Aligning Tangent Spaces with Parallel Transport

26 concepts20 papers1 proposal

In standard deep learning, moving a feature or a filter across an input is trivial. On a flat 2D image grid, a convolutional filter slides from pixel to pixel without changing its meaning because the underlying space has a global coordinate system—"up" means the exact same thing everywhere.

However, when deep learning expands to non-Euclidean domains—such as 3D surface meshes, graphs, or curved Riemannian manifolds—this fundamental assumption breaks down. On a curved surface, there is no global coordinate system or universal definition of direction (Poulenard et al.). If you place an arrow on a sphere and slide it to another point along two different paths, it will end up pointing in two different directions. This geometric phenomenon introduces a profound challenge for neural networks: how do we share weights, pass directional messages, or compare features across a curved space?

The solution lies in a classical concept from differential geometry: parallel transport. By treating parallel transport as a core computational layer, recent research provides a mathematically rigorous way to align directional information across curved domains, unlocking consistent weight sharing, message passing, and trajectory analysis.

The Problem of Tangent Space Misalignment

To understand why parallel transport is necessary, we must first look at how vectors exist on curved surfaces. A vector at any given point on a manifold does not float freely in space; it lives in a flat, localized vector space called a tangent space.

Because every point on a curved surface has its own unique tangent space, vectors located at different points cannot be directly added, subtracted, or compared (Battiloro et al.). If a neural network attempts to aggregate directional features from neighboring vertices on a mesh without adjusting for this curvature, the signals will conflict and cancel out, destroying fine geometric details (De Haan et al.). This lack of a shared reference frame makes achieving translation equivariance on curved surfaces exceptionally difficult (Yang et al.).

Flat Space (Euclidean):
[Point A: Vector Up]  ---> Slide --->  [Point B: Vector Up]  (Identical Frame)

Curved Space (Manifold):
[Point A: Tangent Space A] ---> ??? ---> [Point B: Tangent Space B] (Misaligned Frames)

The Mechanism: Bridging Tangent Spaces

Parallel transport acts as a mathematical bridge between distinct tangent spaces. It defines a way to move a vector along a specific path (such as a shortest-path geodesic) while keeping it as "straight" or "parallel" as possible relative to the surface geometry.

As the vector travels, parallel transport automatically rotates it to compensate for the underlying curvature. This operation effectively translates a vector-valued feature from a source node's local coordinate frame (or "gauge") into a target node's frame (Park et al.). By mapping local tangent spaces across a manifold to a common reference point, parallel transport enables geometrically valid subtraction, distance calculations, and feature aggregation (Su et al.; Myhre et al.).

Weight Sharing and Surface Convolutions

The most immediate application of this mechanism is enabling weight sharing for convolutions on surfaces. In a standard CNN, weight sharing relies on translating a filter across a grid. On a manifold, Cohen et al. formalized filter movement as parallel transport.

When a learned filter is broadcast from a central point to its neighbors, parallel transport preserves its geometric orientation relative to the surface. This ensures gauge equivariance: a pattern learned in one region of a manifold will be recognized identically in another, regardless of how the local coordinate systems are oriented (Cohen et al.). Architectures like Harmonic Surface Networks (Wiersma et al.) and Multi-directional Geodesic Neural Networks (Poulenard et al.) leverage this to maintain rotation-equivariant features, allowing directional signals to propagate through deep layers without losing their orientation.

Message Passing and Attention in Geometric Networks

Beyond convolutions, parallel transport has become vital for message passing in Graph Neural Networks (GNNs) and attention mechanisms in Transformers operating on non-Euclidean data.

  • Anisotropic Message Passing: In Gauge Equivariant Mesh CNNs (De Haan et al.), parallel transport rotates vector features during message passing, allowing the network to capture direction-sensitive geometric details that standard isotropic graph methods ignore. For simulating physical dynamics governed by partial differential equations (PDEs) on curved surfaces, Liao et al. use parallel transport to broadcast parameters within local subtrees, ensuring that the operations respect the local curvature.
  • Manifold Attention: The Gauge Equivariant Transformer (He et al.) incorporates parallel transport into its self-attention layers. This allows the model to calculate attention weights between distant points on a 3D shape while accounting for the geometric rotation that occurs across the surface.
  • Discrete Graph Bundles: The concept also extends to abstract graph structures. Bamberger et al. introduce path-independent parallel transport on flat vector bundles to prevent the signal degradation and feature convergence typically associated with deep GNNs. Similarly, Sun et al. utilize parallel transport to maintain geometric consistency when passing messages across hyperbolic and spherical spaces in graph foundation models.

Trajectory and Statistical Analysis

Parallel transport is equally essential when analyzing trajectories or statistical features that evolve over time on Riemannian manifolds.

When tracking deforming shapes or human motion, the velocity vectors at different time steps reside in entirely different tangent spaces, making standard aggregate statistical analysis mathematically invalid (the authors of the ICCV 2017 paper). By employing parallel transport, these velocity vectors can be moved along the trajectory to a single reference tangent space.

This alignment enables robust, rate-invariant trajectory analysis for tasks like visual speech recognition (Su et al.) and preserves the discriminative power of statistical correlations across different spatial or temporal locations (Nguyen). Furthermore, Brunton et al. demonstrate that parallel transport can be used to propagate local shape correspondences across deforming geometries, allowing mappings to flow naturally across a surface while adapting to local features.

Key Foundational Work

Rate-Invariant Analysis of Trajectories on Riemannian Manifolds with Application in Visual Speech Recognition (2014) — contributed 2 concept elements to this direction.
A Low-Dimensional Representation for Robust Partial Isometric Correspondences Computation (2013) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_parallel_transport_weight_sharing_423a37e4,
  title = {parallel transport weight sharing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/parallel-transport-weight-sharing/rdr_423a37e40c9d4c27a35adf3517e0818a},
  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_parallel_transport_weight_sharing_423a37e4_v1,
  title = {parallel transport weight sharing},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/parallel-transport-weight-sharing/rdr_423a37e40c9d4c27a35adf3517e0818a/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake