Sie sind auf Seite 1von 14

StaisticalPapers48,235-248(2007)

Statistical Papers
Springer-Verlag 2007

2 m41 designs with minimum aberration


or weak minimum aberration
P e n g - F e i Li ~ 2, M i n - Q i a n Liu 1, R u n - C h u Z h a n g I
Department of Statistics, School of Mathematical Sciences and LPMC,
Nankai University, Tianjin 300071, China
(corresponding author, e-mail: mqliu@nankai.edu.cn)
2 Department of Statistics and Actuarial Science, University of Waterloo,
Waterloo, Ontario N2L 3G1, Canada
Received: July 1, 2004; revised version: May 9, 2005
For measuring the goodness of 2m41 designs, Wu and Zhang (1993)
proposed the minimum
aberration (MA) criterion. MA 2"~41 designs have
been constructed using the idea of complementary designs when the number
of two-level factors, m, exceeds y,n where n is the total number of runs. In
5n
this paper, the structures of MA 2m41 designs are obtained when m > 5-~'
Based on these structures, some methods are developed for constructing
MA 2"~41 designs for T~
5n < m < yn as well as for yn _< m < n. When m <
- 5n
1W,
there is no general method for constructing MA 2m41 designs. In this case,
we obtain lower bounds for A3o and A31, where A30 and A31 are the numbers
of type 0 and type 1 words with length three respectively. And a method
for constructing weak minimum aberration (WMA) 2m41 designs (A3o and
A31 achieving the lower bounds) is demonstrated. Some MA or W M A 2"~41
designs with 32 or 64 runs are tabulated for practical use, which supplement
the tables 'in Wu and Zhang (1993), Zhang and Shao (2001) and Mukerjee
and Wu (2001).
K e y w o r d s Minimum aberration; Resolution; Weak minimum aberration; Wordlength pattern.
2000 M a t h e m a t i c s

Subject Classifications:

62K15, 62K05.

1 Introduction
Regular two-level fractional factorial designs are the most commonly used
designs for factorial experiments. A 2 m-p design denotes a design with m
two-level factors and can be constructed by using a defining relation. The

236
numbers 1, 2 , . . . , m attached to the factors are called letters and a product
(juxtaposition) of any subset of these letters is called a word. The number
of letters in a word is called the length of the word. Associated with every
2 m-p design is a set of p words called generators. The set of distinct words
formed by all possible products involving the p generators gives the defining
relation of the design. Let A i ( D ) denote the number of words of length
i in the defining relation of a 2 m-p design D, then the vector W ( D ) =
(A3(D), A 4 ( D ) , . . . , A,~(D)) is called the wordlength pattern of D.
An important characteristic of 2 m-p design is its resolution. The resolution of a 2 m-p design is defined to be the smallest r such that A r > 0
(Box and Hunter, 1961). A 2 m-p design with resolution r is usually denoted
by 2rm-p. In such a design, no c-factor effect is confounded with any other
effect containing less than r - c factors. Experimenters always prefer to use
a design with the maximum resolution. However, not all 2 m-p designs with
the same resolution are equally good. To further discriminate 2 m-p designs,
Fries and Hunter (1980) proposed the m i n i m u m aberration (MA) criterion.
That is,
D e f i n i t i o n 1 For two designs D1 and D 2 , suppose r is the smallest value
such that Ar(D~) ~ Ar(D2). D1 is said to have less aberration than D2 if
At(D1) < A~(D2). If no design has less aberration than D1, then D1 is said
to have m i n i m u m aberration (MA).
Let's illustrate these concepts through the following example.
E x a m p l e 1 Suppose 1, 2, 3, 4, 5 are five independent factors. Consider the
following two 2 7-2 designs:
D1:1,2,3,4,5,6=123,
7=245,
D2 : 1, 2, 3, 4, 5, 6 = 1234, 7 = 2345.
The defining relations and wordlength patterns of these two designs are
given by
D1 : I = 1236 = 2457 = 134567, and W ( D I ) = (0,2,0,1,0),
D2 : I = 12346 = 23457 = 1567,
and W(D2) = (0, 1, 2, 0, 0).
Thus both D1 and D2 have resolution IV, but D2 has less aberration than
D1.
There have been extensive discussions on MA designs in the literature, such as Chen and Wu (1991), Chen (1992, 1998), Chen, Sun and
Wu (1993), Chen and Hedayat (1996), Tang and Wu (1996), Cheng and
Mukerjee (1998), Cheng, Steinberg and Sun (1999) and Butler (2003).
However, in some experiments there are some factors that have four levels. Addelman (1962) constructed this kind of design from two-level designs
by the method of replacement. Wu (1989) improved Addelman's construction method by introducing the method of grouping. Wu, Zhang and Wang
(1992) extended Wu's grouping scheme to cover more general designs. Given

237
the many 2m4 k designs (designs with m two-level factors and k four-level factors) constructed from the above methods, Wu and Zhang (1993) proposed
the minimum aberration (also denoted by MA) criterion for measuring their
goodness. For practical use, we are mainly concerned with the construction
of MA 2"~41 designs here, but the methods can be easily extended to the
construction of MA 2m42 o r 2rn81 designs.
Now we briefly describe the construction of 2m41 designs by the method
of replacement and the MA criterion proposed by Wu and Zhang (1993).
Throughout this paper let q = m - p be the number of independent factors,
rt : 2q be the total number of runs, and let the term two-factor interactions
mean the interactions between the two-level factors only. Note that a design
with n runs and rn factors can be denoted by an n x m matrix, and the rows
and columns are identified with the runs and factors, respectively. Thus in
what follows in this paper, we will not differentiate between the factor and
column. We can represent the n - 1 columns in a saturated two-level design
(denoted by Ha) with n runs by the q independent columns denoted by
1, 2 , . . . , q and their interactions of order 2 to q, that is 12, 1 3 , . . . , 12.. q
(Wu and Zhang, 1993). Any three columns of the form (a,b, ab), where
ab is the interaction column between two-level columns a and b, can be
replaced by a four-level column without affecting the orthogonality (Addelman, 1962). The replacement is done according to the rule shown in Table 1.

Table 1 Rule for replacing any three columns of the form (a, b, ab) by a four-level
column

ab

0
0

0
1

0
1

1
1

0
1

1
0

four-level column
0
1
2
3

Note that when the two levels 0 and 1 are replaced by 1 and - 1 for the twolevel columns, the rule shown in Table 1 is still valid for the replacement.
To discriminate 2m41 designs, we now introduce the MA criterion. Let D
be a 2m41 design, where the two-level column is represented by c 1 , . . . , c,~,
the four-level column is represented by A = (al, a2, a3 = ala2), and the cj's
and ai's are chosen from H,~. In this paper, we assume any two columns
of cj's and ai's are distinct. There are two types of words in the defining
relation of this design. The first involves only the cj's, which is called type 0.
The second involves one of ai's and some of the cj's, which is called type 1.
Because ala2aa = I, any two ai's that appear in a word can be represented
by the third ai. Therefore these two types exhaust the possibilities. For a
2m41 design D, let A~o(D) and Ail(D) respectively be the number of type
0 and type 1 words of length i in the defining relation of D. The vector
W ( D ) = ( A a ( D ) , A 4 ( D ) , . . . ) is called the wordlength pattern of D, where
Ai(D) = { A i o ( D ) , A i l ( D ) } for i >_ 3. The MA criterion is widely used for

238
measuring 2"~41 designs. In practice, A30(D) and A31(D) are of main interest. If MA designs are hard to be constructed, then a modified version of
the MA criterion can be used, that is the weak minimum aberration (WMA)
criterion.
D e f i n i t i o n 2 Let D1 and D2 be two 2m41 designs and r be the smallest
value such that At(D1) ~ At(D2). If Aio(D1) < Ai0(D2), or Ai0(D1) =
Ai0(D2) but Ail(D1) < Ail(D2), then D1 has less aberration than D2. If
no design has less aberration than D1, then D1 is said to have minimum
aberration (MA). A 2m41 design is said to have weak minimum aberration
(WMA) if its A30 and A31 are minimized sequentially.
To illustrate the concepts for 2m41 designs, let's see the following example.
E x a m p l e 2 Based on D1 and D2 in Example 1, let A = (1, 2, 12) be the
four-level column, the following two 2541 designs with 32 runs can be obtained:
D 3 : A , 3, 4, 5, 6 = 1 2 3 , 7 = 2 4 5 ,
Da : A, 3, 4, 5, 6 = 1234, 7 = 2345,
where 3, 4, 5, 6, 7 represent the five two-level columns. Denoting 1, 2, 12 by
al, a2 and a 3 respectively, we have the defining relations of D3 and D4, that
is
D3 : I = a 3 3 6
= a 2 4 5 7 =a134567,
D4 : I = a3346 = a23457 = a1567.
So, A31(D3) = 1, A41(D3) = 1, A61(D3) = 1, A41(D4) = 2, A51(D4) = 1
and also D4 has less aberration than D3. From Wu and Zhang (1993), we
can also know D4 has MA.
There are only two papers in the literature that consider MA 2m41
designs. Mukerjee and Wu (2001) considered the situation when m > ~ and
Zhang and Shao (2001) constructed MA 2m41 designs for m = q, q + 1. In
this paper, alternative methods are developed which allow MA designs with
n runs and m > Tg5~two-level factors to be constructed. When m <_ 5nl__g,
lower bounds of A30 and A31 are obtained and a method for constructing
W M A designs is demonstrated. In Section 2, some existing results and useful
lemmas are given. Theorems on constructing MA and W M A 2m41 designs
are presented in Sections 3 and 4 respectively. Examples are provided for
illustration after the theorems. Section 5 contains some conclusions and
discussions. Some newly obtained designs are presented in the Appendix.

2 Existing results and some useful lemmas


For the convenience of presenting the existing and new results, in this section
and the subsequent sections, we assume the two levels of any two-level factor
are labelled as 1 and - 1 . First we consider the construction of saturated
resolution IV two-level designs with n runs and ~ factors (denoted by Fn).

239

It follows from Margolin (1969) that a saturated resolution IV design must


be the blocked combined design of H~. That is to say, Fn
Define Gn =

H a2
1~2 )
-H~ -19

(H~)
H~_ , then H~ -- (F~, G~). The following four properties of
2

F~ and G~ are very useful for constructing MA and WMA 2~41 designs.
L e m m a 1 (i) I f f l E Fn and f2 E Fn, then f l f 2 c G~, where f l f 2 is the
interaction column between f l and f2.
(ii) I f f l E F~ and f2 E G~, then f l f 2 E F~, where f l f 2 is the interaction
column between f l and f 2 .
(iii) If D is a 2~v-(m-q) design with n runs, then D should be selected from
n
Fn and A i ( D ) = 0 for odd i, when ~5 n + 1 < m < -~.
(iv) If D is a 2~i-('~-q) design with n runs, m > ~, and A3(D) is minimized,
then D can be written in the form D = (F~, do), where do comes from
Gn.

Properties (i) and (ii) are obvious, so we omit their proofs. Property (iii)
can be found in Butler (2003) and Property (iv) is obtained by Chen and
Hedayat (1996).
Butler (2003) developed a new method for constructing MA 2 r e - ( m - q )
designs. Some of the results are very useful for constructing MA 2m41 designs, which are summarized in the following lamina.
L e m m a 2 Let D be a 2 re-(m-q) design with r >_ III. Define T = (tij) =
D D ' and M k ( D ) = n -2 E~,j=I tkj Then the following equations hold.
k-2

(i) M k ( D ) = k ! A k ( D ) + ~ i = 3 ~ k i ( m ) A i ( D ) + constant for k > 3, where


s k i ( m ) ' s are constants only depending on i, k, m and n. Especially,
M4(D) = 24A4(D) + constant and M6(D) = 720A6(D) + (360m 960)A4 (D) + constant.
(ii) If D c Fn, then M k ( D ) = M k ( F n \ D ) + constant for even k and
M k ( D ) = 0 for odd k.
(iii) If D = (F~,d0) and do c Gn, then M k ( D ) = Mk(do) + constant for
k>_3.
The constants in the above lemma may depend on k, m and n, but not
on the particular choice of D, and the summation in (i) equals zero when
the superscript is less than the subscript. In the following of this paper,
constants and the summations have the same properties respectively. From
(i) and (ii) of the above lemma and some simple calculations, we have
L e m m a 3 Let D be a 2m-(m-q) design and D C Fn, then
A4(D) = A a ( F n \ D ) + constant, and
n _ m ) A 4 ( F n \ D ) + constant.
A6(D) = A 6 ( F n \ D ) + (-~

240

Proof From (i) of Lemma 2, we get M4(D) = 24Aa(D) + constant and


Ma (F,~\ D ) = 24A4 (F~ \ D ) + constant. Following the relationship between
M4(D) and Ma(F~\D) in (ii) of Lemma 2, we can easily obtain the first
equation. The second equation follows similarly. []
In fact, following the equations in Lemma 2, we could express Ai(D) for
i _> 8 and even i. However such details are rarely needed in practical use.
3 M A 2m41 d e s i g n s
In this section and the subsequent section, let D be a 2m41 design, Do be
the 2m-(m-q) design and A = (al,a2,a3 = ala2) be the four-level factor.
The technique for constructing MA 2"~41 designs with y6
~ + 5 1 < ~ m < ~ is
different from the case when rn _> ~,n so we partition this section into two
subsections.
3.1 2m41 designs with rn >
Firstly, let's see the worldlength pattern of D, the first component is A30 (D),
which is equal to A3(Do). So we must minimize A3(Do) in the first step.
When m > ~ and A30(D) is minimized, from Lemma 1, we can write Do
in the following form: Do = (F~, do), where do C Gn. When m = ~, there
exist designs with resolution IV (Bose, 1947) and Fn is the unique 2~v- ( ~ - z )
design up to isomorphism. Therefore al, a2 and a3 must come from Gn and
D = (Fn,do, A) when m _> ~. Let d = (d0,A), then d is a 2"~-~41 design
with n runs. Following the above argument and Lemma 2, we obtain the
following theorem for constructing MA 2m41 designs.
T h e o r e m 1 Let D be a 2m41 design. If D satisfies the following two con-

ditions:
(i) There exists a 2"~-~41 design d with n runs satisfying D = ( F n , d),
(ii) d is an MA design,

then D is an MA design.
P r o o f Let Dr = (Do,at) for r = 1,2,3, then Aio(D) = Ai(Do) and
Ail(D) = ~ r 3= l Ai(Dr) - 3Ai(D0). From (i) of Lemma 2, we have
i-2

Mi(Do) = i!Ai(Do) + E aiz(m)Al(Do) + constant, and


l=3
3

E
r=l

i--2

Mi(Dr) = i! E Ai(Dr) + E ail(rn + 1) E


r=l

/=3

At(Dr) + constant.

r=l

Then minimizing (A30 (D), A~I (D), A40 (D), Aal ( D ) , . . . ) sequentially is equivalent to minimizing (M3(D0) , Y~'.3r=1 M3(Dr), M4(D0), E r 3= l M 4 ( D r ) , . . . ) sequentially. Let do be the collection of the two-level columns in d, A =

241

(ax, a2, a3 = ala2) be the four-level factor in d and dr = (do, at) for r =
1, 2, 3. T h e n minimizing (A30 (d), A31(d), A40 (d), A41 ( d ) , . . . ) sequentially is
equivalent to minimizing (M3(do), }-:~= 1 M3(dr), M4 (do), }-~3=1 M 4 ( d r ) , . . . ) .
From (iii) of L e m m a 2, we obtain

Mi(do) = Mi(Do) + constant, and


3

Mi(dr) = E Mi(Dr) + constant.

r=l

Then

r~l

from the above argument,

we can easily have the assertion.

Now, lets's show how to construct an MA


the above method.

[]

22041 design with 32 runs by

Example
3 Let i, 2, 3,4 and 5 be the five independent columns of F32,
and F32 = {1,2,3,4,5,6 = 123,7 = 124,8 = 125,9 = 134, t0 = 135, tl =
145, t2 = 234, t3 = 235, t4 = 245, t5 = 345, t6 = 12345}, where ti represents
the (I0 + i)th factor, for example t2 means the 12th factor. Obviously,
12, 13, 14, 15 are four independent columns of G32. Let A = (12, 13, 23) be
a four-level factor, do = (14, 15, 1245, 34) be the collection of 4 two-level
factors and d -- (do, A). From Wu and Zhang (1993), d is an MA
2441
design. So D -- (F32, d) is an MA 22041 design.

Remark 1 When we use the above method, one thing we should note is
that there are only q - 1 independent columns in Gn, which can be verified
from the structure of Gn. So d is an MA design among all 2m-~41
designs
with n runs that contain at most q - 1 independent columns. Actually, we
can construct d in this way: first construct an MA 2m-~41
design d* with
gn runs, then let d =

d*

, d is what we want. Here, we can easily see t h a t

d and d* have the same defining relation.

Remark 2 Mukerjee and W u (2001) also studied the case when m _> ~,
especially w h e n n - 4 - m = 1 , 2 , . . . , 1 2 a n d n - m = 2
r-w forr<qand
w = 0, 1, 2, 3. T h e n when n = 64, their m e t h o d will be difficult to construct
M A 2m41 designs for m = 36, 37,..., 47. However, from the discussion in
R e m a r k 1, if we can construct MA 2m41 designs for m = 4, 5 , . . . , 15 with
32 runs, the construction becomes easier following our m e t h o d . 32-run M A
2~41 designs with m = 4, 5 , . . . , 9 can be found in W u and Zhang (1993).
A n d 32-run M A 2~41 designs for m = 10, 1 1 , . . . , 15 can be constructed by
our following m e t h o d s and will be tabulated in the Appendix.

3.2 2'~41 designs with m

<

Before presenting the construction


that is the following theorem.

method,

let's see the structure of D first,

242
8

P.-F. Li et al.

T h e o r e m 2 Let D be a 2m41 design with $-~ < m < ~ - 2. If D is an M A


design, then D satisfies the two conditions: Do C Fn and two of al, a2 and
a3 are in I n .
Proof When ~ < m < ~ - 2, the m a x i m u m resolution of Do is IV, which
can be inferred from Corollary 3 of Chen and Hedayat (1998). If D is an
MA design, then Do has resolution IV and Do C Fn (Lemma 1). From
L e m m a 1, we know t h a t two or none of al, a2 and a3 are in Fn. Now we
only need to prove that if all ai ~ F,~, then we can find a design t h a t has
less aberration t h a n D. Since m _< ~ - 2, there exist two columns a and b
satisfying a, b ~ Fn and a, b ~ D. Let A* - (a, b, ab) be the four-level factor
and D* = (Do, A*), next we show D* has less aberration t h a n D.
Let Do = (fl, f 2 , . . . , fro) and Di = (aifl, aif2 . . . . , aif,~) for i = 1, 2, 3.
From L e m m a 1, obviously Di C F~. Since ID01 = ]Di I = m, then IDoNDil =
IDol + IDil - IDo U Dil >_ 2m - ~ for i = 1,2,3, where IDil means the
cardinality of Di for i = 0, 1, 2, 3. Since if a i f j = fk for some i, j and k,
then aifk = f j . Therefore, there are at least m - ~ two-factor interactions
that are aliased with each ai for i = 1, 2, 3. T h a t is to say, A3o(D) = 0 and
A31 (D) _> 3 ( m - ~). In D*, there is no two-factor interaction t h a t is aliased
with a and b and at most "~ two-factor interactions t h a t are aliased with
ab. Hence, A30(D*) = 0 and A3z(D*) <_~-.
"~ Because T ~ < m < y 5 ~~ - 2, then
A31(D*) < A31(D). Thus, the proof is completed.
Without loss of generality, we assume al ~ Fn and a2 E Fn. Let F =
Fn\(Do, al, a2) and DR = (F, A). The following l e m m a studies the relationship between the wordlength patterns of D and D R.
L e m m a 4 Let D be a 2rn41 design which satisfies the two conditions in
Theorem 2, then the wordlength patterns of D and DR satisfy the following
equations.
(i)
(ii)
(iii)
(iv)
(v)

A31 (D) = A31 (DR) + constant,


Aao(D) = A4o(DR) + A41(DR) + A31(DR) + constant,
A41 (D) = - A a l ( D R ) - 2An1 (On) + constant,
A51(D) = A51(DR) + (~ - rn)A31(DR) + A41(DR) + constant,
A60(D) = A 6 o ( D R ) + A 6 1 ( D R ) + A 5 1 ( D R ) + ( ~ - m ) ( A 4 o ( D R ) + A 4 1 ( D R ) +
Aaz (DR)) + constant,
(vi) A61 (D) -- - A 6 1 ( D R ) - 2A51 (DR) - (~ - m + 1)A41 (DR) - 2A4o(DR) -(~ -- 2m)A31 (DR) + constant.
Proof Let Di = (Do, ai), Fi = (F, a), for i = 1, 2, D+ = (Do, al, a2), and
F+ = (F, al, a2).
(i) Because there are only words of even-length in F , , Aal(D) is equal
to the number of length 4 words in the defining relation of D+ that contain
ala2. So
Aal(D) = A4(D+) - An(D1) - A4(D2) + A4(Do).

243

From this equation and Lemma 3, we get


A31(D) = A4(F) - A4(F2) - A4(F1) + A4(F+) ~- constant
= A31(DR) + constant.
(ii) From L e m m a 3, we have
A40(D) -- A4(Do) = A4(F+) + constant.
Note t h a t there are three kinds (containing 0, 1, 2 factors of al and a2) of
length 4 words in the defining relation of F+ and the number is A4o(DR),
A41 (DR) and A31 (DR), respectively. Therefore A40(D) = A40(DR)+A41 (DR)
+Aal (DR) + constant.
(iii) Since Do C Fn and al, a2 C Fn, then A41(D) is equal to the number
of length 4 words in the defining relation of D+ that contain al or a2. T h a t
is to say, A41(D) = A4(D1) + A4(D2) - 2A4(Do). Then from L e m m a 3 and
the equation in (ii), we have
A41(D) = A4(F2) + A4(F1) - 2A4(F+) + constant
= An(F2) + A4(F1) - 2(A40(DR) + A41(DR) + A31(DR)) -~- constant
= -A41 (DR) - 2A31 (DR) + constant.
The last three equations follow similarly.

[]

From this l e m m a and Theorem 2, we can easily obtain the following


theorem, which will be helpful for constructing MA 2m41 designs.
T h e o r e m 3 Let D be a 2m41 design which satisfies the two conditions in
5n
Theorem 2, where -f~
< m ~_ 7n - 2. If D is the unique design (up to isomorphism) to minimize (A31(DR), A40(DR) +A41 (DR), A40(DR), A51(DR),
A60(DR) + A61(DR), A 6 0 ( D R ) , " ") sequentially, then D is an MA design.
In the above theorems, we don't consider the case m = 7n - 1. In this
case, if D is an MA design, then Do C F~ and ai ~t F~ for i =1, 2, 3.
When n > 8, all 2 ~ v 1 - ( ~ - 1 - q ) designs are isomorphic (Chen and Cheng,

2000),thuswecanselectDo

= (_H~).Thendifferentchoicesofal,

a2

and a3 = ala2 in G~ will result in 29-141 designs with the same wordlength
p a t t e r n when n > 8.
Theorem

4 Let Do =

(H~)
-H

and A = (al,a2,a3 = ala2), where n > 8

and ai E G~ for i = 1,2. Then D = (D0,A) is an MA 2~-141 design.


Based on Theorems 3 and 4, we can construct MA 2m41 designs for
5~
~ Now, let's construct some MA designs with 32 runs for
1-~ < m < 7"
illustration. Some MA 2"~41 designs with 32 and 64 runs are shown in the
Appendix.

244
E x a m p l e 4 Suppose 1, 2, 3, 4 and 5 are the five independent columns of
F32, where F32 is the same as t h a t in Example 3.
(i) For m = 14, let A = (1,2, 12) and DR = {A}, which means DR is
a design only containing the four-level factor. T h e n Aij(DR) = 0 for
i > 3, j = 0, 1, and DR satisfies the condition in T h e o r e m 3. Let Do =
F32\{1, 2} = {3, 4, 5 , 1 2 3 , 1 2 4 , 1 2 5 , 1 3 4 , 1 3 5 , 1 4 5 , 2 3 4 , 235,245,345, 12345}
and D = (Do, A), then D is an MA 21441 design.
(ii) For m = 13, let A = (1,2, 12) and DR = {A, 12345}, t h e n DR is a
2141 design with the four-level factor A and two-level factor 12345.
Obviously, D R also satisfies the condition in T h e o r e m 3. Let Do =
F32\{1, 2, 12345} = {3, 4, 5,123,124, 125,134, 135,145,234, 235,245,345}
and D = ( D o , A ) , then D is an MA 21341 design.
(iii) For m = 15, from T h e o r e m 4, let A = (12, 13, 23) be the four-level factor
and Do = F32\{12345} = {1, 2, 3, 4, 5,123,124,125,134, 135,145,234,
235,245,345}, then D = (Do, A) is an M A 21541 design.

4 WMA

2m41 designs

In the above section, we proposed the m e t h o d s for constructing M A 2m41


5n A n d some examples are provided for b e t t e r underdesigns when m > Tg"
standing the m e t h o d s . However, there is no general m e t h o d for constructing
M A 2m41 designs when m < 5,~ In this case, we will consider W M A 2m41
designs, which minimize A30 and A31 sequentially. Firstly, we will derive
lower b o u n d s of A3o and A31, which will serve as b e n c h m a r k s for searching
W M A designs. T h e n , we will propose a m e t h o d to construct W M A 2m41
designs.
Theorem

( m - ~ +n

5 I f D is an MA 2m41 design, then A30(D) = 0 and A a l ( D ) =


1) f o r n~ l <- m < n- 7
2.

Proof W h e n ~n _ 1 _< m _< n - 2, there exist 2~v-(m-q) designs (Bose, 1947).


T h e n if D is an M A 2m41 design, we must have A30(D) = 0. To prove the
second equality, we first prove A31 (D) > (m - ~ + 1) for ~ _ l < m- < - 7n _ 2 .
Let D + = (Do, al, a2), it is a 2 m+2-(m+2-q) design. Next we will only prove
~n + 1 ,
A31(D) _> 1 for m = ~n a n d A a l ( D ) _> 2 f o r m = n + l . W h e n m >
the results follow similarly.
W h e n m = ~, if A a l ( D ) = 0, then D + has resolution at least IV and ala2
is clear (not aliased with any main effect or two-factor interaction), which
contradicts Corollary 3 of Chen and Hedayat (1998). Thus, A31(D) _> 1 for
m = ~ . Let D - i be the 2"~-141 design obtained by deleting the ith two-level
column in D for i = 1 , . . . , m . W h e n m = ~ + 1, if A a l ( D ) = 1, we have
A a l ( D - i ) <_ A a l ( D ) = 1. Since D - i is a 2~41 design, then A 3 1 ( D - i ) _> 1.
Hence, A31(D-i) = 1, which means the ith factor d o e s n ' t a p p e a r in the
length three words of type 1 for i = 1 , . . . , m , thus A31(D) -- 0, which
contradicts A31(D) = 1. So A a l ( D ) _> 2 for m = ~n + 1.

245

Now let us prove A31(D) _< ( m - ~n + 1) for ~n - - 1 _< m _< ~ - - 2 .


Let F , = { b l , b 2 , . . . b~}. From Chen and Hedayat (1998), we know t h a t
there are ~ - 1 disjoint two-factor interactions in Fn, say b3b4,. , b } - l b ~
w i t h o u t loss of generality, t h a t are aliased with bib2. Let A* = (bl, b2, bib2)
be a four-level factor, D~ = ( b 3 , b s , . . . , b ~ - l )
and D* = (D~,A*). T h e n
A30(D*) = 0 and A a l ( D * ) = 0. By adding b4, b6,.. ' ' b-i - to D* sequentially,
we obtain 2"~41 designs with A30(D*) = 0 and A31(D*) = (m - ~ + 1)
f o r t h _ 1 _< m _< y" - 2. Since D is an MA design, then A30(D) = 0 and
A31(D) _< ( m - ~n + 1) for ~n - 1 <
_ m _< yn - 2. Combining the above
arguments, the proof is completed. []
R e m a r k 3 T h e above t h e o r e m considers the lower b o u n d s of A30 and A31.
Actually, in the proof of this theorem, D* is a W M A 2"~41 design, since
A30(D*) and A31(D*) a t t a i n the lower bounds. From the proof, W M A
2m41 desigps can also be constructed by deleting columns from Fn. Deleting
b4, b6, .. , bn/2 from Fn sequentially will result in such W M A designs. This
m e t h o d can be easily employed, especially when m is large. In the following,
we will illustrate the m e t h o d by constructing W M A 2141 design with 32
runs. Some 64-run W M A designs are tabulated in the Appendix.

E x a m p l e 5 Suppose 1, 2, 3, 4 and 5 are the five independent columns of


F32, where F32 is the same as t h a t in E x a m p l e 3. Let A = (1,2, 12) be
the four-level factor. For constructing W M A 2141 designs, we consider the
two-factor interactions t h a t are aliased with 12. T h a t is,
12 = 36 = 47 = 58 = 9t2 = tot3 = t i t 4 = tst6.
T h e n if we delete {6 = 123, 9 = 134, t3 = 235, t6 = 12345} from F32, we will
get a W M A 2141 design D = (Do, A), where Do = {3, 4, 5,124, 125,135,145,
234, 245,345} is the collection of 10 two-level factors. F r o m Sitter, Chen and
Feder (1997), we can know t h a t this design is also an M A design. Similarly,
we can construct W M A 2m41 designs for m = 7, 8, 9. C o m p a r e d with the
M A designs in W u and Zhang (1993), all designs we obtain are also M A
designs except for m = 8. Thus we will not t a b u l a t e these W M A designs in
the Appendix.

5 Conclusions

and discussions

This paper considers the construction of M A and W M A 2rn41 designs. Some


examples are provided for illustration and some newly constructed designs
are t a b u l a t e d for practical use. If the four-level factor is a blocking factor,
then a slightly different M A criterion can be found in Zhang and P a r k (2000)
or Cheng and W u (2002). In this case, some modifications should be made
to T h e o r e m s 1, 3 and 4. T h e n the results here also supplement the work of
Li, Liu and Zhang (2005), which considers the construction of M A 2 "~-(m-q)
designs with two blocks. T h e m e t h o d s can also be easily extended to the
construction of M A and W M A 2"~42 designs. However, can the m e t h o d s be

246
generalized to construct MA sm(s 2) or sm(s2) 2 designs for genral s? These
is an interesting and open problem for further study.

Acknowledgements
The authors would like to thank the Coordinating Editor and two anonymous referees for their valuable comments and suggestions. This work was
partially supported by the National Natural Science Foundation of China
grants 10171051, 10301015, and the Science and Technology Innovation
Fund of Nankai University.

Appendix
In the following three tables, A always represents the four-level factor and
F64 = {1, 2, 3, 4, 5, 6,123,124,125,126,134,135,136,145,146,156,234,235,
236,245,246,256,345,346,356,456, 12345, 12346, 12356, 12456, 13456, 23456}.

Table 2 32-run MA 2m41 designs for 10 ~ m < 15


m
Two-level factors
10
3,4,5,124,125,135,145,234,245,345
11
3,4,5,123,124,125,135,145,234,245,345
12
3,4,5,123,124,125,134,135,145,234,245,345
13
3,4,5,123,124,125,134,135,145,234,235,245,345
14 3,4,5,123,124,125,134,135,145,234,235,245,345,12345
15 1,2,3,4,5,123,124,125,134,135,145,234,235,245,345

Four-level factor A

(1,2,12)
(1,2,12)
(1,2,12)
(1,2,12)
(1,2,12)
(12,13,23)

Table 3 64-run MA 2m41 designs for 21 < m < 31


m
21
22
23
24
25
26
27
28
29
3O
31

Two-level factors
F64\{ 1, 2, 3, 4, 5, 6,134,246, 12345, 23456, 12456}
F64\{1, 2, 3, 4, 5, 6,134, 156,235,246}
F64\{ 1, 2, 3, 4, 5, 6,134, 156,235}
F64\{ 1, 2, 3, 4, 5, 6, 12345, 23456}
F64\{ 1, 2, 3, 4, 5, 6, 23456}
F64\{1, 2, 3, 4, 5, 6}
F64\{1, 2, 3, 4, 5}
F64\{ 1, 2, 3, 4}
F64\{1, 2, 3}

F~4\{1, 2}
F64\{1}

Four-level factor A

(1,2,12)
(1,2,12)
(1,2,12)
(1,2,12)
(1,2,12)
(1,2,12)
(1,2,12)
(1,2,12)
(1,2,12)
(1,2,12)
(12,13,23)

247
T a b l e 4 64-run WMA 2m41 designs for 16 < m < 20, A = (1, 2, 12)
Two-level factors
16
3,4,5,6,134,135,136,145,146,156,345,346,356,456,13456,123
17
3,4,5,6,134,135,136,145,146,156,345,346,356,456,13456,123,124
18
3,4,5,6,134,135,136,145,146,156,345,346,356,456,13456,123,124,125
19
3,4,5,6,134,135,136,145,146,156,345,346,356,456,13456,123,124,125,126
20 3,4,5,6,134,135,136,145,146,156,345,346,356,456,13456,123,124,125,126,234
m

References
1. Addelman S (1962) Orthogonal main-effect plans for asymmetrical factorial
experiments. Technometrics 4, 21-46
2. Bose RC (1947) Mathematical theory of the symmetrical factorial design.
Sankhy~t 8, 107-166
3. Box, GEP, Hunter, JS (1961) The 2 k-p fractional factorial designs I and II.
Technometrics 3, 311-351 and 449-458
4. Butler NA (2003) Some theory for constructing minimum aberration fractional factorial design. Biometrika 90,233-238
5. Chen H, Cheng CS (2000) Uniqueness of some resolution IV two-level regular
fractional factorial designs. SIAM J Discrete Math 13, 571-575
6. Chen H, Hedayat AS (1996) 2 n-l designs with weak minimum aberration.
Ann Statist 24, 2536-2548
7. Chen H, Hedayat AS (1998) 2 n-m designs with resolution III or IV containing
clear two-factor interactions. J Statist Plann Inference 75, 147-158
8. Chen J (1992) Some results on 2~-k fractional factorial designs and search
for minimum averration designs. Ann Statist 20, 2124-2141
9. Chen J (1998) Intelligent search for 213-6 and 214-7 minimum aberration
designs. Statist Sinica 8, 1265-1270
10. Chen J, Sun DX, Wu CFJ (1993) A catalogue of two-level and three-level
fractional factorial designs with small runs. Internat Statist Rev 61, 131-145
11. Chen J, Wu CFJ (1991) Some results on s n-k fractional factorial designs
with minimum aberration or optimal moments. Ann Statist 19, 1028-1041
12. Cheng CS, Mukerjee R (1998) Regular fractional factorial designs with minimum aberration and maximum estimation capacity. Ann Statist 26, 22892300
13. Cheng CS, Steiberg DM, Sun DX (1999) Minimum aberration and model
robustness for two-level fractional factorial designs. J Roy Statist Soc Set B
61, 85-93
14. Cheng SW, Wu CFJ (2002) Choice of optimal blocking schemes in two-level
and three-level designs. Technometrics 44, 2549-2559
15. Fries A, Hunter WG (1980) Minimum aberration 2 k-p designs. Technometrics
22,601-608
16. Li PF, Liu MQ, Zhang RC (2005) Choice of optimal initial designs in sequential experiments. Metrika, in press
17. Margolin BH (1969) Resolution IV fractional factorial designs. J Roy Statist
Soc Ser B 31,514-523
18. Mukerjee R, Wu CFJ (2001) Minimum aberration designs for mixed factorials. Statist Sinica 11,225-239

248
19. Sitter RR, Chen J, Feder A (1997) Fractional resolution and minimum aberration in blocked 2~-p designs. Technometries 39, 382-390
20. Tang B, Wu CFJ (1996) Characterization of minimum aberration 2'~-k designs in terms of their complementary designs. Ann Statist 24, 2549-2559
21. Wu CFJ (1989) Construction of 2m4n via a group scheme. Ann Statist 17,
1880-1885
22. Wu CFJ, Zhang RC (1993) Minimum aberration designs with two-level and
four-level factors. Biometrika 80, 203-209
23. Wu CFJ, Zhang RC, Wang RG (1992) Construction of asymmetrical orthogonal arrays of the type OA(s k, (s ~1),~1 ... (s~),~ ). Statist Sinica 2, 203-219.
24. Zhang RC, Park D (2000) Optimal blocking of two-level fractional factorial
designs. J Statist Plann Inference 91, 107-121
25. Zhang RC, Shao Q (2001) Minimum aberration ($2)S ~-k designs. Statist
Sinica 11, 213-223

Das könnte Ihnen auch gefallen