Concentration inequalities for order statistics
Using the entropy method and Rényi’s representation
Maud Thomas
1in collaboration with Stéphane Boucheron1 1LPMA Université Paris-Diderot
High Dimensional Probability VII Cargèse, May 26 - 30, 2014
Background: order statistics
Sample: X1, . . . ,Xn„i.i.d.F.Order statistics
Xp1qě. . .ěXpnq non-increasing rearrangement ofX1, . . . ,Xn. Xp1q: sample maximum. Xpn{2q: sample median. @k,PtXpkqďtu “ řn i“k `n i ˘ Fiptqp1 ´Fptqqn´i.Classical statistic theoryandExtreme Value Theoryprovide: Asymptotic distributions.
Convergence of moments.
Goal
Background: concentration
Concentration of measure phenomenon
Any function of many independent random variables that does not
depend too much on any of them is concentrated around its mean
value.
Example: Gaussian concentration
X a standard Gaussian vector andZ “fpXq. Poincaré’s inequality: VarrZs ďE}∇f}2. Gross logarithmic Sobolev inequality: EntrZ2
s ď2E}∇f}2. Cirelson’s inequality: PtZ ěEZ `tu ďexpp´t2
{p2L2
qqif
Gaussian case and the Poincaré’s inequality
fpX1, . . . ,Xnq “Xpkq
the rankk order statistic of a sample is a simple function ofn independent random variables.
||∇f|| “1.
Xi are standard Gaussian.
Poincaré’s inequalityñVarrXpkqs ď1.
Extreme Value TheoryñVarrXp1qs “Op1{lognq.
Classical statistic theoryñVarrXpn{2qs “Op1{nq.
We do not understand (clearly)
Order statistics and spacings
Proposition (Boucheron, T. (2012))
For all0ăk ďn{2 VarrXpkqs ďkE ”` Xpkq´Xpk`1q ˘2ı “kEr∆k2s. For allλPR, Ent“eλXpkq‰ :“ λErX pkqeλXpkqs ´EreλXpkqslogEreλXpkqs ď kE“eλXpk`1qψpλpX
pkq´Xpk`1qqq
‰
“ kE“eλXpk`1qψpλ∆kq‰ withψpxq “1` px´1qex.
Remarks
Vk :“k∆2kis called the Efron-Stein estimate of the variance ofXpkq.
Without any assumption such as:
F belongs to themax-domain of attraction of an extreme value distributionG,i.e
lim
nÑ`8F n
panx`bnq “Gpxq
for every continuity pointx ofG.
pXpkqqis a sequence of
extremeorder statistics, ifk fixed,nÑ 8;
centralorder statistics, ifk{nÑpP p0,1qwhile,nÑ 8;
Proof
Efron-Stein inequality (Efron, Stein (1981))
Letf:RnÑRbe measurable, and letZ “fpX1, . . . ,Xnq.
LetZi“fipX1, . . . ,Xi´1,Xi`1, . . . ,Xnqwhere fi:Rn´1ÑRis an arbitrary measurable function.
SupposeZ is square-integrable, then:
VarrZs ďE « n ÿ i“1 pZ´Ziq2 ff .
Modified logarithmic Sobolev inequality (Wu(2000); Massart
(2000))
Letτpxq “ex
´x´1. With the same notations, for any λPR,
Ent“eλZ‰ďE « n ÿ i“1 eλZτp´λpZ´Ziqq ff .
Graphical assessment
+ + + + + + + + + + 1 2 3 4 5Ratio between the Efron-Stein estimate and the variance of the maximum ofn independent Gaussian random variables. n“2p for
p“1, . . . ,10. The asymptote is the line
Rényi’s representation
The order statistics of an exponential sample are distributed as partial sums of independentexponentially distributed random variables.
Rényi’s representation (Rényi (1953))
LetYp1qěYp2qě. . .ěYpnq be the order statistics of an independent
sample of the standard exponential distribution, then
` Ypnq, . . . ,Ypiq, . . . ,Yp1q ˘ „`Enn, . . . , n ÿ k“i Ek k, . . . , n ÿ k“1 Ek k ˘
Quantile transformation
Definition (Quantile function)
FÐppq “inftx:Fpxq ěpu,pP p0,1q .
Notation
Uptq “FÐp1
´1{tq,t P p1,8q.
Representation for order statistics
IfYp1qě. . .ěYpnqare the order statistics of an exponential sample, then
pU˝expqpYp1qq ě. . .ěpU˝expqpYpnqq
Hazard rate, spacings and order statistics
Definition (Hazard rate)
The hazard ratehof a differentiable distribution functionF is defined as:
h“F1{F
“F1{p1
´Fq .
Lemma
The distribution function F has non-decreasing hazard rate h, iffU˝exp
is concave. Indeed,
pU˝expq1“ 1 hpU˝expq .
If the distribution is log-concave, then the associated hazard rate is non-decreasing.
Variance bound for order statistics when the
hazard rate is non-decreasing
Recall: Vk “k∆2k.
Proposition (Boucheron, T. (2012))
If F hasnon-decreasing hazard rateh, then for1ďkďn{2,
Var“Xpkq ‰ ďEVk ď 2 kE „ ´ 1 hpXpk`1qq ¯2 .
Towards an exponential Efron-Stein
inequality
Definition (Exponential Efron-Stein inequality)
LetZ “fpX1, . . . ,XnqwhereX1, . . . ,Xn are independent random
variables andV its Efron-Stein estimate of the variance ofZ.
Z satisfies anexponential Efron-Stein inequalityif for allθ,λą0 such thatλθă1 andE“eλV{θ‰ă 8: logE ” eλpZ´EZqı ď λθ 1´λθlogE ” eλV{θı .
Problem
For an exponential sample,E“eλV{艓 8. ëFind another decoupling inequality.
Decoupling inequality: negative association
Negative association
X andY are negatively associated if for any non-decreasing functionsf,g
ErfpXqgpYqs ďErfpXqsErgpYqs.
Lemma
If the distribution function F has non-decreasing hazard rate, then Xpk`1q
Exponential Efron-Stein inequality for order
statistics
Proposition (Boucheron, T. (2012))
If F hasnon-decreasing hazard rateh, then forλě0, and1ďk ďn{2,logEeλpXpkq´EXpkqq ď λk 2E “ ∆k ` eλ∆k´1˘‰ “ λk 2E «c Vk k ´ eλ?Vk{k´1 ¯ ff .
Gaussian hazard rate
Uptq “ΦÐp1´1{tqfortą1. 0 2 4 6 8 10 −1 0 1 2 3 4 U(e xp(x)) Gaussian distribution 0 2 4 6 8 10 0 1 2 3 4 U(2 e xp(x)) Absolute value of Gaussian random variableVariance of absolute values of Gaussian
random variables
Proposition (Boucheron, T. (2012))
Let ně3, let Xpkqbe the rank k order statistic of absolute values of n
standard independent Gaussian random variables, VarrXpkqs ď 1 klog 2 8 log`2n k ˘
´logp1`k4log log`2n k
˘ q .
For the maximum (k “1), the bound becomes: 1
log 2
8
log 2n´logp1`4 log log 2nq .
Mn : maximum ofnstandard Gaussian r.v
Chatterjee (Talagrand L1-L2 inequality): VarrMns ď 1`log1 n.
Bernstein inequality for the maximum of
absolute values of Gaussian random
variables
Theorem (Boucheron, T. (2012))
For n such that the solution vn of equation16{x`logp1`2{x`4 logp4{xqq “logp2nq
is smaller than 1, for all0ďλă?1 vn, logEeλpXp1q´EXp1qqď vnλ 2 2p1´?vnλq .