A Sheaf-Theoretic Construction of Shape Space

Shreya Arya, Justin Curry, Sayan MukherjeeFound. Comput. Math 2022

TL;DR

This paper introduces a rigorous mathematical framework for "shape space"—the collection of all possible shapes—using the language of sheaf theory. By treating the Persistent Homology Transform (PHT) as a "homotopy sheaf," the authors prove that you can effectively "glue" together the topological summaries of smaller parts to reconstruct the summary of a larger, complex whole.

What Problem Does This Paper Address?

In fields like computer vision, medical imaging, and evolutionary biology, we often need to compare shapes. Traditional methods usually fall into two camps: landmark-based (placing dots on a shape) or diffeomorphism-based (warping one shape into another).

The fatal flaw of these traditional approaches is that they typically assume all shapes in a dataset have the same topology. If you are comparing a normal brain artery to one with an extra branch (a change in topology), these models break down. This paper seeks an algebraic construction of shape space that can handle shapes with different topologies simultaneously while providing a "local-to-global" property—allowing researchers to analyze a complex object by looking at its simpler pieces.

Methods

The authors combine two powerful mathematical tools:

  1. Persistent Homology Transform (PHT): This summarizes a shape by "scanning" it from every possible direction and recording how its topological features (holes, components) appear and disappear. Crucially, the PHT is injective, meaning no information is lost during this transformation.
  2. Sheaf Theory: Traditionally used in algebraic geometry and topology, sheaves provide a way to track local data and ensure it is consistent when "glued" together.

The core technical contribution is proving that the PHT defines a homotopy sheaf on the poset of constructible sets. This involves showing that the PHT satisfies Čech descent, a generalized version of the inclusion-exclusion principle. They also utilize $\infty$-category theory (in the appendix) to handle the subtle "gluing" issues that arise in derived categories.

Datasets

As a theoretical mathematics paper, it does not use standard machine learning benchmarks. Instead, it uses:

  • Synthetic Geometric Shapes: The authors illustrate their theorems using the letters 'A' and 'V', circles ($S^1$), spheres ($S^2$), and polyhedra.
  • Motivation from Biological Data: The paper cites databases like MorphoSource and Phenome10K (CT scans of skulls) and studies of fruit fly wings as the real-world motivation for needing a framework that handles topological variation.

Key Insights and Ideas

  • From Fiber Bundles to Sheaves: Traditional shape spaces are often modeled as fiber bundles, which require "local triviality" (everything looks the same locally). Sheaves replace this with "local continuity," allowing the "rank" or complexity of a shape to change from point to point.
  • The Nerve Lemma for PHT: The authors prove a "Nerve Lemma," which is a massive computational win. It suggests that you can calculate the high-dimensional topological features of a complex shape using only the 0-th homology (connected components) of its simpler, overlapping pieces.
  • Stability is Key: For any summary to be useful in the real world, it must be stable. The authors prove that if two shapes are "close" (via a controlled homotopy), their PHT summaries are also close.
  • Manifold Approximation: They show that any smooth manifold can be reliably approximated by a polyhedron (a collection of flat faces), and the PHT of that polyhedron will be an accurate proxy for the manifold's PHT.

Key Results

  • Theorem 1.1 (Main Result): Formally establishes that the mapping from a constructible set to its PHT is a homotopy sheaf. This provides the mathematical "permission" to glue topological summaries together.
  • Theorem 1.3 (Stability): Quantifies that the PHT is stable under small perturbations. Specifically, if a homotopy moves points by no more than $\epsilon$, the resulting PHTs are $\epsilon$-close.
  • Computational Efficiency: The framework provides a path toward parallelizing PHT computations by breaking a shape into a cover and processing pieces independently.

Limitations and Future Work

  • The Inverse Problem: While the PHT is injective (it's a "lossless" summary), the paper notes that we still don't have a perfect "oracle" to tell us if a random collection of barcodes actually corresponds to a real shape. Reconstructing the shape from its PHT remains a difficult open problem.
  • Optimal Transport: The authors suggest that the next step is connecting this sheaf-theoretic model to "optimal transport," which would allow for more nuanced "morphing" between shapes in this new space.
  • Algorithmic Implementation: While the theory is robust, translating high-level $\infty$-category theory into efficient code for 3D medical imaging is an ongoing challenge.

Summary updated

Abstract

We present a sheaf-theoretic construction of shape space -- the space of all shapes. We do this by describing a homotopy sheaf on the poset category of constructible sets, where each set is mapped to its Persistent Homology Transform (PHT). Recent results that build on fundamental work of Schapira have shown that this transform is injective, thus making the PHT a good summary object for each shape. Our homotopy sheaf result allows us to "glue" PHTs of different shapes together to build up the PHT of a larger shape. In the case where our shape is a polyhedron we prove a generalized nerve lemma for the PHT. Finally, by re-examining the sampling result of Smale-Niyogi-Weinberger, we show that we can reliably approximate the PHT of a manifold by a polyhedron up to arbitrary precision.
BibTeX
@article{arya2022sheaf,
  author = {Shreya Arya and Justin Curry and Sayan Mukherjee},
  title = {{A Sheaf-Theoretic Construction of Shape Space}},
  journal = {Found. Comput. Math},
  year = {2022},
}
Figure