Professional Documents
Culture Documents
Uncertainty Propagation On A Nonlinear Measurement Model Based On Taylor Expansion
Uncertainty Propagation On A Nonlinear Measurement Model Based On Taylor Expansion
Uncertainty Propagation On A Nonlinear Measurement Model Based On Taylor Expansion
Abstract
In this paper, the propagation of uncertainty on a nonlinear measurement model is presented using a higher-order Taylor
series. As the derived formula is based on a Taylor series, it is necessary to compute the partial derivatives of the non-
linear measurement model and the correlation among the various products of the input variables. To simplify the
approximation of this formula, most previous studies assumed that the input variables follow independent Gaussian dis-
tributions. However, in this study, we generate multivariate random variables based on copulas and obtain the covar-
iances among the products of various input variables. By applying the derived formula to various cases regardless of the
error distribution, we obtained the results that coincide with those of a Monte-Carlo simulation. To apply high-order
Taylor expansion, the nonlinear measurement model should be continuous within the range of the input variables to
allow for differentiation, and be an analytic function in order to be represented by a power series. This approach may
replace some time-consuming Monte-Carlo simulations by choosing the appropriate order of the Taylor series, and can
be used to check the linearity of the uncertainty.
Keywords
uncertainty, nonlinear, Monte-Carlo, Taylor
Creative Commons CC BY: This article is distributed under the terms of the Creative Commons Attribution 4.0 License
(https://creativecommons.org/licenses/by/4.0/) which permits any use, reproduction and distribution of the work without
further permission provided the original work is attributed as specified on the SAGE and Open Access pages (https://us.sagepub.com/en-us/nam/
open-access-at-sage).
210 Measurement and Control 54(3-4)
more accurately than using the technique recom- where xk is the input variable, zk is the error in xk, and
mended in GUM. Zhang extended the Taylor series to y is the output of the function. f is a real-valued func-
an infinite number of terms and proposed some meth- tion that is differentiable on x1, ., xm; for simplicity,
ods to calculate the variance of the system.11 Mekid f(x1, ., xm) will be represented as f. Thus, the covar-
and Vaja12 calculated the uncertainty of the measure- iance using the first-order Taylor expansion is given by:
ment model up to the third order using the skewness !
and kurtosis of the input variables. Xue et al.13 derived X m
∂fi X m
∂fj
the variances and covariances of the model based on cov(yi , yj )’cov fi + zk , f j + zp
k=1
∂xk p=1
∂xp
the infinite expansion of the Taylor series. These !
approaches assume that the input variables are inde- Xm
∂fi Xm
∂fj
= cov zk , zp
pendent (or dependent) normal distributions, thus ∂xk ∂xp
k=1 p=1
removing terms that are difficult to calculate analyti- * + * +
cally. Thus, they cannot be applied when the input vari- Xm
∂fi Xm
∂fj Xm
∂fi
= zk zp zk
ables do not obey a normal distribution and/or are k=1
∂xk p = 1 ∂xp k=1
∂xk
correlated with each other. * +
Xm
∂fj
In this paper, we re-derive the Taylor series for the zp
covariance of a nonlinear measurement model regard- p=1
∂xp
less of distribution and correlation of the variables. We Xm
∂fi X m
∂fj
also show the method how to apply the derived for- = zk zp h zk i zp
mula to simple numerical simulation. In the numerical k=1
∂x k p=1 ∂x p
∂p fi (x1 , , xn ) ∂q fj (x1 , , xn )
Numerical partial differentiation
cov(zi1 zip , zj1 zjq ) The n-point stencil is widely used to obtain numerical
∂xi1 ∂xip ∂xj1 ∂xjq
derivatives. For example, the first partial derivative
ð10Þ
using a 5-point stencil is as follows:
As a result, the covariance of the nonlinear measure- 0 1
ment model can be represented as the covariance of the f (x)’ ðfðx 2hÞ 8fðx hÞ + 8fðx + hÞ fðx + 2hÞÞ
12h
various products of input variables with their partial ð11Þ
derivatives. Note that the analytic solution for the cov-
ariance of the product of variables has not known when The n-point stencil can also be used to compute deriva-
they are not normal distribution. This makes equation tives on the multivariable. However, we use a multi-
(10) difficult to apply to the real problem. Equation dimensional m-point stencil to simplify the numerical
(10), however, can be applied using a numeric calcula- implementation, as shown in Figure 1. For O variables,
tion if we can compute the numerical partial derivatives we require nO points. For the first-order partial
and the numerical covariance for the product of input derivatives, equation (11) is applied along each axis. As
variables. We will show this in following section. the order of the differentiation increases, such as for
212 Measurement and Control 54(3-4)
∂2f/∂2x, equation (11) is modified appropriately; the Table 1. Distribution of test variable.
formula can be found in Fornberg.14 Note that this
approach does not require additional induction to Distribution Denote Mean Variance
derive the appropriate formulas. Normal N(m, s2) m s2
Exponential Exp(l) l21 l22
Uniform U(a, b) 1 1 2
2 (a + b) 12 (b a)
Covariance of various products of input variances Gamma G(k, u) kupffiffiffi ku2
Rayleigh R(s) s p2 4p 2
2 s
Equation (10) requires the covariances of various prod-
ucts of input variables, such as cov(zi1 zip , zj1 zjq ).
Although there have been some studies on the covar- where rs and r are Spearman’s and Pearson’s correla-
iance between products of random variables,15,16 we tion coefficients, respectively. In this paper, to focus on
found that these approximations are not sufficiently the numerical approach, equation (12) is applied repeat-
accurate for this study. Instead, we generated multi- edly until the generated random variables meet the tar-
variate random variables from the covariances of input get covariances. The detailed procedure is descried in
variables using the relevant copula and found the cov- the appendix.
ariances of the products of various combinations.
In applying the copula, multivariate normal random
Results
numbers are generated using Cholesky decomposition
and converted to uniform random vectors.17 These vec- We tested the covariances of the nonlinear measure-
tors are then transformed back to random numbers ment model derived from the previous section using
with the target marginal distributions using the inverse four nonlinear functions and five random variable dis-
cumulative distribution function in MATLAB.18 tributions. The nonlinear functions are y1 =
Depending on the marginal distribution, there may be x13 + x22 + 2, y2 = x12 + x22, y3 = x12 x2, and
a difference in the correlations between the target and y4 = x1x2. The distributions considered here are the
the generated variables because of the nonlinearity of normal, exponential, uniform, gamma, and Rayleigh
the transform. To overcome this issue, we apply distributions; details are summarized in Table 1. We
Spearman’s ranked correlation. In a normal distribu- calculated the covariance of y1 and y2 and the covar-
tion, the Pearson correlation coefficient is related to iance of y3 and y4 for the various distributions of input
Spearman’s rho as:19 variables using a MC simulation and first-, second-,
and third-order Taylor approximations. The MC simu-
p
lations consisted of 1,000,000 runs; all results are sum-
r = 2 sin rs ð12Þ
6 marized in Table 2.
Table 2. Comparison of covariance using the MC simulation and the Taylor expansion.
Type Distribution of Covariance Monte-Carlo First order Second order Third order
x1 and x2 of input
S(y1 , y2 ) N(0, 2), N(0, 3) 2:0031 1:2050 137:82 20:84 4:45 4:33
310 5 18:09 20:98 137:82 20:84
1:2050 3:0045
20:84 31:90
4:33 4:22
20:98 31:89
20:84 31:90
pffiffiffi pffiffiffi
Exp(1/ 2), Exp(1/ 3) 2:0010 1:2166 6308:6 1146:5 158:9 107:3 1416:1 645:7 6308:6 1146:5
1:2166 2:9960
1146:5 350:7
107:3 75:9
645:7 350:7
1146:5 350:7
pffiffiffi pffiffiffi
U(– 6, 6), U(–3, 3) 2:0013 1:2032 38:17 8:24 2:56 3:02 7:21 8:22 38:17 8:24
3105
pffiffiffiffiffiffiffiffi 1:2032 2:9982
8:24 12:43
3:02 3:83
8:22 12:43
8:24 12:43
G(5, 2=3), G(3, 1) 2:0009 1:2086 8031:0 1675:0 2353:1 753:6 5430:2 1427:2 8031:0 1675:0
1:2086 3:0074 1675:0 486:9 753:6 280:7 1427:2 486:9
1675:0 486:9
qffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffi
R 4
, R 6 2:0016 1:2260 4339:9 1135:1 1455:1 592:4 2964:6 968:9 4339:9 1135:1
(4p) (4p)
1:2260 3:0021
1135:1 402:8
592:4 278:5
968:9 402:8
1135:1 402:8
pffiffiffi
S(y3 , y4 ) N(0, 2), Exp(1/ 3) 1:9979 1:2009 140:44 42:00 0:0 0:01 23:89 8:50 140:44 42:00
1:2009 3:0004
42:00 17:40
0:01 5:99
8:50 17:40
42:00 17:40
pffiffiffi
Exp(1/ 2), U(–3, 3) 2:0037 1:1713 477:39 71:90 12:01 8:50 74:07 31:26 477:39 71:90
1:1713 3:0000
71:90 14:35
8:50 6:00
31:26 14:35
71:90 14:35
pffiffiffi pffiffiffi
U(– 6, 6), G(3, 1) 2:0003 1:2096 66:46 27:19 0:00 0:04 28:78 6:86 66:46 27:19
1:2096 3:0030 27:19 25:72 0:04 18:01 6:86 25:72
27:19 25:72
pffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffi
G(5, 2=3), R 6 2:0021 1:2202 4640:6 650:2 1686:6 354:4 3381:7 561:8 4640:6 650:2
(4p)
1:2202 2:9946
650:2 102:1
354:4 77:4
561:8 102:1
650:2 102:1
qffiffiffiffiffiffiffiffiffi
R( (4p)4
), N(0, 3) 2:0019 1:2265 730:04 140:57 160:36 59:31 389:57 105:27 730:04 140:57
1:2265 3:0025 140:57 30:79 59:31 21:94 105:27 30:79 140:57 30:79
Gu et al. 213
20. Marsden JE and Tromba A. Vector calculus. 5th ed. New ing mean of 0 are generated. The generated z is then con-
York, NY: W. H. Freeman, 2003. verted back into a uniform random vector u. In order to
21. Jargon JA, Wang C-M and Hale PD. A robust algorithm convert u back to a random number R with the desired
for eye-diagram analysis. J Lightwave Technol 2008; target marginal distribution, we use MATLAB’s inverse
26(21): 3592–3600. cumulative distribution function (CDF) function.
22. Devroye L. General principles in random variate genera- According to the inversion method, applying the inverse
tion. In: Devroye L (ed.) Non-uniform random variate gen- CDF of the distribution F to the uniform random vari-
eration. New York, NY: Springer, 1986, pp.27–82. able U(0,1) produces a random number with exactly F
distribution, which maintains the correlation coefficient
Appendix of z.22 If the covariance of the random number R is simi-
Figure 5 shows the procedure and MATLAB code for lar to that of TargetCov, these numbers are used both
generating multivariate random number used in both MC simulation and calculation of the covariance between
MC simulation and calculation of the covariance between product of input variables. If the difference of two covar-
products of input variables. First, the correlation coeffi- iance is large, apply equation (6) and repeat the above
cient CorCoeff is calculated by dividing the target covar- procedure. Thus, there are slightly differences between
iance TargetCov by each standard deviation. Then, based ‘‘covariance of input’’ and ‘‘Target covariance’’ as shown
on the calculated CorCoeff, multivariate numbers z hav- in Table 2.
Figure 5. Procedure and MATLAB code for generating multivariable random numbers.