Abstract
Mappable nearly orthogonal arrays were recently proposed as a new class of space-filling designs for computer experiments. Inspired by mappable nearly orthogonal arrays, we propose several new classes of space-filling designs. The corresponding construction methods are provided. The resulting designs are more space-filling than mappable nearly orthogonal arrays while accommodating a large number of factors. In addition to the space-filling properties, the column orthogonality is also desirable for designs of computer experiments. Among the new constructed designs, one class is column-orthogonal, and the other two classes, providing many new column-orthogonal designs, are nearly column-orthogonal in the sense that each column is column-orthogonal to a large proportion of the other columns. The constructed designs are good choices for computer experiments due to their attractive space-filling properties and column orthogonality. The proposed construction methods are flexible in the choices of an orthogonal array and/or a strong orthogonal array and their usefulness is appealing. Many newly constructed space-filling designs are tabulated. The expansive replacement method and the generalized doubling play key roles in the constructions.
Similar content being viewed by others
References
Chen HH, Cheng CS (2006) Doubling and projection: a method of constructing two-level designs of resolution IV. Ann Stat 34:546–558
Fang KT, Li R, Sudjianto A (2006) Design and modeling for computer experiments. Chapman and Hall/CRC, Boca Raton
Fang KT, Liu MQ, Qin H, Zhou YD (2018) Theory and application of uniform experimental designs. Springer-Science Press, Singapore, Beijing
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
He Y, Cheng CS, Tang B (2018) Strong orthogonal arrays of strength two plus. Ann Stat 46:457–468
Hedayat AS, Sloane NJA, Stufken J (1999) Orthogonal arrays: theory and applications. Springer, New York
Johnson ME, Moore LM, Ylvisaker D (1990) Minimax and maximin distance designs. J Stat Plann Inference 26:131–148
Li W, Liu MQ, Yang JF (2021) Column-orthogonal nearly strong orthogonal arrays. J Stat Plann Inference 215:184–192
Li W, Liu MQ, Yang JF (2022) Construction of column-orthogonal strong orthogonal arrays. Stat Papers 63:515–530
Lin CD, Mukerjee R, Tang B (2009) Construction of orthogonal and nearly orthogonal Latin hypercubes. Biometrika 96:243–247
Liu H, Liu MQ (2015) Column-orthogonal strong orthogonal arrays and sliced strong orthogonal arrays. Stat Sin 25:1713–1734
Liu H, Sun F, Lin DKJ, Liu MQ (2023) On construction of mappable nearly orthogonal arrays with column orthogonality. Commun Math Stat. https://doi.org/10.1007/s40304-023-00333-x
McKay MD, Beckman RJ, Conover WJ (1979) Comparison of three methods for selecting values of input variables in the analysis of output from a computer code. Technometrics 21:239–245
Mukerjee R, Sun F, Tang B (2014) Nearly orthogonal arrays mappable into fully orthogonal arrays. Biometrika 101:957–963
Owen AB (1992) Orthogonal arrays for computer experiments, integration and visualization. Stat Sin 2:439–452
Santner TJ, Williams BJ, Notz WI (2018) The design and analysis of computer experiments, 2nd edn. Springer, New York
Shi C, Tang B (2019) Design selection for strong orthogonal arrays. Can J Stat 47:302–314
Shi C, Tang B (2020) Construction results for strong orthogonal arrays of strength three. Bernoulli 26:418–431
Sun F, Tang B (2017a) A method of constructing space-filling orthogonal designs. J Am Stat Assoc 112:683–689
Sun F, Tang B (2017b) A general rotation method for orthogonal Latin hypercubes. Biometrika 104:465–472
Tang B (1993) Orthogonal array-based Latin hypercubes. J Am Stat Assoc 88:1392–1397
Wang C, Yang J, Liu MQ (2022) Construction of strong group-orthogonal arrays. Stat Sin 32:1225–1243
Zhou YD, Tang B (2019) Column-orthogonal strong orthogonal arrays of strength two plus and three minus. Biometrika 106:997–1004
Acknowledgements
The authors are grateful to two anonymous referees for their insightful comments and suggestions, and would like to thank Professor Boxin Tang for his valuable comments and suggestions. This work was supported by the National Natural Science Foundation of China (Grant Nos. 12131001, 12201042, 12271270, 12226343 and 11871033), Tianjin Natural Science Foundation of China (Grant No. 20JCYBJC01050), China Postdoctoral Science Foundation (Grant No. 2021M700430), and National Ten Thousand Talents Program of China. 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.
Appendix: Proofs of theorems
Appendix: Proofs of theorems
Proof of Theorem 1
From Construction 1, we have that A is an OA(\(\lambda s^2,s^{m_1})\) and B is an LH\((s, m_2)\) based on an SOA\((s,m_2,p^2,2+)\). Due to the expansive replacement method, any two columns from the same group for D must be two columns of a repeated B up to a row permutation, implying that D satisfies Definition 1(c) according to the definition of the SOA\((s,m_2,p^2,2+)\). Let \((d_1, d_2)\) be any two columns from different groups for D. According to the expansive replacement method, \((d_1, d_2)\) must be two columns of an OA(\(\lambda s^2,s^{m_1})\) up to a level permutation. Since any level permutation of an OA does not change its strength, \((d_1, d_2)\) is an OA with s levels. Therefore, D satisfies Definitions 1(a) and 1(b). This completes the proof. \(\square \)
Proof of Theorem 2
We only need to prove that the resulting design, say D, from Theorem 2 is column-orthogonal as its stratifications follow from Theorem 1. As B is column-orthogonal, then any two columns from the same group for D is column-orthogonal due to the expansive replacement method. According to the proof of Theorem 1, any two columns from different groups for D is an OA with s levels, and then is also column-orthogonal. This completes the proof. \(\square \)
In order to prove Theorem 3, we first give Lemmas 1, 2 and 3. From Step 1 of Construction 2, we have the following result.
Lemma 1
(Sun and Tang 2017a) Any four columns, derived by taking two columns from one group \(C_{i_1}\) and two columns from another group \(C_{i_2}\) with \(i_1 \ne i_2\), must form an OA\((\lambda s^2, 4, p,4)\).
According to the specific structures of E and F in Step 2 of Construction 2, we have the following result.
Lemma 2
Any three columns, derived by taking two columns from E and one column from F, must form an OA\((\lambda p s^2, 3, p,3)\).
Lemma 2 allows many three-column arrays whose columns come from E and F to have strength three, and then allows the constructed D in (6) to have the stratification property of Definition 2(c).
Note that any linear level permutation of C in Step 2 of Construction 2 does not change its strength such that \(u + C\) in F has the same strength as C for \(u=1, \dots , p-1\). From Lemma 1, a direct result is given below.
Lemma 3
Any four columns, derived by taking two columns from one group \(E_{i_1}\) and two columns from another group \(F_{i_2}\) with \(i_1 \ne i_2\), must form an OA\((\lambda p s^2,4, p,4)\).
Lemma 3 allows many four-column arrays whose columns come from E and F to have strength four, and then allows design D in (6) to have the stratification property of Definition 2(b).
Proof of Theorem 3
From (5), we have \(D^*= HR\), where \(H=(H_1, \dots , H_q)\) and \(R=\textrm{diag}\{U,\dots ,U\}\) with U repeating q times. By Lemma 2, we see that (E, F) is an orthogonal array of strength 2, implying that H is column-orthogonal. By noting that U is also column-orthogonal, we have \({D^*}^T {D^*}=(HR)^T HR = R^T (H^T H) R = c_1 R^T R=c_2I_{2q}\), where \(c_1\) and \(c_2\) are two constants, and \(I_{2q}\) is the identity matrix of order 2q. This shows that \(D^*\) in (5) is column-orthogonal, so is D in (6).
Before showing the stratifications of D, we first consider two mappings \(h_1(x)\) and \(h_2(x)\). Note that any column d of D has the following form:
where \((b, b',b'')\) up to a column permutation is \((e_{ik}, f_{ik'}, e_{jl})\), where \(k'=k+1\) if \(k\le m_2-1\) and \(k'=1\) if \(k = m_2\). Consider the mapping
which collapses the \(p^3\) levels in \(\Omega (p^3)\) into the \(p^2\) levels in \(\Omega (p^2)\). We need to show that \(h_1(d)=bp+b'\), which means that the column d becomes the column \(bp+b'\) after the mapping \(h_1(x)\) is applied to each entry of d. By letting \(r=b+(p-1)/2\), \(r'= b' + (p-1)/2\) and \(r''=\pm b'' + (p-1)/2\), we have
Since all entries of \((b, b', \pm b'')\) are in \(\Omega (p)\), all entries of \((r, r', r'')\) must take values from \(\{0, 1, \dots , p-1\}\). We then have \(h_1(d)= r p + r' - (p^2-1)/2= bp+b' \). We next consider the mapping
which collapses the \(p^3\) levels in \(\Omega (p^3)\) into the p levels in \(\Omega (p)\). By the above similar discussion, we have \(h_2(d)= b\).
We now prove the stratifications of D in Definition 2(b). Consider two columns \(d_1\) and \(d_2\) of D as in (8), and write them as
Let us show that the array \((d_1, d_2)\) can be collapsed into an OA\((\lambda p s^2, 2, p^2, 2)\), that is to say, \((h_1(d_1),\) \(h_1(d_2))=(b_1 p+b_1', b_2p+b_2')\) is an OA\((\lambda p s^2,2,p^2,2)\). In fact, this is true by noting the following two facts: (i) from Lemma 3, \((b_1, b_2, b_2, b_2')\) is an OA\((\lambda p s^2, 4, p, 4)\); (ii) \(p^3x_1+p^2x_2+px_3+x_4\) establishes a one-to-one correspondence between the \(p^4\) pairs \((x_1, x_2, x_3, x_4)\) with \(x_1, x_2,x_3,x_4 \in \Omega (p)\) and the \(p^4\) levels in \(\Omega (p^4)\). This shows that D satisfies Definition 2(b).
Next, we prove the stratifications of D in Definition 2(c). Let us show that the array \((d_1, d_2)\) can be collapsed into an OA\((\lambda p s^2,2,{p^2 \times p}, 2)\) and an OA\((\lambda p s^2,2,{p \times p^2}, 2)\), that is to say, \((h_1(d_1), h_2(d_2))=(b_1 p + b_1', b_2)\) and \((h_2(d_1), h_1(d_2))\) \(=(b_1, b_2p + b_2')\) are an OA\((\lambda p s^2,2,{p^2 \times p}, 2)\) and an OA\((\lambda p s^2,2,{p \times p^2},\) 2), respectively. In fact, this is true by noting the following two facts: (i) from Lemma 2, \((b_1, b_1', b_2)\) and \((b_1, b_2,b_2')\) are two OA\((\lambda p s^2, 3, p, 3)\)’s; (ii) \(px_1+x_2\) establishes a one-to-one correspondence between the \(p^2\) pairs \((x_1, x_2)\) with \(x_1, x_2 \in \Omega (p)\) and the \(p^2\) levels in \(\Omega (p^2)\). This shows that D satisfies Definition 2(c). This completes the proof. \(\square \)
Proof of Theorem 4
From the structures of \(C^{(1)}\) and \(C^{(2)}\), we have that the array \((c_{kj}^{(1)}, c_{ij}^{(1)}, c_{ij}^{(2)})\) is a repeated full factorial and is hence an OA\((\lambda s^3,3,s,3)\) for any \(k \ne i\). From the expansive replacement method, \(c_{kl}^{(1)}\) can be obtained by permuting levels in \(c_{kj}^{(1)}\) for \(l\ne j\). Because of this, the array \((c_{kl}^{(1)}, c_{ij}^{(1)}, c_{ij}^{(2)})\) with \(l \ne j\) is also an OA\((\lambda s^3,3,s,3)\). Then we have that the array \((c_{kl}^{(1)}, c_{ij}^{(1)}, c_{ij}^{(2)})\) has strength 3 for any \(k \ne i\). According to Theorem 4 of Zhou and Tang (2019), Definitions 3(a) and 3(b) can be verified for design D in (7). Since B is an LH\((s, m_2)\) based on an OA\((s, p^{m_2})\), any two columns from the same group for D can be collapsed into an OA\((\lambda s^3, p^2)\), implying that D satisfies Definition 3(c). This completes the proof. \(\square \)
Proof of Theorem 5
We only need to prove that the resulting design, say D, from Theorem 5 is column-orthogonal as its stratifications follow from Theorem 4. From Lemma 6.27 of Hedayat et al. (1999) and the definition of X, we see that \((GD_{\textrm{I}}(X)\), \(GD_{\textrm{II}}(X))\) in Step 1 of Construction 3 is an OA(\(\lambda s^3,s^{2m_1})\). As B is column-orthogonal, then \((C^{(1)}, C^{(2)})\) is column-orthogonal according to the proof of Theorem 2, and the inner product of any two columns from the centered \((C^{(1)}, C^{(2)})\) equals zero. From (7), let \((d_i, d_j)\) be any two columns from the centered D, then we have \(d_i=sc_i^{(1)} + c_i^{(2)}\) and \(d_j=sc_j^{(1)} + c_j^{(2)}\), where \(c_i^{(1)}, c_i^{(2)}, c_j^{(1)}\) and \(c_j^{(2)}\) are from the centered \((C^{(1)}, C^{(2)})\), implying that \(d_i^Td_j=0\) due to the column orthogonality of \((C^{(1)}, C^{(2)})\). This completes the proof. \(\square \)
Rights and permissions
Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.
About this article
Cite this article
Li, W., Liu, MQ. & Yang, JF. Several new classes of space-filling designs. Stat Papers 65, 357–379 (2024). https://doi.org/10.1007/s00362-023-01402-7
Received:
Revised:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00362-023-01402-7