 3 2 1 0 0 LO 2.44 1. iteration -5 10 10 -4 -3 10 10 -2 10 -1 1 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 Figure 5: (Colour online) Trained 1. iteration LO map. Red curves are map F2p ’s, green curves init F2p ’s and black curves rest of the database F2p ’s. The number in each cell is the χ2 /N of the map PDF. Table 1 lists the χ2 /N values the original init PDFs obtain within the MIXPDF framework described above. Comparison of these values with the values in Fig. 5 reveals that some of the map PDFs, as also some of the database PDFs, have gained a χ2 /N comparable to or better than that of the init PDFs. Inspired by the progress, we start a second fitting iteration by selecting the 5 best PDF sets from the 25+5+100 PDF sets of the first iteration as our new init PDFs (which are now properly normalized after the first iteration) to generate database PDFs for a whole new SOM. Since the best PDF candidate from the first iteration is matched on this new map as an unmodified init PDF, it is guaranteed that the χ2 /N as a function of the iteration either decreases or remains the same. We keep repeating the iterations until the χ2 /N saturates. Fig. 6 (Case 1) shows the χ2 /N as a function of iterations for the best PDF on the trained map, for the worst PDF on the map and for 12 PDF‡ Alekhin CTEQ6 CTEQ5 CTEQ4 MRST02 GRV98 LO χ2 /N 3.34 1.67 3.25 2.23 2.24 8.47 NLO χ2 /N 29.1 2.02 6.48 2.41 1.89 9.58 Table 1: χ2 /N for different MIXPDF input PDF sets against all the datasets used (H1, ZEUS, BCDMS, N=709). the worst of the 5 PDFs selected for the subsequent iteration as an init PDF. The final χ2 /N of these runs are listed in Table 2 (first row) as Case 1 and Fig. 7 shows these results (black solid line), together with original starting sets at the initial scale (note the different scaling for gluons in LO and in NLO figures). For 10 repeated LO runs we obtain χ¯2 /N=1.208 and σ=0.029. Let us now analyze in more detail how the optimization proceeds. Figs. 8,9 show the LO maps also for the 2. and 3. iterations. On the first iteration the init PDFs, the shapes of which were taken from existing parametrizations, fall in the cells (0,1) (CTEQ5), (1,3) (Alekhin), (1,4) (MRST02), (2,3) (CTEQ6) and (3,0) (GRV98), so the modern sets, Alekhin, MRST02 and CTEQ6, group close to each other, i.e. the shapes of the observables they produce are very similar, as expected. The best 5 PDFs selected as the 2. iteration init PDFs also come from this neighbourhood, 3 of them from the cell (1,3) and the other 2 from the cell (2,3). Two of these selected sets are map PDFs, two are database PDFs and also the original init PDF CTEQ6 survived for the 2. iteration. At the end of the 2. iteration the init PDFs, which originated from the neighbouring cells, are scattered in the cells (0,1), (0.3), (1,0) (CTEQ6), ‡ These are the χ2 /N for the initial scale PDF sets taken from the quoted parametrizations and evolved with CTEQ6 DGLAP settings, the heavy flavours were set to be zero below their mass thresholds, no kinematical cuts or normalization factors for the experimental data were imposed, and no correlated systematic errors of the data were used to compute the χ2 /N . We do not claim these values to describe the quality of the quoted PDF sets. 13 4.0 4.0 LO 3.5 NLO Case 1: best of 5 worst of 5 worst of map 5 5, Nstep=5 3.0 Case 1: Case 2: 3.0 Case 2: 2 2.0 best of 5 worst of 5 worst of map 2.5 2.0 1.5 1.5 1.0 1.0 0.5 0.5 0 1 2 3 4 5 6 7 8 Iteration 0 1 2 3 4 5 6 7 8 Iteration Figure 6: (Colour online) χ2 /N of the MIXPDF runs as a function of the iteration. (2,2) and (2,3) and the best 5 PDFs produced during this iteration are in the cells (4,1) (database PDF), (4,0) (map PDF), (3,2) (map PDF), (2,2) (database PDF) and (3,0) (map PDF). After the 3. iteration the above best 5 PDFs of 2. iteration are in the cells (2,0), (3,1), (0.2), (0.4) and (3,2) and the new best 5 PDFs produced are all map PDFs with 4 of them in neighbouring cell pairs. Map PDFs provide complicated linear combinations of the database PDFs and obviously play an important role in the algorithm. The size of the map dictates how much the neighbouring map vectors differ from each other. Since the PDFs in the same cell are not required to be similar, only the observables constructed from them are, a cell or a neighbourhood may in principle contain a spread of PDFs with a spread of χ2 /N’s. However, since our selection criteria for the init PDFs was based on the best χ2 /N only, it is inevitable that the observables on the map become increasingly similar as the iterations go by, and the χ2 /N flattens very fast as can be seen from Fig. 6. As a consequence we quickly lose the variety in the shapes of the PDFs as the iterations proceed, and on the final iteration all the PDFs on the map end up being very similar. 14 /N best of 5 worst of 5 worst of map 2.5 2 /N 3.5 best of 5 worst of 5 worst of map 5 5, Nstep=5 SOM 5x5 5x5 5x5 5x5 5x5 5x5 5x5 5x5 15x15 15x15 15x15 15x15 Nstep 5 5 5 5 5 5 10 40 5 5 5 5 # init 5 5 10 10 15 20 10 10 5 5 30 30 # database 100 100 100 100 100 100 100 100 900 900 900 900 Case 1 2 1 2 1 1 1 1 1 2 1 2 LO χ2 /N 1.19 1.37 1.16 1.49 1.16 1.17 1.16 1.20 1.22 1.31 1.16 1.25 NLO χ2 /N 1.28 1.44 1.25 1.43 1.45 1.30 1.25 l.53 Table 2: χ2 /N against all the datasets used (H1, ZEUS, BCDMS) for some selected MIXPDF runs. The MIXPDF algorithm obviously has several other weaknesses too. Among them are how the result would change if we started with another first iteration PDF selection, and what are the effects of changing the size of the map, number of the database PDFs and init PDF sets and size of Nstep ? In general, how much the final result depends on the choices that we make during the fitting process? Let us now study some of these questions a bit. Since we have chosen our evolution settings to be those of CTEQ6’s, it necessarily becomes a favoured set (although we don’t impose any kinematical cuts on the experimental data). Therefore we performed another LO and NLO runs, with CTEQ6 now replaced with CTEQ4 [44]. The results of these runs are reported in Table 2 (2. row) and in Fig. 6 (χ2 /N) and Fig. 7 (the PDFs) as Case 2 . The Case 2 clearly produces worse results. Without an input from CTEQ6 we automatically lose all the low gluons at small-x -type of results in NLO, for example. Fig. 10 addresses the issue of choosing different Nstep , for LO Case 1 and Case 2. The solid (dotted) lines show the best (worst) init PDF selected 15 4.0 4.0 Q=1.3 GeV CTEQ6 Alekhin MRST02 GRV98 CTEQ5 CTEQ4 LO 3.5 3.0 0.25*xg 2.5 Q=1.3 GeV CTEQ6 Alekhin MRST02 GRV98 CTEQ5 CTEQ4 NLO 0.45*xg 5 5, Nstep= 5 Case 1 Case 2 5 5, Nstep= 5 Case 1 Case 2 15 15, Nstep= 5 Case 1 Case 2 15 15, Nstep= 5 Case 1 Case 2 3.5 3.0 2.5 2.0 2.0 1.5 1.0 xuV xu 1.5 1.0 xuV xu 0.5 0.0 -5 10 0.5 0.0 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 -5 10 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 Figure 7: (Colour online) MIXPDF results together with the input PDF sets. from each iteration for several Nstep selections. In Case 2 the best results are obtained with small number of training steps, whereas Case 1 does not seem to benefit from a longer SOM training. Keeping the stochastical nature of the process in our minds, we may speculate that the seemingly opposite behaviour for the Case 1 and Case 2 results from the fact that it is more probable to produce a good set of database PDFs in Case 1 than in Case 2. If the database is not so good to begin with, the longer training contaminates all the map PDFs with the low quality part of the database PDFs. Table 2 also showcases best results from a variety of MIXPDF runs where we have tried different combinations of SOM features. It is interesting to notice that increasing the size of the map and database does not necessarily lead to a better performance. Instead the number of the init PDFs on later iterations (always 5 for the first iteration) seem to be a key factor, it has to be sufficiently large to preserve the variety of database the PDFs but small enough to consist of PDFs with good quality. In our limited space of database candidates (∼ 65 = 7776 possibilities) the optimal set of variables 16 1 2 3 4 1.43 1.76 1.70 1.72 1.38 1.59 1.57 1.65 1.51 1.37 1.46 1.53 1.39 1.42 1.35 1.36 1.33 1.41 1.60 1.33 1.47 1.65 1.67 1.67 4 3 2 1 0 0 LO 1.41 2. iteration -5 10 10 -4 -3 10 10 -2 10 -1 1 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 Figure 8: (Colour online) Trained 2. iteration LO map, curves and numbers as in Fig.5. for the MIXPDF run should not be not impossible to map down. The method we used to generate the sample data is very simple indeed. The number of all the possible candidate database PDFs is not very large to begin with, so the quality of the final results strongly depends on the quality of the input for the SOM. Since the map PDFs are obtained by averaging with the training samples, and the non-valence flavours are scaled by a common factor when imposing the sumrules, the map PDFs tend to lie in between the init PDFs. Therefore map PDFs with extreme shapes are never produced, and thus never explored by the algorithm. A method which relies on sampling existing parametrizations on a SOM is inconvenient also because it is not readily applicable to multivariable cases. For the case of the PDFs it is sufficient to have a value for each flavour for a discrete set of x-values, but for a multivariable cases, such as nPDFs, or GPDs the task of keeping track of the grid of values for each flavour in each x for several different values of the additional variables (e.g. A and Z 17 1 2 3 4 1.26 1.26 1.28 1.36 1.35 1.26 1.26 1.31 1.33 1.36 1.29 1.25 1.30 1.37 1.36 1.32 1.27 1.35 1.35 1.41 1.29 1.27 1.30 1.42 4 3 2 1 0 0 LO 1.34 3. iteration -5 10 10 -4 -3 10 10 -2 10 -1 1 1.4 1.2 1.0 0.8 0.6 0.4 0.2 0.0 Figure 9: (Colour online) Trained 3. iteration LO map, curves and numbers as in Fig.5. for nuclei, and the skewness, ξ and the squared 4-momentum transfer, t for GPDs) is computationally expensive. In principle, SOM can keep track of the interrelations of the map vectors, and knowing the parametrization for the init PDFs, it would be possible to construct the parametrization for the map PDFs. That would, however, lead to very complicated parametrizations, and different nPDF, GPD etc. parametrizations are presently not even either measured or defined at the same initial scale. Despite of its problems, on a more basic level, MIXPDF does have the the desirable feature that it allows us to use SOM as a part of the PDF optimization algorithm in such a way that we cluster our candidate PDFs on the map, and select those PDFs which i) minimize a chosen fitness function, e.g. χ2 , when compared against experimental data, and ii) have some desired feature which can be visualized on the map or used as a clustering criterion. Therefore, in the following Section, we keep building on this example. 18 2.2 2.2 LO, 5 5 LO, 5 5 Nstep: Case 1 Case 2 best of 5, solid lines worst of 5, dashed lines 2.0 1.8 1.6 best of 5, solid lines worst of 5, dashed lines 1.4 1.4 1.2 1.2 1.0 1.0 0 1 2 3 4 5 6 7 8 Iteration 0 1 2 3 4 5 6 7 8 Iteration Figure 10: (Colour online) LO χ2 /N for 5 × 5 SOM runs with different Nstep . 4 ENVPDF algorithm Most of the problems with the MIXPDF algorithm originate from the need to be able to generate the database PDFs in an unbiased way as possible, and at the same time to have a variety of PDF candidates available at every stage of the fitting procedure. Yet, one needs to have control over the features of the database PDFs that are created. To accomplish this, we choose, at variance with the “conventional” PDFs sets or NNPDFs, to give up the functional form of PDFs and rather to rely on purely stochastical methods in generating the initial and training samples of the PDFs. Our choice is a GA-type analysis, in which our parameters are the values of PDFs at the initial scale for each flavour at each value of x where the experimental data exist. To obtain control over the shape of the PDFs we use some of the existing distributions to establish an initial range, or envelope, within which we sample the database PDF values. Again, we use the Case 1 and 2 PDF sets (CTEQ6, CTEQ5, CTEQ4, MRST02, Alekhin and GRV98) as an initialization guideline. We construct our initial PDF generator first to, for each flavour separately, select ran19 /N 2 5 10 20 30 40 2 /N 1.8 1.6 Nstep: 5 10 20 30 40 2.0 domly either the range [0.5, 1], [1.0, 1.5] or [0.75, 1.25] times any of the Case 1 (or 2) PDF set. Compared to MIXPDF algorithm we are thus adding more freedom to the scaling of the database PDFs. Next the initial generators generate values for each xdata § using uniform, instead of Gaussian, distribution around the existing parametrizations, thus reducing direct bias from them. Gaussian smoothing is applied to the resulting set of points, and the flavours combined to form a PDF set such that the curve is linearly interpolated from the discrete set of generated points. The candidate PDF sets are then scaled to obey the sumrules as in MIXPDF algorithm. In order to obtain a reasonable selection of PDFs to start with, we reject candidates which have χ2 /N > 10 (computed as in MIXPDF algorithm). To further avoid direct bias from the Case 1 and 2 PDFs, we don’t include the init PDFs into the training set for the first iteration as we did in MIXPDF case. For a N × N SOM we choose the size of the database to be 4N 2 . During the later iterations we proceed as follows: At the end of each iteration we pick from the trained N ×N SOM 2N best PDFs as the init PDFs. These init PDFs are introduced into the training set alongside with the database PDFs, which are now constructed using each of the init PDFs in turn as a center for a Gaussian random number generator, which assigns for all the flavours for each x a value around that same init PDF such that 1 − σ of the generator is given by the spread of the best PDFs in the topologically nearest neighbouring cells. The object of these generators is thus to refine a good candidate PDF found in the previous iteration by jittering it’s values within a range determined by the shape of other good candidate PDFs from the previous iteration. The generated PDFs are then smoothed and scaled to obey the sumrules. Sets with χ2 /N > 10 are always rejected. We learnt from the MIXPDF algorithm that it is important to preserve the variety of the PDF shapes on the map, so we also keep Norig copies of the first iteration generators in our generator mix. Table 3 lists results from a variety of such runs. The results do not seem § To ensure a reasonable large-x behaviour for the PDFs, we also generate with the same method values for them in a few x-points outside the range of the experimental data. We also require the PDFs, the gluons especially, to be positive for simplicity. 20 to be very sensitive to the number of SOM training steps, Nstep , but are highly sensitive to the number of first iteration generators used in subsequent iterations. Although the generators can now in principle produce an infinite number of different PDFs, the algorithm would not be able to radically change the shape of the database PDFs without introducing a random element on the map. Setting Norig > 0 provides, through map PDFs, that element, and keeps the algorithm from getting fixed to a local minimum. The ENVPDF algorithm is now more independent from the initial selection of the PDF sets, Case 1 or 2, than MIXPDF, since no values of e.g. the CTEQ6 set in the original generator are ever introduced on the map directly. SOM Nstep 5x5 5 5x5 10 5x5 20 5x5 30 5x5 40 5x5 5 5x5 20 5x5 5 5x5 10 5x5 15 15x15 5 15x15 5 Norig 2 2 2 2 2 0 0 2 2 2 6 6 Case 1 1 1 1 1 1 1 2 2 2 1 2 LO χ2 /N 1.04 1.10 1.10 1.10 1.08 1.41 1.26 1.14 1.12 1.18 1.00 1.13 NLO χ2 /N 1.08 1.25 1.07 1.18 Table 3: χ2 /N against all the datasets used (H1, ZEUS, BCDMS) for variety of ENVPDF runs. Fig. 11 shows the χ2 /N as a function of iteration for 5x5 LO and NLO, both Case 1 and Case 2, runs, where Nstep = 5 and Norig = 2. Clearly the ENVPDF runs take multiple number of iterations for the χ2 /N to level compared to the MIXPDF runs, and they are therefore more costly in time. With the ENVPDF algorithm, however, the χ2 /N keeps on slowly improving even after all the mother PDFs from the same iteration are equally good fits. For a larger 15x15 SOM the number of needed iterations remains as large. 21 3.0 3.0 LO 2.5 NLO Case 1: 5 5, Nstep=5 Case 1: 5 5, Nstep=5 best of 10 worst of 10 Case 2: Case 2: 2.0 2 best of 10 worst of 10 /N best of 10 worst of 10 2 /N 2.0 2.5 best of 10 worst of 10 1.5 1.5 1.0 1.0 0.5 0.5 0 5 10 15 20 25 30 35 40 45 50 Iteration 0 5 10 15 20 25 30 35 40 45 50 Iteration Figure 11: (Colour online) χ2 /N of the ENVPDF runs as a function of the iteration. Fig. 12 shows some of the Case 1 LO ENVPDF results at the initial scale Q = 1.3 GeV (left panel), and evolved up to Q = 3 GeV (right panel). The reference curves shown are also evolved as in MIXPDF. Although the initial scale ENVPDF results appear wiggly, they smooth out soon because of the additional well known effect of QCD evolution. In fact, the initial scale curves could be made smoother by applying a stronger Gaussian smoothing, but this is not necessary, as long as the starting scale is below the Qmin of the data. The evolved curves preserve the initially set baryon number scaling within 0.5% and momentum sumrule within 1.5% accuracy. Also, the results obtained from a larger map tend to be smoother since the map PDFs get averaged with a larger number of other PDFs. Studying the relation between the redundant wiggliness of our initial scale PDFs and possible fitting of statistical fluctuations of the experimental data is beyond the scope of this paper. The NLO Case 1 and 2 results are presented in Fig. 13. The trend of the results is clearly the same as in MIXPDF case, CTEQ6 is a favoured set, and especially the PDFs with gluons similar to those of CTEQ6’s have good χ2 /N. 22 We did not study the effect of modifying the width or the shape of the envelope in detail here, but choosing the envelope to be the wider or narrower than 1 − σ for the Gaussian generate seem to lead both slower and poorer convergence. Also, since we are clustering on the similarity of the observables, the same cell may in theory contain the best PDF of the iteration and PDFs which have χ2 /N as large as 10. Therefore the shape of the envelope should be determined only by the curves with promising shapes. 4.0 2.5 LO LO 3.5 Q=1.3 GeV CTEQ6 MRST02 CTEQ4 2.0 0.25*xg 1.5 Q=3.0 GeV CTEQ6 MRST02 CTEQ4 0.1*xg 5 5, Nstep= 5 Case 1 Case 2 5 5, Nstep= 5 Case 1 Case 2 15 15, Nstep= 5 Case 1 Case 2 15 15, Nstep= 5 Case 1 Case 2 3.0 2.5 2.0 1.0 1.5 xuV xu 1.0 xu 0.5 xuV 0.5 0.0 -5 10 0.0 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 -5 10 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 Figure 12: (Colour online) LO ENVPDF results at the initial scale, and at Q = 3.0 GeV. Next we want to study the PDF uncertainty using the unique means the SOMs provide to us even for a case of PDFs without a functional form. Since we have only used DIS data in this introductory study, we are only able to explore the small-x uncertainty for now. Figs. 14 (LO) and 15 (NLO) showcase, besides our best results, the spread of all the initial scale PDFs with χ2 /N ≤ 1.2, that were obtained during a 5 × 5 (left panel) and 15 × 15 (right panel) SOM run. Since the number of such PDF sets is typically of the order of thousands, we only plot the minimum and maximum of the bundle 23 4.0 3.0 Q=1.3 GeV CTEQ6 MRST02 CTEQ4 2.5 2.0 15 15, Nstep= 5 Case 1 Case 2 NLO NLO 3.5 Q=3.0 GeV CTEQ6 MRST02 CTEQ4 0.25*xg 5 5, Nstep= 5 Case 1 Case 2 0.85*xg 3.0 5 5, Nstep= 5 Case 1 Case 2 2.5 2.0 1.5 15 15, Nstep= 5 Case 1 Case 2 1.5 1.0 xuV 1.0 xu xuV 0.5 xu 0.0 -5 10 0.5 0.0 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 -5 10 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 Figure 13: (Colour online) NLO ENVPDF results at the initial scale, and at Q = 3.0 GeV. of curves. Since the total number of experimental datapoints used is ∼ 710, the spread ∆χ2 /N ∼ 0.2 corresponds to a ∆χ2 ∼ 140. Expectedly, the smallx gluons obtain the largest uncertainty for all the cases we studied. Even though a larger SOM with a larger database might be expected to have more variety in the shapes of the PDFs, the χ2 /N ≤ 1.2 spreads of the 5 × 5 and 15 × 15 SOMs are more or less equal sized (the apparent differences in sizes at Q = Q0 even out when the curves are evolved). Both maps therefore end up producing the same extreme shapes for the map PDFs although a larger map has more subclasses for them. Remarkably then, a single SOM run can provide a quick uncertainty estimate for a chosen ∆χ2 without performing a separate error analysis. Due to the stochastical nature of the ENVPDF algorithm, we may well also study the combined results from several separate runs. It is especially important to verify the stability of our results, to show that the results are indeed reproducible instead of lucky coincidences. Left panels of Figs 16 (LO) and 24 3.0 3.0 LO LO 2.5 2.5 Q=1.3 GeV CTEQ6 MRST02 2.0 Q=1.3 GeV CTEQ6 MRST02 5 5, Nstep= 5 Case 1 2 /N 1.2 0.25*xg 2.0 15 15, Nstep= 5 Case 1 2 /N 1.2 0.25*xg 1.5 1.5 1.0 1.0 xuV xuV xu xu 0.5 0.0 -5 10 0.5 0.0 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 -5 10 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 Figure 14: (Colour online) LO ENVPDF best result and the χ2 /N ≤ 1.2 spread of results. 17 (NLO) present the best results, and the combined χ2 /N ≤ 1.2 spreads for 10 repeated 5 × 5, Nstep = 5 runs at the initial scale. The average χ2 /N and the standard deviation σ for these runs are in LO (NLO) are 1.065 and 0.014 (1.122 and 0.029), corresponding to ∆χ2 ∼ 10 (20) for LO (NLO). The right panels of the same Figs 16, 17 show the 10 best result curves and the χ2 /N ≤ 1.2 spreads evolved up to Q = 3.0 GeV. Clearly the seemingly large difference between the small-x gluon results at the initial scale is not statistically significant, but smooths out when gluons are evolved. Thus the initial scale wiggliness of the PDFs is mainly only a residual effect from our method of generating them and not linked to the overtraining of the SOM, and we refrain from studying cases where stronger initial scale smoothing is applied. Therefore our simple method of producing the candidate PDFs by jittering random numbers inside a predetermined envelope is surprisingly stable when used together with a complicated PDF processing that SOMs provide. 25 3.5 3.5 Q=1.3 GeV CTEQ6 MRST02 NLO 3.0 NLO 5 5, Nstep= 5 Case 1 2 /N 1.2 2.5 0.85*xg Q=1.3 GeV CTEQ6 MRST02 3.0 15 15, Nstep= 5 Case 1 2 /N 1.2 0.85*xg 2.5 2.0 2.0 1.5 1.5 1.0 1.0 xuV 0.5 xuV xu 0.5 xu 0.0 -0.5 -5 10 0.0 -0.5 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 -5 10 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 Figure 15: (Colour online) NLO ENVPDF best result and the χ2 /N ≤ 1.2 spread of results. 5 Future of the SOMPDFs So far we have shown a relatively straightforward method of obtaining stochastically generated, parameter-free, PDFs, with an uncertainty estimate for a desired ∆χ2 . On every iteration using our competitive learning algorithm, the selection of the winning PDFs was based on the χ2 /N alone, and the fitting procedure was fully automated. In our MIXPDF algorithm the SOMs were used merely as a tool to create new combinations, map PDFs, of our input database. The ENVPDF algorithm also used the topology of the map to determine the shape of the envelope, within which we sampled the database PDFs. We reiterate that our initial study was aimed at observing and recording the behavior of the SOM as an optimization tool. Many of the features of our results could not in fact be predicted based on general assumptions. The proposed method can be extended much further than that. The automated version of the algorithm could be set to sample a vector consisting of PDF 26 4.0 3.0 LO LO 3.5 2.5 Q=1.3 GeV CTEQ6 MRST02 2.0 0.25*xg Q=3.0 GeV CTEQ6 MRST02 0.1*xg 5 5, Nstep= 5 Case 1 2 /N 1.2 3.0 5 5, Nstep= 5 Case 1 2 /N 1.2 2.5 2.0 1.5 1.5 1.0 xuV 0.5 1.0 xu xu xuV 0.5 0.0 -5 10 0.0 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 -5 10 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 Figure 16: (Colour online) LO ENVPDF best results and the χ2 /N ≤ 1.2 spreads of results from 10 separate runs. parameters, instead of values of PDFs in each value of x of the data. That would lead to smooth, continuous type of solutions, either along the lines of global analyses, or NNPDFs using N SOMs for N Monte-Carlo sampled replicas of the data. Since the solution would be required to stay within an envelope of selected width and shape given by the map, no restrictions for the parameters themselves would be required. For such a method, all the existing error estimates, besides an uncertainty band produced by the map, would be applicable as well. What ultimately sets the SOM method apart from the standard global analyses or NNPDF method, however, are the clustering and visualization possibilities that it offers. Instead of setting Mdata = L1 and clustering according to the similarity of the observables, it is possible to set the clustering criteria to be anything that can be mathematically quantified, e.g. the shape of the gluons or the large-x behaviour of the PDFs. The desired feature of the PDFs can then be projected out from the SOM. Moreover, by combining the 27 4.0 3.5 Q=1.3 GeV CTEQ6 MRST02 NLO 3.0 NLO 3.5 5 5, Nstep= 5 Case 1 2 /N 1.2 2.5 Q=3.0 GeV CTEQ6 MRST02 0.25*xg 0.85*xg 3.0 5 5, Nstep= 5 Case 1 2 /N 1.2 2.0 2.5 2.0 1.5 1.0 1.5 xuV 0.5 1.0 xu xu xuV 0.0 -0.5 -5 10 0.5 0.0 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 -5 1 10 2 5 -4 10 2 5 10 -3 2 x 5 -2 10 2 5 -1 10 2 5 1 Figure 17: (Colour online) NLO ENVPDF best results and the χ2 /N ≤ 1.2 spreads of results from 10 separate runs. method with an interactive graphic user interface (GUI), it would be possible to change and control the shape and the width of the envelope as the minimization proceeds, to guide the process by applying researcher insight at various stages of the process. Furthermore, the uncertainty band produced by the SOM as the run proceeds, could help the user to make decisions about the next steps of the minimization. With GUI it would be e.g. possible to constrain the extrapolation of the NN generated PDFs outside the x-range of the data without explicitly introducing terms to ensure the correct smalland large-x behaviour as in NNPDF method (see Eq.(87) in [23]). The selection of the best PDF candidates for the subsequent iteration could then be made based on the user’s preferences instead of solely based on the χ2 /N. That kind of method in turn could be extended to multivariable cases such as nPDFs and even GPDs and other not so well-known cases, where the data is too sparse for stochastically generated, parameter-free, PDFs. Generally, any PDF fitting method involves a large number of flexible points 28 “opportunities for adapting and fine tuning”, which act as a source for both systematical and theoretical bias when fixed. Obvious optimization method independent sources of theoretical bias are the various parameters of the DGLAP equations, inclusion of extra sources of Q2 -dependence beyond DGLAP-type evolution and the data selection, affecting the coverage of different kinematical regions. SOMs themselves, and different SOMPDF algorithm variations naturally also introduce flexible points of their own. We explored a little about the effects of choosing the size of the SOM and the number of the batch training steps Nstep . There are also plenty of other SOM properties that can be modified, such as the shape of the SOM itself. We chose to use a rectangular lattice, but generally the SOM can take any shape desired. For demanding vizualisation purposes a hexagonal shape is an excellent choice, since the meaning of the nearest neighbours is better defined. The SOMPDF method, supplemented with the use of a GUI, will allow us to both qualitatively and quantitatively study the flexible points involved in the PDFs fitting. More complex hadronic matrix elements, such as the ones defining the GPDs, are natural candidates for future studies of cases where the experimental data are not numerous enough to allow for a model independent fitting, and the guidance and intuition of the user is therefore irreplaceable. The method we are proposing is extremely open for user interaction, and the possibilities of such a method are widely unexplored. Acknowledgements We thank David Brogan for helping us to start this project. This work was financially supported by the US National Science Foundation grant no.0426971. HH was also supported by the U.S. Department of Energy, grant no. DE-FG02-87ER40371. SL is supported by the U.S. Department of Energy, grant no. DE-FG02-01ER41200. 29 References [1] P. M. Nadolsky et al., arXiv:0802.0007 [hep-ph]. [2] A. D. Martin, R. G. Roberts, W. J. Stirling and R. S. Thorne, Eur. Phys. J. C 23 (2002) 73 [arXiv:hep-ph/0110215]. [3] A. D. Martin, R. G. Roberts, W. J. Stirling and R. S. Thorne, Phys. Lett. B 604 (2004) 61 [arXiv:hep-ph/0410230]. [4] A. D. Martin, W. J. Stirling, R. S. Thorne and G. Watt, Phys. Lett. B 652 (2007) 292 [arXiv:0706.0459 [hep-ph]]. [5] S. Alekhin, Phys. Rev. D 68 (2003) 014002 [arXiv:hep-ph/0211096]. [6] S. Alekhin, K. Melnikov and F. Petriello, Phys. Rev. D 74 (2006) 054033 [arXiv:hep-ph/0606237]. [7] S. Chekanov et al. [ZEUS Collaboration], Eur. Phys. J. C 42 (2005) 1 [arXiv:hep-ph/0503274]. [8] C. Adloff et al. [H1 Collaboration], Eur. Phys. J. C 30 (2003) 1 [arXiv:hep-ex/0304003]. [9] J. Pumplin et al., Phys. Rev. D 65 (2002) 014013 [arXiv:hepph/0101032]. [10] A. D. Martin, R. G. Roberts, W. J. Stirling and R. S. Thorne, Eur. Phys. J. C 35 (2004) 325 [arXiv:hep-ph/0308087]. [11] A. D. Martin, R. G. Roberts, W. J. Stirling and R. S. Thorne, Eur. Phys. J. C 28 (2003) 455 [arXiv:hep-ph/0211080]. [12] K. J. Eskola, H. Paukkunen and C. A. Salgado, JHEP 0807 (2008) 102 [arXiv:0802.0139 [hep-ph]]. 30 [13] K. J. Eskola, V. J. Kolhinen, H. Paukkunen and C. A. Salgado, JHEP 0705 (2007) 002 [arXiv:hep-ph/0703104]. [14] M. Hirai, S. Kumano and T. H. Nagai, Phys. Rev. C 70 (2004) 044905 [arXiv:hep-ph/0404093]. [15] M. Hirai, S. Kumano and T. H. Nagai, Phys. Rev. C 76 (2007) 065207 [arXiv:0709.3038 [hep-ph]]. [16] M. Vanderhaeghen, P. A. M. Guichon and M. Guidal, Phys. Rev. D 60 (1999) 094017 [arXiv:hep-ph/9905372]. [17] S. Ahmad, H. Honkanen, S. Liuti and S. K. Taneja, Phys. Rev. D 75 (2007) 094003 [arXiv:hep-ph/0611046]. [18] S. Ahmad, H. Honkanen, S. Liuti and S. K. Taneja, arXiv:0708.0268 [hep-ph]. [19] D. Stump et al., Phys. Rev. D 65 (2002) 014012 [arXiv:hep-ph/0101051]. [20] J. Pumplin, AIP Conf. Proc. 792, 50 (2005) [arXiv:hep-ph/0507093]. [21] A. Djouadi and S. Ferrag, Phys. Lett. B 586 (2004) 345 [arXiv:hepph/0310209]. [22] M. Ubiali, arXiv:0809.3716 [hep-ph]. [23] R. D. Ball et al., arXiv:0808.1231 [hep-ph]. [24] L. Del Debbio, S. Forte, J. I. Latorre, A. Piccione and J. Rojo [NNPDF Collaboration], JHEP 0703 (2007) 039 [arXiv:hep-ph/0701127]. [25] W. T. Giele, S. A. Keller and D. A. Kosower, arXiv:hep-ph/0104052. [26] G. Cowan, Prepared for 14th International Workshop on Deep Inelastic Scattering (DIS 2006), Tsukuba, Japan, 20-24 Apr 2006 31 [27] G. D’Agostini, “Bayesian Reasoning in Data Analysis: A Critical Introduction”, World Scientific Publishing, Singapore (2003). [28] T. Kohonen, Self-organizing Maps, Third Edition, Springer (2001). [29] http://www.cis.hut.fi/projects/somtoolbox [30] J. S. Lange, H. Freiesleben and P. Hermanowski, Nucl. Instrum. Meth. A 389 (1997) 214. [31] K. H. Becks, J. Drees, U. Flagmeyer and U. Muller, Nucl. Instrum. Meth. A 426 (1999) 599. [32] J. S. Lange, C. Fukunaga, M. Tanaka and A. Bozek, Nucl. Instrum. Meth. A 420 (1999) 288. [33] Charu C. Aggarwal , Alexander Hinneburg , Daniel A. Keim, Proceedings of the 8th International Conference on Database Theory, p.420-434, January 04-06, 2001 [34] http://www.ai-junkie.com/ann/som/som1.html and http://davis.wpi.edu/~ matt/courses/soms/ [35] C. Adloff et al. [H1 Collaboration], Eur. Phys. J. C 21 (2001) 33 [arXiv:hep-ex/0012053]. [36] A. C. Benvenuti et al. [BCDMS Collaboration], Phys. Lett. B 223 (1989) 485. [37] A. C. Benvenuti et al. [BCDMS Collaboration], Phys. Lett. B 237 (1990) 592. [38] S. Chekanov et al. [ZEUS Collaboration], Eur. Phys. J. C 21 (2001) 443 [arXiv:hep-ex/0105090]. 32 [39] J. Pumplin, D. R. Stump, J. Huston, H. L. Lai, P. Nadolsky and W. K. Tung, JHEP 0207 (2002) 012 [arXiv:hep-ph/0201195]. [40] http://www.nikhef.nl/~ h24/qcdnum/ [41] H. L. Lai et al. [CTEQ Collaboration], Eur. Phys. J. C 12 (2000) 375 [arXiv:hep-ph/9903282]. [42] A. D. Martin, R. G. Roberts, W. J. Stirling and R. S. Thorne, Phys. Lett. B 531 (2002) 216 [arXiv:hep-ph/0201127]. [43] M. Gluck, E. Reya and A. Vogt, Eur. Phys. J. C 5 (1998) 461 [arXiv:hepph/9806404]. [44] H. L. Lai et al., Phys. Rev. D 55, 1280 (1997) [arXiv:hep-ph/9606399]. 33 

Power Grid Simulation using Matrix Exponential Method with Rational Krylov Subspaces arXiv:1309.5333v3 [cs.CE] 14 Oct 2013 Hao Zhuang, Shih-Hung Weng, and Chung-Kuan Cheng Department of Computer Science and Engineering, University of California, San Diego, CA, 92093, USA hao.zhuang@cs.ucsd.edu, s2weng@ucsd.edu, ckcheng@ucsd.edu Abstract— One well-adopted power grid simulation methodology is to factorize matrix once and perform only backward/forward substitution with a deliberately chosen step size along the simulation. Since the required simulation time is usually long for the power grid design, the costly factorization is amortized. However, such fixed step size cannot exploit larger step size for the low frequency response in the power grid to speedup the simulation. In this work, we utilize the matrix exponential method with the rational Krylov subspace approximation to enable adaptive step size in the power grid simulation. The kernel operation in our method only demands one factorization and backward/forward substitutions. Moreover, the rational Krylov subspace approximation can relax the stiffness constraint of the previous works [12][13]. The cheap computation of adaptivity in our method could exploit the long low-frequency response in a power grid and significantly accelerate the simulation. The experimental results show that our method achieves up to 18X speedup over the trapezoidal method with fixed step size. I. I NTRODUCTION Power grid simulation is a very essential and computational heavy tasks during VLSI design. Given current stimulus and the power grid structure, designers could verify and predict the worst-case voltage noise through the simulation before signing off their design. However, with the huge size of modern design, power grid simulation is a timeconsuming process. Moreover, manifesting effects from the package and the board would require longer simulation time, e.g., up to few µs, which worsens the performance of the power grid simulation. Therefore, an efficient power grid simulation is always a demand from industry. Conventionally, the power grid simulation is based on the trapezoidal method where the major computation is to solve a linear system by either iterative approaches or direct methods. The iterative methods usually suffer from the convergence problem because of the ill-conditioned matrix from the power grid design. On the other hand, the direct methods, i.e., Cholesky or LU factorizations, are more general for solving a linear system. Despite the huge memory demanding and computational effort, with a carefully chosen step size, the power grid simulation could perform only one factorization at the beginning while the rest of operations are just backward/forward substitutions. Since a power grid design usually includes board and package models, a long simulation time is required to manifest the lowfrequency response. Hence, the cost of expensive factorization can be amortized by many faster backward/forward substitutions. Such general factorization and fixed step size strategy[14][15][16][17] is widely adopted in industry. The matrix exponential method (MEXP) for the circuit simulation has better accuracy and adaptivity because of the analytical formulation and the scaling invariant Krylov subspace approximation[12][13]. Unlike the fixed step size strategy, MEXP could dynamically adjust the step size to exploit the low-frequency response of the power grid without expensive computation. However, the step size in MEXP is limited by the stiffness of circuit. This constraint would drag the overall performance of MEXP for the power grid simulation. In this work, we tailor MEXP using rational Krylov subspace for the power grid simulation with adaptive time stepping. The rational Krylov subspace uses (I − γA)−1 as the basis instead of A used in the conventional Krylov subspaces, where I is an identity matrix and γ is a predefined parameter. The rational basis limits the spectrum of a circuit system, and emphasizes small magnitude eigenvalues, which are important to exponential function, so that the exponential of a matrix can be accurately approximated . As a result, MEXP with rational Krylov subspace can enjoy benefits of the adaptivity and the accuracy of MEXP. Even though the rational Krylov subspace still needs to solve a linear system as the trapezoidal method does, MEXP can factorize the matrix only once and then constructs the rest of rational Krylov subspaces by backward/forward substitutions. Therefore, MEXP can utilize its capability of adaptivity to accelerate the simulation with the same kernel operations as the fixed step size strategy. Overall, our MEXP enables adaptive time stepping for the power grid simulation with only one LU factorization, and allows scaling large step size without compromising the accuracy. The experimental results demonstrate the effectiveness of MEXP with adaptive step size. The industrial power grid designs can be accelerated 17X on average compared to the trapezoidal method. The rest of paper is organized as follows. Section II presents the background of the power grid simulation and MEXP. Sections III and IV show the theoretical foundation of the rational Krylov subspace and our adaptive step scheme for the power grid simulation, respectively. Section V presents experimental results of several industrial power grid designs. Finally, Section VI concludes this paper. II. P RELIMINARY A. Power Grid Formulation A power grid can be formulated as a system of differential equations via modified nodal analysis as below: Cẋ(t) = −Gx(t) + Bu(t), (1) where matrix C describes the capacitance and inductance, matrix G represents the resistance and the incidence between voltages and currents, and matrix B indicates locations of the independent sources. Vector x(t) describes the nodal voltages and branch currents at time t, and vector u(t) represents the corresponding supply voltage and current sources associated to different blocks. In this work, we assume those input sources are given and in the format of piece-wise linear. B. Matrix Exponential Method MEXP[12][13] is based on the analytical solution of (1). With initial solution from the DC analysis via direct[3] or iterative approaches[7], the equation of MEXP from t to t+h can be expressed as Z h x(t + h) = eAh x(t) + eA(h−τ ) b(t + τ )dτ. (2) 0 where A = −C−1 G, and b(t) = C−1 Bu(t), when C is not singular. Assuming that input u(t) is piece-wise linear (PWL), we integrate the last term in (2) analytically, turning the solution into the sum of three terms associated with matrix exponential operators, x(t + h) = eAh x(t) + (eAh − I)A−1 b(t) + (eAh − (Ah + I))A−2 b(t + h) − b(t) . (3) h Eqn. (3) has three matrix exponential terms, which are generally referred as ϕ-functions of the zero, first and second order [6]. It has been shown in [1, Theorem 2.1] that one can obtain the sum of them in one shot by computing the exponential of a slightly larger (n+p)×(n+p) matrix, where n is the dimension of A, and p is the order of the ϕ-functions (p = 2 in (3)). Thus, (3) can be rewritten into     e x(t) , (4) x(t + h) = In 0 eAh e2 with e = A J=   A 0 0 0 W J  1 , 0  , W= e2 = h  b(t+h)−b(t) h 0 1 b (t)  i (5) To keep the notations simple, we use v to represent [x(t) e2 ]T in the rest of paper, respectively. Note that the kernel computation of MEXP is to derive the exponential of a matrix, i.e., eA v, which is approximated by the Krylov subspace method in works [13][12]. The Krylov subspace method has better scalability mainly from its sparse matrix-vector product centric computation. However, such approximation is only better for those eigenvalues with small magnitude, which means the maximum step size of MEXP in a stiff circuit has to be constrained in order to maintain the accuracy of approximation. In the following section, we will present how the rational basis could relax the stiffness constraint. III. MEXP WITH R ATIONAL K RYLOV S UBPSACE In [13][12], Eqn. (4) is calculated via the Krylov subspace method using Arnoldi process. The subspace is defined as e v) = span{v, Av, e ··· ,A e m−1 v}, Km (A, (6) where v is an initial vector. The Arnoldi process approximates the eigenvalues with large magnitude well. But when handling a stiff circuit system, the formed matrix usually contains many eigenvalues e with small magnitude. Besides, eAh is mostly determined by the eigenvalues with smallest magnitudes and their corresponding invariant subspaces. In this scenario, due to the existence of eigenvalues e the Arnoldi process for Eqn. (6) requires with large magnitude in A, large m to capture the important eigenvalues (small magnitudes) and invariant spaces for exponential operator. Therefore, the time steps in MEXP has to be small enough to capture the important eigenvalues. This suggests us transforming the spectrum to intensify those eigenvalues with small magnitudes and corresponding invariant subspaces. We make such transformation based on the idea of rational Krylov subspace method[5][11]. The details are presented in the following subsections. A. Rational Krylov Subspaces Approximation of MEXP For the purpose of finding the eigenvalues with smallest magnitude e −1 , instead of using A e first, we uses a preconditioner (I − γ A) directly. It is known as the rational Krylov subspace[5][11]. The formula for the rational Krylov subspace is e −1 , v) = span{v, (I − γ A) e −1 v, · · · , Km ((I − γ A) e −(m−1) v}, (I − γ A) (7) where γ is a predefined parameter. The Arnoldi process constructs Vm+1 and Hm+1,m , and the relationship is given by e −1 Vm = Vm Hm,m + hm+1,m vm+1 eTm , (I − γ A) (8) where em is the m-th unit vector. Matrix Hm,m is the first m × m square matrix of an upper Hessenberg matrix of Hm+1,m , and hm+1,m is its last entry. Vm consists of [v1 , v2 , · · · , vm ] , and vm+1 is its last vector. After re-arranging (8) and given a time step e h, the matrix exponential eAh v can be calculated as e eAh v e e ≈ Vm Vm T eAh v = kvk2 Vm Vm T eAh Vm e1 = kvk2 Vm eαHm,m e1 , e (9) h e m,m = I − H−1 where H m,m , α = γ is the adjustable parameters for control of adaptive time step size in Section IV. Note that in practice, instead of computing (I − γA)−1 directly, we only need to solve (C + γG)−1 Cv, which can be achieved by one LU factorization at beginning. Then the construction of the following subspaces is by backward/forward substitutions. This strategy is also presented in [5][11]. Intuitively, the “shift-andinvert” operation would intensify the eigenvalues with small magnitudes and minify the eigenvalues with large magnitudes. By doing so, the Arnoldi process could capture those eigenvalues important to the exponential operator, which originally cannot be manifested with small m in the conventional Krylov subspace. We would like to point e out that the error bound for Eqn. (9) does not longer depend on kAhk e as [13]. It is only the first (smallest magnitude) eigenvalue of A. We observe that large α provides less absolute error under the same dimension m. An intuitive explanation is also given by [11], the larger α combined with exponential operators, the relatively smaller portion of the eigenvalues with smallest magnitude determine the final vector. Within the assumption of piecewise linear in Eqn. (3), our method can step forward as much as possible to accelerate simulation, and still maintain the high accuracy. The sacrifice resides in the small time step when more eigenvalues determine the final vector. So we should choose a appropriate parameter γ or increase the order m to balance the accuracy and efficiency. Even though the increasing m results more backward/forward substitutions, the m is still quite small in the power grid simulation. Therefore, it does not degrade our method too much. The formula of posterior error estimation is required for controlling adaptive step size. We use the formula derived from [11], kvk2 e m,m αH e m+1 eTm H−1 e1 (10) hm+1,m (I − γ A)v m,m e γ The formula provides a good approximation for the error trend with respect to m and α in our numerical experiment. B. Block LU factorization err(m, α) = In practical numerical implementation, in order to avoid direct inversion of C to form A in Eqn. (7), the equation (C + γG)−1 C is used. Correspondingly, for Eqn. (4), we uses the equations e = where C h  Bu(t+h)−Bu(t) h e − γ G) e −1 C e (C   C 0 −G e = ,G 0 I 0 i Bu (t) (11)  f W f = , and W J The Arnoldi process based on Eqn. (11) actually only requires to solve vk+1 with vk . The linear system is expressed as e − γ G)v e k+1 = Cv e k, (C (12) e − γG e = LU. C (13) where vk and vk+1 are k-th and (k + 1)-th basis in the rational f changes with inputs during the simulation, the Krylov subspace. If W Arnoldi process has to factorize a matrix every time step. However, e stay the same for this linear it is obvious that the majority of G system. To take advantage of this property, a block LU factorization is devised here to avoid redundant calculation. The goal is to obtain the lower triangular L and the upper triangular U matrices: At the beginning of simulation, after LU factorization of C + γG = Lsub Usub , we obtain the lower triangular sub-matrix Lsub , and upper triangular sub-matrix Usub . Then Eqn. (13) only needs updating via     f Lsub 0 Usub −γL−1 sub W , (14) , U= L= 0 I 0 IJ where IJ = I − γJ is an upper triangular matrix. Assume we have vk , the following equations further reduce operation L−1 sub and construct vector vk+1 : z1 = [C, 0]vk , z2 = [0, I]vk ; y2 = f I−1 J z2 , Lsub Usub y1 = z1 + γ W y2 . Then, we obtain vk+1 = [y1 , y2 ]T . By doing this, it only needs one LU factorization at the beginning of simulation, and with cheap updates for the L and U at each time step during transient simulation. IV. A DAPTIVE T IME S TEP C ONTROL The proposed MEXP can significantly benefit from the adaptive time stepping because the rational Krylov subspace approximation relaxes the stiffness constraint as well as preserves the scaling invariant property. As a result, MEXP can effortlessly adjust the step size to different scale, during the simulation. Such adaptivity is particularly helpful in the power grid where the voltage noise includes the high- to low-frequency responses from die, package and board. Our adaptive step scheme is to step forward as much as possible so that MEXP can quickly finish the simulation. With the insight from Eqn. (9), MEXP can adjust α to calculate results of required step sizes with only one Arnoldi process. However, even though the rational Krylov subspace could scale robustly, the step size in MEXP is restrained from input sources. As shown in Eqn. (3), MEXP has to guarantee constant slope during a stepping, and hence the maximum allowed step size h at every time instant is limited. Our scheme will first determine h from inputs at time t and construct the rational Krylov subspace from x(t). Then, x within interval [t, t + h] are calculated through the step size scaling. Algorithm 1 shows MEXP with adaptive step control. In order to comply with the required accuracy during the simulation, the allowed error err(m, α) at certain time instant t is defined as err ≤ ETT ol h where ET ol is the error tolerance in the whole simulation process, T is the simulation time span, h is the step size at time t, and err is the posterior error of MEXP from Eqn. (10). Hence, when we construct the rational Krylov subspace, we will increase m until the err(m, α) satisfies the error tolerance. The complexity of MEXP with adaptive time stepping is mainly determined by the total number of required backward/forward substitutions during the Psimulation process. The number of total substitution operations is N i=0 mi where N is total time steps, and mi is required dimension of the rational Krylov subspace at time step i. Compared to the trapezoidal method where the number of substitution operations depends only on the fixed step size, MEXP could use less substitution operations as long as the maximum allowed step size h is much larger than the fixed step size. Algorithm 1: MEXP with Adaptive Step Control Input: C, G, B, u(t), τ , error tolerance ET ol and simulation time T Output: x(t) 1 t = 0; x(0) = DC analysis; 2 [Lsub , Usub ] =LU(C + γG); 3 while t ≤ T do 4 Compute maximum allowed step size h from u(t) and α = hγ ; 5 Construct Hm,m , Vm,m , err by Arnoldi process and (10) until err(m, α) ≤ ETT ol h; 6 Compute x(t + h) by (9); 7 t = t + h; 8 end Our experiments in the following section demonstrates it is usually the case for the power grid simulation. V. E XPERIMENTAL R ESULTS In this section, we compare performance of the power grid simulation by MEXP and the trapezoidal method (TR). MEXP with adaptive step size control follows Algorithm 1. We predefine γ e.g. 10−10 here. and restrict the maximum allowed step size within 1ns to have enough time instants to plot the figure. It is possible to have more fine-grain time instants, e.g., 10ps, with only negligible cost by adjusting α in Eqn. (9). TR is in fixed step size h in order to minimize the cost of LU factorization. Both methods only perform factorization once for transient simulation, and rest of operations is mainly backward/forward substitution. We implement both methods in MATLAB and use UMFPACK package for LU factorization. Note that even though previous works[2][9] show that using iterative approach in TR could also achieve adaptive step control, long simulation time span in power grid designs make direct method with fixed step size more desirable[14][15][17]. The experiments are performed on a Linux workstation with an Intel Core i7-920 2.67GHz CPU and 12GB memory. The power grid consists of four metal layers: M1, M3, M6 and RDL. The physical parameters of each metal layer is listed in Table I. The package is modeled as an RL series at each C4 bump, and the board is modeled as a lumped RLC network. The specification of each PDN design is listed in Table II where the size of each design ranges from 45.7K to 7.40M. TABLE I W IDTHS AND PITCHES OF METAL LAYERS IN THE PDN DESIGN (µm). pitch 2.5 M1 width 0.2 pitch 8.5 M3 width 0.25 pitch 30 M6 width 4 RDL pitch width 400 30 TABLE II S PECIFICATIONS OF PDN D ESIGNS Design D1 D2 D3 D4 Area (mm2 ) 0.352 1.402 2.802 5.002 #R 23221 348582 1468863 3748974 #C 15193 228952 965540 2467400 #L 15193 228952 965540 2464819 #Nodes 45.7K 688K 2.90M 7.40M In order to characterize a PDN design, designers can rely on the simulation result of impulse response of the PDN design. Many previous works[4][10] have proposed different PDN analysis based on the impulse response. The nature of impulse response of the PDN design, which contains low-, mid- and high-frequency components, TABLE III S IMULATION RUNTIME OF PDN DESIGNS Design DC(s) D1 D2 D3 D4 0.710 12.2 69.6 219 TR (h = 10ps) LU(s) Total 0.670 44.9m 15.6 15.4h 91.6 76.9h 294 204h Our MEXP (γ = 10−10 ) LU(s) Total Speedup 0.680 2.86m 15.7 15.5 54.6m 16.9 93.3 4.30h 17.9 299 11.3h 18.1 VI. C ONCLUSION For large scale power grid simulation, we propose an MEXP framework using two methods rational Krylov subspace approximation and adaptive time stepping technique. The former method can relax stiffness constraint of [12][13]. The later one helps adaptively exploit low-, mid-, and high-frequency property in simulation of industrial PDN designs. In the time-consuming impulse response simulation, the proposed method achieve more than 15 times speedup on average over the widely-adopted fixed-step trapezoidal method. MEXP TR HSPICE 0.1 0.09 0.08 0.07 Voltage can significantly enjoy the adaptive step size in MEXP. We would also like to mention that the impulse response based analysis is not only for the PDN design, but also for worst-case eye opening analysis in the high speed interconnect [8]. The impulse response can be derived from the simulation result of a step input from 0V to 1V with a small transition time. Hence, we inject a step input to each PDN design and compare the performance of MEXP and TR. The transition time of the step input and the simulation time span is 10ps and 1µs for observing both high- and low-frequency responses. Table III shows the simulation runtime of MEXP and TR where the fixed step size is set as 10ps to comply with the transition time. In the table, “DC”, “LU” and “Time” indicate the runtime for DC analysis, LU factorization and the overall simulation, respectively. DC analysis is also via the LU factorization. We can also adopt other techniques[14][15][16] to improve the performance of DC analysis for both methods. It is noted that these cases are very stiff and with singular matrix C. We do not use the method[12][13] on the benchmarks, because that it cannot handle the singular C in these industrial PDN design without regularization. It is worth pointing out that even after regularization[13], the stiffness still causes large m series for matrix exponential evaluation. For example, we construct a simple RC mesh network with 2500 nodes. The extreme values of this circuit are Cmin = 5.04 × 10−19 , Cmax = 1.00 × 10−15 , Gmin = 1.09 × 10−2 , and Gmax = 1.00 × 102 . The corresponding maximum eigenvalue of −C−1 G is −1.88 × 109 and minimum eigenvalue is Re(λmin ) −3.98 × 1017 . The stiffness is Re(λ = 2.12 × 108 . During max ) simulation of 1ns time span, with a fixed step size 10ps, MEXP based on [12][13] costs average and peak dimensions of Krylov subspace mavg = 115, and mpeak =264, respectively. Our MEXP uses rational Krylov subspaces, which only need mavg =3.11, mpeak =10 and lead to 224X speedup in total runtime. In these test cases, our MEXP has significant speedup over TR because it can adaptively exploit much large step size to simulate the design whereas TR can only march with 10ps time step for whole 1µs time span. The average speedup is 17X. Fig. 1 shows the simulation result of design D1 at a node on M1. As we can see, the result by MEXP and TR are very close to the result of HSPICE, which is as our reference result here. The errors of MEXP and TR to HSPICE are 7.33×10−4 and 7.47×10−4 . This figure also demonstrates that a PDN design has low-, mid- and high-freqeuncy response so that long simulation time span is necessary, meanwhile, small time steps are required during the 20ns in the beginning. 0.06 0.05 0.04 0.07 0.03 0.06 0.02 0.05 0 2 4 0 0 6 −8 0.01 x 10 0.1 0.2 0.3 Fig. 1. 0.4 0.5 Time 0.6 0.7 0.8 0.9 1 −6 x 10 Result of D1 VII. ACKNOWLEDGEMENTS This work was supported by NSF CCF-1017864. R EFERENCES [1] A. H. Al-Mohy and N. J. Higham. Computing the action of the matrix exponential, with an application to exponential integrators. SIAM Journal on Scientific Computing, 33(2):488–511, 2011. [2] T.-H. Chen and C. C.-P. Chen. Efficient large-scale power grid analysis based on preconditioned krylov-subspace iterative methods. In Proc. Design Automation Conference, pages 559–562, 2001. [3] T. A. Davis. Direct Method for Sparse Linear Systems. SIAM, 2006. [4] P. Du, X. Hu, S.-H. Weng, A. Shayan, X. Chen, A. Ege Engin, and C. K. Cheng. Worst-case noise prediction with non-zero current transition times for early power distribution system verification. In Intl. Symposium on Quality Electronic Design, pages 624–631, 2010. [5] I. Moret and P. Novati. Rd-rational approximations of the matrix exponential. BIT Numerical Mathematics, 44(3):595–615, 2004. [6] J. Niesen and W. M. Wright. A Krylov subspace algorithm for evaluating the ϕ-functions appearing in exponential integrators. ACM Trans. Math. Software. in press. [7] Y. Saad. Iteravite Methods for Sparse Linear Systems. SIAM, 2003. [8] R. Shi, W. Yu, Y. Zhu, C. K. Cheng, and E. S. Kuh. Efficient and accurate eye diagram prediction for high speed signaling. In Proc. Intl. Conf. Computer-Aided Design, pages 655–661, 2008. [9] H. Su, E. Acar, and S. R. Nassif. Power grid reduction based on algebraic multigrid principles. In Proc. Design Automation Conference, pages 109–112, 2003. [10] W. Tian, X.-T. Ling, and R.-W. Liu. Novel methods for circuit worst-case tolerance analysis. IEEE Trans. on Circuits and Systems I: Fundamental Theory and Applications, 43(4):272–278, 1996. [11] J. van den Eshof and M. Hochbruck. Preconditioning lanczos approximations to the matrix exponential. SIAM Journal on Scientific Computing, 27(4):1438–1457, 2006. [12] S.-H. Weng, Q. Chen, and C. K. Cheng. Circuit simulation using matrix exponential method. In Intl. Conf. ASIC, pages 369–372, 2011. [13] S.-H. Weng, Q. Chen, and C. K. Cheng. Time-domain analysis of largescale circuits by matrix exponential method with adaptive control. IEEE Trans. Computer-Aided Design, 31(8):1180–1193, 2012. [14] X. Xiong and J. Wang. Parallel forward and back substitution for efficient power grid simulation. In Proc. Intl. Conf. Computer-Aided Design, pages 660–663, 2012. [15] J. Yang, Z. Li, Y. Cai, and Q. Zhou. Powerrush: Efficient transient simulation for power grid analysis. In Proc. Intl. Conf. Computer-Aided Design, pages 653–659, 2012. [16] T. Yu, M. Wong, et al. Pgt solver: an efficient solver for power grid transient analysis. In Proc. Intl. Conf. Computer-Aided Design, pages 647–652, 2012. [17] M. Zhao, R. V. Panda, S. S. Sapatnekar, and D. Blaauw. Hierarchical analysis of power distribution networks. IEEE Trans. Computer-Aided Design, 21(2):159–168, 2002. 

A Novel Weighted Distance Measure for Multi-Attributed Graph Muhammad Abulaish, SMIEEE Jahiruddin∗ Department of Computer Science South Asian University New Delhi-21, India abulaish@sau.ac.in Department of Computer Science Jamia Millia Islamia New Delhi-25, India jahir.jmi@gmail.com 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 permissions@acm.org. 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 

A Machine Learning Approach to Optimal Tikhonov Regularization A Machine Learning Approach to Optimal Tikhonov Regularization I: Affine Manifolds Ernesto De Vito (corresponding author) devito@dima.unige.it arXiv:1610.01952v2 [math.NA] 12 Oct 2017 DIMA, Università di Genova, Via Dodecaneso 35, Genova, Italy Massimo Fornasier massimo.fornasier@ma.tum.de Technische Universität München, Fakultät Mathematik, Boltzmannstrasse 3 D-85748, Garching bei München, Germany Valeriya Naumova valeriya@simula.no Simula Research Laboratory, Martin Linges vei 25, Fornebu, Norway Abstract Despite a variety of available techniques the issue of the proper regularization parameter choice for inverse problems still remains one of the relevant challenges in the field. The main difficulty lies in constructing a rule, allowing to compute the parameter from given noisy data without relying either on any a priori knowledge of the solution or on the noise level. In this paper we propose a novel method based on supervised machine learning to approximate the high-dimensional function, mapping noisy data into a good approximation to the optimal Tikhonov regularization parameter. Our assumptions are that solutions of the inverse problem are statistically distributed in a concentrated manner on (lowerdimensional) linear subspaces and the noise is sub-gaussian. We show that the number of previously observed examples for the supervised learning of the optimal parameter mapping scales at most linearly with the dimension of the solution subspace. Then we also provide explicit error bounds on the accuracy of the approximated parameter and the corresponding regularization solution. Even though the results are more of theoretical nature, we present a recipe for the practical implementation of the approach, we discuss its computational complexity, and provide numerical experiments confirming the theoretical results. We also outline interesting directions for future research with some preliminary results, confirming their feasibility. Keywords: Tikhonov regularization, parameter choice rule, sub-gaussian vectors, high dimensional function approximations, concentration inequalities. 1. Introduction In many practical problems, one cannot observe directly the quantities of most interest; instead their values have to be inferred from their effects on observable quantities. When this relationship between observable Y and the quantity of interest X is (approximately) linear, as it is in surprisingly many cases, the situation can be modeled mathematically by the equation Y = AX (1) for A being a linear operator model. If A is a “nice”, easily invertible operator, and if the data Y are noiseless and complete, then finding X is a trivial task. Often, however, the 1 De Vito, Fornasier and Naumova mapping A is ill-conditioned or not invertible. Moreover, typically (1) is only an idealized version, which completely neglects any presence of noise or disturbances; a more accurate model is Y = AX + η, (2) in which the data are corrupted by an (unknown) noise. In order to deal with this type of reconstruction problem a regularization mechanism is required (Engl et al., 1996). Regularization techniques attempt to incorporate as much as possible an (often vague) a priori knowledge on the nature of the solution X. A well-known assumption which is often used to regularize inverse problems is that the solution belongs to some ball of a suitable Banach space. Regularization theory has shown to play its major role for solving infinite dimensional inverse problems. In this paper, however, we consider finite dimensional problems, since we intend to use probabilistic techniques for which the Euclidean space is the most standard setting. Accordingly, we assume the solution vector X ∈ Rd , the linear model A ∈ Rm×d , and the datum Y ∈ Rm . In the following we denote with kZk the Euclidean norm of a vector Z ∈ RN . One of the most widely used regularization approaches is realized by minimizing the following, so-called, Tikhonov functional min kAz − Y k2 + α kzk2 . z∈Rd (3) with α ∈ (0, +∞). The regularized solution Z α := Z α (Y ) of such minimization procedure is unique. In this context, the regularization scheme represents a trade-off between the accuracy of fitting the data Y and the complexity of the solution, measured by a ball in Rd with radius depending on the regularization parameter α. Therefore, the choice of the regularization parameter α is very crucial to identify the best possible regularized solution, which does not overfit the noise. This issue still remains one of the most delicate aspects of this approach and other regularization schemes. Clearly the best possible parameter minimizes the discrepancy between Z α and the solution X α∗ = arg min α∈(0,+∞) kZ α − Xk. Unfortunately, we usually have neither access to the solution X nor to information about the noise, for instance, we might not be aware of the noise level kηk. Hence, for determining a possible good approximation to the optimal regularization parameter several approaches have been proposed, which can be categorized into three classes • A priori parameter choice rules based on the noise level and some known “smoothness” of the solution encoded in terms, e.g., of the so-called source condition (Engl et al., 1996); • A posteriori parameter choice rules based on the datum Y and the noise level; • A posteriori parameter choice rules based exclusively on the datum Y or, the so-called, heuristic parameter choice rules. 2 A Machine Learning Approach to Optimal Tikhonov Regularization For the latter two categories there are by now a multitude of approaches. Below we recall the most used and relevant of them, indicating in square brackets their alternative names, accepted in different scientific communities. In most cases, the names we provide are the descriptive names originally given to the methods. However, in a few cases, there was no original name, and, to achieve consistency in the naming, we have chosen an appropriate one, reflecting the nature of the method. We mention, for instance, (transformed/modified) discrepancy principle [Raus-Gfrerer rule, minimum bound method]; monotone error rule; (fast/hardened) balancing principle also for white noise; quasi-optimality criterion; L-curve method; modified discrepancy partner rule [Hanke-Raus rule]; extrapolated error method; normalized cumulative periodogram method; residual method; generalized maximum likelihood; (robust/strong robust/modified) generalized cross-validation. Considering the large number of available parameter choice methods, there are relatively few comparative studies and we refer to (Bauer and Lukas, 2011) for a rather comprehensive discussion on their differences, pros and contra. One of the features which is common to most of the a posteriori parameter choice rules is the need of solving (3) multiple times for different values of the parameters α, often selected out of a conveniently pre-defined grid. In this paper, we intend to study a novel, data-driven, regularization method, which also yields approximations to the optimal parameter in Tikhonov regularization. After an offline learning phase, whose complexity scales at most algebraically with the dimensionality of the problem, our method does not require any additional knowledge of the noise level; the computation of a near-optimal regularization parameter can be performed very efficiently by solving the regularization problem (3) only a moderated amount times, see Section 6 for a discussion on the computational complexity. In particular cases, no solution of (3) is actually needed, see Section 5. Not being based on the noise level, our approach fits into the class of heuristic parameter choice rules (Kindermann, 2011). The approach aims at employing the framework of supervised machine learning to the problem of approximating the highdimensional function, which maps noisy data into the corresponding optimal regularization parameter. More precisely, we assume that we are allowed to see a certain number n of examples of solutions Xi and corresponding noisy data Yi = AXi + ηi , for i = 1, . . . , n. For all of these examples, we are clearly capable to compute the optimal regularization parameters as in the following scheme (X1 , Y1 ) → α∗1 = arg α∈(0,+∞) (X2 , Y2 ) → α∗2 = arg α∈(0,+∞) (Xn , Yn ) → α∗n = arg α∈(0,+∞) ... ... min kZ α (Y1 ) − X1 k min kZ α (Y2 ) − X2 k min kZ α (Yn ) − Xn k (??, Y ) → ᾱ Denote µ the joint distribution of the empirical samples (Y1 , α∗1 ), . . . , (Yn , α∗n ). Were its conditional distribution R ∞ µ(· | Y ) with respect to the first variable Y very much concentrated (for instance, when 0 (α − ᾱ)q dµ(α | Y ) is very small for q ≥ 1 and for variable Y ), then 3 De Vito, Fornasier and Naumova we could design a proper regression function R : Y 7→ ᾱ := R(Y ) = Z 0 ∞ αdµ(α | Y ). Such a mapping would allow us, to a given new datum Y (without given solution!), to associate the corresponding parameter ᾱ not too far from the true optimal one α∗ , at least with high probability. We illustrate schematically this theoretical framework in Figure 1. Figure 1: Learning optimal regularization parameters from previously observed samples by approximation of the regression function R. At a first glance, this setting may seem quite hopeless. First of all, one should establish the concentration of the conditional distribution generating α∗ given Y . Secondly, even if we assume that the regression function R is very smooth, the vectors Y belong to the space Rm and the number of observations n required to learn such a function need to scale exponentially with the dimension m (Novak and Woźniakowski, 2009). It is clear that we cannot address neither of the above issues in general. The only hope is that the solutions are statistically distributed in a concentrated manner over smooth sets of lower dimension h ≪ m and the noise has also a concentrated distribution, so that the corresponding data Y are concentrated around lower-dimensional sets as well. And luckily these two assumptions are to a certain extent realistic. By now, the assumption that the possible solutions belong to a lower-dimensional set of Rd has become an important prior for many signal and image processing tasks. For instance, were solutions natural images, then it is known that images can be represented as nearly-sparse coefficient vectors with respect to shearlets expansions (Kutyniok and Labate, 2012). Hence, in this case the set of possible solutions can be stylized as a union of lower-dimensional linear subspaces, consisting of sparse vectors (Mallat, 2009). In other situations, it is known that the solution set can be stylized, at least locally, as a smooth lower-dimensional nonlinear manifold V (Allard et al., 2012; Chen et al., 2013; Little et al., 2017). Also in this case, at least locally, it is possible to approximate the solution set by means of affine lower-dimensional sets, representing tangent spaces to the manifold. Hence, the a priori knowledge that the solution is belonging to some special (often nonlinear) set 4 A Machine Learning Approach to Optimal Tikhonov Regularization should also be taken into account when designing the regularization method. In this paper, we want to show very rigorously how one can construct, from a relab to the regression tively small number of previously observed examples, an approximation R function R, which is mapping a noisy datum into a good approximation to the optimal Tikhonov regularization parameter. To this end, we assume the solutions to be distributed sub-gaussianly over a linear subspace V ⊂ Rd of dimension h ≪ m and the noise η to be also sub-gaussian. The first statistical assumption is perhaps mostly technical to allow us to provide rigorous estimates. Let us describe the method of computation as follows. We introduce the m × m noncentered covariance matrix built from the noisy measurements n X bn = 1 Yi ⊗ Yi , Σ n i=1 b n the projections onto the vector space spanned by the first most relevant and we denote by Π b eigenvectors of Σn . Furthermore, we set α bn ∈ (0, +∞) as the minimizer of b n Y k2 , min kZ α − A† Π α∈(0,+∞) where A† is the pseudo-inverse. We define b )=α R(Y bn and we claim that this is actually a good approximation, up to noise level, to R as soon as n is large enough, without incurring in the curse of dimensionality (i.e., without exponential dependency of the computational complexity on d). More precisely, we prove that, for a 2 given τ > 0, with probability greater than 1 − 6e−τ , we have that 1 ∗ kZ αbn − Xk ≤ kZ α − Xk + B(n, τ, σ), σd where σd is the smallest positive singular value of A. Let us stress that B(n, τ, σ) gets b ) = α actually small for small σ and for n = O(m × h) (see formula (33)) and R(Y bn is σ-optimal. We provide an explicit expression for B in Proposition 6. In the special case where A = I we derive a bound on the difference between the learned parameter α bn and the optimal parameter α∗ , see Theorem 12, justifying even more precisely the approximation b )=α R(Y bn ≈ α∗ = R(Y ). The paper is organized as follows: After introducing some notation and problem set-up in the next section, we provide the accuracy bounds on the learned estimators with respect to their distribution dependent counterparts in Section 3. For the special case A = I we provide an explicit bound on the difference between the learned and the optimal regularization parameter and discuss the amount of samples needed for an accurate learning in Section 4. We also exemplify the presented theoretical results with a few numerical illustrations. Section 5 provides explicit formulas by means of numerical linearization for the parameter learning. Section 6 offers a snapshot of the main contributions and presents a list of open questions for future work. Finally, Appendix A and Appendix B contain some background information on perturbation theory for compact operators, the sub-gaussian random variables, and proofs of some technical theorems, which are valuable for understanding the scope of the paper. 5 De Vito, Fornasier and Naumova 2. Setting This section presents some background material and sets the notation for the rest of the work. First, we fix some notation. The Euclidean norm of a vector v is denoted by kvk and the Euclidean scalar product between two vectors v, w by hv, wi. We denote with S d−1 the Euclidean unit sphere in Rd . If M is a matrix, M T denotes its transpose, M † the pseudo-inverse, M †k = (M † )k and kM k its spectral norm. Furthermore, ker M and ran M are the null space and the range of M respectively. For a square-matrix M , we use Tr(M ) to denote its trace. If v and w are vectors (possibly of different length), v ⊗ w is the rank one matrix with entries (v ⊗ w)ij = vi wj . Given a random vector ξ ∈ Rd , its noncentered covariance matrix is denoted by Σξ = E[ξ ⊗ ξ], which is a positive matrix satisfying the following property ran Σξ = (ker Σξ )⊥ = span{x ∈ Rd | P[ξ ∈ B(x, r)] > 0 ∀r > 0}, (4) here B(x, r) denotes the ball of radius r with the center at x. A random vector ξ is called sub-gaussian if 1 1 (5) kξkψ2 := sup sup q − 2 E[|hξ, vi|q ] q < +∞. v∈S d−1 q≥1 The value kξkψ2 is called the sub-gaussian norm of ξ and the space of sub-gaussian vectors becomes a normed vector space (Vershynin, 2012). Appendix B reviews some basic properties about sub-gaussian vectors. We consider the following class of inverse problems. Assumption 1 In the statistical linear inverse problem Y = AX + σW, the following conditions hold true: a) A is an m × d-matrix with norm kAk = 1; b) the signal X ∈ Rd is a sub-gaussian random vector with kXkψ2 = 1; c) the noise W√∈ Rm is a sub-gaussian centered random vector independent of X with √ kW kψ2 = 1/ 2 and with the noise level 0 < σ < 2; d) the covariance matrix ΣX of X has a low rank matrix, i.e., rank(ΣX ) = h ≪ d. We add some comments on the above conditions. The normalisation assumptions on kAk, kXkψ2 and kW kψ2 are stated only to simplify the bounds. They can √ always be satisfied by rescaling A, X and W and our results hold true by replacing σ with 2kW kψ2 kAk−1 kXk−1 ψ2 σ. The upper bound on σ reflects the intuition that σW is a small perturbation of the noiseless problem. 6 A Machine Learning Approach to Optimal Tikhonov Regularization Condition d) means that X spans a low dimensional subspace of Rd . Indeed, by (4) condition d) is equivalent to the fact that the vector space V = ran ΣX = span{x ∈ Rd | P[X ∈ B(x, r)] > 0 for all r > 0} (6) is an h-dimensional subspace and h is the dimension of the minimal subspace containing X with probability 1, i.e., h = min dim K, K where the minimum is taken over all subspaces K ⊂ Rd such that P[X ∈ K] = 1. We write a . b if there exists an absolute constant C such that a ≤ Cb. By absolute we mean that it holds for all the problems Y = AX + σW satisfying Assumption 1, in particular, it is independent of d, m and h. The datum Y depends only on the projection X † of X onto ker A⊥ and Z α as solutions of (3) also belong to ker A⊥ . Therefore, we can always assume, without loss of generality, for the rest of the paper that A is injective by replacing X with X † , which is a sub-gaussian random vector, and Rd with ker A⊥ . Since A is injective, rank(A) = d and we define the singular value decomposition of A by (ui , vi , σi )di=1 , so that A = U DV T or Avi = σi ui , i = 1, . . . , d, where σ1 ≥ σ2 ≥ · · · ≥ σd > 0. Since kAk = 1, clearly σ1 = 1. Furthermore, let Q be the projection onto the span{u1 , . . . , ud }, so that QA = A, and we have the decomposition Q = AA† . (7) Recalling (6), since A is now assumed injective and ΣAX = E[AX ⊗ AX] = AΣX AT , then W = ran ΣAX = (ker ΣAX )⊥ = AV, (8) and, by condition d) in Assumption 1, we have as well dim W = h. We denote by Π the projection onto W and by p = max{i ∈ {1, . . . , d} | Πui 6= 0}, (9) so that, with probability 1, ΠAX = AX and X= p X hX, vi ivi . (10) i=1 Finally, the random vectors η = σW , AX, and Y are sub-gaussian and take value in Rm , W, and Rm , respectively, with kAXkψ2 ≤ kAT kkXkψ2 = 1 kY kψ2 ≤ kAXkψ2 + σkW kψ2 ≤ 2 7 (11) De Vito, Fornasier and Naumova √ since, by Assumption 1, kAk = 1 and σ ≤ 2. For any t ∈ [0, 1] we set Z t as the solution of the minimization problem
min t kAz − Y k2 + (1 − t) kzk2 , z∈Rd (12) which is the solution of the Tikhonov functional min kAz − Y k2 + α kzk2 . z∈Rd with α = (1 − t)/t ∈ [0, +∞]. For t < 1, the solution is unique, for t = 1 the minimizer is not unique and we set Z 1 = A† Y. The explicit form of the solution of (12) is given by Z t = t(tAT A + (1 − t)I)−1 AT Y (13) d X tσi 2 + (1 − t) hY, ui ivi tσ i i=1   d X tσi2 tσi = hX, vi i + 2 hη, ui i vi , tσi2 + (1 − t) tσi + (1 − t) i=1 = which shows that Z t is also a sub-gaussian random vector. We seek for the optimal parameter t∗ ∈ [0, 1] that minimizes the reconstruction error min kZ t − Xk2 . t∈[0,1] (14) Since X is not known, the optimal parameter t∗ can not be computed. We assume that we have at disposal a training set of n-independent noisy data Y1 , . . . , Yn , where Yi = AXi + σWi , and each pair (Xi , Wi ) is distributed as (X, W ), for i = 1, . . . , n. We introduce the m × m empirical covariance matrix n X bn = 1 Σ Yi ⊗ Yi , n (15) i=1 b n the projections onto the vector space spanned by the first h-eigenvectors and we denote by Π b of Σn , where the corresponding (repeated) eigenvalues are ordered in a nonincreasing way. b n in terms of spectral gap at Remark 1 The well-posedness of the empirical realization Π the h-th eigenvalue will be given in Theorem 3, where we show that for n large enough the h + 1-th eigenvalue is strictly smaller than the h-th eigenvalue. 8 A Machine Learning Approach to Optimal Tikhonov Regularization We define the empirical estimators of X and η as b = A† Π b nY X so that, by Equation (7), b n Y ), ηb = (Y − Π and b + Qb AX η = QY. (16) (17) Furthermore, we set b tn ∈ [0, 1] as the minimizer of b 2. min kZ t − Xk t∈[0,1] b is close to X, we expect that the solution Z btn has a reconstruction error close to the If X minimum value. In the following sections, we study the statistical properties of b tn . However, we first provide some a priori information on the optimal regularization parameter t∗ . 2.1 Distribution dependent quantities We define the function t 7→ kR(t)k2 , where R(t) = Z t − X t ∈ [0, 1] is the reconstruction error vector. Clearly, the function t 7→ kR(t)k2 is continuous, so that a global minimizer t∗ always exists in the compact interval [0, 1]. Define for all t ∈ [0, 1] the d × d matrix d X (tσi2 + 1 − t) vi ⊗ vi , B(t) = tA A + (1 − t)I = T i=1 which is invertible since A is injective and its inverse is B(t)−1 = d X tσi2 i=1 1 vi ⊗ vi . +1−t Furthermore, B(t) and B(t)−1 are smooth functions of the parameter t and B ′ (t) = (AT A − I) (B(t)−1 )′ = −B(t)−2 (AT A − I). (18) Since Y = AX + η, expression (13) gives R(t) = tB(t)−1 AT Y − X = tB(t)−1 AT (AX + η) − X = B(t)−1 (tAT AX − B(t)X + tAT η) = B(t)−1 (−(1 − t)X + tAT η). Hence, kR(t)k2 = kB(t)−1 (−(1 − t)X + tAT η)k2  d  X −(1 − t)ξi + tσi νi 2 , = tσi2 + (1 − t) i=1 9 (19) De Vito, Fornasier and Naumova where for all i = 1, . . . , d ξi = hX, vi i νi = hη, ui i. In order to characterize t∗ we may want to seek it among the zeros of the following function H(t) = 1 d kZ t − Xk2 = hR(t), R′ (t)i. 2 dt Taking into account (18), the differentiation of (19) is given by R′ (t) = B(t)−1 AT Y − tB(t)−2 (AT A − I)AT Y (20) = B(t)−2 (B(t) − t(AT A − I))AT Y = B(t)−2 AT Y, so that H(t) = hAB(t)−3 (−(1 − t)X + tAT η), AX + ηi = = = d X i=1 d X i=1 d X σi −(1 − t)ξi + tσi νi (ξi σi + νi ) (tσi2 + (1 − t))3 σi ξi (ξi σi + νi ) (σi νi ξi−1 + 1)t − 1 (1 − (1 − σi2 )t)3 σi αi hi (t), i=1 where αi = ξi (σi ξi + νi ) and hi (t) = We observe that (σi νi ξi−1 +1)t−1 . (1−(1−σi2 )t)3 a) if t = 0 (i.e., α = +∞), B(0) = I, then H(0) = −kAXk2 − hAX, ηi, which is negative if kΠηk ≤ kAXk, i.e., for σ≤ kAXk . kΠW k Furthermore, by construction, E[H(0)] = − Tr(ΣAX ) < 0; b) if t = 1 (i.e., α = 0), B(1) = AT A and H(1) = hA(AT A)−3 AT η, AX + ηi = k(AAT )† ηk2 + h(AAT )† η, (AT )† Xi, 10 (21) A Machine Learning Approach to Optimal Tikhonov Regularization which is positive if k(AAT )† ηk ≥ k(AT )† Xk, for example, when σ ≥ σd kXk . |hW, ud i| Furthermore, by construction, E[H(1)] = Tr(Σ(AAT )† η ) > 0. Hence, if the noise level satisfies kAXk kXk ≤σ≤ |hW, ud i| kΠW k the minimizer t∗ is in the open interval (0, 1) and it is a zero of H(t). If σ is too small, there is no need of regularization since we are dealing with a finite dimensional problem. ∗ On the opposite side, if σ is too big, the best solution is the trivial one, i.e., Z t = 0. 2.2 Empirical quantities We replace X and η with their empirical counterparts defined in (16). By Equation (17) and reasoning as in Equation (19), we obtain bn (t) = Z t − X b R b = tB(t)−1 AT QY − X b + tAT ηb), = B(t)−1 (−(1 − t)X and bn (t)k2 = kB(t)−1 (−(1 − t)X b + tAT ηb)k2 kR !2 d X −(1 − t)ξbi + tσi νbi = , 2 + (1 − t) tσ i i=1 where for all i = 1, . . . , d b vi i ξbi = hX, Clearly, and νbi = hb η , ui i. b′ (t) = R′ (t) = B(t)−2 AT QY. R n From (21), we get c′ n (t)i b n (t) = hR bn (t), R H (23) b + tAT ηb), AT AX b + AT ηb)i = hB(t)−3 (−(1 − t)X = = d X −(1 − t)ξbi + tσi νbi i=1 d X i=1 (22) (1 − (1 − σi2 )t)3 σi α bi b hi (t), 11 (ξbi σi2 + σi νbi ), De Vito, Fornasier and Naumova where α bi = ξbi (σi ξbi + νbi ) and b hi (t) = (σi νbi ξbi−1 +1)t−1 . (1−(1−σi2 )t)3 b n , which can be useful as a different numerical An alternative form in terms of Y and Π implementation, is b n (t) = hB(t)−1 (tAT (Y − Π b n Y ) − (1 − t)A† Π b n Y ), B(t)−2 AT Y i H b n Y ) − (1 − t)QΠ b n Y , (tAAT + (1 − t)I)†3 QY i = htAAT (Y − Π b n Y ) − (1 − t)Π b n Y , (tAAT + (1 − t)I)†3 QY i. = htAAT (Y − Π bn (t)k2 always exists in [0, 1] and, for σ As for t∗ , the minimizer b tn of the function t 7→ kR in the range of interest, it is in the open interval (0, 1), so that it is a zero of the function b n (t). H 3. Concentration inequalities In this section, we bound the difference between the empirical estimators and their distribution dependent counterparts. By (8) and item d) of Assumption 1, the covariance matrix ΣAX has rank h and, we set λmin to be the smallest non-zero eigenvalue of ΣAX . Furthermore, we denote by ΠY the projection from Rm onto the vector space spanned by the eigenvectors of ΣY with corresponding eigenvalue greater than λmin /2. The following proposition shows that ΠY is close to Π if the noise level is small enough. Proposition 2 If σ 2 < λmin /4, then dim ran ΠY = h and kΠY − Πk ≤ 2σ 2 . λmin (24) Proof Since AX and W are independent and W has zero mean, then ΣY = ΣAX + σ 2 ΣW . Since ΣW is a positive matrix and W is a sub-gaussian vector satisfying (11), with the choice q = 2 in (5), we have kΣW k = sup hΣW v, vi = sup E[hW, vi2 ] ≤ 2kW k2ψ2 = 1, v∈S m−1 (25) v∈S m−1 so that kΣY − ΣAX k ≤ σ 2 < λmin /4. We now apply Proposition 15 with A = ΣAX and eigenvalues (αj )j and projections1 (Pj )j , and B = ΣY with eigenvalues (βℓ )ℓ and projections (Qℓ )ℓ . We choose j such that αj = λmin so that αj+1 = 0, Pj = Π and, by (8), dim ran Pj = dim ran Π = dim ran ΣAX = h. 1. In the statement of Proposition 15 the eigenvalues are counted without their multiplicity and ordered in a decreasing way and each Pj is the projection onto the vector space spanned by the eigenvectors with corresponding eigenvalue greater or equal than αj . 12 A Machine Learning Approach to Optimal Tikhonov Regularization Then there exists ℓ such that βℓ+1 < λmin /2 < βℓ , so that Qℓ = ΠY and it holds that dim ran ΠY = dim ran Pj = h. Finally, (45) implies (24) since αh+1 = 0. b n is the projection onto the vector space spanned by the first h-eigenvectors Recall that Π b of Σn defined by (15). 2 b n coincides with the proTheorem 3 Given τ > 0 with probability greater than 1−2e−τ , Π b n with corresponding eigenvalue jection onto the vector space spanned by the eigenvectors of Σ greater than λmin /2. Furthermore r  1 τ m 2 b kΠn − Πk . + √ +σ , (26) λmin n n provided that √ n & ( m + τ )2 max σ2 < Proof λmin . 8   64 ,1 λ2min We apply Theorem 20 with ξi = Yi and r τ m + √ ≤ min{1, λmin /8} ≤ 1, δ=C n n (27) (28) 2 by (25). Since δ2 ≤ δ, with probability greater than 1 − 2e−τ , b n − ΣAX k ≤ kΣ b n − ΣY k + kΣY − ΣAX k kΣ  r τ m +√ + σ2 ≤C n n λmin λmin λmin ≤ + = , 8 8 4 where the last inequality follows by (28). b n and As in the proof of Proposition 2, we apply Proposition 15 with A = ΣAX , B = Σ αj = λmin to be the smallest eigenvalue of ΣAX , so that Pj = Π and dim ran Pj = h. b n such that βℓ+1 < λmin /2 < βℓ and Then, there exists a unique eigenvalue βℓ of Σ b dim ran Qℓ = dim ran Pj = h. Then Qℓ = Πn and (45) implies (26). Note that the constant C depends on kY kψ2 ≤ 2 by (11), so that it becomes an absolute constant, when considering the worst case kY kψ2 = 2. If n and σ satisfy (27), the above proposition shows that the empirical covariance matrix b n has a spectral gap around the value λmin /2 and the number of eigenvector with corΣ b n is uniquely defined. responding eigenvalue greater than λmin /2 is precisely h, so that Π Furthermore, the dimension h can be estimated by observing spectral gaps in the singular b n. value decomposition of Σ If the number n of samples goes to infinity, bound (26) does not converge to zero due to term proportional to the noise level σ. However, if the random noise W is isotropic, we can improve the estimate. 13 De Vito, Fornasier and Naumova 2 Theorem 4 Assume that ΣW = Id. Given τ > 0 with probability greater than 1 − 2e−τ , r  1 τ m b n − Πk . kΠ +√ , (29) λmin n n provided that √ 16 n & ( m + τ )2 {1, 2 } λmin λmin σ2 < . 2 (30) Proof As in the proof of Proposition 2, we have that ΣY = ΣAX + σ 2 ΣW = ΣAX + σ 2 Id, where the last equality follows from the assumption that the noise is isotropic. Hence, the matrices ΣY and ΣAX have the same eigenvectors, whereas the corresponding eigenvalues are shifted by σ 2 . Taking into account that λmin is the smallest non-zero eigenvalue of ΣAX and denoted by (αj )N j=1 the eigenvalues of ΣY , it follows that there exists j = 1, . . . , N such that α1 > α2 > αj = λmin + σ 2 αj+1 = . . . = αN = σ 2 . Furthermore, denoted by Pj the projection onto the vector space spanned by the eigenvectors with corresponding eigenvalue greater or equal than αj , it holds that Π = Pj . By assumption σ 2 < λmin /2, so that ΠY = Pj = Π and, hence, dim ran Pj = h. 2 As in the proof of Theorem 3, with probability 1 − 2e−τ ,  r τ αh − αh+1 λmin m b n − ΣY k ≤ C +√ }< , < min{1, kΣ n n 4 4 b n such where n is large enough, see (30). Then, there exists a unique eigenvalue βℓ of Σ λmin b n and (45) that βℓ+1 < 2 + σ 2 < βℓ and dim ran Qℓ = dim ran Pj = h. Then Qℓ = Π implies (29). We need the following technical lemma. 2 Lemma 5 Given τ > 0, with probability greater than 1 − 4e−τ , simultaneously it holds √ √ √ √ kY k . ( h + σ m + τ ) kΠW k . ( h + τ ). (31) kXk . ( h + τ ) Proof Since X is a sub-gaussian random vector taking values in V with h = dim V, taking into account that kXkψ2 = 1, bound (50) gives √ kXk ≤ 9( h + τ ), 2 with probability greater than 1 − 2e−τ . Since W is a centered sub-gaussian random vector taking values in Rm and kW kψ2 ≤ 1, by (51) √ kW k ≤ 16( m + τ ), 14 A Machine Learning Approach to Optimal Tikhonov Regularization 2 with probability greater than 1 − e−τ . Since kAk = 1 and kY k ≤ kAXk + σkW k ≤ kXk + σkW k, 2 the first two bounds in (31) hold true with probability greater than 1−3e−τ . Since ΠW is a centered sub-gaussian random vector taking values in W with h = dim W, and kΠW kψ2 ≤ 1, by (51) √ kΠW k ≤ 16( h + τ ), 2 with probability greater than 1 − e−τ . As a consequence, we have the following bound. Proposition 6 Given τ > 0, if n and σ satisfy (27), then with probability greater than 2 1 − 6e−τ b n )Y − Πηk . B(n, τ, σ), k(Π − Π (32) where r   √ hm σ2 √ 1 m σ3 √ √ h+ h+ m+ +σ B(n, τ, σ) = + λmin n λmin n λmin λmin r r     m m 1 1 1 1 σ2 √ . +τ +σ 1+ + τ2 + λmin n λmin n λmin λmin n 1 Proof Clearly, (33) b n )Y − Πηk ≤ kΠ − Π b n kkY k + σkΠW k. k(Π − Π If (27) holds true, bounds (26) and (31) imply r  √ √ √ τ m 1 2 b + √ + σ ( h + σ m + τ ) + σ( h + τ ), k(Π − Πn )Y − Πηk . λmin n n 2 with √ probability√ greater than 1− 6e−τ . By developing the brackets and taking into account that h + m ≤ 2m, (32) holds true. Remark 7 Usually in machine learning bounds of the type (32) are considered in terms of their expectation, e.g., with respect to (X, Y ). In our framework, this would amount to the following bound i h b n )Y − Πηk Y1 , . . . , Yn . E k(Π − Π r  √ √ √ 1 τ m 2 . + √ + σ ( h + σ m) + σ h, λmin n n obtained by observing that E[kY k] ≤ E[kAkkXk] + σE[kW k], E[kXk]2 ≤ E[kXk2 ] = Tr(ΣX ) ≤ 2hkXk2ψ2 . h, and, by a similar computation, E[kW k] . √ E[kΠW k] . m √ h. Our bound (32) is much stronger and it holds in probability with respect to both the training set Y1 , . . . Yn and the new pair (X, Y ). 15 De Vito, Fornasier and Naumova Our first result is a direct consequence of the estimate (32). 2 Theorem 8 Given τ > 0, with probability greater than 1 − 6e−τ , 1 B(n, τ, σ) σd kQb η − Qηk . B(n, τ, σ) 1 ∗ b kZ tn − Xk − kZ t − Xk . B(n, τ, σ) σd bn (t)k − kR(t)k| . 1 B(n, τ, σ) sup |kR σd 0≤t≤1 b − Xk . kX provided that n and σ satisfy (27). Proof By the first identity of (10) so that b = A† ΠAX − A† Π b n (AX + η) X −X b n )AX + A† (Π − Π b n )η − A† Πη = A† (Π − Π   b n )Y − Πη , = A† (Π − Π b ≤ kX − Xk (34) 1 b n )Y − Πηk. k(Π − Π σd An application of (32) to the previous estimate gives the first bound of the statement. Similarly, we derive the second bound as follows. Equations (17), (7), and (34) yield b Qη − Qb η = A(X − X)   b = Q (Π − Πn )Y − Πη . b as we show below. The other bounds follow by estimating them by multiples of kX − Xk b By definition of tn b b b + kX − Xk b kZ tn − Xk ≤ kZ tn − Xk ∗ b + kX − Xk, b ≤ kZ t − Xk ∗ b ≤ kZ t − Xk + 2kX − Xk. Furthermore,   b n )Y − Πη , bn (t) − R(t) = X − X b = A† (Π − Π R (35) and triangle inequality gives bn (t)k − kR(t)k| ≤ sup kR bn (t) − R(t)k = kX − Xk. b sup |kR 0≤t≤1 0≤t≤1 All the remaining bounds in the statement of the theorem are now consequences of (32). 16 A Machine Learning Approach to Optimal Tikhonov Regularization Remark 9 To justify and explain the consistency of the sampling strategy for approximation of the optimal regularization parameter t∗ , let us assume that n goes to infinity bn (t)k conand σ vanishes. Under this theoretical assumption, Theorem 8 shows that kR vergences uniformly to kR(t)k with high probability. The uniform convergence implies the Γ-convergence (see Braides, 2001), and, since the domain [0, 1] is compact, Theorem 1.22 in Braides (2001) ensures that   b lim inf kRn (t)k − inf kR(t)k = 0. n→+∞ σ→0 0≤t≤1 0≤t≤1 While the compactness given by the Γ-convergence guarantees the consistency of the approximation to an optimal parameter, it is much harder for arbitrary A to provide an error bound, depending on n. For the case A = I in Section 4.4 we are able to establish very precise quantitative bounds with high probability. Remark 10 Under the conditions of Theorem 8 for all i = 1, . . . , d it holds as well 1 |ξi − ξbi | . B(n, τ, σ) σi |νi − νbi | . B(n, τ, σ). These bounds are a direct consequence of Theorem 8. The following theorem is about the uniform approximation to the derivative function H(t). 2 Theorem 11 Given τ > 0, with probability greater than 1 − 10e−τ ,   σ √ 1 √ b ( h + τ) + 4 ( d + τ) sup |Hn (t) − H(t)| . B(n, τ, σ) σp3 σd 0≤t≤1 provided that n and σ satisfy (27), where p is defined in (9). Proof Equations (22) and (35) give b n (t) − H(t) = hR bn (t) − R(t), R′ (t)i H b n )Y − Πη, (AT )† B(t)−2 AT Y i = h(Π − Π b n )Y − Πη, (tAAT + (1 − t)I)−2 QY i, = h(Π − Π where we observe that tAAT + (1 − t)I is invertible on ran Q and AT Y = AT QY . Hence, b n (t) − H(t)| ≤ k(Π − Π b n )Y − Πηk× |H
× k(tAAT + (1 − t)I)−2 AXk + σk(tAAT + (1 − t)I)−2 QW k . Furthermore, recalling that AX = ΠAX and Πui = 0 for all i > p, (31) implies that 1 √ σp kXk . ( h + τ) (tσp2 + (1 − t))2 σp3 1 √ 1 ( d + τ) kQW k . k(tAAT + (1 − t)I)−2 QW k ≤ 2 (tσd + (1 − t))2 σd4 k(tAAT + (1 − t)I)−2 AXk ≤ 17 De Vito, Fornasier and Naumova 2 hold with probability greater than 1 − 4e−τ . Bound (32) provides the desired claim. The uniform approximation result of Theorem 11 allows us to claim that any b tn ∈ [0, 1] ∗ b b such that Hn (tn ) = 0 can be attempted as a proxy for the optimal parameter t , especially if it is the only root in the interval (0, 1). b n a sum of d rational functions of polynomial numerator of degree 1 Nevertheless, being H and polynomial denominator of degree 3, the computation of its zeros in [0, 1] is equivalent to the computation of the roots of a polynomial of degree 3(d − 1) + 1 = 3d − 2. The computation cannot be done analytically for d > 2, because it would require the solution of a polynomial equation of degree larger than 4. For d > 2, we are forced to use numerical methods, but this is not a great deal as by now there are plenty of stable and reliable routines to perform such a task (for instance, Newton method, numerical computation of the eigenvalues of the companion matrix, just to mention a few). We provide below relatively simple numerical experiments to validate the theoretical results reported above. In Figure 2 we show optimal parameters t∗ and corresponding apb n (t) = 0 proximations b tn (computed by numerical solution to the scalar nonlinear equation H on [0, 1]), for n different data Y = AX + η. The accordance of the two parameters t∗ and b tn is visually very convincing and their statistical (empirical) distributions reported in Figure 3 are also very close. 1.0 0.8 0.6 0.4 0.2 10 20 30 40 50 Figure 2: Optimal parameters t∗ and corresponding approximations b tn (n = 1000) for 50 d m different new data Y = AX +η for X ∈ R and η ∈ R Gaussian vectors, d = 200, m = 60 and A ∈ R60×200 . Here we assumed that X ∈ V for V = span{e1 , . . . , e5 }. We designed the matrix in such a way that the spectrum is vanishing, i.e., σmin ≈ 0. Here we considered as noise level σ = 0.03, so that the optimal parameter t∗ is rather concentrated around 0.7. The accordance of the two parameters t∗ and b tn is visually very convincing. In the next two sections we discuss special cases where we can provide even more precise statements and explicit bounds. 18 A Machine Learning Approach to Optimal Tikhonov Regularization 4 6 5 3 4 2 3 2 1 1 0.2 0.4 0.6 0.8 1.0 1.2 0.2 0.4 0.6 0.8 1.0 1.2 Figure 3: Empirical distribution of the optimal parameters t∗ (left) and the corresponding empirical distribution of the approximating parameters b tn (right) for 1000 randomly generated data Y = AX + η with the same noise level. The statistical accordance of the two parameters t∗ and b tn is shown. 4. Toy examples In the following we specify the results of the previous in simple cases, which allow to understand more precisely the behavior of the different estimators presented in Theorem 8. 4.1 Deterministic sparse signal and Bernoulli noise in two dimensions b Z t∗ , Z btn are all performing From Theorem 8 one might conclude that the estimators X, an approximation to X with a least error proportional to B(n, τ, σ). With this in mind, b is a sufficient effort, with no need of one might be induced to conclude that computing X b t n tn . In this section considering the empirical estimator Z , hence no need for computing b we discuss precisely this issue in some detail on a concrete toy example, for which explicit computations are easily performed. Let us consider a deterministic sparse vector X = (1, 0)T ∈ R2 and a stochastic noise W = (W1 , W2 )T ∈ R2 with Bernoulli entries, i.e., Wi = ±1 with probability 1/2. This noise distribution is indeed Subgaussian with zero mean. For the sake of simplicity, we fix now an orthogonal matrix A ∈ R2×2 , which √ we express in terms of its vector-columns as 2 with√probability 1 and kXk = 1. Hence, we follows: A = (A1 |A2 ). Notice that kW k = √ may want to consider a noise level σ ∈ [−1/ 2, 1/ 2]. In view of the fact that the sign of σ would simply produce a symmetrization of the signs of the component of the noise, we consider from now on only the cases where W1 ≥ 0. The other cases are obtained simply by changing the sign of σ. We are given measurements Y = AX + σW. While X is deterministic (and this is just to simplify the computations) Y is stochastic given the randomness of the noise. Hence, noticing that Y = AX + σW = A1 + σW and observing that EW = 0, we obtain Z Z ΣY = Y Y T dP(W ) = (A1 + σW )(A1 + σW )T dP(W ) = A1 AT1 + σ 2 I. 19 De Vito, Fornasier and Naumova It is also convenient to write ΣY by means of its complete singular value decomposition   T  1 + σ2 0 A1 ΣY = (A1 A2 ) . 0 σ2 AT2 4.2 Exact projection The projection onto the first largest singular space of ΣY is simply given by Πξ = hξ, A1 iA1 . This implies that ΠY = A1 + σhW, A1 iA1 . From now on, we denote σi = σhW, Ai i, i = 1, 2. Let us stress that these parameters indicate how strong the particular noise instance W is correlated with W = span{A1 = AX}. If |σ1 | is large, then there is strong concentration of the particular noise instance on W and necessarily |σ2 | is relatively small. If, vice versa, |σ2 | is large, then there is poor correlation of the particular noise instance with W. We observe now that, being A orthogonal, it is invertible and I = AT A = A−1 A = (A−1 A1 A−1 A2 ). This means that A−1 A1 = (1, 0)T and A−1 A2 = (0, 1)T and we compute explicitly X̄ = A−1 ΠY = A−1 A1 + σ1 A−1 A1 = (1 + σ1 , 0)T . It is important to notice that for σ1 small X̄ is indeed a good proxy for X and it does not belong to the set of possible Tikhonov regularizers, i.e., the minimizers of the functional tkAZ − Y k22 + (1 − t)kZk22 (36) for some t ∈ [0, 1]. This proxy X̄ for X approximates it with error exactly computed by √ (37) kX̄ − Xk = |σ1 | ≤ 2|σ|. The minimizer of (36) is given by Z t = t(tAT A + (1 − t)I)−1 AT Y  T  A1 = t(t(AT − A−1 )A + I)−1 (A1 + σW ). AT2 In view of the orthogonality of A we have AT = A−1 , whence t(AT − A−1 )A = 0, and the simplification  T  A1 t Z = t (A1 + σW ) AT2 = t(1 + σ1 , σ2 )T 20 A Machine Learning Approach to Optimal Tikhonov Regularization Now it is not difficult to compute R(t)2 = kZ t − Xk22 = (t(1 + σ1 ) − 1)2 + t2 σ22 = ((1 + σ1 )2 + σ22 )t2 − 2t(1 + σ1 ) + 1 This quantity is optimized for 1 + σ1 (1 + σ1 )2 + σ22 t∗ = The direct substitution gives R(t∗ ) = p |σ2 | (1 + σ1 )2 + σ22 And now we notice that, according to (37) one can have either kX̄ − Xk = |σ1 | < p or kX̄ − Xk = |σ1 | > p |σ2 | (1 + σ1 )2 |σ2 | (1 + σ1 )2 ∗ + σ22 = kZ t − Xk + σ22 = kZ t − Xk ∗ very much depending on σi = σhW, Ai i. Hence, for the fact that X̄ does not belong to the ∗ set of minimizers of (36), it can perfectly happen that it is a better proxy of X than Z t . Recalling that X̄ = (1 + σ1 , 0)T and Z t = t(1 + σ1 , σ2 )T , let us now consider the error R̄(t)2 = kZ t − X̄k2 = ((1 + σ1 )2 + σ22 )t2 − 2t(1 + σ1 )2 + (1 + σ1 )2 . This is now opimized for t̄ = (1 + σ1 )2 (1 + σ1 )2 + σ22 Hence we have R(t̄)2 = kZ t̄ − Xk2 = ((1 + σ1 )2 + σ22 )t̄2 − 2t̄(1 + σ1 ) + 1 = or σ12 (1 + σ1 )2 + σ22 (1 + σ1 )2 + σ22 p σ 2 (1 + σ1 )2 + σ22 R(t̄) = p 1 (1 + σ1 )2 + σ22 √ In this case we notice that, in view of the fact that σ12 ≤ 1 for |σ| ≤ 1/ 2 one can have only p p σ12 (1 + σ1 )2 + σ22 (1 + σ1 )2 + σ22 /σ12 p kX̄ − Xk = |σ1 | ≤ p = |σ | = kZ t̄ − Xk 1 2 2 2 2 (1 + σ1 ) + σ2 (1 + σ1 ) + σ2 independently of σi = σhW, Ai i. The only way of reversing the inequality is by allowing noise level σ such that σ12 > 1, which would mean that we have noise significantly larger than the signal. 21 De Vito, Fornasier and Naumova 4.2.1 Concrete examples Let us make a few concrete examples. Let us recall our assumpution that√sign W1 = +1 W while the one of W2 is kept arbitrary. Suppose that A1 = kW k , then σ1 = σ 2 and σ2 = 0. In this case we get √ σ 2 = kX̄ − Xk = kZ t̄ − Xk, √ W and X̄ = Z t̄ . The other extreme case is given by A2 = kW , then σ1 = 0 and σ2 = ±σ 2. k p In this case X̄ = X, while R(t̄) = R(t∗ ) = |σ| 2/(1 + 2σ 2 ) > 0. An intermediate case is given precisely by A = I, which we shall investigate in more generality in Section 4.4. For this choice we have σ1 = σ and σ2 = ±σ. Notice that the sign of σ2 does not matter in the expressions of R(t̄) and R(t∗ ) because it appears always squared. Instead the sign of σ1 matters and depends on the choice of the sign of σ. We obtain R(t∗ ) = p and R(t̄) = s |σ| (1 + σ)2 + σ 2 , σ 2 (2 + (σ(2 + σ)) , 1 + 2σ(1 + σ) and, of course kX̄ − Xk = |σ|. In this case, R(t∗ ) ≤ kX̄ − Xk if and only if σ ≥ 0, while - unfortunately - R(t̄) > kX̄ − Xk for all |σ| < 1 and the inequality gets reversed only if |σ| ≥ 1. 0.8 °X-X°= Σ¤ 0.6 R Ht* L= Out[99]= Σ IΣ2 +H1+ΣL2 M 0.4 RHtL= Σ2 H2+Σ H2+ΣLL 1+2 Σ H1+ΣL 0.2 0.0 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 Figure 4: Comparison of kX̄ − Xk, R(t∗ ), and R(t̄) as a function of σ for σ1 = σ. In Figure 5 we illustrate the case of σ1 = 1.2σ and in Figure 6 the case of σ1 = 0.7σ. Let us now focus for a moment on the purely denoising case A = I. In this case, the projection Π is actually the projection onto the subspace where X is defined. While X̄ belongs also to 22 A Machine Learning Approach to Optimal Tikhonov Regularization 1.4 1.2 °X-X°=1.2 Σ¤ 1.0 0.8 R Ht* L= Out[123]= 1- H1+1.2 ΣL2 0.56 Σ2 +H1+1.2 ΣL2 0.6 RHtL= 0.4 1- 2 H1+1.2 ΣL3 0.56 Σ2 +H1+1.2 ΣL2 + H1+1.2 ΣL4 0.56 Σ2 +H1+1.2 ΣL2 0.2 0.0 -1.0 -0.5 0.0 0.5 1.0 Figure 5: Comparison of kX̄ − Xk, R(t∗ ), and R(t̄) as a function of σ for σ1 = 1.2σ. 1.0 0.8 °X-X°=0.7 Σ¤ 0.6 R Ht* L= Out[134]= 1- 0.4 RHtL= 1- H1+0.7 ΣL2 H1+0.7 ΣL2 +1.51 Σ2 2 H1+0.7 ΣL3 H1+0.7 ΣL2 +1.51 Σ2 + H1+0.7 ΣL4 H1+0.7 ΣL2 +1.51 Σ2 0.2 0.0 -1.0 -0.5 0.0 0.5 1.0 Figure 6: Comparison of kX̄ − Xk, R(t∗ ), and R(t̄) as a function of σ for σ1 = 0.7σ. that subspace, this is not true for Z t in general. Since we do dispose of (an approximation) to Π when A = I, it would be better to compare ∗ kX̄ − Xk, kΠZ t − Xk, kΠZ t̄ − Xk, as functions of σ. In general, we have RΠ (t)2 = kΠZ t − Xk2 = (1 + σ1 )2 t2 − 2t(1 + σ1 ) + 1, and
σ22 − σ1 (σ1 + 1)2
, R (t̄) = σ22 + (σ2 + 1)2 Π RΠ (t∗ ) = σ22 σ22
. + (σ1 + 1)2 The comparison of these quantities is reported in Figure 7 and one can observe that, at ∗ least for σ ≥ 0, both ΠZ t and ΠZ t̄ are better approximations of X than X̄. 23 De Vito, Fornasier and Naumova 1.2 1.0 °X-X°= Σ¤ 0.8 RP Ht* L= Out[199]= 0.6 RP HtL= 0.4 Σ2 IΣ2 +H1+ΣL2 M Σ2 -Σ H1+ΣL2 IΣ2 +H1+ΣL2 M 0.2 0.0 -1.0 0.0 -0.5 0.5 1.0 Figure 7: Comparison of kX̄ − Xk, RΠ (t̄), and R(t∗ ) as a function of σ for σ1 = σ. 4.3 Empirical projections b n . According In concrete applications we do not dispose of Π but only of its empirical proxy Π to Theorem 4 they are related by the approximation  r m 1 τ b √ , kΠ − Πn k ≤ + λmin n n with high probability (depending on τ > 0), where m = 1 in our example and λmin = 1 is the smallest nonzero eigenvalue of ΣAX = A1 AT1 . Hence, we compute and, since A is an orthogonal matrix b = A−1 Π b n Y, X b b n )Y k kX − Xk = kA−1 (Π − Π b n )Y k = k(Π − Π ≤ ≤ Here we used 1+τ √ kA1 + σW k n √ 1+τ 1+τp √ 1 + 2σ 2 + 2σ1 ≤ √ (1 + 2σ). n n kA1 + σW k2 = kA1 k2 + +σ 2 kW k2 + 2σhA1 , W i = 1 + 2σ 2 + 2σ1 . In view of the approximation relationship above, we can express now where p ǫ21 + ǫ22 ≤ 1+τ √ (1 n + √ b = (1 + σ1 + ǫ1 , ǫ2 )T , X 2σ). Then b = kX − Xk q (σ1 + ǫ1 )2 + ǫ22 . 24 A Machine Learning Approach to Optimal Tikhonov Regularization One has also b 2 = (t(1 + σ1 ) − (1 + σ1 + ǫ1 ))2 + (tσ2 − ǫ2 )2 Rn (t)2 = kZ t − Xk = t2 ((1 + σ1 )2 + σ22 ) − 2t(σ2 ǫ2 + (1 + σ1 )ǫ1 + (1 + ǫ1 )2 ) + const. Hence, its optimizer is given by (σ2 ǫ2 + (1 + σ1 )ǫ1 + (1 + ǫ1 )2 ) b tn = (1 + σ1 )2 + σ22 and R(b tn ) = kZ b tn − Xk = s (σ22 (1 + ǫ2 )2 + 2ǫ2 (1 + σ1 )(ǫ1 + σ1 )σ2 + (1 + σ1 )2 (ǫ1 + σ1 )2 ) (1 + σ1 )2 + σ22 4.3.1 Concrete example We compare in Figure 8 the behavior of the difference of the errors b − kZ btn − Xk kX − Xk b tn depending on σ, for σ1 = 1.3σ and n = 100. In this case, the empirical estimator √ Z is a b for all noise levels σ ∈ [−0.07, 1/ 2], while significantly better approximation to X than X √ b X keeps being best estimator, e.g., for σ ∈ [−1/ 2, −0.1]. 0.05 0.00 -0.05 ` ` H°X-Xn °-RHtn LL Out[29]= -0.10 -0.15 -0.6 -0.4 -0.2 0.0 0.2 0.4 0.6 b Figure 8: The difference of the errors kX − Xk− kZ btn − Xk as a function of σ for σ1 = 1.3σ. √ b > kZ btn − Xk. For σ ∈ [−0.07, 1/ 2] one has kX − Xk From these simple two dimensional toy examples and related numerical experiments we deduce the following general principles: ∗ • The performances of the estimators X̄, Z t , Z btn depend very much on how the noise is distributed with respect to the subspace W to which AX belongs and its signature; 25 De Vito, Fornasier and Naumova if the noise is not very much concentrated on such subspace, then X̄ tends to be a better estimator, see Fig. 6, while a concentrated noise would make X̄, Z btn rather ∗ equivalent and Z t the best one for some ranges of noise, see Fig. 4 and Fig. 5. • If the number n of samples is not very large (in the example above we considered b nY = X b ≈ X̄ = A−1 ΠY would be still too n = 100), so that the approximation A−1 Π b for significant ranges of noise, indepenrough, then the estimator Z btn may beat X dently of its instance correlation with W, see Fig. 8. b Z btn perform better as Since it is a priori impossible to know precisely in which situation X, estimators (as it depends on the correlation with W of the particular noice instance), in the practice it will be convenient to compute them both, especially when one does not dispose of a large number n of samples. 4.4 The case A = I As yet one more example, this time in higher dimension, we consider the simple case where m = d and A = I, so that W = V. In this case, we get that R(t) = −(1 − t)X + tη. If Y 6= 0, an easy computation shows that the minimizer of the reconstruction error kR(t)k2 is   hY, Xi ∗ ∗ , (38) t = t (Y, X) = ϕ hY, Y i where   0 if s ≤ 0 ϕ(s) = s if 0 < s < 1 .   1 if s ≥ 1 If Y = 0, the solution Z t does not depend on t, so that there is not a unique optimal parameter and we set t∗ = 0. We further assume that X is bounded from 0 with high probability, more precisely,   1 P[kXk < r] ≤ 2 exp − 2 . (39) r This assumption is necessary to avoid that the noise is much bigger than the signal. 2 Theorem 12 Given τ ≥ 1, with probability greater than 1 − 6e−τ ! r e √ d τ 1 2 ∗ ( h + τ ), + √ + σ + σ ln |b tn − t | ≤ λmin n σ n (40) provided that   √ 64 2 ,1 n & ( d + τ ) max λ2min ) (r λmin 1−16τ 2 . ,e σ < min 8 26 (41a) (41b) A Machine Learning Approach to Optimal Tikhonov Regularization Proof Without loss of generality, we assume that λmin ≤ 8. Furthermore, on the event {Y = 0}, by definition t∗ = b tn = 0, so that we can further assume that Y 6= 0. Since ϕ is a Lipschitz continuous function with Lipschitz constant 1, b − Xi| |hY, X |b tn − t∗ | ≤ kY k2 b n )Y − Πηk k(Π − Π ≤ kY k b n )k + σ kΠW k , ≤ k(Π − Π kY k where the second inequality is consequence of (34). Since (41a) and (41b) imply (27) and m = d, by (26) we get ! r 1 τ d b n − Πk . kΠ + √ + σ2 λmin n n 2 with probability greater than 1 − 2e−τ . It is now convenient to denote the probability distribution of X as ρX , i.e., X ∼ ρX . Fixed r > 0, set Ω = {kXk < r} ∪ {2σhX, W i < −kXk2 /2}, whose probability is bounded by P[Ω] ≤ P[kXk < r] + P[4σhX, W i < −kXk2 , kXk ≥ r] Z P[4σhx, W i < −kxk2 ] dρX (x) = P[kXk < r] + kxk≥r ≤ P[kXk < r] + Z  kxk2 exp − 256σ 2 kxk≥r  r2 ≤ P[kXk < r] + exp − 256σ 2   dρX (x) , where we use (46c) with ξ √= −W (and the fact that W and X are independent), τ = kxk/(16σ) and kW kψ2 = 1/ 2. With the choice r = 16τ / ln(e/σ), we obtain  τ2 P[Ω] ≤ P[kXk < 16τ / ln(e/σ)] + exp − 2 2 σ ln (e/σ) ≤ P[kXk < 16τ / ln(e/σ)] + exp(−τ 2 ),  where σ 7→ σ ln(e/σ) is an increasing positive function on (0, 1], so that it is bounded by 1. Furthermore, by (41b), i.e., 16τ / ln(e/σ) ≤ τ1 , we have P[kXk < 16τ / ln(e/σ)] ≤ P [kXk < 27 1 ] ≤ 2 exp(−τ 2 ), τ De Vito, Fornasier and Naumova 2 by Assumption (39). Hence it holds that P[Ω] ≤ 3e−τ . On the event Ωc kY k2 = kXk2 + 2σhX, W i + σ 2 kW k2 ≥ kXk2 + 2σhX, W i ≥ kXk2 /2 ≥ r 2 /2 ≃ τ 2 / ln2 (e/σ) ≥ 1/ ln2 (e/σ) since τ ≥ 1. Finally, (51) with ξ = ΠW ∈ W yields √ kΠW k . ( h + τ ) with probability greater than 1 − exp(−τ 2 ). Taking into account the above estimates, we conclude with probability greater than 1 − 4 exp(−τ 2 ) that e √ kΠW k . ln ( h + τ ). kY k σ Then, with probability greater than 1 − 6 exp(−τ 2 ), we conclude the estimate ∗ |b tn − t | . 1 λmin r d τ + √ + σ2 n n ! √ + σ ln(e/σ)( h + τ ). Remark 13 The function ln(e/σ) can be replaced by any positive function f (σ) such that σf (σ) is an infinitesimal function bounded by 1 in the interval (0, 1]. The condition (41b) becomes (r ) λmin σ < min ,1 f (σ) ≥ 16τ 2 , 8 and, if f is strictly decreasing, σ < min (r ) λmin , 1, f −1 (16τ 2 ) . 8 Theorem 12 shows that if the number n of examples is large enough and the noise level is small enough, the estimator b tn is a good approximation of the optimal value t∗ . Let us stress very much that the number n of samples needed to achieve a good accuracy depends at most algebraically on the dimension d, more precisely n = O(d). Hence, in this case one does not incur in the curse of dimensionality. Moreover, the second term of the error estimate (40) gets smaller for smaller dimensionality h. Remark 14 If there exists an orthonormal basis (ei )i of Rd , such that the random variables hW, e1 i, . . . , hW, ed i are independent with E[hW, e1 i2 ] = 1, then Rudelson and Vershynin 28 A Machine Learning Approach to Optimal Tikhonov Regularization √ (2013, Theorem 2.1) showed that kW k concentrates around d with high probability. Reasoning as in the proof of Theorem 12, by replacing kXk2 with σ 2 kW k2 , with high probability it holds that ! r 1 τ d 1 √ τ |b tn − t∗ | . + √ + σ2 + √ ( h + τ ) √ λmin n n d d without assuming condition (39). In Figures 9–12, we show examples of numerical accordance between optimal and estimated regularization parameters. In this case, the agreement between optimal parameter t∗ and learned parameter b tn is overwhelming. 1.0 0.8 0.6 0.4 0.2 10 20 30 40 50 Figure 9: Optimal parameters t∗ and corresponding approximations b tn (n = 1000) for 50 different data Y = X + η for X and η generated randomly with Gaussian distributions in Rd for d = 1000. We assume that X ∈ V for V = span{e1 , . . . , e5 }. 2.5 2.5 2.0 2.0 1.5 1.5 1.0 1.0 0.5 0.5 0.2 0.4 0.6 0.8 1.0 1.2 0.2 0.4 0.6 0.8 1.0 1.2 Figure 10: Empirical distribution of the optimal parameters t∗ (left) and the corresponding empirical distribution of the learned parameters b tn (right) for 500 randomly generated data Y = X + η with the same noise level. 29 De Vito, Fornasier and Naumova 5. An explicit formula by linearization b n (t) = 0 by analytic methods for d > 2 While it is not possible to solve the equation H(t) ≈ H in the general case, one might attempt a linearization of this equation in certain regimes. It is well-known that the optimal Tikhonov regularization parameter α∗ = (1−t∗ )/t∗ converges to 0 for vanishing noise level and this means that t∗ = t∗ (σ) → 1 as σ → 0. Hence, if the matrix A has a significant spectral gap, i.e., σ1 ≥ σ2 ≥ · · · ≥ σd ≫ 0 and σ ≈ 0 is small enough, then σi ≫ (1 − t∗ ), (42) and in this case (σi νbi ξbi−1 + 1)t − 1 (σi νbi ξbi−1 + 1)t − 1 b ≈ , hi (t) = ((1 − t) + tσi2 )3 σi6 t ≈ t∗ . The above linear approximation is equivalent to replacing B(t)−1 with B(1)−1 = (AT A)−1 and Equation (23) is replaced by the following proxy (at least if t ≈ t∗ ) b lin (t) = htAAT (Y − Π b n Y ) − (1 − t)Π b n Y , (AAT )†3 Y i H n b + tAT ηb), AX b + ηb)i = hA(AT A)−3 (−(1 − t)X   b + AT ηb), AX b + ηbi t − hA(AT A)−3 X, b AX b + ηbi = hA(AT A)−3 (X ! d d X X α bi α bi −1 b (σ ν b ξ + 1) t − = i i i 5 5. σ σ i i i=1 i=1 b lin (t) is The only zero of H n b n Y , (AAT )†3 Y i hΠ b nY ) + Π b n Y , (AAT )†3 Y i hAAT (Y − Π b n Y , (AAT )†2 Y i hY − Π =1− b nY ) + Π b n Y , (AAT )†3 Y i hAAT (Y − Π b AX b + ηbi hA(AT A)−3 X, = b + AT ηb), AX b + ηbi hA(AT A)−3 (X b + ηbi h(AAT )†2 ηb, AX =1− b + AT ηb), AX b + ηbi hA(AT A)−3 (X Pd −5 bi i=1 σi α = Pd −5 bi (σi νbi ξbi−1 + 1) i=1 σi α Pd −4 bi νbi i=1 σi α = 1 − Pd . −5 bi (σi νbi ξbi−1 + 1) i=1 σi α b tlin n = In Figure 11, we present the comparison between optimal parameters t∗ and their approxi∗ mations b tlin n . Despite the fact that the gap between σd and 1 − t is not as large as requested in (42), the agreement between t∗ and b tlin n keeps rather satisfactory. In Figure 12, we report the empirical distributions of the parameters, showing essentially their agreement. 30 A Machine Learning Approach to Optimal Tikhonov Regularization 1.0 0.8 0.6 0.4 0.2 10 20 30 40 50 Figure 11: Optimal parameters t∗ and corresponding approximations b tn (n = 1000) for 50 different data Y = AX +η for X and η generated as for the experiment of Figure 2. Here we considered a noise level σ = 0.006, so that the optimal parameter t∗ can be very close to 0.5 and the minimal singular value of A is σd ≈ 0.7. 3.0 2.5 2.5 2.0 2.0 1.5 1.5 1.0 1.0 0.5 0.5 0.2 0.4 0.6 0.8 1.0 1.2 0.2 0.4 0.6 0.8 1.0 1.2 Figure 12: Empirical distribution of the optimal parameters t∗ (left) and the corresponding empirical distribution of the approximating parameters b tlin n (right) for 1000 randomly generated data Y = AX + η with the same higher noise level. The statistical accordance of the two parameters t∗ and b tlin n is shown. 6. Conclusions and a glimps to future directions Motivated by the challenge of the regularization and parameter choice/learning relevant for many inverse problems in real-life, in this paper we presented a method to determine the parameter, based on the usage of a supervised machine learning framework. Under the assumption that the solution of the inverse problem is distributed sub-gaussianly over a small dimensional linear subspace V and the noise is also sub-gaussian, we provided a rigorous theoretical justification for the learning procedure of the function, which maps given noisy data into an optimal Tikhonov regularization parameter. We also presented and discussed explicit bounds for special cases and provided techniques for the practical implementation 31 De Vito, Fornasier and Naumova of the method. Classical empirical computations of Tikhonov regularization parameters, for instance the discrepancy principle, see, e.g., (Bauer and Lukas, 2011; Engl et al., 1996), may be assuming the knowledge of the noise level σ and use regularized Tikhonov solutions for different choices of t Z t = t(tAT A + (1 − t)I)−1 AT Y, to return some parameter b tσ which ideally provides σ-optimal asymptotic behavior of Z btσ for noise level σ → 0. To be a bit more concrete, by using either bisection-type methods or by fixing a pretermined grid of points T = {t(j) : j = 1, 2, . . . }, one searches some t(j̄) ∈ T (j̄) tσ := t(j̄) . For each t = t(j) the computation of Z t has for which kAZ t − Y k ≈ σ and set b cost of order O(d2 ) as soon as one has precomputed the SVD of A. Most of the empirical estimators require then the computation of Nit instances of Z t for a total of O(Nit d2 ) operations to return b tσ . Our approach does not assume knowledge of σ and it fits into the class of heuristic parameter choice rules, see (Kindermann, 2011); it requires first to have precomputed the b n by computation of one single h-truncated SVD2 of the empirical empirical projection Π b n of dimension m × m, where m is assumed to be significantly smaller covariance matrix Σ than d. In fact, the complexity of methods for computing a truncated SVD out of standard books is O(hm2 ). When one uses randomized algorithms one could reduce the complexity even to O(log(h)m2 ), see (Halko et al., 2011). Then one needs to solve the optimization problem. b 2. min kZ t − Xk (43) t∈[0,1] b n (t), This optimization is equivalent to finding numerically roots in (0, 1) of the function H which can be performed in the general case and for d > 2 only by iterative algorithms. They also require sequential evaluations of Z t of individual cost O(d2 ), but they converge usually ∗ very fast and they are guaranteed by our results to return Z bt ≈ Z t . If we denote Nit◦ the number of iterations needed for the root finding with appropriate accuracy, we obtain a total complexity of O(Nit◦ d2 ). In the general case, our approach is going to be more advantageous with respect to classical methods if Nit◦ ≪ Nit . However, in the special case where the noise level is relatively small compared to the minimal positive singular value of A, e.g., σd ≫ σ, then our arguments in Section 5 suggest that one can approximate the b n (t) and hence t∗ with the smaller cost of O(d2 ) and no need of several relevant root of H evaluations of Z t . As a byproduct of our results we also showed that b = A† Π b nY X (44) b n and is also a good estimator and this one requires actually only the computation of Π then the execution of the operations in formula (44) of cost O(md) (if we assumed the precomputation of the SVD of A). 2. One needs to compute h singular values and singular vectors up to the first significant spectral gap. 32 A Machine Learning Approach to Optimal Tikhonov Regularization Our current efforts are devoted to the extension of the analysis to the case where the underlying space V is actually a smooth lower-dimensional nonlinear manifold. This extension will be realized by firstly approximating the nonlinear manifold locally on a proper decomposition by means of affine spaces as proposed in (Chen et al., 2013) and then applying our presented results on those local linear approximations. Another interesting future direction consists of extending the approach to sets V, unions of linear subspaces as in the case of solutions expressible sparsely the respect to certain dictionaries. In this situation, one would need to consider different regularization techniques and possibly non-convex non-smooth penalty quasi-norms. For the sake of providing a first glimps on the feasibility of the latter possible extension, we consider below the problem of image denoising. In particular, as a simple example of the image denoising algorithm, we consider the wavelet shrinkage (Donoho and Johnstone, 1994): given a noisy image Y = X + σW (already expressed in wavelet coordinates), where σ is the level of Gaussian noise, the denoised image is obtained by Z α = Sα (X) = arg min kZ − Y k2 + 2αkZkℓ1 , Z where k · kℓ1 denotes the ℓ1 norm, which promotes a sparse representation of the image with respect to a wavelet decomposition. Here, as earlier, we are interested in learning the highdimensional function mapping noisy images X into their optimal shrinkage parameters, i.e., an optimal solution of kZ α − Xk2 → minα . Employing a properly modified version of the procedure described in this paper, we are obtaining very exciting and promising results, in particular that the optimal shrinkage parameter α essentially depends nonlinearly on very few (actually 1 or 2) linear evaluations of Y . This is not a new observation and it is a data-driven verification of the well-known results of (Donoho and Johnstone, 1994) and (Chambolle et al., 1998), establishing that the optimal parameter depends essentially on two meta-features of the noisy image, i.e., the noise level and its Besov regularity. In Figure 13 and Figure 14 we present the numerical results for wavelet shrinkage, which show that our approach chooses a nearly optimal parameter in terms of peak signal-to-noise ratio (PSNR) and visual quality of the denoising. Acknowledgments E. De Vito is a member of the Gruppo Nazionale per l’Analisi Matematica, la Probabilità e le loro Applicazioni (GNAMPA) of the Istituto Nazionale di Alta Matematica (INdAM). M. Fornasier acknowledges the financial support of the ERC-Starting Grant HDSPCONTR “High-Dimensional Sparse Optimal Control”. V. Naumova acknowledges the support of project “Function-driven Data Learning in High Dimension” (FunDaHD) funded by the Research Council of Norway. Appendix A. Perturbation result for compact operators We recall the following perturbation result for compact operators in Hilbert spaces (Anselone, 1971) and (Zwald and Blanchard, 2006, Theorem 3), whose proof also holds without the assumption that the spectrum is simple (see Rosasco et al., 2010, Theorem 20). 33 De Vito, Fornasier and Naumova Ground truth Noisy image 20 20 40 40 60 60 80 80 100 100 120 120 20 40 60 80 100 120 20 40 60 80 100 120 Reconstructed image with learned parameter Reconstructed image with optimal parameter 20 20 40 40 60 60 80 80 100 100 120 120 20 40 60 80 100 120 20 40 60 80 100 120 Figure 13: Numerical Experiments for Wavelet Shrinkage. Proposition 15 Let A and B be two compact positive operators on a Hilbert space H L and denote by (αj )N j=1 and (βℓ )ℓ=1 the corresponding families of (distinct) strictly positive eigenvalues of A and B ordered in a decreasing way. For all 1 ≤ j ≤ N , denote by Pj (resp. Qℓ with 1 ≤ ℓ ≤ L) the projection onto the vector space spanned by the eigenvectors of A (resp. B) whose eigenvalues are greater or equal than αj (respect. βℓ ). Let j ≤ N such that α −α kA − Bk < j 4 j+1 , then there exists ℓ ≤ L so that αj + αj+1 < βℓ 2 2 kQℓ − Pj k ≤ kA − Bk αj − αj+1 βℓ+1 < (45) dim Qℓ H = dim Pj H. If A and B are Hilbert-Schmidt, the operator norm in the above bound can be replaced by the Hilbert-Schmidt norm. 34 A Machine Learning Approach to Optimal Tikhonov Regularization 45 PSNR for denoised image with learned PSNR for denoised image with optimal PSNR for noisy image 40 35 30 25 20 15 0 10 20 30 40 50 60 Figure 14: PSNR between the ground truth image and its noisy version (green line); the ground truth and its denoised version with the optimal parameter (blue line) and the learned parameter (red line). The results are presented for 60 random images. In the above proposition, if N or M are finite, αN +1 = 0 or βL+1 = 0. We note that, if the eigenvalues of A or B are not simple, in general ℓ 6= j. However, the above result shows that, given j = 1, . . . , N there exists a unique ℓ = 1, . . . , L such that dim Qℓ H = dim Pj H and βℓ is the smallest eigenvalue of B greater than (αj + αj+1 )/2. Appendix B. Sub-gaussian vectors We recall some facts about sub-gaussian random vectors and we follow the presentation in (Vershynin, 2012), which provides the proofs of the main results in Lemma 5.5. Proposition 16 Let ξ be a sub-gaussian random vector in Rd . Then, for all τ > 0 and v ∈ Rd P[|hξ, vi| > 3kξkψ2 kvkτ ] ≤ 2 exp(−τ 2 ). (46a) Under the further assumption that ξ is centered, then E[exp(τ hξ, vi)] ≤ exp 8τ 2 kξk2ψ2 √ P[hξ, vi > 4 2kξkψ2 kvkτ ] ≤ exp(−τ 2 ).
(46b) (46c) Proof We follow the idea in Vershynin (2012, Lemma 5.5) of explicitly computing the constants. By rescaling ξ to ξ/kξkψ2 , we can assume that kξkψ2 = 1. 35 De Vito, Fornasier and Naumova Let c > 0 be a small constant to be fixed. Given, v ∈ S d−1 , set χ = hξ, vi, which is a real sub-gaussian vector. By Markov inequality, P[|χ| > τ ] = P[c|χ|2 > cτ 2 ] = P[exp(c|χ|2 ) > exp(cτ 2 )] 2 ≤ E[exp(c|χ|2 )]e−cτ . By (5), we get that E[|χ|q ] ≤ q q/2 , so that E[exp(c|χ|2 )] = 1 + +∞ k X c k=1 k! E[|χ|2k ] ≤ 1 + +∞ X (2ck)k k=1 k! +∞ ≤1+ 1X 2c (2ce)k = 1 + , e 1 − 2ce k=1 where we use the estimate k! ≥ e(k/e)k for k ≥ 1. Setting c = 1/9, 1 + 2c 1−2ce < 2, so that P[|χ| > 3τ ] ≤ 2 exp(−τ 2 ), so that (46a) is proven. Assume now that E[ξ] = 0. By (5.8) in Vershynin (2012) +∞  X τ E[exp( χ)] ≤ 1 + e k=2 |τ | √ k k , (47) and, by (5.9) in Vershynin (2012), 2 exp τ C 2
 +∞  X C|τ | 2h √ . ≥1+ h h=1 (48) If |τ | ≤ 1, fix an even k ≥ 2 so that k = 2h with h ≥ 1. The k-th and (k + 1)-th terms of the series (47) is       k+1    |τ | k |τ | C|τ | k |τ | k |τ | 2h 2 √ + √ ≤ √ ≤2 √ = 2h h √ , C 2 k+1 k k 2h k where the right hand side is the h-th term of the series (48) and the last inequality holds true if C ≥ 1. Under this assumption
τ |τ | ≤ 1. (49) E[exp( χ)] ≤ exp τ 2 C 2 e If |τ | ≥ 1, fix an odd k ≥ 3, so that k = 2h − 1 with h ≥ 2, then the k-th and (k + 1)-th terms of the series (47) is !h  √  k+1        |τ | k+1 |τ | 2h |τ | 2h |τ | 2h |τ | k √ √ ≤2 √ + √ = ≤ √ , C 2 (2h − 1) k+1 k k h h where the right hand side is the h-th term of the series (48) and the inequality holds true provided that √ √ 2 2 2h 2 = , C ≥ sup 3 h≥2 (2h − 1) 36 A Machine Learning Approach to Optimal Tikhonov Regularization which holds true with the choice C = 1. Furthermore, the term of (47) with k = 2 is clearly bounded by term of (48) with h = 1, so that we get
τ E[exp( χ)] ≤ exp τ 2 e |τ | ≥ 1. Together with bound (49) with C = 1, the abound bound gives (46b) since e2 < 8. Finally, reasoning as in the proof of (46a) and by using (46b) P[χ > τ ] = P [exp(cχ) > exp(cτ )] ≤ E[exp(cχ)]e−cτ ≤ exp(8c2 − cτ ), which takes the minimum at c = τ /16. Hence, P[χ > τ ] = exp(− τ2 ). 32 Remark 17 Both (46a) and (46b) (for a suitable constants instead of kξkψ2 ) are sufficient conditions for sub-gaussianity and (46b) implies that E[ξ] = 0, see (Vershynin, 2012, Lemma 5.5). The following proposition bounds the Euclidean norm of a sub-gaussian vector. The proof is standard and essentially based on the results in Vershynin (2012), but we were not able to find the precise reference. The centered case is done in Rigolet (2015). Proposition 18 Let ξ a sub-gaussian random vector in Rd . Given τ > 0 with probability 2 greater than 1 − 2e−τ √ √ (50) kξk ≤ 3kξkψ2 ( 7d + 2τ ) ≤ 9kξkψ2 ( d + τ ). 2 If E[ξ] = 0, then with probability greater than 1 − e−τ r √ √ 7 d + 2τ ) ≤ 16kξkψ2 ( d + τ ). kξk ≤ 8kξkψ2 ( 2 (51) Proof As usual we assume that kξkψ2 = 1. Let N be a 1/2-net of S d−1 . Lemmas 5.2 and 5.3 in Vershynin (2012) give kξk ≤ 2 maxhξ, vi v∈N |N | ≤ 5d . Fixed v ∈ N , (46a) gives that P[|hξ, vi| > 3t] ≤ 2 exp(−t2 ). 37 De Vito, Fornasier and Naumova By union bound P[kξk > 6t] ≤ P[max|hξ, vi| > 3t] ≤ 2|N | exp(−t2 ) ≤ 2 exp(d ln 5 − t2 ). v∈N p Bound (50) follows with the choice t = τ + 7d/4 taking into account that t2 − d ln 5 > τ 2 since ln 5 < 7/4. Assume that ξ is centered and use (46c) instead of (46a). Then √ √ P[kξk > 8 2t] ≤ P[max|hξ, vi| > 4 2t] ≤ |N | exp(−t2 ) ≤ exp(d ln 5 − t2 ). v∈N As above, the choice t = τ + p 7d/4 provides the bound (51). Remark 19 Compare with Theorem 1.19 in Rigolet (2015), noting that by (46b) the parameter σ in Definition 1.2 of Rigolet (2015) is bounded by 4kξkψ2 . The following result is a concentration inequality for the second momentum of subgaussian random vector, see Theorem 5.39 and Remark 5.40 in Vershynin (2012) and footnote 20. Theorem 20 Let ξ ∈ Rd be a sub-gaussian vector random vector in Rd . Given a family ξ1 , . . . , ξn of random vectors independent and identically distributed as ξ, then for τ > 0 n P[k 1X 2 ξi ⊗ ξi − E[ξ ⊗ ξ]k > max{δ, δ2 }] ≤ 2e−τ n i=1 where δ = Cξ r τ d +√ n n ! , and Cξ is a constant depending only on the sub-gaussian norm kξkψ2 . References William K. Allard, Guangliang Chen, and Mauro Maggioni. Multi-scale geometric methods for data sets. II: Geometric multi-resolution analysis. Appl. Comput. Harmon. Anal., 32 (3):435–462, 2012. ISSN 1063-5203. doi: 10.1016/j.acha.2011.08.001. Philip M. Anselone. Collectively compact operator approximation theory and applications to integral equations. Prentice-Hall Inc., Englewood Cliffs, N. J., 1971. Frank Bauer and Mark A. Lukas. Comparing parameter choice methods for regularization of ill-posed problems. Math. Comput. Simul., 81(9):1795–1841, 2011. Andrea Braides. Gamma-convergence for beginners. Lecture notes. http://www.mat.uniroma2.it/ braides/0001/dotting.html, 2001. 38 Available on line A Machine Learning Approach to Optimal Tikhonov Regularization Antonin Chambolle, Ronald DeVore, Nam-yong Lee, and Bradley Lucier. Nonlinear wavelet image processing: variational problems, compression, and noise removal through wavelet shrinkage. IEEE Trans. Image Process., 7(3):319–335, 1998. Guangliang Chen, Anna V. Little, and Mauro Maggioni. Multi-resolution geometric analysis for data in high dimensions. In Excursions in harmonic analysis. Volume 1, pages 259– 285. New York, NY: Birkhäuser/Springer, 2013. David Donoho and Iain Johnstone. Biometrika, 81:425–55, 1994. Ideal spatial adaptation by wavelet shrinkage. Heinz W. Engl, Martin Hanke, and Andreas Neubauer. Regularization of inverse problems. Dordrecht: Kluwer Academic Publishers, 1996. N. Halko, P.G. Martinsson, and J.A. Tropp. Finding structure with randomness: probabilistic algorithms for constructing approximate matrix decompositions. SIAM Rev., 53 (2):217–288, 2011. ISSN 0036-1445; 1095-7200/e. doi: 10.1137/090771806. Stefan Kindermann. Convergence analysis of minimization-based noise level-free parameter choice rules for linear ill-posed problems. ETNA, Electron. Trans. Numer. Anal., 38: 233–257, 2011. ISSN 1068-9613/e. Gitta Kutyniok and Demetrio Labate, editors. Shearlets. Multiscale analysis for multivariate data. Boston, MA: Birkhäuser, 2012. Anna V. Little, Mauro Maggioni, and Lorenzo Rosasco. Multiscale geometric methods for data sets. I: Multiscale SVD, noise and curvature. Appl. Comput. Harmon. Anal., 43(3): 504–567, 2017. ISSN 1063-5203. doi: 10.1016/j.acha.2015.09.009. Stéphane Mallat. A wavelet tour of signal processing. The sparse way. 3rd ed. Amsterdam: Elsevier/Academic Press, 3rd ed. edition, 2009. ISBN 978-0-12-374370-1/hbk. Erich Novak and Henryk Woźniakowski. Optimal order of convergence and (in)tractability of multivariate approximation of smooth functions. Constr. Approx., 30(3):457–473, 2009. Philippe Rigolet. 18.S997: High dimensional statistics. Lecture notes. Available on line http://www-math.mit.edu/ rigollet/PDFs/RigNotes15.pdf, 2015. Lorenzo Rosasco, Mikhail Belkin, and Ernesto De Vito. On learning with integral operators. J. Mach. Learn. Res., 11:905–934, 2010. Mark Rudelson and Roman Vershynin. Hanson-Wright inequality and sub-Gaussian concentration. Electron. Commun. Probab., 18(82):1–9, 2013. Roman Vershynin. Introduction to the non-asymptotic analysis of random matrices. In Compressed sensing, pages 210–268. Cambridge Univ. Press, Cambridge, 2012. Laurent Zwald and Gilles Blanchard. On the convergence of eigenspaces in kernel principal component analysis. In Y. Weiss, B. Schölkopf, and J. Platt, editors, Advances in Neural Information Processing Systems 18, pages 1649–1656. MIT Press, Cambridge, MA, 2006. 39 

Accepted as a workshop contribution at ICLR 2015 P URINE : A BI - GRAPH BASED DEEP LEARNING FRAME - arXiv:1412.6249v5 [cs.NE] 16 Apr 2015 WORK Min Lin1,2 , Shuo Li3 , Xuan Luo3 & Shuicheng Yan2 1. Graduate School of integrated Sciences and Engineering 2. Department of Electrical and Computer Engineering National University of Singapore 3. Zhiyuan College, Shanghai Jiao Tong University {linmin, eleyans}@nus.edu.sg li3shuo1@gmail.com roxanneluo@sjtu.edu.cn A BSTRACT In this paper, we introduce a novel deep learning framework, termed Purine. In Purine, a deep network is expressed as a bipartite graph (bi-graph), which is composed of interconnected operators and data tensors. With the bi-graph abstraction, networks are easily solvable with event-driven task dispatcher. We then demonstrate that different parallelism schemes over GPUs and/or CPUs on single or multiple PCs can be universally implemented by graph composition. This eases researchers from coding for various parallelization schemes, and the same dispatcher can be used for solving variant graphs. Scheduled by the task dispatcher, memory transfers are fully overlapped with other computations, which greatly reduces the communication overhead and helps us achieve approximate linear acceleration. 1 I NTRODUCTION The need for training deep neural networks on large-scale datasets has motivated serveral research works that aim to accelerate the training process by parallelising the training on multiple CPUs or GPUs. There are two different ways to parallelize the training. (1) Model parallelism: the model is distributed to different computing nodes (Sutskever et al., 2014) (2) Data parallelism: different nodes train on different samples for the same model (Seide et al., 2014; Chilimbi et al., 2014). Some of the works even use a hybrid of them (Krizhevsky, 2014; Dean et al., 2012; Le, 2013). For data parallelism, there are also two schemes regarding communication between the peers. (1) the allreduce approach where all updates from the peers are aggregated at the synchronization point and the averaged update is broadcasted back to the peers (Seide et al., 2014; Krizhevsky, 2014). (2) the parameter server approach handles the reads and writes of the parameters asynchronously (Dean et al., 2012; Le, 2013; Chilimbi et al., 2014). Efficient implementations of the various parallelization schemes described by previous works are non-trivial. To facilitate the implementation of various parallelization schemes, we built a bigraph-based deep learning framework called “Purine”. It is named “Purine”, which is an analog of caffeine in molecular structure, because we benefited a lot from the open source Caffe framework (Jia et al., 2014) in our research and the math functions used in Purine are ported from Caffe. 2 B I -G RAPH A BSTRACTION Purine abstracts the processing procedure of deep neural networks into directed bipartite graphs (Bi-Graphs). The Bi-Graph contains two types of vertices, tensors and operators. Directed edges are only between tensors and operators and there is no interconnection within tensors or operators. Figure 1 illustrates the Bi-Graph for the convolution layer defined in Caffe. All feed-forward neural nets can be represented by a directed acyclic bipartite graph, which can be solved by a universal task dispatcher. There are several works that use similar abstractions. For 1 Accepted as a workshop contribution at ICLR 2015 Top Top ΔTop Add bias ΔTop Bias gradient ΔBias Bias Convolution Layer Conv w.r.t bottom Conv Bottom ΔBottom Conv w.r.t weight ΔWeight Weight Bottom ΔBottom (b) Bipartite Graph (a) Caffe Convolution Layer Figure 1: (a) shows the convolution layer defined in Caffe together with its inputs and outputs. (b) is the corresponding bipartite graph that describes the underlying computation inside the convolution layer. There are two types of vertices in the Bi-Graph. Boxes represent data tensors and the circles represent operators. Operators are functions of the incoming tensors and the results of the functions are put in the outgoing tensors. example, the dataflow graph in Dryad (Isard et al., 2007) and Pig Latin (Olston et al., 2008) are the same as the Bi-Graph abstraction introduced in this paper. Graphlab (Low et al., 2010) proposed a more general abstraction which is applicable to iterative algorithms. However, these systems are designed for general problems and do not support GPU. Theano (Bergstra et al., 2010) compiles math expressions and their symbolic differentiations into graphs for evalutation. Though it supports GPU and is widely used for deep learning, the ability to parallelize over multiple GPUs and over GPUs on different machines is not complete. 2.1 TASK D ISPATCHER Purine solves the Bi-Graph by scheduling the operators within the Bi-Graph with an event-driven task dispatcher. Execution of an operator is triggered when all the incoming tensors are ready. A tensor is ready when all its incoming operators have completed computation. The computation of the Bi-Graph starts from the sources of the graph and stops when all the sinks are reached. This process is scheduled by the task dispatcher. 2.2 I TERATIONS Although it has been argued in (Low et al., 2010) that the directed acyclic graph could not effectively express iterative algorithms as the graph structure would depend on the number of iterations. We overcome this by iteration of the graphs. Because the task dispatcher waits until all the sinks of the graph are reached, it acts as a synchronization point. Thus parallelizable operations can be put in a single graph, while sequential tasks (iterations) are implemented by iteration of graphs. A concrete example is shown in Figure 2. 3 PARALLELIZATION Parallelization of the Bi-Graph on a cluster of CPUs or GPUs or mixed can be easily implemented by introducing a “location” property for the tensors and operators. The “location” property uniquely identifies the computation resource (CPUs/GPUs) on which a tensor/operator should be allocated. The “location” property comprises two fields: hostname and device id. In a multi-machine cluster, hostname identifies the machine that the vertice resides on. Device id specifies whether the tensor/operator should be allocated on CPU or GPU and the ordinal of the GPU if there are multiple GPUs installed on a single machine. Besides the “location” property, another property “thread” is assigned to operators because both CPU and GPU support multithreading. Operators with the same 2 Accepted as a workshop contribution at ICLR 2015 (a) params (b) params new params DNN new params Swap (c) DNN, Swap, DNN, Swap, ... ... Figure 2: Implementation of SGD by Graph iteration. Every iteration of SGD calculates a modification of the network parameter, and updates the parameter before the next iteration. Since direct updating the network parameter would form a cyclic loop in the graph, it is dissected into two parts. (a) The DNN graph calculates the updated parameters and places them in “new params”, (b) The swap graph will swap the memory address of the “new” and “old” parameters. As a whole, SGD is implemented by iterating the two graphs as in (c). thread id will be queued in the same thread, while those with different ids are parallelized whenever possible. It is up to the user to decide the assignment of the graph over the computation resources. 3.1 C OPY O PERATOR In the multidevice setting, data located on one device are not directly accessible by operators on another. Thus a special “Copy” operator is introduced to cross the boundary, connecting parts of the Bi-Graph on individual devices. The Copy operators, just like other operators, are scheduled by the task dispatcher. Therefore it is straightforward to overlap copy operations with other computation tasks by assigning different threads to them. 3.2 TASK D ISPATCHER In the case of single machine and multiple devices, only one dispatcher process is launched. Operators are associated to their threads and scheduled by the global task dispatcher. In the case of multiple machines and multiple devices, individual dispatcher processes are launched on each of the machines. Copy operators that copy data from machine A to machine B are sinks on machine A and sources on machine B. This way, each machine only needs to schedule its own subgraph and no global scheduling mechanism or communication between dispatchers is necessary. 3.3 M ODEL PARALLELISM We demonstrate how model parallelism can be implemented in Purine by taking a two-layer fully connected neural network as example. It can be extended to deeper networks easily. As is shown in Figure 3, execution of the two-layer network can be divided into three sequential steps. They are labeled as A, B, C correspondingly. To keep resources busy all the time, the network is replicated three times and executed in order. 3.4 DATA PARALLELISM Data parallelism is a simple yet straightforward way to parallelize deep networks. In data parallelism, computation peers each keep a replicate of the deep network. The communication between peers can be either synchronous or asynchronous. In the synchonous case, the gradients from peers are gathered by the parameter server. The updated parameter is calculated and copied back to all the peers. A hybrid of data parallelism and model parallelism has previously been proposed by Krizhevsky (2014) in which the convolution layers use data parallelism and fully connected layers use model parallelism. This is based on the observation that the number of parameters is large and thus the communication cost is big for fully connected layers. The hybrid approach greatly reduces the communication cost. A different approach to reduce communication overhead is to overlap the data 3 Accepted as a workshop contribution at ICLR 2015 (a) B Loss Device 1 Label Softmax Layer Device 2 Softmax Copy Op Softmax Diff Copied Blob on another location A C (b) A1 A2 A3 A1 A2 A3 ...... B1 B2 B3 B1 B2 B3 ...... C1 C2 C3 C1 C2 C3 ...... Inner Product Layer Iterations Figure 3: Implementing model parallelism in Purine. (a) The two-layer network can be divided into three subgraphs which execute in sequence. (b) The network is replicated three times and executed in order. transfer with computations. Double buffering is proposed by Seide et al. (2014) to break a minibatch in half and exchange the gradients of the first half while doing computaion of the second half. With the scheduling of the task dispatcher in Purine, we propose a more straightforward way to hide the communication overhead. We show that data parallelism is feasible even for fully connected layers, especially when the network is very deep. Since the fully connected layers are usually at the top of the neural networks, exchange of the parameter gradients can be overlapped with the backward computation of the convolution layers. As is shown in Figure 4, exchange of gradients in the higher layer can be overlapped with the computation of lower layers. Gradient exchange of lower layers could be less of a problem because they usually have a much smaller number of parameters. Weights Gradients Gradient exchange Forward Backward Green arrows can overlap in time Figure 4: Overlapping communication with computation. 4 Accepted as a workshop contribution at ICLR 2015 4 R ESULTS We carried out experiments on the Purine framework with data parallelism on GoogLeNet (Szegedy et al., 2014). Data parallelism often results in larger batch sizes, which are unfavorable for SGD convergence demonstrated by previous studies. In this paper we ignored the possible change in convergence rate but instead studied how much more data can be processed per unit time with the parallelization. We compared the number of images processed per second for GoogLeNet with different numbers of GPUs for data parallelism. The batch size we use is 128 per GPU. There are 3 GPUs installed on each workstation, interconnected with 10 Gigabit Ethernet. As is shown in Table 1, the speed increases linearly with more GPUs added. The speed is faster than the previous version of this paper because we upgraded the CUDNN library to version 2, which is faster compared to version 1. Note that the machines are connected by 10 gigabit ethernet and thus data on GPU need to go through CPU memory to be tranferred over the ethernet. Even with this limitation, the speed up is linear thanks to the overlapping of communication with computation. Table 1: Number of images per second with increasing number of GPUs. (GPU number smaller than or equal to 3 are tested on single machine. Performances of GPU number larger than 3 are on different machines interconnected with 10 Gigabit Ethernet.) Number of GPUs Images per second 1 112.2 2 222.6 3 336.8 6 673.7 9 1010.5 12 1383.7 Running GoogLeNet with 4 GPUs on a single machine is profiled and shown in Figure 5. It can be seen that the memory copy of model parameters between CPUs and GPUs is fully overlapped with the computations in the backward pass. The only overhead is in the first layer of the network, which results in the gap between iterations. It is favorable to have small batch size in stochastic gradient descent. However, regarding parallelism, it is more favorable to have larger batch size and thus higher computation to communication ratio. We searched for the minumum batch size possible to achieve linear speed up by exploring different batch sizes. Memcpy CPU to GPU Memcpy GPU to CPU CUDA Kernel Call Forward Start of graph Backward End of graph Figure 5: Profiling results of Purine. Memory copies (row 1 and 2) are overlapped with computation (row 3). The only overhead is the memory copy of first convolution layer, which results in the gap between iterations. Table 2: Number of images per second with 12 GPUs and different batch sizes. Batch size per GPU Images per second Acceleration Ratio 128 1383.7 12 64 1299.1 11.26 56 1292.3 11.20 48 1279.7 11.09 40 1230.2 10.66 32 1099.8 9.53 Table 2 shows the processing speed with different batch sizes. The acceleration ratio is not reduced much with batch size 64 as compared to 128. We can still achieve 9.53 fold acceleration with 12 GPUs when the batch size is set to 32. 5 Accepted as a workshop contribution at ICLR 2015 R EFERENCES James Bergstra, Olivier Breuleux, Frédéric Bastien, Pascal Lamblin, Razvan Pascanu, Guillaume Desjardins, Joseph Turian, David Warde-Farley, and Yoshua Bengio. Theano: a CPU and GPU math expression compiler. In Proceedings of the Python for Scientific Computing Conference (SciPy), June 2010. Oral Presentation. Trishul Chilimbi, Yutaka Suzue, Johnson Apacible, and Karthik Kalyanaraman. Project adam: Building an efficient and scalable deep learning training system. In Proceedings of the 11th USENIX conference on Operating Systems Design and Implementation, pages 571–582. USENIX Association, 2014. Jeffrey Dean, Greg Corrado, Rajat Monga, Kai Chen, Matthieu Devin, Mark Mao, Andrew Senior, Paul Tucker, Ke Yang, Quoc V Le, et al. Large scale distributed deep networks. In Advances in Neural Information Processing Systems, pages 1223–1231, 2012. Michael Isard, Mihai Budiu, Yuan Yu, Andrew Birrell, and Dennis Fetterly. Dryad: distributed dataparallel programs from sequential building blocks. In ACM SIGOPS Operating Systems Review, volume 41, pages 59–72. ACM, 2007. Yangqing Jia, Evan Shelhamer, Jeff Donahue, Sergey Karayev, Jonathan Long, Ross Girshick, Sergio Guadarrama, and Trevor Darrell. Caffe: Convolutional architecture for fast feature embedding. In Proceedings of the ACM International Conference on Multimedia, pages 675–678. ACM, 2014. Alex Krizhevsky. One weird trick for parallelizing convolutional neural networks. arXiv preprint arXiv:1404.5997, 2014. Quoc V Le. Building high-level features using large scale unsupervised learning. In Acoustics, Speech and Signal Processing (ICASSP), 2013 IEEE International Conference on, pages 8595– 8598. IEEE, 2013. Yucheng Low, Joseph Gonzalez, Aapo Kyrola, Danny Bickson, Carlos Guestrin, and Joseph M Hellerstein. Graphlab: A new framework for parallel machine learning. arXiv preprint arXiv:1006.4990, 2010. Christopher Olston, Benjamin Reed, Utkarsh Srivastava, Ravi Kumar, and Andrew Tomkins. Pig latin: a not-so-foreign language for data processing. In Proceedings of the 2008 ACM SIGMOD international conference on Management of data, pages 1099–1110. ACM, 2008. Frank Seide, Hao Fu, Jasha Droppo, Gang Li, and Dong Yu. 1-bit stochastic gradient descent and its application to data-parallel distributed training of speech dnns. In Fifteenth Annual Conference of the International Speech Communication Association, 2014. Ilya Sutskever, Oriol Vinyals, and Quoc VV Le. Sequence to sequence learning with neural networks. In Advances in Neural Information Processing Systems, pages 3104–3112, 2014. Christian Szegedy, Wei Liu, Yangqing Jia, Pierre Sermanet, Scott Reed, Dragomir Anguelov, Dumitru Erhan, Vincent Vanhoucke, and Andrew Rabinovich. Going deeper with convolutions. arXiv preprint arXiv:1409.4842, 2014. 6