1. Introduction
Computer experiments have been widely used recently to explore complex systems in many fields; space-filling and column orthogonality are desirable properties in the design of computer experiments [
1]. The space-filling property, which measures the uniformity of the design points in the experimental region, is a fundamental criterion for evaluating designs for computer experiments. Latin hypercube designs (LHDs), proposed by [
2], are widely used space-filling designs for computer experiments. An LHD with
n runs and
m factors, denoted as LHD
, is an
matrix with each column a permutation of
n equally spaced levels. Such a design achieves the maximum stratification in each dimension. Based on the effect sparsity principle [
3], for a high-dimensional design region, only a handful of the factors are expected to be active. In [
4], the author proposed orthogonal array (OA)-based LHDs which improve the low-dimensional projection properties of random LHDs. In [
5,
6], the authors discussed space-filling designs with good projection properties in low dimensions. Recently, ref. [
7] introduced strong orthogonal arrays, and [
8] proposed mappable nearly orthogonal arrays. Both of these two kinds of arrays have better space-filling properties than ordinary orthogonal arrays.
Column orthogonality is a desirable property for LHDs; when a linear model is fitted, this property ensures that the estimates of the main effects are uncorrelated. In addition, orthogonality can be viewed as a stepping stone to space-filling designs when Gaussian process models are considered [
9]. There are many ways to construct orthogonal LHDs (OLHDs); see, e.g., [
10,
11,
12,
13,
14,
15] and the references therein. Among them, the method of rotation has attracted widespread attention. In the extant literature, few works have simultaneously considered both the space-filling property and column orthogonality. In [
16], the authors constructed OLHDs which achieved stratifications on
or
grids, with most column pairs achieving stratifications on
grids. In [
17], the authors provided column-orthogonal designs (ODs) with two-dimensional stratifications. In [
18], the authors studied ODs with two-dimensional and three-dimensional stratifications, while [
15] proposed ODs with multi-dimensional stratifications.
The goal of the present paper is to construct ODs with stratifications on finer grids, i.e., , , and . We first introduce a new class of OLHDs with runs by rotating s-level OAs, where can be any positive even integer not smaller than 4 and . We additionally introduce a new class of ODs with flexible run sizes. All these designs guarantee desirable two-dimensional space-filling properties. Most column pairs of the resulting designs can achieve stratifications on grids. Moreover, many column pairs can satisfy stratifications on and grids. Furthermore, the resulting ODs can have or levels and mixed levels.
The rest of this paper is organized as follows:
Section 2 introduces preliminaries used in this paper;
Section 3 proposes the general construction method for OLHDs and extends it to accommodate more factors;
Section 4 concentrates on the construction of ODs with
levels,
levels, and mixed levels; finally, concluding remarks are provided in
Section 5. All proofs are deferred to
Appendix A.
2. Definitions and Notation
We use to denote a balanced design of n runs and m factors, with each of the levels from . When all the instances of are equal to s, the design is a symmetric balanced design . Further, if , it is an LHD, denoted as LHD.
A mixed-level orthogonal array (OA) with strength t and levels , denoted as OA, satisfies the requirement that all possible level combinations for any columns t occur with the same frequency. When all are equal to s, the array is symmetric and denoted as OA. For an OA, it must have for some integer , which is the index of the OA.
For an array with n runs and m factors, we say it achieves a stratification on an grid for some if the corresponding p columns of it can be collapsed into an OA .
The correlation between two vectors
and
is defined as
where
and
. The average correlation of a design
is defined as
Two vectors are said to be column-orthogonal if the correlation between them is 0. A design is said to be column-orthogonal, denoted as OD, if any two of its columns are column-orthogonal. Obviously, any OA with is an OD. Similarly, we have OLHD .
To facilitate the study of orthogonality, we sometimes center the
levels of an OD
into
Let be a Galois field of order and let be a Galois field of order s. We denote an matrix with entries from as , which is called a difference scheme if it satisfies the requirement that, for any i and j with , the vector difference of the ith and jth columns contains every element of equally often.
For two matrices
and
with entries from
, we define
where + is the addition defined on
.
For any design
A with entries from
, let
; we define
for
, and define
where
,
denotes the largest integer less than or equal to
h and
denotes the
lth column in the difference scheme
for
.
For a prime power
, let
be an OA
with entries from
and let
D be a difference scheme
. We now create
for
and define
For a prime s and integer d with and , we denote the d columns of an -run full factorial design as . Any generated column including each column of can be denoted as for some , and corresponds to a nonzero element in . Here, let .
As discussed in [
11], the corresponding columns of the first
non-zero elements of
,
modulo
form a regular design
D, where
is a primitive polynomial of order
d. Any
d consecutive columns of
D form a full factorial design, denoted as
and
mod
,
. Let
, defining
for
,
. Then, we have
and
Without particular explanation, in this paper, s is a prime, , d is an integer with , , and . We provide an illustrative example in the following.
Example 1. For and , we denote the full factorial design as . Here, , with the primitive polynomial . Then, , , , modulo are 1, x, , , which correspond to columns , , , . Similarly, we have the elements of modulo . For example, is obtained by modulo . The obtained full factorial designs are shown in Table 1, where is the jth column in for and . 3. Construction of Orthogonal LHDs
This section first introduces a rotation method in Algorithm 1 to construct OLHDs with attractive stratification properties, then generalizes the method to enlarge the columns of these LHDs.
To make it easier for readers to understand the algorithm, we provide the flowchart in
Figure 1 to explain the algorithm.
To measure the stratification properties of a design with
m columns, we define the following two proportions:
where
is the number of column pairs that achieve stratifications on
grids and
is the number of column pairs that achieve stratifications on
and
grids. The properties of the designs in Algorithm 1 are summarized in Theorem 1.
Algorithm 1 Construction of OLHDs |
- Input:
, , and integer k. - 1:
Let as defined in ( 6) and replace the levels of in each with . - 2:
Obtain a difference scheme . For a given k, let with for , where is defined in ( 2). - 3:
Replace the levels of with entries from as in ( 1) and denote the resulting design as . - 4:
Obtain , where , , and is defined in ( 3). - Output:
Design Z.
|
Theorem 1. Design Z in Algorithm 1 is an OLHD where and , and has the following properties:
- (1)
Any two columns achieve a stratification on an or grid;
- (2)
The proportion of column pairs achieving stratifications on grids satisfies ;
- (3)
The proportion of column pairs achieving stratifications on and grids satisfies .
For this, we use the following illustrative example.
Example 2. For and , we denote the four independent columns as , , , , while the generated columns of these four columns are denoted as , , , , , , , , , , . We have with and the primitive polynomial . It is easy to obtain three full factorial designs , , and . Then, we have , , and . Thus, , which is displayed in Table 2. In this way, we obtain a difference scheme , denoted aswith , where for . We can obtain by replacing the levels of with entries from . Then, we rotate by to generate an OLHD, where The resulting OLHD is displayed in Table A1 of Appendix B. From Table 3, it is apparent that 260 out of all 276 (i.e., ) column pairs achieve stratifications on grids, more than of column pairs achieve stratifications on or grids, and of column pairs achieve stratifications on or grids. To illustrate the projection property of the resulting OLHD, we display the pairwise scatter plots of the (1–5)th columns of the design in Figure 2. From Figure 2, it can be seen that all the column pairs of the first five columns achieve stratifications on grids, and most of the same column pairs achieve stratifications on and grids. Other column pairs perform similarly. Example OLHDs constructed by Algorithm 1 are listed in
Table 4. Without loss of generality, we only show the lower bounds of
and
in
Table 4, which are denoted as
and
, respectively.
As shown in
Table 4, the lower bounds of
are very close to 1, while that of
is close to 1 when the run size is large, which means that nearly all the column pairs achieve stratifications on
grids and that most column pairs achieve stratifications on
and
grids.
A comparison between the OLHDs obtained using Algorithm 1 and the OLHDs in [
14,
15,
16] is presented in
Table 5. Compared with this class of designs, the resulting OLHDs satisfy two-dimensional space-filling properties on finer grids, i.e.,
,
, and
. The OLHDs in [
14] satisfy stratifications on
and
grids, while the designs in [
16] satisfy stratifications on
grids. Thus, the OLHDs based on Algorithm 1 satisfy better two-dimensional space-filling properties. Moreover, these OLHDs are able to accommodate more factors than OLHDs in [
15], and can fill the gap between the run sizes of the available OLHDs in [
16]. For example, we can construct OLHDs of 64 and 1024 runs, while such designs are not available in [
16].
Furthermore, we can construct OLHDs with more columns through Algorithm 2. The flowchart of Algorithm 2 is shown in
Appendix B.
Algorithm 2 Enlarging the columns of OLHDs |
- Input:
, , and integer k. - 1:
Let be an OLHD . For , obtain matrix , by replacing the s levels of B with , respectively, where B = is the same as in Section 2. - 2:
For , obtain from per ( 6) in Section 2 and replace the levels of each with . - 3:
For a given integer k and , let with for , where is defined in ( 2). - 4:
For , replace the levels of with entries from in ( 1) and denote the resulting design as . Construct , where , , and is the same as in Algorithm 1. - Output:
Design M.
|
Corollary 1. Design M obtained by Algorithm 2 is an OLHD where and . Each sub-design achieves the same stratifications with Z in Algorithm olhds1 for . At least column pairs of M achieve stratifications on grids in all the two dimensions.
In Step 1, we can choose OLHDs obtained by [
15,
19] when
5, 7, 11, and 17, respectively. According to Algorithm 2, we can obtain OLHDs with
t sub-designs, with each satisfying the same stratification properties in Theorem 1. Moreover, when
, per Algorithm 2 we can obtain an OLHD
that can accommodate more columns than OLHD
in [
14] and has more attractive space-filling properties.
4. Construction of Orthogonal Designs
This section introduces three rotation methods for constructing ODs. The first two methods can construct ODs with
and
levels, respectively, while the third can obtain mixed-level ODs. We first present the construction of ODs with
levels and investigate their properties. The construction method is provided in Algorithm 3, and the flowchart is shown in
Appendix B.
Algorithm 3 Construction of -level ODs |
- Input:
and . - 1:
Let with , and let D be a difference scheme with entries from as defined in Section 2. Replace the levels of in each with and define
- 2:
Divide each into g or groups for as if or if where each has two columns. Then, order ’s as
- 3:
Replace the levels of each with entries from and denote the resulting design as . Take two successive instances of at a time in the order given in ( 8), and obtain sets of four columns, denoted as . Combine for together:
- 4:
Create
where
is a rotation matrix up to a constant. - Output:
Design .
|
Based on the form of the rotation matrix, it is easy to see that any column
x in
obtained in (
9) has the following form:
where
e and
are the two columns in some
with
for some
l; here,
is a column which is not in
. We call
e the leading column of
x to facilitate later study. Now, we can consider the mapping
with
collapsing the
levels in
into
levels in
. For example, when
, the 64 levels are collapsed into 4 levels by the mapping, as follows:
Then, we consider the mapping
with
collapsing the
levels in
into
levels in
. For example, when
, the 64 levels are collapsed into 16 levels by the mapping, as follows:
The resulting design is orthogonal and achieves stratifications on or grids; most column pairs can achieve stratifications on grids. Moreover, column pairs can achieve stratifications on and grids as well. We can summarize the properties of in the following theorem.
Theorem 2. Design in (9) is an OD, where and ; can be partitioned into disjoint groups of columns, each with the following properties: - (1)
Any two distinct columns achieve a stratification on an or grid;
- (2)
Most column pairs achieve stratifications on grids, and the proportion is not less than ;
- (3)
The proportion of column pairs achieving stratifications on and grids satisfies .
From Theorem 2, it can be understood that the obtained ODs have appealing stratification properties. For example, for and at least of all column pairs of can achieve stratifications on grids. Furthermore, many column pairs of achieve stratifications on finer and grids (). The lower bound of this proportion is relatively loose. Below, we provide an illustrative example.
Example 3. Consider the same conditions in Example 2 with , , and ; we can obtain and the difference scheme . From Step 1, we have , where Then, we divide into two groups, as follows: for , where each has two columns, and order the as follows: We replace the levels of each with entries from and denote the resulting design as . Taking two successive instances of at a time in the order given in (13), we obtain sets of four columns, denoted as . Then, we can obtain an OLHD throughwhich is displayed in Table A2 of Appendix B. The stratification properties of are summarized in Table 6. It can be seen that this design has the same number of column pairs achieving stratification on a grid as the one in Example 2, and has more column pairs achieving stratifications on or grids than the one in Example 2 with . Furthermore, by calculation, we can say that the obtained OLHD achieves stratifications on a grid in 140 out of all 276 (i.e., ) and on a grid in100 out of all 276 (i.e., ). The OAs and difference schemes used in the construction are available in [
20] and the library of OAs (
http://neilsloane.com/oadir/index.html, accessed on 16 March 2023). It is easy to show that
is an OLHD
when
.
Table 7 summarizes example ODs constructed by Algorithm 3. Their space-filling properties are characterized by
and
. Similar to
Section 3, we only list the lower bounds of
and
in
Table 7, denoted as
and
, respectively. As shown in
Table 7, the lower bounds of
are very close to 1 and those of
are quite large in most cases, which means that nearly all the column pairs of these ODs achieve stratifications on
grids, and that most column pairs achieve stratifications on finer
and
grids as well.
Next, we introduce the construction of ODs with levels with the same space-filling properties as the designs obtained in Algorithm 3. The construction method is provided in Algorithm 4.
Algorithm 4 Construction of -level ODs |
- Input:
and . - 1:
Let matrices B, F, D, , be the same as in Algorithm 3. - 2:
For , define with . - 3:
Create
where
where is repeated times and all off-diagonal sub-matrices of V are zero matrices. - Output:
Design Y.
|
For the resulting design, it is easy to obtain the following theorem.
Theorem 3. Design Y in (14) is an OD with , that can be partitioned into disjoint groups of columns with the same stratification properties as in Theorem 2. Compared with the ODs constructed in Algorithm 3, design Y has lower levels and can accommodate more columns than design in Algorithm 3 when d is an odd number. Now, we turn to an illustrative example.
Example 4. Considering the same conditions in Example 3, we first obtain the Ei,js. For , we define and . We can construct an OD bywhich is displayed in Table A3 of Appendix B. It can be seen that the design points are well-scattered in the two-dimensional projections of the resulting OD, and it has the same space-filling properties as the OD constructed in Example 3. Table 8 lists example ODs obtained by Algorithm 4. Compared with the ODs constructed using Algorithm 3, the obtained designs have lower levels and more columns when
d is odd. The resulting designs have the same stratification properties as the ODs constructed by Algorithm 3. These ODs have more flexible run sizes than the OLHDs obtained by Algorithm 1.
Now, we consider the construction of mixed-level ODs, which are very useful when the factors cannot have the same number of levels. The construction method is provided in Algorithm 5.
Algorithm 5 Construction of mixed-level ODs |
- Input:
and . - 1:
Let matrices B, F, D, , and be the same as in Algorithm 3. Then, order the E i,js as in ( 8). First, take two successive instances at each time in the above list and take a total of times (where ) to sets of four columns each. Center the levels of each column into in ( 1) and denote these sets as . - 2:
Take one at a time in the remaining list of ( 8), thus sets of two columns each can be obtained. Similarly center the levels of each column into and denote these sets as . - 3:
Create
where and are the rotation matrices provided in ( 3). - Output:
Design H.
|
For the resulting design H, the following theorem holds.
Theorem 4. Design H in (15) is an OD with that can be partitioned into disjoint groups of columns with the same stratification properties as design in Algorithm 3. Table 9 summarizes example ODs constructed by Algorithms 3–5 for practical needs.
5. Conclusions, Limitations, and Future Research
In this paper, we have proposed a new rotation method to generate OLHDs that can achieve stratifications on or grids; moreover, most column pairs can achieve stratifications on grids, and a large portion of column pairs can achieve stratifications on and grids. Furthermore, we introduce a new class of space-filling ODs with levels, levels, and mixed levels, which can guarantee desirable stratifications in two dimensions.
It is worth noting that the resulting OLHDs and ODs enjoy stratifications on finer grids that cannot be satisfied by the existing space-filling designs. To the best of our knowledge, this is a new development in the literature. The theoretical constructions are well established. All these properties make the resulting designs competitive for computer experiments. The proposed designs are constructed systematically, without relying on any optimization algorithm, and the methods are efficient from a time perspective.
Next, we provide a simple simulation example to illustrate the performance of the resulting OLHDs from the model perspective. First, we use the following three methods to screen the active effects: the least absolute shrinkage and selection operator (LASSO) in the ‘glmnet’ R package, the smoothly clipped absolute deviation (SCAD) in the ‘ncvreg’ R package, and the stepwise linear model regression following the AIC criterion; the first two methods were recently used in [
21], and further details can be found there. Suppose the true model is
where
,
, and the random error
. We use the first twelve columns of the designed OLHD
in Example 2 to generate the responses. All screening results are provided in
Table 10, where the truly active factors are marked with the superscript ‘‘a’’ and identified ‘‘active factors’’ are indicated with ‘‘•’’.
From
Table 10, it can be seen that each method identifies eleven active effects, and the three methods all obtain the same ten truly active effects. Next, we entered all of these thirteen active effects into the model in order to test the significance of the coefficients.
Table 11 shows the estimates and significance test results.
From
Table 11, the true active effects can be correctly identified as
. Then, we fitted the model in (
17), with the
of the model being 0.9378.
This result is very close to the real model.
The computation was implemented on a personal computer with an Intel i5-4210H CPU and 2.90 GHz, which needed 0.384 seconds to generate the design, screen the active effects, and fit the model.
Due to the utilization of the rotation method, the run sizes of the obtained designs are restricted to prime powers, and certain two-dimensional stratification properties are not satisfied by all the column pairs, only a large proportion of them. Due to time restrictions, we do not provide an empirical example here. These issues are, however, deserving of future work. In this paper, we consider the space-filling properties measured by the two-dimensional stratifications; however, other criteria, for example, the maximin distance criterion, can be suitable choices as well. Constructions of column-orthogonal designs using the maximin distance and flexible run sizes are interesting topics for future research.