# News for November 2018

Alas! Sorry for the delay, but November was a good month with five seven papers. We missed a few papers from September, so please look at the updated post for September as well. (Update: Thanks to Omri Ben-Eliezer and an anonymous poster for pointing out two missed papers.)

From Local to Robust Testing via Agreement Testing, by Irit Dinur, Tali Kaufman, and Noga Ron-Zewi (ECCC). Consider a code, which is simply a subset of $$C \subseteq \mathbb{F}^n_q$$. (Typically, we think of linear codes, which form a linear subspace.) The study of locally testable codes is a rich area. Consider an input $$x \in \mathbb{F}^n_q$$. A local tester for code/property $$C$$ queries, according to some distribution, a subset $$K$$ of $$Q$$ coordinates in $$x$$, and rejects if the “view” $$x|_K$$ is inconsistent with any codeword. The rejection probability should be proportional to the distance of $$x$$ to $$C$$, denoted $$dist(x,C)$$. A robust tester demands the average distance of $$x|_K$$ to the corresponding property $$C|_K$$ must be proportional to $$dist(x,C)$$. This is a much stronger demand than that of just local testing. The main result is local testability implies robust testability for the special class of lifted codes. The proof goes via agreement testing, where an algorithm gets access to some fragments of $$x$$, and must decide if these fragments are consistent with a codeword.

Testing local properties of arrays, by Omri Ben-Eliezer (ECCC). Consider a function $$f:[n]^d \to \Sigma$$, where $$\Sigma$$ is finite. Now, suppose we define a property $$\mathcal{P}$$ with respect to a set of forbidden $$[k]^d$$ (consecutive) subarrays. Such a property is called $$k$$-local. The inspiration for this work is a line of research on testing of image properties. Note that for $$k=2,3$$, this notion subsumes a number of properties, such as monotonicity, $$c$$-Lipschitz, submodularity, convexity, etc. The main result is a one-sided, non-adaptive tester for all such properties. Ignoring $$\epsilon, k$$ dependencies, the query complexity is $$O(\log n)$$ for $$d=1$$ and $$O(n^{d-1})$$ for $$d > 1$$. Remarkably, there is no dependence on the alphabet size $$|\Sigma|$$. Moreover, the bound above is optimal for one-sided, non-adaptive testers. The basic idea to pick blocks of various sizes, and query all points on the boundary. Then, the tester can determine (by sort of brute force search) if this is consistent with some function in $$\mathcal{P}$$. The key is to prove that for a function that is far from $$\mathcal{P}$$, there are many blocks that cannot be consistent with $$\mathcal{P}$$.

Erasures vs. Errors in Local Decoding and Property Testing, by Sofya Raskhodnikova, Noga Ron-Zewi and Nithin Varma (ECCC). The main theme of this paper is the notion of erasure-resilient testing. Consider property $$\mathcal{P}$$. Suppose $$\alpha$$-fraction of the coordinates of input $$x$$ is hidden. Thus, queries to these coordinates of $$x$$ return nothing. Our aim is to accept if there is some “filling in” of the hidden coordinates that make $$x$$ satisfy $$\mathcal{P}$$, and reject if for all fillings, $$x$$ remains from from $$\mathcal{P}$$. If $$\alpha = 0$$, this is standard property testing. Suppose there exists an $$(\alpha, \alpha+\epsilon)$$-tolerant tester for $$\mathcal{P}$$. (I’m fudging factors a bit here for readability.) We get an $$\epsilon$$-tester with $$\alpha$$-fraction of erasures, by simply filling $$x$$ in arbitrarily. So this model is in between vanilla and tolerant testing. The main result is a separation. There is a property that is constant-query testable under erasures, but requires $$n^{\Omega(1)}$$ queries to tolerant test (with the above parameters). One of the main technical results is a list decoder for the Hadamard code that can tolerate erasures of any $$\alpha \lt 1$$.

Domain Reduction for Monotonicity Testing: A o(d) Tester for Boolean Functions on Hypergrids, by Hadley Black, Deeparnab Chakrabarty, and C. Seshadhri (ECCC). Consider monotonicity testing of Boolean functions of the hypergrid, $$f:[n]^d \to \{0,1\}$$. It is known that there are $$O(d)$$ testers with complexity independent of $$n$$, while for $$n=2$$, there are $$o(d)$$ testers. The main result is a tester with both these properties. The paper gives an $$O(d^{5/6})$$ tester for such functions. The main technical ingredient is a domain reduction theorem. Consider restrictions of $$f$$ to a uniform random $$[poly(d\epsilon^{-1})]^d$$ hypergrid. The paper proves that if $$f$$ is $$\epsilon$$-far from monotone, then (with high probability) so is the restriction. Thus, for the purposes of monotonicity testing, one can simply assume that $$n = poly(d)$$. This gives a black-box method to get testers independent of $$n$$.

On the Testability of Graph Partition Properties, by Yonatan Nakar and Dana Ron (ECCC). Consider property testing on dense graphs. The seminar result of Goldreich-Goldwasser-Ron studies graph partitioning properties. Such a property is defined as follows. Fix constant $$k$$. We wish to partition the vertices of graph $$G$$ into $$V_1, V_2, \ldots, V_k$$. For each $$i$$, $$|V_i|/n \in [\rho^L_i, \rho^U_i]$$. Furthermore, for each $$i, j$$, the number of edges between $$V_i, V_j$$ must be in $$[\rho^L_{i,j}n^2, \rho^U_{i,j}n^2]$$. (All the $$\rho$$s are parameters of the property.) This subsumes properties like $$k$$-colorability, and the classic result is a $$poly(1/\epsilon)$$-tester for these properties. Note that the edge conditions are absolute, with respect to $$n^2$$, and not with respect to $$|V_i|\cdot |V_j|$$. This paper allows for relative bounds of the form $$[\rho^L_{i,j}|V_i|\cdot |V_j|, \rho^U_{i,j}|V_i|\cdot |V_j|]$$. This is far more expressive, subsuming properties like having large cliques, or large complete bipartite graphs. One of the main results is a (two-sided) $$poly(1/\epsilon)$$ tester for all such properties. Furthermore, there is a characterization of such properties that are one-sided testable in $$poly(1/\epsilon)$$ queries. For such properties, it is proven that the density conditions (between $$V_i, V_j$$) must either be unconstrained or have $$\rho^L_{i,j} = \rho^U_{i,j}$$ be either 1 or 0.

Limits of Ordered Graphs and Images, by Omri Ben-Eliezer, Eldar Fischer, Amit Levi, and Yuichi Yoshida (arXiv). The theory of graph limits is a deep topic, with connections to property testing, regularity lemmas, and analysis of real-world networks. Roughly speaking, consider a sequence $$\{G_n\}$$ of graphs where the (relative) frequency of any fixed subgraph converges. Then, we can define a limit of the graphs, called a graphon, which is a measurable function $$W:[0,1]^2 \to [0,1]$$. This paper discovers appropriate limits of ordered graphs. Note that images can be represented as ordered graphs. An ordered graph is a symmetric function $$G:[n]^2 \to \{0,1\}$$ (with the trivial automorphism group). The graphon definitions and limits do not generalize here. There is a sequence of ordered graphs where the ordered subgraph frequencies converge, but one cannot associate any graphon as the limit. This paper discovered a new object called an orderon, which is a function $$W:([0,1]^2)^2 \to [0,1]$$. Orderons can be shown to be the limits of sequences of ordered graphs.

Two Party Distribution Testing: Communication and Security, by Alexandr Andoni, Tal Malkin, and Negev Shekel Nosatzki (arXiv). This paper gives a new take on distribution testing. Suppose Alice and Bob have each have access to $$t$$ samples from (unknown) distributions $$p$$ and $$q$$ respectively. Their aim is to distinguish $$p = q$$ from $$\|p-q\|_{TV} > \epsilon$$. How much communication is required? In the standard setting, it is known that $$n^{2/3}$$ (ignoring $$\epsilon$$) is necessary and sufficient. In the communication setting, the bound turns out to basically be $$\Theta((n/t)^2)$$. Observe how for $$t=n^{2/3}$$, the communication is $$n^{2/3}$$ (saying that Alice is simply going to send all her samples to Bob). But for larger $$t$$, the communication decreases. Since Alice can learn more about her distribution, she can use sketching techniques to reduce her communication. There is an analogous result for testing independence of two distributions. Furthermore, one can even require these protocols to be secure, so Alice and Bob learn nothing about each others samples (other than the final answer).

# News for October 2018

As October draws to a close, we are left with four new papers this month.

Testing Matrix Rank, Optimally, by Maria-Florina Balcan, Yi Li, David P. Woodruff, Hongyang Zhang (arXiv). This work investigates the problem of non-adaptively testing matrix properties, in both the standard query model and the more general sensing model, in which the algorithm may query the component-wise inner product of the matrix with “sensing” matrices. It proves tight upper and lower bounds of $$\tilde \Theta(d^2/\varepsilon)$$ for the query model, and eliminating the dependence on $$\varepsilon$$ in the sensing model. Furthermore, they introduce a bounded entry model for testing of matrices, in which the entries have absolute value bounded by 1, in which they prove various bounds for testing stable rank, Schatten-$$p$$ norms, and SVD entropy.

Testing Halfspaces over Rotation-Invariant Distributions, by Nathaniel Harms (arXiv). This paper studies the problem of testing from samples whether an unknown boolean function over the hypercube is a halfspace. The algorithm requires $$\tilde O(\sqrt{n}/\varepsilon^{7})$$ random samples (which has a dependence on $$n$$ which is tight up to logarithmic factors) and works for any rotation-invariant distribution, generalizing previous works that require the distribution be Gaussian or uniform.

Testing Graphs in Vertex-Distribution-Free Models, by Oded Goldreich (ECCC). While distribution-free testing has been well-studied in the context of Boolean functions, it has not been significantly studied in the context of testing graphs. In this context, distribution-free roughly means that the algorithm can sample nodes of the graph according to some unknown distribution $$D$$, and must be accurate with respect to the measure assigned to nodes by the same distribution. The paper investigates various properties which may be tested with a size-independent number of queries, including relationships with the complexity of testing in the standard model.

A Theory-Based Evaluation of Nearest Neighbor Models Put Into Practice, by Hendrik Fichtenberger and Dennis Rohde (arXiv). In the $$k$$-nearest neighbor problem, we are given a set of points $$P$$, and the answer to a query $$q$$ is the set of the $$k$$ points in $$P$$ which are closest to $$q$$. This paper considers the following property testing formulation of the problem: given a set of points $$P$$ and a graph $$G = (P,E)$$, is each point $$p \in P$$ connected to its $$k$$-nearest neighbors, or is it far from being a $$k$$NN graph? The authors prove upper and lower bounds on the complexity of this problem, which are both sublinear in the number of points $$n$$.

# News for September 2018

Last month was slower than usual, with only one property testing paper. We did miss a few papers, so hopefully this list is now up to date.

Property testing and expansion in cubical complexes, by David Garber, Uzi Vishne (arXiv). Consider the question of testing if an arbitrary function $$f\colon V\times V \to\{-1,1\}$$ is of the form $$f(x,y) = h(x)h(y)$$ for some $$h\colon V\to\{-1,1\}$$. An intuitive one-sided test, shown to work by Lubotzky and Kaufman (2014), is to pick uniformly random $$x,y,z\in V$$ and check that $$f(x,y)f(y,z)f(z,x)=1$$. This paper considers the high-dimensional generalization of testing the property that a function$$f\colon V\times V \times V\times V \to\{-1,1\}$$ is of the form $$f(w,x,y,z) = \alpha\cdot h(w,x)h(y,x)h(y,z) h(w,z)$$, for some $$h\colon V\times V\to\{-1,1\}$$ and sign $$\alpha\in\{-1,1\}$$. The authors derive necessary and sufficient conditions for testability of this property, by formulating it in the language of incidence geometry and exploiting this connection.

Local Computation Algorithms for the Lovász Local Lemma, by Dimitris Achlioptas, Themis Gouleakis and Fotis Iliopoulos (arXiv). There has been significant work in the past decade on constructive versions of the Lovász Local Lemma (LLL), since the seminal work of Moser-Tardos. This paper designs news Local Computation Algorithms (LCAs) for the LLL. It’s best to consider the problem of $$k$$-SAT. Consider a $$k$$-CNF $$\phi$$ with $$n$$ variables, $$m$$ clauses, where every variable is in at most $$d$$ clauses. By the LLL, if $$d \leq 2^k/ke$$, then $$\phi$$ is satisfiable. An LCA would compute any bit of a satisfying assignment, by making sublinear queries into $$\phi$$. This was first studied by Rubinfeld-Tamir-Vardi-Xie. Their LCA would make polylogarithmically many queries, but requires a stronger condition that what LLL achieves. This paper gives the first sublinear LCA with precisely the LLL conditions, though the number of queries is $$n^\beta$$ (for $$\beta \lt 1$$). The main result is an LCA for an abstract LLL formulation, that also leads to LCAs for graph coloring. Roughly speaking, for a graph with maximum degree $$\Delta$$ where all neighborhoods are sufficiently far from cliques, the LLL shows that the chromatic number bound of $$\Delta + 1$$ can be beaten. This result gives an LCA for graph coloring under these LLL conditions.

Sublinear Time Low-Rank Approximation of Distance Matrices, by Ainesh Bakshi and David P. Woodruff (arXiv). Consider two sets of points $$P$$ and $$Q$$ in a metric space, with $$m$$ and $$n$$ points respectively. The $$m \times n$$ distance matrix $$A$$ contains all pairwise distances between them. This paper studies approximating $$A$$ using a low rank representation, without reading all the entries in $$A$$. The main result is as follows. For rank parameter $$k$$, let $$A_k$$ be the closest (by Frobenius norm) rank-$$k$$-approximation to $$A$$. There is a $$O(m^{1+\gamma} + n^{1+\gamma}poly(k\epsilon^{-1}))$$ (for arbitrary $$\gamma > 0$$) algorithm that outputs a rank $$k$$-matrix $$B$$ with the following property: $$\|A-B\|^2_F \leq \|A-A_k\|^2_F + \epsilon \|A\|^2_F$$. Interestingly, there is a lower bound showing that a $$o(mn)$$ algorithm cannot get a multiplicative approximation. One technical ingredient is a method to sample column norms of $$A$$, under an approximate triangle inequality constraint. This allows one to compute smaller matrices that approximate $$A$$, on which one can directly compute an approximate rank-$$k$$ decomposition.

On Solving Linear Systems in Sublinear Time, by Alexandr Andoni, Robert Krauthgamer, Yosef Pogrow (arXiv). Solving Laplacian linear systems is an immensely deep area, with lots of exciting recent work. This paper studies sublinear algorithms for such problems. Consider a Laplacian matrix $$L$$ (think of $$I – A/d$$, for adjacency matrix $$A$$ of a $$d$$-regular graph). The aim is to solve $$Lx = b$$, for $$x, b \in {\mathbb R}^n$$. Let the solution be $$x^*$$. The main result shows that one can approximate any entry in sublinear time. Specifically, for any coordinate $$i$$, one can output an approximate $$\hat{x_i}$$ such that $$|\hat{x_i} – x^*_i| \leq \|x^*\|_\infty$$. The running time is essentially $$d\epsilon^{-2}\kappa^3$$, where $$\kappa$$ is a bound on the condition number of $$L$$. There are generalizations for Symmetrically Diagonally Dominant (SDD) matrices, a generalization of Laplacians. There is an $$\Omega(n^{1/d^2})$$ lower bound for solving general PSD systems, and a lower bound showing that $$\Omega(\kappa^2)$$ queries into $$b$$ are necessary.

# News for August 2018

Three papers this month close out Summer 2018.

Test without Trust: Optimal Locally Private Distribution Testing, by Jayadev Acharya, Clément L. Canonne, Cody Freitag, and Himanshu Tyagi (arXiv). This work studies distribution testing in the local privacy model. While private distribution testing has recently been studied, requiring that the algorithm’s output is differentially private with respect to the input dataset, local privacy has this requirement for each individual datapoint. The authors prove optimal upper and lower bounds for identity and independence testing, using a novel public-coin protocol named RAPTOR which can outperform any private-key protocol.

Testing Graph Clusterability: Algorithms and Lower Bounds, by Ashish Chiplunkar, Michael Kapralov, Sanjeev Khanna, Aida Mousavifar, and Yuval Peres (arXiv). This paper studies the problem of testing whether a graph is $$k$$-clusterable (based on the conductance of each cluster), or if it is far from all such graphs — this is a generalization of the classical problem of testing whether a graph is an expansion. It manages to solve this problem under weaker assumptions than previously considered. Technically, prior work embedded a subset of the graph into Euclidean space and clustered based on distances between vertices. This work uses richer geometric structure, including angles between the points, in order to obtain stronger results.

Near log-convexity of measured heat in (discrete) time and consequences, by Mert Saglam (ECCC). Glancing at the title, it might not be clear how this paper relates to property testing. The primary problem of study is the quantity $$m_t = uS^tv$$, where $$u, v$$ are positive unit vectors and $$S$$ is a symmetric substochastic matrix. This quantity can be viewed as a measurement of the heat measured at vector $$v$$, after letting the initial configuration of $$u$$ evolve according to $$S$$ for $$t$$ time steps. The author proves an inequality which roughly states $$m_{t+2} \geq t^{1 – \varepsilon} m_t^{1 + 2/t}$$, which can be used as a type of log-convexity statement. Surprisingly, this leads to lower bounds for the communication complexity of the $$k$$-Hamming problem, which in turns leads to optimal lower bounds for the complexity of testing $$k$$-linearity and $$k$$-juntas.

# News for July 2018

Three Four papers for July. New sublinear graph algorithms, distribution testing under new models, and sublinear matrix algorithms. Onward ho…
(Sorry Amit, for missing your paper on sublinear matrix algorithms.)

Metric Sublinear Algorithms via Linear Sampling, by Hossein Esfandiari and Michael Mitzenmacher (arXiv). Consider a weighted clique $$G = (V,E)$$ where $$V$$ is a set of points in a metric space and edge weights are metric distances. In this setting, sublinear algorithms are those that make $$o(n^2)$$ edge queries. This paper studies problems like densest subgraph and maxcut in this setting. The key method is a sparsifying algorithm that achieves the following. (I paraphrase their language.) Consider a positive parameter $$\alpha$$, and let $$w(e)$$ denote the weight of edge $$e$$. The aim is to get a subgraph $$H$$ that contains every edge $$e$$ (in $$G$$) with independent probability $$\min(w(e)/\alpha, 1)$$. Furthermore, this subgraph should be obtained in time linear in the number of edges in $$H$$ (hence the title of the paper). For problems like 1/2-approximating the densest subgraph and PTASes for maxcut, the results show that for a carefully chosen $$\alpha$$, approximate solutions on $$H$$ give solutions of comparable quality on $$G$$. These results cleanly generalize to settings where edge weights satisfy triangle inequality with some multiplicative penalty.

Sublinear Algorithms for ($$\Delta$$ + 1) Vertex Coloring, by Sepehr Assadi, Yu Chen, and Sanjeev Khanna (arXiv). Arguably, the first thing you learn about vertex coloring is that a graph with maximum degree $$\Delta$$ admits a $$(\Delta+1)$$-coloring, that can be found in linear time. But what about sublinear time/space? I like this! You take a simple classical fact, throw in sublinear constraints, and it opens up a rich theory. This paper shows a non-adaptive $$O(n^{3/2})$$-time algorithm for this problem, and gives a nearly matching lower bound. There are also results for streaming and parallel computation, but let’s focus on the sublinear result. It is remarkable that there is no loss in colors in going to sublinear time. (In contrast, the papers shows an $$\Omega(n^2)$$ lower bound for constructing a maximal matching.) The main technical tool is a list coloring result, where each vertex is given a list of colors and much choose its own from that list. Obviously, if each list is $$[\Delta + 1]$$, such a coloring is possible. The paper proves that even if each list is an independent $$O(\log n)$$-sized sample of $$[\Delta+1]$$, a valid coloring is still possible. The final algorithm is pretty involved, and uses this meta-algorithm as a building block.

Anaconda: A Non-Adaptive Conditional Sampling Algorithm for Distribution Testing, by Gautam Kamath and Christos Tzamos (ECCC). The standard model for distribution testing is access to samples from the unknown distribution $$\mathcal{D}$$ with support $$[n]$$. This has attracted much attention with a rich set of results, and the complexity classic problems of uniformity, identity, and equivalence are well understood. But there are alternate models, such as the model of conditional samples (Chakraborty-Fischer-Goldhirsh-Matsliah ’13 and Canonne-Ron-Servedio ’14). For any subset $$S \subseteq [n]$$, we can get a random sample from $$\mathcal{D}$$ restricted to $$S$$. This adds an algorithmic dimension to distribution testing. This paper studies the power of non-adaptive conditional (NACOND) queries. The main result is that uniformity, identity, and equivalence are testable with $$\mathrm{poly}(\log n)$$ queries. (There are existing $$\Omega(\log n)$$ lower bounds for all these problems.) The heart of these algorithms is a procedure ANACONDA that tries to find a set $$S$$ where some element has a high probability, relative to the mass of $$S$$.

Sublinear-Time Quadratic Minimization via Spectral Decomposition of Matrices, by Amit Levi and Yuichi Yoshida (arXiv). When it comes to fundamental problems, it’s hard to beat quadratic minimization. Given a matrix $$A \in \mathbb{R}^{n\times n}$$, we wish to find $$v \in \mathbb{R}^n$$ that minimizes $$v^TAv$$. (This is basically a singular value/vector problem.) One may have additional terms in the objective, depending on $$v^Tv$$ or $$v^Tb$$ (for fixed vector $$b$$). This paper gives sublinear algorithms for this problem. A natural approach is to simply subsample $$k$$ rows and columns to get submatrix $$B$$, solve the problem for $$B$$, and hope for the best. This idea has a rich history from seminal work of Frieze-Kannan. Recently, Hayashi-Yoshida show that constant $$k$$ (only depending on error parameters) suffice for getting a non-trivial approximation for this problem. Unfortunately, the solution error depends on the $$\ell_\infty$$-norm of the solution. This paper shows that for polylogarithmic $$k$$, one can get an error depending on the $$\ell_2$$-norm of the solution. This is a significant improvement, especially for sparse solution vectors. The main technical workhorse is a new matrix decomposition theorem, that shows how any matrix can be written as a sum of a few block matrix, and a low-norm “error” matrix. Admirably, the paper shows a number of experiments,
showing the effectiveness of this technique for eigenvalue computations. It’s very nice to see how ideas from sublinear algorithms might have a practical impact.

# News for June 2018

The summer gets off to a flying start, with three property testing papers, spanning differential privacy, distribution testing, and juntas in Gaussian space!

On closeness to $$k$$-wise uniformity, by Ryan O’Donnell and Yu Zhao (arXiv)
In this paper, the authors consider the following structural question about probability distributions over the Boolean hypercube $$\{-1,1\}^n$$: ” what is the relation between total variation distance $$\delta$$ to $$k$$-wise independence, and bound $$\varepsilon$$ on the Fourier coefficients of the distribution on degrees up to $$k$$?”

While this question might seem a bit esoteric at first glance, it has direct and natural applications to derandomization, and of course to distribution testing (namely, to test $$k$$-wise independence and its generalization, $$(\varepsilon, k)$$-wise independence of distributions over the hypercube).

The main contribution here is to improve (by a $$(\log n)^{O(k)}$$ factor) the bounds on $$\delta (n,k,\varepsilon)$$ over the previous work by Alon et al. [AAK+07], making them either tight (for $$k$$ even) or near-tight. To do so, the authors introduce a new hammer to the game, using linear programming duality in the proof of both their upper and lower bounds.

Property Testing for Differential Privacy, by Anna Gilbert and Audra McMillan (arXiv)
Differential privacy, as introduced by Dwork et al., needs no introduction. Property testing, especially on this website, needs even less. What about a combination of the two? Namely, given black-box access to an algorithm claiming to perform a differentially private computation, how to test whether this is indeed the case?

Introducing and considering this quite natural question for the first time [01/31/2019: see erratum below], this work shows, roughly speaking, that testing differential privacy is hard. Specifically, they show that for many notions of differential privacy (pure, approximate, and their distributional counterparts), testing is either impossible or possible but not with a sublinear number of queries (even when the tester is provided with side information about the black-box). In other terms, as the authors put it: trusting the privacy of an algorithm “requires compromise by either the verifier or algorithm owner” (and, in the latter case, even then it’s not a simple matter).

Is your data low-dimensional?, by Anindya De, Elchanan Mossel, and Joe Neeman (arXiv)
(Well, is it?) To state it upfront, I am biased here, as it is a problem I was very eager to see investigated to begin with. To recap, the question is as follows: “given query access to some unknown Boolean-valued function $$f\colon \mathbb{R}^n \to \{-1,1\}$$ over the high-dimensional space $$\mathbb{R}^n$$ endowed with the Gaussian measure, how can one check whether $$f$$ only depends on “few” (i.e., $$k \ll n$$) variables?”

This is the continuous, Gaussian version of the (quite famous) junta testing problem, which has gathered significant attention over the past years (the Gaussian version has, to the best of my knowledge, never been investigated). Now, the above formulation has a major flaw: specifically, it is uninteresting. In Gaussian space*, who cares about the particular basis I expressed my input vector in? So a more relevant question, and that that the authors tackle, is the more robust and natural one: “given query access to some unknown Boolean-valued function $$f\colon \mathbb{R}^n \to \{-1,1\}$$ over the high-dimensional space $$\mathbb{R}^n$$ endowed with the Gaussian measure, how can one check whether $$f$$ only depends on a low-dimensional linear combination of the variables?” Or, put differently, does all the relevant information for $$f$$ live in a low-dimensional subspace?

De, Mossel, and Neeman show how can do this, non-adaptively, with a query complexity independent of the dimension $$n$$ (hurray!), but instead polynomial in $$k$$, the distance parameter $$\varepsilon$$, and the surface area $$s$$ of $$f$$. And since this last parameter may seem quite arbitrary, they also proceed to show that a polynomial dependence in this $$s$$ is indeed required.

*”In Gaussian space, no one can hear you change basis?”

Erratum (01/31/2019): It was brought to our attention that our overview of “Property Testing for Differential Privacy” was overlooking a key part of the literature; specifically, a work of Dixit et al. (TCC 2013)  which introduces this very question. From the abstract:

How does one ensure that [those third-parties] have implemented their algorithms in a way which meet the specifications of the privacy requirements? […] In this work, we propose a new approach to the above problem which we call privacy testing. We do this by formulating the above problem in the well-studied framework of property testing.

# News for May 2018

Six papers for May, including new models, hierarchy theorems, separation results, resolution of conjectures, and a lot more fun stuff. A lot of things to read this month!

Lower Bounds for Tolerant Junta and Unateness Testing via Rejection Sampling of Graphs, by Amit Levi and Erik Waingarten (ECCC). This paper proves a number of new lower bounds for tolerant testing of Boolean functions, including non-adaptive $$k$$-junta testing and adaptive and non-adaptive unateness testing. Combined with upper bounds for these and related problems, these results establishes separation between the complexity of tolerant and non-tolerant testing for natural properties of Boolean functions, which have so far been elusive. As a technical tool, the authors introduce a new model for testing graph properties, termed the rejection sampling model. In this model, the algorithm queries a subset $$L$$ of the vertices, and the oracle will sample an edge uniformly at random and output the intersection of the edge endpoints with the query set $$L$$. The cost of an algorithm is measured as the sum of the query sizes. In order to prove the above lower bounds (in the standard model), they show a non-adaptive lower bound for testing bipartiteness (in their new model).

Hierarchy Theorems for Testing Properties in Size-Oblivious Query Complexity, by Oded Goldreich (ECCC). This work proves a hierarchy theorem for properties which are independent of the size of the object, and depend only on the proximity parameter $$\varepsilon$$. Roughly, for essentially every function $$q : (0,1] \rightarrow \mathbb{N}$$, there exists a property for which the query complexity is $$\Theta(q(\varepsilon))$$. Such results are proven for Boolean functions, dense graphs, and bounded-degree graphs. This complements hierarchy theorems by Goldreich, Krivelevich, Newman, and Rozenberg, which give a hierarchy which depends on the object size.

Finding forbidden minors in sublinear time: a $$O(n^{1/2+o(1)})$$-query one-sided tester for minor closed properties on bounded degree graphs, by Akash Kumar, C. Seshadhri, and Andrew Stolman (ECCC). At the core of this paper is a sublinear algorithm for the following problem: given a graph which is $$\varepsilon$$-far from being $$H$$-minor free, find an $$H$$-minor in the graph. The authors provide a (roughly) $$O(\sqrt{n})$$ time algorithm for such a task. As a concrete example, given a graph which is far from being planar, one can efficiently find an instance of a $$K_{3,3}$$ or $$K_5$$ minor. Using the graph minor theorem, this implies analogous results for any minor-closed property, nearly resolving a conjecture of Benjamini, Schramm and Shapira.

Learning and Testing Causal Models with Interventions, by Jayadev Acharya, Arnab Bhattacharyya, Constantinos Daskalakis, and Saravanan Kandasamy (arXiv). This paper considers the problem of learning and testing on causal Bayesian networks. Bayesian networks are a type of graphical model defined on a DAG, where each node has a distribution defined based on the value of its parents. A causal Bayesian network further allows “interventions,” where one may set nodes to have certain values. This paper gives efficient algorithms for learning and testing the distribution of these models, with $$O(\log n)$$ interventions and $$\tilde O(n/\varepsilon^2)$$ samples per intervention

Property Testing of Planarity in the CONGEST model, by Reut Levi, Moti Medina, and Dana Ron (arXiv). It is known that, in the CONGEST model of distributed computation, deciding whether a graph is planar requires a linear number of rounds. This paper considers the natural property testing relaxation, where we wish to determine whether a graph is planar, or $$\varepsilon$$-far from being planar. The authors show that this relaxation allows one to bypass this linear lower bound, obtaining a $$O(\log n \cdot \mathrm{poly(1/\varepsilon))}$$ algorithm, complemented by an $$\Omega(\log n)$$ lower bound.

Flexible models for testing graph properties, by Oded Goldreich (ECCC). Usually when testing graph properties, we assume that the vertex set is $$[n]$$, implying that we can randomly sample nodes from the graph. However, this assumes that the tester knows the value of $$n$$, the number of nodes. This note suggests more “flexible” models, in which the number of nodes may be unknown, and we are only given random sampling access. While possible definitions are suggested, this note contains few results, leaving the area ripe for investigation of the power of these models.

# News for April 2018

Three Four papers for April: a new take on linearity testing, results on sublinear algorithms with advice, histogram testing, and distributed inference problems. (Edit: Sorry Clément, for missing your paper on distributed inference!)

Testing Linearity against Non-Signaling Strategies, by Alessandro Chiesa, Peter Manohar, and Igor Shinkar (ECCC). This paper gives a new model for property testing, through the notion of non-signaling strategies. The exact definitions are quite subtle, but here’s a condensed view. For $$S \subseteq \{0,1\}^n$$, let an $$S$$-partial function be one that is only defined on $$S$$. Fix a “consistency” parameter $$k$$. Think of the “input” as a collection of distributions, $$\{\mathcal{F}_S\}$$, where each $$|S| \leq k$$ and $$\mathcal{F}_S$$ is a distribution of $$S$$-partial functions. We have a local consistency requirement: $$\{\mathcal{F}_S\}$$ and $$\{\mathcal{F}_T\}$$ must agree (as distributions) on restrictions to $$S \cap T$$. In some sense, if we only view pairs of these distributions of partial functions, it appears as if they come from a single distributions of total functions. Let us focus on the classic linearity tester of Blum-Luby-Rubinfeld in this setting. We pick random $$x, y, x+y \in \{0,1\}^n$$ as before, and query these values on a function $$f \sim {\mathcal{F}_{x,y,x+y}}$$. The main question addressed is what one can say about $$\{\mathcal{F}_S\}$$, if this linearity test passes with high probability. Intuitively (but technically incorrect), the main result is that $$\{\mathcal{F}_S\}$$ is approximated by a “quasi-distribution” of linear functions.

An Exponential Separation Between MA and AM Proofs of Proximity, by Tom Gur, Yang P. Liu, and Ron D. Rothblum (ECCC). This result follows a line of work on understanding sublinear algorithms in proof systems. Think of the standard property testing setting. There is a property $$\mathcal{P}$$ of $$n$$-bit strings, an input $$x \in \{0,1\}^n$$, and a proximity parameter $$\epsilon > 0$$. We add a proof $$\Pi$$ that the tester (or the verifier) is allowed to use, and we define soundness and completeness in the usual sense of Arthur-Merlin protocols. For a $$\mathbb{MA}$$-proof of proximity, the proof $$\Pi$$ can only depend on the string $$x$$. In a $$\mathbb{AM}$$-proof of proximity, the proof can additionally depend on the random coins of the tester (which determine the indices of $$x$$ queried). Classic complexity results can be used to show that the latter subsume the former, and this paper gives a strong separation. Namely, there is a property $$\mathcal{P}$$ where any $$\mathbb{MA}$$-proof of proximity protocol (or tester) requires $$\Omega(n^{1/4})$$-queries of the input $$x$$, but there exists an $$\mathbb{AM}$$-proof of proximity protocol making $$O(\log n)$$ queries. Moreover, this property is quite natural; it is simply the encoding of permutations.

Testing Identity of Multidimensional Histograms, by Ilias Diakonikolas, Daniel M. Kane, and John Peebles (arXiv). A distribution over $$[0,1]^d$$ is a $$k$$-histogram if the domain can be partitioned into $$k$$ axis-aligned cuboids where the probability density function is constant. Recent results show that such histograms can be learned in $$k \log^{O(d)}k$$ samples (ignoring dependencies on accuracy/error parameters). Can we do any better for identity testing? This paper gives an affirmative answer. Given a known $$k$$-histogram $$p$$, one can test (in $$\ell_1$$-distance) whether an unknown $$k$$-histogram $$q$$ is equal to $$p$$ in (essentially) $$\sqrt{k} \log^{O(d)}(dk)$$ samples. There is a nearly matching lower bound, when $$k = \exp(d)$$.

Distributed Simulation and Distributed Inference, by Jayadev Acharya, Clément L. Canonne, and Himanshu Tyagi (arXiv   ECCC). This papers introduces a model of distributed simulation, which generalizes distribution testing and distributed density estimation. Consider some unknown distribution $$\mathcal{D}$$ with support $$[k]$$, and a “referee” who wishes to generate a single sample from $$\mathcal{D}$$ (alternately, she may wish to determine if $$\mathcal{D}$$ has some desired property). The referee can communicate with “players”, each of whom can generate a single independent sample from $$\mathcal{D}$$. The catch is that each player can communicate at most $$\ell$$ < $$log_2k$$ bits (otherwise, the player can simply communicate the sampled element). How many players are needed for the referee to generate a single sample? The paper first proves that this task is basically impossible with a (worst-case) finite number of players, but can be done with expected $$O(k/2^\ell)$$ players (and this is optimal). This can plugged into standard distribution testing results, to get inference results in this distributed, low-communication setting. For example, the paper shows that identity testing can be done with $$O(k^{3/2}/2^\ell)$$ players.

# News for March 2018

March has been a relatively slow month for property testing, with 3 works appearing online. (If you notice we missed some, please leave a comment pointing it out)

Edge correlations in random regular hypergraphs and applications to subgraph testing, by Alberto Espuny Díaz, Felix Joos, Daniela Kühn, and Deryk Osthus (arXiv). While testing subgraph-freness in the dense graph model is now well-understood, after a series of works culminating in a complete characterization of the testing problems which admit constant-query testers, the corresponding question for hypergraphs is far from resolved. In this paper, the authors develop new techniques for the study of study of random regular hypergraphs, which imply new testing results for subhypergraph-freeness testing, improving on the state-of-the-art for some parameter regimes (e.g., when the input graph satisfies some average-degree condition).

Back from hypergraphs to graphs, we also have:
The Subgraph Testing Model, by Oded Goldreich and Dana Ron (ECCC). Here, the authors introduce a new model for property testing of graphs, where the goal is to test if an unknown subgraph $$F$$ of an explicitly given graph $$G=(V,E)$$ satisfies the desired property. The testing algorithm is provided access to $$F$$ via membership queries, i.e., through query access to the indicator function $$\mathbf{1}_F\colon E \to \{0,1\}$$. (In some very hazy sense, this is reminiscent of the active learning or testing frameworks, where one gets more or less free access to unlabeled data but pays to see their label.) As a sample of the results obtained, the paper establishes that this new model and the bounded-degree graph model are incomparable: there exist properties easier to test in one model than the other, and vice-versa — and some properties equally easy to test in both.

And finally, to drive home the point that “models matter a lot,” we have our third paper:
Every set in P is strongly testable under a suitable encoding, by Irit Dinur, Oded Goldreich, and Tom Gur (ECCC). It is known that the choice of representation of the objects has a large impact in property testing: for instance, the complexity of testing a given property can change drastically between the dense and bounded-degree graph models. This work provides another example of such a strong dependence on the representation: while membership to some sets in $$P$$ is known to be hard to test, the authors here prove that, for every set $$S\in P$$, there exists a (polynomial-time, invertible) encoding $$E_S\colon \{0,1\}^\ast\to \{0,1\}^\ast$$ such that testing membership to $$S$$ under this encoding is easy. (They actually show even stronger a statement: namely, that under this encoding the set admits a “proximity-oblivious tester,” that is a constant-query testing algorithm which rejects with probability function of the distance to $$S$$.)

Finally, on a non-property testing note: Edith Cohen, Vitaly Feldman, Omer Reingold, and Ronitt Rubinfeld recently wrote a pledge for inclusiveness in the TCS community, available here: https://www.gopetition.com/petitions/a-pledge-for-inclusiveness-in-toc.html
If you haven’t seen it already, we encourage you to read it.

Update: Fixed a mistake in the overview of the second paper; as pointed out by Oded in the comments, the main comparison was between the new model and the bounded-degree graph model, not the dense graph one.

# News for February 2018

February had a flurry of conference deadlines, and they seem to have produced six papers for us to enjoy, including three on estimating symmetric properties of distributions.

Locally Private Hypothesis Testing, by Or Sheffet (arXiv). We now have a very mature understanding of the sample complexity of distributional identity testing — given samples from a distribution $$p$$, is it equal to, or far from, some model hypothesis $$q$$? Recently,  several papers have studied this problem under the additional constraint of differential privacy. This paper strengthens the privacy constraint to local privacy, where each sample is locally noised before being provided to the testing algorithm.

Distribution-free Junta Testing, by Xi Chen, Zhengyang Liu, Rocco A. Servedio, Ying Sheng, and Jinyu Xie (arXiv). Testing whether a function is a $$k$$-junta is very well understood — when done with respect to the uniform distribution. In particular, the adaptive complexity of this problem is $$\tilde \Theta(k)$$, while the non-adaptive complexity is $$\tilde \Theta(k^{3/2})$$. This paper studies the more challenging task of distribution-free testing, where the distance between functions is measured with respect to some unknown distribution. The authors show that, while the adaptive complexity of this problem is still polynomial (at $$\tilde O(k^2)$$), the non-adaptive complexity becomes exponential: $$2^{\Omega(k/3)}$$. In other words, there’s a qualitative gap between the adaptive and non-adaptive complexity, which does not appear when testing with respect to the uniform distribution.

The Vertex Sample Complexity of Free Energy is Polynomial, by Vishesh Jain, Frederic Koehler, and Elchanan Mossel (arXiv). This paper studies the classic question of estimating (the logarithm of) the partition function of a Markov Random Field, a highly-studied topic in theoretical computer science and statistical physics. As the title suggests, the authors show that the vertex sample complexity of this quantity is polynomial. In other words, randomly subsampling a $$\mathrm{poly}(1/\varepsilon)$$-size graph and computing its free energy gives a good approximation to the free energy of the overall graph. This is in contrast to more general graph properties, for the vertex sample complexity is super-exponential in $$1/\varepsilon$$.

Entropy Rate Estimation for Markov Chains with Large State Space, by Yanjun Han, Jiantao Jiao, Chuan-Zheng Lee, Tsachy Weissman, Yihong Wu, and Tiancheng Yu (arXiv). Entropy estimation is now quite well-understood when one observes independent samples from a discrete distribution — we can get by with a barely-sublinear sample complexity, saving a logarithmic factor compared to the support size. This paper shows that these savings can also be enjoyed in the case where we observe a sample path of observations from a Markov chain.

Local moment matching: A unified methodology for symmetric functional estimation and distribution estimation under Wasserstein distance, by Yanjun Han, Jiantao Jiao, and Tsachy Weissman (arXiv). Speaking more generally of the above problem: there has been significant work into estimating symmetric properties of distributions, i.e., those which do not change when the distribution is permuted. One natural method for estimating such properties is to estimate the sorted distribution, then apply the plug-in estimator for the quantity of interest. The authors give an improved estimator for the sorted distribution, improving on the results of Valiant and Valiant.

INSPECTRE: Privately Estimating the Unseen, by Jayadev Acharya, Gautam Kamath, Ziteng Sun, and Huanyu Zhang (arXiv). One final work in this area — this paper studies the estimation of symmetric distribution properties (including entropy, support size, and support coverage), but this time while maintaining differentially privacy of the sample. By using estimators for these tasks with low sensitivity, one can additionally obtain privacy for a low or no additional cost over the non-private sample complexity.