Spatially Aware Persistent Homology

16 concepts6 papers1 proposal

Persistent homology (PH) is celebrated for its ability to capture the global structure of data—such as loops, cavities, and connected components—independent of coordinate transformations. However, this coordinate invariance is a double-edged sword. Standard PH can tell you that a topological feature exists, but it completely discards where it is located.

A emerging research direction addresses this limitation by embedding spatial and coordinate information directly into topological descriptors, creating spatially aware persistent homology.

The Spatial Blind Spot of Traditional Topology

Imagine a 3D shape generation task or a medical image segmentation model. If a model predicts a structural hole in the top-left corner of an image, but the ground truth has a hole in the bottom-right corner, standard topological metrics will treat the prediction as perfectly accurate.

As noted by Trofimov et al., traditional topological data analysis tools focus purely on the existence of features rather than their specific spatial coordinates. Consequently, they cannot distinguish between two states that contain identical numbers of topological features in entirely different locations. This lack of localization makes it incredibly difficult to apply targeted topological corrections to specific regions of a generated 3D shape, a limitation highlighted by Chen et al..

This loss of spatial information is fundamental to how persistence diagrams are constructed. In a rigorous look at this inverse problem, Curry demonstrated in The Fiber of the Persistence Map for Functions on the Interval that the spatial arrangement of topological features—such as the relative positioning of peaks and valleys—is precisely the information discarded when a function is summarized into a persistence barcode. Multiple functions can be algebraically identical yet spatially unique.

Standard PH:
[Spatial Domain] ──(Discards Coordinates)──> [Persistence Diagram] (Only counts features)

Spatially Aware PH:
[Spatial Domain] ──(Preserves Coordinates)──> [Localized Diagram] (Features + Positions)

To make topology useful for high-precision downstream tasks like image segmentation and 3D modeling, researchers have developed two primary mechanisms to inject spatial awareness into persistent homology: fused filtrations and patch-based spatial mapping.

Mechanism 1: Fused Filtrations

The most elegant way to force a persistence diagram to remember space is to change how the mathematical filtration is constructed. Instead of filtering data purely by a single value (such as pixel intensity), we can fuse the data value with spatial coordinates.

Oner et al. introduced this concept by combining thresholding-based filtration with a height function. In their extended work, Persistent Homology With Improved Locality Information for More Effective Delineation, the authors formalize this fused filtration function as:

$$F(Y, s)[p] = \mathbf{1}(Y[p] + \rho(p) < s)$$

Where $Y[p]$ is the pixel value at position $p$, and $\rho(p)$ is a spatial coordinate function (such as a height function or a directional ramp).

By adding the spatial term $\rho(p)$, features that possess identical intensities but reside at different positions $p$ are forced to be born or die at different times $s$ during the filtration. This effectively "spreads" the topological features across the persistence diagram's plane based on their physical location. The resulting descriptor becomes sensitive to both the presence and the exact location of topological features, providing granular guidance for deep learning models.

Mechanism 2: Patch-Based Spatial Mapping

Another intuitive approach to restoring spatial awareness is to explicitly localize the topological computation through domain partitioning.

Instead of computing a single global persistence diagram for an entire dataset, Gu et al. propose partitioning an image into uniform local patches. In TopoImages, the authors compute topological summaries for each individual segment and then map these localized descriptors back onto the original spatial grid.

This creates a multi-channel "topology image" that preserves the exact coordinates of structural features like loops or irregular connectivity. By maintaining the spatial relationship between structural features and their coordinates within an image, the model can easily identify localized anomalies—such as irregular connectivity in specific sections of a medical lesion—which global descriptors would otherwise obscure.

Why Spatially Aware Topology Works

By bridging the gap between abstract geometry and spatial coordinates, localized persistent homology allows machine learning models to leverage the robust structural guarantees of topology without losing the spatial precision required for computer vision.

When integrated into deep learning frameworks, it transforms topological descriptors from passive global summaries into actionable, spatially localized signals. This enables loss functions to backpropagate errors to the exact pixel or voxel where a topological artifact occurs, resulting in highly accurate, topologically correct segmentations and generations.

Key Foundational Work

Localized Persistent Homologies for more Effective Deep Learning (2022) — contributed 6 concept elements to this direction.
The Fiber of the Persistence Map for Functions on the Interval (2017) — 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_fused_filtration_for_topological_descrip_69e299b7,
  title = {fused filtration for topological descriptors},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fused-filtration-for-topological-descriptors/rdr_69e299b76f9f45a993c301b1e03b8c47},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_fused_filtration_for_topological_descrip_69e299b7_v1,
  title = {fused filtration for topological descriptors},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/fused-filtration-for-topological-descriptors/rdr_69e299b76f9f45a993c301b1e03b8c47/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake