Abstract
Maximin distance designs and orthogonal designs have become increasingly popular in computer and physical experiments. The construction of such designs is challenging, especially under the maximin distance criterion. This paper studies a class of Latin hypercube designs by calculating the minimum distances between design points. We derive a general formula for the minimum intersite distance of this kind of design. The row pairs with the minimum intersite distance are also specified. The results show that such kind of Latin hypercube design is asymptotically optimal under both the maximin distance criterion and the orthogonality criterion.
Similar content being viewed by others
References
Ba S, Myers WR, Brenneman WA (2015) Optimal sliced Latin hypercube designs. Technometrics 57(4):479–487
Fang K-T, Li R, Sudjianto A (2006) Design and modeling for computer experiments. CRC Press, Boca Raton
Grosso A, Jamali A, Locatelli M (2009) Finding maximin Latin hypercube designs by iterated local search heuristics. Eur J Oper Res 197(2):541–547
He X (2019) Interleaved lattice-based maximin distance designs. Biometrika 106(2):453–464
Jin R, Chen W, Sudjianto A (2005) An efficient algorithm for constructing optimal design of computer experiments. J Stat Plan Inference 134(1):268–287
Johnson ME, Moore LM, Ylvisaker D (1990) Minimax and maximin distance designs. J Stat Plan Inference 26(2):131–148
Joseph VR, Hung Y (2008) Orthogonal-maximin Latin hypercube designs. Stat Sin 18:171–186
Joseph VR, Gul E, Ba S (2015) Maximum projection designs for computer experiments. Biometrika 102(2):371–380
Liefvendahl M, Stocki R (2006) A study on algorithms for optimization of Latin hypercubes. J Stat Plann Inference 136(9):3231–3247
Lin CD, Kang L (2016) A general construction for space-filling Latin hypercubes. Stat Sin 26:675–690
Lin CD, Tang B (2015) Latin hypercubes and space-filling designs. In: Dean A, Morris J, Stufken J, Bingham D (eds) Handbook of design and analysis of experiments. Chapman and Hall/CRC, New York, pp 593–625
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(2):239–245
Morris MD, Mitchell TJ (1995) Exploratory designs for computational experiments. J Stat Plan Inference 43(3):381–402
Santner TJ, Williams BJ, Notz WI (2003) The design and analysis of computer experiments. Springer, NewYork
Sun F, Liu M-Q, Lin DK (2009) Construction of orthogonal Latin hypercube designs. Biometrika 96(4):971–974
Sun F, Tang B (2017) A method of constructing space-filling orthogonal designs. J Am Stat Ass 112(518):683–689
Tang B (1994) A theorem for selecting oa-based Latin hypercubes using a distance criterion. Commun Stat Theory Methods 23(7):2047–2058
Van Dam ER, Husslage B, Den Hertog D, Melissen H (2007) Maximin Latin hypercube designs in two dimensions. Oper Res 55(1):158–169
Van Dam ER, Rennen G, Husslage B (2009) Bounds for maximin Latin hypercube designs. Oper Res 57(3):595–608
Viana FA, Venter G, Balabanov V (2010) An algorithm for fast optimal Latin hypercube design of experiments. Int J Numer Meth Eng 82(2):135–156
Wang L, Yang J-F, Lin DK, Liu M-Q (2015) Nearly orthogonal Latin hypercube designs for many design columns. Stat Sin 25(4):1599–1612
Wang Y, Yang J, Xu H (2018) On the connection between maximin distance designs and orthogonal designs. Biometrika 105(2):471–477
Xiao Q, Xu H (2018) Construction of maximin distance designs via level permutation and expansion. Stat Sin 28(3):1395–1414
Ye QK, Li W, Sudjianto A (2000) Algorithmic construction of optimal symmetric Latin hypercube designs. J Stat Plan Inference 90(1):145–159
Zhou Y, Xu H (2015) Space-filling properties of good lattice point sets. Biometrika 102(4):959–966
Zhu H, Liu L, Long T, Peng L (2012) A novel algorithm of maximin Latin hypercube design using successive local enumeration. Eng Optim 44(5):551–564
Acknowledgements
We are grateful to the editor, the associate editor and two anonymous referees for their insightful comments and suggestions, which have led to significant improvements. This work was supported by the National Natural Science Foundation of China (Grant Nos. 11871033, 11871052 and 11971204), the Natural Science Foundation of Tianjin (20JCYBJC01050) and the Fundamental Research Funds for the Central Universities (63211090). 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
Appendix
A.1. Proof of Lemma 4
First, based on Property 2, we can easily obtain the \(\phi \) values for the following three row pairs
and
Next, we confirm that the value \(\phi (m_1,m_{2^c+1})\) is the maximum \(\phi \) value and the values \(\phi (m_1, m_{2^c+2^{c-1}+1})\), \(\phi (m_{2^{c-1}+1}, m_{2^c+1})\) are the second-largest \(\phi \) value. According to Properties 1 and 2 , it is easy to see that the maximum \(\phi \) value is the \(\phi \) value of the first row and the (\(2^c+1\))th row. We will focus on the second-largest \(\phi \) value.
For any \(i,j \ne 1,2^c+1\), since the row vector \(m_i\) has a half of elements positive and others negative, so half of the \(2^{c+1}\) elements of the sum formula \(\phi (m_i, m_j)\) are negative. Then, if possible, the maximum value of \(\phi (m_i,m_j)\) is
By Property 1, we can know the sum formula \(\phi (m_1,m_i)\) has \(2^c+2^{c-1}\) positive elements. For the values \(\phi (m_{2^c+1},m_i)\), \(\phi (m_i,m_1)\), and \(\phi (m_i,m_{2^c+1})\), we have a similar conclusion. By the (i) and (iii) of Property 2, we have
where the equality achieves at \(\phi (m_1, m_{2^c+2^{c-1}+1})\) and \(\phi (m_{2^{c-1}+1}, m_{2^c+1})\). Also, we note that
Summarizing the above results, we can see that the second-largest \(\phi \) value can only be achieved at \(\phi (m_1, m_{2^c+2^{c-1}+1})\) and \(\phi (m_{2^{c-1}+1}, m_{2^c+1})\). This completes the proof. \(\square \)
A.2. Proof of Theorem 1
For \(c=2\), it is easy to show that the minimum \(L_2\)-distance between any two distinct rows in \(H_2\) is \(d_2(h_i,h_j)=2\), where \(\left| i-j \right| =2^{c}\); and the second-smallest \(L_2\)-distance is \(d_2(h_1,h_8) = d_2(h_2,h_7) = \frac{1}{3} \times 2^6 - \frac{3}{4} \times 2^4 + \frac{1}{6} \times 2^2 = 10\). In the following, we only consider the case of \(c > 2\).
For convenience, we always assume \(i < j\) and denote \({\mathscr {A}}=\{ 1,2,\dots , 2^{c} \}\) and \({\mathscr {B}}=\{ 2^{c}+1,2^{c}+2,\dots , 2^{c+1} \}\). We calculate the \(L_2\)-distances between any two distinct rows in \(H_c\) by dividing i and j into the following four cases.
-
(i)
If \(i,j \in {\mathscr {A}} \,\mathrm{or}\, i,j \in {\mathscr {B}}, \,\mathrm{and}\, \vert i-j \vert =2^{c-1} \), by Lemma 3, we have
$$\begin{aligned} d_2(h_i,h_j) = \sum _{r=1}^{2^{c-1}} \vert h_{ir}-h_{jr} \vert ^2 = 4 \times \frac{2^{3(c-1)+2}-2^{c-1}}{3} = \frac{2}{3}\times 2^{3c} - \frac{2}{3}\times 2^{c}. \end{aligned}$$ -
(ii)
If \(i,j \in {\mathscr {A}} \,\mathrm{or}\, i,j \in {\mathscr {B}}, \,\mathrm{and}\, \vert i-j \vert \ne 2^{c-1} \), by Lemma 3, we have
$$\begin{aligned} d_2(h_i,h_j) =4 \times \frac{2^{3(c-1)+2}-2^{c-1}}{6} =\frac{1}{3}\times 2^{3c} - \frac{1}{3}\times 2^{c}. \end{aligned}$$ -
(iii)
If \(i \in {\mathscr {A}}, j \in {\mathscr {B}} \,\mathrm{and}\, \vert i-j \vert =2^{c} \), we have
$$\begin{aligned} d_2(h_i,h_j) = \sum _{r=1}^{2^{c-1}} \vert h_{ir}-h_{jr} \vert ^2 = \sum _{r=1}^{2^{c-1}} \left| (2m_{ir}-\frac{1}{2}) - (2m_{ir}+\frac{1}{2}) \right| ^2 = 2^{c-1}. \end{aligned}$$ -
(iv)
If \(i \in {\mathscr {A}}, j \in {\mathscr {B}} \,\mathrm{and}\, \vert i-j \vert \ne 2^{c} \), we have
$$\begin{aligned} \begin{aligned} d_2(h_i,h_j)&= \sum _{r=1}^{2^{c-1}} \vert h_{ir}-h_{jr} \vert ^2\\&= \sum _{r=1}^{2^{c-1}} \left| (2m_{ir}-\frac{1}{2}) - (2m_{j-2^{c},r}+\frac{1}{2}) \right| ^2\\&= 4d_2(m_i,m_{j-2^{c}}) - 4\phi (m_i,m_{j-2^{c}}) + 2^{c-1}. \end{aligned} \end{aligned}$$If \(\left| i-(j-2^{c}) \right| = 2^{c-1}\), then we have
$$\begin{aligned} d_2(h_i,h_j)&=\; 4 \times \frac{2^{3(c-1)+2}-2^{c-1}}{3} -4\phi (m_i,m_{j-2^{c}}) + 2^{c-1} \nonumber \\&\geqslant \, 4 \times \frac{2^{3(c-1)+2}-2^{c-1}}{3} -4\phi (m_1,m_{1+2^{c-1}}) + 2^{c-1} \nonumber \\&=\; 4 \times \frac{2^{3(c-1)+2}-2^{c-1}}{3} -4\times 2^{2(c-1)} + 2^{c-1} \nonumber \\&=\; \frac{2}{3}\times 2^{3c} - 2^{2c} -\frac{1}{6} \times 2^c. \end{aligned}$$(6)The equality (6) holds if and only if \(i = 1\) and \(j = 1+2^{c-1}+2^c\).
If \(\left| i-(j-2^{c}) \right| \ne 2^{c-1}\), then we have
$$\begin{aligned} d_2(h_i,h_j)&=\; 4 \times \frac{2^{3(c-1)+2}-2^{c-1}}{6} -4\phi (m_i,m_{j-2^{c}}) + 2^{c-1} \nonumber \\&\geqslant \, 4 \times \frac{2^{3(c-1)+2}-2^{c-1}}{6} -4\times \frac{3}{4}\times 2^{2(c-1)} + 2^{c-1} \nonumber \\&=\; \frac{1}{3}\times 2^{3c} - \frac{3}{4} \times 2^{2c} +\frac{1}{6} \times 2^c. \end{aligned}$$(7)The equality (7) holds if and only if one of the following two conditions holds: (a) \(i = 1\) and \(j = 2^c+2^{c-1}+2^{c-2}+1\); and (b) \(i = 2^{c-2}+1\) and \(j = 2^c+2^{c-1}+1 \).
By summarizing the above results, when \(c \geqslant 2\), we can see that the minimum \(L_2\)-distance between any two distinct rows in \(H_c\) is \(d_2(h_i,h_j) = 2^{c-1}\), where \(\left| i-j \right| =2^{c} \); and the second-smallest \(L_2\)-distance is \(d_2(h_1,h_{2^{c}+2^{c-1}+2^{c-2}+1}) = d_2(h_{2^{c-2}+1},h_{2^{c}+2^{c-1}+1}) = \frac{1}{3}\times 2^{3c} - \frac{3}{4} \times 2^{2c} +\frac{1}{6} \times 2^c\). All other \(L_2\)-distances between row pairs are larger than \(\frac{1}{3}\times 2^{3c} - \frac{3}{4} \times 2^{2c} +\frac{1}{6} \times 2^c\). This completes the proof. \(\square \)
A.3. Proof of Theorem 2
By Lemma 3, we know that, for any two rows of \(M_c\), their \(L_2\)-distance is
By Theorem 1, we know that the minimum \(L_2\)-distance between two distinct rows in \(H_{c}\) is \(2^{c-1}\). This value is the \(L_2\)-distance between \(h_i\) and \(h_j\) with \(\vert i-j \vert = 2^c\). So when \(i,j \in \{ 1,2,\dots , 2^{c+1} \} \,\mathrm{and}\, \vert i-j \vert = 2^c\), the \(L_2\)-distance between two rows \(p_i\) and \(p_j\) is
By Theorem 1, we know that the second-smallest \(L_2\)-distance between two distinct rows of \(H_{c}\) is \(d_2(h_1,h_{2^{c}+2^{c-1}+2^{c-2}+1}) = d_2(h_{2^{c-2}+1},h_{2^{c}+2^{c-1}+1}) = \frac{1}{3}\times 2^{3c} - \frac{3}{4} \times 2^{2c} +\frac{1}{6} \times 2^c.\) So when \(i,j \in \{ 1,2,\dots , 2^{c+1} \} \,\mathrm{and}\, \vert i-j \vert \ne 2^c\), the minimum \(L_2\)-distance between two rows \(p_i\) and \(p_j\) is
By summarizing above results, we know that the \(L_2\)-distance of the design \(P_c\) is \(2^{3c} - \frac{3}{4} \times 2^{2c}\), which is either \(d_2(p_1, p_{2^{c}+2^{c-1}+2^{c-2}+1})\) or \(d_2(p_{2^{c-2}+1}, p_{2^{c}+2^{c-1}+1})\). This completes the proof. \(\square \)
A.4. Proof of Theorem 3
According to Theorem 2, the \(L_2\)-distance efficiency of the design \(P_c\) is
which clearly tends to one as c increases infinitely. \(\square \)
Rights and permissions
About this article
Cite this article
Pang, T., Wang, Y. & Yang, JF. Asymptotically optimal maximin distance Latin hypercube designs. Metrika 85, 405–418 (2022). https://doi.org/10.1007/s00184-021-00833-2
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00184-021-00833-2