The bandwidth selection problem in kernel density estimation has a topological solution.
Hyndman, Kandanaarachchi, and Turner (arXiv:2603.22636) update the lookout algorithm for anomaly detection by choosing kernel bandwidth using Rips death diameters — a concept from persistent homology. In standard KDE, bandwidth is a tuning parameter that trades off resolution against noise. Too narrow and every point is its own cluster; too wide and genuine outliers disappear into smoothed background. The usual methods (Silverman's rule, cross-validation) optimize for density estimation accuracy, not anomaly detection.
Note (2026-11-12): checked against the paper's full text. The essay presents the choice of a Rips-death-diameter bandwidth as the update, but the original lookout algorithm (Kandanaarachchi and Hyndman, 2022) already used a bandwidth based on persistent homology; this paper studies the theory of that choice (consistency of the kernel density estimate, with a condition on admissible bandwidths) and proposes a modified lookout with a different bandwidth that is still based on the death diameters, with a multivariate scaling, and reports better performance than the previous version. The bandwidth in the paper is a single scalar (a bandwidth matrix proportional to the identity) read from the dimension-zero barcode at the largest gap between successive death diameters, not a bandwidth that adapts to each region's local connectivity as the essay says. 'The anomalies are the points that die early' is the essay's own phrase; the paper declares anomalies by fitting a generalized Pareto distribution to the smallest leave-one-out density estimates.
Rips death diameters come from a different optimization target entirely. Build a Rips complex — connect points within radius ε, track which topological features (connected components, loops) appear and die as ε grows. The “death diameter” of a connected component is the scale at which it merges with another component. This is the scale at which a cluster stops being isolated. Using these diameters as bandwidth parameters means the kernel width adapts to the local topology of the data — each region gets a bandwidth that reflects its actual connectivity structure.
The theoretical contribution: the resulting KDE is consistent — it converges to the true density as sample size grows. The multivariate scaling approach is both robust and computationally efficient, avoiding the curse of dimensionality that plagues naive multivariate bandwidth selection.
The deeper insight is a pattern: topology provides natural scales that statistical methods need but lack. KDE needs a bandwidth. Persistent homology provides a principled one. The information was always in the data's shape — the death diameters encode exactly the spatial scales that matter for distinguishing structure from noise. The anomalies are the points that die early.