EFFICIENT GEOMETRIC ALGORITHMS FOR STOCHASTIC AND CATEGORICAL DATA

Abstract

Modern algorithmic frameworks must contend with data that is both massive and inherently uncertain. This dissertation explores efficient geometric algorithms for handling such data across two regimes: temporal evolution and spatial classification.

In the first part, we address "temporal uncertainty" by developing algorithms to track evolving data. We introduce a framework for tracking labels on trees and moving points in the Euclidean plane, establishing the necessary speedup factors required to maintain an accurate hypothesis of the ground truth. We extend this to track evolving probability distributions under a local-motion model, using the Kullback-Leibler (KL) divergence to bound the information loss over time.

In the second part, we address "spatial uncertainty" by developing classifiers for data constrained to bounded domains, such as the probability simplex. We investigate Support Vector Machines (SVMs) under a non-Euclidean geometry called the Hilbert metric. This is a projective geometry that like the KL divergence is sensitive to domain boundaries. We prove that the Hilbert SVM problem is of LP-type, allowing for exact solutions, and we present a polynomial-time approximation algorithm that overcomes the curse of dimensionality inherent in exact approaches. Together, these results provide a robust algorithmic foundation for managing uncertainty in dynamic and bounded geometric systems.