The Complexity of Dynamic Matrix Problems

Introduction

Modern data analysis and optimization rely on our ability to process rapidly evolving dynamic datasets, often involving matrix operations in very high dimensions. Dynamic data structures enable fast information retrieval on these huge databases by maintaining implicit information on the underlying data. As such, understanding the power and limitations of dynamic (matrix) data structures is a fundamental question in theory and practice.

Despite decades of research, there are still very basic dynamic problems whose complexity is (exponentially) far from understood. Bridging this gap is one of the centerpieces of this proposal.

Advancing Dynamic Data Structures

The second theme of this proposal is advancing the nascent role of dynamic data structures in continuous optimization. For over a century, the traditional focus of optimization research was on minimizing the rate of convergence of local-search methods.

The last ∼3 years have witnessed the dramatic potential of dynamic data structures in reducing the cost-per-iteration of (Newton type) optimization algorithms. This indicates that the bottleneck to accelerating literally thousands of algorithms is the efficient maintenance of dynamic matrix functions.

This new framework is only at its early stages but has already led to breakthroughs on decade-old problems in computer science. This proposal will substantially develop this interdisciplinary theory and identify the mathematical machinery which would lead to ultra-fast first and second-order convex optimization.

Non-Convex Setting

In the non-convex setting, this proposal demonstrates the game-changing potential of dynamic data structures and algebraic sketching techniques in achieving scalable training and inference of deep neural networks, a major challenge of modern AI.

Our program is based on a novel connection of kernel methods and compressed sensing techniques for approximate matrix multiplication.