estimator
Citation for published version (APA):
Sander, P. C. (1979). The covariance matrix of a multivariate locally best unbiased estimator. (TH Eindhoven. THE/BDK/ORS, Vakgroep ORS : rapporten; Vol. 7903). Technische Hogeschool Eindhoven.
Document status and date: Published: 01/01/1979 Document Version:
Publisher’s PDF, also known as Version of Record (includes final page, issue and volume numbers) Please check the document version of this publication:
• A submitted manuscript is the version of the article upon submission and before peer-review. There can be important differences between the submitted version and the official published version of record. People interested in the research are advised to contact the author for the final version of the publication, or visit the DOI to the publisher's website.
• The final author version and the galley proof are versions of the publication after peer review.
• The final published version features the final layout of the paper including the volume, issue and page numbers.
Link to publication
General rights
Copyright and moral rights for the publications made accessible in the public portal are retained by the authors and/or other copyright owners and it is a condition of accessing publications that users recognise and abide by the legal requirements associated with these rights. • Users may download and print one copy of any publication from the public portal for the purpose of private study or research. • You may not further distribute the material or use it for any profit-making activity or commercial gain
• You may freely distribute the URL identifying the publication in the public portal.
If the publication is distributed under the terms of Article 25fa of the Dutch Copyright Act, indicated by the “Taverne” license above, please follow below link for the End User Agreement:
www.tue.nl/taverne Take down policy
If you believe that this document breaches copyright please contact us at: openaccess@tue.nl
providing details and we will investigate your claim.
,
ARW
16
BDK
7903
The covariance matrix of a multivariate locally best unbiased estimator
by Peter C. Sander
Eindhoven University of Technology Netherlands
April 1979
The covariance matrix of a multivariate locally best unbiased estimator
Barankin (1949) has given the variance of a univariate locally best unbiased estimator. In this note the covariance matrix is given of a multivariate locally best unbiased estimator. The Cramer-Rao and the Chapman-Robbins-Kiefer bounds appear as special cases.
1. INTRODUCTION
Let X be a random variable defined on a measure space
(X, A).
Let p = {p(.le)le E ~} be a family of probability densities of X withrespect to the sigma finite measure ~; ~ is an index g (e) : = (g 1 (e), .•• , g (e» t wi th g. : ~ -+ lR (i
=
1,m ~
Estimates of gee) are denoted by o(x)
=
(01 (x), .•. ,
set. Define ••• , m). o (x»t with
m
o. : X -+ lR (i= 1, ••• , m) ~-measurable; 0 (X) is the corresponding
~
estimator.
The class of unbiased estimators o (X) of gee) is denoted by Z,
while Z(80), eO
E~,
denotes the subclass of Z with the propertyE(o~(x)leO)
< 00. (i=
1, ... , TIl).~
Definition 1. oO(X) is a locally best unbiased estimator (LBUE) of
g (8) ';n ~
e
O, . ~ f t e o h f 1 1 " ow~ng lS true f or every a=
( aI' ••• , a ) m tE lR m
and
ii) var ([ a.o.(x)le )
m o o
~ var ([ a.o.(x)le ) for alIam
°
E Z.1 ~ ~ 1 ~ ~
o
Let the covariance matrix of O(X) E Z with respect to p(.le) be given by
[o(e) =
«cr~
.(e»). If oO(X) E Z(80), then oO(X) is a LBUE iff1.J m (1) [ m L i,j=l
o
°
ma.a.cr .. (e) for all a E lR , 0 E Z.
i,j=l ~ J ~J
°
F ormu a 1 ( 1) means [
° (
e 0) - [° (
e 0 ) . ~s pos~t~ve sem~-. . . d e f' . ~n~te, i.e.°
LO (eO) ~ [o(eO).
The following results are of immediate use (cf. Rao (1973) pp. 317-318).
AMS 1970 subject classifications. Primary 62FlO; Secondary 62H99.
Key words and phrases. Locally best unbiased estimator, covariance matrix, Cramer-Rao bound.
Theorem 1.
°
Then 0 (X)
°
cov(n (X),
Assume m = 1 and oO(X) E z(eO).
is a LBUE of g(e) in eO iff
o
(X) leO) =°
for all 6(X) E Z(eO). Corollary 1i) 6(X) is a LBUE of gee) in eO iff 0i(X) is a LBUE of gi(e) in eO (i = I, ... , m)
°
ii) if there are two LBUEs of g(e) in e , they are the same except for
°
a set of p(.le ) measure zero.
o
From this corollary it follows that LBUEs can be constructed co-ordinate wise, following Barankin (1949).
2. LINEAR TRANSFORMATIONS
2 .
Let L (X, v) be the family of all square integrable real valued functions with respect to a a-finite measure v on X, with inner product
and norm (f, g) :=
f
f g X Ilfll = (f, f)!. 2 d v for f, gEL (X, v),By Riesz'representation theorem with every bounded linear functional H
O L
2
(X, v) + lR there corresponds an adjoint function hO E L2(X, v),
such that
(2) HO(g) =
f
g hO dv, for all g E L2(X, v).X
Let H(g) := (HI (g), ••• , Hm (g» t with Hi (i = I, ••• , m) a bounded linear, functional with adjoint hi E L2(X; v). Then we will call H : L2(X, v) +
lR~
a bounded linear functional as well, with adjoint Ht
:=
(hi' ••• , hm).~dr:
.,~
'.,
.l
2 m
Lemma 1. Let H : L (X, v) + 1R be a bounded linear functional with d · · t ( h ) a JOLnt H
=
hI' ... , m' Define S :=
«
f
h. h . dv ) )~
. I L J L ,J= X Thenii) if there is an m x m matrix A with
then
A ~ S.
Proof
i) Let a € 1Rm. By Schwarz' inequality
(4) Hence a
ts
aJ
Ix12
dv~
X H). I f (3) then D: a.f
xh. dv]2 for all x E L2(X, v) • L l. Xfor all x E L2(X, v),x =f 0 and all a E 1Rm.
From (4) it follows that for an arbitrary but fixed a E 1Rm and
m x
=
1::i=l
a.h.
II
x112
ats
a = a H(x)[H(x) ] a. t t Hence, for this choice of x,(6) a A a - a S a t t ~ O.
Since (6) is true for all a ~ lRm the proof is complete.
D
Notes
I. Since S
=
«
f
h.h.dv»~
. I' it is logical to denote S by HHt. ~ J ~ ,J=X
This will be done in what remains.
2. If a matrix S satisfies i) and ii) of lemma 1, the following notation will be used
D
2 .1
Let K be a linear subspace of L
(X,
v) and K the orthogonal complement of K in L 2 (X, v). Let F : K -+ lR m be a bounded linear functional.Define the bounded linear extension F of F by
e
(7)
Lemma 2. Let K be a linear subspace of L2(X, v) and let F
a bounded linear functional with extension F .
m K -+ lR be e Then Proof By lemma
Especially
Assume there is an m x m matrix T such that
Then
and hence, since Ilxll = Ilx} + x211
~
Ilx}11 (Xl E K, x2 E Kl.),The result follows from lemma 1. D
3. THE COVARIANCE MATRIX OF A LBUE Define the measure v by
v(A)
=
f
p(.leO)d~
for all A EA.A Assume from now on:
I. p(xle)/p(xleO) 18 defined
v -
a.e. for all e E Q and this quotient is denoted by TI(xle)II.TI(.ls) E L2(X, v) for all
e
E Q •2
Define K as the linear subspace of L (X, v) spanned by finite linear combinations of the TI(.le), e E Q.
Lemma 3. Let F : {~(.Ie) Ie € Q} + IRm be defined by F(~(.le» ; gee).
Then the propositions i) and ii) are equivalent:
i) there is a bounded linear transformation A L 2 (X, \) + IR m such that A(n(.le»; F(n(.IS»
ii) there is a positive semidefinite matrix M such that
(8)
for all n € :IN, a = (aI' •.• , an) E IR nand e 1 ,
...
,
n .
Furthermore, let z ; E a.~(.le1) and define ~ 1 1 m K + IR by (9) ~ (z) n . := E a.F(~(.le1». ] 1
iii) I f ii), then
e
is well-defined, Furthermore MO (8) and M ;::: MO' Proof i) ~ ii). Choose M=
AAt, 'V n 'VI
"~J' iii) Assume z = E a. ~(,6 )
] J 'V and z = z. Then by (8)hence ~ is well-defined. Since ~ is a bounded linear transformation the results follow from lemma 2,
ii) ~ i). ~ meets the requirements of i),
e
o
We are now able to give the covariance matrix of a multivariate locally best unbiased estimator.
Theorem 2. Let ¢, ¢e and M
O be defined as in lemma 3. Then i) Z(¢O) f
0
iff(10)
f or a 11 n E IN, a = ( aI' ••• , an ) E lR, n
e
1 , ••• , e n E Q.°
0°
ii) I f 8 (x) E Z (e ) then I (e ) ~ HO.
iii)Z(eo) f
0
iff there is one and only one (p(.leo)a~e.)
°
0 80°
8 (x) E z(e ) with. I (e)
=
MO. Proofi) Assume 8(x) E. Z(e ) (the "only if" part). Then,
o
.
~n an obviousnotation, Hence (11 ) n . ~ I a.g(e ) 1 ~
=
J
X n . 8 I a·7f(.le~)d\). 1 ~ t n i u i tu (I a.g(e »(I a.g(e » u
1 ~ ~
=
utef
8f
ai7f(·le~)d\»(
f
X X n . 2III
a·7f(.le~)11 1 ~From lemma 3 it follows that (10) ~s true for M = MO'
Now the "if" part. By Riesz' representation theorem there is a 0° := (0°1' ... , oO)t with o? E L2(X, v) (i = I, ..• , m) such that
m ~
(I 2) <p (f)
e
Since <p (TI(.IS»
=
g(S) from (12) it follows that oO(X) E Z(SO).e
ii) This follows immediately from (II) and lemma 3.
iii) If Z(SO) f
~
then 0° defined by (12) meets the requirements. The unicity is a result of corollary I.4. THE CRAMER-RAO BOUND
The Cramer-Rao bound is a special case of theorem 2. This can be seen from what follows.
Theorem 3. Assume i) Z(SO) f ~
ii)
n
c lR k, Q is openiii) p(xIS) f
°
~-ae for all SEQD
t t
iv) if S
=
(SI' ••• , Sk) E Q, S(i,h):= (SI' ... , Si-I' Si+h , Si+I' ' ' ' , Sk)v) vi)
and h, a I' .•. , ak E lR, then
lim
f {
~
a. (TI(.IS) - TI(.IS(i, h»)/h}2 dv h+OX I ~=
f· {
lim~
a. (TI(.IS) - TI(.IS(i, h»)/h}2 dvX h+O I ~ M
O is positive definite
F :=
«
f
(oTI(.ls)/os.)(aTI(.ls)/oS.)dv»kand
G :=
are well-defined and F is non singular. Then
Proof. By theorem 2
for all n E IN, a
=
(aJ, Especially for a and
e
l ,(13) sup
m
UEm
Note that there is an
...
,
...
,
a)E m
n ,e
1 , ... , enEn,
uE mm.
n en fixed t m x m matrix B such that MO=
B B.Define n . 1 V := 2: a.g(e ) 1 1 P := (B-1)t v g := Bu.
From
it follows that
hence
(14) sup utvvtu/(utMOU) vtM~1 v.
u
From (13) and (14) it follows that
n i t -1 (2: a.g(e» MO 1 ~ n . (2: a.g(8~» ::::; 1 ~
for. all a = (aI' •• " a ) E lR n . and e 1 ,
n
In particular
...
,
e
n r.E .. ,
k 0 0 . t -J k 0 0
(l; a.[g(6 ) - g(e (1,h»J/h) MO (.2: ai[g(6 ) - g(e (i,h»J/h)
i=l ~ . . ~=I
::::; II
Hence with iv)
k
15) (2: a.
1 ~
t
=
a F aLet I be the identity matrix of order n, and B an arbitrary m x k
n
From (15) it follows that
Gt M-I
o
G :s:;, Fhence
The result now follows from (16).
5. FINAL REMARKS
In lemma 3 MO
~s
defined by MO =~e~et,
in other words(17) sup
n,a,€l
It is easily seen that (17) is equivalent to
(18)
with
t [E
1g(V)-E2g(V)][EIg(V)-E2g(V)]
MO = sup, "'I""'2
f { f
p (xI
v)d["'l (v) -"'2 (v) J) 2/p
(xl eO)d" (x) X Q E.g(V) 1.=
f
G g(v) d w. (v) ~o
and in which the supremum 1.S taken over all measures wI and w
2 with wI ~ w
2 and such that the integrand in the denumerator is defined
lJ - a.e . •
(g(eo)-E g(V»(g(eO)-E g(V»t 8
°
o o · 00 E (8 ) ~ sup ----..;:~---;::"--....;~--2 00 E p(XIV) E [_00...----,,:_ ] eO p(xleO) - 1This last result is a logical generalization of the well-known one-dimensional Chapman-Robbins-Kiefer theorem (cf. Chapman and Robbins (1951) and Kiefer (1952».
REFERENCES
[IJ Barankin, E.W. (1949). Locally best unbiased estimates, Ann. Math. Statist. 20 477-501
[2J Chapman, D.G. and Robbins, H. (1951). Minimum variance estimation without regularity assumptions.
Ann. Math. Statist. 22 581-586
[3J Kiefer, J. (1952). On minimum variance estimators.
Ann. Math. Statist. 23 627-632
[4J Rao, C.R. (1973). Linear statistical inference and its applications (second edition). Wiley, New York.
Department of Industrial Engineering Eindhoven University of Technology 5600 MB Eindhoven