Abstract
Strong orthogonal arrays (SOAs) have received more and more attention recently since they enjoy more desirable space-filling properties than ordinary orthogonal arrays. Among them, the SOAs of strength \(2+\) are the most advisable as they satisfy the same two-dimensional space-filling property as SOAs of strength 3 while having more columns for given run sizes. In addition, column-orthogonality is also a desirable property for designs of computer experiments. Existing column-orthogonal SOAs of strength \(2+\) have limited columns. In this paper, we propose a new class of space-filling designs, called group SOAs of strength \(2+\), and provide construction methods for such designs. The proposed designs can accommodate more columns than column-orthogonal SOAs of strength \(2+\) for given run sizes while satisfying similar stratifications and retaining a high proportion of column-orthogonal columns. Orthogonal arrays and difference schemes play important roles in the construction. The construction procedures are easy to implement and a large amount of group SOAs with \(s^2\) levels are constructed where \(s \ge 2\) is a prime power. In addition, the run sizes of the constructed designs are s times the ones of the orthogonal arrays used in the construction procedure. Thus they are relatively flexible.
Similar content being viewed by others
References
Bingham D, Sitter RR, Tang B (2009) Orthogonal and nearly orthogonal designs for computer experiments. Biometrika 96:51–65
Cheng CS, He Y, Tang B (2021) Minimal second order saturated designs and their applications to space-filling designs. Stat Sin 31:867–890
Fang KT, Li R, Sudjianto A (2006) Design and modeling for computer experiments. Chapman & Hall/CRC, Boca Raton
He Y, Cheng CS, Tang B (2018) Strong orthogonal arrays of strength two plus. Ann Stat 46:457–468
He Y, Tang B (2013) Strong orthogonal arrays and associated Latin hypercubes for computer experiments. Biometrika 100:254–260
He Y, Tang B (2014) A characterization of strong orthogonal arrays of strength three. Ann Stat 42:1347–1360
Hedayat AS, Sloane NJA, Stufken J (1999) Orthogonal arrays: theory and applications. Springer, New York
Lin CD, Tang B et al (2015) Latin hypercubes and space-filling designs. In: Dean A (ed) Handbook of design and analysis of experiments. Chapman & Hall/CRC, Boca Raton, pp 593–626
Liu H, Liu MQ (2015) Column-orthogonal strong orthogonal arrays and sliced strong orthogonal arrays. Stat Sin 25:1713–1734
McKay MD, Beckman RJ, Conover WJ (1979) A comparison of three methods for selecting values of input variables in the analysis of output from a computer code. Technometrics 21:239–245
Owen AB (1992) Orthogonal arrays for computer experiments, integration, and visualization. Stat Sin 2:439–452
Santner TJ, Williams BJ, Notz WI (2003) The design and analysis of computer experiments. Springer, New York
Tang B (1993) Orthogonal array-based Latin hypercubes. J Am Stat Assoc 88:1392–1397
Zhou Y, Tang B (2019) Column-orthogonal strong orthogonal arrays of strength two plus and three minus. Biometrika 106:997–1004
Acknowledgements
This work was supported by the National Natural Science Foundation of China (Grant Nos. 12131001, 11771219, 11771220 and 11871288), National Ten Thousand Talents Program, and Tianjin Development Program for Innovation and Entrepreneurship. The authorship is listed in alphabetic order.
Author information
Authors and Affiliations
Corresponding author
Additional information
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Appendices
Appendix
A.1 Some multiplication tables of GF(s)
A.2 Proofs of theorems
Proof of Theorem 1 First, in order to prove that the constructed design D can be collapsed into an OA(2n, u, 2, 2), we need to prove that A is an OA of strength 2. As the Kronecker sum with modulo 2 of an OA of strength 2 and the following matrix
is still an OA of strength 2, then A is an OA of strength 2.
To prove that any two distinct columns from the same group can be collapsed into an OA\((2n,2,4\times 2, 2)\) and an OA\((2n,2,2\times 4, 2)\), we need further prove that \((a_{kq}, b_{kq}, a_{kl})\) is an OA of strength 3, \(1\le k\le \lfloor m/2\rfloor +1\), \(q\ne l\). For \(k=1\), \((a_{1q}, b_{1q}, a_{1l})\) can be written as
As \(c_q\) is a two-level vector with entries 0 and 1, \(c_q+1_n\) can be obtained from \(c_q\) by replacing entry 0 by 1, and 1 by 0. It is easy to see that the array is an OA of strength 3. For \(k=2, \ldots , \lfloor m/2\rfloor +1\),
Due to the property of \(c_i\), we can obtain that it is an OA of strength 3. For \((a_{k2}, b_{k2}, a_{k1})\), the proof is similar.
To prove (i) of Theorem 1, we need to prove \((a_{1q}, b_{1q}, a_{kl})\) and \((a_{kl}, b_{kl}, a_{1q})\) are OAs of strength 3 when \(q\ne k-1\) and \(m-k+2\). \((a_{1q}, b_{1q}, a_{kl})\) and \((a_{kl}, b_{kl}, a_{1q})\) have the following expressions:
As \(c_i\), \(c_j\) and \(c_k\) are columns of OA of strength 2, we can obtain that both expressions are OAs of strength 3. Then for an odd m, the total number of pairs of columns that can achieve stratifications on \(4 \times 2\) and \(2\times 4\) grids is at least \((3m-3)(m-1)/2\), so \(p_1\ge (3m-3)/(4m-2)\); and for an even m, it is at least \((3m-4)m/2\), thus \(p_1\ge (3m-4)/(4m-2)\).
To prove (ii) of Theorem 1, without loss of generality, we assume that the levels of A and B have been centered. Then the inner product of any two columns of D \(d_{pq}^Td_{kl}=4a_{pq}^Ta_{kl}+2a_{pq}^Tb_{kl}+2b_{pq}^Ta_{kl}+b_{pq}^Tb_{kl}\), where \(p\ne k\) or \(q\ne l\). It is equal to zero if and only if \({a_{pq}}\) and \(a_{kl}\), \({a_{pq}}\) and \(b_{kl}\), \({b_{pq}}\) and \(a_{kl}\), \({b_{pq}}\) and \(b_{kl}\) are column-orthogonal. It is clear that \(a_{pq}^Ta_{kl}=b_{pq}^Tb_{kl}=0\). For the other two terms, without loss of generality, we assume \(p<q\) when \(p\ne q\). There are three cases:
-
(1)
when \(p=1\) and \(q=k-1\), \(a_{pq}^Tb_{kl}=-2n\), \(b_{pq}^Ta_{kl}=0\), then \(\rho (d_{pq}, d_{kl})=-4n/10n=-0.4\);
-
(2)
when \(p=1\) and \(q=m-k+2\), \(a_{pq}^Tb_{kl}=0\), \(b_{pq}^Ta_{kl}=-2n\), then \(\rho (d_{pq}, d_{kl})=-4n/10n=-0.4\);
-
(3)
for all other conditions, \(a_{pq}^Tb_{kl}=b_{pq}^Ta_{kl}=0\).
Thus, we can get \(p_2=(2m-3)/(2m-1)\). The proof is completed.
Proof of Theorem 2. To prove that the constructed design can be collapsed into an OA(sn, u, s, 2), we need to prove that A is an OA of strength 2. Note that any two distinct columns of A is
This form can be divided into two cases:
-
(1)
when \(i\ne j\), then \((c_{i}^l, c_{j}^{l'})\) is an OA, \((a_{pq}, a_{kl})\) is s times replication of it, so it is also an OA;
-
(2)
when \(i= j\), then \((c_{i}^l, c_{j}^{l'})=(c_{i}^l, c_{i}^{l'})\), as \(c_{i}^l\) and \(c_{i}^{l'}\) are obtained from \(c_{i}\) by replacing the levels with the elements of the difference scheme H, \((c_{i}^l, c_{i}^{l'})\) is a difference scheme. For any \(\sigma ,\sigma '\in GF(s)\), we must show that the number of runs \((\sigma , \sigma ')\) is equal to n/s. We know that n/s entries in \(c_{i}^l-c_{i}^{l'}\) are equal to \(\sigma -\sigma '\). Without loss of generality, we assume that the run \((\sigma , \sigma ')\) occurs in \((c_{i}^l, c_{i}^{l'})\), then there is a unique row \((\sigma , \sigma ')\) for each occurrence of \(\sigma -\sigma '\) in \(c_{i}^l-c_{i}^{l'}\), and the row will not occur in \((c_{i}^l+j1_n, c_{i}^{l'}+j1_n)\), \(j=2,\ldots ,s-1\), hence \((a_{pq}, a_{kl})\) is an OA.
To prove that any two distinct columns \(d_{kp}\) and \(d_{kq}\) from the kth group can be collapsed into an OA\((sn,2,s^2\times s, 2)\) and an OA\((sn,2,s\times s^2, 2)\), we need to prove that \((a_{kp}, b_{kp}, a_{kq})\) and \((a_{kq}, b_{kq}, a_{kp})\) are both OAs of strength 3. Let us first consider \((a_{kp}, b_{kp}, a_{kq})\) which has three cases.
Case 1:
As \((c_k^l, c_{m+1-k}^{l'})\) is an OA, it consists of the \(s^2\) different 2-tuples with \(n/s^2\) times each. It is obvious that \(c_{m+1-k}^{l'}\) is a level permutation of \(c_{m+1-k}^{l}\), hence \((c_k^l, c_{m+1-k}^{l}, c_{m+1-k}^{l'})\) consists of the \(s^2\) different 3-tuples \(n/s^2\) times each. It is similar for \((c_k^l+j1_n, c_{m+1-k}^{l}, c_{m+1-k}^{l'}+j1_n)\), \(j\in \{1,\ldots ,s-1\}\). For the same reason, \((c_{m+1-k}^{l}, c_{m+1-k}^{l'}+i1_n)\) and \((c_{m+1-k}^{l}, c_{m+1-k}^{l'}+j1_n)\) consist of different 2-tuples when \(i\ne j\), where \(i,j \in GF(s)\). Hence there are \(s^3\) different 3-tuples in \((a_{kp}, b_{kp}, a_{kq})\) with \(n/s^2\) times each, \((a_{kp}, b_{kp}, a_{kq})\) is an OA of strength 3.
Case 2:
As \(c_{m+1-k}^{l}\) is a level permutation of itself, this is a special case of Case 1.
Case 3:
It is obvious that \((c_k^l, c_k^{l'})\) consists of the s different 2-tuples n/s times each. As \((c_k^{l}, c_{m+1-k}^{l})\) is an OA, \((c_k^l, c_{m+1-k}^{l}, c_{k}^{l'})\) consists of the \(s^2\) different 3-tuples with \(n/s^2\) times each. It is similar for \((c_k^l+j 1_n, c_{m+1-k}^{l}, c_{k}^{l'}+j 1_n)\), \(j\in \{1,\ldots ,s-1\}\). As \((c_k^l, c_k^{l'})\) is a difference scheme, \((c_k^l+ i 1_n, c_k^{l'}+i 1_n)\) and \((c_k^l+ j 1_n, c_k^{l'}+j 1_n)\) consist of different combinations when \(i\ne j\), where \(i,j \in GF(s)\). Hence there are \(s^3\) different 3-tuples in \((a_{kp}, b_{kp}, a_{kq})\) with \(n/s^2\) times each, \((a_{kp}, b_{kp}, a_{kq})\) is an OA of strength 3.
For \((a_{kq}, b_{kq}, a_{kp})\), it is similar. Thus, we conclude that any two distinct columns from the same group can be collapsed into an OA\((sn,2,s^2\times s, 2)\) and an OA\((sn,2,s\times s^2, 2)\) now.
For (i) and (ii) of Theorem 2, when m is even, \(\lceil m/2\rceil =m/2\), the resulting design is a GSOA\((ns, (s^2)^{2e\cdot (m/2)},2+)\) with the total number of columns being me, the number of column pairs achieving stratifications on \(s^2\times s\) and \(s\times s^2\) grids is at least \(me(2e-1)/2\), so \(p_1\ge (2e-1)/(me-1)\); the number of column-orthogonal column pairs is \(me^2(m-1)/2\), then \(p_2=e(m-1)/(me-1)\). When m is odd, \(\lceil m/2\rceil =(m+1)/2\), the resulting design is a GSOA\((ns, (s^2)^{2e\cdot (m-1)/2+1},2+)\) with the total number of columns being \((m-1)e+1\), the number of column pairs achieving stratifications on \(s^2\times s\) and \(s\times s^2\) grids is at least \(e(2e-1)(m-1)/2\), so \(p_1\ge (2e-1)/(me-e+1)\); the number of column-orthogonal column pairs is \(e(m-1)(me-2e+2)/2\), then \(p_2=(me-2e+2)/(me-e+1)\).
The proof of (iii) of Theorem 2 is similar to that of Theorem 1. Any two columns of design D, \({d_{pq}}\) and \(d_{kl}\) are column-orthogonal if and only if \({a_{pq}}\) and \(a_{kl}\), \({a_{pq}}\) and \(b_{kl}\), \({b_{pq}}\) and \(a_{kl}\), \({b_{pq}}\) and \(b_{kl}\) are column-orthogonal. Without loss of generality, we assume that the levels of A and B have been centered. It is easy to see, \(a_{pq}^Ta_{kl}=a_{pq}^Tb_{kl}=b_{pq}^Ta_{kl}=0\). For \(b_{kl}^T{b_{pq}}\), there are two cases:
-
(1)
\(b_{kl}^T{b_{pq}}=0\) when \(p\ne k\);
-
(2)
when \(p=k\), without loss of generality, we assume \(q<l\). For this case,
$$\begin{aligned} b_{kl}^T{b_{kq}}= {\left\{ \begin{array}{ll} \ 0, &{} {\text{ if } 1\le q\le e,\ e+1\le l\le 2e}\\ nv_{ql},&{} {\text{ if } 1\le q<l \le e \text{ or } e+1\le q< l\le 2e,} \end{array}\right. } \end{aligned}$$(1)where \(v_{ql}=\langle h_{q({mod}\ e)},h_{l({mod}\ e)}\rangle \). So the correlation of \(d_{kq}\) and \(d_{kl}\) is
$$\begin{aligned} \rho _{(d_{kq},d_{kl})}= {\left\{ \begin{array}{ll} 12v_{ql}/f,&{} {\text{ if } s \text{ is } \text{ odd },}\\ \ \ 3v_{ql}/f,&{} {\text{ if } s \text{ is } \text{ even },} \end{array}\right. } \end{aligned}$$(2)where \(f=s(s^2-1)(s^2+1)\).
The proof is completed.
Rights and permissions
About this article
Cite this article
Liu, M., Liu, MQ. & Yang, J. Construction of group strong orthogonal arrays of strength two plus. Metrika 85, 657–674 (2022). https://doi.org/10.1007/s00184-021-00843-0
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00184-021-00843-0