Research map / Learning theory and algorithms / Matrix and tensor methods
Random matrix theory improved Fréchet mean of symmetric positive definite matrices
Florent Bouchard, Ammar Mian, Malik Tiomoko, Guillaume Ginolhac, Frederic Pascal
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a random-matrix-theory corrected Fréchet (Karcher) mean for SPD matrices that optimizes an RMT-corrected squared Fisher distance directly from raw data, and plugs it into nearest centroid and K-means classifiers.
Problem
Computing Fréchet means of SPD covariance matrices when only raw data with low sample support (n comparable to p) is available, where SCM-based means are biased and unreliable.
Approach
Replace the squared Fisher distance in the Fréchet mean objective with an RMT-corrected consistent estimator computed from the sample covariances of the raw data, and optimize it via Riemannian gradient descent in one step; also improves the covariance estimator via a new stopping criterion (alpha/p threshold) instead of squaring the distance estimator.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the RMT-corrected Fréchet mean and distance instead of SCM-based means and distance | Nearest Centroid on EEG motor imagery datasets · GrosseWentrup09, Schirmeister17, Cho17, Lee19 | accuracy: 0.632 -> 0.638; 0.597 -> 0.603; 0.615 -> 0.622; 0.666 -> 0.66 | Mixed |
| Use the RMT-corrected Fréchet mean and distance instead of SCM-based means and distance | K-means on hyperspectral datasets · Indian Pines, Salinas, Pavia, KSC | accuracy: Indian Pines: 0.385 -> 0.454, 0.357 -> 0.413, 0.377 -> 0.453; Salinas: 0.542 -> 0.777, 0.525 -> 0.746, 0.497 -> 0.632; P | Improved |
| Use the RMT-corrected Fréchet mean and distance instead of SCM-based means and distance | K-means on hyperspectral datasets · Indian Pines, Salinas, Pavia, KSC | mIoU: Indian Pines: 0.278 -> 0.367, 0.229 -> 0.284, 0.253 -> 0.285; Salinas: 0.382 -> 0.631, 0.34 -> 0.532, 0.317 -> 0.461; Pa | Improved |
| Use the RMT-based covariance estimator instead of SCM and linear Ledoit-Wolf | Covariance estimation · Synthetic data at low sample support | squared Fisher distance | Improved |
| Use the RMT-based covariance estimator instead of nonlinear Ledoit-Wolf | Covariance estimation · Synthetic data | squared Fisher distance | Hurt |
| Use the RMT Fréchet mean instead of two-step means computed from covariance estimators | Synthetic Fréchet mean estimation · Simulated covariance matrices; especially moderately low n and larger K | squared Fisher distance | Improved |
Surprising findings
- The RMT covariance estimator, despite being more expensive, does not outperform non-linear Ledoit-Wolf in covariance estimation, so the authors advise against using it for covariance estimation alone.
- Non-linear Ledoit-Wolf failed to produce SPD matrices on some EEG and hyperspectral data, making it unusable there.
Experimental setup
Datasets: GrosseWentrup2009 Schirmeister2017 Cho2017 Lee2019 Indian Pines Salinas Pavia KSC synthetic Gaussian data
Benchmarks: MOABB motor imagery EEG datasets (GrosseWentrup2009, Schirmeister2017, Cho2017, Lee2019) Indian Pines Salinas Pavia KSC
Metrics: classification accuracy squared Fisher distance mean intersection over union
Baselines: SCM Ledoit-Wolf linear shrinkage Ledoit-Wolf non-linear shrinkage nearest centroid with SCM mean K-means with SCM mean
Builds on: RMT-corrected squared distance estimator (Mestre/Vallet-style) Ledoit-Wolf shrinkage QuEST Riemannian gradient descent
Closest papers by meaning
- Random Matrix Improved Covariance Estimation for a Large Class of Metrics (ICML 2019 · similarity 0.85)
- Intrinsic Riemannian Cross-covariance for Manifold-valued Random Objects (NeurIPS 2026 · similarity 0.82)
- Fréchet Regression on the Bures-Wasserstein Manifold (NeurIPS 2026 · similarity 0.82)
- Spectral Estimation with Free Decompression (NeurIPS 2025 · similarity 0.82)
- Wrapped Gaussian on the manifold of Symmetric Positive Definite Matrices (ICML 2025 · similarity 0.82)
- The Underlying Universal Statistical Structure of Natural Datasets (ICML 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
