Metric-Aware PCA unifies linear methods with geometric deep learning

Michael Leznik's new paper, 'Metric-Aware PCA as a Linear Instance of Geometric Deep Learning,' formally reframes Principal Component Analysis through the lens of symmetry and equivariance. MAPCA replaces PCA's standard variance-maximization with a parameterized positive-definite metric matrix. This metric acts as a geometric prior, inducing an orthogonal symmetry group under which MAPCA solutions are equivariant and the resulting spectrum is invariant. The key insight: MAPCA's defining constraint is the linear analogue of Schur-type weight constraints used in equivariant neural networks. A uniqueness theorem proves that Invariant PCA (IPCA)—recovered by diagonalizing the metric—is the only linear data-derived metric in the family that remains equivariant under arbitrary diagonal rescaling, matching the classic variance-maximization criterion under normalization.

Beyond theoretical alignment, Leznik extends the framework into three practical bridges: kernel PCA as the nonlinear extension, spectral graph methods as MAPCA on graph-structured data, and a deep MAPCA construction that integrates into deep equivariant networks. This positions MAPCA as a fundamental building block connecting classical dimensionality reduction with modern geometric deep learning. For practitioners, it provides a rigorous, interpretable way to incorporate symmetry priors into linear models, opening the door to more principled data preprocessing and representation learning in domains with known invariances, such as computer vision, physics, and graph analytics.