The first month of 2020 is behind us, and it’s already looking good! Four papers last month, spanning quantum, graphs, and probability distributions.

On Efficient Distance Approximation for Graph Properties, by Nimrod Fiat and Dana Ron (arXiv). In the dense graph (i.e., adjacency-matrix) model, one is given a distance parameter \(\varepsilon\) and granted query access to the adjacency matrix of a graph \(G\), and seeks to determine something about the distance of \(G\) to a prespecified property \(\mathcal{P}\) of interest (i.e., the fraction of the matrix that needs to be changed for \(G\) to satisfy the property). Testing requires to distinguish whether that distance is zero, or at least \(\varepsilon\); many results over the past years have shown that many properties can be tested with a number of queries depending only on \(\varepsilon\) (and not on \(n=|G|\). This work focuses on the harder problem of distance estimation, or, equivalently, tolerant testing: that is, estimate this distance up to \(\pm\varepsilon\). The authors introduce a general framework to get distance approximation algorithms from a “complicated” property by decomposing it into simpler properties and using algorithms for those. Applying this framework to a few flagship properties, they then show that \(P_3\)-freeness, induced \(P_4\)-freeness, and cordiality are properties which have efficient distance estimation algorithms (independent of \(n\), and polynomial in \(1/\varepsilon\)).

Minimax Optimal Conditional Independence Testing, by Matey Neykov, Sivaraman Balakrishnan, and Larry Wasserman (arXiv). Given i.i.d. draws from a triple \((X,Y,Z)\), how hard is it to check whether \(X \perp Y \mid Z\), that is, “\(X\) and \(Y\) are independent conditioned on \(Z\) (or far from it)?” This is the problem on conditional independence testing, which was covered back in the days for the case where \(X,Y,Z\) are discrete. Well, this new work takes the fight out of the discrete world: extending the results and techniques from the discrete case, it provides optimal bound for the continuous case: where \(Z\) is on \([0,1]\), and then when all three \(X,Y, Z\) are continuous.

How symmetric is too symmetric for large quantum speedups?, by Shalev Ben-David and Supartha Podder (arXiv); and Can graph properties have exponential quantum speedup?, by Andrew M. Childs and Daochen Wang (arXiv). Both these works independently study the relation between the (bounded-error) randomized and quantum query complexities of any graph property \(f\), in the dense graph (adjacency-matrix) model. In particular, how much advantage do quantum algorithms provide for those?
As it turns out, not so much: for those functions, both papers show the two quantities are always polynomially related (\(R(f) \leq Q(f) \leq O(R(f)^6))\)) in the dense-graph model. As a corollary, this implies that testing any such property won’t benefit much from quantum (that is, at most polynomially)…. at least in this model. In the adjacency list model (also known as the bounded-degree graph model), the first paper conjectures that exponential query complexity improvements are possible; and the second paper provides an example, establishing it. Altogether, this fully settles an open problem of Ambainis, Childs, and Liu, and Montanaro and de Wolf.

Happy new year! And now for the last post of 2019 papers. We have found a diverse collection of six papers, ranging from classic property testing topics to new perspectives on sublinear computation.

Sublinear Optimal Policy Value Estimation in Contextual Bandits by Weihao Kong, Gregory Valiant, Emma Brunskill (arXiv). This isn’t our usual sublinear paper, but it is definitely of interest to us sublinear folk. Let’s start with a stripped down definition of the problem (or rather, game). There are \(K\) “arms”, where the \(i\)th arm is represented by an unknown vector in \(\beta_i \in \mathbb{R}^d\). We are presented with a “context”, which is a vector \(x \in \mathbb{R}^d\). Our job is to choose an arm \(i \in [K]\). We get the reward \(x \cdot \beta_i\) (with some noise added). The contexts appears from a known distribution. To aid us, we observe the rewards of \(N\) iid contexts, so we observe a total of \(T = KN\) rewards. There has been much work on figuring out the minimum value of \(T\) required to learn the optimal policy. One requires at least \(d\) (the dimension) samples to estimate any of the arm vectors. This papers shows that one can actually estimate the expected reward of the optimal policy, without being able to describe it, with sublinear in \(d\) (technically, \(\widetilde{O}(\sqrt{d})\)) samples. We see this a lot in property testing, where producing the “optimal” solution for a problem requires linear-in-dimension samples, but estimating the optimal value is much cheaper (consider, for example, the situation of linearity testing, where we wish to find the closest linear function).

Sublinear Time Numerical Linear Algebra for Structured Matrices by Xiaofei Shi and David P. Woodruff (arXiv). This follows the recent linear of advances in sublinear time linear algebra. Given a matrix \(A \in \mathbb{R}^{n \times d}\), the aim is to get algorithms that only look at \(o(nnz(A))\) entries (where \(nnz(A)\) is the number of non-zeroes, or the support). Consider the classic talk of low rank approximation. Unfortunately, suppose one entry is extremely large, and the others are extremely small. One has to find this large entry for any reasonable approximation, which (in the worst-case) requires \(nnz(A)\) queries into \(A\). Thus, previous papers make structural assumption (such as, \(A\) being a Vandermonde matrix) to get sublinear bounds. This paper gives a clean black box method to get a variety of such results. Basically, one can replace the usual \(nnz(A)\) term in many algorithms, by \(T(A)\), which is the time to compute the matrix-vector product \( Ay\), for \(y \in \mathbb{R}^d\). In many cases \(T(A) = \widetilde{O}(n)\), which can be significantly smaller than \(nnz(A)\). This paper gives such results for low-rank approximations and many regression problems.

Robust and Sample Optimal Algorithms for PSD Low-Rank Approximation by Ainesh Bakshi, Nadiia Chepurko, and David P. Woodruff. Consider the low rank problem discussed above. As mentioned in the previous paragraph, we need structural assumptions on \(A\). Previous results gave sublinear time low-rank approximations assuming that \(A\) is positive semidefinite (PSD). The aim is to get a rank \(k\) matrix \(B\) such that \(\|A-B\|^2_2\) is at most \((1+\epsilon)\)-times the optimal such approximation. The previous algorithm of Musco-Woodruff makes \(\widetilde{O}(nk/\epsilon^{2.5})\) queries in to \(A\), while there is a lower bound of \(\Omega(nk/\epsilon)\). This gap between the complexities is resolved in this paper with an upper bound of \(\widetilde{O}(nk/\epsilon)\) queries.

Constructive derandomization of query algorithms by Guy Blanc, Jane Lange, and Li-Yang Tan (arXiv). This paper discusses an intriguing angle to sublinear question: when can they be derandomized? Abstractly, consider a randomized algorithm \(R\) that makes \(q\) queries. Think of \(R\) as a function \(R(x,r)\), where \(x\) is the input, and \(r\) is the randomness. We would like to design a deterministic algorithm \(D\) making, ideally, close to \(q\) queries and approximates \(\mathop{E}_r[R(x,r)]\). For starters, consider some distribution over \(x\), and suppose we want \(\mathbb{E}_x[D(x) – \mathbb{E}_r[R(x,r)]] < \epsilon\). By (the easy direction of) Yao’s minimax lemma, one can show the existence of such an algorithm \(D\) that makes \(O(q/\epsilon)\) queries. But how to explicitly construct it? Indeed, the first result of this paper gives a “meta-algorithm” that takes as input the description of \(R\) (which is of size \(N\)), has running time \(poly(N)2^{O(q/\epsilon)}\) and outputs a description of \(D\). When \(R\) satisfies the stronger property of “bounded error”, one can get a \(O(q^3)\)-query algorithm \(D\) that approximates \(\mathop{E}_r[R(x,r)]\) for all \(x\) (again, the existence is proven by a classic theorem of Nisan). Overall, this paper gives a method to derandomize sublinear time algorithms, and I wonder if there could be some applications of this method for proving lower bounds. After all, Yao’s minimax theorem is the tool for property testing lower bounds, and any perspective on Yao’s theorem is likely of relevance.

Testing Membership for Timed Automata by Richard Lassaigne and Michel de Rougemont (arXiv). Property testing for regular languages is a classic result in sublinear algorithms. This paper focuses on the more complex notion of timed automata. The technical definition is quite complicated, but here’s an overview. There is a finite automaton and a collection of “clocks”. Imagine a string being processed, where each alphabet symbol appears with a new timestamp. Thus, the input word is called a “timed word”. The transitions of the automaton involve the new symbol read, as well as constraints involving the clock times and the timestamp. Thus, we can enforce conditions like “only transition if another symbol is read within a single time unit”. In general, deciding whether a timed word is accepted by a timed automaton is NP-complete. This papers studies the property testing viewpoint. The paper gives a new definition of “timed edit distance” between timed words. The main result shows that one can distinguish time words accepted by a timed automaton from words that are far (according to timed edit distance), by querying a constant number of word positions.

On the query complexity of estimating the distance to hereditary graph properties by Carlos Hoppen, Yoshiharu Kohayakawa, Richard Lang, Hanno Lefmann, and Henrique Stagni (arXiv). This paper concerns the classic setting of property testing of dense graphs. It is well-known that all hereditary graph properties are testable, and are moreover, one can estimate the distance to the property in time that only depends on \(\varepsilon\). Unfortunately, the queries complexities have large tower dependencies on \(\varepsilon\), arising from the use of the Szemeredi regularity lemma. The question of property testing in dense graphs can be reduced to finding “removal” lemmas (such as the classic triangle remove lemma). Such a lemma states that if at least \(\varepsilon n^2\) edges need to be removed from \(G\) to destroy all “forbidden subgraphs”, then there must be “many” forbidden subgraphs in \(G\). There is much recent research on finding families of forbidden subgraphs, where the “many” (in the above statement) is at least \(poly(\varepsilon)\) times the trivial upper bound. This paper shows that one can also estimate the distance to any hereditary property, in a query complexity that depends directly on the corresponding removal lemma parameters. As a compelling example, one can estimate the distance to being chordal in \(\exp(1/\varepsilon)\) queries, a significant improvement over standard tower bounds.

We hit the mother-lode of property testing papers this month. Stick with us, as we cover 10 (!) papers that appeared online in November.

EDIT: We actually have 11 papers, check out Optimal Adaptive Detection of Monotone Patterns at the bottom.

Testing noisy linear functions for sparsity, by Xue Chen, Anindya De, and Rocco A. Servedio (arXiv). Given samples from a noisy linear model \(y = w\cdot x + \mathrm{noise}\), test whether \(w\) is \(k\)-sparse, or far from being \(k\)-sparse. This is a property testing version of the celebrated sparse recovery problem, whose sample complexity is well-known to be \(O(k\log n)\), where the data lies in \(\mathbb{R}^n\). This paper shows that the testing version of the problem can be solved (tolerantly) with a number of samples independent of \(n\), assuming technical conditions: the distribution of coordinates of \(x\) are i.i.d. and non-Gaussian, and the noise distribution is known to the algorithm. Surprisingly, all these conditions are needed, otherwise the dependence on \(n\) is \(\tilde \Omega(\log n)\), essentially the same as the recovery problem.

Pan-Private Uniformity Testing, by Kareem Amin, Matthew Joseph, Jieming Mao (arXiv). Differentially private distribution testing has now seen significant study, in both the local and central models of privacy. This paper studies a distribution testing in the pan-private model, which is intermediate: the algorithm receives samples one by one in the clear, but it must maintain a differentially private internal state at all time steps. The sample complexity turns out to be qualitatively intermediate to the two other models: testing uniformity over \([k]\) requires \(\Theta(\sqrt{k})\) samples in the central model, \(\Theta(k)\) samples in the local model, and this paper shows that \(\Theta(k^{2/3})\) samples are necessary and sufficient in the pan-private model.

Almost Optimal Testers for Concise Representations, by Nader Bshouty (ECCC). This work gives a unified approach for testing for a plethora of different classes which possess some sort of sparsity. These classes include \(k\)-juntas, \(k\)-linear functions, \(k\)-terms, various types of DNFs, decision lists, functions with bounded Fourier degree, and much more.

Unified Sample-Optimal Property Estimation in Near-Linear Time, by Yi Hao and Alon Orlitsky (arXiv). This paper presents a unified approach for estimating several distribution properties with both near-optimal time and sample complexity, based on piecewise-polynomial approximation. Some applications include estimators for Shannon entropy, power sums, distance to uniformity, normalized support size, and normalized support coverage. More generally, results hold for all Lipschitz properties, and consequences include high-confidence property estimation (outperforming the “median trick”) and differentially private property estimation.

Testing linear-invariant properties, by Jonathan Tidor and Yufei Zhao (arXiv). This paper studies property testing of functions which are in a formal sense, definable by restrictions to subspaces of bounded degree. This class of functions is a broad generalization of testing whether a function is linear, or a degree-\(d\) polynomial (for constant \(d\)). The algorithm is the oblivious one, which simply repeatedly takes random restrictions and tests whether the property is satisfied or not (similar to the classic linearity test of BLR, along with many others).

Approximating the Distance to Monotonicity of Boolean Functions, by Ramesh Krishnan S. Pallavoor, Sofya Raskhodnikova, Erik Waingarten (ECCC). This paper studies the following fundamental question in tolerant testing: given a Boolean function on the hypercube, test whether it is \(\varepsilon’\)-close or \(\varepsilon\)-far from monotone. It is shown that there is a non-adaptive polynomial query algorithm which can solve this problem for \(\varepsilon’ = \varepsilon/\tilde \Theta(\sqrt{n})\), implying an algorithm which can approximate distance to monotonicity up to a multiplicative \(\tilde O(\sqrt{n})\) (addressing an open problem by Sesh). They also give a lower bound demonstrating that improving this approximating factor significantly would necessitate exponentially-many queries. Interestingly, this is proved for the (easier) erasure-resilient model, and also implies lower bounds for tolerant testing of unateness and juntas.

Testing Properties of Multiple Distributions with Few Samples, by Maryam Aliakbarpour and Sandeep Silwal (arXiv). This paper introduces a new model for distribution testing. Generally, we are given \(n\) samples from a distribution which is either (say) uniform or far from uniform, and we wish to test which is the case. The authors here study the problem where we are given a single sample from \(n\) different distributions which are either all uniform or far from uniform, and we wish to test which is the case. By additionally assuming a structural condition in the latter case (it is argued that some structural condition is necessary), they give sample-optimal algorithms for testing uniformity, identity, and closeness.

Random Restrictions of High-Dimensional Distributions and Uniformity Testing with Subcube Conditioning, by Clément L. Canonne, Xi Chen, Gautam Kamath, Amit Levi, and Erik Waingarten (ECCC, arXiv). By now, it is well-known that testing uniformity over the \(n\)-dimensional hypercube requires \(\Omega(2^{n/2})\) samples — the curse of dimensionality quickly makes this problem intractable. One option is to assume that the distribution is product, which causes the complexity to drop to \(O(\sqrt{n})\). This paper instead assumes one has stronger access to the distribution — namely, one can receive samples conditioned on being from some subcube of the domain. With this, the paper shows that the complexity drops to the near-optimal \(\tilde O(\sqrt{n})\) samples. The related problem of testing whether a distribution is either uniform or has large mean is also considered.

Property Testing of LP-Type Problems, by Rogers Epstein, Sandeep Silwal (arXiv). An LP-Type problem (also known as a generalized linear program) is an optimization problem sharing some properties with linear programs. More formally, they consist of a set of constraints \(S\) and a function \(\varphi\) which maps subsets of \(S\) to some totally ordered set, such that \(\varphi\) possesses monotonicity and locality properties. This paper considers the problem of testing whether \(\varphi(S) \leq k\), or whether at least an \(\varepsilon\)-fraction of constraints in \(S\) must be removed for \(\varphi(S) \leq k\) to hold. This paper gives an algorithm with query complexity \(O(\delta/\varepsilon)\), where \(\delta\) is a dimension measure of the problem. This is applied to testing problems for linear separability, smallest enclosing ball, smallest intersecting ball, smallest volume annulus. The authors also provide lower bounds for some of these problems as well.

Near-Optimal Algorithm for Distribution-Free Junta Testing, by Xiaojin Zhang (arXiv). This paper presents an (adaptive) algorithm for testing juntas, in the distribution-free model with one-sided error. The query complexity is \(\tilde O(k/\varepsilon)\), which is nearly optimal. Algorithms with this sample complexity were previously known under the uniform distribution, or with two-sided error, but this is the first paper to achieve it in the distribution-free model with one-sided error.

Optimal Adaptive Detection of Monotone Patterns, by Omri Ben-Eliezer, Shoham Letzter, Erik Waingarten (arXiv). Consider the problem of testing whether a function has no monotone increasing subsequences of length \(k\), versus being \(\varepsilon\)-far from having this property. Note that this is a generalization of testing whether a function is monotone (decreasing), which corresponds to the case \(k = 2\). This work shows that the adaptive sample complexity of this problem is \(O_{k,\varepsilon}(\log n)\), matching the lower bound for monotonicity testing. This is in comparison to the non-adaptive sample complexity, which is \(O_{k,\varepsilon}((\log n)^{\lfloor \log_2 k\rfloor})\). In fact, the main result provides a certificate of being far, in the form of a monotone increasing subsequence of length \(k\).

Last month was a lean month, with only three papers: one on direct product testing, one on finding forbidden patterns in a sequence, and one (an update of a paper which we had missed in the Spring) on quantum distribution testing.

Direct sum testing – the general case, by Irit Dinur and Konstantin Golubev (ECCC). Say a function \(f\colon \prod_{i=1}^d [n_i] \to \mathbb{F}_2\) is a direct product if it can be factored as \(f(x_1,\dots,x_d)=\sum_{i=1}^d f_i(x_i)\), where \(f_i\colon [n_i]\to\mathbb{F}_2\).
This paper provides a 4-query tester (i.e., a proximity oblivious tester (POT)) for the direct product property, reminiscent of (and relying on) the BLR linearity test: specifically, draw two subsets \(S,T\subseteq [d]\) and two inputs \(x,y\in \prod_{i=1}^d [n_i]\) u.a.r., and accept iff
\(f(x)+f(x_Sy)+f(x_Ty)+f(x_{S\Delta T}y) = 0\,.\)
The main theorem of the paper is to show that the probability that this simple test rejects is lower bounded (up to a constant factor) by the distance of \(f\) to direct-product-ness. (The authors also provide a different POT making \(d+1\) queries, but with a simpler analysis.)

Finding monotone patterns in sublinear time, by Omri Ben-Eliezer, Clément Canonne, Shoham Letzter, and Erik Waingarten (ECCC). Given a function \(f\colon [n]\to\mathbb{R}\), a monotone subsequence of size \(k\) is a \(k\)-tuple of indices \(i_1 < \dots <i_k\) such that \(f(i_j) < f(i_{j+1})\) for all \(j\). This work considers (non-adaptive) one-sided testing of monotone-subsequence-freeness, or, equivalently, the task of finding such a monotone subsequence in a function promised to contain many of them. (This, in particular, generalizes the problem of one-sided monotonicity testing, which is the case \(k=2\).) The main result is a full characterization of the query complexity of this question (for constant \(k\)): strange as the exponent may seem, \(\Theta_\varepsilon( (\log n)^{\lfloor \log_2 k\rfloor} )\) queries are necessary and sufficient. The proof relies on a structural dichotomy result, stating that any far-from-free sequence either contains “easy to find” increasing subsequences with increasing gaps between the elements, or has a specific hierarchical structure.

Quantum Closeness Testing: A Streaming Algorithm and Applications, by Nengkun Yu (arXiv). This paper is concerned with quantum distribution testing in the local model, which only allows a very restricted (albeit, as the author argues, more natural and easier to implement) type of measurements, and is particularly well-suited to a streaming setting. The main contribution of this paper is to show a connection to classical distribution testing, allowing one to obtain quantum distribution testing upper bounds from their classical distribution testing counterparts. In more detail, the paper shows that, from local measurements to two \(d\)-dimensional quantum states \(\rho,\sigma\), one can provide access to two classical distributions \(p,q\) on \(\approx d^2\) elements such that (i) \(\| p-q\|_2 \approx \|\rho-\sigma\|_2/d\) and (ii) \(\| p\|_2,\| q\|_2 = O(1/d)\).
Using this connection, the paper proceeds to establish a variety of upper bounds for testing several distribution properties in the local quantum model.

Five Six papers this month: results on testing separations, linearity testing in \(\mathbb{R}^n\), testing for regular languages, graph property testing, topological property testing, and Boolean rank.

Hard properties with (very) short PCPPs and their applications, by Omri Ben-Eliezer, Eldar Fischer, Amit Levi, and Ron D. Rothblum (arXiv). Probably, the most significant takeaway from this work is a (largest possible) separation between standard and tolerant property testing. PCPPs (Probabilistically Checkable Proofs of Proximity) are the “NP” variant of property testing, where the tester is aided by a proof string. Consider property \(\mathcal{P}\). If \(x \in \mathcal{P}\), there must be a proof string that makes the tester accept (with probability 1). If \(x\) is far from \(\mathcal{P}\) (in the usual property testing sense), for any proof string, the tester must reject with sufficiently high probability. PCPPs have played a role in the classical constructions of PCPs, but have also found uses in getting a better understanding of property testing itself. And this paper shows how PCPP constructions can be used to get property testing separations. The main result in this paper is a property \(\mathcal{P}\) that (basically) requires \(\Omega(n)\) queries to “property test”, but has a PCPP system where the proof length is \(\widetilde{O}(n)\). (\(n\) is the input length.) The main construction uses collections of random linear codes. Significantly, these constructions show a strong separation between standard vs tolerant property testing, and standard vs erasure-resilient property testing. (The latter is a recent variant by Dixit et al, where certain parts of the input are hidden from the tester.) There is a property that is testable in a constant number of queries, but requires \(\widetilde{\Omega}(n)\) queries to test tolerantly (for any non-trivial choice of parameters). An analogous result holds for erasure-resilient testing.

Distribution-Free Testing of Linear Functions on R^n, by Noah Fleming and Yuichi Yoshida (arXiv). Linearity testing is arguably the canonical problem in property testing, yet there is still much to be learned about it. This paper considers functions \(f: \mathbb{R}^n \to \mathbb{R}\), and the distribution-free setting. (In this setting, distance is measured according is an unknown distribution \(\mathcal{D}\) over the input, and the tester can access samples from this distribution. For \(\mathbb{R}^n\), the “standard” distribution would the \(n\)-dimensional Gaussian.) The main result is that linearity testing can be done in the distribution-free setting with \(\widetilde{O}(1/\varepsilon)\) queries, assuming that the distribution is continuous. The primary technical tool, an interesting result in its own right, is that additivity \((f(x+y) = f(x) + f(y))\) can be tested in \(\widetilde{O}(1/\varepsilon)\) queries. The significance of the testing result is cemented by an \(\Omega(n)\) lower bound for sample-based testers.

Sliding window property testing for regular languages by Moses Ganardi, Danny Hucke, Markus Lohrey, Tatiana Starikovskaya (arXiv). Fix a regular language \(\mathcal{R}\). Consider the streaming model, and the basic question of recognizing whether the string (being streamed) is in \(\mathcal{R}\). Simple, you will say! Run the DFA recognizing \(\mathcal{R}\) in constant space. Now, suppose there is a sliding window length of \(n\). The aim is to determine if the past \(n\) symbols (the “active window”) form a string in \(\mathcal{R}\). Suprisingly (at least to me), there is a full characterization of the space required for randomized algorithms, and (depending on \(\mathcal{R}\)), it is either \(\Theta(1)\), \(\Theta(\log\log n)\), \(\Theta(\log n)\), or \(\Theta(n)\). In the interest of beating these lower bounds, suppose we wish to property test on the active window. It turns out the answer is quite nuanced. There are deterministic \(O(\log n)\)-space testers and randomized two-sided \(O(1/\varepsilon)\)-space testers for all regular languages. For randomized one-sided testers, there are multiple possibilities for the optimal space complexity, and there is a full characterization of these regular languages.

A characterization of graph properties testable for general planar graphs with one-sided error (It’s all about forbidden subgraphs) by Artur Czumaj and Christian Sohler (arXiv). Property testing of sparse graphs has been receiving more attention, but most results focus on the bounded degree setting. Unfortunately, many of these results break quite dramatically on sparse graphs with unbounded degrees. This paper focuses on property testing, within the class of unbounded degree planar graphs. (Meaning, the input is always assumed to be planar.) The results achieve a significant goal: as the title suggests, there is a complete characterization of properties that are constant-query testable with one-sided error. The easier part is in showing that all such properties can be reduced to testing \(H\)-freeness. The harder (remarkable) result is \(H\)-freeness can be tested in general planar queries with constant queries. (This is non-trivial even for triangle-freeness.) And, as is easy to conjecture but hard to prove, these results carry over for all minor-closed families. As a small indication of the challenge, most testers for bounded-degree graphs work by doing constant depth BFSes. When high degree vertices are present, this method fails, and we really need new ideas to deal with such graphs.

Near Coverings and Cosystolic Expansion – an example of topological property testing by Irit Dinur and Roy Meshulam (ECCC). In most algebraic settings, property testing results can be seen as local to global theorems. When do local constraints on a large object imply a global condition? This paper gives a topological instantiation of this phenomenon. We need to define the cover of a simplicial complex \(X\). For concreteness, think of a 2D simplicial complex \(X\), which is a hypergraph with hyperedges of size at most 3, where subsets of hyperedges are also present. A 2-cover is a simplicial complex \(X’\) with the following property. It has two copies of each vertex of \(X\). Each hyperedge of \(X\) must have two “corresponding” disjoint copies in \(X’\). Let the copies of vertex \(v\) be \(v_0, v_1\). Then, for every hyperedge (say) \((u,v,w)\) of \(X\), there must be two disjoint hyperedges in \(X’\) involving copies of the corresponding vertices. One can consider the property testing twist: if the neighborhoods of “most” vertices \(v\) in \(X\) satisfy these condition (with respect to the neighborhoods of the copies of \(v\) in \(X’\)), then is \(X’\) close to being a cover of \(X\)? Indeed, this paper proves that such a “property testing condition” holds iff \(X\) is a high-dimensional expander.

Property testing of the Boolean and binary rank by Michal Parnas, Dana Ron, and Adi Shraibman (arXiv). The Boolean rank of a matrix \(M\) is a fundamental quantity that appears in many lower bound constructions. (Recall that an \(n \times m\) Boolean matrix \(M\) has a rank \(r\) if \(M\) can be expressed as \(X \cdot Y\), where \(X \in \mathbb{F}_2^{n \times d}\) and \(Y \in \mathbb{F}_2^{d \times m}\).) In the real-valued setting, results show that one can property test rank in \(poly(d/\varepsilon)\) queries. This paper proves an analogous result for the Boolean rank. There is a surprise element here: over reals, the rank can be computed in polynomial time, and many of the geometric intuitions can be brought over to the property testing problem. On the other hand, the Boolean rank is NP-hard to compute exactly, yet we can still get a tester with \(poly(d)\) query complexity. The paper also gives results for binary rank. For the binary rank, we require the component matrices \(X, Y\) to be Boolean, but algebraic operations are over the reals. In the case, the tester has query complexity \(2^{2d}\) (with varying dependencies on \(\varepsilon\) for adaptive/non-adaptive testers). The intriguing open problem is whether \(poly(d)\)-query testers exist for binary rank.

A comparatively slow month, as summer draws to a close: we found three papers online. Please let us know if we missed any! (Edit: And we added two papers missed from June.)

Testing convexity of functions over finite domains, by Aleksandrs Belovs, Eric Blais, and Abhinav Bommireddi (arXiv). This paper studies the classic problem of convexity testing, and proves a number of interesting results on the adaptive and non-adaptive complexity of this problem in single- and multi-dimensional settings. In the single-dimensional setting on domain \([n]\), they show that adaptivity doesn’t help: the complexity will be \(O(\log n)\) in both cases. However, in the simplest two-dimensional setting, a domain of \([3] \times [n]\), they give a polylogarithmic upper bound in the adaptive setting, but a polynomial lower bound in the non-adaptive setting, showing a strong separation. Finally, they provide a lower bound for \([n]^d\) which scales exponentially in the dimension. This leaves open the tantalizing open question: is it possible to avoid the curse of dimensionality when testing convexity?

Testing Isomorphism in the Bounded-Degree Graph Model, by Oded Goldreich (ECCC). This work investigates the problem of testing isomorphism of graphs, focusing on the special case when the connected components are only polylogarithmically large (the general bounded-degree case is left open). One can consider when a graph is given as input, and we have to query a graph to test if they are isomorphic. This can be shown to be equivalent (up to polylogarithmic factors) to testing (from queries) whether a sequence is a permutation of a reference sequence. In turn, this can be shown to be equivalent to the classic distribution testing question of testing (from samples) whether a distribution is equal to some reference distribution. The same sequence of equivalences almost works for the case where there is no reference graph/sequence/distribution, but we only have query/query/sample access to the object. The one exception is that the reduction doesn’t work to reduce from testing distributions to testing whether a sequence is a permutation, due to challenges involving sampling with and without replacement. However, the author still shows the lower bound which would be implied by such a reduction by adapting Valiant’s proof for the distribution testing problem to this case.

Learning Very Large Graphs with Unknown Vertex Distributions, by Gábor Elek (arXiv). In this note, the author studies a variant of distribution-free property testing on graphs, in which (roughly) neighboring vertices have probabilities of bounded ratio, and a query reveals this ratio. Applications to local graph algorithms and connections to dynamical systems are also discussed.

EDIT: We apparently missed two papers from June — the first paper was accepted to NeurIPS 2019, the second to COLT 2019.
The Broad Optimality of Profile Maximum Likelihood, by Yi Hao and Alon Orlitsky (arXiv). Recently, Acharya, Das, Orlitsky, and Suresh (ICML 2017) showed that the Profile Maximum Likelihood (PML) estimator enables a unified framework for estimating a number of distribution properties, including support size, support coverage, entropy, and distance to uniformity, obtaining estimates which are competitive with the best possible. The approach is rather clean: simply estimate the PML of the distribution (i.e., the maximum likelihood distribution of the data, if the the labels are discarded and only the multiplicities of elements are kept), and apply the plug-in estimator (i.e., if you want to estimate entropy, compute the entropy of the resulting PML distribution). The present work shows that PML is even more broadly applicable — such an approach applies to any property which is additive, symmetric, and appropriately Lipschitz. They also show specific results for many other properties which have been considered in the past, including Rényi entropy, distribution estimation, and identity testing.

Sample-Optimal Low-Rank Approximation of Distance Matrices by Piotr Indyk, Ali Vakilian, Tal Wagner, David Woodruff (arXiv). Getting a rank \(k\) approximation of an \(n \times m\) matrix \(M\) is about as classic a problem as it gets. Suppose we wanted a running time of \(O(n+m)\), which is sublinear in the matrix size. In general, this is not feasible, since there could be a single large entry that dominates the matrix norm. This paper studies the case where the matrix is itself a distance matrix. So there is an underlying point set in a metric space, and the \(i, j\)th entry of \(M\) is the distance between the $i$th and $j$th point. Previous work showed the existence of \(O((n+m)^{1+\gamma})\) time algorithms (for arbitrary small constant $\gamma > 0$, with polynomial dependence on \(k\) and error parameters). This work gives an algorithm that runs in \(\widetilde{O}(n+m)\) time. The main idea is to sample the rows and columns according to row/column norms.

We saw quite an activity last month, with a total of 9 papers on property testing or related areas.

Let’s start with a paper on PCPs:
Revisiting Alphabet Reduction in Dinur’s PCP, by Venkatesan Guruswami, Jakub Opršal, and Sai Sandeep (ECCC). As mentioned in the title, this work provides an alternate proof of one of the two parts of the Dinur’s PCP theorem proof. In that proof, the argument goes by alternating two main steps: (i) amplifying the soundness gap, at the price of increasing the alphabet size (and the instance size); (ii) reducing the alphabet size, at the price of increasing the instance size. From the observation that step (i) creates some structure in the instances that can be leveraged, the authors provide an alternative construction for (ii), leading to an arguably simpler analysis of the soundness of the underlying test. A nice throwback to some of the roots of property testing.

Then, four papers on distribution testing:
Testing Mixtures of Discrete Distributions, by Maryam Aliakbarpour, Ravi Kumar, Ronitt Rubinfeld (arXiv). One of the most studied problems in distribution testing is identity testing (a.k.a., goodness-of-fit): given a reference distribution \(q\) over a domain of size \(n\) and i.i.d. samples from an unknown \(p\), decide between \(p=q\) and \(d_\textrm{TV}(p,q) > \varepsilon\). While this is now well-understood and can be done with a strongly sublinear number of samples (namely, \(\sqrt{n})\)), the case of tolerant testing is much sadder: Valiant and Valiant proved that \(\Theta(n/\log n)\) samples were required for distinguishing \(d_\textrm{TV}(p,q) < \varepsilon/2\) from \(d_\textrm{TV}(p,q) > \varepsilon\). So, noise-tolerance is almost as hard as it gets… except, as this works suggests and studies, if one has some guarantees on the noise! What if the noise in the completeness case was just that \(q\) is perturbed by some (known, or available via samples as well) noise coming from a second distribution \(\eta\)? I.e., the question is now to test whether \(p\) if of the form \(\alpha q + (1-\alpha) \eta\) for some \(\alpha\in [0,1]\), or if it is \(\varepsilon\)-far from all mixtures of \(q\) and \(\eta\). The authors have various results on this question and some extensions, but the upshot is: “with structured noise, strongly sublinear sample complexity is once again possible.”

Can Distributed Uniformity Testing Be Local?, by Uri Meir, Dor Mintzer, and Rotem Oshman (ACM PODC Proceedings). More on identity testing, specifically uniformity testing. No noise here, but another constraint: the samples are distributed. There are \(k\) players, each holding \(q\) i.i.d. samples from a distribution \(p\) over \([n]\); each can send \(\ell=1\) bit to a central referee, in a non-interactive fashion. Is \(p\) the uniform distribution, or is it \(\varepsilon\)-far from it? The authors here consider the case where \(k,n,\varepsilon\) are fixed, and prove lower bounds on the number \(q=q(k,n,\varepsilon)\) of samples each player must hold in order to perform this uniformity testing task. (Note: their lower bounds hold even in the public-randomness setting.) Motivated by the LOCAL model, they focus on 3 cases of interest, for which they obtain tight or near-tight bounds on \(q\) (in view of upper bounds from previous work of a subset of the authors): (1) when the referee’s decision has to be the \(\textsf{AND}\) of all \(n\) bits she receives; (2) when the referee can do something more involved, and use a threshold function; and (3) when the referee can use an arbitrary function of those \(n\) messages. Underlying those lower bounds is a neat application of Boolean Fourier analysis, recasting the analysis of the standard “Paninski” lower bound instance and the player’s decision functions as a question on Boolean functions.

Domain Compression and its Application to Randomness-Optimal Distributed Goodness-of-Fit, by Jayadev Acharya, Clément Canonne, Yanjun Han, Ziteng Sun, and Himanshu Tyagi (arXiv,ECCC). Remember the paper just above? Now, focus on the case (3), where the referee can use any function of the message, allow each player to send \(\ell\) bits instead of one, but give only one sample (instead of \(q\)) to each player. This becomes a setting we have talked about before on this blog (specifically, with three papers on these three months). Those three papers established the optimal number of player \(k\), in terms of the domain size \(n\), the distance parameter \(\varepsilon\), and the number of bits per player \(\ell\), in both the private-coin and public-coin settings. This new work interpolates between those two extremes, and gives the tight answer to the general question: if there are \( 0\leq s \leq \log n\) bits of public randomness available, what is the optimal number of players to perform identity testing? Behind the upper bounds is a new notion of (randomness-efficient) domain compression they introduce: how to reduce the domain size of the probability distributions while preserving the \(\ell_1\) distances between them as much as possible?

Towards Testing Monotonicity of Distributions Over General Posets, by Maryam Aliakbarpour, Themis Gouleakis, John Peebles, Ronitt Rubinfeld, and Anak Yodpinyanee (arXiv). Away from identity testing, let’s now consider another central question, testing monotonicity of distributions. Specifically, a distribution \(p\) over a given poset \((\mathcal{X}, \prec)\) is said to be monotone if \(p(x) \leq p(y)\) whenever \(x\prec y\). While the question is fully understood for the familiar case of the line \(\mathcal{X} = \{1,2,\dots,n\}\), the case of general posets is much murkier, and more or less uncharted. This paper initiates a general study of the question, improving on previously known bounds for the matching poset and Boolean hypercube, and providing several new results and techniques for the general case, to establish both upper and lower bounds on the sample complexity.

And now… graphs!
Expansion Testing using Quantum Fast-Forwarding and Seed Sets, by Simon Apers (arXiv). This paper is concerned with (bicriteria) testing of vertex expansion of a given graph \(G=(V,E)\) in the bounded-degree graph model. Loosely speaking, given a value \(\Phi\) and query access (i.e., sampling a node u.a.r., querying the degree of a node, and querying the \(i\)-th neighbor of a node) to a graph \(G\) with maximum degree \(d=\Theta(1)\), the goal is to distinguish between (i) \(G\) has vertex expansion at most \(\Phi\) and (ii) \(G\) is \(\varepsilon\)-far from any \(G’\) with vertex expansion at most \(\Phi^2\) (simplifying and dropping some technicalities). The previous state-of-the-art was a \(\tilde{O}_{\varepsilon}(n^{1/2}/\Phi^2)\) query complexity for classical algorithms, and a \(\tilde{O}_{\varepsilon}(\min(n^{1/3}/\Phi^2),n^{1/2}/\Phi^2))\) query complexity upper bound for quantum testers. The current work improves on this by providing a \(\tilde{O}_{\varepsilon}(n^{1/3}/\Phi)\) quantum tester. Graphs expand—and so does the gap between classical and quantum testing.

Walking Randomly, Massively, and Efficiently, by Jakub Łącki, Slobodan Mitrović, Krzysztof Onak, and Piotr Sankowski (arXiv). One very useful primitive, in many graph property testing results and of course well beyond property testing, is the ability to perform random walks on graphs. In the case of large, distributed (or merely impractically massive) graphs, however, it is not clear how to implement this primitive efficiently. This work addresses this question in the MPC (Massive Parallel Computation) model, and shows how to perform independently, from all of the \(n\) nodes of a graph (i.e., machine), an \(\ell\)-length random walk with only \(O(\log \ell)\) rounds of communication and \(O(n^\alpha)\) space per machine, for any constant \(\alpha \in(0,1)\). This round complexity matches a known (conditional) lower bound, and thus the result is as good as it gets — further, the authors obtain such MPC algorithms for both undirected and directed graph, somehow leveraging the latter case to handle the (significantly more complicated) directed case. As an application of this efficient random walk primitive, they provide MPC property testing algorithms for both bipartiteness and vertex expansion (same definition as in the paper above: vertex expansion \(\Phi\) vs. \(\varepsilon\)-far from any \(G’\) with vertex expansion \(\Phi^2\)) in the general graph model. (Beyond property testing, another application—the main one in the paper— is computing PageRank in both graphs and directed graphs.)

A Lower Bound on Cycle-Finding in Sparse Digraphs, by Xi Chen, Tim Randolph, Rocco Servedio, and Timothy Sun (arXiv). In this paper, the authors tackle the problem of one-sided testing of acyclicity of directed graphs, in the bounded-degree graph model (equivalently, on the task of finding a cycle in a digraph promised to be far from acyclic). Previously, an \(\Omega(\sqrt{n})\) lower bound for this task had been shown by Bender and Ron, based on a birthday-paradox-type argument; this work improves on this hardness result, showing that \(\tilde{\Omega}(n^{5/9})\) queries are necessary. Interestingly, whether achieving any \(o(n)\) query complexity is possible remains open.

Constant-Time Dynamic \((\Delta+1)\)-Coloring and Weight Approximation for Minimum Spanning Forest: Dynamic Algorithms Meet Property Testing, by Monika Henzinger and Pan Peng (arXiv). Finally, the last paper covered this month is concerned with dynamic graph algorithms: where updates to a graph (which can be both edge additions and deletions) come sequentially, and the objective is to maintain, e.g., a coloring, or an approximation of some function of the graph, at any time using as little time per update. This paper advocates the use of techniques from property testing (loosely speaking, as their locality and query-efficiency translates to fast update times) to design better dynamic graph algorithms. It illustrates this idea by obtaining several new results, in particular an (amortized) \(O(1)\)-update-time randomized algorithm to maintain a \((\Delta+1)\)-coloring in a graph with degree bound \(\Delta\). The algorithm relies on a random ranking of the nodes which, combined with a suitable randomized update rule for recoloring a vertex when needed, ensure that the recolorings will not (with high probability) “propagate” to much. This is the first time that this idea, previously used in several testing and sublinear-time papers, is used in the context of dynamic graph algorithms—adding an edge between the two areas, hopefully only the first of many.

Update: we did miss a paper! Make it ten in July…
Nearly optimal edge estimation with independent set queries, by Xi Chen, Amit Levi, Erik Waingarten (arXiv). Consider the following type of access to an unknown graph \(G=(V,E)\) on \(n=|V|\) nodes: on query \(S\subseteq V\), the oracle returns 1 if \(S\) is is an independent set, and 0 otherwise. (This sounds quite a bit like what is allowed in the group testing literature.) How many such queries are necessary to get a multiplicative estimate of the number of edges, \(m=|E|\)? In this paper, the authors improve on the previous bound of \(\min(\sqrt{m}, n^2/m)\) (ignoring logarithmic factors), obtaining an \(\min(\sqrt{m}, n/m)\) upper bound — complemented by a (nearly, up to those pesky log factors) matching lower bound.

If we missed a paper this month, or represented some result above, please mention it in the comments.

We’ve got four papers this month. A mix of distribution testing, matrix problems, and a different paper on the power of sampling.

On the Complexity of Estimating the Effective Support Size, by Oded Goldreich (ECCC). In distribution testing, a classic problem is that of approximating the support size. By a (now) classic result of Valiant-Valiant, the complexity of this problem is \(\Theta(n/\log n)\). This paper raises the question of approximating the “effective” support, and that too with more than just samples from the distribution. The \(\epsilon\)-support of discrete distribution \(\mathcal{D}\) is the smallest support among any distribution \(\mathcal{D}’\) such that \(\|\mathcal{D} – \mathcal{D}’\|_1 \leq \epsilon\) (or TV-distance). Denote this as \(supp_\epsilon(\mathcal{D})\). One can also consider a bicriteria version. Given approximation parameter \(f\) and thresholds \(\epsilon_1 < \epsilon_2\), we need to provide a number in the range \([supp_{\epsilon_2}(\mathcal{D}), f\cdot supp_{\epsilon_1}(\mathcal{D})]\). The primary model studied allows for random samples and evaluation queries (where one gets the probability of any known element of the domain). In this model, for arbitrary \(\epsilon_1, \epsilon_2\), there is a continuum of algorithms, trading off query complexity with approximation. At one end, for \(f = O(\log n)\), the query complexity is \(\widetilde{O}(1/\epsilon_1)\). At the other end, for \(f=1\), the query complexity is \(O(\log^* n)\). (Here, \(n = supp_{\epsilon_1}(\mathcal{D})\).) There are lower bounds showing the necessity of evaluation queries for subpolynomial query complexities.

Communication and Memory Efficient Testing of Discrete Distributions, by Ilias Diakonikolas, Themis Gouleakis, Daniel M. Kane, and Sankeerth Rao (arXiv). This paper adds additional computational constraints to the distribution testing problem. In the streaming setting, the algorithm is only allowed a single pass over the random samples. It has \(m\) bits of storage, much smaller than \(n\), the distribution size. The aim is to minimize the number of samples required for uniformity (and closeness) testing. For uniformity testing in the streaming setting, the paper gives an algorithm that uses \(\widetilde{O}(m + n/m)\) samples. The standard collision algorithm requires \(\Theta(\sqrt{n})\) storage (to store the samples), while this result gives a non-trivial bound for \(m \ll \sqrt{n}\). There are lower bounds showing that these bounds are basically tight. In the distributed distribution testing problem, there are a number of processors, each holding \(\ell\) samples. A referee asks a question to the processor, whose answers are broadcast to everyone. The aim is to minimize communication cost. For uniformity testing in this setting, there is a protocol using \(O(\sqrt{n\log n/\ell})\) bits of communication. As a sanity check, note that for \(\ell = \sqrt{n}\), one processor could simply run the collision algorithm locally to report the result.

Querying a Matrix through Matrix-Vector Products by Xiaoming Sun, David P. Woodruff, Guang Yang, and Jialin Zhang (arXiv). Consider \(n \times d\) matrix \(M\) over some field \(\mathbb{F}\). One gets access to this matrix through matrix-vector products, and wishes to test some matrix property. This is a natural model in many settings, and generalizes the classic setting of query access. A subtle point is that one can only right multiply with “query” vectors, and there are problems where left multiplication can change the complexity. (A nice example in the paper is testing if a square matrix is symmetric. With both left and right multiplications, this is easy, since we can directly access rows and columns. By only accessing columns, this is non-trivial.) This paper studies a number of problems, broadly classified as linear algebra problems, statistics problems, and graph problems. Some highlights are: testing symmetry can be done in \(O(1)\) queries, and the maximum eigenvalue can be approximated in \(O(\log n)\) queries adaptively (but there is an \(\Omega(n)\) non-adaptive lower bound). For graph problems, here’s an interesting discovery. If \(M\) is the adjacency matrix, connectivity requires \(\Omega(n/\log n)\) queries. But if \(M\) is the signed edge-vertex matrix, then this can be done in \(poly(\log n)\) queries. This paper provides a number of interesting directions and problems to study.

The Adversarial Robustness of Sampling by Omri Ben-Eliezer and Eylon Yogev (arXiv). This isn’t a property testing paper, but how can one ignore a paper on understanding the power of sampling? It’s convenient to think of the following setting. An algorithm gets a stream of \(n\) numbers, and has to answer some questions about the stream (say, the median or other quantiles). The simplest strategy is to take a small random sample of the stream. But what if an adversary was generating the stream, depending on the current sample? Under what circumstances is the sample still “representative” of the stream? The specific results of this paper require getting into set systems and VC dimension, which I’ll leave for the sake of simplicity. Let’s go back to the median question. To get an approximate median, a constant number of samples suffice. A consequence of the main result is that if one takes \(O(\log n)\) samples, then this is robust to adversarial streams. On the other hand, lower bounds show that constant sized samples can be fooled by an adversary. The paper is a really interesting read, and is a nice take on “standard facts” that we take for granted.

We were able to find four new papers for May 2019 — as usual, please let us know if we missed any!
EDIT: We did, in fact, miss one paper, which is the bottom one listed below.

On Local Testability in the Non-Signaling Setting, by Alessandro Chiesa, Peter Manohar, and Igor Shinkar (ECCC). This paper studies testability of a certain generalization of (distributions over) functions, known as \(k\)-non-signalling functions, objects which see use in hardness of approximation and delegation of computation. Prior work by the authors show the effectiveness of the linearity test in this setting, leading to the design of PCPs. On the other hand, in this work, the authors show that two types of bivariate tests are ineffective in revealing low-degree structure of these objects.

Computing and Testing Small Vertex Connectivity in Near-Linear Time and Queries, by Danupon Nanongkai, Thatchaphol Saranurak, and Sorrachai Yingchareonthawornchai (arXiv). This work, apparently simultaneous with the one by Forster and Yang that we covered last month, also studies the problem of locally computing cuts in a graph. The authors also go further, and study approximation algorithms for the same problems. Inspired by the connections to property testing in the work of Forster and Yang, they apply these approximation algorithms to get even more query-efficient algorithms for the problems of testing \(k\)-edge- and \(k\)-vertex-connectivity.

Testing Graphs against an Unknown Distribution, by Lior Gishboliner and Asaf Shapira (arXiv). This paper studies graph property testing, under the vertex-distribution-free (VDF) model, as recently introduced by Goldreich. In the VDF model, rather than the ability to sample a random node, the algorithm has the ability to sample a node from some unknown distribution, and must be accurate with respect to the same distribution (reminiscent of the PAC learning model). In Goldreich’s work, it was shown that every property which is testable in the VDF model is semi-hereditary. This work strengthens this statement and proves a converse, thus providing a characterization: a property is testable in the VDF model if and only if it is both hereditary and extendable. These descriptors roughly mean that the property is closed under both removal and addition of nodes (with the choice of addition of edges in the latter case). This is a far simpler characterization than that of properties which are testable in the standard model, which is a special case of the VDF model.

Private Identity Testing for High-Dimensional Distributions, by Clément L. Canonne, Gautam Kamath, Audra McMillan, Jonathan Ullman, and Lydia Zakynthinou (arXiv). This work continues a recent line on distribution testing under the constraint of differential privacy. The settings of interest are multivariate distributions: namely, product distributions over the hypercube and Gaussians with identity covariance. An application of a statistic of CDKS, combined with a Lipschitz extension from the set of datasets likely to be generated by such structured distributions, gives a sample-efficient algorithm. A time-efficient version of this extension is also provided, at the cost of some loss in the sample complexity. Some tools of independent interest include reductions between Gaussian mean and product uniformity testing, balanced product identity to product uniformity testing, and an equivalence between univariate and “extreme” product identity testing.

Testing Bipartitness in an Augmented VDF Bounded-Degree Graph Model, by Oded Goldreich (arXiv). Another work on the vertex-distribution-free (VDF) model, as described above. In this one, Goldreich considers an augmentation of the model, where the algorithm is further allowed to query the probability of each node. With this augmentation, he gives \(\tilde O(\sqrt{n})\)-time algorithms for testing bipartiteness and cycle-free-ness, where \(n\) is the “effective support” of the distribution. That is, \(n\) is the number of nodes in the graph after discarding the nodes with minimal probability until \(\varepsilon/5\) mass is removed.