Professional Documents
Culture Documents
To cite this article: Boxin Tang (1993) Orthogonal Array-Based Latin Hypercubes, Journal of the American Statistical
Association, 88:424, 1392-1397
Taylor & Francis makes every effort to ensure the accuracy of all the information (the “Content”) contained
in the publications on our platform. However, Taylor & Francis, our agents, and our licensors make no
representations or warranties whatsoever as to the accuracy, completeness, or suitability for any purpose of
the Content. Any opinions and views expressed in this publication are the opinions and views of the authors,
and are not the views of or endorsed by Taylor & Francis. The accuracy of the Content should not be relied
upon and should be independently verified with primary sources of information. Taylor and Francis shall not be
liable for any losses, actions, claims, proceedings, demands, costs, expenses, damages, and other liabilities
whatsoever or howsoever caused arising directly or indirectly in connection with, in relation to or arising out of
the use of the Content.
This article may be used for research, teaching, and private study purposes. Any substantial or systematic
reproduction, redistribution, reselling, loan, sub-licensing, systematic supply, or distribution in any
form to anyone is expressly forbidden. Terms & Conditions of access and use can be found at http://
www.tandfonline.com/page/terms-and-conditions
Orthogonal Array-Based Latin Hypercubes
Boxin TANG*
~ ~~~
In this article, we use orthogonal arrays (OA's) to construct Latin hypercubes. Besides preserving the univariate stratificationproperties
of Latin hypercubes, these strength r OA-based Latin hypercubes also stratify each r-dimensional margin. Therefore, such OA-based
Latin hypercubes provide more suitable designs for computer experimentsand numerical integration than do general Latin hypercubes.
We prove that when used for integration, the sampling scheme with OA-based Latin hypercubes offers a substantial improvement
over Latin hypercube sampling.
KEY WORDS: Computer experiments; Numerical integration; Orthogonal arrays.
definition of OA's, a Latin hypercube is an OA of strength a Latin hypercube randomly from the collection of Latin
one, OA(n, rn, n, 1). hypercubes { U } (each such Latin hypercube having an
Current users of Latin hypercubes randomly select per- equal chance to be drawn) such that (Z(u0)) forms an
mutation columns (Welch and Sacks 1990), usually with a OA( n, rn, s, r ) that is equivalent to the given A . (Two OA's
supplemental control of the correlations among the per- are said to be equivalent if one can be obtained by permuting
mutation columns, in the hope of getting better uniformity the rows, columns, and symbols of the other.) This view is
properties for multivariate margins. The Latin hypercubes particularly useful for the derivations of various probabilities
to be constructed here possess the r-variate uniformity prop- connected with U sampling. We postpone the study of the
erties inherent in a strength r OA used for construction. theoretical properties of U sampling to Section 3. The re-
Let A be an OA(n, m , s, r ) . For each column of A , we mainder of this section presents some examples of U designs.
replace the As'-' positions with entry k by a permutation of
Example 1. We use OA(4, 2, 2, 2) to construct a four-
( k - l)As'-'+ l , ( k - 1)Asr-'+2, . . . , ( k - l)As'-'+As'-'
point U design, U . For comparisons, we also give a four-
= kAs'-', for all k = 1, . . . , s. After the replacement pro-
point LHD, L, that is not a U design. The two designs are
cedure is done for all m columns of A , the newly obtained
matrix, denoted by U , is evidently a Latin hypercube. It 1 2 3 4 1 2 3 4 '
enjoys the merits of A in achieving uniformity in each r-
variate margin. Note that U becomes A under the element-
u=[2 ,]' L=[l 3]
Z ( i ) = ri/Asr-9, i = I , . . . , n, (1)
Example 2. We construct two U designs in this ex-
ample, one based on OA(9, 2, 3, 2) and the other on
where rxl is the smallest integer 2 x . OA(8, 2, 2, 2). We give their graphical representations in
The construction of OA-based Latin hypercubes depends Figure 2.
on the existence of the corresponding OA. For related results
on OA's, we refer to Plackett and Burman (1946), Rao Example 3. We consider a more realistic situation
( 1947), Bose and Bush ( 1952), Dey ( 1985), and de Launey in this example. A random U design is generated using
(1986). OA-based Latin hypercubes are referred to as U de- OA(49, 8, 7, 2), and the painvise plots of the columns of
signs hereafter when they are used for designing experiments. the design are given in Figure 3. To allow comparison, we
U designs should be more appropriate designs than general also provide in Figure 4 the plots of this sort for a random
L H D s for computer experiments. They may also be useful LHD. It is seen that the design points in the former look
for physical experiments, in which some factors have many more uniform than those in the latter.
levels to be accommodated so that the number of experi- 3. SOME RESULTS ON U SAMPLING
ments required by a suitable OA may be too large to be
practical. In this section, we study U sampling with associated OA's
The nonuniqueness of U designs based on a single OA being OA( s 2 , m , s, 2), as the use of such arrays leads to the
poses the problem of choosing a desirable U design. A paper most economical sample size, n = s 2 , for a given s. (For
in preparation by Tang and Wu discusses this problem, using a given s, the U sample is obtained by taking a fixed
both distance and correlation criteria. We do not give details OA(s2, rn, s, 2) and then applying the method of Section
here but merely point out this important issue. 2.) The basic ideas should cany through for general arrays,
We now consider how to use OA-based Latin hypercubes however, although derivations will become very complicated.
for numerical integration. For a given OA( n , rn, s, r ) = A, An exception is Theorem 2, in which general OA's are con-
we randomize its rows, columns, and symbols to obtain a sidered.
randomized OA. Then for each column of this randomized
array, we replace the As'-' positions with entry k by a random
permutation (with each such permutation having an equal
probability of being taken) of ( k - 1) As'-' + 1, ( k - 1) As'-'
+
I
+ 2, . . . , ( k - l)Xs'-' As'-' = kAs'-', for all k = 1, . . . , I
I
s. This procedure generates a random OA-based Latin hy- I
I
k -
percube, denoted by U = (u,,). Now suppose that X;'
-
I
Unif((i - l ) / n , i / n ) , i = 1 , . . . , n , j = 1, . . . , m , are
all generated independently. Then the n points to be used
for integration are formed by X, = (XL,,, X & , . . . , X L ) , I I I I I
use of U sampling, we can in fact omit the step of random- Figure 1. A Four-Point U Design and a Four-Point LHD. The points in
izing the rows of A . It is included simply to make the U (a) represent the U design; the points in (b) , the LHD. On intuitive grounds,
the former should be preferred to the latter. Indeed, suppose that the area
sample X I , . . . , X, exchangeable and hence to ease the study
within the dashed line box is the design region. The design points in (a)
of the theoretical properties of U sampling. are more uniformly scattered than those in (b) in the two-dimensional
The description of U sampling can be made more compact - in the sense that each small dashed line box in fa)
reoion . , contains exactlv
using the mapping Z introduced in ( 1). We may simply draw one design point.
1394 Journal of the American Statistical Association, December 1993
where PIX --
* X Pm and Ql X ---
X Qm are the small (1 992) independently obtained a result similar to Corollary
and large cells to which XI belongs and, for example, con- 2 by using randomized OA's, with the aid of the results
ditioning on c( PJ) means conditioning on XI € c( PI). of Patterson (1954). Owen's sampling procedure may be
The proof of Theorem 1 is given in Appendix A. If f i s briefly described as follows. Suppose that A = (a,) is an
continuous on [0, l]", consequently also bounded, then we O A ( n , m , s, r ) with its symbols randomized and that X :
can easily show the following result using Theorem 1. - Unif(0, l), i = 1, . . . , n , j = 1, . . . , m , are generated
independently. Then the n points Owen used for integration
Corollary 1. If f i s continuous on [0, l]", then under
are the rows of the matrix ( ( a , - X { )/s). To compare U
U sampling we have
sampling with Owen's procedure, let Vl and V2be the vari-
m
ances of using U sampling and Owen's procedure. Then
cov[f(xl~,f(x2)l= -n-l c var[J(XJ)I
J=I
we can show the following theorem.
m Theorem 2. When f is additive-that is, whenf( X ) = p
- n-I 2 var[f;,(x',XI)] + o ( n - l ) , + CJmlJ ( XI)-we have
l<J
V2 - v,
where J and AJ are main effects and interactions as defined m
in Section 1. = n-1 C E { v a r [ E ( ~ ( X J ) I X J ~ P C Q ) l ~ '>~O Q, l }
Under LHS, we know that cov[f(Xl), f(X2)l = -n-'
Downloaded by [Moskow State Univ Bibliote] at 07:40 14 January 2014
]=I
+
X CJml var[J(XJ)] o ( n - ' ) (Stein 1987). Therefore, Cor-
ollary 1 indicates a stronger negative correlation between
-
where X J Unif[O, 11 and P E F', Q E Q.
Thus Theorem 2 shows that U sampling gives a smaller
f(XI) and f(X2) under U sampling than under LHS in the
asymptotic sense.
r
variance of than does Owen's procedure whenever f i s
additive. Note that Theorem 2 is true for any OA. We prove
Corollary 2. If f i s continuous on [0, I]", then under Theorem 2 in Appendix B.
U sampling we have Before concluding this section, we note the close connec-
m tions between our results and the work in the numerical
var( F) = n-'a2 - n-I 2 var[J(XJ)] analysis literature on quasi-Monte Carlo. Hua and Wang
J=I (198 1) gave a systematic account of quasi-Monte Carlo
m methods. For historical developments, we refer to the review
- +
2 var[f;,(X',XJ)] o ( n - ' ) . papers by Niederreiter (1978, 1988). Our method has two
1x1 features not shared by quasi-Monte Carlo methods. First,
This follows from Corollary 1 and our Assertion. OA-based Latin hypercubes stress low-dimensional margins
If we write and specifically achieve uniformity in each t-dimensional
margin ( t I r ) . Second the uniformity properties of OA-
m m
f(') = p +
]=I
J ( X J )+
[<I
cf;,(X', based Latin hypercubes are in fact finite sample properties
+ I(')> in the sense that, for example, each c(PJ)contains exactly
one point. I am currently exploring further connections be-
then it is easy to check that the terms on the right side of tween OA-based Latin hypercubes and quasi-Monte Carlo
the equality are uncorrelated with each other. Therefore, we methods, both theoretically and numerically, and hope to
obtain the corresponding variance decomposition report them in a future communication.
m m
a2 = C var(J> + C var(JJ) + var[r(X)I, 4. GENERALIZATION
]=I [GI
and consequently the variance of under U sampling isr Generalizations of OA-based Latin hypercubes can be
+
n-' var [ r( X )] o( n-' ). This generalizes Stein's result. Owen made easily using asymmetrical OA's. Relevant references
on asymmetrical OA's include Rao ( 1973), Dey ( 1985), Wu
(1989), Wang and Wu (199l), and Wu, Zhang, and Wang
(1992). The notation OA( n , r ; sI,. . . ,sm)is used to denote
an asymmetrical OA of size n and strength r , with sJ levels
in thejth column, j = 1, . . . , m , where the sj's may not be
distinct. Let A be a given OA(n, r ; sl, . . . , sm). Then an
LHD, U = ( uU), is said to be an asymmetrical U design
based on A , if ( Z j (u;,)) = A , where Z,( i) is defined to be
risj/nl. The generalization of U sampling can be made ac-
cordingly. A simple example illustrates the underlying idea.
(4 (b) Example 4. An asymmetrical U design is constructed
Figure 5. A Six-Point U Design Together With a Six-Point LHD. The
using the OA(6, 2; 2, 3), and is represented graphically in
points in (a) represent the U design; the points in (b), the LHD. Note that Figure 5 . An LHD that is not an asymmetrical u design is
each small dashed line box in (a) contains exactly one design point. also given in Figure 5.
1396 Journal of the American Statistical Association, December 1993
APPENDIX A: PROOF OF THEOREM 1 Case 2. The row ( bl I , b12,. . . , b I m is ) in the first block, and
the row (bZ1,b22, . . . , 62,) is in the second block, as bll # b21.
Several lemmas are needed for proving Theorem 1 . Here again, the symbols bI2, . . . , b l , must all be in different rows
Letcl = P I X --- X P, be asmall cell in @ and L = Ql X --. in the second block. Because the row ( bZl, b22, . . . , 62,) already
X Q, be the large cell containing cI. Further, let G = ( UTl c( P j ) ) occupies one row in the second block, there are (s - l)(,-') ways
U ( U z j c(Qi X Q j ) ) .For U sampling using OA(s2, rn, s, 2 ) , we to arrange the symbols b I 2 ,. . . , b l , in the second block.
have P( C2 = c2 I CI = cI) = 0 for all c2 C G. We derive below an
expression for P(C2 = c2I CI = c1), for c2 C G', which plays a key It is well known that for the existence of an OA( sz, rn, s, 2 ) , it
role in the proof of Theorem 1 . The following lemma is obvious + +
is necessary that rn Is 1 . If rn = s 1 , then Lemma 2 is obviously
by the definition of U sampling. true, because it is impossible to arrange b12,. . . , bl, in the second
block for Case 2. Now suppose that rn Is.
Lemma 1. We have
Case 1. We now have b I 2., . . ,bl, placed in the first and second
P(C2 = C21CI = CI)
blocks, and a subrow ( b22,. . . , b2,) of different symbols in the
second row of the first block. Because rn Is, at least one of the
rows of the second block (without loss of generality, say, the first
row) has none of b I 2 ,bl3, . . . , blmin it. Let ( b 2 ,b3, . . . , 6,) be
an assignment of symbols to that row.
Case 2. Here we have b12,. . . ,bl, placed in the first and second
where a and b are constants. Note that G' = (nSl c(Qj)') blocks and a subrow ( b22,. . . , b2,) of different symbols in the first
Downloaded by [Moskow State Univ Bibliote] at 07:40 14 January 2014
U ('-'TI (dQj )\ G)). row of the second block. Let ( b2, . . . , b,) be an assignment of
In Lemmas 2 and 3 we derive the expressions for a and 6 . symbols to the second row of the first block.
Lemma 2. We have At this point, each Case 1 assignment compatible with the given
A could have been amved at as a Case 2 assignment, and vice versa.
a-s-rn+l
-_ To be specific, with any particular arrangement for the symbols
b S b12,. . . , b l mlet
, BI for Case I and B2for Case 2 be the collections
of OA's that can be obtained by filling the remaining positions with
Proof: Let cI = ( u l l , u12, . . . , u l m ) and c2 = ( u ~ 2422,
,
. . . , u2,).
We want to count the number of Latin hypercubes U appropriate symbols. For any B E B1,if we switch the two symbols
= ( uu), with the first and second rows fixed to be cI and c2, such
bj and bZj in column j , j = 2, . . . , rn and then supplement the
that (Z(u,)) is an OA(s2, rn, s, 2 ) that is equivalent to the given symbol-switchingprocedure with a suitably chosen permutation of
array, A, where Z is defined in (1). the rows, we will obtain an OA in B 2 .This defines a mapping from
B, to B2.It is easy to check that this mapping actually establishes
Step 1. We first obtain the elements uil,i = 3, . . . , n, in the first a one-to-one correspondence between BIand B2.Thus the numbers
column. The number of ways for doing this is ( n - 2)!. of OA's in BI and in 3 2 are equal. Note that for any BI E B I ,the
Step 2. Now consider an OA(s2, rn, s , 2 ) , B = (b,) that is equiv- array in B2(say, B2) given by the mapping is equivalent to B1. So
alent to A , such that the symbols in the first and second rows are either both BI and B2 are equivalent to the given A or neither is
given by bu = Z(u,), i = 1 , 2, j = 1, . . . , rn, and the symbols in equivalent to A. Therefore, BI and B2 contain the same number
the first column are given by b,1 = Z ( u i l ) , i = 3, . . . , n . of OKs that are equivalent to A.
In understanding Step 2, it is helpful to think of there being rns Step 3. Obtain uu, i 2 3, j 2 2 , suth that Z ( u,) = b,. It is easily
symbols T,, j = 1, . . . , rn and T = 1 , . . . , s; then bij = 7,if column seen that the number of ways of getting such uu's is the same for
j has level T in the ith row. Thus the symbols look like 2 , , 3 2 , 1 3 , both cases.
and so on.
In summary, we conclude that the number of Latin hyper-
For the rn pairs ( blj, b2J)of symbols, j = 1 , . . . , rn, two cases
cubes based on A , with the first two rows equal to cI and c2, is
arise:
Case I . When c2 C UjEl (c(Qj)\G), there must be one pair
- + -
ks. .(s - rn 2 ) for Case 1 and k(s - 1 ) - -(s - rn 1 ) for +
Case 2. Therefore,
having the same symbol in it. Without loss of generality, suppose
a k(s- l ) . . . ( s - m + 1 ) -
- s-rn+ 1
that this is the first pair, which implies bll = b21,and b , # b2j,
j = 2 , ..., m .
-=
b -
ks. .(s - rn + 2 ) S
i
m
s2 - ( m - I ) ~
SZ- 1
, for x~ f3 c ( ~ j ) = ,
j=I
g(xlC1 = CI) =
s2 + ( m - 1)s
-
m
sz- 1
, for x E u
j=I
(c(Qj)\G). where Xi. Unif((k - l ) / s , k / s ) . With this observation, the
proof of Theorem 2 is straightforward.
We are ready to prove Theorem 1.
Proof of Theorem 1. Without loss of generality, we assume p [Received July 1991. Revised January 1993.1
= 0 in the proof. We now have