CS Paper Radar

Rectangular matrix multiplication from polynomial products

Przemyslaw Uznanski

cs.DS ·submitted 2026-10-07 17:46 UTC · v2
theoryhighly cited

Abstract

In this note, we extend the matrix-multiplication analysis by OpenAI to rectangular products and prove $ω(1,k,1)\le2$ for $0\le k\le\frac{1}{2}$, and $ω(1,k,1)\le1+k+\frac{1}{4k}$ for $k\ge\frac{1}{2}$. In particular, $ω(1, \frac{1}{2},1)=2$ and the dual exponent satisfies $α\ge\frac{1}{2}$. The proof uses explicit degenerations of convolution tensors, multinomial counting, and Strassen's spectral theorem. An angular constraint on three dot-product parameters gives both the rectangular curve and a bound for $ω(a,b,c)$. As an application, Zwick's algorithm for all-pairs shortest paths in directed unweighted graphs runs in $O(n^{2.5})$ time. Combining the rectangular bound with the $(\min,+)$-product improvement of Alman and Vassilevska Williams further gives $O(n^{2.4999})$ running time.

Closest papers

Nearest neighbours by meaning across the whole index.

Multiplying matrices via polynomials: exposition and barriers

cs.CC math.RAmath.RT ·2026-10-21 ·Alimzhan Amanov, Maxim van den Berg, Jeroen Zuiddam
highly cited similarity 0.70
Abstract

We give an exposition of OpenAI's recent proof that the exponent $ω_\mathbb{C}$ of matrix multiplication over the complex numbers satisfies $ω_\mathbb{C}\leq9/4$. Our version of the proof gives $ω_\mathbb{F}\leq9/4$ over every field $\mathbb{F}$. We also treat the rectangular matrix multiplication exponent $ω_\mathbb{F}(a,b,c)$. Here we have $ω_\mathbb{F}(1,k,1)=2$ for $0\leq k\leq1/2$ and $ω_\mathbb{F}(1,k,1)\leq1+k+1/(4k)$ for $k\geq1/2$, and, for real numbers $a\geq b\geq c\geq0$, $ω_\mathbb{F}(a,b,c)=a+b$ if $c(a+b)\leq ab$ and $ω_\mathbb{F}(a,b,c)\leq(a+b+c)/2+(ab/c+ac/b+bc/a)/4$ otherwise. Our exposition is in the language of degeneration inequalities among polynomial multiplication tensors and Strassen's asymptotic spectrum duality. We moreover prove an operational barrier (which can be thought of as coming from an operational asymptotic spectrum) showing that the current analysis of all these upper bounds on $ω_\mathbb{F}$ and $ω_\mathbb{F}(a,b,c)$ is optimal and that in fact the current degeneration ingredients cannot improve the bounds.

0.46

Faster Algorithms for Disjoint Paths in DAGs and Two Disjoint Shortest Paths

cs.DS ·2026-11-01 ·Yang Cao, Haoran Qi, Hanzhi Wang
similarity 0.63
Abstract

For $k$ vertex-disjoint paths with prescribed endpoints in a directed acyclic graph (DAG), the $O(mn^{k-1})$-time pebbling algorithm of Fortune, Hopcroft, and Wyllie [TCS 1980] has remained the fastest for every $k\geq3$ for over four decades, even when the paths must be shortest. We break this bound for every $k\geq3$ with randomized algorithms running in expected $O_k(m+n^{k-3+ω+\varepsilon})$ time for every $\varepsilon>0$, where $ω<2.372$ is the matrix multiplication exponent; this is $O_k(m+n^{k-0.62})$. For two disjoint shortest paths, we find the paths as fast as we detect them: in expected linear time on DAGs and on undirected graphs with polynomially bounded integer weights, and in expected $O(\min\{mn,n^{ω+\varepsilon}\})$ time on directed graphs whose cycles have positive weight, which on dense graphs is $o(mn)$ even for detection. All algorithms evaluate a polynomial whose monomials are the solutions: a new first-collision circuit speeds up the evaluation in DAGs, and a correlated extraction step at cuts crossed once by every path finds a solution at the cost of a constant number of evaluations.

0.20

Combinatorial Boolean Matrix Multiplication in $n^3 / 2^{Ω(\sqrt{\log n})}$ Time

cs.DS ·2026-11-04 ·Piaoyang Cui, Yiming Liu, David P. Woodruff
similarity 0.61
Abstract

For over fifty years, combinatorial algorithms for Boolean matrix multiplication (BMM) and triangle detection saved only polylogarithmic factors over cubic time. At STOC 2024, Abboud, Fischer, Kelley, Lovett, and Meka broke this barrier with a deterministic $O(n^3 / 2^{Ω((\log n)^{1/7})})$-time algorithm, and asked whether $n^3 / 2^{Ω(\sqrt{\log n})}$ time is achievable. We answer this question in the affirmative, giving deterministic combinatorial algorithms for both BMM and triangle detection running in $O(n^3 / 2^{Ω(\sqrt{\log n})})$ time. Following Abboud et al., we decompose $(A,B)$ into pairs of submatrices $(A_k,B_k)$, called \emph{pieces}, each of which is either sparse or pseudorandom in a sense that forces $A_k\star B_k$ to be nearly all-ones on its vertex sets. Whereas their decomposition produces $2^{O(d^7)}$ pieces at regularity parameter $d$, we obtain $2^{O(d^2)}$ pieces, directly improving the runtime exponent from $(\log n)^{1/7}$ to $\sqrt{\log n}$. Prior work found pseudorandom pieces by repeatedly restricting to denser submatrices, paying worst-case vertex losses at each step that compounded across nested recursion levels, together with an expensive derandomization step. We eliminate the extra powers of $d$ by matching the vertex loss at each step to the size of its density jump so losses telescope across scales, derandomizing witness search via $L_d$-norm estimation so the sampler length stays $2^{O(d)}$, and bounding overlap across the recursion with a fractional density-weighted volume budget rather than raw submatrix volume.

0.20

An Exposition of the $\widetilde{O}(\log^{1/4} n)$ Bound for the Komlós Problem

cs.DM math.COcs.DSmath.PR ·2026-08-28 ·Nikhil Bansal, Haotian Jiang
publishedtheory similarity 0.59
Abstract · An extended abstract of this work appeared in STOC 2026: https://arxiv.org/pdf/2508.03961. The present article treats only the Komlós proble…

A conjecture of Komlós states that the combinatorial discrepancy of any matrix $A\in\mathbb R^{m\times n}$ whose columns have Euclidean norm at most one is bounded by a universal constant. We prove that the combinatorial discrepancy of every such matrix is at most $O((\log n)^{1/4}(\log\log n)^{7/4})$. This is the first asymptotic improvement over the $O(\sqrt{\log n})$ bound established by Banaszczyk [Banaszczyk, Random Struct.\ Algorithms, 1998], and it refutes a conjecture of Hajela [Hajela, European J.\ Combin., 1988] that a lower bound of order $Ω(\sqrt{\log n})$ should hold.

0.38

Fast Spectral Signing for Vector Balancing

cs.DS ·2026-09-24 ·Xiaoyu Li
theoryhighly cited similarity 0.58
Abstract

The Komlós conjecture, now a theorem, asserts that whenever the columns of a matrix $A\in\mathbb{R}^{m\times n}$ have Euclidean norm at most one, some signs $\varepsilon\in\{-1,1\}^n$ make every coordinate of $A\varepsilon$ bounded by an absolute constant. Guo, Fang, and Lu gave the first polynomial-time algorithm for finding such signs, a deterministic spectral signing procedure with discrepancy $8272$ and running time $O((mn^9+n^{10})\log(m+n))$. We give a deterministic algorithm that finds signs with $\|A\varepsilon\|_\infty<99$ using $O(mn+n^{ω+2}\log^3 n)$ arithmetic operations, where $ω>2$ is any fixed attainable matrix-multiplication exponent; with the current bounds on $ω$ this is $\widetilde O(mn+n^{4.372})$. Our algorithm uses the same framework: it rounds a single fractional coloring and watches all rows through the top eigenvalue of a Gram matrix of energy-corrected barriers. Steps follow flat directions, rescaled so that no barrier near its threshold moves faster than a constant, and the regularizer grows as coordinates freeze; together these bound the number of updates by $O(n^2\log n)$. A weak quadratic charge on the tracked row sums leaves $O(n\log^2 n)$ rows to evaluate at any time, and a motion clock bounds when any other row could approach its barrier. Each update is a short sequence of matrix products. Its direction is read off by conditional expectations from a polynomial soft projector, its small constraint residual is repaired in affine row representations, and one identity accounts for every change of representation. For rational input the algorithm has polynomial bit complexity.

0.32

Matrix Multiplication from Polynomial Packings: A Constructive and Elementary Proof of $ω\le9/4$

cs.CC ·2026-10-20 ·Rafael Pass
similarity 0.58
Abstract

We give a self-contained constructive proof that the matrix multiplication exponent satisfies $ω\le9/4$. In a recent breakthrough, OpenAI~\cite{original} established this bound in the model of exact scalar arithmetic over the complex numbers, using non-constructive arguments in Strassen's asymptotic spectral framework. We explicitly construct a Strassen-style matrix multiplication identity and obtain the bound by ordinary block recursion. For every $γ>0$, the resulting algorithm multiplies two $n\times n$ integer matrices using $O(n^{9/4+γ})$ integer arithmetic operations. As an independent contribution, we develop a framework of \emph{polynomial packings}---methods for reducing many polynomial multiplications to fewer ones using linear operations. We isolate and generalize a construction of OpenAI~\cite{original} to obtain a simple packing, and show how packings can be composed to amplify their ``compression rate", which finally yields Strassen-like identities.

0.07

Thin matrix products up to D <= N^0.172 and lossless reductions for 3SUM, Exact Triangle and APSP, with 3SUM over abelian groups and a dichotomy for coordinatewise problems

cs.DS ·2026-10-12 ·Ruhollah Majdoddin
theorycode/data similarity 0.57
Abstract · 130 pages. Lean formalisation and scripts: https://github.com/Majdoddin/formal-math

Alman and Vassilevska Williams (arXiv:2610.06783) gave the first polynomial improvements over the textbook algorithms for 3SUM and APSP: deterministic algorithms with exponents 1.9992 for 3SUM, 2.9983 for Exact Triangle and 2.99942 for APSP. Their algorithm has two steps: a matrix theorem, computing N^2/sqrt(D) prescribed entries of the product of an N x D and a D x N integer matrix in time N^2/D^gamma, gamma > 0, when D <= N^eps, for every eps < 0.1204; and reductions from the three problems to the lopsided sparse triangle problem. We improve both. With Schönhage's five-term partial-product tensor, the matrix theorem holds for every eps < 0.172. The preprint's reductions keep a quarter of the matrix theorem's saving for 3SUM and a half for Exact Triangle and APSP; ours keep all of it, by calling the counting version of the triangle problem adaptively and building composite moduli one prime at a time. Together this gives deterministic 3SUM in n^{1.9945+o(1)}, and Exact Triangle, (min,+)-product and APSP in n^{2.9945+o(1)}. We also improve on the preprint's proof sketch for 3XOR and solve it deterministically in O(n^{1.99451}). Which other problems on three lists does the method make truly subquadratic? It solves 3SUM over F_p^d for every fixed prime p with 3XOR's exponents, and over every finite abelian group of exponent p^2, such as Z_4^d, in O(n^{1.99726}). For coordinatewise problems with a fixed relation P on every coordinate triple, a certified exhaustive computation shows that every Boolean P, and every functional P over three letters, either is the solution set of an equation over Z, F_2 or F_3, and the problem is then truly subquadratic, or has a gadget for Orthogonal Vectors, and the problem is then hard under the Orthogonal Vectors hypothesis. Over four letters, just four functional relations, up to symmetry, satisfy neither condition.

0.58

Improving the matrix multiplication exponent with modern optimization and AlphaEvolve

cs.DS cs.AIcs.CCcs.LG ·2026-08-17 ·Emilien Dupont, Marvin Eisenberger, Borislav Kozlovskii, Abbas Mehrabian, Francisco J. R. Ruiz +5 more
theoryhighly cited similarity 0.57
Abstract

The current best bounds on the matrix multiplication exponent $ω$ are obtained through a refinement of the laser method called combination loss analysis (Duan et al., 2022; Williams et al., 2024; Alman et al., 2025). In this note, we address the optimization problem at the core of this approach and propose several improvements. First, we reformulate the optimization problem allowing us to solve it in a larger setting than was previously possible. Second, we leverage recent advances in machine learning to design a new optimization algorithm for this problem. Finally, we refine the resulting optimization algorithm with AlphaEvolve. Our combined approach yields an upper bound of $ω$ < 2.371177, improving the previous best bound of 2.371339.

0.58