Work

Replication code and benchmark summaries for: Exact Subspace Bures-Wasserstein Barycenters and Grassmannian Debiasing for Spiked High-Dimensional Covariances

Year: 2026
Type: software
Abstract: Standalone Python/NumPy/SciPy replication code and benchmark summaries (v1.0.0). Bures-Wasserstein distances and barycenters struggle on sample covariance matrices in high dimensions. In particular, rank-deficient sample subspaces rarely intersect, and fixed-point barycenter updates take cubic time in dimension at each step. We prove two exact results for spiked covariance models. When centered degrees of freedom fall below one quarter of the dimension, random subspace misalignment inflates the expected empirical distance between two independent samples from one isotropic law above their true distance to a point mass. Meanwhile, for spiked positive-definite matrices with isotropic orthogonal variance, the exact barycenter and all pairwise distances live on the joint spike subspace and scalar square roots outside it with zero approximation error. Combining this subspace method with radial Huber trace calibration and Marchenko-Pastur debiasing reduces mean error under heavy tails from 8.09 for dense iteration to 2.93. Our estimator outperforms Ledoit-Wolf Log-Euclidean averaging while running 102.7x faster at dimension 512. On motor-imagery electroencephalography in three frequency bands, the subspace classifier matches dense iteration to machine precision with a 3.45x speedup, and single-trial accuracy shows no significant differences across subjects.
Language: English
Open Access status: green
OpenAlex ID: W7236699718
Cites:
Cited by:
Related to: