Riemannian Learning on Symmetric Positive Definite Manifolds

This direction is split into other directions: SPD manifold representation. Its last version stays readable and citable here.

Version 1 of 126 papers26 conceptsfrozen 2026-10-04

In traditional machine learning, we treat data as points in a flat, Euclidean space. We calculate distances with straight lines and find averages by summing vectors. However, many real-world signals—from the firing patterns of neurons to the volatility of stock markets—are best described not by raw values, but by the correlations between them.

These correlations are represented as Symmetric Positive Definite (SPD) matrices. While you can flatten a matrix into a long vector, doing so destroys its mathematical essence. As Fatih Porikli argues in Learning on Manifolds, these matrices do not live in a flat world; they inhabit a curved Riemannian manifold.

The Geometry of Correlation

An SPD matrix (like a covariance matrix) must satisfy two strict rules: it must be symmetric, and all its eigenvalues must be positive. In Euclidean space, if you "average" two SPD matrices, you might stay within these bounds, but as you move toward the edges of the space, standard operations can lead to "swelling" effects or matrices that are no longer positive definite.

By treating the space of SPD matrices as a curved manifold, researchers like Ravishankar Sivalingam in Positive definite dictionary learning for region covariances preserve the non-Euclidean geometry of the data. This approach ensures that any statistical operation—like interpolation or averaging—results in a mathematically valid matrix that respects the physical constraints of the original signal.

Why Curvature Matters for Robustness

One of the most compelling reasons to use Riemannian geometry is its inherent robustness to noise. In SPDFusion, Huan Kang and colleagues demonstrate that representing image features as SPD matrices captures second-order statistics that are far more stable than raw pixels.

This geometric perspective provides a form of "invariance." For instance, in Brain-Computer Interfaces (BCIs), EEG signals are notoriously noisy and vary between subjects. Xiaoxi Wei and the BEETL Competition organizers showed that by mapping signal spatial characteristics to an SPD manifold, models can align distributions across different people more effectively than in Euclidean space. Similarly, Motoaki Kawanabe uses this geometry to handle "label shift," where the frequency of certain brain states changes between training and testing.

Bridging the Gap: Metrics and Embeddings

To perform learning on a manifold, we need a way to measure distance. The "straightest" path between two points on a curve is a geodesic.

  1. Log-Euclidean Metrics: Ruiping Wang and Zhiwu Huang introduced Log-Euclidean Metric Learning (LEML), which maps matrices into a tangent space using the matrix logarithm. This allows for fast, efficient computation while maintaining the manifold's structure. Joseph S. Picard applied this to radar signal processing, finding it significantly more stable in low-signal environments.
  2. Siegel Embeddings: Often, we care about both the average (mean) and the correlation (covariance). Marcelo Hartmann proposes using Siegel embeddings to fuse these two moments into a single point on a higher-dimensional SPD manifold, allowing the model to treat the entire distribution as one geometric entity.
  3. Kernels: For tasks like gesture recognition, Lei Zhang 0036 uses the Stein kernel to map manifold points into a high-dimensional feature space where "one-shot" learning becomes possible (BoMW).

Deep Learning on Manifolds

The frontier of this research is integrating these geometric insights into neural networks. Rather than just using SPD matrices as inputs, researchers like Zhi Gao are building Deep Networks on SPD Manifolds. These networks use specialized layers that aggregate local features into global covariance descriptors, preserving structural relationships that standard pooling layers (like Max-Pooling) typically discard.

This shift from "flat" to "curved" learning is proving essential in fields where the relationship between variables is more important than the variables themselves—whether detecting structural cracks in infrastructure (Sanduni Jayasinghe), tracking financial volatility (Chulwoo Han), or mapping the dynamic connectivity of the human brain (Dehan Kong).

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_riemannian_learning_on_symmetric_positiv_4df3c26c,
  title = {Riemannian Learning on Symmetric Positive Definite Manifolds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/riemannian-learning-on-symmetric-positive-definite-manifolds/rdr_4df3c26c77c24597b322601ab28e1863},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_riemannian_learning_on_symmetric_positiv_4df3c26c_v1,
  title = {Riemannian Learning on Symmetric Positive Definite Manifolds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/riemannian-learning-on-symmetric-positive-definite-manifolds/rdr_4df3c26c77c24597b322601ab28e1863/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}