1. Introduction
Fractional factorial designs are among the most popular experimental designs in various fields. The minimum aberration criterion [
1] and its extension, the generalized minimum aberration criterion [
2,
3] are commonly used for comparing fractional factorials.
If the goal is to conduct a sensitivity analysis between inputs and outputs, computer experiments are commonly used, especially if the input-output relation of experiments is likely to have some curvature. With the rapid increase in computational power, more and more large fractional factorial designs are used in large-scale computer experiments in practice. For example, researchers at Johns Hopkins University initially employed a design with 512 runs followed by 352 additional runs to resolve the aliasing of two-factor interactions in a ballistic missile defense project [
4]. The second scenario was explored using a resolution V design with 4096 runs obtained using SAS’s PROC FACTEX. Another example is reported in [
5] that designs with over 600 runs, and as many as 53 parameters were used in computer simulations at Los Alamos National Laboratory. Bettonvil and Kleijnen [
6] discussed a case study on the CO
2 greenhouse effect using a deterministic simulation model with 281 factors [
7]. Kleijnen et al. [
8] applied sequential bifurcation to a practical discrete event simulation of a supply chain centered around the Ericsson company in Sweden, involving 92 factors. Motivated by practical applications, the construction method for designs with a large size (a large number of runs and/or a large number of factors while keeping the run size relatively small) is urgently needed, and it is an important issue. In particular, the construction of large-size designs from small designs has attracted more and more attention. For two-level designs, doubling plays an important role in the construction of two-level designs of resolution IV [
9]. Given a two-level regular fractional factorial design of resolution IV, the method of doubling produces another design of resolution IV which doubles both the run size and the number of factors of the initial design. One can refer to [
10,
11,
12,
13] for more details about doubling.
When both the factorial main effects and some quadratic effects need to be detected, it is very necessary to apply multi-level designs for that purpose. Since three-level designs are the most commonly used designs with factor levels higher than two, the three-level fractional factorials constructed in this paper provide an alternative for this demand in practical applications, such as elemental factorial analysis of nanostructure congeners, capturing curvature or active pure-quadratic effects of quality control, and so on. For designs with more than two levels, based on the level permutation method [
14,
15,
16,
17,
18,
19], the doubling process of two-level designs has been naturally extended to three-level designs. A method of tripling for three-level designs, which triples both the run size and the number of factors of the initial three-level design, is proposed by combining all possible level permutations of its initial design in Ou et al. [
20] and Li and Qin [
21], respectively.
In this paper, we aim to explore the additional properties of triple designs using indicator functions, which provide a closer look at triple designs. The indicator function has been adopted by Fontana et al. [
22] to study the two-level factorial designs. It allows us to discuss not only the regular factorial designs but also non-regular factorial designs. The indicator function has become a powerful tool for studying general two-level factorial designs; see, for example [
12,
23,
24,
25,
26,
27,
28,
29,
30,
31,
32]. Furthermore, Cheng and Ye [
14], Pistone and Rogantin [
33,
34], and Pang and Liu [
35] established that general fractional factorial designs, three or higher levels or multilevel, can also be represented with indicator functions.
The contribution of this paper is twofold. First, the closer relationship between a triple design and its initial design is built with indicator functions. It is shown that the indicator function of a triple design is decided uniquely by one of its initial designs. The internal structure of a triple design is explored from the word characteristic of its indicator function, and a new look of triple designs is provided. Second, the properties of a triple design and its projections, such as resolution and orthogonality, are studied by the expression of the indicator function. Given a three-level fractional factorial design of resolution III (IV), we show that its triple design is a design of resolution III, and the projections of a triple design also is a design of resolution III (IV). These theoretical results provide a solid foundation for the tripling construction method for a design with a large size, in which the constructed designs have better properties, such as high resolution and orthogonality, and are recommended for use in practice. The triple designs discussed in this paper are competitive in large-scale computer experiments, such as aerospace, quantum communication, intelligent manufacturing, and so on.
The paper is organized as follows. In
Section 2, some notations and preliminaries are included. In
Section 3, the indicator function of a triple design is expressed based on the indicator function of its initial design, and the section provides a closer look at the internal structure of a triple design by its indicator function. In
Section 4, the close relationships between a triple design and its initial design are built from properties such as resolution and orthogonality. Finally, some conclusions are given in
Section 5. For clarity, we have placed all the proofs in
Appendix A.
2. Notations and Preliminaries
Let
be a
full factorial design [
36] with
s three-level factors, where the three levels of each factor are
, and
, i.e., evenly spaced solutions of
on the unit circle in the complex plane
. Accordingly, the design points of
are just the solutions of the polynomial system {
} on
. Under this level coding strategy, the polynomial representation of the indicator function benefits from its cube-free property. An
n-run unreplicated three-level fractional factorial design
is regarded as a subset of
, each row of
corresponds to a run and each column of
to an experimental factor in the design. Let
be the set of
U-type designs with
n runs and
s three-level factors. A design
in
can be presented as an
matrix with entries
(or equivalently with entries
), where each entry appears equally often in each column of
. If all the possible
level combinations corresponding to any
t columns of design
appear equally often, design
is called an orthogonal array [
37] of strength
t and denoted by
.
The indicator function
of design
, due to [
22,
26], is defined as a function on
full factorial design
such that
Under the constraint
, the indicator function
of design
can be uniquely cube-free represented by the complex polynomial function defined on
as
where
L is the set of all
s tuples
, that is,
and
where
is the conjugate of
. Therefore, an indicator function of design
has the unique cube-free polynomial representation on
.
The coefficients
of
reflect some basic information of design
. In particular,
, where
n is the run size of
. In other words,
is just the ratio between the number of points of
and the number of points of
. The coefficients
of
satisfy
. A design is a regular design if and only if
for any
. A word of the design
is defined as the term with a non-zero coefficient (except the constant) in the indicator function
of design
. Following Li et al. [
24] for two-level designs, the length of a word
is defined as
, where
represents the number of letters in the word
, i.e., the number of nonzero elements in
. The length of the shortest word of
is called the generalized resolution of design
.
The definition of an indicator function follows immediately from the following lemma. The proof of the lemma is straightforward and is omitted here.
Lemma 1. Let and be indicator functions of two disjoint designs and , respectively. The indicator function of design is then
Following Cheng and Ye [
14], the generalized word-length pattern and generalized minimum aberration criterion of three-level design
are defined as follows.
Definition 1. Let be an n runs s three-level factors fractional factorial design, is its indicator function. The generalized word-length pattern of is defined aswhere is the complex module. The generalized minimum aberration criterion is to sequentially minimize for . The resolution of equals the smallest t such that . Remark 1. The definition of the generalized word-length pattern of in Definition 1 is equivalent to the definition in Xu and Wu [3]. Example 1. Consider a regular three-level design with defining relations , where (or ) are the facotor labels of design . Accordingly, the definition contrast subgroup of is . The indicator function of isFrom the expression of given above, one can easily find that design is a regular design since for any . Moreover, following Definition 1, the generalized word-length pattern of is . Suppose
is a three-level design with
n runs
s three-level factors, then there are six kinds of level permutations of
, which are listed in
Table 1.
Ou et al. [
20] and Li and Qin [
21] proposed a new concept named the tripling of three-level design
based on all of the possible level permutations of
shown in
Table 1, which is defined below.
Definition 2 (Ou et al. [
20])
. Suppose is a three-level design with n runs s three-level factors, are the level permutations of as listed in Table 1. The matrix is defined as triple design of . 3. Indicator Function of Triple Design
In this section, we aim to explore the link between a triple design and its initial design by using the tool of the indictor function, which provides a closer look at triple design .
Denote
,
and
as the column blocks of
, and
,
as the row blocks of
. Based on the polynomial form of the indicator function of a fractional factorial design in (
1), the indicator functions of
can be written as
where
is the corresponding run in
for given
.
Define
. Similarly, the indicator function of
can be written as
where
for
, and the indicator functions of
can be written as
For any
and
, define
For any
, there exists
such that
. Accordingly, the indicator functions of
in (
6) can be rewritten as
where
By Lemma 1, the indicator function of
in (
5) can be rewritten as
where
For any , define , and denote , , that is, is the number of j in . For any and , define , and , . Moreover, denote , and .
Based on the above notations, the following three lemmas provide some properties of the term
in (
4).
Lemma 2. Let be an n runs s three-level factors fractional factorial design. For and , we have
(a) if , ; (b) if , ;
(c) if , ; (d) if , .
Lemma 3. Let be an n runs s three-level factors fractional factorial design.
(a) For and , we have when one of the following conditions satisfies:
(i) and ; (ii) and ; (iii) and .
(b) For and , we have when one of the following conditions satisfies:
(i) and ; (ii) and ; (iii) and ; (iv) and ; (v) and ; (vi) and .
Lemma 4. Let be an n runs s three-level factors fractional factorial design, and respectively be the indicator functions of and , then .
Based on Lemmas 2–4, the following two theorems provide the relationships between the coefficients
’s of
given in (
7) and the coefficients
’s of the indicator function
of
.
Theorem 1. Let be an n runs s three-level factors fractional factorial design, be the indicator function of , then for any and , we have
(a) if for , ;
(b) if for , , particularly, if , ;
(c) if for , , particularly, if , ;
(d) if for , .
Remark 2. Theorem 1 shows the close relationship between the initial design and the subdesigns of the triple design , . Since , the word in with coefficient if and only if for , the word in with coefficient if and only if for . Based on Theorem 1, one can easily obtain some properties of the projection designs of , which is given in Theorem 5.
Theorem 2. Let be an n runs s three-level factors fractional factorial design, be the indicator function of , then for any , we have
(a) when one of the following three conditions satisfies:
(i) and for ; (ii) and for ;
(iii) and for .
(b) when one of the following six conditions satisfies:
(i) and for ; (ii) and for ;
(iii) and for ; (iv) and for ;
(v) and for ; (vi) and for .
Based on Theorems 1 and 2, the following theorem gives the expression of the indicator function of triple design from the indicator function of its original design .
Theorem 3. Let be an n runs s three-level factors fractional factorial design, be the indicator function of , then the indicator function in (8) of the triple design of can be expressed as follows Remark 3. Theorem 3 gives the analytical relationship between the indicator functions of initial design and its triple design . One can easily find that the coefficients of the indicator function of are completely decided by the coefficients of the indicator function of .
4. Some Basic Properties of Triple Designs
In this section, some good properties of are provided. Based on these good properties, one can easily construct large designs with resolution III or IV.
The following result, whose two-level design version of doubling can also be found in Chen and Cheng [
9], reveals the crucial role played by the method of tripling in constructing designs of resolution III.
Theorem 4. If is a design of resolution III, then is also a design of resolution III. Particularly, if is a regular design of resolution III, then is also a regular design of resolution III.
Remark 4. Theorem 4 shows that if the resolution of is III, the resolution of triple design remains the same as its initial design . Theorem 4 has no counterpart for designs of higher resolution than III. In fact, if the resolution of is higher than III, the resolution of can only achieve III. For any such that and , for and for (or for and for ), , and the coefficient of is . Hence, must contain word(s) with a length of 3.
In the following, the designs with resolution IV are constructed by the projections of
. Denote the projection designs of
as
Theorem 5. If is a design of resolution III (or IV), then are also designs of resolution III (or IV) for . Particularly, If is a regular design of resolution III (or IV), then are also regular designs of resolution III (or IV) for .
Remark 5. Theorem 5 shows that if the resolution of is III, the resolution of the projection designs and of triple design remains the same as its original design . Theorem 5 has no counterpart for designs of higher resolution than IV. In fact, if the resolution of is higher than IV, the resolution of and can only achieve IV. For any such that and , for and . According to Theorem 1, the coefficient of in the indicator function of is , namely, must contains word(s) with a length of 4. Hence, the resolution of can only achieve IV. The same is true for and .
From Theorems 4 and 5, the following result is obvious.
Corollary 1. If is an orthogonal array of strength 2, then both and are orthogonal arrays of strength 2 for .