proof of Bennett inequality


By Chernoff-Cramèr inequalityMathworldPlanetmath (http://planetmath.org/ChernoffCramerBound), we have:

Pr{i=1n(Xi-E[Xi])>ε}exp[-supt0(tε-ψ(t))]

where

ψ(t)=i=1n(lnE[etXi]-tE[Xi]).

Keeping in mind that the condition

Pr{|Xi|M}=1 i

implies that, for all i,

E[|Xi|k]Mk k0

(see here (http://planetmath.org/RelationBetweenAlmostSurelyAbsolutelyBoundedRandomVariablesAndTheirAbsoluteMoments) for a proof) and since lnxx-1 x>0, and

E[|X|k]Mk  E[|X|k]E[X2]Mk-2 k2,kN

(see here (http://planetmath.org/AbsoluteMomentsBoundingNecessaryAndSufficientCondition) for a proof), one has:

ψ(t) = i=1n(lnE[etXi]-tE[Xi])
i=1nE[etXi]-tE[Xi]-1
= i=1nE[k=0(tXi)kk!]-tE[Xi]-1
= i=1n(k=0tkE[Xik]k!)-tE[Xi]-1
= i=1n(k=2tkE[Xik]k!)
i=1n(k=2tkE[|Xi|k]k!)
i=1n(k=2tkE[Xi2]Mk-2k!)
= k=2tkMk-2i=1nE[Xi2]k!
= v2M2k=2(tM)kk!
= v2M2[exp(tM)-tM-1]

One can now write

supt0(tε-ψ(t))supt0(tε-v2M2(etM-tM-1))=supt>0[v2M2(M2εv2t-(etM-tM-1))].

By elementary calculus, we obtain the value of t that maximizes the expression in round brackets:

topt=1Mln(1+Mεv2)

which, once plugged into the bound, yields

Pr{i=1n(Xi-E[Xi])>ε}exp[-v2M2((1+Mεv2)ln(1+Mεv2)-Mεv2)].

Observing that (1+x)ln(1+x)-xx2ln(1+x) x0 (see here (http://planetmath.org/ASimpleMethodForComparingRealFunctions)), one gets the sub-optimal yet more easily manageable formula:

Pr{i=1n(Xi-E[Xi])>ε}exp[-ε2Mln(1+εMv2)].
Title proof of Bennett inequality
Canonical name ProofOfBennettInequality
Date of creation 2013-03-22 16:12:27
Last modified on 2013-03-22 16:12:27
Owner Andrea Ambrosio (7332)
Last modified by Andrea Ambrosio (7332)
Numerical id 22
Author Andrea Ambrosio (7332)
Entry type Proof
Classification msc 60E15