Concept

Hoeffding inequality

A bound on how far a sum of independent bounded quantities strays from its average. For n terms each between 0 and 1, the chance of straying by t falls exponentially in t squared over n; it uses only each term's range, so it is simple and often far from sharp.

Named by 3 essays across one field — each of them below, with the objects they name alongside it.

How far each estimate can stray, when rare jumps are possible. The error that the sample mean and the median of 12 block means stay within, at confidence levels from 80% to 99.9%, over 60000 repeated samples of 120 draws from normal draws with rare large jumps, with Chebyshev's guarantee for the mean.

The median of many small averages

Knowing only that a quantity has a finite spread, the plain average of n samples can be promised to within σ/√(nδ) with confidence 1 − δ, and no better — Chebyshev's bound is tight, and rare large jumps achieve it. Cut the same samples into a dozen blocks, average each block, and take the median of the averages, and the promise improves to within about σ√(log(1/δ)/n). Nothing about the data has been assumed beyond the spread; only the way of combining it has changed.

probability · Concentration
Tails of a sample drawn with and without replacement, against two bounds. t 0.000: without 1.00e+0, with 1.00e+0, Hoeffding 1.00e+0, Serfling 1.00e+0; t 0.075: without 2.65e-1, with 3.88e-1, Hoeffding 1.00e+0, Serfling 9.56e-1; t 0.150: without 1.35e-2, with 5.57e-2, Hoeffding 3.31e-1, Serfling 1.05e-1; t 0.225: without 1.10e-4, with 3.02e-3, Hoeffding 3.48e-2, Serfling 2.62e-3; t 0.300: without 1.19e-7, with 8.09e-5, Hoeffding 1.49e-3, Serfling 1.50e-5.

Drawn without putting back

Every concentration bound on this shelf assumes the draws are independent. A real sample is not: a pollster does not ring the same person twice, and every ball taken from an urn changes what is left in it. The dependence runs the helpful way. Wassily Hoeffding proved in 1963 that a sample drawn without replacement is at least as concentrated as one drawn with it, for every convex measure of spread at once — and the variance falls by an exact factor that reaches zero when the whole urn is taken.

probability · Concentration
Tails of a sum of rare events: exact, and three bounds. 10: exact 5.42e-1, Bennett 1.00e+0, Bernstein 1.00e+0, Hoeffding 1.00e+0; 15: exact 8.34e-2, Bennett 3.39e-1, Bernstein 3.42e-1, Hoeffding 9.95e-1; 20: exact 3.44e-3, Bennett 2.09e-2, Bernstein 2.35e-2, Hoeffding 9.80e-1; 25: exact 4.64e-5, Bennett 3.66e-4, Bernstein 5.50e-4, Hoeffding 9.56e-1; 30: exact 2.46e-7, Bennett 2.34e-6, Bernstein 6.10e-6, Hoeffding 9.23e-1; 35: exact 5.88e-10, Bennett 6.45e-9, Bernstein 3.92e-8, Hoeffding 8.82e-1; 40: exact 7.03e-13, Bennett 8.70e-12, Bernstein 1.67e-10, Hoeffding 8.35e-1; 45: exact 4.56e-16, Bennett 6.27e-15, Bernstein 5.21e-13, Hoeffding 7.83e-1; 50: exact 1.71e-19, Bennett 2.59e-18, Bernstein 1.27e-15, Hoeffding 7.26e-1.

Charged for the variance, not the range

Hoeffding's inequality knows one thing about each term of a sum: the interval it lies in. For ten thousand coins that each land heads once in a thousand, that makes it promise almost nothing — a 92% chance of thirty heads, when the truth is two in ten million. Tell the bound each term's variance as well and it changes character: Bernstein's and Bennett's inequalities decay like the normal curve while the deviation is small and like a Poisson tail beyond, and for rare events they are millions of times sharper.

probability · Concentration

Named alongside it

The objects these essays reach for when they reach for this one.

VarianceConcentration inequalityMoment generating functionTail boundBinomial distributionChebyshev inequalityConfidenceConvexityCorrelationEstimatorExtremal exampleHeavy tails

All concepts