Doctoral Thesis: Linear Algebraic Techniques in Algorithms and Complexity


Event Speaker: 

Josh Alman

Event Location: 


Event Date/Time: 

Thursday, August 29, 2019 - 3:00pm


We develop linear algebraic techniques in algorithms and complexity, and apply them to a variety of different problems. We focus in particular on matrix multiplication algorithms, which have surprisingly fast running times and can hence be used to design fast algorithms in many settings, and matrix rank methods, which can be used to design algorithms or prove lower bounds by analyzing the ranks of matrices corresponding to computational tasks.

First, we study the design of matrix multiplication algorithms. We define a new general method, called the Universal Method, which subsumes all the known approaches to designing these algorithms. We then design a suite of techniques for proving lower bounds on the running times which can be achieved by algorithms using many tensors and the Universal Method. Our main limitation result is that a large class of tensors generalizing the Coppersmith-Winograd tensors (the family of tensors used in all record-holding algorithms for the past 30+ years) cannot achieve a better running time for multiplying n by n matrices than O(n^{2.168}).

Second, we design faster algorithms for batch nearest neighbor search, the problem where one is given sets of data points and query points, and one wants to find the most similar data point to each query point, according to some distance measure. We give the first subquadratic time algorithm for the exact problem in high dimensions, and the fastest known algorithm for the approximate problem, for various distance measures including Hamming and Euclidean distance. Our algorithms make use of new probabilistic polynomial constructions to reduce the problem to the multiplication of low-rank matrices.

Third, we study rigid matrices, which cannot be written as the sum of a low rank matrix and a sparse matrix. Finding explicit rigid matrices is an important open problem in complexity theory with applications in many different areas. We show that the Walsh-Hadamard transform, previously a leading candidate rigid matrix, is in fact not rigid. We also give the first nontrivial construction of rigid matrices in a certain parameter regime with applications to communication complexity, using an efficient algorithm with access to an NP oracle.  

Thesis Supervisors: Professors Ryan Williams and Virginia Vassilevska Williams