Geometry Aware Dimensionality Reduction for Riemannian Manifolds

This direction is merged into another direction: manifold-preserving dimensionality reduction. Its last version stays readable and citable here.

Version 1 of 128 papers34 conceptsfrozen 2026-10-04

When we reduce the dimensionality of data, we usually think of "flattening" it—projecting high-dimensional points onto a lower-dimensional Euclidean plane. However, much of the data in modern computer vision and signal processing does not live in flat space. Instead, it resides on curved surfaces known as Riemannian manifolds.

Common examples include Symmetric Positive Definite (SPD) matrices (like covariance matrices used in brain-computer interfaces) and Grassmann manifolds (which represent linear subspaces). If you treat these curved surfaces as flat, you introduce geometric distortions. It is like trying to represent the Earth on a flat map: you can preserve some properties, but you inevitably stretch or tear the underlying relationships.

The core idea of geometry-aware dimensionality reduction is to map data from a high-dimensional manifold directly to a lower-dimensional manifold of the same type, rather than flattening it into Euclidean space.

The Problem with Flattening

Standard techniques like Principal Component Analysis (PCA) or linear projections assume that the "distance" between two points is a straight line. But for SPD matrices, the space is curved. As Sylvain Chevallier and Florian Yger point out in Geometry-aware visualization of high dimensional Symmetric Positive Definite matrices, flattening these manifolds causes a "swelling effect." This occurs when the calculated dispersion of data points is artificially inflated because the Euclidean straight line "cuts through" the manifold's curvature rather than following it.

By preserving the manifold structure, researchers like Mathieu Salzmann and Mehrtash Harandi demonstrated in From Manifold to Manifold: Geometry-Aware Dimensionality Reduction for SPD Matrices that we can achieve massive computational speedups and higher accuracy by staying "on the manifold."

Bilinear Mapping as a Geometric Bridge

The most common mechanism for reducing the dimension of an SPD matrix $X$ is a bilinear transformation: $f(X) = W X W^T$.

To ensure the output remains a valid SPD matrix, the weight matrix $W$ cannot be arbitrary. In A Riemannian Network for SPD Matrix Learning, Zhiwu Huang and Luc Van Gool constrain $W$ to a Stiefel manifold—meaning $W$ must have semi-orthogonal columns ($W^T W = I$).

This constraint is the "secret sauce." It ensures that the transformation behaves like a rotation and scaling that respects the positive-definiteness of the matrix. This allows a neural network to learn optimal spatial filters that shrink the data size without destroying the signal's intrinsic covariance structure. This is particularly useful in EEG analysis, where Florent Bouchard and Marco Congedo showed that one can reduce electrode counts from 118 down to 32 while maintaining classification accuracy.

Deep Learning on Manifolds

The evolution of this field has moved from shallow, hand-crafted projections to "Riemannian Networks." Architectures like SPDNet and LieNet (introduced by Zhiwu Huang and colleagues) replace standard neural network layers with geometry-aware equivalents:

  1. BiMap Layers: Use the bilinear mapping described above to reduce dimensionality.
  2. ReEig Layers: Act like a ReLU for matrices, rectifying eigenvalues to ensure they stay positive and non-zero, preventing the manifold from "collapsing."
  3. LogEig Layers: Map the data to a tangent space (a local flat approximation) only at the very end of the network to allow for standard classification.

These networks, such as the one proposed in Deep Manifold Learning of Symmetric Positive Definite Matrices with Application to Face Recognition by Su Jia and Zhen Dong, allow the model to learn discriminative features directly on the curved surface.

Beyond SPD: Grassmann and Lie Groups

The same intuition applies to other geometries. In video-based face recognition, a video can be viewed as a linear subspace—a point on a Grassmann manifold. Ruiping Wang and Zhiwu Huang developed Projection Metric Learning (PML) to map these high-dimensional subspaces to lower-dimensional ones.

As explored in Projection Metric Learning on Grassmann Manifold with Application to Video based Face Recognition, this approach avoids the "curse of dimensionality" by learning a mapping between two Grassmannians. Similarly, for 3D skeleton-based action recognition, Zhiwu Huang and Luc Van Gool utilized Lie Groups to represent rotations, using geometric pooling to compress temporal trajectories without losing the structural integrity of the human body's movement.

Why It Works

Geometry-aware dimensionality reduction works because it respects the constraints of the data. When data represents a physical or mathematical property—like the power of a signal (which must be positive) or the orientation of a joint (which must be a rotation)—Euclidean methods can produce "impossible" intermediate states. By constraining the transformation to map from manifold to manifold, we ensure that every step of the dimensionality reduction process produces a mathematically valid and physically meaningful representation.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_geometry_aware_dimensionality_reduction__fd2db0e3,
  title = {Geometry Aware Dimensionality Reduction for Riemannian Manifolds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometry-aware-dimensionality-reduction-for-riemannian-manifolds/rdr_fd2db0e3e05c405d9568a02a93a2c41f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_geometry_aware_dimensionality_reduction__fd2db0e3_v1,
  title = {Geometry Aware Dimensionality Reduction for Riemannian Manifolds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geometry-aware-dimensionality-reduction-for-riemannian-manifolds/rdr_fd2db0e3e05c405d9568a02a93a2c41f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}