Acta Mathematica Sinica, English Series Springer-Verlag Berlin Heidelberg & The Editorial Office of AMS 2015

Similar documents
UNIFORM FRACTIONAL FACTORIAL DESIGNS

Some Nonregular Designs From the Nordstrom and Robinson Code and Their Statistical Properties

Interaction balance in symmetrical factorial designs with generalized minimum aberration

Moment Aberration Projection for Nonregular Fractional Factorial Designs

USING REGULAR FRACTIONS OF TWO-LEVEL DESIGNS TO FIND BASELINE DESIGNS

Some optimal criteria of model-robustness for two-level non-regular fractional factorial designs

MINIMUM MOMENT ABERRATION FOR NONREGULAR DESIGNS AND SUPERSATURATED DESIGNS

Optimal Selection of Blocked Two-Level. Fractional Factorial Designs

An Approach to Constructing Good Two-level Orthogonal Factorial Designs with Large Run Sizes

GENERALIZED RESOLUTION AND MINIMUM ABERRATION CRITERIA FOR PLACKETT-BURMAN AND OTHER NONREGULAR FACTORIAL DESIGNS

CONSTRUCTION OF OPTIMAL MULTI-LEVEL SUPERSATURATED DESIGNS. Hongquan Xu 1 and C. F. J. Wu 2 University of California and University of Michigan

Connections between the resolutions of general two-level factorial designs

Minimax design criterion for fractional factorial designs

A General Criterion for Factorial Designs Under Model Uncertainty

Projection properties of certain three level orthogonal arrays

Algorithmic Construction of Efficient Fractional Factorial Designs With Large Run Sizes

Forms of four-word indicator functions with implications to two-level factorial designs

Maximal Rank - Minimum Aberration Regular Two-Level Split-Plot Fractional Factorial Designs

COMPROMISE PLANS WITH CLEAR TWO-FACTOR INTERACTIONS

CONSTRUCTION OF NESTED (NEARLY) ORTHOGONAL DESIGNS FOR COMPUTER EXPERIMENTS

SOME NEW THREE-LEVEL ORTHOGONAL MAIN EFFECTS PLANS ROBUST TO MODEL UNCERTAINTY

Bounds on the maximum numbers of clear two-factor interactions for 2 (n 1+n 2 ) (k 1 +k 2 ) fractional factorial split-plot designs

Characterizations of indicator functions of fractional factorial designs

A note on optimal foldover design

Construction of column-orthogonal designs for computer experiments

UCLA Department of Statistics Papers

Minimum Aberration and Related Designs in Fractional Factorials. 2 Regular Fractions and Minimum Aberration Designs

HIDDEN PROJECTION PROPERTIES OF SOME NONREGULAR FRACTIONAL FACTORIAL DESIGNS AND THEIR APPLICATIONS 1

Construction of some new families of nested orthogonal arrays

CONSTRUCTION OF OPTIMAL MULTI-LEVEL SUPERSATURATED DESIGNS. University of California, Los Angeles, and Georgia Institute of Technology

CONSTRUCTION OF SLICED ORTHOGONAL LATIN HYPERCUBE DESIGNS

An Algorithm for Constructing Orthogonal and Nearly Orthogonal Arrays with Mixed Levels and Small Runs

Optimal blocking of two-level fractional factorial designs

Optimal Fractional Factorial Plans for Asymmetric Factorials

arxiv: v1 [stat.me] 16 Dec 2008

Citation Statistica Sinica, 2000, v. 10 n. 4, p Creative Commons: Attribution 3.0 Hong Kong License

CONSTRUCTION OF SLICED SPACE-FILLING DESIGNS BASED ON BALANCED SLICED ORTHOGONAL ARRAYS

Indicator Functions and the Algebra of the Linear-Quadratic Parametrization

University, Wuhan, China c College of Physical Science and Technology, Central China Normal. University, Wuhan, China Published online: 25 Apr 2014.

Classification of three-word indicator functions of two-level factorial designs

Statistica Sinica Preprint No: SS R2

A GENERAL CONSTRUCTION FOR SPACE-FILLING LATIN HYPERCUBES

Some characterizations of affinely full-dimensional factorial designs

E(s 2 )-OPTIMALITY AND MINIMUM DISCREPANCY IN 2-LEVEL SUPERSATURATED DESIGNS

ORTHOGONAL ARRAYS CONSTRUCTED BY GENERALIZED KRONECKER PRODUCT

Construction and analysis of Es 2 efficient supersaturated designs

Mixture Designs Based On Hadamard Matrices

Construction of optimal Two- Level Supersaturated Designs

All Good (Bad) Words Consisting of 5 Blocks

18Ï È² 7( &: ÄuANOVAp.O`û5 571 Based on this ANOVA model representation, Sobol (1993) proposed global sensitivity index, S i1...i s = D i1...i s /D, w

A Coset Pattern Identity between a 2 n p Design and its Complement

On the simplest expression of the perturbed Moore Penrose metric generalized inverse

A NEW CLASS OF NESTED (NEARLY) ORTHOGONAL LATIN HYPERCUBE DESIGNS

Optimal Foldover Plans for Two-Level Fractional Factorial Designs

Construction of optimal supersaturated designs by the packing method

arxiv: v1 [math.co] 27 Jul 2015

D-Optimal Designs for Second-Order Response Surface Models with Qualitative Factors

Definitive Screening Designs

CONSTRUCTION OF NESTED ORTHOGONAL LATIN HYPERCUBE DESIGNS

A UNIFIED APPROACH TO FACTORIAL DESIGNS WITH RANDOMIZATION RESTRICTIONS

FRACTIONAL FACTORIAL SPLIT-PLOT DESIGNS WITH MINIMUM ABERRATION AND MAXIMUM ESTIMATION CAPACITY

QUASI-ORTHOGONAL ARRAYS AND OPTIMAL FRACTIONAL FACTORIAL PLANS

A Short Overview of Orthogonal Arrays

ON THE RELATIVE GENERALIZED HAMMING WEIGHTS OF A 4-DIMENSIONAL LINEAR CODE AND A SUBCODE WITH DIMENSION ONE

By Ming-Chung Chang and Ching-Shui Cheng Academia Sinica and University of California, Berkeley

A new family of orthogonal Latin hypercube designs

proposed. This method can easily be used to construct the trend free orthogonal arrays of higher level and higher strength.

Optimal Two-Level Regular Fractional Factorial Block and. Split-Plot Designs

On the construction of asymmetric orthogonal arrays

arxiv: v1 [math.ra] 27 Jul 2013

Resolvable partially pairwise balanced designs and their applications in computer experiments

Florida State University Libraries

FACTOR SCREENING AND RESPONSE SURFACE EXPLORATION

Houston Journal of Mathematics. c 2008 University of Houston Volume 34, No. 4, 2008

On the decomposition of orthogonal arrays

A General Criterion for Factorial Designs Under Model Uncertainty

ORTHOGONAL ARRAYS OF STRENGTH 3 AND SMALL RUN SIZES

Samurai Sudoku-Based Space-Filling Designs

Sliced Minimum Aberration Designs for Four-platform Experiments

Structure Functions for Regular s l m Designs with Multiple Groups of Factors

Neighbor Sum Distinguishing Total Colorings of Triangle Free Planar Graphs

Representations of disjoint unions of complete graphs

Statistica Sinica Preprint No: SS

AMBIGUOUS FORMS AND IDEALS IN QUADRATIC ORDERS. Copyright 2009 Please direct comments, corrections, or questions to

Dipartimento di Matematica

Construction of Mixed-Level Orthogonal Arrays for Testing in Digital Marketing

A RESOLUTION RANK CRITERION FOR SUPERSATURATED DESIGNS

INTELLIGENT SEARCH FOR AND MINIMUM ABERRATION DESIGNS

Research Article On Polynomials of the Form x r f (x (q 1)/l )

Analysis Methods for Supersaturated Design: Some Comparisons

The Structure of Minimal Non-ST-Groups

Asymptotic behavior for sums of non-identically distributed random variables

Minimum Aberration and Related Criteria for Fractional Factorial Designs

A NEW ALGORITHM FOR OBTAINING MIXED-LEVEL ORTHOGONAL AND NEARLY-ORTHOGONAL ARRAYS

Maximal perpendicularity in certain Abelian groups

Designing Two-level Fractional Factorial Experiments in Blocks of Size Two

Weizhen Wang & Zhongzhan Zhang

A SIMPLE METHOD FOR CONSTRUCTING ORTHOGONAL ARRAYS BY THE KRONECKER SUM

Proof of a Conjecture on Monomial Graphs

ON D-OPTIMAL DESIGNS FOR ESTIMATING SLOPE

Transcription:

Acta Mathematica Sinica, English Series Jul., 2015, Vol. 31, No. 7, pp. 1163 1170 Published online: June 15, 2015 DOI: 10.1007/s10114-015-3616-y Http://www.ActaMath.com Acta Mathematica Sinica, English Series Springer-Verlag Berlin Heidelberg & The Editorial Office of AMS 2015 Some Properties of β-wordlength Pattern for Four-level Designs Wei Wei SHENG Xia Ming LI 1) Yu TANG School of Mathematical Sciences, Soochow University, Suzhou 215006, P. R. China E-mail : shengweiwei333@163.com xiamlee@suda.edu.cn ytang@suda.edu.cn Abstract Fractional factorial designs have played a prominent role in the theory and practice of experimental design. For designs with qualitative factors under an ANOVA model, the minimum aberration criterion has been frequently used; however, for designs with quantitative factors, a polynomial regression model is often established, thus the β-wordlength pattern can be employed to compare different fractional factorial designs. Although the β-wordlength pattern was introduced in 2004, its properties have not been investigated extensively. In this paper, we will present some properties of β-wordlength pattern for four-level designs. These properties can help find better designs with quantitative factors. Keywords β-wordlength pattern, four-level, fractional factorial design MR(2010) Subject Classification 62K15 1 Introduction Fractional factorial designs have been extensively used in industrial, agricultural and scientific experiments. If we use an appropriate design to conduct experiments, we cannot only reduce the cost, but also keep high efficiency. In this sense, it is of great importance to provide a good criterion to compare different designs. For regular designs with qualitative factors, the minimum aberration (MA) criterion proposed by Fries and Hunter [6] has been widely employed. Later on, Deng and Tang [3] and Tang and Deng [10] generalized the criterion for non-regular designs with two levels. Using coding theory, Xu and Wu [13] proposed the generalized minimum aberration and justified it for the designs with qualitative factors under the ANOVA model. In fact, as pointed in Wu and Hamada [12], the idea behind the (generalized) MA criterion is the hierarchical ordering principle, i.e., lower order effects are more likely to be important than higher order effects and effects with the same order are considered to have equal importance. Some other criteria are also proposed in literature, such as the minimum generalized aberration (see Ma and Fang [7]), the general minimum lower order confounding (see Zhang et al. [14]) and the minimum hybrid aberration (Pang and Liu [9]). More discussions about the criteria for designs with qualitative factors can be found in Mukerjee and Wu [8], Wu and Hamada [12] and references therein. Received November 30, 2013, revised August 21, 2014, accepted October 8, 2014 Supported by NSFC (Grant No. 11271279), NSF of Jiangsu Province (Grant No. BK2012612) and Qing Lan Project 1) Corresponding author

1164 Sheng W. W., et al. For designs with quantitative factors, things become more complicated. As Cheng and Wu [1] and Fang and Ma [5] pointed out, designs with the same wordlength patterns (thus cannot be distinguished under the MA criterion) may have different statistical inference abilities. When multiple-level quantitative factors are involved, normally a polynomial regression model (or a response surface model) will be established to analyze data collected by a fractional factorial design. In such cases, level permutation of factors can alter the geometrical structure of the design and ultimately change the efficiency when the coefficients of the regression model are estimated. In order to systematically compare designs with quantitative factors, Cheng and Ye [2] proposed the β-wordlength pattern and classified geometrically non-isomorphic designs. Under their framework, lower degree effects are more likely to be important than higher degree effects and effects with the same degree are considered to have equal importance, thus designs with better β-wordlength pattern are recommended. In [2], β-wordlength pattern was originally defined using the concept of indicator function. Although the relationship between them is quite clear, it is not easy to analyze more properties of β-wordlength pattern due to the complicated definition of indicator function itself. Recently, Tang and Xu [11] simplified the expression of β-wordlength pattern, and obtained some interesting results related to the properties of β-wordlength pattern for three-level regular designs. The current paper continues their approach and provides some results for four-level regular designs. The rest of the paper is organized as follows. In Section 2, some basic concepts and notations are introduced. Section 3 discusses a simple case for 4 n 1 regular designs and Section 4 generalizes the situation for 4 n k regular designs. Finally, the last section gives some conclusion and discussion. 2 Concepts and Notations AdesignD, denoted by (N,s n ), with N runs and n factors, each with s levels, is an N n matrix, whose entries take values from the set of residue classes modulo p, i.e., Z p = {0, 1,...,s 1}. Let p 0 (x) 1andp j (x) beaj-th polynomial on Z s,where1 j s 1, such that s 1 0, if i j; p i (x)p j (x) = s, if i = j. x=0 The set {p 0 (x),p 1 (x),...,p s 1 (x)} is called an orthogonal polynomial basis (see Draper and Smith [4, Chapter 22]). For a design D =(d il ) N n,letf 1,...,F n be its n factors. When j 1 + + j n = j, F j 1 1 Fj n n is called an interaction with degree j. The orthogonal polynomial contrast coefficient of F j 1 1 Fj n n is an N 1 vector, whose i-th element is p j1 (d i1 ) p jn (d in ). Following [13], for a design D, denoted as (N,s n ), consider the ANOVA model, Y = X 0 α 0 + X 1 α 1 + + X n α n + ɛ, where Y is the N 1 response vector, α 0 is the intercept, X 0 is an N 1 all-one vector, α j is the vector of all interactions of j-th order, X j is the orthogonal contrast coefficient matrix of α j and ɛ represents the random error. Let n j =(s 1) j ( n j ), X j =(x (j) ik ) N n j,

Some Properties of β-wordlength Pattern for Four-level Designs 1165 and n j N A j (D) =N 2 k=1 x (j) 2 ik i=1 for j =0, 1,...,n. (2.1) Then the vector (A 1 (d),...,a n (d)) is called the (generalized) wordlength pattern of design D. For two designs D (1) and D (2),wecallD (1) haslessaberrationthand (2) if there exists a positive integer r, 1 r n, such that A r (D (1) ) <A r (D (2) ), and A i (D (1) )=A i (D (2) ), i =1,...,r 1. If there does not exist any other design, which has less aberration than D (1), then D (1) is said to be a (generalized) minimum aberration design. Following [2], for a design D, denoted as (N,s n ), consider the polynomial model, Y = Z 0 θ 0 + Z 1 θ 1 + + Z K θ K + ɛ, where Y is the N 1 response vector, θ j is the vector of all interactions of degree j, Z j is the orthogonal contrast coefficient matrix of θ j and ɛ represents the random error. Let Z j =(z (j) ak ) N n, j where n j represents the number of all interactions with degree j and n j N β j (D) =N 2 z (j) 2 ak for j =0, 1,...,K. (2.2) k=1 a=1 Then the vector (β 1 (d),...,β K (d)) is called the β-wordlength pattern of design D. HereK = n(s 1) is the highest degree of all interactions. Cheng and Ye [2] suggested that a good design with quantitative factors should sequentially minimize (β 1,...,β K ). Noticing that, for j 1 + + j n = j, z (j) ak = p j 1 (d a1 ) p jn (d an ), we have N n 2 β j (D) =N 2 p jl (d al ) for j =0, 1,...,K. (2.3) 0 j 1,...,jn s 1 j 1 + +jn=j a=1 l=1 When a design with qualitative factors is considered, levels of the factors may not be restricted on the set of residue classes Z p. For example, regular fractional factorial designs listed in textbooks are often constructed on finite fields. Let q beaprimeoraprimepowerandgf(q) be the finite field of order q. For a positive integer s, denote V (s, q) as the s-dimensional linear space, formed by all s-th row vectors on GF(q). Obviously, the cardinality of V (s, q) isq s. A design D is called to be a full factorial design if it contains all elements of V (s, q) as its runs. AdesignD is called to be a regular design if all its runs form a subspace of V (s, q). A regular deign D, denoted by q s k,representsa1/q k -fraction of the full factorial design, and all its q s k runs form a group. When we consider a design with quantitative factors and calculate its β-wordlength pattern, theentriesofthedesignmustbevaluesinz p. If the level q is a prime, it is well known that Z p and GF(q) are isomorphic, thus we can map all entries of a regular design to Z p and then calculate its β-wordlength pattern directly. If the level q is a prime power, things become complicated, because Z p and GF(q) are no longer isomorphic. In order to calculate the β- wordlength pattern of a regular design with a prime power level, we must firstly find a suitable map to transform all the entries of the design from GF(q) toz q. Obviously, different maps will lead to different β-wordlength patterns.

1166 Sheng W. W., et al. 3 Properties of β-wordlength Pattern for a 4 n 1 Design Table 1 A regular 4 3 1 design with β 3 (D)=0 0 0 0 0 1 2 0 1 1 0 ξ ξ 0 ξ 2 ξ 2 1 0 1 1 1 0 1 ξ ξ 2 1 ξ 2 ξ ξ 0 ξ ξ 1 ξ 2 ξ ξ 0 ξ ξ 2 1 ξ 2 0 ξ 2 ξ 2 1 ξ ξ 2 ξ 1 ξ 2 ξ 2 0 0 0 1 0 2 0 0 3 3 1 1 1 1 0 2 1 2 3 1 3 0 2 1 0 2 0 3 2 2 2 2 3 1 3 1 3 3 0 0 3 2 1 3 3 2 Typically, to construct a four-level regular design 4 n k, we first define a primitive element of GF(4), ξ, satisfying ξ 2 + ξ +1 = 0. A regular 4 n k design is formed by taking n k primitive columns and other k columns of their linear combinations. For example, the left side of Table 1 is aregular4 3 1 design, which takes the last column as the sum of the first two primitive columns. Notice here all operations are performed over GF(4). Now comes the problem: different ways of assigning levels to be 0, 1, 2, 3 lead to different β-wordlength patterns. For a regular design, we always have β 0 =1andβ 1 = β 2 = 0; but in general, β 3 of the design is not zero if its levels are assigned improperly. In fact, for regular 4 n 1 minimum aberration designs, we have the following lemma. Lemma 3.1 Let D be a regular 4 n 1 minimum aberration design on GF(4). Let D 0 be a design by conducting linear permutation of the dependent column of D. Letφ(x) be a bijection map from GF(4) to Z 4. Denote by φ(d 0 ) the design obtained by mapping all elements of D 0 using φ. Thenβ n (φ(d 0 )) 0. Proof Without loss of generality, assume the sum of each row of D 0 is b 0,whereb 0 GF(4). When n is odd, if b 0 {φ 1 (0),φ 1 (1)}, thenineachrowofd 0, there must be odd number of x, where x {φ 1 (0),φ 1 (1)}. Otherwise, if there exists a row of D 0 containing even number of elements denoted by x in {φ 1 (0),φ 1 (1)} and odd number of elements denoted by y in {φ 1 (2),φ 1 (3)}, then the sum of these x is either 0 or φ 1 (0)+φ 1 (1) (= φ 1 (2)+φ 1 (3), as φ 1 (0) + φ 1 (1) + φ 1 (2) + φ 1 (3) = 0); the sum of these y s is either φ 1 (2) or φ 1 (3). So the total sum of the row is either φ 1 (2) or φ 1 (3), which is a contradiction. So in each row of φ(d 0 ), there must be odd number of x,wherex {0, 1}. Noticing that D 0 is a 4 n 1 minimum

Some Properties of β-wordlength Pattern for Four-level Designs 1167 aberration design, we have A j (φ(d 0 )) = 0 for j =1, 2,...,n 1, which makes β j (φ(d 0 )) = 0 for j =1, 2,...,n 1. According to the definition of β-wordlength pattern, we have N n 2 β n (φ(d 0 )) = N 2 p 1 (d al ). (3.1) a=1 l=1 Since p 1 (0) and p 1 (1) are both less than 0, N n a=1 l=1 p 1(d al ) < 0, which makes β n (φ(d 0 )) 0. Similarly, if b 0 {φ 1 (2),φ 1 (3)}, thenineachrowofd 0, there must be even number of x, where x {φ 1 (0),φ 1 (1)}, which will make N n, ifb 0 {0,φ 1 (0) + φ 1 (1)}, then N a=1 n a=1 l=1 p 1(d al ) > 0. In the same vein, for even n l=1 p 1(d al ) > 0; otherwise, N n a=1 l=1 p 1(d al ) < 0. Lemma 3.1 tells us that if we use the same function to map levels of each column to 0, 1, 2, 3, then β n (D 0 ) 0. However, if we assign levels as the right design of Table 1, its β 3 is zero. The assignment method used in Table 1 can be generalized. In fact, we can obtain the following result related to regular minimum aberration 4 n 1 designs. Theorem 3.2 By assigning levels of a regular minimum aberration 4 n 1 design, there always exists a design D, satisfying β n (D) =0. Proof Let x i, i =1, 2,...,n 1 ben 1 primitive columns. The n-th column is formed by taking the sum of the primitive columns. Now we assign the levels, (0, 1,ξ,ξ 2 ), of the first n 2primitive columns as (0, 1, 2, 3); of the (n 1)-th primitive column as (1, 0, 2, 3) and of the last n-th column as (2, 1, 0, 3). When n is odd, for any row (η 1,...,η n 2,η n 1,η n ) in the original design, where η i GF(4), there exists a distinct row (ξ 2 +η 1,...,ξ 2 +η n 2,ξ+η n 1, 1+η n ), as they share the same sum. Let (z 1,...,z n 2,z n 1,z n ) be the corresponding row of (η 1,...,η n 2,η n 1,η n ) after conducting the above level assignment strategy. Then row (ξ 2 + η 1,...,ξ 2 + η n 2,ξ + η n 1, 1+η n ) will correspond to (3 z 1,...,3 z n 2, 3 z n 1, 3 z n ), which is the resultant row when conducting the reflection operator to (z 1,...,z n 2,z n 1,z n ). Thus, n i=1 p 1(3 z i )= n i=1 p 1(z i ). All rows of D can then be partitioned into reflection pairs, which leads to β n (D) = 0. Similarly, when n is even, for any row (η 1,...,η n 3,η n 2,η n 1,η n ) in the original design, there exists a distinct row (η 1,...,η n 3,ξ 2 + η n 2,ξ+ η n 1, 1+η n ) with the same sum. Let (z 1,...,z n 2,z n 1,z n ) be the corresponding row of (η 1,...,η n 3,η n 2,η n 1,η n ). Then (η 1,...,η n 3,ξ 2 +η n 2,ξ+η n 1, 1+η n ) will correspond to (z 1,...,z n 3, 3 η n 2, 3 z n 1, 3 z n ). Since n i=n 2 p 1(3 z i )= n i=n 2 p 1(z i ), the contribution of the above paired rows vanishes when β n (D) iscalculated. 4 Properties of β-wordlength Pattern for a 4 n k Design To generalize the idea in Section 3 to find a good map for 4 n k designs, we need more concepts. Let D be an s-level design on Z p. Mapping each run of D from {z 1,z 2,...,z n } to {s 1 z 1,s 1 z 2,...,s 1 z n } forms a new design, which is called the mirror-image of D. Definition 4.1 When the mirror-image of a design D is itself, the design D is said to be mirror-symmetric. For a mirror-symmetric design, it is easy to show the following property. Theorem 4.2 If D is mirror-symmetric, then for any odd j, β j (D) =0.

1168 Sheng W. W., et al. Proof The result simply follows from the fact that the orthogonal polynomial p m (x) isanodd function for any odd m, andp m (x) isanevenfunctionforanyevenm. For m =1,ξ and ξ 2, we define three maps from GF(4) to Z 4, such that if φ m (x) =z, then φ m (x + m) =3 z, wherex GF(4) and z Z 4. Now we consider 4 n k designs. Without loss of generality, assume the first n k columns are independent and the last k columns are linear combinations of the independent ones. Firstly, we have the following theorem. Theorem 4.3 Let D be a 4 n k design. Denote its j-th column by C j,where1 j n and assume the first n k columns are independent. For 1 j k, if C n k+j = C i1 + C i2 + + C iaj + ξ(c t1 + C t2 + + C tbj )+ξ 2 (C q1 + C q2 + + C qcj ), where i 1,...,i aj,t 1,...,t bj,q 1,...,q cj are all distinct positive integers in [1,n k], anda j,b j,c j are not all odd or not all even, then there exists a map ϕ from GF(4) to Z 4, such that β m (ϕ(d)) = 0, wherem is odd. Proof Letusfirstassumea j,b j,c j are odd, odd and even, respectively. To define the map of the design from GF(4) to Z 4, we consider two cases. For the first n k independent columns C j,where1 j n k, weusethemapφ ξ 2 defined in Section 3, while for the last k dependent columns C n k+j,where1 j k, weusethemapφ ξ. For any row (η 1,...,η n k,η n k+1,...,η n ) in the original design, where η i GF(4), there exists another row (ξ 2 + η 1,...,ξ 2 + η n k,η n k+1,...,η n), where η n k+j and η n k+j satisfy and a j η n k+j = η n k+j = (η is + ξ 2 )+ξ a j η is + ξ b j η ts + ξ 2 c j η qs b j (η ts + ξ 2 )+ξ 2 c j a j = η is + ξ b j η ts + ξ 2 c j = η n k+j + ξ 2 + ξ 3 +0 = η n k+j + ξ 2 +1 = η n k+j + ξ. (η qs + ξ 2 ) η qs + a j ξ 2 + ξ b j ξ 2 + ξ 2 c j ξ 2 Considering the maps φ ξ 2 for the first n k columns and φ ξ for the last k columns we use here, we know that if denote by (z 1,...,z n k,z n k+1,...,z n ) the corresponding row of (η 1,...,η n k,η n k+1,...,η n ) after conducting the above level assignment strategy, then there exists another row (3 z 1,...,3 z n k, 3 z n k+1,...,3 z n ) as the corresponding row of (ξ 2 + η 1,...,ξ 2 + η n k,ξ + η n k+1,...,ξ + η n ). That is to say, the transformed design is mirror-symmetric. So the result follows from Theorem 4.2. For other cases of a j,b j and c j, we can prove the result similarly, except for the different choices of the maps for the last k dependent columns C n k+j,where1 j k. More strictly, if (a j,b j,c j ) is (odd, even, odd) or (even, odd, even), then we use the map φ 1 ;if(a j,b j,c j )is

Some Properties of β-wordlength Pattern for Four-level Designs 1169 (even, odd, odd) or (odd, even, even), then we use the map φ ξ ;if(a j,b j,c j ) is (even, even, odd) or (odd, odd, even), then we use the map φ ξ 2. Notice that in Theorem 4.3, if a j,b j,c j are all odd or all even, then we cannot assign the same map, i.e., φ ξ 2 in the proof of Theorem 4.3, to each independent columns. Whether we can find a suitable map to make the transformed design mirror-symmetric depends on the structure of the original design. For example, the left design 4 3 1 in Table 1 is defined by C 3 = C 1 + C 2. Here a 1 =2andb 1 = c 1 = 0 are all even, but we can assign φ ξ 2, φ ξ and φ 1 to the three columns respectively to make the transformed design mirror-symmetric. However, we cannot find such a map for the following saturated design 4 5 3 in Table 2, which is defined by C 3 = C 1 + C 2 ; C 4 = C 1 + ξ C 2 ; C 5 = C 1 + ξ 2 C 2. Table 2 A regular 4 5 3 design 0 0 0 0 0 0 1 1 ξ ξ 2 0 ξ ξ ξ 2 1 0 ξ 2 ξ 2 1 ξ 1 0 1 1 1 1 1 0 ξ 2 ξ 1 ξ ξ 2 ξ 0 1 ξ 2 ξ 0 ξ 2 ξ 0 ξ ξ ξ ξ 1 ξ 2 0 1 ξ ξ 0 1 ξ 2 ξ ξ 2 1 ξ 2 0 ξ 2 0 ξ 2 ξ 2 ξ 2 ξ 2 1 ξ 1 0 ξ 2 ξ 1 0 ξ ξ 2 ξ 2 0 ξ 1 5 Conclusion and Discussion In this paper, we show that for a regular 4 n 1 minimum aberration design on GF(4), if we map all the elements of the designs to Z 4 using a unified transformation, we cannot get a design with β n = 0. However, if we map different columns with different transformations, we can choose an appropriate map to make many transformed 4 n k designs mirror-symmetric, which means β j s of the transformed designs are all zeros for odd j. Such results can help find designs with better β-wordlength patterns. Generalizing these properties to all 4 n k designs seems challenging, but is well worth further research.

1170 Sheng W. W., et al. References [1] Cheng, S. W., Wu, C. F. J.: Factor screening and response surface exploration (with discussion). Statist. Sinica, 11, 553 604 (2001) [2] Cheng, S. W., Ye, K. Q.: Geometric isomorphism and minimum aberration for factorial designs with quantitative factors. Ann. Statist., 32, 2168 2185 (2004) [3] Deng, L. Y., Tang, B.: Generalized resolution and minimum aberration criteria for Plackett Burman and other nonregular factorial designs. Statist. Sinica, 9, 1071 1082 (1999) [4] Draper, N. R., Smith, H.: Applied Regression Analysis (3rd Edition), Wiley, New York, 1998 [5] Fang, K. T., Ma, C. X.: Uniform and Orthogonal Designs (in Chinese), Science Press, Beijing, 2001 [6] Fries, A., Hunter, W. G.: Minimum aberration 2 k p designs. Technometrics, 22, 601 608 (1980) [7] Ma, C. X., Fang, K. T.: A note on generalized aberration in factorial designs. Metrika, 53, 85 93 (2001) [8] Mukerjee, R., Wu, C. F. J.: A Modern Theory of Factorial Design, Springer, New York, 2006 [9] Pang, F., Liu, M. Q.: Indicator function based on complex contrasts and its application in general facotrial designs. J. Statist. Plann. Inference, 140, 189 197 (2010) [10] Tang, B., Deng, L. Y.: Minimum G 2 -aberration for nonregular fractional factorial designs. Ann. Statist., 27, 1914 1926 (1999) [11] Tang, Y., Xu, H.: Permuting regular fractional factorial designs for screening quantitative factors. Biometrika, 101, 333 350 (2014) [12] Wu, C. F. J., Hamada, M.: Experiments: Planning, Analysis and Parameter Design Optimization (2nd Edition), Wiley, New York, 2009 [13] Xu, H., Wu, C. F. J.: Generalized minimum aberration for asymmetrical fractional factorial designs. Ann. Statist., 29, 1066 1077 (2001) [14] Zhang, R. C., Li, P., Zhao, S. L., et al.: A general minimum lower-order confounding criterion for two-level regular designs. Statist. Sinica, 18, 1689 1705 (2008)