Extrinsic Geometry for Learning on Grassmann Manifolds

This direction is merged into another direction: geodesic computation overhead. Its last version stays readable and citable here.

Version 1 of 115 papers19 conceptsfrozen 2026-10-04

In computer vision and machine learning, we often represent complex data—like a video clip or a set of face images—as a linear subspace. These subspaces do not live in a standard "flat" Euclidean space; instead, they inhabit a curved mathematical object called a Grassmann manifold.

While these manifolds capture the intrinsic geometry of data, they are notoriously difficult to work with. Traditional "intrinsic" methods require computing Riemannian gradients and logarithmic maps, which are iterative, computationally expensive, and often lack closed-form solutions. A growing body of research suggests a more efficient alternative: extrinsic methods. Instead of navigating the curve, we embed the manifold into a higher-dimensional, flat Euclidean space where standard linear algebra reigns supreme.

The Projection Mapping Intuition

The core idea behind the extrinsic approach is to find a mapping that flattens the manifold without destroying its geometric relationships. For Grassmann manifolds, which consist of $k$-dimensional subspaces, the most effective mapping is the projection matrix.

As explored by Mehrtash Harandi, Chunhua Shen, and colleagues in Extrinsic Methods for Coding and Dictionary Learning on Grassmann Manifolds, a subspace represented by an orthonormal basis $X$ can be mapped to the space of symmetric matrices via the operation $f(X) = XX^T$.

This mapping is isometric: it preserves the "projection distance" between subspaces. By transforming a point on a curved manifold into a symmetric matrix in a higher-dimensional Euclidean space, researchers can use the standard Frobenius norm to measure distances. This simple shift allows complex geometric operations to be replaced by efficient linear algebra, such as eigenvalue decomposition.

Bypassing the Logarithm Map

Why go to the trouble of embedding? In intrinsic manifold learning, moving between the manifold and its "tangent space" requires the Riemannian Logarithm and Exponential maps. These are the "bridges" that allow optimization, but they are computationally heavy.

In Dictionary Learning and Sparse Coding on Grassmann Manifolds: An Extrinsic Solution, Mehrtash Harandi and Conrad Sanderson demonstrate that the extrinsic approach bypasses these maps entirely. By working in the embedded space of symmetric matrices, they derived closed-form solutions for sparse coding. This makes algorithms significantly faster and more stable than their intrinsic counterparts, which must iterate through tangent spaces at every step.

Deep Learning on Curved Spaces

The extrinsic philosophy has also revolutionized deep learning for non-Euclidean data. Standard neural networks expect flat input. To process manifold-valued data, Zhiwu Huang, Jiqing Wu, and Luc Van Gool introduced GrNet in Building Deep Networks on Grassmann Manifolds.

GrNet uses a "Projection Mapping Layer" to transform orthonormal matrices into a flat representation. This allows the network to respect the Grassmannian geometry in its early layers while still utilizing standard fully connected layers and softmax classifiers at the output. By "flattening" the manifold at the right moment, the authors enabled end-to-end training for tasks like action recognition and face verification.

Generalizing the Extrinsic Framework

The utility of extrinsic embedding extends beyond simple classification. It has become a foundational tool for various complex optimization problems:

Summary

The shift from intrinsic to extrinsic methods represents a move toward computational pragmatism. By treating a curved manifold as a subset of a larger, flat space, researchers like Harandi, Huang, and Wang have unlocked the ability to use the most powerful tools in the machine learning arsenal—SGD, kernels, and deep networks—on data that was previously too complex to handle efficiently. The intuition is simple: if the world is too curved to walk on, find a way to look at it as a flat map.

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_extrinsic_geometry_for_learning_on_grass_b977e45e,
  title = {Extrinsic Geometry for Learning on Grassmann Manifolds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/extrinsic-geometry-for-learning-on-grassmann-manifolds/rdr_b977e45e789046cfa2dd4b94fde72202},
  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_extrinsic_geometry_for_learning_on_grass_b977e45e_v1,
  title = {Extrinsic Geometry for Learning on Grassmann Manifolds},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/extrinsic-geometry-for-learning-on-grassmann-manifolds/rdr_b977e45e789046cfa2dd4b94fde72202/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}