On the Saddlepoint Approximation for the Residual Sum of Squares in Special Linear Heteroscedastic Models
Viktor Sn´ıˇzek
Examensarbete i matematisk statistik, 20 po¨ang Handledare: Silvelyn Zwanzig
Maj 2003
Matematiska institutionen
Uppsala universitet
Contents
1 Abstract 4
2 Introduction 5
3 Thanks 6
4 Linear model 7
4.1 Heteroscedastic linear model . . . 7
4.2 Saddlepoint computation in a special case . . . 10
4.3 Homoscedastic linear model . . . 14
4.4 General p-sample regression model . . . 16
4.4.1 Special p-sample case . . . 18
4.4.2 Special two-sample case . . . 20
5 Simulations 22 5.1 Setting the amount of simulated data set . . . 23
5.2 The SA function . . . 23
5.3 Simulation in homoscedastic case . . . 25
5.3.1 Small sample size and low number of repetitions . . . 25
5.3.2 Different sample size with the same number of repeti- tions . . . 27
5.4 Simulations in heteroscedastic case . . . 29
5.4.1 Small sample size and also low number of repetitions . 29 5.4.2 Sufficient amount of repetitions with different sample size . . . 31
2
CONTENTS 3
5.4.3 Big difference between variances . . . 33
5.4.4 Small difference between variances . . . 36
5.5 Summarizing . . . 37
6 Conclusions 38 6.1 Outlook . . . 39
A R procedures 40 A.1 Rsim . . . 40
A.2 SA . . . 40
A.3 sim . . . 41
A.4 Trunsim . . . 42
A.5 Convolute . . . 43
A.6 simdiff . . . 44
Abstract
There are several aims of this work. First some simplifications of the general saddlepoint approximation have to be done. Secondly the conditions for the explicit solution of the saddlepoint equations are discussed and finally the simulations are performed and compared to the approximations.
4
Chapter 2
Introduction
This paper is based on results of Saddlepoint approximation (SA) of RSS published in Journal of Multivariate Analysis by Alexander V. Ivanov and Silvelyn Zwanzig under the title Saddlepoint Expansions in Linear Regression.
The main aim is to check out the saddlepoint approximation. It is nec- essary to find simplifications of the theoretical SA for some special models and to proof the theoretical formulas by simulations.
The solution of the saddlepoint equations is explicitly known only for the homoscedastic model. In this case it’s also known that the RSS is χ2(n − p) distributed. We are interested in finding the solution of the sad- dlepoint equations in the heteroscedastic model. The results are summed up in Theorem 2 where necessary conditions are presented. Under these conditions the solution of the saddlepoint equations can be explicitly solved for p components, only one component is expressed implicitly. This essen- tially simplify the approximation formulas in the paper of A. V. Ivanov and S. Zwanzig. In Theorem 3 is shown that p-sample model fulfills conditions of Theorem 2 if the weights and variances in each sample are equal. But weights and variances in different samples can be different.
The second part of the paper will give us the answer if the theoretical research is on the correct way. The simulations use the 2-sample regression linear model with different starting conditions. After the series of simula- tions we discovered some practical problems with the saddlepoint approxi- mation that we should be aware of. Such problems can appear when we have just a small sample size of the data. The SA function gives big functional values for increasing negative values. Nevertheless this is not the problem of the theory but the computation. The simulation chapter shows ways how to solve it. Summarizing, I found the saddlepoint approximation as a really good approximation for the 2-sample case.
5
Thanks
This paper was written during my study stay at the Uppsala university at the Mathematical institution. I would like to express my warmest thanks to Silvelyn Zwanzig who advised me the whole year and the co-operation with her has always taken place in a very friendly and nice atmosphere.
The stay was only possible thanks to the Svenska institutet’s scholarship and the co-operation between it and the Mathematical institution namely Allan Gut who become my contact professor.
Thanks to J¨org Polzehl, Weierstraß Institute for Applied Analysis and Stochastics, Berlin, for the R procedure convolute and for advice with the kernel density estimator.
6
Chapter 4
Linear model
4.1 Heteroscedastic linear model
Consider the heteroscedastic linear regression model as:
yi,n= xTi,nβ + εi,n, (4.1) where
εi,n= σi,n²i, ²i ∼ N (0, 1) i.i.d. (4.2)
yi,n are observations. The n × p design matrix is given by X =
xT1
... xTn
.
β is a p × 1 vector of fixed unknown parameters.
Further we suppose W = (wij) to be a known symmetric diagonal n × n matrix of weights, where 0 < wmin< wi,n< wmax.
We should also have the variances bounded using the condition:
wi,nσi,n2 = 1 + ∆2i,n, (4.3) where max
1≤i≤n∆2i,n≤ ∆2 < ∞.
In the homoscedastic case we are using the simplified form:
wi,nσi,n2 = 1 + ∆2n, (4.4) with ∆2n≤ ∆2< ∞.
Let ¯σn2 = n1 Pn
i=1
wi,nσ2i,n.
7
Let’s present two regularity conditions:
There exist constants λx> 0 and cx< ∞ such that limn→∞λmin
à 1 n
Xn i=1
xi,nxTi,n
!
> λx, (4.5)
and limn→∞max
i≤n kxikmax< cx. (4.6) Under (4.5) and (4.6) there exist a positive constant λ0 = λxwmin and a bounded constant c0= cxwmax such that
λminI(n)> λ0 and tr(I(n)) < c0, where
I(n)= 1
nXTW X = 1 n
Xn i=1
wi,nxi,nxTi,n (4.7) is a weighted Fisher matrix.
The weighted residual sum of squares is:
RSS = min
β
Xn i=1
wi,n(yi,n− xTi,nβ)2.
Theorem 1. Under assumptions of (4.2), (4.5), (4.6) it holds
P µ√
n µ1
nRSS − ¯σ2n
¶
> u
¶
= Z∞ u
Z
fhet(y)(1 + Rn(y))dy2dy1
with y = µ y1
y2
¶
∈ Rp+1, y1∈ R, y2 ∈ Rp and with fhet(y) given in (4.11);
supy∈Y|Rn(y)| = O(n−1) for Y = n
y : kˆθ(y)k ≤ c o
, where ˆθ(y) ∈ Rp+1 is the exact solution of (4.18)
Proof. The proof of this theorem is shown in [1] (p. 203). ¥ Lemma 1. Let
Si(θ1) = −1
2ln(1 − 2wi,nσi,n2 θ1).
Then
si(θ1) = wi,nσi,n2
1 − 2wi,nσ2i,nθ1 = Si0(θ1), (4.8) s0i(θ1) = 2s2i(θ1), (4.9) s00i(θ1) = 8s3i(θ1). (4.10) The prime denotes the derivation with respect to θ1.
4.1. HETEROSCEDASTIC LINEAR MODEL 9
Proof.
Si0(θ1) = ∂
∂θ1 µ
−1
2ln(1 − 2wi,nσi,n2 θ1)
¶
=
= −1 2
1
1 − 2wi,nσi,n2 θ1(−2wi,nσ2i,n) = si(θ1), s0i(θ1) = ∂
∂θ1
à wi,nσ2i,n 1 − 2wi,nσ2i,nθ1
!
=
= −wi,nσi,n2
(1 − 2wi,nσi,n2 θ1)2(−2wi,nσi,n2 ) = 2s2i(θ1).
¥ The saddlepoint approximation density looks like:
fhet(y) = exp (κ(ˆθ) −√
nˆθTF−1(y))
(2π)p+12 |Chet(ˆθ)|12 , (4.11) where
κhet(θ) = −n¯σ2nθ1+ Xn i=1
Si(θ1) + nθ2TI(n)−1B(θ1)I(n)−1θ2, (4.12)
with θ = µ θ1
θ2
¶
∈ Rp+1, θ1 ∈ R, θ2 ∈ Rp.
Si(θ1), si(θ1) and s0i(θ1) are defined in Lemma 1.
B(θ1) = 1 n
Xn i=1
si(θ1)
2 wi,nxi,nxTi,n, (4.13)
s(θ1) = 1 n
Xn i=1
si(θ1). (4.14)
F−1(y) = Ã
y1+ n−12y2TI(n)y2 y2
!
. (4.15)
Chet(θ) =
à s0(θ1) + θ2TI(n)−1B00(θ1)I(n)−1θ2 2θT2I(n)−1B0(θ1)I(n)−1 2I(n)−1B0(θ1)I(n)−1θ2 2I(n)−1B(θ1)I(n)−1
!
. (4.16)
The saddlepoint ˆθ = ˆθ(y) ∈ Rp+1 is defined as the exact solution of the equation system
mhet(ˆθ) = F−1(y), (4.17)
where
mhet(ˆθ) =√ n
à −¯σ2n+ s(ˆθ1) + ˆθ2TI(n)−1B0(ˆθ1)I(n)−1θˆ2 2I(n)−1B(ˆθ1)I(n)−1θˆ2.
!
. (4.18)
To summarize all the previous formulas connected to the saddlepoint density in (4.11) we have:
exp (κ(ˆθ) −√
nˆθTF−1(y)) = exp(−n¯σn2θˆ1+Pn
i=1
Si(ˆθ1) + nˆθT2I(n)−1B(ˆθ1)I(n)−1θˆ2
−n
³θˆ1(−¯σn2+ s(ˆθ1) + ˆθ2TI(n)−1B0(ˆθ1)I(n)−1θˆ2) + 2ˆθT2I(n)−1B(ˆθ1)I(n)−1θˆ2
´ ) =
= exp(Pn
i=1
Si(ˆθ1) − nˆθ1s(ˆθ1) − nˆθ2TI(n)−1
³
B(ˆθ1) + ˆθ1B0(ˆθ1)
´
I(n)−1θˆ2) (4.19) and using the following formula for the determinant calculation of the matrix divided into segments
|C| = |C22||C11− C12C22−1C12T| (4.20) for a C matrix C =
µ C11 C12 C12T C22
¶ , we have:
|Chet(ˆθ)| = |2I(n)−1B(ˆθ1)I(n)−1|(s0(ˆθ1) + ˆθT2I(n)−1B00(ˆθ1)I(n)−1θˆ2
−4
2θˆT2I(n)−1B0(ˆθ1)I(n)−1I(n)B−1(ˆθ1)I(n)I(n)−1B0(ˆθ1)I(n)−1θˆ2) =
= |2I(n)−1B(ˆθ1)I(n)−1|(s0(ˆθ1) + ˆθT2I(n)−1
³
B00(ˆθ1) − 2B0(ˆθ1)B−1(ˆθ1)B0(ˆθ1)
´
I(n)−1θˆ2).
(4.21)
4.2 Saddlepoint computation in a special case
In this section we try to solve the system of the saddlepoint equation in (4.17) for the heteroscedastic model and to find out the conditions for finding the explicit solution ˆθ.
Theorem 2. Provided that following conditions are fulfilled 1. B0(θ1) = 4B(θ1)I(n)−1B(θ1) ∀θ1 ∈ Θ,
2. B−1(ˆθ1) = 2I(n)−1ΣI(n)−1− 4ˆθ1I(n)−1, for ˆθ1 solved in (4.17) and Σ inde- pendent of ˆθ
it follows that
4.2. SADDLEPOINT COMPUTATION IN A SPECIAL CASE 11 1. θˆ1: s(ˆθ1) = y1
√n + ¯σn2 (4.22) is the first component (implicit),
2. and
θˆ2 = 1 2√
nI(n)B−1(ˆθ1)I(n)y2 (4.23) is the second, p-dimensional component (explicit) of the solution of (4.17)
3. the saddlepoint density in (4.11) can be split into
fspec(y1, y2) = g(ˆθ1(y1))ϕ0,Σ(y2) (4.24) with
ϕ0,Σ(y2) = exp(−12y2TΣy2)
(2π)p2|Σ|−12 , the density of Np(0, Σ), (4.25) and
g(ˆθ1(y1)) =
exp(Pn
i=1
Si(ˆθ1) − nˆθ1˜s)
(2π|CspecΣ|)12 , (4.26)
where
|Cspec| = s0(ˆθ1)|2I(n)−1B(ˆθ1)I(n)−1|. (4.27) Proof. We will compare left and right side of the system
Ã
y1+ n−12y2TI(n)y2 y2
!
=√ n
à −¯σn2+ s(ˆθ1) + ˆθ2TI(n)−1B0(ˆθ1)I(n)−1θˆ2 2I(n)−1B(ˆθ1)I(n)−1θˆ2.
!
(4.28) and we directly get that
y2 = 2√
nI(n)−1B(ˆθ1)I(n)−1θˆ2, (4.29) equivalently: ˆθ2 = 1
2√
nI(n)B−1(ˆθ1)I(n)y2. (4.30) We’ll put the result in (4.29) into the first term on the left in (4.28) and compare to the first term on the right side:
y1+ 4√
nˆθT2I(n)−1B(ˆθ1)I(n)−1B(ˆθ1)I(n)−1θˆ2 = −√
n¯σn2+√ ns(ˆθ1)
+ √
nˆθT2I(n)−1B0(ˆθ1)I(n)−1θˆ2.(4.31)
Comparison shows that:
y1 = −√
n¯σn2+√
ns(ˆθ1), (4.32) equivalently: s(ˆθ1) = y1
√n+ ¯σn2 = ˜s. (4.33)
Comparing the terms with ˆθ2 inside in (4.31) we get the condition (1) for the solution of this equation system:
B0(ˆθ1) = 4B(ˆθ1)I(n)−1B(ˆθ1). (4.34)
We’ll try to insert this result into the density calculation in (4.19) but we’ll concentrate on the quadratic form with ˆθ2:
nˆθT2I(n)−1
³
B(ˆθ1) + ˆθ1B0(ˆθ1)
´
I(n)−1θˆ2 = nˆθT2I(n)−1
³
B(ˆθ1) + 4ˆθ1B(ˆθ1)I(n)−1B(ˆθ1)
´
I(n)−1θˆ2 =
= 1
4y2TI(n)B−1(ˆθ1)
³
B(ˆθ1) + 4ˆθ1B(ˆθ1)I(n)−1B(ˆθ1)
´
B−1(ˆθ1)I(n)y2 =
= 1 4y2T
³
I(n)B−1(ˆθ1)I(n)+ 4ˆθ1I(n)
´
y2. (4.35)
The computation in (4.35) with the condition (2) will give the following result:
1 4yT2
³
I(n)B−1(ˆθ1)I(n)+ 4ˆθ1I(n)
´ y2= 1
2y2TΣy2.
Now let’s try to compute the determinant |Cspec| in (4.21) using the condition (1) and the result of ˆθ2 in (4.30):
|Cspec| = |2I(n)−1B(ˆθ1)I(n)−1|(s0(ˆθ1) + ˆθ2TI(n)−1
³
B00(ˆθ1) − 2B0(ˆθ1)B−1(ˆθ1)B0(ˆθ1)
´
I(n)−1θˆ2) =
= |2I(n)−1B(ˆθ1)I(n)−1|(s0(ˆθ1) + 1
4ny2TI(n)B−1(ˆθ1)I(n)I(n)−1
³
B00(ˆθ1) − 32B(ˆθ1)I(n)−1B(ˆθ1)B−1(ˆθ1)B(ˆθ1)I(n)−1B(ˆθ1)
´
I(n)−1I(n)B−1(ˆθ1)I(n)y2) =
= |2I(n)−1B(ˆθ1)I(n)−1|(s0(ˆθ1) + 1
4ny2TI(n)B−1(ˆθ1) (4.36)
³
B00(ˆθ1) − 32B(ˆθ1)I(n)−1B(ˆθ1)I(n)−1B(ˆθ1)
´
B−1(ˆθ1)I(n)y2).
We will compute B00(θ1) for any θ1∈ Θ by the per partes method (product
4.2. SADDLEPOINT COMPUTATION IN A SPECIAL CASE 13 rule) and under the condition (1):
B00(θ1) = (B0(θ1))0= [4
³
B(θ1)I(n)−1
´
B(θ1)]0 = 4(B0(θ1)I(n)−1B(θ1) + +B(θ1)I(n)−1B0(θ1)) = 4(4B(θ1)I(n)−1B(θ1)I(n)−1B(θ1) +
+4B(θ1)I(n)−1B(θ1)I(n)−1B(θ1)) = 32B(θ1)I(n)−1B(θ1)I(n)−1B(θ1).
And the Cspec determinant will be:
|Cspec| = s0(ˆθ1)|2I(n)−1B(ˆθ1)I(n)−1|. (4.37)
Summarizing these results we get this equation for the density:
fspec(y) = exp
µ n P
i=1
Si(ˆθ1) − nˆθ1s −˜ 12yT2Σy2
¶
(2π)p+12 |Cspec|12 . (4.38)
We are able to split the density into two parts: first one depending only on ˆθ1 – (4.26) and the second part depending only on y2. The second part has p-dimensional Normal distribution with Σ variance matrix as shown in
(4.25). ¥
Corollary 1. Under the conditions of Theorem 2 we have
P µ√
n µ1
nRSS − ¯σn2
¶
> u
¶
= Z∞ u
g(z) dz + Rest,
where g(z) = g(ˆθ1(y1)) and Rest = R∞
u
Rfspec(y)Rn(y)dy2dy1.
Proof. Hence Theorem 1 we have:
Z
fspec(y1, y2) dy2= g(ˆθ1(y1)) Z
ϕ0,Σ(y2) dy2
| {z }
=1
= g(ˆθ1(y1)).
¥
4.3 Homoscedastic linear model
Now let’s consider the homoscedastic linear model. It means that the model holds the condition in (4.4) and that is why the formulas can be simplified:
¯
σ2n = wi,nσi,n2 , (4.39) s(θ1) = si(θ1) = σ¯2n
1 − 2¯σn2θ1 ∀i = 1, . . . , n, (4.40) B(θ1) = 1
2s(θ1)I(n), (4.41)
B0(θ1) = s2(θ1)I(n), (4.42) B00(θ1) = 4s3(θ1)I(n). (4.43) All the necessary conditions in Theorem 2 should be fulfilled and we’ll prove it:
• Condition (1):
4B(θ1)I(n)−1B(θ1) = 41
4s2(θ1)I(n)I(n)−1I(n)= s2(θ1)I(n)= B0(θ1).
• Condition (2):
B−1(˜θ1) = 2 1
s(˜θ1)I(n)−1= 21 − 2¯σn2θ˜1
¯
σ2n I(n)−1= 2 1
¯
σn2I(n)−1− 4˜θ1I(n)−1 =
= 2I(n)−1ΣI(n)−1− 4˜θ1I(n)−1 with Σ = σ¯12
nI(n).
From Theorem 2 it follows that the saddlepoint equations can be solved and, as we shall see, even explicitly for the first component ˆθ1.
Using (4.40) and the results from Theorem 2 (4.23) and (4.22) we get:
¯ σn2
1 − 2¯σ2nθ˜1 = s(˜θ1) = ˜s = ¯σn2+ 1
√ny1, (4.44)
θ˜1 = 1 2
µ 1
¯ σn2 −1
˜ s
¶
= 1 2¯σn2
y1
√n¯σn2+ y1, (4.45) θ˜2 = 1
2√ n
2
˜
sI(n)I(n)−1I(n)y2 = 1
√n˜sI(n)y2. (4.46)
From (4.27) we have:
|Chom(˜θ)| = 2˜s2|21
2˜sI(n)−1I(n)I(n)−1| = 2˜sp+2|I(n)−1|. (4.47)
4.3. HOMOSCEDASTIC LINEAR MODEL 15 Because of Theorem 2 we are able to split the density
fhom(y) = exp
³
−n2 lnσ¯˜sn2 − n2
³ ˜s
¯ σ2n − 1
´
−2¯σ12
nyT2I(n)y2
´
((2˜s)p+2πp+1|I(n)−1|)12 into two parts:
ϕ(y2) = exp¡
−12yT2Σy2¢
³
(2π¯σ2n)p|I(n)−1|
´1
2
and
g(y1) = exp
³
−n2
³ s˜
¯ σ2n − 1
´´
³ 4π˜s2
³ s˜
¯ σ2n
´p´1
2 ³
¯ σn2
˜ s
´n
2
= exp
³
−n2
³ s˜
¯ σn2 − 1
´´
³ 4π¯σ4n¯σs˜24
n
³ s˜
¯ σ2n
´p´1
2 ³
¯ σ2n
˜ s
´n
2
=
= 1
2¯σn2√ π
µ ˜s
¯ σn2
¶n−p
2 −1
exp µ
−n 2
µ s˜
¯ σ2n − 1
¶¶
. (4.48)
Finally after inserting ˜s in (4.33) into (4.48) we have:
g(y1) = 1 2¯σn2√
π µ
1 + y1
√n¯σ2n
¶n−p
2 −1
exp µ
−
√ny1 2¯σ2n
¶
. (4.49) Let’s compare this formula with the exact distribution.
We know that Y = RSSσ2 ∼ χ2 with n − p degrees of freedom. The linear transformation has to be used.
Y = aX + b.
In this case we have X =√
n µRSS
n − ¯σ2n
¶
= RSS
√n −√
n¯σn2 = σ¯n2
√n RSS
¯ σn2
| {z }
Y
−√ n¯σn2,
Y =
√n
¯ σ2n
|{z}
a
X + n|{z}
b
.
The density will be computed using this formula:
fY(x) = 1
|a||{z}fX
=g
(x − b a ).
It means:
fY(x) = σ¯n2
√n 1 2¯σ2n√
π µ
1 + x − n
√n¯σn2
¯ σn2
√n
¶n−p
2 −1
exp µ
−
√n(x − n) 2¯σ2n
¯ σn2
√n
¶
=
= 1
2√ nπ
³ x n
´n−p
2 −1
exp µ
−1
2(x − n)
¶
=
= 1
2√
πnn−p−12 e−n2xn−p2 −1exp
³
−x 2
´
. (4.50)
Remind the Stirling formula: Γ(x) = q2π
x
¡x
e
¢x¡
1 + O(x−1)¢ . The χ2 distribution with N degrees of freedom has the density:
fN(y) = 1 2N2Γ¡N
2
¢ yN2−1exp−y2 = 1 2N2
q4π N
¡N
2e
¢N
2
yN2−1exp−y2 ¡
1 + O(N−1)¢
=
= 1
2√
πNN −12 e−N2 yN2−1exp−y2 ¡
1 + O(N−1)¢
. (4.51)
We also know that
(n − p)−n−p−12 = n−n−p−12 (1 + O(n−n)). (4.52) Letting N = n − p we got the same result as in (4.50).
4.4 General p-sample regression model
We assume a p-sample model and comparing to the general model in (4.1) we can use even simpler formula:
yi= µj + εi,j, (4.53)
where j = 1 . . . p and i = 1 . . . n.
Definition 1. Let’s define an index set J(j) : i ∈ J(j) ⇔ xi,j = 1, for i = 1 . . . n and j = 1 . . . p. |J(j)| = nj corresponds to nj observations in the jth sample,P
nj = n.
We further assume σ(j)= σi for ∀i ∈ J(j), j = 1 . . . p. It means that all the observations of the same sample have the same variance.
4.4. GENERAL P -SAMPLE REGRESSION MODEL 17 The model in (4.53) has a n × p dimensional design matrix
X =
1 0 0
... ...
1 0 ... 0 1
... · · · ... 1 0
0 1
... ...
0 0 1
.
The p-dimensional matrix I(n) is diagonal:
I(n)= 1
nXTW X = 1 n
w(1) 0
. ..
0 w(p)
,
where w(j)= P
i∈J(j)
wi.
Some simplifications as in the homoscedastic case can be made even here:
Matrix B(ˆθ1):
B(ˆθ1) = 1 2n
P
i∈J(1)
wisi(ˆθ1) 0 . ..
0 P
i∈J(p)
wisi(ˆθ1)
. (4.54)
All its derivations are diagonal with the following terms B0(ˆθ1) = diag(b01(ˆθ1), . . . , b0p(ˆθ1))
with b0j(ˆθ1) = 1 n
X
i∈J(j)
wis2i(ˆθ1), (4.55) B00(ˆθ1) = diag(b001(ˆθ1), . . . , b00p(ˆθ1))
with b00j(ˆθ1) = 4 n
X
i∈J(j)
wis3i(ˆθ1). (4.56)
κsam(ˆθ) = −n¯σn2θˆ1+ Xn i=1
Si(ˆθ1) + n3 Xp j=1
θˆ2,j2 bjw−2(j), (4.57)
Csam(ˆθ) =
s0(ˆθ1) + n2 Pp
j=1
θˆ2,j2 wb200j
(j) 2n2θˆ2,1wb201
(1) · · · 2n2θˆ2,pwb20p (p)
2n2 bw21 (1)
0 . ..
2n2 bw2p (p)
(4.58) is a symmetric matrix. Using the formula in (4.21) for the determinant calculation we get:
|Csam(ˆθ)| = n2p Yp j=1
2bj w2(j)
s0(ˆθ1) + n2 Xp j=1
θˆ2,j2 b00j
w(j)2 − 2n2 Xp j=1
θˆ2,j2 b02j bjw2(j)
.(4.59)
msam(ˆθ) =√ n
−¯σn2+ s(ˆθ1) + n2 Pp
j=1
θˆ22,jwb20j (j)
2n2θˆ2,1b1w−2(1) ... 2n2θˆ2,pbpw−2(p)
. (4.60)
Using these results we can write (4.11) as:
fsam(y) = exp
ÃPn i=1
Si(ˆθ1) − nˆθ1s(ˆθ1) − n3 Pp
j=1
θˆ22,jbjw(j)−2− n3θˆ1 Pp
j=1
θˆ22,jb0jw−2(j)
!
(2π)p+12 |Csam(ˆθ)|12 . (4.61)
4.4.1 Special p-sample case
Theorem 3. Let’s suppose the p-sample regression model. The condition on weights:
wk= wl ∀k, l ∈ J(j), j = 1 . . . p (4.62) is the necessary condition for fulfilling the prerequisites of Theorem 2.
Proof. The necessary conditions from Theorem 2 will be proofed:
• Condition (1):
b0j = 1 n
X
i∈J(j)
wi
à wiσ(j)2 1 − 2wiσ2(j)θˆ1
!2
= 4 1 4n2
X
i∈J(j)
wi wiσ2(j) 1 − 2wiσ2(j)θˆ1
2
n w(j),