UC Berkeley

Broadly speaking, this thesis investigates mathematical questions motivated by computer science. The involved topics include communication avoiding algorithms, classical analysis, convex geometry, and low-rank matrix approximation. In total, the thesis consists of four self-contained sections, each adapted from papers the author has been a part of.

The first two sections are both motivated by the Brascamp-Lieb inequalities, which are also often referred to as Hölder-Brascamp-Lieb inequalities. These inequalities have featured prominently in recent theoretical computer science work, due to connections to geometric complexity theory, harmonic analysis, communication-avoidance, and many other areas. Moreover, work generalizing the inequalities in various ways, such as to nonlinear versions, has been impactful to the study of differential equations.

Section 1 studies the application of Hölder-Brascamp-Lieb (HBL) inequalities to the design of communication optimal algorithms. In particular, it describes optimal tiling (blocking) strategies for nested loops that lack data dependencies and exhibit affine memory access patterns. The problem roughly amounts to maximizing the volume of an object provided some of its linear images have bounded volume. The methods used are algorithmic.

Another reason for the interest in these inequalities is because they are an interesting test case for non-convex optimization techniques. The optimal constant for a particular instance of the inequality is given by solving a non-convex optimization problem that is still highly structured. Of particular relevance to this thesis is that it can be formulated as a geodesically-convex problem, considered in the context of the manifold of positive definite matrices of determinant $1$. Even using the methods of Section 1, the procedure is not necessarily polynomial time, and this motivates further study of geodesic convexity.

This lead to the work of Section 2, which discusses a notion of halfspace for Hadamard manifolds that is natural in the context of convex optimization. For this notion of halfspace, we generalize a classic result of Grunbaum, which itself is a corollary of Helly's theorem. Namely, given a probability distribution on the manifold, there is a point for which all halfspaces based at this point have at least 1/(n+1) of the mass, n being the dimension of the manifold. As an application, the gradient oracle complexity of geodesic convex optimization is polynomial in the parameters defining the problem. In particular it is polynomial in -log(epsilon), where epsilon is the desired error. This is a step toward the open question of whether such an algorithm exists.

The remaining two sections of the paper present a different research direction, randomized numerical linear algebra. Numerical linear algebra has long been an important part of scientific computing. Due to the current trend of increasing matrix sizes and growing importance of fast, approximate solutions in industry, randomized methods are quickly increasing in popularity. Sections 3 and 4 in this thesis aim to show that randomized low-rank approximation algorithms satisfy many of the properties of classical rank-revealing factorizations.

Section 3 introduces a Generalized Randomized QR-decomposition (RURV) that may be applied to arbitrary products of matrices and their inverses, without needing to explicitly compute the products or inverses. This factorization is a critical part of a communication-optimal spectral divide-and-conquer algorithm for the nonsymmetric eigenvalue problem. In this paper, we establish that this randomized QR-factorization satisfies the strong rank-revealing properties. We also formally prove its stability, making it suitable in applications. Finally, we present numerical experiments which demonstrate that our theoretical bounds capture the empirical behavior of the factorization.

Section 4 concerns a Generalized LU-Factorization (GLU) for low-rank matrix approximation. We relate this to past approaches and extensively analyze its approximation properties. The established deterministic guarantees are combined with sketching ensembles satisfying Johnson-Lindenstrauss properties to present complete bounds. Particularly good performance is shown for the sub-sampled randomized Hadamard transform (SRHT) ensemble. Moreover, the factorization is shown to unify and generalize many past algorithms. It also helps to explain the effect of sketching on the growth factor during Gaussian Elimination.