Abstract
We define notions of differentiability for maps from and to the space of persistence barcodes. Inspired by the theory of diffeological spaces, the proposed framework uses lifts to the space of ordered barcodes, from which derivatives can be computed. The two derived notions of differentiability (respectively from and to the space of barcodes) combine together naturally to produce a chain rule that enables the use of gradient descent for objective functions factoring through the space of barcodes. We illustrate the versatility of this framework by showing how it can be used to analyze the smoothness of various parametrized families of filtrations arising in topological data analysis.
Introduction
In Topological Data Analysis (TDA), persistence barcodes serve as powerful descriptors of shape and structure. However, their utility in optimization tasks—such as tuning the parameters of a filtration to minimize a loss function—has been hindered by the mathematical nature of the barcode space ($Bar$). Because $Bar$ is a space of multisets where points can emerge or vanish, it does not possess the structure of a smooth manifold. This research establishes a rigorous foundation for differential calculus on barcodes, moving beyond heuristic approaches to a formal framework based on diffeology.
Methodology: Lifts and Diffeology
The core challenge is defining a derivative on a space that lacks a standard smooth structure. The authors address this by "lifting" barcodes to a higher-dimensional ordered space where standard calculus applies.
As shown in Figure 1, a map $B$ from a parameter space $\mathcal{M}$ to $Bar$ is considered differentiable if it can be locally represented by a smooth lift $\tilde{B}$ into the space of ordered barcodes ($\mathbb{R}^{2m} \times \mathbb{R}^n$). By equipping $Bar$ with a diffeology—a generalization of differential geometry for non-manifold spaces—the authors prove a formal chain rule. This allows the differential of a composite loss function $\mathcal{L} \circ B$ to be computed as the product of individual derivatives, even when the intermediate step involves barcodes.
Applications and Results
The framework's versatility is demonstrated across various TDA applications. A key finding is generic differentiability: while barcode maps may not be differentiable everywhere, they are differentiable on an open and dense subset of the parameter space. This ensures that gradient descent algorithms are unlikely to encounter non-differentiable points in practice.
Figure 2 details how this framework applies to diverse tasks, from topological regularizers in image segmentation to shape matching using bottleneck distances. For instance, the differentiability of a Rips filtration depends on having distinct pairwise distances between points. The paper also confirms that popular vectorizations like Persistence Images are infinitely differentiable, provided their weighting functions are smooth.
Conclusion
This work provides the theoretical "missing link" for gradient-based optimization in TDA. By formalizing the relationship between parameter changes and barcode variations, it validates existing heuristic methods and opens the door for more sophisticated topological machine learning models. Future research may extend this to handle "singular" points where differentiability fails, potentially using sub-differential calculus to capture critical topological transitions.
Summary updated