FOLIA OECONOMICA 285, 2013
[159]
Małgorzata Graczyk
*OPTIMALITY OF SPRING BALANCE
WEIGHING DESIGNS
Abstract. In this paper, the problem of existence of optimal spring balance weighing designs is discussed. Two optimality criterions are compared and the appropriate optimality conditions are presented.
Key words: A–optimal design, D–optimal design, spring balance weighing design.
I. INTRODUCTION
We consider the model of spring balance weighing design
y XwE ,
Cov
e
2G
, (1)where
y
is an n1 random vector of the observations. Each observation meas-ures the sum of the measurements of objects taken to the combination.
0
,
1
t nΦ
X
,Φ
nt
0
,
1
denotes the class ofn
t
matricesX
x
ij ,n
i
1
,
2
,...,
,j
1
,
2
,...,
t
, having entriesx
ij
1
or 0,w
w
1,
w
2,...,
w
t
'is a vector representing true measurements of objects and e is the n1 random vector of errors,E
e
0
n,E
ee
'
2G
, where G is positive definite knownmatrix. If the designs matrix
X
is of full column rank, then allw
j are estimable and the variance matrix of their best linear unbiased estimator is
2
X
'G
1X
1In the theory of optimal designs, very important role plays the matrix G. It describe the relations between measurement errors. Each form of that matrix requires specific investigations.
* Ph.D., Department of Mathematical and Statistical Methods, Poznań University of Life
II. OPTIMALITY CRITERIA
Kiefer (1974) considered the
φ
p (forp
0
) optimality criteria in the form
1
for
tr
0
for
det
φ
1 1p
p
pΩ
Ω
, (2)where
Ω
X
'X
. The D-, A- criteria of optimality are the same as 0φ
and φ1, respectively. Minimizingφ
p is equivalent to maximizing Pukelsheim’sφ
p (see Pukelsheim (1993)), which is also defined for
1
p
0
. Here, we gener-alize the optimum criterion given by Kiefer (1974). For positive definite matrixG of known elements, the information matrix for estimating w is
X
G
X
Ω
G
' 1 . Thus
1
for
tr
0
for
det
φ
1 1p
p
p G GΩ
Ω
and we have the following definition.
Definition 2.1. Any nonsingular spring balance weighing design
0
,
1
t n
Φ
X
with the variance matrix of errors
2G
is said to bei) A-optimal if and only if tr
ΩG 1 is minimal, ii) D-optimal if and only if det
ΩG 1 is minimal.III. THE DESIGN MATRIX
In Jacroux and Notz (1983) the optimal designs are presented. Theorem 3.1. Let
t
be odd and let the condition
'
'4
1
t t tt
n
t
1
1
I
X
X
(4)be satisfied. Thus
X
Φ
nt
0
,
1
is A-optimal and D-optimal.It is worth pointing out that in the class
X
Φ
nt
0
,
1
for a given n andt
, the number n
t1 4t 1 must be integer. This requirement is very restricted as we are not able to construct optimal design for any n andt
and it is needed toselect an optimal weighing design. Accordingly, let
Ψ
n1 t
0
,
1
denotes the class of matrices of spring balance weighing designs in that (4) is true, i.e. for anyX
1 Ψ
n1t
0
,
1
the condition
'
1 ' 14
1
1
t t tt
n
t
1
1
I
X
X
(5) is satisfied.Special attention in this paper is given to the determining optimal design in the class
X
Φ
nt
0
,
1
, in that optimal design satisfying conditions given by Jacroux and Notz (1983) doesn’t exist. Because of this the motivation of this paper is to consider the situation in that the measurements are taken in two dif-ferent conditions or on two difdif-ferent installations, n1 measurements are taken with variance
2 and one, additionally measurement in other conditions with varianceg
1
2. Thus ' 1 1 1 1 g n n n 0 0 I G . (6)According to (6), we consider
X
Φ
nt
0
,
1
in the form
'1x
X
X
, (7)where
X
1 Ψ
n1t
0
,
1
satisfies (5),x
is t1 vector of elements equal to 1 or 0. The form (7) could be interpreted as construction the design matrixX
in the classΦ
nt
0
,
1
based on matrix X1 from the classΨ
n1t
0
,
1
, which is A- and D-optimal. In the other words, the problem is how to add to n1 meas-urements one additionally measurement to become optimal design. Theorems given by Katulska and Przybył (2007) and Graczyk (2011) will be required to prove the main result of this paper.Theorem 3.1. Katulska and Przybył (2007). Let
t
be odd. If2
1
'
t
t1
x
then any nonsingular spring balance weighing design
X
Φ
nt
0
,
1
in the form (7) with the variance matrix of errors
2G
in (6) is regular D-optimal.Theorem 3.2. Graczyk (2011). Let
t
be odd. The designX
Φ
nt
0
,
1
in the form (7) with the variance matrix of errors
2G
in (6) is A-optimal ifi) for fixed
g
,g
0
,
P
0
, if2
1
'
t
t1
x
,ii) for fixed
g
,g
L
s,
P
s
, if t
t
s
2
1
'
1
x
,iii) for fixed
g
,
,
2 1 pL
g
, if x'1t 1,iv) for fixed
g
P
s, if t
t
s
2
1
'1
x
or t
t
s
2
1
'1
x
, where
1
1
0
p
p
n
P
,
2
1
2
1
1
3
4
1
2
1
s
p
s
p
p
n
s
s
p
p
L
s ,
2
1
2
1
1
1
4
2
1
s
p
s
p
p
n
s
ps
p
P
s ,2
3
,...,
2
,
1
p
s
.Theorem 3.3. Let
X
Φ
nt
0
,
1
be nonsingular spring balance weighing design in the form (7) for that the condition (5) is satisfied.i) If
g
0
,
P
0
and2
1
'
t
t1
x
, thenX
Φ
nt
0
,
1
is A- and D-optimal.ii) If
g
L
s,
P
s
and moreover if a. t
t
s
2
1
'1
x
thenX
Φ
nt
0
,
1
is A-optimal, b.2
1
'
t
t1
x
thenX
Φ
nt
0
,
1
is D-optimal. iii) If
,
2 1 pL
g
and moreover if a. x'1t 1 thenX
Φ
nt
0
,
1
is A-optimal,b.
2
1
'
t
t1
x
thenX
Φ
nt
0
,
1
is D-optimal. iv) Ifg
P
s and moreovera. t
t
s
2
1
'1
x
or t
t
s
2
1
'1
x
thenX
Φ
nt
0
,
1
is A-optimal, b.2
1
'
t
t1
x
thenX
Φ
nt
0
,
1
is D-optimal.Proof. For positive definite matrix G in (6) of known elements, let us con-sider the design
X
1 Ψ
n1t
0
,
1
which satisfies Condition (5), i.e. X1 is A- and D-optimal. Based on that design, we formX
Φ
nt
0
,
1
in (7). According to optimality criterion we considerΩ
G. We have' 1 ' 1 1 '
G
X
X
X
xx
X
Ω
G
g
. (8)The proof falls naturally into two parts. First we determine D-optimal de-sign. We count det
ΩG 1. Katulska and Przybył (2007) showed that
tt
n
t
n
gt
t
4
1
1
1
1
1
det
Ω
G . (9)and, moreover they showed that maximum of (9), i.e. minimum of det
ΩG 1 is attained if and only if2
1
'
t
t1
x
. Thus if2
1
'
t
t1
x
thenX
Φ
nt
0
,
1
is D-optimal. Next we check if inΦ
nt
0
,
1
exists A-optimal design. Thus we count tr
ΩG 1 and we determinex
for which minimum of tr
ΩG 1 is at-tained. From Graczyk (2011) we have, ifg
0 P
,
0
then
1 1
1
1 1 4 tr 2 3 1 n tg n t t g n t G Ω (10) and ifg
P
0 then
1
1
1
4
tr
1
n
t
t
t
GΩ
. (11) Ifg
L
s,
P
s
then
Ms n t s t g t n t t 1 1 1 2 4 1 1 4 tr 2 2 2 3 1 G Ω , (12) where
1 1
2 1
2 3
2 2 1 2 s t s t tg n t t s t t Ms and ifg
P
s then
1 1
2 1
1 4 1 2 4 tr 1 2 s n t s s t t G Ω . (13) If
,
2 1 pL
g
then
1 1
2 1
1 4 1 1 4 tr 2 2 2 1 s n t t g n t t t G Ω . (14)For
g
0 P
,
0
, minimum of (10) is attained if and only if2
1
'
t
t1
x
. Hence if t
t
s
2
1
'1
x
thenX
Φ
nt
0
,
1
is A-optimal. Ifg
P
0 then the equality in (11) is fulfilled if2
1
'
t
t1
x
or2
1
'
t
t1
x
. Forg
L
s,
P
s
minimum of (12) is attained if and only if
2
1
'
t
t1
x
. Ifg
P
s then the equality in (13) is true if t
t
s
2
1
'1
x
or t
t
s
2
1
'1
x
. In this case
0
,
1
t nΦ
X
is A-optimal. We have, if t
t
s
2
1
'1
x
thenX
Φ
nt
0
,
1
is A-optimal. If
g
P
s then the equality in (13) is true if t
t
s
2
1
'1
x
ors
t
t
2
1
'1
x
. In this caseX
Φ
nt
0
,
1
is A-optimal. Finally, if
,
2 1 pL
g
then minimum of (14) is attained if and only if ' 1 t 1 x . Thus if 1 ' t 1x then
X
Φ
nt
0
,
1
is A-optimal. Hence the theorem.Now, we form the design matrix
X
1 Ψ
n1t
0
,
1
based on the incidence matrix of balanced incomplete block design with the parametersv
,
b
,
r
,
k
,
, see Raghavarao and Padgett (2005), asX
1
N
'.Lemma 3.1. If N is the incidence matrix of balanced incomplete block de-sign with the parameters
1)
v
q
4
1
, b2
4q1
, r2
2q1
,k
q
2
1
,
q
2
1
or 2)v
q
4
1
,b
q
4
1
,r 2
q
,k 2
q
,
q
,q
1
,
2
,...
. then forX
1
N
' the condition (5) is fulfilled.The proof is left for the reader.
From now on, we consider
X
1
N
', where N is the incidence matrix of balanced incomplete block design with the parameters given in Lemma 3.1. The parameter connected with precision of measurementsg
is given. The values ofg
determining respectively intervals are the same as in Theorem 3.2. Thus in next corollaries, according to the value ofg
we give the conditions determining optimal designX
in the classΦ
nt
0
,
1
.Corollary 3.1. If
g
0 P
,
0
and if2
1
'
t
t1
x
, thenX
Φ
b1t
0
,
1
in the form (7) with the variance matrix of errors
2G
in (6) is A- and D-optimal.Corollary 3.2. If
g
L
s,
P
s
and moreover if i) t
t
s
2
1
'
1
x
thenX
Φ
b1t
0
,
1
in the form (7) with the variance matrix of errors
2G
in (6) is A-optimal,ii)
2
1
'
t
t1
x
thenX
Φ
b1t
0
,
1
in the form (7) with the variance ma-trix of errors
2G
in (6) is D-optimal.Corollary 3.3. If
,
2 1 pL
g
and moreover if i) ' 1 t 1x then
X
Φ
b1t
0
,
1
in the form (7) with the variance matrix of errors
2G
in (6) is A-optimal, ii)if2
1
'
t
t1
x
thenX
Φ
b1t
0
,
1
in the form (7) with the variance matrix of errors
2G
in (6) is D-optimal.Corollary 3.3. If
g
P
s and moreover i) t
t
s
2
1
'1
x
or t
t
s
2
1
'1
x
thenX
Φ
b1t
0
,
1
in the form (7) with the variance matrix of errors
2G
in (6) is A-optimal,ii)
2
1
'
t
t1
x
thenX
Φ
b1t
0
,
1
in the form (7) with the variance ma-trix of errors
2G
in (6) is D-optimal.IV. EXAMPLE
Let us consider experiment in that using n11 measurements we determine unknown weights of t5 objects. For the construction of design matrix we use the incidence matrix N of balanced incomplete block design with the parame-ters v5, b10, r6, k 3,
3 given as
0
0
1
0
1
1
0
1
1
1
0
1
0
1
0
1
1
0
1
1
1
0
0
1
1
0
1
1
0
1
1
1
1
0
0
0
1
1
1
0
1
1
1
1
1
1
0
0
0
0
N
. Thus if
2
1
,
0
g
then X2 is A- and D-optimal, if
2
45
,
2
1
g
then X2 is D-optimal and X3 isA-optimal, if
,
2
45
g
then X2 is D-optimal and X4 is A-optimal,more-over if
2
1
2
45
g
then X2 is D-optimal and X3, X4 are A-optimal, where2 X , X3, X4
Φ
115
0
,
1
are given in as
0
0
0
1
0
1
1
0
1
1
1
1
0
1
0
1
0
1
1
0
1
1
1
1
0
0
1
1
0
1
1
0
1
0
1
1
1
0
0
0
1
1
1
0
1
1
1
1
1
1
1
0
0
0
0
' 2X
,
0
0
0
1
0
1
1
0
1
1
1
1
0
1
0
1
0
1
1
0
1
1
1
1
0
0
1
1
0
1
1
0
1
0
1
1
1
0
0
0
1
1
1
0
0
1
1
1
1
1
1
0
0
0
0
' 3X
,
0
0
0
1
0
1
1
0
1
1
1
1
0
1
0
1
0
1
1
0
1
1
0
1
0
0
1
1
0
1
1
0
1
0
1
1
1
0
0
0
1
1
1
0
0
1
1
1
1
1
1
0
0
0
0
' 4X
. REFERENCESGraczyk M. (2011), Some notes about spring balance weighing design, To appear.
Jacroux M., Notz W. (1983), On the optimality of spring balance weighing designs, The Annals of
Statistics, 11, 970–978.
Katulska K., Przybył K. (2006), On certain D-optimal spring balance weighing designs. Journal of
Statistical Theory and Practice, 1, 393–404.
Kiefer J. (1974), General equivalence theory for optimum designs. The Annals of Statistics 2, 849–879.
Pukelsheim F. (1993), Optimal design of experiment. John Wiley and Sons, New York.
Raghavarao D., Padgett L.V. (2005), Block Designs, Analysis, Combinatorics and Applications, Series of Applied Mathematics, 17, Word Scientific Publishing Co. Pte. Ltd.
Małgorzata Graczyk
OPTYMALNOŚĆ W SPRĘŻYNOWYCH UKŁADACH WAGOWYCH
W pracy przedstawiono zagadnienie A- i D-optymalności sprężynowego układu wagowego. Rozważania teoretyczne zostały zobrazowane przykładem konstrukcji macierzy układu w oparciu o macierze incydencji układów zrównoważonych o blokach niekompletnych.