Formula Lab

Same evidence, nine ways to decide how much of it to trust. Wassily's jar uses the Hoeffding bound; the other eight are here to compare against it.

Wassily uses · Hoeffding

Evidence

Shared with the proof panel on Overview

2,346
0.580
28
30.8%
n₊ 723 · n₋ 1,623

Proven floor vs tokens sampled

AUC 0.580 · d 28 · survivors 30.8% · hover for values

0.400.450.500.550.600.651205001k2.5k5k10k24kn (log)target 0.60measured 0.580Hoeffding

All bounds, ranked by floor

n = 2,346 · AUC 0.580 · δ = 0.05 · highlighted: Wassily's formula

FormulaCreditεFloorJar
Wilcoxon
Wilcoxon–Mann–Whitney normal interval
Frank Wilcoxon; Mann & Whitney; Hanley & McNeil
1945 / 1947 / 1982
0.02130.5587
59%
Bayes
Bayes–Laplace posterior credible floor
Thomas Bayes; Pierre-Simon Laplace
1763 / 1774
0.02150.5585
59%
HoeffdingWassily
Hoeffding bound for U-statistics
Wassily Hoeffding
1963
0.04550.5345
34%
Bernstein
Bernstein inequality
Sergei Bernstein
1924
0.04770.5323
32%
Chebyshev
Chebyshev–Cantelli inequality
Pafnuty Chebyshev; Francesco Cantelli
1867 / 1928
0.05650.5235
23%
Kolmogorov (DKW)
Kolmogorov–Smirnov / DKW band
Andrey Kolmogorov; Dvoretzky–Kiefer–Wolfowitz; Massart
1933 / 1956 / 1990
0.09180.4882
0%
VC bound
Vapnik–Chervonenkis capacity bound
Vladimir Vapnik & Alexey Chervonenkis
1971
0.27370.3063
0%
VC ∧ bootstrap
min(VC floor, bootstrap floor)
Vapnik–Chervonenkis + Bradley Efron
1971 / 1979
Bootstrap
Percentile bootstrap
Bradley Efron
1979