Abstract
Computer experiments play a crucial role when physical experiments are expensive or difficult to be carried out. As a kind of designs for computer experiments, maximin distance designs have been widely studied. Many existing methods for obtaining maximin distance designs are based on stochastic algorithms, and these methods will be infeasible when the run size or number of factors is large. In this paper, we propose some deterministic construction methods for maximin \(L_2\)-distance designs in two to five dimensions based on densest packings. The resulting designs have large \(L_2\)-distances and are mirror-symmetric. Some of them have the same \(L_2\)-distances as the existing optimal maximin distance designs, and some of the others are completely new. Especially, the resulting 2-dimensional designs possess a good projection property.
Similar content being viewed by others
References
Ba S, Myers WR, Brenneman WA (2015) Optimal sliced Latin hypercube designs. Technometrics 57(4):479–487
Conway JH, Sloane NJA (1998) Sphere packings, lattices and groups. Springer, New York
Draguljić D, Santner TJ, Dean AM (2012) Noncollapsing space-filling designs for bounded nonrectangular regions. Technometrics 54(2):169–178
Fang KT, Li R, Sudjianto A (2005) Design and modeling for computer experiments. Chapman & Hall/CRC, Boca Raton
Grosso A, Jamali ARMJU, Locatelli M (2009) Finding maximin Latin hypercube designs by iterated local search heuristics. Eur J Oper Res 197(2):541–547
He X (2017) Rotated sphere packing designs. J Am Stat Assoc 112:1612–1622
Husslage BGM (2006) Maximin designs for computer experiments. Dissertation, Tilburg University
Husslage BGM, Rennen G, van Dam ER, Hertog DD (2011) Space-filling Latin hypercube designs for computer experiments. Optim Eng 12:611–630
Johnson ME, Moore LM, Ylvisaker D (1990) Minimax and maximin distance designs. J Stat Plann Inference 26:131–148
Joseph VR, Hung Y (2008) Orthogonal-maximin Latin hypercube designs. Stat Sin 18:171–186
Moon H, Dean A, Santner T (2011) Algorithms for generating maximin Latin hypercube and orthogonal designs. J Stat Theory Pract 5:81–98
Morris MD, Mitchell TJ (1995) Exploratory designs for computer experiments. J Stat Plann Inference 43:381–402
Oler N (1961) A finite packing problem. Canadian Math Bull 4:153–155
Pronzato L, Müller WG (2012) Design of computer experiments: space filling and beyond. Stat Comput 22(3):681–701
Ranjan P, Haynes R, Karsten R (2011) A computationally stable approach to Gaussian process interpolation of deterministic computer simulation data. Technometrics 53:366–378
Santner T, Williams B, Notz W (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
Tang Y, Xu H (2014) Permuting regular fractional factorial designs for screening quantitative factors. Biometrika 101:333–350
van Dam ER, Husslage B, Den Hertog D, Melissen H (2007) Maximin Latin hypercube designs in two dimensions. Oper Res 55:158–169
Wang L, Xiao Q, Xu H (2018) Optimal maximin \(L_1\)-distance Latin hypercube designs based on good lattice point designs. Ann Stat 46:3741–3766
Xiao Q, Xu H (2017) Construction of maximin distance Latin squares and related Latin hypercube designs. Biometrika 104:455–464
Xiao Q, Xu H (2018) Construction of maximin distance designs via level permutation and expansion. Stat Sin 28:1395–1414
Ye KQ, Li W, Sudjianto A (2000) Algorithmic construction of optimal symmetric Latin hypercube designs. J Stat Plann Inference 90:145–159
Zhou YD, Xu H (2015) Space-filling properties of good lattice point sets. Biometrika 102:959–966
Acknowledgements
The authors thank the associate editor and two anonymous referees for their valuable comments and suggestions. This work was supported by the National Natural Science Foundation of China (Grant Nos. 11771220 and 11871288), National Ten Thousand Talents Program, Tianjin Development Program for Innovation and Entrepreneurship, Tianjin “131” Talents Program, and Natural Science Foundation of Tianjin (19JCZDJC31100). The first two authors contributed equally to this work.
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.
Electronic supplementary material
Below is the link to the electronic supplementary material.
Supplementary material
The supplementary material includes the figures for densest packings A2 and D3 and more comparisons among the resulting designs and other types of designs with two to five dimensions.
Appendix
Appendix
The points of \(P_{0i},i=1,\ldots ,4\), in Algorithm 1.
where
Proof of Proposition 1
According to the definition of C, we have
Thus, the points in C can be divided into two categories, which are points on the lines \(x=i\) for \(i\in {\mathbb {Z}}\) and points on the lines \(x=i+1/2\) for \(i\in {\mathbb {Z}}\). Combining with the relationship between two coordinates, it is easy to obtain the expressions of \(C_1\) and \(C_2\) in Proposition 1. Moreover, for any point in C, \({\mathbf{p}}_0=(x_0,y_0)\), according to the structure of C, it is obvious that the neighboring points of \({\mathbf{p}}_0\) are
Then, the \(L_2\)-distances between \({\mathbf{p}}_0\) and these points are always one. The proof is completed. \(\square \)
Proof of Corollary 1
We have \(X=C\cap [-K,K]^2=(C_1\cup C_2)\cap [-K,K]^2=X_1\cup X_2\), where
Thus, the number of points in X is equal to the sum of the numbers of points in \(X_1\) and \(X_2\), i.e.
According to the coordinates of points in X, it is obvious that both \(X_1\) and \(X_2\) are symmetric against two axes and the origin. Thus, if there is a point \(\mathbf{a }=(x_1,x_2)\in X_1\) then the points \(\mathbf{b }=(x_1,-x_2)\), \(\mathbf{c }=(-x_1,x_2)\) and \(\mathbf{d }=(-x_1,-x_2)\) all belong to \(X_1\), which implies that \(X_1\) is column-orthogonal and mirror-symmetric. Similarly, \(X_2\) is column-orthogonal and mirror-symmetric. Then, we have that \(X=X_1\cup X_2\) is column-orthogonal and mirror-symmetric. \(\square \)
Proof of Proposition 2
Note that
where \(G_0^*=G_0\cdot R\). In other words, obtaining \(C^*\) by rotating C is equivalent to generating \(C^*\) by the rotated generator matrix \(G_0^*\). Thus, we only need to consider the property about the rotation of generator matrix \(G_0\).
(1) If \(\theta =\pi /6\), we have
Then,
where \(z_1^*=z_2\) and \(z_2^*=z_1\). It implies that \(C^*\) is equivalent to exchanging the two columns of C in the sense of reordering the rows.
(2) If \(\theta =\pi /3\), we have
Then,
where \(z_1^*=-z_2\) and \(z_2^*=z_1+z_2\). It implies that \(C^*\) is equivalent to C in the sense of reordering the rows. In these two cases, \(C^*\) and C are isomorphic. The proof is completed. \(\square \)
Proof of Theorem 1
The mirror-symmetrical property of P can be easily obtained from the property of \(P_0\). Thus, we only prove the distance property of P here. Combining with Proposition 1 and the relationship between P and \(P_0\), we have \(d(P)=d(P_0)/(2l)=1/(2l)\) and \({\mathrm{mpd}}(P)={\mathrm{mpd}}(P_0)/(2l)\). So, we only need to obtain \({\mathrm{mpd}}(P_0)\). Since there is a slight difference in the points for the cases of \(K\in {\mathbb {N}}\) and \(K+1/2 \in {\mathbb {N}}\), the proofs need to be considered separately. Here we only discuss the first case, and the other case follows similarly.
To prove the first case, we define a way of rotation, called non-crossed rotation. Given an \(n\times k\) matrix \(X=(x_{il})\), let \(X^*=(x_{il}^*)=XR\) with rotation angle \(\theta \). If for any i, j, l satisfying \(x_{il}<x_{jl}\), there is always \(x_{il}^*<x_{jl}^*\), then we call this rotation is non-crossed. The rotation used in Algorithm 1 is a non-crossed rotation. According to Corollary 1 and the definition of non-crossed rotation, there are two kinds of projection distances on the first dimension and two kinds on the second dimension between any two neighboring points in \(X^*\), which are
respectively. To show these projection distances more clearly, see Fig. 2. Points with the same shape have the same horizontal coordinates before rotating. In this example, there are two kinds of projection distances on the first dimension, which are
And there are two kinds of projection distances on the second dimension, which are
Thus, to achieve equal spacing on the first dimension, let \(d_{1,1}=d_{1,2}\). Therefore, we have the value of \(\theta \) shown in Algorithm 1, i.e, \(\theta =\arctan \left[ \sqrt{3}\left( 3+2\left\lfloor \frac{K}{\sqrt{3}} -\frac{1}{2}\right\rfloor +2\left\lfloor \frac{K}{\sqrt{3}} \right\rfloor \right) \right] ^{-1} \). Based on values of \(n,\theta \) and \(l=K\cos \theta -\sqrt{3}t\sin \theta \), we have \(l=(\sqrt{3}/2)(n-1)\sin \theta \). At this time, it is easy to see that \({\mathrm{mpd}}(P_0)=\min \{d_{1,1},d_{2,1},d_{2,2}\}=\sin \theta \). Thus, we have
\(\square \)
Proof of Proposition 3
The coordinates of points in \(P_0\) can be obtained easily. Now, let us calculate the number of points in \(P_0\). According to the coordinates of points, there are only two cases of points: (1) \(x_1,x_2,x_3\) are all even, (2) only one of \(x_1,x_2,x_3\) is even and the others are odd. Thus, we have
\(\square \)
Proof of Corollary 2
It is easy to obtain the coordinates and the number of points of \(P_0^c\), so we omit it here. Conway and Sloane (1998) showed that the packing radius, the radius of sphere, of lattice \(D_3\) is \(r=1/\sqrt{2}\), and the kissing number is \(\tau = 12\). To get the \(L_2\)-distance of \(P_0^c\), we only need to illustrate that the structure of \(D_3^c\) is the same as that of \(D_3\). Based on the coordinates of \(D_3\), it can be found that the points in \(D_3\) can be divided into two parts \(D_{31}\) and \(D_{32}\), whose coordinates are
Obviously, \(D_3^c\) has a similar interleaved structure as follows:
where \(D_3^c=D_{31}^c\cup D_{32}^c\). Thus, the points in \(D_3^c\) can be regarded as being obtained from \(D_3\) with shifting, which means that the radius of sphere and kissing number of \(D_3^c\) are equal to those of \(D_3\). So, the \(L_2\)-distance of \(P_0^c\) is \(d=2r=\sqrt{2}\), which is the same as that of \(P_0\). \(\square \)
Proof of Corollary 3
According to Corollary 2, we have that the \(L_2\)-distance is \(d_0=\sqrt{2}\) for both \(P_0\) and \(P_0^c\). Thus, it is easy to obtain that \(d(P)=d_0/s=\sqrt{2}/s\) since \(P=P_0/s\). Moreover, since the orthogonality property of P is the same as that of \(P_0\), we only need to prove that \(P_0\) is a column-orthogonal and mirror-symmetric design. Let \(P_0^*=P_0+ (-s/2)\). According to the coordinates of points in \(P_0^*\), it is easy to find that if there is a point \(\mathbf{a }=(x_1,x_2,x_3)\in P_0^*\) then the points \(\mathbf{b }=(-x_1,x_2,x_3)\), \(\mathbf{c }=(x_1,-x_2,x_3)\), \(\mathbf{d }=(x_1,x_2,-x_3)\), \(\mathbf{e }=(-x_1,-x_2,x_3)\), \(\mathbf{f }=(-x_1,x_2,-x_3)\), \(\mathbf{g }=(x_1,-x_2,-x_3)\) and \(\mathbf{h }=(-x_1,-x_2,-x_3)\) all belong to \(P_0^*\). With a simple calculation, it can be seen that the inner product of any two distinct columns in \(P_0^*\) is zero, which means that \(P_0\) is column-orthogonal. Moreover, \(P_0^*\) can be written as \(P_0^*=(\mathbf{X }^T,-\mathbf{X }^T)^T\), which indicates that \(P_0\) is a mirror-symmetric design. \(\square \)
Rights and permissions
About this article
Cite this article
Yang, L., Zhou, Y. & Liu, MQ. Maximin distance designs based on densest packings. Metrika 84, 615–634 (2021). https://doi.org/10.1007/s00184-020-00788-w
Received:
Published:
Issue Date:
DOI: https://doi.org/10.1007/s00184-020-00788-w