no. 17 · Wednesday 18 November 2026
On the Benign Optimization Landscape and Geometry of Heteroscedastic Probabilistic PCA
Abstract
In this work, we study the global optimization landscape and local geometry of heteroscedastic probabilistic principal component analysis (HePPCA). Since the signal strengths and noise variances are often unknown in practice, we analyze a formulation that uses fixed plug-in estimates of these parameters. We first show that the population HePPCA problem is an instance of a homogeneous quadratic program with orthogonality constraints (QPOC). For QPOC, we first derive a necessary and sufficient condition for a critical point to be a global maximizer and explicitly construct a positive-curvature direction at every nonglobal critical point. This establishes a benign optimization landscape of QPOC in which every critical point is either a global maximizer or a strict saddle point. Building on this analysis, we establish quantitative curvature bounds for population HePPCA, yielding local geodesic strong concavity near every global maximizer and uniformly positive curvature in suitable tangent directions at all nonglobal critical points. Combining these results with a deterministic perturbation argument and sample covariance concentration bounds, we derive a sufficient sample-size condition under which empirical HePPCA inherits the benign landscape and local geodesic strong concavity with high probability. Finally, we present numerical experiments to corroborate our theoretical results.
- comments
- 33 pages, 4 figures
- fields
- Optimization and Control