Preserving Geodesic Distances on Data Manifolds

63 concepts55 papers

When analyzing high-dimensional data—whether it is images of faces, text embeddings, or acoustic features of speech—we often assume that standard geometric metrics like Euclidean distance or cosine similarity accurately reflect true similarity. However, high-dimensional data rarely fills its ambient space uniformly. Instead, it typically concentrates on lower-dimensional, highly curved surfaces known as manifolds.

In a seminal paper, Mapping a Manifold of Perceptual Observations, Joshua B. Tenenbaum illustrated why standard metrics fail in these settings. Imagine a sheet of paper rolled into a spiral (a "Swiss roll"). Two points on adjacent layers of the spiral may be physically close in the three-dimensional ambient space, but to travel from one to the other along the surface of the paper requires walking all the way around the curve. Measuring the straight-line Euclidean distance takes an illegal "shortcut" through the empty space, misrepresenting the true intrinsic relationship between the points.

This phenomenon appears across diverse domains. For instance, Mingyu You et al. note in Manifolds Based Emotion Recognition in Speech that acoustic features for speech emotion recognition form complex, non-linear spaces where straight-line metrics lose vital structural information. Similarly, Alexander Modell et al. demonstrate in The Origins of Representation Manifolds in Large Language Models that large language models represent continuous concepts (like time or color) as geometric manifolds, meaning that standard linear metrics like cosine similarity can distort the true conceptual distance.

Approximating Geodesic Distance via Neighborhood Graphs

To preserve the true geometry of a manifold, we need to measure geodesic distance—the length of the shortest path between two points confined entirely to the manifold's surface. Because we rarely possess an explicit mathematical formula for the underlying manifold, we must approximate it from discrete data points.

The core mechanism relies on a simple but powerful insight from differential geometry: curved spaces look locally flat. While global Euclidean distance fails, the Euclidean distance between a point and its immediate neighbors remains a highly accurate approximation of the geodesic distance.

This motivates a three-step graph-based approximation framework:

  1. Construct a Neighborhood Graph: Connect each data point to its nearest neighbors (using a $k$-nearest neighbor graph or an $\epsilon$-radius neighborhood). Weight each edge by the standard Euclidean distance between the connected points.
  2. Compute Shortest Paths: For any two distant points on the manifold, find the shortest path between them through the graph. Because the path is forced to hop from neighbor to neighbor, it is constrained to follow the winding contours of the manifold surface.
  3. Construct the Geodesic Matrix: Accumulate these shortest-path lengths into a global distance matrix that replaces the flawed ambient metrics.

Different algorithms implement this shortest-path step depending on the architecture. For example, Rishi Jha et al. utilize Dijkstra's algorithm in On Geodesic Distances and Contextual Embedding Compression for Text Classification to compute these paths for compressing text embeddings. In the context of structural graph learning, Ryoma Sato demonstrates in Graph Neural Networks can Recover the Hidden Features Solely from the Graph Structure that graph neural networks can implicitly compute these distances using message-passing variants of the Bellman-Ford algorithm to recover hidden geometric features from structure alone. To handle noisy or functional data, Chloé Dimeglio et al. propose in A robust algorithm for template curve estimation based on manifold embedding using a Minimum Spanning Tree to adaptively define connectivity scales, minimizing bias compared to fixed-neighborhood approaches.

Unfolding the Manifold

Once the global geodesic distance matrix is constructed, the next challenge is to find a low-dimensional representation that honors these distances. This process is often conceptualized as "unfolding" or "flattening" the manifold into a lower-dimensional Euclidean space without tearing it.

In the classic Isomap algorithm introduced by Joshua B. Tenenbaum in Mapping a Manifold of Perceptual Observations, this flattening is achieved via Multidimensional Scaling (MDS). MDS ingests the matrix of approximated geodesic distances and optimizes a global coordinate system that minimizes the discrepancy between the low-dimensional Euclidean distances and the high-dimensional geodesic distances.

Alternative optimization strategies offer different geometric guarantees. Amir Globerson et al. frame the problem in Visualizing pairwise similarity via semidefinite programming as a convex semidefinite program that maximizes the variance (or "spread") of the embedding while maintaining pairwise similarity constraints, preventing the representation from collapsing. To bridge the gap between non-linear manifold discovery and out-of-sample generalization, Deng Cai et al. introduce Isometric Projection, which learns a linear transformation matrix that approximates the global geodesic structure, combining the computational efficiency of linear methods with the topological awareness of non-linear approaches.

Beyond Dimensionality Reduction

While originally designed for visualization and dimensionality reduction, preserving geodesic distances has evolved into a foundational design pattern across modern machine learning paradigms.

Multi-modal Alignment

When aligning disparate modalities like text and images, linear alignment strategies often struggle with non-linear semantic spaces. Shibin Mei et al. introduce a geodesic perspective in GeoMM: On Geodesic Perspective for Multi-modal Learning, replacing standard cosine similarity with graph-based shortest paths. This allows models to better differentiate between semantically distinct samples that appear superficially close in a joint embedding space. Similarly, Cencheng Shen et al. utilize shortest-path distances in Manifold matching using shortest-path distance and joint neighborhood selection to align cross-modal datasets by capturing their shared underlying shape.

Robust Representation and Classification

In zero-shot cross-lingual spoken language understanding, Xuxin Cheng et al. leverage geodesic-based contrastive learning in Cyclical Contrastive Learning Based on Geodesic for Zero-shot Cross-lingual Spoken Language Understanding to map sentences across languages, ensuring that semantic relatedness respects the intrinsic geometry of the embedding space. For pattern classification, Ming-Hsuan Yang et al. demonstrate in Extended Isomap for Pattern Classification that using geodesic distances as features directly feeds richer structural information into downstream classifiers like Fisher Linear Discriminants, significantly boosting performance on face and digit recognition tasks.

By shifting the analytical lens from straight-line ambient metrics to graph-approximated geodesic paths, these techniques allow algorithms to respect the true topological constraints of the data, bypassing the curse of dimensionality by operating natively within the data's intrinsic geometry.

Key Foundational Work

Mapping a Manifold of Perceptual Observations (1997) — contributed 2 concept elements to this direction.
Semantic geodesic maps: a unifying geometrical approach for studying the structure and dynamics of single trial evoked responses (2002) — contributed 1 concept elements to this direction.
Extended Isomap for Pattern Classification (2002) — 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_geodesic_distance_preservation_c39bfd60,
  title = {geodesic distance preservation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geodesic-distance-preservation/rdr_c39bfd6031d5478abbac16e03b77b458},
  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_geodesic_distance_preservation_c39bfd60_v1,
  title = {geodesic distance preservation},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/geodesic-distance-preservation/rdr_c39bfd6031d5478abbac16e03b77b458/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake