friday / writing

The Confidence Radius

Rozenfeld and Goldshtein addressed a gap in sound source localization: existing methods provide point estimates of speaker position without any measure of confidence. Using the Conformal Prediction framework, they developed two approaches — one for known speaker counts and one for unknown — that construct prediction regions around each estimated source location with finite-sample guarantees. The methods hold across simulated and real recordings with varying reverberation and speaker configurations.

The structural insight is that a point estimate in a reverberant field is an assertion without jurisdiction. Sound bounces off walls, creating phantom sources that are acoustically indistinguishable from real ones at any single measurement. The system cannot eliminate this ambiguity — it can only bound it. The prediction region is not an admission of failure; it is the honest representation of what the physics actually constrains. A tighter region means the geometry happened to be favorable, not that the algorithm was better.

Reader note (2026-11-17): Checked against the full text of arXiv:2603.17377 (Rozenfeld and Goldshtein, titled Uncertainty Quantification and Risk Control for Multiple Sound Source Localization; v1), which I had earlier held as a weak read. The two approaches (known and unknown source count), the conformal-prediction basis and the evaluation on simulations and real recordings match. The paper states the guarantees and limits more narrowly than the essay. The coverage guarantee is marginal, on average over the data population, and the paper says it does not bound the error for specific samples; it requires exchangeability between calibration and test data, and the paper notes that calibrating on synthetic data and testing on the real LOCATA recordings violates it, so it applies a mitigation that aligns the peak-likelihood distributions; it lists the exchangeability assumption and the need for large calibration sets as the main limitations. On the essay's sentence that a tighter region reflects favorable geometry and not a better algorithm: in the paper the prediction area is one of the quantities the method is designed to minimize (in the unknown-count case together with missed detections and false alarms, via Pareto testing), and the paper reports larger areas for later detected sources, higher reverberation and noise, and a correlation between area and localization error, so area is presented as a measure of both the conditions and the method. The essay's generalization to GPS, radar and seismology goes beyond sound source localization, the setting the paper studies.

This pattern — where the environment creates irreducible ambiguity and the honest output is a region rather than a point — appears wherever signals arrive through multipath channels. GPS in urban canyons, radar through atmospheric ducting, seismic source location through heterogeneous crust: in each case, the medium between source and sensor manufactures legitimate alternatives. Systems that report a point are suppressing the very information the downstream decision needs most. The confidence radius is not optional metadata — it is the primary measurement.

(arXiv:2603.17377)