Accurate Evaluation of Quickest Changepoint Detectors via Censoring-Aware Estimators — reproduction

OpenReview LhGxRnGmGJ

ClaimVerdictHeadline
1 bias decays exponentially in Nfalsifiedpower law, slope −0.92 (R²0.90) vs exponential R²0.51
2 KM-ADD bias boundreproduced0.031% relative bias at N=6,400
3 truncation bias < conventionalreproduced~100× smaller on 3 detection laws
4 asymptotically unbiased if support ≤ boundaryreproduced−0.003 within vs −22.4 beyond, at any N
5 reduced variance vs LBfalsifiedKM variance is 2.4× higher; RMSE 46× better
6 accurate from 10% to 90% censoringfalsified0.85% at 48% censoring, 24% at 87%