After a few slow months, property testing is back with a bang! This month we have a wide range on results ranging from classic problems like junta testing to new models of property testing.

Tolerant Junta Testing and the Connection to Submodular Optimization and Function Isomorphism, by Eric Blais, Clément L. Canonne, Talya Eden, Amit Levi, and Dana Ron (arXiv). The problem of junta testing requires little introduction. Given a boolean function \(f:\{-1,+1\}^n \mapsto \{-1,+1\}\), a \(k\)-junta only depends on \(k\) of the input variables. A classic problem has been that of property testing \(k\)-juntas, and a rich line of work (almost) resolves the complexity to be \(\Theta(k/\epsilon)\). But what of tolerant testing, where we want to tester to accept functions close to being a junta? Previous work has shown that existing testers are tolerant, but with extremely weak parameters. This paper proves: there is a \(poly(k/\epsilon)\)-query tester that accepts every function \(\epsilon/16\)-close to a \(k\)-junta and rejects every function \(\epsilon\)-far from being a \(4k\)-junta. Note that the “gap” between the junta classes is a factor of 4, and this is the first result to achieve a constant gap. The paper also gives testers when we wish to reject functions far from being a \(k\)-junta (exactly matching the definition of tolerant testng), but the tester has an exponential dependence on \(k\). The results have intriguing connections with isomorphism testing, and there is a neat use of constrained submodular minimization.

Testing Pattern-Freeness, by Simon Korman and Daniel Reichman (arXiv). Consider a string \(I\) of length \(n\) and a “pattern” \(J\) of length \(k\), both over some alphabet \(\Sigma\). Our aim is to property test if \(I\) is \(J\)-free, meaning that \(I\) does not contains \(J\) as a substring. This problem can also be generalized to the 2-dimensional setting, where \(I\) and \(J\) are matrices with entries in \(\Sigma\). This formulation is relevant in image testing, where we need to search for a template pattern in a large image. The main results show that these testing problems can be solved in time \(O(1/\epsilon)\), with an intriguing caveat. If the pattern \(J\) has at least \(3\) distinct symbols, the result holds. If \(J\) is truly binary, then \(J\) is not allowed to be in a specified set of forbidden patterns. The main tool is a modification lemma that shows how to “kill” a specified occurrence of \(J\) without introducing a new one. This lemma is not true for the forbidden patterns, resulting in the dichotomy of the results.

Erasure-Resilient Property Testing, by Kashyap Dixit, Sofya Raskhodnikova, Abhradeep Thakurta, and Nithin Varma (arXiv). Property testing begins with a query model for \(f: \mathcal{D} \mapsto \mathcal{R}\), so we can access any \(f(x)\). But what if an adversary corrupted some fraction of the input? Consider monotonicity testing on the line, so \(f: [n] \mapsto \mathbb{R}\). We wish to test if \(f\) is \(\epsilon\)-far from being monotone, but the adversary corrupts/hides an \(\alpha\)-fraction of the values. When we query such a position, we receive a null value. We wish to accept if there exists some “filling” of the corrupted values that makes \(f\) monotone, and reject if all “fillings” keep \(f\) far from monotone. It turns out the standard set of property testers are not erasure resilient, and fail on this problem. This papers gives erasure resilient testers for properties like monotonicity, Lipschitz continuity, and convexity. The heart of the techniques involves a randomized binary tree tester for the line that can avoid the corrupted points.

Partial Sublinear Time Approximation and Inapproximation for Maximum Coverage, by Bin Fu (arXiv). Consider the classic maximum coverage problem. We have \(m\) sets \(A_1, A_2, \ldots, A_m\) and wish to pick the \(k\) of them with the largest union size. The query model allows for membership queries, cardinality queries, and generation of random elements from a set. Note that the size of the input can be thought of as \(\sum_i |A_i|\). Can one approximate this problem without reading the full input? This paper gives a \((1-1/e)\)-factor approximation that runs in time \(poly(k)m\log m\), and is thus sublinear in the input size. The algorithm essentially implements a greedy approximation algorithm in sublinear time. It is shown the the linear dependence on \(m\) is necessary: there is no constant factor approximation that runs in \(q(n) m^{1-\delta}\) (where \(n\) denotes the maximum cardinality, \(q(\cdot)\) is an arbitrary function, and \(\delta > 0\)).

Local Testing for Membership in Lattices, by Karthekeyan Chandrasekaran, Mahdi Cheraghchi, Venkata Gandikota, and Elena Grigorescu (arXiv). Inspired by the theory of locally testable codes, this paper introduces local testing in lattices. Given a set of basis vectors \(b_1, b_2, \ldots\) in \(\mathbb{Z}^n\), the lattice \(L\) is the set of all integer linear combinations of the basic vectors. This is the natural analogue of linear error correcting codes (where everything is done over finite fields). Given some input \(t \in \mathbb{Z}^n\), we wish to determine if \(t \in L\), or is far (defined through some norm) from all vectors in \(L\), by querying a constant number of coordinates in \(t\). We assume that \(L\) is fixed, so it can be preprocessed arbitrarily. This opens up a rich source of questions, and this work might be seen as only the first step in this direction. The papers shows a number of results. Firstly, there is a family of “code formula” lattices for which testers exists (with almost matching lower bounds). Furthermore, with high probability over random lattices, testers do not exist. Analogous to testing codes, if a constant query lattice tester exists, then there exists a canonical constant query tester.

This month isn’t the most exciting for property testing. Just one paper to report (thanks to Clément for catching!), of which only a minor part is actually related to property testing.

Approximating the Spectral Sums of Large-scale Matrices using Chebyshev Approximation, by Insu Han, Dmitry Malioutov, Haim Avron, and Jinwoo Shin (arXiv). The main results of the paper deal with approximating the trace of matrix functions. Consider symmetric matrix \(M \in \mathbb{R}^{d\times d}\), with eigenvalues \(\lambda_1, \lambda_2, \ldots, \lambda_d\). The paper gives new algorithms to approximate \(\sum_i f(\lambda_i)\). Each “operation” of the algorithm is a matrix-vector product, and the aim is to minimize the number of such operations. The property testing application is as follows. Suppose we wish to test if \(M\) is positive-definite (all \(\lambda_i > 0\)). We consider a matrix \(\epsilon\)-far from being positive-definite if the smallest eigenvalue is less than \(-\epsilon \|M\|_F = -\epsilon \sum_i \lambda^2_i\). The main theorems in this paper yield a testing algorithm (under this definition of distance) that makes \(o(d)\) matrix-vector products. While this is not exactly sublinear under a standard query access model, it is relevant when \(M\) is not explicitly represented and the only access is through such matrix-vector product queries.

News for February 2016! Another take on distribution testing, exciting stuff about generating random graphs, and distributed property testing.

Sublinear Random Access Generators for Preferential Attachment Graphs by Guy Even, Reut Levi, Moti Medina, and Adi Rosén (ArXiv). A major aspect of research in “network science” and applied large graph analysis is random graph models. A classic example is the Barabasi-Albert preferential attachment (PA) model, that attempts to model graphs that appear in the real-world. The model is inherently sequential, in that vertices “arrive” in order and connect to other vertices according to a specified distribution. It was not clear how to generate a PA graph in parallel, or how one can construct a local view of a PA graph without constructing all of it. Not clear, that is, until this paper. This really cool result shows the following. Suppose we want to construct an \(n\) vertex PA graph. One can basically query the adjacency list of any vertex in \(poly(\log n)\) time, without constructing the graph! The PA graph is implicitly constructed through the queries. This is really interesting because PA graphs (at least conceptually) are created sequentially, and the order of vertex arrival fundamentally affects the graph structure.

The Uniform Distribution is Complete with Respect to Testing Identity to a Fixed Distribution by Oded Goldreich (ECCC) . This is a follow up to a recent result by Diakonikolas and Kane (covered last month!). This paper reduces testing equality (of a distribution) to a fixed distribution, to the special case to when the fixed distribution is uniform. The notions of reduction in distribution testing were exploited by Diakonikolas and Kane to get optimal testers. While this new paper does not give new bounds, it gives a clean and simplified approach for testing distribution equality. This paper came out of Oded’s lecture notes, and has a nice expository feel to it.

Fast Distributed Algorithms for Testing Graph Properties by Keren Censor-Hillel, Eldar Fischer, Gregory Schwartzman, and Yadu Vasudev (ArXiv) . Connections between distributed algorithms and property testing have been unearthed in the past (most explicitly by Parnas and Ron). This paper directly solves graph property testing problems in the distributed setting. For dense graph testing, it is known that any (constant-query) testing algorithm algorithm can be made non-adaptive, and thus can be simulated in the distributed setting. The results for testing bipartiteness in the sparse graph model are quite interesting, because existing property testers use random walks. This paper gives a distributed implementation of multiple random walks from all vertices, and controls the total congestion of the implementation. This leads to a \(O(\log n)\)-round bipartiteness tester.

(Updating post with an additional paper that we missed in our first posting. Sorry! Feel free to email us at little.oh.of.n@gmail.com to inform us of papers we should mention.)

We’ve got exciting new in November! Optimal results for testing of monotone conjunctions, a new lower bound for monotonicity testing (yay!), and new lower bounds for Locally Testable Codes.

Tight Bounds for the Distribution-Free Testing of Monotone Conjunctions by Xi Chen and Jinyu Xie (arxiv). Consider functions \(f: \{0,1\}^n \rightarrow \{0,1\}\) (amen), and the basic property of being a monotone conjunction. Our notion of distance is with respect to a distribution \(\mathcal{D}\), so the distance between functions \(f\) and \(g\) is \(\mathop{Pr}_{x \sim \mathcal{D}} [f(x) \neq g(x)]\). In the distribution-free testing model, the tester does not know the distribution \(\mathcal{D}\), but is allowed samples from \(\mathcal{D}\). When \(\mathcal{D}\) is uniform, this coincides with standard property testing. There can be significant gaps between standard and distribution-free testing, evidenced by conjunctions. Parnas, Ron, and Samorodnitsky proved that monotone conjunctions can be tested in the vanilla setting in time independent of \(\mathcal{n}\), while Glasner-Servedio prove a \(\Omega(n^{1/5})\) lower bound for distribution-free testing. This paper provides a one-sided, adaptive distribution-free tester that makes \(\widetilde{O}(n^{1/3})\) queries, and they also prove a matching (up to polylog terms and \(\epsilon\)-dependencies) two-sided, adaptive lower bound. This is a significant improvement on the previous upper bound of \(\widetilde{O}(\sqrt{n})\) of Dolev-Ron, as well as over the previous lower bound. Furthermore, the results of this paper hold for general conjunctions.

A Polynomial Lower Bound for Testing Monotonicity by Aleksandrs Belovs and Eric Blais (arxiv). Surely the reader needs little introduction to testing monotonicity of Boolean functions, and a previous Open Problem post should help. A quick summary: we want to test monotonicity of \(f:\{0,1\}^n \rightarrow \{0,1\}\). The best upper bound is the \(\widetilde{O}(\sqrt{n})\) non-adaptive, one-sided tester of Khot et al. There is a matching, non-adaptive lower bound (up to polylog terms) by Chen et al, implying the best-known adaptive lower bound of \(\Omega(\log n)\). Can adaptivity help? This paper proves an adaptive lower bound of \(\Omega(n^{1/4})\). Exciting! The approach of Chen et al is to focus on monotonicity of linear threshold functions (technically, regular LTFS, where the weights are bounded). The authors show that this property can be tested in \(\mathop{poly}(\log n)\) time, shooting down hopes of extending the LTF approach for adaptive lower bounds. The key insight is to work with the distribution of Talagrand’s random DNF instead, which is the most noise sensitive DNF. (Talagrand strikes again. He helps the upper bound, he helps the lower bound.) Perturbations of this DNF lead to non-monotone functions, the paper proves that these distributions cannot be distinguished in \(o(n^{1/4})\) queries.

Lower bounds for constant query affine-invariant LCCs and LTCs by Arnab Bhattacharyya and Sivakanth Gopi (arxiv). Think of any code as a set/property of codewords in the domain \(\Sigma^N\). Such a code is an \(r\)-LTC if it has an \(r\)-query property tester. An \(r\)-LCC is has the property that, given any \(x \in \Sigma^n\) sufficiently close to a codeword, one can determine any coordinate of the codeword using \(r\)-queries. A fundamental question is to understand the length of LCCs and LTCS, or alternately (fixing \(\Sigma^N\)) determining the largest possible set of codewords. Existing constructions typically have much symmetry, either linear or affine invariance. (Check out Arnab Bhattacharyya’s survey on affine invariant properties for more details.) It is convenient to think of any codeword as a function \(f: [N] \rightarrow \Sigma\), and furthermore, think of \([N]\) as a vector space \(\mathbb{K}^n\). The best known construction of Guo, Kopparty, and Sudan yields (affine-invariant) LCCs of size \(\exp(\Theta(n^{r-1}))\) and LTCs of size \(\exp(\Theta(n^{r-2}))\) (many dependences on \(\mathbb{K}\), the rate, etc. are hidden in the big-Oh). This paper show that these bounds are actually the best achievable by any affine-invariant code. (Previous lower bounds of Ben-Sasson and Sudan only held for linear codes.) The intriguing and wide-open question is to prove such lower bounds without affine invariance.

Alas! A dry month, with no property testing or sublinear papers. Maybe next month…

The summer month of July brings to us exciting work on distribution testing, scale-free graph testing, personal PageRank estimation, and (hold your breath!) cake cutting. Let’s hear more…

Avid followers of PTReview need little introduction to distribution testing, since we’ve seen much work on this topic over the past year. We have two concurrent results that really push the state-of-the-art of distribution testing.
Testing Shape Restrictions of Discrete Distributions by Clément Canonne, Ilias Diakonikolas, Themis Gouleakis, and Ronitt Rubinfeld (arxiv). Most (sublinear) distribution testing results work for structured classes, like monotone, k-modal, Poisson, multinomial etc. distributions. This paper gives an overarching framework that unifies these concepts in terms of succinctness. Consider a class \(\mathbb{C}\) of distributions over \([n]\). This class is succinct if for any distribution \(\mathcal{D} \in \mathbb{C}\), one can partition \([n]\) into a “small” number of intervals such that \(\mathcal{D}\) is “approximately constant” with each interval. The main punchline is that any such class has a sublinear sample distribution tester. Amazingly, the generic tester given is optimal (up to poly log factors and the distance parameter)! Much previous work is subsumed as a corollary of this main theorem, and the paper also gives a lower bound framework to get matching lower bounds.

Optimal Testing for Properties of Distributions by Jayadev Acharya, Constantinos Daskalakis, and Gautam Kamath (arxiv). This paper concurrently gets many of the results of the previous paper using completely different techniques. The focus is on monotone, product, and log-concave distributions, where the paper gives completely optimal testers (both in terms on the domain size and the distance parameter). The key insight is to focus on the \(\chi^2\)-distance between distributions, though distribution testing typically uses the total-variation (or \(\ell_1\)) distance. The tester is also robust/tolerant, in that yes instances include distributions close to the class considered.

Bidirectional PageRank Estimation: From Average-Case to Worst-Case by Peter Lofgren, Siddhartha Banerjee, and Ashish Goel (arxiv). Not your typical sublinear algorithms paper, but that never stopped us. Personalized PageRank (PPR) is a fundamental similarity notion between vertices in a graph, and is (sort of) the probability that a short random walk from \(u\) ends at \(v\). Naturally, estimating the PPR value from \(u\) to \(v\) can be done in time \(O(1/p)\), where \(p\) is the PPR value. (Simply take many random walks from \(u\), and see how often you hit \(v\).) Can you do sublinear in this value? The answer is yes, when the graph is undirected and has bounded degree. One can do this in \(O(1/\sqrt{p})\) time, a significant improvement. It basically uses bidirectional search, by doing forward walks from \(u\) and backward walks from \(v\). These ideas were developed in earlier work of Lofgren et al, based on true-blooded sublinear graph algorithms for expander reconstruction by Kale, Peres, and Seshadhri, which are further based on truer-blooded property testing questions by Goldreich and Ron on expansion testing. The best thing, these algorithms work in practice!