This month’s post will liberally link to previous posts, because it turns out we have written enough in the past to motivate this month’s work. Our news this March has more monotonicity testing, non-deterministic testing, and image testing. Let me test your patience no more…

by Marek Karpinski and Ronald Markó (arXiv). The notion of nondeterministic graph property testing was introduced by Lovász and Vestergombi and further developed by Gishboliner and Shapira, and the authors. Our August 2014 news post explains these results, rather (ahem) beautifully, so check that out. This result generalizes nondeterministic property testing to hypergraphs, and proves the equivalence to deterministic (standard) property testing. Unlike their previous result, this does not give an explicit function for the complexity of a deterministic tester for a specified nondeterministically testable property. (We only know that it is “some” constant.) This is analogous to the original work of Lovász and Vestergombi, which was non-constructive. These non-constructive results use limit theorems for dense (hyper)graphs. This certainly leaves the open question of getting explicit complexity bounds.

Happy new year! I just looked back at our archives, and saw that PTReview has been on from July 2013. I’m happy (and maybe mildly surprised) that it’s still going strong. And so the chronicling of \(o(n)\) continues…

Much coolness we have to report: permutation testing, linearity testing, distribution testing, and monotonicity testing. Without further ado:

Large permutations and parameter testing by Roman Glebov, Carlos Hoppen, Tereza Klimosova, Yoshiharu Kohayakawa, Daniel Kral, and Hong Liu (arXiv). Just as typical dense graph testing involves checking properties on a random, constant-sized induced subgraph, we can look at permutation testers that test properties of permutations by sampling sub-permutations. The theory of dense graph testing is closely tied to the Szemeredi regularity lemma, the notion of graph limits, and the behavior of subgraph densities. (Check out this survey by Borgs et al.) An analogous theory for permutations has been built by a subset of the authors in a series of papers (survey). There is a notion of permutation properties/parameters that are forcible, meaning that two permutations that have similar frequencies of (a fixed, finite set of) subpermutations have close values of the parameter. This seems naturally connected to testing. Indeed, a permutation parameter that is forcible can be approximated in constant-time by simply approximating the frequencies of the subpermutations. This paper shows that converse is false: there is an constant-time approximable property that is not forcible.

A Self-Tester for Linear Functions over the Integers with an Elementary Proof of Correctness by Sheela Devadas and Ronitt Rubinfeld (arXiv). The venerable property of linearity testing needs no introduction, and has enough history to fill up this month’s post (and the next and the next). But what if the domain is n-bit integers and we care about running time in terms of bit operations? Suppose we wish to check if a program supposed to compute \(f(x) = b\cdot x\) (for fixed, known \(b\)). Inputs like \(2^k\) are easy to evaluate by the checker, and could be used to get faster checkers. This paper gives testers for linearity, multivariate linear functions, and multilinear functions, where computation time is always linear in sample complexity.

\(\ell_p\) Testing and Learning of Discrete Distributions by Bo Waggoner (arXiv). Distribution testing needs little introduction, and we have seen much progress over our time in PTReview. Let’s start with the basics. The seminal work of Batu et al. and later Paninksi showed that the sample complexity of testing uniformity of a distribution \(\mathcal{D}\) over universe \([n]\) is \(\Theta(\sqrt{n}/\epsilon^2)\). Meaning, the sample complexity of checking \(\|\mathcal{D} – \mathcal{U}\|_1 > \epsilon\) (where \(\mathcal{U}\) is uniform on \([n]\)) is \(\Theta(\sqrt{n}/\epsilon^2)\). But what if we had \(\ll \sqrt{n}/\epsilon^2\) samples? From prior work, nothing could be inferred. (Update: As Ilias Diakonikolas pointed out to me, the \(\ell_2\) results were previously known, both by the Batu et al. work and a recent paper by Chan et al. that settles the \(\ell_2\) question.) This paper changes that, and says that we can still infer something about other \(\ell_p\) norms, \(\|\mathcal{D} – \mathcal{U}\|_p\). What impressed me about this paper is the detailed understand of the interplay between \(n, \epsilon, p\). For example, the sample complexity of uniformity testing over \(\ell_p\)-norm for any \(p > 1\) is independent of \(n\). There are many, many results in this paper with an intriguing threshold phenomenon at \(p=2\). For distribution testing in practice, I would think that this result would be of much significance.

New algorithms and lower bounds for monotonicity testing by Xi Chen, Rocco A. Servedio, and Li-Yang Tan (arXiv). Ah yes, Boolean monotonicity testing. Consider the standard coordinate wise partial order on \(\{0,1\}^n\), given by \(\prec\). A function \(f:\{0,1\}^n \rightarrow \{0,1\}\) is monotone if \(\forall x \prec y, f(x) \leq f(y)\). The complexity of property testing (Boolean) monotonicity is one of those tantalizing, simple-to-state questions that is still quite open. I’ll spare you the full story and the epsilons, but here’s the status. The best upper bound is a non-adaptive, one-sided \(O(n^{7/8})\) tester by Chakrabarty and Seshadhri. The best lower bound is a non-adaptive, one-sided lower bound of \(\Omega(\sqrt{n})\) by Fischer et al. This implies an \(\Omega(\log\log n)\) lower bound for general testers. This paper changes all of this. The authors prove a \(\Omega(n^{1/5})\) lower bound for two-sided non-adaptive testers, leading to an exponentially better \(\Omega(\log n)\) lower bound for general testers. The main insight is to focus on monotone vs non-monotone families of linear threshold functions, and show that procedures making few (non-adaptive) queries cannot distinguish between such families. The main hammer is to use recent Central Limit Theorems. As an added bonus, this paper improves the upper bound for monotonicity testing to \(O(n^{5/6})\), with a modified tester (and better analysis) of Chakrabarty and Seshadhri. But can the lower bound be improved?

Boolean function monotonicity testing requires (almost) \(n^{1/2}\) non-adaptive queries by Xi Chen, Anindya De, Rocco A. Servedio, and Li-Yang Tan (arXiv). Yes, the lower bound can be improved. This paper gets the non-adaptive two-sided bound to (almost) \(\Omega(\sqrt{n})\), matching the one-sided Fischer et al bound. The paper proves improved Central Limit Theorems, tailored for this application. The authors, and I with them, believe that this is the true complexity. At that intriguing note, we end 2014!

We’ve got a nice collection of papers this August, ranging from graph property testing, locally testable codes, to new lower bound approaches.

Deconstructions of Reductions from Communication Complexity to Property Testing using Generalized Parity Decision Trees by Roei Tell (ECCC). Blais, Brody, and Matulef (BBM) introduced the beautiful and powerful method of using communication complexity to derive property testing lower bounds. This result is all about understanding this approach is greater depth, and really trying to get at the crux of the reductions. The BBM approach uses rather heavy hammers (communication complexity lower bounds), so could we get away with weaker hammers to get property testing lower bounds? The key insight is that less powerful models can be reduced to existing property testing problems, and these are equivalent to the BBM reductions. While this (as of now) does not lead to newer bounds, it does shed significant light on existing bounds. The weaker model is essentially that of linear decision trees. Such reductions were also discovered by Brushundi, Chakraborty, and Kulkarni, proving \(\Omega(k)\) lower bounds for testing \(k\)-linearity.

Locally Correctable and Testable Codes Approaching the Singleton Bound by Or Meir (ECCC). Locally Testable Codes (LTCs) are codes that have a constant (or polylogarithmic) time tester. It is known that such codes with minor caveats cannot have a constant rate (a number of papers on this topic: Katz and Trevisan, Dinur and Kaufman, Ben-Sasson and Viderman). But when the query complexity can be \(n^\beta\) (where \(n\) is the codeword size), the rate can be arbitrarily close to \(1\) (Viderman, and more refs in the paper). This result takes this fact to the very extreme. The Singleton Bound is a lower bound between the tradeoff between rate and relative distance of a code. Amazingly, this lower bound tradeoff can be achieved by an LTC (and also Locally Correctable Code) testable with \(n^\beta\) queries!

Complexity of Nondeterministic Graph Parameter Testing by Marek Karpinksi and Ronald Mark&#243 (arXiv). It is easiest to think of nondeterminstic graph testing with the example of testing bipartiteness for dense graphs. Suppose we wish to test \(G\) for bipartiteness. As a “certificate”, we are provided a candidate partition \(V = V_1 \cup V_2\). There is a “simple” constant-time tester that can property testing if \(G\) is bipartite with respect to this partition. (Just sample two uniform random vertices, and check if the edge between them respects the bipartition.) So we say that bipartiteness is non-determinstically testable: if \(G\) is far from being bipartite, no certificate can convince the simple tester. The full definition is somewhat complex, but it is basically the same idea. (In this context, “testablility” always refers to constant time testable.) Lovász and Vestergombi proved the striking result that any non-deterministically testable property is also deterministically testable, but their proof did not give an explicit bound on the query complexity. Gishboliner and Shapira fixed that, and gave a tower-type bound for the query complexity. So if the query complexity of the underlying “simple” tester is \(q\), then the non-deterministic property has a tester whose query complexity is a tower of height \(q\). This results goes much further, and shows the final query complexity is at most triply exponential in \(q\). This is quite significant, since it is uses a weaker regularity approach than previous results.

May seems to be a slow month for property testing, with only one paper on locally testable codes. Nonetheless, we have unearthed a paper from April (missed despite our best efforts) on a related topic of hidden set approximations.

The Power of an Example: Hidden Set Size Approximation Using Group Queries and Conditional Sampling by Dana Ron and Gilad Tsur (arXiv). Consider a known universe \(U\) and an unknown subset \(S\). Our aim is to approximately determine \(|S|\). We can perform subset queries: for a query subset \(T\), an oracle returns whether \(T \cap S \neq \emptyset\) or (a more powerful model) the oracle returns a uniform random element in \(T \cap S\). This paper gives a detailed study of this problem under various settings. There is a fascinating array of upper and lower bounds, including situations where \(U\) is ordered and \(T\) must be an interval, the difference between adaptivity and non-adaptivity, arbitrary subset queries, and much more. The topic seems to be quite rich, and it should yield some nice sets of problems to study further. It appears that some lower bounds are quite open. Students, pay attention!

Limitations on Testable Affine-Invariant Codes in the High-Rate Regime by Venkatesan Guruswami, Madhu Sudan, Ameya Velingker, and Carol Wang (ECCC). Locally Testable Codes (LTCs) are codes that (typically) have a constant time tester. Interestingly, recent work on testing codes using a linear number of queries (some small constant fraction of the code length) has connections to the Small Set Expansion (SSE) problem, which in turn is related to the famous Unique Games Conjecture (UGC). A paper by Barak et al. gives a construction of a small set expander (a graph where all small sets have high conductance) where the Laplacian has many small eigenvalues (so it’s “difficult” to find a low conductance cut). This construction uses high-rate LTCs that are linear time testable. In some sense, the better the rate, the better construction one gets. This paper asks how far this idea can go. So the idea is to construct the best rate linear-time testable LTC. Reed-Muller codes fall in this category, but their rate is quite far from optimal. Recent work by Guo et al. gives a slightly better construction, but this is far from the (near) optimal rate achieved by BCH codes (which are not known to be testable). All in all, this paper shows that for affine-invariant codes, the Guo et al. construction is essentially the best testable high-rate code one can get. I have obviously done no justice to the depth of the connections, the intricacies of the parameters, or the overall coolness of this subject. So go read the paper!

So what did we see in February 2014?

Local Algorithms for Sparse Spanning Graphs by Reut Levi, Dana Ron, and Ronitt Rubinfeld (arXiv). Given a graph \(G = (V,E)\), the aim is the find a small sparse spanning subgraph \(G’\) in the local algorithm setting. So, given any edge \((u,v) \in E\), we must determine in sublinear time whether this edge exists in \(G’\). The main result is an \(\Omega(|V|)\) lower bound, and a matching algorithm when \(G\) is either an expander or hyperfinite.

A Characterization of Locally Testable Affine-Invariant Properties via Decomposition Theorems by Yuichi Yoshida (arXiv). This result gives a complete characterization of (two-sided) locally testable affine-invariant properties. These are based on existing decomposition theorems for functions into structured and pseudorandom parts. This result is analogous to that of Alon, Fischer, Newman, and Shapira (link) on characterizations for testability over dense graphs.

Testing probability distributions underlying aggregated data by Clément Canonne and Ronitt Rubinfeld (arXiv). There has been much recent work on distribution testing, and this work discusses stronger models where better algorithms can be obtained. Given a distribution \(D\) over \([n]\), we can get samples from this distribution, and also query the probability \(D(i)\) of any \(i \in [n]\). With these “dual” queries, one gets distribution testers for a variety of problems surpassing previous lower bounds. Interestingly, this is stronger than the conditional model introduced recently (here and here).

Strong Locally Testable Codes with Relaxed Local Decoders by Oded Goldreich, Tom Gur, and Ilan Komargodski (ECCC). For a locally testable code (LTC), the property of being a codeword is efficiently (constant time) testable. A locally decodable code (LDC) is a code from which individual bits of the original message can be recovered in constant time. A relaxed local decoder is allowed to declare failure for recovered a small fraction of message bits. This allows for bypassing the difficult (and prominent) open problem of constructing small LDCs. This result gives a construction of near-linear LTCs with a relaxed local decoder. This result leads to a property that is hard to test but easy to verify (in the MA proofs of proximity sense, refer to last month’s open problem).