A note on the estimation of Lp-norms

A note on the estimation of Lp-norms

Statistics & Probability North-Holland Letters 28 September 15 (1992) 131-133 A note on the estimation Michael 1992 of P-norms Weba Institut f...

176KB Sizes 1 Downloads 58 Views

Statistics & Probability North-Holland

Letters

28 September

15 (1992) 131-133

A note on the estimation Michael

1992

of P-norms

Weba

Institut fiir Mathematische Stochastik, Unicersitiit Hamburg, Germany Received

August

1991

Abstract: The Lp’-norm of a random variable with unknown distribution is determined Under a mild moment condition, expansions of bias and variance are established.

Keywords: L”-norm;

statistical

Let X be a random

of the substitution

estimator.

point estimation.

variable

on a probability

Ilxllp=(EIIXI”])“p plays an important problem: we want to of X is unknown or (possibly simulated) same distribution as

by means

space (0,

&, P> with values

in [w. The LP-norm

(l
with error bounds. Consider the following role in stochastics, e.g., in connection make use of II X IIp but an explicit formula is not available because the distribution too complicated. In that case, 11X I(,, must be determined empirically by means of a random sample. Provided X,, X,, . . . , X,, are i.i.d. random variables having the X, the substitution principle suggests to estimate I1X II p by the quantity

iP,n

is the pth root of z,, = (l/n)Cr=, I X, I ‘. Since z?,, is a ‘reasonable’ estimator of E[ I X I “I the problem of computing bias and variance of 6p,n can be viewed as a problem of statistical point estimation. In order to describe bias and variance of t?p,n in the non-trivial case p > 1 - the case p = 1 will be excluded in the sequel - one may try to utilize well-known results such as the theorem on p. 353 of Cramer (1958) or Theorem 5.la on p. 109 of Lehmann (1983). These theorems need some boundedness conditions which are, in general, violated because the function r(t) = t ‘jp, t 2 0, is rather unpleasant: r and all its derivatives are unbounded on (0, m>, and r is not differentiable at the origin. It is the purpose of this note to establish expansions of bias and variance of 6p,n where X is assumed to fulfil a moment condition. The symbol i; will stand for the jth central moment of I X I “, i.e.,

Correspondence to: Michael Hamburg 13, Germany. 0167.7152/92/$05.00

Weba,

Institut

0 1992 - Elsevier

fiir Mathematische

Science

Publishers

Stochastik,

Universitat

B.V. All rights reserved

Hamburg,

Bundesstrasse

55, W-2000

131

Volume

15, Number

2

STATISTICS

& PROBABILITY

28 September

LETTERS

1992

Theorem. Assume that X has the property P(X = 0) < 1 and satisfies the moment condition E[ I x I 4p+a] < m for p > 1 and some real (Y > 0. Then bias and variance of 6p,n admit the representations E[ ip,n] - IIX IIp = al/n

+ az/n2 + o(l/n2)

(1)

and

VX[ Sp,n] = b,/n

+ b,/n’

+ 0(1/n*),

(2)

respectively, where the constants are given by (1 -a) a, = ~ 2P2 a2 =

. 12. [IX Ilj-2p,

(1 -P)(l-2P)

* !tj.

6P3

1 b1= P2

. &.

11X

b =

(1 -PI

2

P3

1j2-2p P

II x II;-“” +

(l-p)(l-2p)(l-3p)

.[*

IIxII,_4p 2'

8P4

P

'



.~3.IIxllp2-3P+

(1-p;f4-5p)

.&

IIx

If-"".

r(t) = t l/p , t > 0, can be represented

Proof. For each fixed C > 0, the function

according

to

r(t)=r(~)+r’(~).(t-~)+~r”(~).(t-5)2+~r”’(5).(t-g)3+~r’iv’(~).(t-~)4 + 4(l, where

t) . (t

- 0”

c#&J, t) satisfies t’el 4(L,

By definition

t> = $(l?

r> = 0.

of r(t), we also have sup I4(L

t)l
I>0

Setting

Zi = I X, I p for i E N, z,, = (l/n)C~zIZj

J!?[(Z~)~‘~] =J1lp + +r”(J)

*

and 5 = EIZil =

Ekzln lj21 +6

l+rr(

+&r(‘“)( 5) . +Zl-l>“] n3 i

++i, In order

to verify (1) it remains lilimn2.E[

132

4(l,

l)

II X II:, one obtains .

+I-

03] n2

3(n-1)(-$~l-i)2])2 +

n3

I

z,)*(Gq4]. to show

.Fn). (Zn -i.)“]

= 0.

(3)

Volume

15, Number

For each It-f\ >6

2

STATISTICS

0, there implies

exists a number

E >

It-51

~ IW> Hence

t)

i

6 > 0 with

28 September

LETTERS

I c#&J,t) 1 < F for

I t - J I < 6. On the other

1992

hand,

a’p i

6

& PROBABILITY

. sup I&L

s) I.

S>O

the relation

SUPl4(57 s)l <&f

S>O

.

p/P

I t -

i

I a’P

is valid for all t a 0, and one can conclude

The inequality

of Marcinkiewicz-Zygmund

E

lz,

-

jiJia”]

=

yields

o(~-~-w’(~P)).

1

(See Chow and Teicher

(1988, p. 3681.1 Therefore,

lim sup n2. IE[~(I,Z,).(Z,,-1)4]I~C.& n~m holds for some constant C 2 0. This proves (3) because E > 0 was arbitrary. Repeating the above considerations with r* instead of r, we obtain an expansion equation (2) follows if this expansion is combined with equation (1). q Under

the assumptions E [ (ip,, -

In the special

of the theorem

the mean

square

error can be represented

of E[(.?fn>2/J’]. Then

by

II x II p,2] = b,/n + (4 + b,)/n2 + 0(1/d).

case p = 2, the leading

coefficient

b, takes the simple

form

b, = f.Var[X*]/E[X*].

References Chow, Y.S. and H. Teicher (1988), Probability (Springer, New York). Cram&, H. (1958), Mathematical Methods of (Princeton, Univ. Press, Princeton, NJ).

Theov

Lehmann, E.L. New York).

(19831,

Theory

of Point

Estimation

(Wiley,

Statistics

133