HELSINKI UNIVERSITY OF TECHNOLOGY Department of Engineering Physics and Mathematics

HELSINKI UNIVERSITY OF TECHNOLOGY
Department of Engineering Physics and Mathematics
Systems Analysis Laboratory
Mat-2.4108 Independent Research Projects in Applied Mathematics
Sample Size Requierement for Monte Carlo - simulations
using Latin Hypercube Sampling
Anna Matala
60968U
20.5.2008
1
Contents
1 Introduction
3
2 Monte Carlo Integration
3
3 Latin Hypercube Sampling (LHS)
5
3.1
Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
5
3.2
Variance analysis . . . . . . . . . . . . . . . . . . . . . . . . . . .
6
3.3
Estimate of variance . . . . . . . . . . . . . . . . . . . . . . . . .
8
4 Sample size
8
5 Examples
12
5.1
Example of sampling with lognormal distribution . . . . . . . . .
12
5.2
Example of multinormal distribution . . . . . . . . . . . . . . . .
15
5.3
Example of binary variables . . . . . . . . . . . . . . . . . . . . .
19
5.4
Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
23
6 Conclusions
23
2
1
Introduction
The goal of probabilistic fire risk analysis is to discover the probabilities of the
harmful consequences of fire. The variating initial and boundary conditions,
that lead to growth of fire, are taken into account statistically. The consequences are represented as a limit state function. The probability is calculated
by integrating the limit state function over multidimensional joint distribution.
In practice, the integration is made by using Monte Carlo method, because the
physical models of fire are non-linear and often really complex. In Probabilistic Fire Simulator (PFS, developed by VTT) [1] the sampling can be done by
Simple Random sampling (also known as Monte Carlo sampling) or Latin Hypercube sampling.
The purpose of this work is to study the accuracy of Latin Hypercube sampling
and to find a simple manner to evaluate the sample size. Estimates and experiences are searched from literature. Finally, the results are tested in simple
simulations.
2
Monte Carlo Integration
Random sampling has been used for solving numerical problems as early as 1777
when Comte de Buffon made his experiment by dropping needle to the board
with parallel lines with same distance [2]. If the length of the needle is L and
the distance between lines d (d>L), the probability that the needle intersects a
line is
p=
2L
.
πd
(1)
Later Laplace realized that this method could be used for estimating the value
of π.
Modern Monte Carlo (SRS) method was born in 1940s when Stanislaw Ulam,
John von Neumann and others started to use random numbers to examine
physics from the stochastic perspective. One of the most effective use of Monte
Carlo method is to evaluate definite integrals which analytically would be too
difficult to find [3]. If the problem is
3
Z
b
f (x)dx,
y=
(2)
a
the solution can found by ’hit-or-miss’ (acceptance-rejection) method: A box is
drawn around the function from a to b and from 0 to y0 (y0 >f(x)). After this,
N random numbers are taken from a uniform distribution and plotted (as in
Figure 1). A number of points that fall below f(x) are then counted and marked
with N0 . Now the estimate of integral can be calculated using equation
yest =
N0
(y0 (b − a)).
N
(3)
The estimate is getting more and more precise when N is increased and will
converge to the correct answer when N → ∞. In this example, x had only one
dimension for simplicity, but the method can be successfully applied to multidimensional problems as well.
Figure 1: ’Hit or miss’ method in integration of curve f(x).
4
3
3.1
Latin Hypercube Sampling (LHS)
Method
The Latin Hypercube Sampling [4]-[6] is a type of stratified Monte Carlo sampling. All the areas of the sample space are represented by input values. To
generate a sample size N from K variables x=[x1 , x2 , ..., xK ]T with probability
density function f(x), the procedure is as follows. The range of each variable is
portioned into N non overlapping intervals of equal probability 1/N (Figure 2).
From each interval one value is selected randomly according to the probability
density of the interval. The N values of x1 are paired in a random manner with
values of x2 (Figure 3), these pairs are then paired similarly with values of x3
and so on, until N samples of n variables are formed. So created cube has N K
cells which cover the sample space. Denote the input vector X=[x1 ,...,xN ].
Figure 2: Density function of variable xi divided to N=5 non overlapping intervals.
5
Figure 3: Pairing x1 with x2 when N=5.
3.2
Variance analysis
Let h(x) be the transformation function of input x [4],[7]. The estimate of
expectation value E(h(x)) is
h = N −1
N
X
h(xj ).
(4)
j=1
The variance of the estimate depends on the sample method used. For Simple
Random sampling the variance is
var(h) = N −1 var(h(x))
(5)
and for the Latin Hypercube sampling
var(h) = N −1 var(h(x)) +
N −1
cov(h(x1 ), h(x2 )).
N
(6)
Thus the LHS sampling decreases the variance if and only if cov(h(x1 ),h(x2 ))<0.
In their article McKay et al. [4] proved that the covariance is negative whenever
h(X) is monotonic in each of its components. However, in many problems the
monotonicity does not hold. Stein [7] showed 1987 that when N → ∞, the
covariance term is asymptotically non positive. Define function
Z
gk (xk ) =
h(x)
K
Y
i=1,i6=k
6
dFi (xi )
(7)
and {XjN }(where j=1,...,N and N=1,2,...). Stein’s Theorem 1 shows that if
E(h2 )< ∞ and N → ∞,
K
X
K
cov(h(X)1N , h(X)2N ) = E(h)2 − N −1
N
Z
gk (x)2 dFk (x) + o(N −1 ). (8)
k=1
By Jensens’s inequality,
Z
Z
gk (x)2 dFk (x) ≥ ( gk (x)dFk (x))2 = (Eh)2 ,
(9)
so the highest-order term in expansion of cov(h(x1 ),h(x2 )) is non positive. That
means, that
limN →∞ N cov(h(x1 ), h(x2 )) ≤ 0.
(10)
This can be written easier if we define
ha (x) =
K
X
gk (xk ) − (K − 1)E(h)
(11)
k=1
and
r(x) = h(x) − ha (x).
(12)
The function ha (x) is the best additive fit to h(x) which means
Z
Z
2
r (x)dF (x) ≤
(h(x) −
K
X
hk (xk ))2 dF (x).
(13)
k=1
Now an approximation of variance can be written as N → ∞,
var(N −1
N
X
h(XjN )) = N −1
Z
r(x)2 dF (x) + o(N −1 ).
(14)
j=1
This is relevant whenever N is much larger than K, so the sample size is greater
than the number of variables.
7
3.3
Estimate of variance
A simple way to get an estimate to the variance of average is called replicated
Latin hypercube sampling [7]. The idea is to produce several independent LHS
samples and then compare the variances between samples. If the number of
samples is α and size of them M, N=α·M. If
hi = M −1
M
X
h(Xij ),
(15)
j=1
then an unbiased estimator for the variance is
var(α
−1
α
X
Pα
α
X
hi
1
2
[
hi ) =
hi − ( i=1 )2 ].
α(α − 1) i=1
α
i=1
(16)
This is exactly same as the sample variance of mean for normal sampling. Now
if α is too small, the estimate of variance is not precise. On the other hand,
increasing α while N is fixed increases this estimator of variance. However, as
long as ratio
4
M
K
is large, the increase in variance is small.
Sample size
In literature almost no recommendations or estimates for the sample size of LHS
sampling were proposed. However, it is possible the study the examples and apply them to fire-Monte Carlo simulation. In fire simulations, the most limiting
factor is time. Often the samples are simulated by Fire Dynamics Simulator
(FDS) and that makes the Monte Carlo simulation quite slow. As an example
of order of magnitude, already 1000 samples can require computation time of
several months. Typically number of variables is less than 10. All the following
examples are about the LHS sampling unless else is mentioned.
Stein [7] gives one example where ratio
M
K =10.
In his example N=1000, M=200,
K=20 and α=5. For fire simulation, this could mean 5 independent samples of
size 100 makes total sample size of 500. In his other example there are 6 variables and the estimator of variance is calculated by 100 independent replications
of samples of size 100. Now
M
K
would be nearly 17. This is unnecessary high
value and clearly impossible for fire simulation. However, according to the first
8
example it is possible to get reliable results by using values α ≥ 5 and
M
K
> 10.
Jianfang Jia et al. [8] give an example of NF-κB signaling pathway model. They
did global sensitivity analysis using 64 independent variables and 1000 samples,
so the ratio of samples and variables is 15.6. Applied to 10 variables, this example would suggest sample size 156.
As mentioned before, the latin hypercube sampling converges at least as fast
as simple random sampling if N is big enough. Swidzinski et al. [9] go further
stating that LHS could be nearly five times more effective in yield estimation
than traditional sampling methods. They justify it by their study at the Texas
A&M. In the tests, they generated 10 iterations of samples of size 200 using
both LHS and simple random sampling. After that, sample size of 1000 was
generated 10 times using simple random sampling. The results showed that
variation in the yield estimate during the 10 iterations were similar when LHS
sample size was 200 and simple random sample was 1000. The theory of simple
random sampling is better known, so this gives suggestive size of LHS samples.
Eduard Hofer [10] studied computationally intensive models, where sample sizes
were 100 as maximum and the samples were generated by Simple Random sampling. As the distributions of parameters are usually not known, the fractiles of
5% and 95% can not be immidiately obtained, and hence Monte Carlo simulations are used. The resulting distributions are often long-tailed so the fractile
estimates are highly dependent of the sample size N. Therefore, the possible
impact of the sampling error needs to be quantified. That can be done by calculating statistical tolerance limits (u(%), v(%)) where u is the percentage of
the combined influence of the quantified uncertainties and v is the confidence
level of it. In other words, with confidence level v(%) we can say that u(%) of
the samples are inside the tolerance limits. The tolerance interval is chosen to
be from the smallest to the largest observation in the sample.
Hofer gives an example: Two sided statistical tolerance limits (95%, 95%) for
Simple Random sample require only 93 model runs. This is a result of nonparametric statistics and depends only on the percentages u and v, not on the
number of uncertainties taken into account. Besides, it does not assume any
particular type of underlying distribution. This result is general and really
promissing from the point of view of fire simulation, especially because LHS
9
Table 1: Sample sizes for two-sided nonparametric tolerance limits. 1-α is the
confidence coefficient and q is the fraction of the population that lies between
the smallest and the largest observation of the sample.
q
0.5 0.7 0.750 0.8 0.85 0.9 0.95
0.975
0.98
0.99
1-α
0.5
3
6
7
9
11
17
34
67
84
168
0.7
5
8
10
12
16
24
49
97
122
244
0.75
5
9
10
13
18
27
53
107
134
269
0.8
5
9
11
14
19
29
59
119
149
299
0.85
6
10
13
16
22
33
67
134
168
337
0.9
7
12
15
18
25
38
77
155
194
388
0.95
8
14
18
22
30
46
93
188
236
473
0.975
9
17
20
26
35
54
110
221
277
555
0.98
9
17
21
27
37
56
115
231
290
581
0.99
11
20
24
31
42
64
130
263
330
662
0.995
12
22
27
34
47
72
146
294
369
740
0.999
14
27
33
42
58
89
181
366
458
920
sampling can be expected to show less variability than simple random sampling.
The solution for this example can be found from Table 1 originally extracted
from Conover’s book ([11]: Appendix A6). The table shows the sample sizes
required according some confidence coefficients and population fractions. If the
Table 1 does not offer direct answer, it is possible to make an approximation
N≈
1
1+q 1
χ(1−α)
+ (r + m − 1),
4
1−q 2
(17)
where r is the number of observation of lower bound of tolerance limit and
(N + 1 − m) the number of observation of upper bound. The sample must be
assorted to ascending order. r=m=1 is commonly used, so the lower bound is the
smallest observation and the upper bound the largest. χ1−α is the 1−α quantile
of a chi-square random variable with 2(r+m) degrees of freedom. Revising the
example before causes the same result:
N≈
1
1 + 0.95 1
· 9.488 ·
+ (1 + 1 − 1) = 93.008.
4
1 − 0.95 2
However, the statistical tolerance limits cannot be computed from LHS sample.
The reason is that the sample estimates of the 95% fractile can be significantly
10
below the true 95% fractile for long-tailed distributions. Without statistical
tolerance limits it is not possible to know how likely this shortcoming is and
therefore the analysis has to be made using simple random sampling. Of course,
after finding sample size N, it can be considered as upper limit of real LHS sample size needed.
11
5
5.1
Examples
Example of sampling with lognormal distribution
The following example is about samples of lognormal distribution. Lognormal
distribution is chosen because it is important in fire simulations.
Lognormal variables can be written y=ex where x is normally distributed with
mean α and variance β. Density function of lognormal distribution [12] is
f (y) =
with mean µ = eα+β
2
/2
1
√
βy 2π
e
−
(ln(y)−α)2
2β 2
2
(18)
2
and variance σ 2 = e2α+β (eβ − 1). The mean and
variance of ln(y) (α, β 2 respectively) can be calculated backwards using following
equations:
1
σ2
α = ln(µ) − ln( 2 + 1)
2
µ
(19)
and
β 2 = ln(
σ2
+ 1).
µ2
(20)
Monte Carlo simulations were made with sample sizes N=50, 100, 500, 1000,
5000 and 10000, using both Simple Random sampling (SRS) and Latin Hypercube sampling (LHS). The idea was to compare mean and variance of samples.
The lognormal distribution used had mean 27.4 and variance 16, which result
(equations (19) and (20)) α=3.3000 and β 2 =0.0211. The sample means and
variances are calculated using equations [13]
x
¯=
N
1 X
xi
N i=1
(21)
and
N
s2 =
1 X
(xi − x
¯ )2 .
N − 1 i=1
This is valid for both SRS and LHS sampling [4].
12
(22)
Figure 4: Sample means as function of sample size N.
In Figures 4 and 5 are shown sample means and variances with SRS and LHS
sampling. In simple case of only one variable, the LHS sampling is better with
every N. The differences between the correct and estimated mean and variance
values are listed in table 2.
N
Table 2: Difference of mean and variance.
∆µ (SRS)
∆σ 2 (SRS)
∆µ (LHS)
∆σ 2 (LHS)
50
0.169231071
3.549920858
0.039925728
0.068954216
100
0.333796062
2.874371726
0.007375298
0.155918985
500
0.235102161
0.102067908
0.000254559
0.032888354
1000
0.071753115
0.424214694
0.000419834
0.003795222
5000
0.012967127
0.153119925
0.000180192
0.004282639
10000
0.023052468
0.054337911
0.0000689647
0.00197263
13
Figure 5: Sample variances as function of sample size N.
14
5.2
Example of multinormal distribution
Consider next a situation where variable X depends on five independent, normal
distributed variables (with parameters µi , σi ). The x is defined as
x=
5
X
xi ,
(23)
i=1
where xi ∼N(µi ,σi ). The parameters used in this example are listed in table 3
and the simulations were made with N=50, 100, 500, 1000 and 5000. Sum of norP
mal distributed variables is also normal distrbuted with parameters µ =
µi
pP
2
σi . In this example, the mean µ is 16.5 and standard deviation
and σ =
σ=5.7879 (variance 33.4998). In Figures 6 and 7 are shown the sample mean
and variance with different sample sizes. It seems that LHC is closer to real
parameter values than SRS with each N.
Table 3: Parameters of variables.
i µi
σi
1
1
0.5
2
2.5
2
3
-4
5
4
10
2
5
7
0.5
Study now the distributions of samples using the Pearson’s chi-square test. That
compares sample distribution F(x) to some known distribution G(x). The null
hypothesis is H0 : F(x)=G(x). To calculate the p-value, the sample has to be
divided in m categories. The χ2 is
χ2 =
m
X
(Oi − Ei )2
Ei
i=1
,
(24)
where Oi is frequency in category i and Ei =NP, the expected frequency of category i (calculated from known distribution). The degree of freedom is df=m-l -1,
where l is the number of parameters in known distribution. The prerequisites
are N>50, Ei >2 for all i and only 20 % of categories have Ei <5. If p-value is
less than chosen risk level, H0 will be refuted. The risk level is normally 5 %.
15
Figure 6: Mean of sample as function of sample size N.
Null hypothesis is now
H0 : F(x)=N(16.5,5.7879).
Divide distribution into 8 classes that are listed in table 4. The probabilities
come from the standard normal distribution table, where z =
x−µ
σ .
It can be
seen that now both prerequisites are filled, because when N≥ 50, sample size is
expected to be >5 in every cathegory.
The χ2 -values with different sample sizes are presented in table 5. Now df =
8 − 2 − 1 = 5. In χ2 -table the maximum value to be accepted with risk level 5
% is 11.07. In this example, all the χ2 -values are in the acceptance region, so
all the samples normally distributed with parameters µ = 16.5 and σ = 5.7879.
In this case, with five independent variables, sample size 50 was enough in both
SRS and LHS sampling.
16
Figure 7: Variance of sample as function of sample size N.
Cathegory
x
P
I
[-∞,9.09)
0.1003
II
[9.09,12)
0.1174
III
[12,14)
0.1159
IV
[14,16)
0.1305
V
[16,18)
0.1385
VI
[18,20)
0.1231
VII
[20,23)
0.1429
VIII
[23, ∞)
0.1314
Table 4: Categories of normal distribution.
17
N
Table 5: χ2 values.
χ2 (SRS)
χ2 (LHS)
50
10.37789289
9.148392622
100
10.21332038
2.156814332
500
3.388478928
4.64044467
1000
6.533513491
5.451614406
5000
3.706511017
6.611654417
10000
2.131056413
0.857279318
18
5.3
Example of binary variables
Consider a nuclear power plant where in case of fire the failure time t1 of some
important component is lognormally distributed with parameters µ1 and σ1 .
The failure can be prevented by the fire service if they arrive before the failure
happens. The arriving time of the fire service (t2 ) is lognormally distributed
with parameters µ2 and σ2 . Now the interesting factor would be to know what
is the probability that the failure can be prevented, or that t2 < t1 . This is
actually the probability p(A) of Bernoulli distribution, where X=1 if A happens
and X=0 if not. In lognormal distribution negative values are not allowed, so
the probability of P(t2 <t1 ) has to be calculated from joint distribution. The
joint lognormal distribution is
f (t1 , t2 ) = f (t1 )f (t2 ) =
1
√
β1 t1 2π
e
−
(ln(t1 )−α1 )2
2
2β1
·
1
√
β2 t2 2π
−
e
(ln(t2 )−α2 )2
2
2β2
,
(25)
where α and β are calculated from equations (19) and (20). Now we are interested in the probability of t2 < t1 , and that can be found by integrating
Z
∞
Z
t1
F (t2 < t1 ) =
f (t1 )f (t2 )dt2 dt1 .
0
(26)
0
This can be integrated numerically e.g. by matlab using function called ’trapz’.
In this example µ1 =12, σ1 =3, µ2 =13 and σ2 =5, so α1 =2.455, β2 =0.246, α2 =2.496
and β2 =0.371. The distributions are plotted in Figure 8. Expected probability
is calculated by integrating the joint distribution, and has value of 0.463.
Confidence intervals for Bernoulli distribution are calculated as
r
pˆ ± zα/2
pˆ(1 − pˆ)
,
n
(27)
where 1-α is the confidence level (often 95 %), zα/2 ∼N(0,1), pˆ is the estimated
probability and n is the sample size. This is valid only for simple random sampling. In m independent repetitions, the real value of p should be inside the
confidence intervals m·p times in an average.
Repeat the simulations 14 times (averages as a function of sample size N are
shown in Figure 9), and calculate the confidence intervals as in equation (27).
19
Figure 8: Lognormal distributions f1 (α1 ,β1 ) and f2 (α1 ,β2 ).
The idea is now to compare the confidence intervals of Monte Carlo samples
to similar calculated intervals of Latin Hypercube samples. These intervals are
not valid confidence intervals of LHS sample, but comparing these intervals can
give more information. Choose 95 % confidence level, so Z0.025 = 1.96. The
confidence intervals were calculated to each simulation, and then counted the
times the real value of p was included in the confidence interval in 14 repetitions.
The results are listed in table 6. It would be interesting to test null-hypothesis
H0 : P(’The real value of p is inside the 95 % confidence interval in LHS sampling’)=0.95, but that is not possible for two reasons. First, to make that test,
m should be big. The rule is that
mˆ
p ≥10
m(1-ˆ
p)≥10.
The estimates are near 0.95 so the minimum m should be 200 which is too much
considering that with PFS (Probabilistic Fire Simulator) the maximum variable
number is 15 and for each pˆ two variables are needed. This would lead to nearly
30 simulations for each N. However, with smaller number of repetitions it is
possible to get an idea of the result of the test. Another problem is, that the
test variable is
20
N
n (SRS)
n (LHS)
25
13
14
50
12
14
100
12
14
500
14
14
1000
12
14
5000
13
14
10000
13
14
Table 6: n is a number of times that p is inside the confidence intervals in 14
repetitions.
pˆ − p
z=p
,
pˆ(1 − pˆ)/m
(28)
which goes to infinity if pˆ=1. In this case the null-hypothesis would be always
rejected.
After 14 repetitions the estimate of probability is pˆ=1 with each N in LHS
sampling. The test variable is ∞ and the null-hypothesis rejected. For LHS
sampling, the 95 % confidence intervals are significantly smaller than for SRS
sampling, so with every N the LHS sampling gives at least as good results as
the SRS.
21
Figure 9: Averages of 14 independent repetitions.
22
5.4
Discussion
To determine the sample size is not a simple task. Many guides have focused
on the Simple Random sampling, but only one common rule was found from
the literature for Latin Hypercube samples: It gives at least as good results as
Simple Random sampling when the sample size is big enough. The sample size
required may depend on the distribution, number of variables or something else.
The only way to see if the sample has converged, is to make more simulations
and see if the result stays similar. Even this does not work for latin hypercube
sampling, where the sample size should be known beforehand.
In all three examples the latin hypercube samples were closer to the real value
for each N. For five independent and normal distributed variables sample size
50 was enough to say that the sample distribution did not differ from the original distribution significantly. In the last example it was seen that the samples
generated with LHS had smaller confidence intervals, so the results of LHS sampling are more accurate.
6
Conclusions
The target of this work was to study the accuracy of Latin Hypercube sampling
and find some advices how to choose the sample size N. In literature, very few
directions were given, and in the tests it became clear that general rules are
difficult to find. However, both the literature and examples refer to the fact
that Latin Hypercube samples converge with smaller sample size than simple
random samples. The sample sizes needed in simple random sampling can be
used as upper limits for the Latin Hypercube sample size.
References
[1] Hostikka, S., Keski-Rahkonen, O. Probabilistic simulation of
fire scenarios. Nuclear Engineering and Design 224 (2003)
301-311.
[2] Veach, E. Robust Monte Carlo methods for light transport
simulation (1997).
23
[3] Landau, D., Binder, K. A Guide to Monte Carlo Simulations
in Statistical Physics (2000).
[4] McKay M., Beckman, R., Conover, W. A Comparison of
Three Methods for Selecting Values of Input Variables in the
Analysis of Output from a Computer Code. Technometrics,
vol 21, no. 2, May 1979.
[5] Keramat, M., Kielbasa, R. Modified Latin Hypercube Sampling Monte Carlo (MLHSMC) Estimation for Average Quality Index. Analog Integrated Circuits and Signal Processing,
vol. 19, no. 1, April 1999.
[6] Wyss, G., Jorgensen, K. A User’s Guide to LHS: Sandia’s
Latin Hypercube Sampling Software. (1998)
[7] Stein, M. Large Sample Properties of Simulations Using Latin
Hypercube Sampling. Technometrics, vol 29, no.2, May 1987.
[8] Jia, J., Yue, H., Liu, T., Wang, H. Global Sensitivity Analysis of Cell Signalling Transduction Networks Based on Latin
Hypercube Sampling Method. Bioinformatics and Biomedical
Engineering, 2007.
[9] Swidzinski J., Chang, K. A novel nonlinear statistical modeling technique for microwave devices. Microwave Symposium
Digest., 2000 IEEE MTT-S International (0-7803-5687-X)
June 2000. Vol.2;p.887
[10] Hofer, E. Sensitivity analysis in the context of uncertainty
analysis for computationally intensive morels. Computer
Physics Communications 117 (1999) 21-34.
[11] Conover, W. Practical nonparametric statistics, 2nd ed. (Wiley, New York, 1980).
[12] R˚
ade, L., Westergren, B. Mathematics handbook for science
and engineering. (4th edition, 1998).
[13] Laininen, P. Sovelletun todenn¨ak¨oisyyslaskun kaavoja ja
taulukoita.
24
`