arXiv subjects

Yuval Filmus

Publications and source records attributed to Yuval Filmus.

At least 37 records · Page 2Linked to original sources

arXiv2024

Shrinkage under Random Projections, and Cubic Formula Lower Bounds for $\mathsf{AC}^0$

$\newcommand{\ACz}{\mathbf{AC}^0}$ Håstad showed that any De Morgan formula (composed of AND, OR and NOT gates) shrinks by a factor of $\tilde{O}(p^{2})$ under a random restriction that leaves each variable alive independently with probability $p$ [SICOMP, 1998]. Using this result, he gave an $\widetildeΩ(n^{3})$ formula size lower bound for the Andreev function, which, up to lower order improvements, remains the state-of-the-art lower bound for any explicit function. In this paper, we extend the shrinkage result of Håstad to hold under a far wider family of random restrictions and their generalization -- random projections. Based on our shrinkage results, we obtain an $\widetildeΩ(n^{3})$ formula size lower bound for an explicit function computable in $\ACz$. This improves upon the best known formula size lower bounds for $\ACz$, that were only quadratic prior to our work. In addition, we prove that the KRW conjecture [Karchmer et al., Computational Complexity 5(3/4), 1995] holds for inner functions for which the unweighted quantum adversary bound is tight. In particular, this holds for inner functions with a tight Khrapchenko bound. Our random projections are tailor-made to the function's structure so that the function maintains structure even under projection -- using such projections is necessary, as standard random restrictions simplify $\ACz$ circuits. In contrast, we show that any De Morgan formula shrinks by a quadratic factor under our random projections, allowing us to prove the cubic lower bound. Our proof techniques build on Håstad's proof for the simpler case of balanced formulas. This allows for a significantly simpler proof at the cost of slightly worse parameters. As such, when specialized to the case of $p$-random restrictions, our proof can be used as an exposition of Håstad's result.

cs.CC↗

arXiv2024

Proving Unsatisfiability with Hitting Formulas

Hitting formulas have been studied in many different contexts at least since [Iwama,89]. A hitting formula is a set of Boolean clauses such that any two of them cannot be simultaneously falsified. [Peitl,Szeider,05] conjectured that hitting formulas should contain the hardest formulas for resolution. They supported their conjecture with experimental findings. Using the fact that hitting formulas are easy to check for satisfiability we use them to build a static proof system Hitting: a refutation of a CNF in Hitting is an unsatisfiable hitting formula such that each of its clauses is a weakening of a clause of the refuted CNF. Comparing this system to resolution and other proof systems is equivalent to studying the hardness of hitting formulas. We show that tree-like resolution and Hitting are quasi-polynomially separated. We prove that Hitting is quasi-polynomially simulated by tree-like resolution, thus hitting formulas cannot be exponentially hard for resolution, so Peitl-Szeider's conjecture is partially refuted. Nevertheless Hitting is surprisingly difficult to polynomially simulate. Using the ideas of PIT for noncommutative circuits [Raz-Shpilka,05] we show that Hitting is simulated by Extended Frege. As a byproduct, we show that a number of static (semi)algebraic systems are verifiable in a deterministic polynomial time. We consider multiple extensions of Hitting. Hitting(+) formulas are conjunctions of clauses containing affine equations instead of just literals, and every assignment falsifies at most one clause. The resulting system is related to Res(+) proof system for which no superpolynomial lower bounds are known: Hitting(+) simulates the tree-like version of Res(+) and is at least quasi-polynomially stronger. We show an exponential lower bound for Hitting(+).

cs.CC↗

arXiv2024

Sparse juntas on the biased hypercube

We give a structure theorem for Boolean functions on the $p$-biased hypercube which are $ε$-close to degree $d$ in $L_2$, showing that they are close to sparse juntas. Our structure theorem implies that such functions are $O(ε^{C_d} + p)$-close to constant functions. We pinpoint the exact value of the constant $C_d$. We also give an analogous result for monotone Boolean functions on the biased hypercube which are $ε$-close to degree $d$ in $L_2$, showing that they are close to sparse DNFs. Our structure theorems are optimal in the following sense: for every $d,ε,p$, we identify a class $\mathcal{F}_{d,ε,p}$ of degree $d$ sparse juntas which are $O(ε)$-close to Boolean (in the monotone case, width $d$ sparse DNFs) such that a Boolean function on the $p$-biased hypercube is $O(ε)$-close to degree $d$ in $L_2$ iff it is $O(ε)$-close to a function in $\mathcal{F}_{d,ε,p}$.

cs.CC↗

arXiv2024

Optimal sets of questions for Twenty Questions

In the distributional Twenty Questions game, Bob chooses a number $x$ from $1$ to $n$ according to a distribution $μ$, and Alice (who knows $μ$) attempts to identify $x$ using Yes/No questions, which Bob answers truthfully. Her goal is to minimize the expected number of questions. The optimal strategy for the Twenty Questions game corresponds to a Huffman code for $μ$, yet this strategy could potentially uses all $2^n$ possible questions. Dagan et al. constructed a set of $1.25^{n+o(n)}$ questions which suffice to construct an optimal strategy for all $μ$, and showed that this number is optimal (up to sub-exponential factors) for infinitely many $n$. We determine the optimal size of such a set of questions for all $n$ (up to sub-exponential factors), answering an open question of Dagan et al. In addition, we generalize the results of Dagan et al. to the $d$-ary setting, obtaining similar results with $1.25$ replaced by $1 + (d-1)/d^{d/(d-1)}$.

cs.DM↗

arXiv2024

Bandit-Feedback Online Multiclass Classification: Variants and Tradeoffs

Consider the domain of multiclass classification within the adversarial online setting. What is the price of relying on bandit feedback as opposed to full information? To what extent can an adaptive adversary amplify the loss compared to an oblivious one? To what extent can a randomized learner reduce the loss compared to a deterministic one? We study these questions in the mistake bound model and provide nearly tight answers. We demonstrate that the optimal mistake bound under bandit feedback is at most $O(k)$ times higher than the optimal mistake bound in the full information case, where $k$ represents the number of labels. This bound is tight and provides an answer to an open question previously posed and studied by Daniely and Helbertal ['13] and by Long ['17, '20], who focused on deterministic learners. Moreover, we present nearly optimal bounds of $\tildeΘ(k)$ on the gap between randomized and deterministic learners, as well as between adaptive and oblivious adversaries in the bandit feedback setting. This stands in contrast to the full information scenario, where adaptive and oblivious adversaries are equivalent, and the gap in mistake bounds between randomized and deterministic learners is a constant multiplicative factor of $2$. In addition, our results imply that in some cases the optimal randomized mistake bound is approximately the square-root of its deterministic parallel. Previous results show that this is essentially the smallest it can get.

cs.LG↗

arXiv2024

Boolean functions on high-dimensional expanders

We initiate the study of Boolean function analysis on high-dimensional expanders. We give a random-walk based definition of high-dimensional expansion, which coincides with the earlier definition in terms of two-sided link expanders. Using this definition, we describe an analog of the Fourier expansion and the Fourier levels of the Boolean hypercube for simplicial complexes. Our analog is a decomposition into approximate eigenspaces of random walks associated with the simplicial complexes. Our random-walk definition and the decomposition have the additional advantage that they extend to the more general setting of posets, encompassing both high-dimensional expanders and the Grassmann poset, which appears in recent work on the unique games conjecture. We then use this decomposition to extend the Friedgut-Kalai-Naor theorem to high-dimensional expanders. Our results demonstrate that a constant-degree high-dimensional expander can sometimes serve as a sparse model for the Boolean slice or hypercube, and quite possibly additional results from Boolean function analysis can be carried over to this sparse model. Therefore, this model can be viewed as a derandomization of the Boolean slice, containing only $|X(k-1)|=O(n)$ points in contrast to $\binom{n}{k}$ points in the $(k)$-slice (which consists of all $n$-bit strings with exactly $k$ ones).

cs.CC↗

arXiv2023

Bounded Simultaneous Messages

We consider the following question of bounded simultaneous messages (BSM) protocols: Can computationally unbounded Alice and Bob evaluate a function $f(x,y)$ of their inputs by sending polynomial-size messages to a computationally bounded Carol? The special case where $f$ is the mod-2 inner-product function and Carol is bounded to AC$^0$ has been studied in previous works. The general question can be broadly motivated by applications in which distributed computation is more costly than local computation, including secure two-party computation. In this work, we initiate a more systematic study of the BSM model, with different functions $f$ and computational bounds on Carol. In particular, we give evidence against the existence of BSM protocols with polynomial-size Carol for naturally distributed variants of NP-complete languages.

cs.CC↗

arXiv2023

Generalized polymorphisms

We find all functions $f_0,f_1,\dots,f_m\colon \{0,1\}^n \to \{0,1\}$ and $g_0,g_1,\dots,g_n\colon \{0,1\}^m \to \{0,1\}$ satisfying the following identity for all $n \times m$ matrices $(z_{ij}) \in \{0,1\}^{n \times m}$: \[ f_0(g_1(z_{11},\dots,z_{1m}),\dots,g_n(z_{n1},\dots,z_{nm})) = g_0(f_1(z_{11},\dots,z_{n1}),\dots,f_m(z_{1m},\dots,z_{nm})). \] Our results generalize work of Dokow and Holzman (2010), which considered the case $g_0 = g_1 = \cdots = g_n$, and of Chase, Filmus, Minzer, Mossel and Saurabh (2022), which considered the case $g_0 \neq g_1 = \cdots = g_n$.

math.CO↗

arXiv2023

Affine vector space partitions

An affine vector space partition of $\operatorname{AG}(n,q)$ is a set of proper affine subspaces that partitions the set of points. Here we determine minimum sizes and enumerate equivalence classes of affine vector space partitions for small parameters. We also give parametric constructions for arbitrary field sizes.

math.CO↗

arXiv2023

MC-finiteness of restricted set partition functions

A sequence $s(n)$ of integers is MC-finite if for every $m \in \mathbb{N}^+$ the sequence $s^m(n) = s(n) \bmod{m}$ is ultimately periodic. We discuss various ways of proving and disproving MC-finiteness. Our examples are mostly taken from set partition functions, but our methods can be applied to many more integer sequences.

math.CO↗

arXiv2023

Sampling and Certifying Symmetric Functions

A circuit $\mathcal{C}$ samples a distribution $\mathbf{X}$ with an error $ε$ if the statistical distance between the output of $\mathcal{C}$ on the uniform input and $\mathbf{X}$ is $ε$. We study the hardness of sampling a uniform distribution over the set of $n$-bit strings of Hamming weight $k$ denoted by $\mathbf{U}^n_k$ for _decision forests_, i.e. every output bit is computed as a decision tree of the inputs. For every $k$ there is an $O(\log n)$-depth decision forest sampling $\mathbf{U}^n_k$ with an inverse-polynomial error [Viola 2012, Czumaj 2015]. We show that for every $ε> 0$ there exists $τ$ such that for decision depth $τ\log (n/k) / \log \log (n/k)$, the error for sampling $\mathbf{U}_k^n$ is at least $1-ε$. Our result is based on the recent robust sunflower lemma [Alweiss, Lovett, Wu, Zhang 2021, Rao 2019]. Our second result is about matching a set of $n$-bit strings with the image of a $d$-_local_ circuit, i.e. such that each output bit depends on at most $d$ input bits. We study the set of all $n$-bit strings whose Hamming weight is at least $n/2$. We improve the previously known locality lower bound from $Ω(\log^* n)$ [Beyersdorff, Datta, Krebs, Mahajan, Scharfenberger-Fabian, Sreenivasaiah, Thomas and Vollmer, 2013] to $Ω(\sqrt{\log n})$, leaving only a quartic gap from the best upper bound of $O(\log^2 n)$.

cs.CC↗

arXiv2022

MaxSAT Resolution and Subcube Sums

We study the MaxRes rule in the context of certifying unsatisfiability. We show that it can be exponentially more powerful than tree-like resolution, and when augmented with weakening (the system MaxResW), p-simulates tree-like resolution. In devising a lower bound technique specific to MaxRes (and not merely inheriting lower bounds from Res), we define a new proof system called the SubCubeSums proof system. This system, which p-simulates MaxResW, can be viewed as a special case of the semialgebraic Sherali-Adams proof system. In expressivity, it is the integral restriction of conical juntas studied in the contexts of communication complexity and extension complexity. We show that it is not simulated by Res. Using a proof technique qualitatively different from the lower bounds that MaxResW inherits from Res, we show that Tseitin contradictions on expander graphs are hard to refute in SubCubeSums. We also establish a lower bound technique via lifting: for formulas requiring large degree in SubCubeSums, their XOR-ification requires large size in SubCubeSums.

cs.CC↗

arXiv2022

A Resilient Distributed Boosting Algorithm

Given a learning task where the data is distributed among several parties, communication is one of the fundamental resources which the parties would like to minimize. We present a distributed boosting algorithm which is resilient to a limited amount of noise. Our algorithm is similar to classical boosting algorithms, although it is equipped with a new component, inspired by Impagliazzo's hard-core lemma [Impagliazzo95], adding a robustness quality to the algorithm. We also complement this result by showing that resilience to any asymptotically larger noise is not achievable by a communication-efficient algorithm.

cs.LG↗

arXiv2022

Junta threshold for low degree Boolean functions on the slice

We show that a Boolean degree $d$ function on the slice $\binom{[n]}{k}$ is a junta if $k \geq 2d$, and that this bound is sharp. We prove a similar result for $A$-valued degree $d$ functions for arbitrary finite $A$, and for functions on an infinite analog of the slice.

math.CO↗

arXiv2022

Orthogonal basis for functions over a slice of the Boolean hypercube

We present an orthogonal basis for functions over a slice of the Boolean hypercube. Our basis is also an orthogonal basis of eigenvectors for the Johnson and Kneser graphs. As an application of our basis, we streamline Wimmer's proof of Friedgut's theorem for slices of the Boolean hypercube.

math.CO↗

arXiv2022

Simple Algebraic Proofs of Uniqueness for Erdős-Ko-Rado Theorems

We give simpler algebraic proofs of uniqueness for several Erdős-Ko-Rado results, i.e., that the canonically intersecting families are the only largest intersecting families. Using these techniques, we characterize the largest partially 2-intersecting families of perfect hypermatchings, resolving a recent conjecture of Meagher, Shirazi, and Stevens.

math.CO↗

arXiv2021

Boolean functions on $S_n$ which are nearly linear

We show that if $f\colon S_n \to \{0,1\}$ is $ε$-close to linear in $L_2$ and $\mathbb{E}[f] \leq 1/2$ then $f$ is $O(ε)$-close to a union of "mostly disjoint" cosets, and moreover this is sharp: any such union is close to linear. This constitutes a sharp Friedgut-Kalai-Naor theorem for the symmetric group. Using similar techniques, we show that if $f\colon S_n \to \mathbb{R}$ is linear, $\Pr[f \notin \{0,1\}] \leq ε$, and $\Pr[f = 1] \leq 1/2$, then $f$ is $O(ε)$-close to a union of mostly disjoint cosets, and this is also sharp; and that if $f\colon S_n \to \mathbb{R}$ is linear and $ε$-close to $\{0,1\}$ in $L_\infty$ then $f$ is $O(ε)$-close in $L_\infty$ to a union of disjoint cosets.

math.CO↗

arXiv2021

Query-to-Communication Lifting Using Low-Discrepancy Gadgets

Lifting theorems are theorems that relate the query complexity of a function $f:\{0,1\}^{n}\to\{0,1\}$ to the communication complexity of the composed function $f \circ g^{n}$, for some "gadget" $g:\{0,1\}^{b}\times\{0,1\}^{b}\to\{0,1\}$. Such theorems allow transferring lower bounds from query complexity to the communication complexity, and have seen numerous applications in the recent years. In addition, such theorems can be viewed as a strong generalization of a direct-sum theorem for the gadget $g$. We prove a new lifting theorem that works for all gadgets $g$ that have logarithmic length and exponentially-small discrepancy, for both deterministic and randomized communication complexity. Thus, we significantly increase the range of gadgets for which such lifting theorems hold. Our result has two main motivations: First, allowing a larger variety of gadgets may support more applications. In particular, our work is the first to prove a randomized lifting theorem for logarithmic-size gadgets, thus improving some applications of the theorem. Second, our result can be seen as a strong generalization of a direct-sum theorem for functions with low discrepancy.

cs.CC↗