Persistence-based topological optimization: a survey

A team of researchers from DATASHAPE, LIGM, and the University of Tokyo has published a seminal survey paper, 'Persistence-based topological optimization: a survey,' on arXiv. The work synthesizes a decade of research on integrating topological data analysis (TDA), specifically persistent homology, into machine learning optimization pipelines. Persistent homology provides a mathematical framework to quantify the shape and structure of complex data—like point clouds, graphs, or images—extracting robust topological features that can serve as powerful priors. The core challenge the survey addresses is making these topological descriptors compatible with gradient-based optimization, enabling their use as regularizers in loss functions to guide models toward solutions with desired structural properties.

This 5,425 KB document serves as both an accessible introduction for newcomers and a technical reference for practitioners, covering theoretical underpinnings and algorithmic implementations. Crucially, the authors have released an accompanying open-source library that implements the various optimization approaches discussed. This provides a practical 'playground' for data scientists and researchers to experiment with topologically-informed losses, lowering the barrier to applying these advanced techniques. The survey showcases practical applications, demonstrating how topological priors can improve model performance and interpretability in fields where data structure is paramount, from computational biology to material science.