Measurability and continuity of parametric low-rank approximation in Hilbert spaces: linear operators and random variables

Abstract

We develop a unified theoretical framework for low-rank approximation techniques in parametric settings, where traditional methods like Singular Value Decomposition (SVD), Proper Orthogonal Decomposition (POD), and Principal Component Analysis (PCA) face significant challenges due to repeated queries. Applications include, e.g., the numerical treatment of parameter-dependent partial differential equations (PDEs), where operators vary with parameters, and the statistical analysis of longitudinal data, where complex measurements like audio signals and images are collected over time. Although the applied literature has introduced partial solutions through adaptive algorithms, these advancements lack a comprehensive mathematical foundation. As a result, key theoretical questions -- such as the existence and parametric regularity of optimal low-rank approximants -- remain inadequately addressed. Our goal is to bridge this gap between theory and practice by establishing a rigorous framework for parametric low-rank approximation under minimal assumptions, specifically focusing on cases where parameterizations are either measurable or continuous. The analysis is carried out within the context of separable Hilbert spaces, ensuring applicability to both finite and infinite-dimensional settings. Finally, connections to recently emerging trends in the Deep Learning literature, relevant for engineering and data science, are also discussed.