Abstract
We present a robust statistical framework for estimating the number of local maxima (modes) in noisy one-dimensional signals. By introducing "Kolmogorov signatures," we bridge the gap between Topological Data Analysis (TDA) and classical statistical inference. Our approach utilizes the Kolmogorov norm and "taut string" algorithms to provide mode estimates with statistically controllable confidence intervals, eliminating the need for problematic data presmoothing.
Introduction
The estimation of modes in a signal corrupted by noise is a cornerstone of statistical signal processing. While persistent homology in TDA offers a topological approach via barcodes, its reliance on the supremum ($L^\infty$) norm makes it susceptible to outliers and Gaussian noise. Conventional solutions involving kernel smoothing suffer from the bandwidth selection problem, where the choice of smoothing parameter can either obscure true modes or fail to filter noise. This paper proposes a method that provides statistical guarantees without arbitrary preprocessing.
Methodology
Our approach redefines persistence through the lens of metric geometry. We define a metric signature as the distance from a noisy signal to the nearest function with at most $k$ modes.
The Kolmogorov Metric
Instead of comparing functions directly, we compare their antiderivatives. The Kolmogorov distance naturally damps high-frequency noise, providing a more robust metric than the standard sup norm.
Taut Strings and Optimization
To compute these distances efficiently, we utilize "taut strings"—a concept from total variation minimization. By pulling a string tight through a "tube" of radius $\alpha$ around the signal's antiderivative, we find the function that minimizes the number of modes within that distance. We provide an $O(n \log n)$ algorithm for this computation.
Persistence as Distance
A key theoretical contribution is the formalization of persistence values as distances to sets of functions with fewer modes. This allows us to apply tools from metric geometry to TDA.
Statistical Guarantees and Results
We derive a universal threshold, $\tau_n(\alpha)$, allowing for the construction of "honest" confidence regions for the number of modes. Our method provides non-asymptotic exponential deviation inequalities, ensuring reliability for finite sample sizes. Simulation results on standard "blocks" and "bumps" benchmarks demonstrate that Kolmogorov signatures clearly distinguish significant modes from noise, even in challenging environments.
Discussion
While the Kolmogorov norm excels at detecting broad, low-intensity signals, it is less sensitive to narrow, high-amplitude "needles" than the sup norm. Furthermore, while the statistical theory is generalizable, the current $O(n \log n)$ computational efficiency is restricted to 1D signals. Future work will explore hybrid metrics and extensions to higher-dimensional data.
Conclusion
Kolmogorov signatures offer a statistically rigorous and computationally efficient alternative to traditional persistence for mode detection. By treating persistence as a geometric distance, we provide a framework that avoids the pitfalls of manual smoothing and offers clear statistical confidence.
Summary updated