Chapter 3 Comparative Studies of Different UP Methods

There are different kinds of UP methods but Polynomial chaos expansion is a really import
part of uncertainty propagation(UP). This method quantifies uncertainty in system output
performance supported random or noisy inputs, is of great importance for design under uncertainty
particularly for issues with computationally valuable simulation analysis models (such as finite
component analysis and process fluid dynamics). A good kind of probabilistic UP approaches are
developed [2],among that the polynomial chaos growth (PCE) technique could be a rigorous
approach thanks to its strong mathematical basis and skill to provide useful representations of
random quantities. With PCE, the operate with random inputs will be depicted as a stochastic meta
model, supported that lower order applied math moments additionally as dependableness of the
operate output will be derived with efficiency to facilitate the implementation of style optimization
underneath uncertainty eventualities like strong style [46] and reliability-based style. As a well-
known PCE approach, the generalized PCE (𝑔PCE) methodology supported the Askey theme has
been wide applied to UP for its higher accuracy and higher convergence compared to the classic
Wiener PCE with Hermite polynomials, like the propagation of uncertainty in composite
structures, the modelling of input uncertainties and its propagation within the incompressible fluid
flow [47] and therefore the analysis of the dynamic behavior of unsure gear friction system.
Generally, the random input doesn't essentially follow the 5 kinds of probabilistic distributions
(i.e., normal, uniform, exponential, beta and gamma) within the Askey theme. In this case, the
transformation ought to be created to transfer every random input variable to at least one of the 5
distributions, which would degrade the accuracy of PCE significantly. Therefore, the Gram-
Schmidt PCE (GS-PCE) [48] and multi-element PCE (ME-PCE) [11] strategies are developed to
accommodate arbitrary distributions through constructing their own orthogonal polynomials.
instead of bearing on the existing ones within the Askey theme, the orthogonal polynomials in
ME-PCE and GS-PCE are made by the Stieltjes three-term repeat severally. All the PCE strategies
mentioned on top of is made based on the idea that the precise information of the concerned
probability density operates (PDF) of every random input variable exists. However, the PDF of a
random parameter could exist as some information or numerically as an advanced bar graph, like
bi-modal or multi-modal kind, that is commonly troublesome to get the analytical expression of

1
its PDF accurately. On the opposite hand, the knowledge regarding the distribution of information
is often terribly restricted in realistic engineering applications, which cannot be adequate to
produce a full image of their PDFs.

3.1 Introduction:
One of the most important components of uncertainty analysis is the quantification of uncertainties
in the system output performances propagated from uncertain inputs. It is named as uncertainty
propagation (UP). A lot of experiments have been done to develop methods of uncertainty
propagation in various fields such as structural reliability, stochastic mechanics, quality
engineering and many others. As the methods were created with different backgrounds on visions,
it is necessary to test the efficiency and applicability in engineering design. In design under
uncertainty, the role that UP is expected to play varies based on different design scenarios. For
example, in robust optimization design, the interest of UP is to evaluate the low-order moments of
a performance. In reliability-based design, the interest in utility optimization, where the probability
distribution needs to be integrated with the designer’s utility function to maximizing the expected
utility of a product. Four examples are tested to compare the performance of the methods
mentioned, under different levels of function nonlinearly and different types of uncertainty inputs.
The first two examples are to compare the FFNI, UDR and PCE performances. Three and five-
node (𝑚 = 3 𝑜𝑟 5) quadrature rules are used for one-dimensional integration in both FFNI and
UDR. In PCE, the expansion is made up to the 4th order (𝑝 = 4) are tried with coefficients.
calculated with 5𝑛 FFNI. Here, the PCE method is used, the moments are calculated analytically
and the probability of failure is estimated by MCS based on the PCE functions.

3.2 Example 1:
𝑦 = 𝑥𝑘 𝑘 = 1,2, … ,7 (3.1)
This example is used to compare the accuracy of various methods in moment estimation against
the nonlinearity of performance function in a one-dimensional case. I’ll be testing two cases, one
with a normal distribution’s input and another with a lognormal distribution’s input. The problem
is one-dimensional, UDR method.

3.2.1 Case1:
Here 𝑥~𝑁𝑜𝑟𝑚𝑎𝑙(𝜇 = 1, 𝜎 = 0.2). In the following Figs. 3.1 and 3.2, the skewness and kurtosis
calculation Results are depicted with the ration to the exact values obtained analytically. NI in the

2
legend means method using numerical integration. Results of MCS with 1,000𝑘 samples are
provided for comparison. Since the input x follows the normal distribution the Gauss-Hermite
quadrature rule with m-node is used for numerical integration and 2𝑚 − 1 integration order is
expected. It means that when m is 5 the skewness results are exact up to nonlinear order 𝑘 = 3 and
kurtosis results are exact up to 𝑘 = 2. The 5 nodes numerical integration seems sufficient for
calculating the coefficients of PCE up to 4-th order in this case. PCE with coefficients obtained by
numerical integration is compared with PCE with calculated analytically in the Appendix. As the
nonlinearity of 𝑔(𝑥) increases, the coefficients become erroneous for the higher order terms in the
PCE. The reason why higher order expansion is still more accurate under the bigger coefficient
estimation error is that the truncation error is more significant than the numerical error in the
coefficient estimation. It is notable that for any order of expansion, 𝐸[𝑦 (𝑝) ] equals 𝐸[𝑦]. Due to
the orthogonality, the coefficients of low order terms stay the same even if higher order terms are
added. From figure one and two it is noted that for this example, for different orders of nonlinearity
𝑘, the accuracy follows the sequence of MCS (Highest), 4th order PCE, 5 node NI, 3rd PCE, 2nd
order PCE shows slightly better results in skewness and kurtosis calculation than the 5-node
numerical integration while both methods use the same amount of samples. Comparing the results
of different orders of PCE, we can check the amount of truncation error when a lower order of
PCE is used. On the other hand, the higher order expansion we choose, the greater the error might
become in the coefficient of the expansion function. Further study is necessary on choosing the
right order based on the balance between the truncation error and the coefficient estimation error.

3
 Skewness(𝐲 = 𝐱 𝐤 )

Figure 3.1. Comparison of results in skewness estimation

(ration=result/exact value)

Kurtosis (𝒚 = 𝒙𝒌 )

Figure 3.2. Comparison of results in kurtosis estimation

(ration=result/exact value)

4
 Figure 3.3. Probability density function estimated by PCE (k=7)
3.2.2 Case 2:

𝑥~ log 𝑛𝑜𝑟𝑚𝑎𝑙 (𝜇 = 1, 𝜎 = 0.2) (3.2)

Here we will test the same problem with lognormal distribution. In numerical integration and
MCS, no transformation is involved here rather than PCE. A transformation to the standard normal
variable should be used during the calculation of coefficients. In the following example, the
mapping between the normal and the lognormal variable is available as,
𝑥 = 𝑒𝑥𝑝(𝜎̂𝜉 + 𝜇̂ ) 𝜉~𝑁(0, 12 ) (3.3)

𝜇̂ log 𝜇𝑥 − 0.5 log(𝜎𝑥2 + 1), 𝜎̂ = √log(𝜎𝑥2 + 1) (3.4)

Where 𝜇𝑥 and 𝜎𝑥 demotes the mean and standard variation of 𝑥, respectively. The results of the
skewness and kurtosis calculation are summarized in figure three and four. In the numerical
integration, the evaluation points and weights are obtained by solving the moment-matching
𝑘 𝑘
equation, 𝑀𝑘 = (𝑥 − 𝜇) 𝑓(𝑥)𝑑𝑥 = 𝑤 (𝛼𝜎) 𝑘 = 0, . . . ,2𝑚 − 1. Although the results show

5
bigger error for large 𝑘 values (order of nonlinearity) compared to the results of case 1, we can see
that the numerical integration with the moment matching quadrature rule provides accurate results
up to 2𝑚 − 1 polynomial order.

skewness (ex1.2)

Figure 3.4. Comparison of results in skewness estimation

(Ratio=result/exact value)

Kurtosis (ex1.2)

6
 Figure 3.5. Comparison of results in Kurtosis estimation
(ratio=result/exact value)

Figure 3.6.

In the case no 2, we can see the accuracy follows the sequence of MCS(Highest), 5 node NI, 4 th
order PCE, 3rd order PCE, 3 node NI and 2nd order PCE(Lowest). Contrary to the result of the
normal input case, the accuracy of 4th order PCE is worse than the 5 node numerical integration
and that of the 2nd order PCE is worse than 3 node NI, which implies that the transformation does
have a negative effect on the results. The nonlinearity of 𝑦 is amplified with the transformation in
the Eqs. (3.3) and (3.4). When the nonlinearity exceeds the integration order of the quadrature rule,
the coefficient estimation becomes erroneous. When checking the convergence of PDF obtained
by PCE. We can see in that the convergence behavior is not as good as in Case 1 with the normal
distribution.

𝜇̂ log 𝜇𝑥 − 0.5 log(𝜎𝑥2 + 1), 𝜎̂ = √log(𝜎𝑥2 + 1) (2.5)

7
When the nonlinearity exceeds the integration order of the quadrature rule. When checking the
convergence of probability density function(PDF) obtained by PCE, it is shown that the
convergence behavior is not good as Case 1 with normal distributions.

3.3 Example 2:

I’ll take a three dimensional polynomial for this as follows:

𝑦 = 𝑥1𝑎 𝑥2𝑎 + 2𝑥34 (3.6)

Here is 𝑥𝑖 follows the lognormal distribution with the mean 1. Four values of the standard
deviation, 0.1, 0.2, 0.3 and 0.4 are tested to test the effects if variability of inputs on the moment
estimation. Also we intend to observe the effect of interactions among variables on the
performance of univariate dimension reduction method by trying several values of 𝑎 in Eq. (3.5)
When calculating the PC coefficients, the transformation in Eq. (2.6) used again.

3.3.1 Case 1:
𝑎=2 (3.7)
The results of standard deviation and skewness calculation are summarized in Figs. 3.7 and 3.8.
From the figures it is seen that the results of UDR are almost identical with that from the FFNI,
which implies that the univariate decomposition is only valid for 𝑦, that is, the interaction effects
among variables are not significant in this example. From Table 3.1, we note that the number of
function evaluates for UDR is much less than what used for the rest of the methods. The result of
PCE is much worse than that from the 5𝑛 FFNI which is due to the non-normality of the input
random variables as in case 2 of example 1. Transformation does not always result in degradation
of accuracy, however, we can see that care must be taken when integrating a function with
transformed variables. UDR shows excellent efficiency compared to FFNI. On the other hand, the
performance of MCS is even worse compared to the 5𝑛 FFNI and 5𝑛 UDR.
Table 3.1. number of function evaluation in Example. 2
Method MCS 3𝑛 5𝑛 3𝑛 5𝑛 4th PCE
FFNI FFNI UDR UDR
𝐹𝑛 calls 1,000 k 27 125 7 13 125

8
3.3.2 Case 2:
𝑎=3 (3.8)
When 𝑎 is increasing, the interaction effect between 𝑥1 and 𝑥2 becomes more significant. The
analysis results of the standard deviation and skewness are summarized in table 2.2~2.3 The
computational cost is the same as in case 1. If we compare it to the results of case 1, the results of
UDR show considerable discrepancies with those of MCS, and FFNI especially when the standard
deviations (𝜎) of variables are large. From this, we can understand that when the interaction effects
are strong, the error of using UDR increases.
Table 3.2 Standard deviation calculation results for Ex. 2(case 2)
𝜎 MCS 3𝑛 5𝑛 3𝑛 5𝑛 4th PCE
FFNI FFNI UDR UDR
0.1 1.0001 0.9991 0.9993 0.9878 0.9882 0.9993
0.2 2.6855 2.6722 2.6893 2.5633 2.5812 2.6852
0.3 6.5868 6.4262 6.6236 5.9016 6.1112 6.4625
0.4 17.5817 16.2043 17.6456 14.0806 15.6561 15.5570

Table 3.3 Skewness calculation results for Ex 2(case 2)

𝜎 MCS 3𝑛 5𝑛 3𝑛 5𝑛 4th PCE
FFNI FFNI UDR UDR
0.1 1.0540 0.9990 1.0529 0.9344 0.9978 1.0538
0.2 2.9134 2.2052 2.8981 2.1222 2.8800 2.7198
0.3 7.2888 4.0924 8.2545 3.8276 8.7002 5.2246
0.4 24.6707 8.2237 31.3566 13.7713 33.7417 7.6338

9
 STD (Ex 2.1)

Figure 3.7. Comparison of results of STD estimation

(ratio=result/exact value

Skewness (Ex2.1)

Figure 3.8. Comparison of results of skewness estimation

(ratio=result/exact value)

10
Actually, several categories of uncertainty propagation techniques, including a few techniques that
are receiving growing attention in the present time. FFNI, UDR and PCE are examined in depth
to understand the characteristics and limitations of various methods. Similar examinations are
performed utilizing illustrative precedents. In a perfect world, the exhibitions ought to be assessed
under an extensive scope of framework nonlinearity, different conveyances of info arbitrary
factors, and different dimensionality as far as exactness and effectiveness. Thus, it probably won't
be conceivable to pass judgment or rank the techniques with only a few precedents. Be that as it
may, through this correlation contemplate, a few attributes, preferences and drawbacks of every
strategy can be summed up. But from the test results’ accuracy PCE is very close MCS what my
comparisons show here.

11
12