Merging discrete Morse vector fields: a case of stubborn geometric parallelization

TL;DR

This paper explores methods for merging discrete Morse vector fields computed on image patches, addressing challenges in parallel computation and ensuring correct assembly of global fields in discrete Morse theory.

Contribution

It introduces general and specialized merging procedures for discrete Morse vector fields, facilitating parallelization in persistent homology computations.

Findings

Developed general merging algorithms for discrete Morse fields.

Designed efficient merging methods for specific patch coverings.

Ensured correct assembly of global fields from local computations.

Abstract

We address the basic question in discrete Morse theory of combining discrete gradient fields that are partially defined on subsets of the given complex. This is a well-posed question when the discrete gradient field $V$ is generated using a fixed algorithm which has a local nature. One example is ProcessLowerStars, a widely used algorithm for computing persistent homology associated to a grey-scale image in 2D or 3D. While the algorithm for $V$ may be inherently local, being computed within stars of vertices and so embarrassingly parallelizable, in practical use it is natural to want to distribute the computation over patches $P_{i}$, apply the chosen algorithm to compute the fields $V_{i}$ associated to each patch, and then assemble the ambient field $V$ from these. Simply merging the fields from the patches, even when that makes sense, gives a wrong answer. We develop both very general merging procedures and leaner versions designed for specific, easy to arrange covering patterns.