Professional Documents
Culture Documents
of Stochastic Systems
_____________________1
INTRODUCTION
The last few decades have witnessed tremendous developments in the modeling of mechanical and
structural systems, due to advances in computational mechanics. Numerical methods such as finite
elements, finite difference, boundary elements, etc., have reached broad acceptability and wide
coverage of applications. New developments address the solution of complex, non-linear problems.
Multi-physics analyses allow investigation of new, unforeseen interaction effects between structures,
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
_____________________2
soils, fluids, thermo-dynamic and electric effects. Significant developments have also been recently
achieved in modelling uncertainty propagation through mechanical and other types of systems.
The Monte Carlo simulation method remains a popular, yet computationally expensive tool for
analyzing uncertainty propagation through mechanical systems. The computational cost of Monte
Carlo simulation can easily become prohibitive, for highly non-linear problems and complex
geometries. More efficient, intrusive methods have recently been developed, such as the stochastic
finite element method (Ghanem&Spanos, 1991) or stochastic Galerkin Method (Avila et al., 2010).
Intrusive methods have the inconveniency of requiring full re-programming of conventional finite
element software. Hence, non-intrusive Monte Carlo simulation methods remain popular in the
solution of stochastic mechanics problems.
In linear stochastic mechanics problems, the numerical solution of a differential equation is
replaced by the solution of a system of linear algebraic equations (stiffness matrix). In this context,
when Monte Carlo simulation is employed, for each system realization, the stiffness matrix needs to
be evaluated and inverted. Depending on the dimensions of the linear system, and the required
number of samples, this can become computationally intensive. For a linear operator in finite
dimensions, A : n n , the inverse can be represented by the Neumann series, composed of
operators P : n n related to A .In finite dimensions, linear operators are matrices. The
objective of using the Neumann series is to replace the matrix inversions by a truncated series
expansion. However, depending on the number of terms in the Neumann series, the number of
operations to be performed may become larger than required for the direct linear system solution,
when some interactive method is used to compute the inverse. Therefore, in this paper, properties of
the Neumann series are exploited in order to derive, formally, lower and upper bounds for each
realization of the numerical approximation of the stochastic system response. To increase the
efficiency of the proposed scheme, equivalence between norms of finite dimension operators are
considered.
First use of the Neumann series to solve stochastic problems in mechanics goes back to the late
eighties. Yamazaki (1989) employed the Neumann series to obtain samples of the displacement
response, for a plane elasticity problem with random Youngs modulus. Arajo&Awruch (1994)
derived the response on non-linear static and dynamic problems, with random mechanical properties.
Chakraborty &Dey (1996) obtained expected value and variance of displacement responses
considering uncertain geometries and mechanical properties. Chakraborty &Dey (1998) formulated
the dynamic bending of curved beams, with uncertain stiffness parameters. Lei &Qiu (2000)
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
_____________________3
employed the Neumann series to study uncertainty propagation in structures using movement
equations. Chakraborty& Sarkar (2000) estimated statistical moments of the transversal displacement
response of curved beams resting on Winkler foundations. Chakraborty& Bhattacharyya (2002)
obtained response statistics for linear tri-dimensional elasticity problems, representing elastic
properties as Gaussian processes of the continuous. Li et al. (2006) presented a methodology to
obtain solutions of linear systems, based on a discretization of the stochastic problem. Schevenels et
al.(2007) proposed a methodology based on Green functions, which was compared to the Neumann
series, in a wave propagation problem on random Winkler foundation. In all references above, the
Neumann series was employed in a conventional fashion, as an alternative to solve the stochastic
problem. Convergence properties of the Neumann series were not explored in these articles.
The present article advances the state of the art by employing well-known properties of the
Neumann series to derive formal lower and upper bounds for the realizations of a stochastic system
response. To increase the efficiency of the proposed scheme, equivalence between norms of finite
dimension operators are considered. The developed scheme is illustrated by means of application to
twoexample problems, involving beam bending and axial reaction-diffusion.
2.
Several uncertainty propagation problems in mechanics are derived from well-known physical
principles such as the first and second laws of thermo-dynamics, among other. In linear mechanics,
equilibrium equations are generally derived from the minimization of certain functionals. One of the
most popular is the Principle of Minimum Potential Energy, for which the functional minimization
yields the Euler-Lagrange equations and the boundary conditions. In this paper, we consider a linear
elliptic boundary value problem, whose operator appears in beam and plate bending problems, as
well as heat transfer in steady-state conditions. The stochastic problem to be addressed herein is
defined in a probability space ( , F , P ) ,where" "is the space of events, " F "is a -algebra of
events and" P " is a probability measure. The linear uncertainty propagation problem consists in
finding the system response(u), such that:
( ) ( x, ) = f ( x, ) , ( x, ) D ( , F , P ) , a. e.;
, m
subject to specified boundary conditions.
(1)
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
_____________________4
wheref( , )is the source term. The stochastic system response is a consequence of the randomness in
coefficients { } , <m . In problems of mechanics, these coefficients can be related to stiffness (e.g.,
beam bending) or thermal conductivity (heat transfer). In this paper, these coefficients are
represented as parameterized stochastic processes.
The following hypotheses are required to warrant existence and uniqueness of the solution:
H1 : L ( , F , P; H ( D ) ) , 0, a. e. in D, , ;
H2 : f L , F , P; L ( D ) .
(2)
Hypothesis H1 ensures that coefficients { } , <m are differentiable, positive, and uniformly limited
in probability (Babuska et al., 2005). Hypothesis H2 ensures that source terms have finite variance.
From hypothesis H1 and H2, the Lax-Milgram lemma ensures existence and uniqueness of the
solutions, for random samples of system coefficients { } , <m . In this paper, a formal study about
existence and uniqueness of the solutions will not be performed, in particular because we are so far
dealing with a general uncertainty propagation problem in mechanics. However, we emphasize that
such studies are always required, so that approximated numerical solutions be obtained in spaces
spanned by the appropriate functions, consistent with the theoretical problem solutions. For instance,
existence and uniqueness of the solutions are studied for Galerkin solutions of beam bending in
Avila & Beck (2011), plate bending in Avila & Beck (2010) and reaction-diffusion in Avila et al.
(2013).
Numerical solution of uncertainty propagation problems generally involves replacing the
solution of a stochastic differential equation by the solution of a system of algebraic equations. The
Galerkin method is one of the most popular techniques, and will be used herein. Approximated
numerical solutions are obtained from the abstract variational problem (AVP) associated to Eq. (1).
{ (
)}
a ( u, v ) = l ( v ) , v V .
(3)
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
_____________________5
a ( u, v ) = x, ( k ) . u ( x, k ) v ( x ) dx,
, m D
l ( v ) = ( f .v )( x, ) dx + boundary integrals.
(4)
Eq. (3) represents a system of variational equations. For a fixed sample of coefficients
{ ( x,
( k ) )} , solution of the PVA is given as one realization of the operator defined in Eq. (1).
From the PVA the Galerkin solution is constructed. With this purpose, the approximated solution
) . Without loss
of generality, for the purposes of this article, the source term is considered deterministic. Each
realization of system response is obtained from one sample of system parameters,hence
u = u ( x, ( k ) ) . From the AVP and for the kth realization, numerical solutions are obtained via
Galerkin method. From the ensemble of system realizations, expected value and variance of the
approximated numerical solution can be evaluated.
For the kth realization, approximated numerical solutions um are obtained as:
um x, ( k ) = ui ( k ) i ( x ) ,
where ui ( ( k ) )
i =1
i =1
(5)
are the coefficients to be evaluated. For the kth realization, the approximated
v Vm .
a ( um , v ) = l ( v ) ,
(6)
By replacing this approximated solution into Eq. (3), a linear system of algebraic equations is
obtained:
_____________________6
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
K ( ( k ) ) U ( ( k ) ) = F;
(7)
the response vector, whose entries are coefficients of the linear combination in Eq. (5). For the kth
realization, entries of stiffness matrix and source vector in Eq. (7) are:
K ( ) = kij ( ) , kij ( ) = a ( i , j ) ;
mm
F = [ fi ]m1 , fi = l ( i ) .
(8)
Formally, solution of the linear system of equations in Eq. (7) is given by:
U ( ) = K ( )
Where H ( ) = hij ( )
F = H ( ) F ,
mm
(9)
ui ( ) = hij ( ) f j ,
(10)
j =1
FromEqs.(7) and(10), the approximated numerical solution, for the kth realization of the
um x, = ( hij ( ) f j )i ( x ) = ( F H ( ) ) ( x ) .
(11)
i =1 j =1
where : D is the vector containing the interpolating functions. Hence, Eq. (11) is the
approximated Galerkin solution to the kth realization of the system response.
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
_____________________7
In the solution by direct Monte Carlo simulation, estimates of expected value and variance of
{ (
)}
k =1
observes that each system response evaluation requires one inversion of the stiffness matrix, H ( ) .
Computation of this inverse should be avoided, due to the large computational cost involved. For
solution of the linear system (Eq. 9), iterative methods such as Jacobi, Gauss-Seidell or Conjugated
Gradients are generally employed (Kincaid & Cheney, 2002). However, the cost of direct Monte
Carlo simulation can still become prohibitive, for instance for non-linear problems involving large
number of samples. Under such conditions, one alternative to reduce the computational cost is use of
the Neumann series.
3.
dimensional space. In this case, the mathematical object of the operator is a matrix K : n n . The
stiffness matrix, for the kthsystem realization, can be decomposed as follows:
K ( ) = K0 I + P ( ) ,
1
P ( ) = ( K0 ) K ( ) ;
K ( ) = K0 + K ( )
mm
(12)
( )
expected values of the stiffness coefficients . Substituting Eq. (12) in Eq. (11), a formal
solution U = U ( ) is given by:
U ( ) = K ( )
( )
where U0 = K0
then I + P ( )
2012):
F = I + P ( )
U0
(13)
F and U0 = u10 ,, um0 . From the Neumann series theorem, if 0 < P ( ) < 1
exists, and can be represented by the following series (Golub& Van Loan,
_____________________8
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
( I + P ( )) = ( P ( )) .
(14)
q =0
Moreover, the theorem provides an upper bound, for the summation in Neumann series, Eq.
(14),ofmatrix I + P ( )
( I + P ( ))
1
1 P ( )
..
(15)
U ( ) = P ( ) U0 .
q =0
(16)
um x, = ( pij( q ) ( ) u 0j ) i ( x ) = ( U0 )
( P ( )) ( x );
i =1 j =1 q =1
(17)
q =0
where
( P ( ))
= pij(
q)
( )
mm
(18)
umn x, = ( pij( q ) ( ) u 0j ) i ( x ) = ( U0 )
i =1 j =1 q =0
( P ( )) ( x ) .
q =0
(19)
_____________________9
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
Com o resultado relacionado a uma cota superior, para a soma da srie de Neumann, Eq. (15), so
{ (
( ) )}
, das
k =1
4.
In this section the main contribution of this article is presented: based on properties of the Neumann
expansion and classical results of algebra, a methodology is proposed to derive lower and upper
bounds for samples of the response process.The main ideas are: to avoid direct computation of the
linear system of equations (Eq. 9) and to obtain an approximated solution, umn = umn x,
)) , using
a truly small (n=0 orn=1) number of terms in the Neumann expansion, but also introducing a
correction trough an estimate of the summation in Eq. (15).
If
the
set
of
coefficient
samples,
{ ( x ,
( k ) )}
k =1
are
such
that
0 < P ( ( k ) ) < 1, k {1, ..., N } ; then the numerical approximations to the displacement and
temperature fields can be represented in terms of the Neumann series using matrix I + P ( )
,Eq. (19).For thekthresponse realization, and for a fixed x D , the distance between response
realizations obtained by direct Monte Carlo, Eq. (11), and by the Neumann series, Eq. (19), can be
estimated as:
( um umn ) x,
q = n +1
P ( )
(20)
U0 ( x ) .
q = n+1
P ( )
can be
evaluated as:
q =n +1
P ( ) =
P ( )
n +1
1 P ( )
(21)
_____________________10
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
Substituting Eq. (21) in Eq. (20), the distance between the direct Monte Carlo andthe Monte Carlo
Neumann solutions become:
( um umn ) ( x, ) ( n, )
U0 ( x ) , ( x, ) D ( , F , P ) ;
(22)
where
n, =
n +1
( )
P ( )
P
1
(23)
The coefficient ( , ) depends on the number of terms ("n") used in the Neumann series for
( ))
matrix I + P
( )
the lower and upper bounds, for the kthsample, of the response process:
) (
x, um x, x, , x, D ( , F , P ) ;
(24)
where ( , ) and ( , ) are lower and upper bounds, respectively, for response process, defined as:
(
(
)
)
(
(
) (
) (
)
)
x, = umn x, + n, U0 ( x ) ,
x, = umn x, + n, U0 ( x ) , x, D ( , F , P ) ;
(25)
where n, = n, .
With the functions in Eq. (25), it is possible to obtain lower and upper bounds for expected
value and variance of the response process, from the ensemble of realizations of lower and upper
{(
( ) ) , ( x, ( k ) )}
N
k =1
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
_____________________11
In order to compute coefficient " " in Eq. (25), one needs to solve for the norm P , which
may involve high computational cost and hence turn the procedure useless. In order to circumvent
this difficulty, we resort to the equivalence between matrix norms (Golub& Van Loan, 2012):
c P
(26)
P C P ,
where 0 < c C < + ,are equivalence constants. Based on this equivalence, five alternatives are
proposed herein to evaluate P and the coefficient " " :
( ) )
(
( n, ) = 1 m . P
( )
P ( )
( n, ) = 1 P ;
( )
m . P
n +1
;
1
n +1
( ) )
( n, ) = 1 m . P
( )
P ( )
n
,
=
;
( ) 1 P
( )
m . P
n +1
n +1
F
( n, ) =
n +1
{ ( ) }) ;
1 m.( max { p ( ) } )
m.max pij
i,j
ij
i,j
(27)
where
( )
= max pij ;
1 j m
1
i =1
( )
= max pij ;
1i m
j =1
( )
( )
( )
=
F
p ( )
ij
i =1 j =1
(28)
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
_____________________12
Within this paper, for simplicity, the norms in Eq. (27) are referred to as follows: 1 is called
norm one, 2 is norm two, is the infinity norm or norm 3, F is the Frobenius norm or norm 4
and max { } is the maximum norm, or norm 5.
The definition of coefficient " " is based on estimates about the sum in Neumann series; hence
the equivalent norm, ,inEq. (26), must comply with the hypothesis inEq. (17), that is: C P
5.
< 1.
LOWER AND UPPER BOUNDS FOR THE EXPECTED VALUE AND COVARIANCE
OF THE RESPONSE PROCESS
From the ensemble of computed lower and upper bounds for system response, one can compute
lower and upper bounds for the expected value and for the variance of the response process.
({
Let um ( x, ( j ))
N
j =1
of the response process. Then, estimates for expected value and covariance are defined as:
N
1
x
=
(
)
(
)
N um x, ( j ) , x D;
um
i =1
N
( 2)
1
x
,
y
=
(
)
(
)
N 1 um x, ( j ) um y , ( j ) , ( x , y ) D D.
um
j =1
) (
(29)
InsertingEq. (29)in Eq. (25), one obtains lower and upper bounds for the expected value and
covariance of the response process:
( x )
um ( x ) i ( x ) , x D, i = 1, ...,5 ;
i
( 2)
(2)
( 2)
i ( x, y ) um ( x, y ) i ( x, y ) , ( x, y ) D D, i = 1, ...,5.
(30)
In the following, direct Monte Carlo simulation is employed to evaluate the performance of the
proposed lower and upper bounds, in the solution of two linear problems.In order to evaluate the
accuracy of the proposed lower and upper bounds, deviation (error) functions need to be established.
The relative deviation, with respect to direct Monte Carlo simulation, of the k order moment lower
and upper bounds, ( ) , ( ) , i = 1, ...,5 , are given by:
_____________________13
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
k = 100%
(i)
( k ) = 100%
6.
k
( )
) .1 (ki) , i = 1, ...,5;
)
um
k
( )
. 1 ki , i = 1, ...,5.
u( m)
(31)
NUMERICAL RESULTS
In this section, numerical results are presented for two examples, one involving bending of a beam
on Winkler-typefoundation; the other involving reaction and diffusion in a bar. Direct Monte Carlo
simulation is used to compute the reference solution, to which the proposed Monte Carlo - Neumann
bounds are compared. In Direct Monte Carlo, the kth realization of the response process is obtained
from the kth coefficient sample, which consists in inserting in Eq. (9) the sampled values of vectors
= ( k ) , k = ( k ) )
; evaluate from Eq. (10) the stiffness matrix; solve the linear system of
equations (Eq. 9) for U ( ) and compute the coefficients of the linear combinations in Eq. (7).
Monte carlo Neumann bounds are computed for the five equivalent normsshown in Eq. (27), and
for Neumann expansions with one or two terms (n=0 and n=1). Parameterized processes are used to
represent the uncertainty in coefficients. Is all cases, the source term is considered deterministic. All
computer simulations are performed in MATLAB. In order to reduce spurious correlations, Latin
Hypercube Sampling is employed (Olsson &Sandberg, 2002).
In this example, the proposed Monte Carlo - Neumann bounds are computed for a problem involving
bending of an Euler-Bernoulli beam on a Winkler foundation. Uncertainty is considered on the beam
and
foundations
stiffness,
one
at
time.
The
vector
space
2
2
Q = v H 4 ( 0, l ) v ( 0 ) = v ( l ) = 0 d v2 ( 0 ) = d v2 ( l ) = 0 is employed in the solution, and the beam
dx
dx
_____________________14
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
d 2 EI d 2u + u = q, ( x, ) ( 0, l ) , a. e.;
W
dx2
dx2
u ( 0, ) = u ( l , ) = 0;
d 2u
2
= d 2u
= 0, ;
dx2
dx
( 0 ,)
( l ,)
(32)
whereEI is the bending stiffness, W is foundation stiffness coefficient, u is the transversal beam
displacement, q is the load term. Beam length is one meter ( l = 1 m ) , and the deterministic loading
2
2
[ kPa.m ] , x [ 0,1] .
EI ( x, ( ) ) = EI +
( x, ( ) ) = +
W
W
EI
k =1
( )
3
3
where ( EI , EI ) and W , W
2.k 1
2.k
(33)
k =1
( )
( )
foundation stiffness, respectively.The following numerical values are considered for these
parameters:
EI ( x ) = 1400 N.m 2
W ( x ) = 1400 Pa
In
{ ( )}
j
order
to
EI ( x ) = 101 , x [ 0,l ];
(34)
( x ) = 101 , x [ 0, l] .
characterize
= {1 ( j ) , 2 ( j ) , 3 ( j ) , 4 ( j )}
j =1
N
j =1
the
and
uncertainty,
{ ( )}
j
random
= {1 ( j ) , 2 ( j ) , 3 ( j ) , 4 ( j )}
j =1
vectors
N
j =1
are
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
_____________________15
( 2l , , ) , obtained by fixing x =
l
2
problem.
AVILA, este grfico est idntico ao do paper anterior, acredito que vc no atualizou.
ATUALIZAR OU REMOVER!
Figure 2: Convergence of Monte Carlo simulation results for the mean and
variance of mid-spam displacement ( x = 12 ) .
One observes in Figure 2 that for N 5, 000 realizations, the expected value and variance of
mid-spam displacements stabilize at values um ( 12 ) 0.00863 m and u2m ( 12 ) 2.03524 107 m 2 .
Figure 2 shows that using N=30,000 samples is sufficient to obtain reliable results via direct Monte
Carlo simulation.
7.1.1 Example 1a: random beam stiffness
In this example, random beam stiffness is considered, following Eq. (33). Foundation
stiffness is considered deterministic and equal to the mean value in Eq. (33).
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
_____________________16
Figure 3 shows relative deviations of lower and upper bounds for the expected value and
variance of the displacement process, Eq. (31). The deviations are computed by comparing the
proposed Monte Carlo Neumann bounds, for n=1, with the direct Monte Carlo solution, following
Eq. (35). One observes in Fig. 3 that the most accurate bounds are obtained when coefficient " " is
computed using norm 2.The largest deviations are obtained for coefficient " " evaluated by the
maximum norm, ( mx ) , or norm 5. One can also observe that the relative deviations are larger for
variance then for expected values.
AVILA, acredito que na figura de cima os resultados para as normas 2 e 5 esto trocados para o
lowerbound (betas). Alem disto, os alfas esto trocados pelos betas, em abas as figuras, pois os alfas
aparecem como upperbound e os betas como lowerbounds... Deveria ser ao contrrio. Ao refazer
estas figuras, e as demais, sugiro que voc utilize apenas 1 , 2 , 3 , 1 ,2 ,3 para identificar as
linhas, porque os smbolos so redundantes, e j esto no eixo y. Assim evita-se utilizar 3subindices... lembre, isto apenas uma figura, no uma equao matemtica!
_____________________17
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
Table 1 shows numerical values of the relative deviations in expected value and variance,
computed viaMonte Carlo-Neumann using different norms, for n=0 and n=1. The table shows
maximum deviation values within the domain, max { } , and deviations computed for x = 35 m . One
x[0,1]
observes that the deviations in variance are larger than deviations in expected values. Also, one
observes that using n=1 in the Monte Carlo-Neumann solution significantly reduces the relative
deviations. The largest deviations are obtained when coefficient " " is computed using the
maximum norm, or norm 5 (Eq. 27). When the bounds are evaluated by norm 2(Eq. 27), the smallest
deviations are obtained. Also, it is observed that relative deviations are larger for lower bounds than
for upper bounds.
Table 1: Relative deviations in expected value and variance, computed via
Monte Carlo-Neumann using different norms, for n=0 andn=1 (example 1).
ExpectedValue
Norm
mx ( x )
x[0,1]
i
( 53 )
i
mx ( x )
x[0,1]
( 53 )
n=0
26.113274414633
26.10600793330
27.086214428717
27.0855386560489
n=1
6.2521309420642
6.2462587341439
7.24094736673004
7.24084980083362
n=0
10.119519951647
10.113175009457
11.0935297267617
11.0927057322009
_____________________18
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
n=1
0.7428825455645
0.7373148157940
1.73200924279316
1.73190588248217
n=0
28.041455883294
28.034078302682
29.0142669286992
29.0136090254253
n=1
7.1337137575271
7.1277928274081
8.12248053277713
8.12238389409656
n=0
14.151309210285
14.144731961689
15.1250493144435
15.1242626844328
n=1
1.8564968574000
1.8508675818060
2.84556083754879
2.84545864849398
n=0
31.321306114917
31.313739553637
32.2938977836913
32.2932702763821
n=1
8.9579417201056
8.9519199708700
9.94660575758618
9.94651103755849
Variance
Norm
mx (2) ( x)
x[0,1]
(2) ( 35 )
mx ( 2) ( x )
x[ 0,1]
(2) ( 53 )
n=0
26.113274414633
60.5045259492099
27.086214428717
44.8500231574282
n=1
6.2521309420642
13.1987145418507
7.24094736673004
13.5236739469688
n=0
10.119519951647
20.9411702510525
11.0935297267617
21.0605414343467
n=1
0.7428825455645
1.46867650556004
1.73200924279316
3.42271218462158
n=0
28.041455883294
65.7674054359802
29.0142669286992
47.2248604672628
n=1
7.1337137575271
15.2008116118716
8.12248053277713
15.0174013519348
n=0
14.151309210285
30.356495620076
15.1250493144435
27.6155306034285
n=1
1.8564968574000
3.75913781264843
2.84556083754879
5.54235358058716
n=0
31.321306114917
75.8065085937241
32.2938977836913
50.1775672670156
n=1
8.9579417201056
19.6493540549172
9.94660575758618
17.7694632216577
3
4
Using norm 2 leads to the most accurate results, but there is a computational penalty. Table 2
compares the computation time taken to obtain the direct Monte Carlo solutions ( TMC ), and the
solutions via Monte Carlo Neumann with different norms (T1,, T5 ) , Eq. (27), and n values. When
the bounds are evaluated via Neumann series with n=0, the computation time is zero, (TSN = 0 ) ,
1
since ( I + PR ( ) ) I . One observes in Table 2 that for some norms (1, 2 and 5) there is no gain in
computational time for the proposed Monte Carlo Neumann solutions, when compared to direct
Monte Carlo ( TMC ): in fact, computation times have increased significantly. However, when the
_____________________19
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
Frobeniusnorm is used (norm 4 in Eq. 27), there is reduction between nearly two and four times in
computation time. Figure 3 and Table 1 show that accurate bounds are obtained when the Frobenius
norm is used (norm 4 in Eq. 27) to compute coefficient " " , with errors of less than around 3% for
n=1.
Table2: Computation (clock) time to obtain the direct Monte Carlo ( TMC ) and Monte Carlo
Neumann solutions, using different norms ( Ti ) and n values.
Reduction
w.r.t. MC
Norm
Example
TMC
1a
1b
TSN
T1
T2
T3
T4
T5
forT4
n=0
55920.383
15107.916
813.997
202.930
3240.062
4.53x
n=1
183.599
59387.427
17005.949
889.626
215.823
3856.594
2.30x
n=0
58390.597
18364.812
1491.728
207.279
3374.629
4.42x
n=1
170.290
64116.130
16210.235
1070.900
196.287
1600.975
2.50x
918.35
916.48
_____________________20
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
that the deviations in variance are larger than deviations in expected values, but also that deviations
are smaller for example 1b than for example 1a.
Computation times for the direct Monte Carlo solution and for the Monte Carlo-Neumann
solutions using different norms are shown in Table 2. Computation times are similar to those found
for example 1b: for norms 1, 2 and 5, computation times are actually larger. Only when using the
Frobenius norm, norm 4, one obtains a computational advantage.
_____________________21
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
Norm
mx ( x )
x[ 0,1]
i
( 53 )
i
mx ( x )
x[0,1]
( 53 )
n=0
0.55697446743920
0.13062893617796
0.55711673670140
0.13072164182149
n=1
0.00075591311534
0.00014775535719
0.00089769428055
0.00024015834887
n=0
0.32872788297855
0.07708540531248
0.32887015224216
0.07717811095736
n=1
0.00021062957960
0.00001983928823
0.00035241074481
0.00011224228156
n=0
0.67354149997320
0.15797396970294
0.67368376923585
0.15806667534814
n=1
0.00104651391437
0.00021592633521
0.00118829508012
0.00030832932733
n=0
0.32931739240249
0.07722369617095
0.32945966166493
0.07731640181515
n=1
0.00021165717338
0.00002008034767
0.00035343833978
0.00011248333982
n=0
0.95331306724635
0.22360455770487
0.95345533651002
0.22369726334912
n=1
0.00236860847586
0.00052607165923
0.00251038964174
0.00061847465089
Variance
Norm
mx (2) ( x )
x[ 0,1]
(2) ( 35 )
mx ( 2) ( x )
x[ 0,1]
(2) ( 53 )
n=0
1.11833883178381
0.26146127655633
1.1098398121527
0.26123941016273
n=1
0.00151185090597
0.00029551299445
0.00179537725603
0.00048031658758
n=0
0.65888415307608
0.15421129433324
0.65630969544336
0.15431546121246
n=1
0.00042126400376
0.00003967929136
0.00070482114370
0.00022448457887
n=0
1.35305077234131
0.31623814790298
1.34139521400725
0.31584257471974
n=1
0.00209306498271
0.00043185595419
0.00237657229332
0.00061665849960
n=0
0.66006874773700
0.15448818275198
0.65748313449240
0.15459173572919
n=1
0.00042331921414
0.00004016141246
0.00070687632488
0.00022496669832
n=0
1.91978967640438
0.44789524689581
1.89374068842934
0.44670759204049
n=1
0.00473739034083
0.00105215517005
0.00502064515243
0.00123694382023
_____________________22
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
d
du
dx EA dx + u = q, ( x, ) ( 0, l ) ( , F , P ) , a. e.;
u ( 0, ) = u ( l , ) = 0, , a. e.;
(35)
whereEA is the membrane (diffusion) stiffness, is (reaction) stiffness coefficient, u is the axial
displacement, q is the load term. Two cases are considered: 2a) uncertain membrane stiffness
coefficient and 2b) uncertain reaction coefficient. A total of thirty thousand samples are considered
(N=30,000) in the direct Monte Carlo and in the Monte Carlo Neumann solutions presented herein.
EA ( x, ( ) ) = EA + ( 23 . EA ) 2.k 1 ( ) cos
k =1
( k4x ) +
2.k
( ) sin ( k 4x ) ,
(36)
where EA and EA are the mean and the standard deviation of membrane stiffness, respectively.The
following numerical values are considered for these parameters:
EA ( x ) = ?? N.m2
EA ( x ) = 101 , x [0, l] .
(37)
Figure 5 shows the expected value of axial displacement, obtained by direct Monte Carlo
simulation, and the lower and upper bounds computed by the proposed Monte Carlo Neumann
bounds, for n=1, using some of the norms in Eq. (27). Figure 6 shows the errors in expected
values for the computed bounds. Figure 7 shows the variance of axial displacement, obtained by
_____________________23
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
direct Monte Carlo simulation, and the lower and upper bounds for the variance, computed by the
proposed Monte Carlo Neumann bounds, for n=1, using some of the norms in Eq. (27). Figure 8
shows the errors in variance for the computed bounds. It can be observed that accurate bounds for
expected value and variance are obtained using the Frobeniusnorm, or norm 4 in Eq. (27). It is
also observed that errors for expected value are slightly smaller than errors on variance.
Figure 5: Expected value um computed by direct MCS and
proposed bounds, i and , for i = 3, 4 .
i
_____________________24
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
AVILA, estas duas figuras ficaram mais esticadas que as anteriores possvel manter um
padro??
2
Figure 7: Variance u( m) of the displacement response, computed by
2
2
direct MCS and proposed bounds, ( ) and ( ) for i = 3, 4 .
i
Table 4 shows numerical values of the relative deviations in expected value and variance, for
the displacements evaluated at x = 23 m , computed viaMonte Carlo-Neumann using different norms,
for n=0 and n=1. One observes that the deviations in variance are larger than deviations in expected
values. Also, one observes that using n=1 in the Monte Carlo-Neumann solution significantly
reduces the relative deviations. The largest deviations are obtained when coefficient " " is computed
using the maximum norm, or norm 5 (Eq. 27). When the bounds are evaluated by norm 2(Eq. 27),
the smallest deviations are obtained.
_____________________25
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
Norm
( 23 ) [%]
( 2) ( 23 ) [% ]
( 23 ) [%]
n=0
20.0956061561214
19.3236175106023
35.1764346892333
43.0654287180933
n=1
4.42308802116401
3.63497429876076
8.47915276885897
7.50748781845525
n=0
8.61229013089511
7.84030148537598
16.6023598709104
16.0538242493126
n=1
1.18675203868329
0.398638316279554
2.35430501010019
0.7914992100172
n=0
17.8711289455031
17.0991402999832
31.8659492578897
37.5453372949846
n=1
3.63577285775532
2.8476591353521
7.02235209901952
5.83958409768135
n=0
12.170938351111
11.3989497055909
22.5960989232118
24.1869624949976
n=1
2.04510558190445
1.25699185950121
4.0070032341121
2.54709801797477
n=0
24.1738868772815
23.4018982317626
39.5717406704905
54.8598014489074
n=1
6.49294150007064
5.70482777766681
11.8697845644908
12.3433141666401
(2) ( 23 ) [% ]
i
Table 5 compares the computation time taken to obtain the direct Monte Carlo solutions ( TMC ),
and the solutions via Monte Carlo Neumann with different norms (T1,, T5 ) , Eq. (27), and n
values. When the bounds are evaluated via Neumann series with n=0, the computation time is zero,
(TSN = 0 ) , since ( I + PR () )
I . As for Example 1 (Table 2), one observes in Table 5 that for some
norms (1, 2 and 5) computation times have increased significantly for the proposed bounding
solutions, when compared to direct Monte Carlo ( TMC ). However, when the Frobenius norm is used
(norm 4 in Eq. 27), there is a reduction between around two times in computation time (last column
of Table 5). Figures5 to 8 and Table 4 show that accurate bounds are obtained when the Frobenius
norm is used (norm 4 in Eq. 27) to compute coefficient " " , with errors of less than around 4% for
n=1.
Table 5: Computation (clock) time to obtain the direct Monte Carlo ( TMC ) and Monte Carlo
Neumann solutions, using different norms ( Ti ) and n values.
Example
Norm
Reduction
w.r.t. MC
_____________________26
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
TMC
2a
2b
TSN
T1
T2
T3
T4
T5
forT4
n=0
1987.203
2603.017
100.730
175.501
778.726
1,63x
n=1
63,179
2647.960
1263.717
157.701
62.822
226.435
2,27x
n=0
2780.281
1713.171
185.579
57.455
376.025
5,18x
n=1
60,772
2910.432
1745.042
135.976
61.322
808.210
2,44x
286.19
297.77
( x, ( ) ) = +
3
2
2.k 1
2.k
(38)
where and are the mean and the standard deviation of reaction stiffness, respectively.The
following numerical values are considered for these parameters:
( x ) = ?? N.m2
( x ) = 101 , x [0, l] .
(39)
Membrane stiffness is considered deterministic and equal to the mean value in Eq. (37).
Figure 9 shows the expected value of axial displacement, obtained by direct Monte Carlo
simulation, and the lower and upper bounds computed by the proposed Monte Carlo Neumann
bounds, for n=1, using some of the norms in Eq. (27). Figure 10 shows the errors in expected values
for the computed bounds. Figure 11 shows the variance of axial displacement, obtained by direct
Monte Carlo simulation, and the lower and upper bounds for the variance, computed by the proposed
Monte Carlo Neumann bounds, for n=1, using some of the norms in Eq. (27). Figure 12 shows the
errors in variance for the computed bounds. It can be observed that accurate bounds for expected
value and variance are obtained using the Frobeniusnorm, or norm 4 in Eq. (27). As for Example 2a,
one observes that the relative deviations are larger for variance then for expected values. However,
for example 2b the deviations are significantly smaller. Hence, for uncertain reaction stiffness,
_____________________27
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
uncertainty propagation to displacement response is much smaller than for random membrane
stiffness.
Figure 9: Expected value um computed by direct MCS and
proposed bounds, i and , for i = 3, 4 .
i
Figure 10: Relative error functions for expected value, and ,for i = 3, 4 .
i
_____________________28
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
2
Figure 11: Variance u( m) of the displacement response, computed by
2
2
direct MCS and proposed bounds, ( ) and ( ) for i = 3, 4 .
i
Table 6 shows numerical values of the relative deviations in expected value and variance, for
the displacements evaluated at x = 23 m , computed viaMonte Carlo-Neumann using different norms,
for n=0 and n=1. One observes that the deviations in variance are larger than deviations in expected
values. Also, one observes that using n=1 in the Monte Carlo-Neumann solution significantly
_____________________29
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
reduces the relative deviations. The largest deviations are obtained when coefficient " " is computed
using the maximum norm, or norm 5 (Eq. 27). When the bounds are evaluated by norm 2 (Eq. 27),
the smallest deviations are obtained. Small deviations are also obtained when using the Frobenius
norm, norm 4, which is the least expensive to compute, following Table 5.
Table6: Relative deviations in expected value and variance, computed via
Monte Carlo-Neumann using different norms, for n=0 and n=1 (example 2b).
Norm
( 23 ) [%]
i
( 23 ) [%]
(2) ( 23 ) [% ]
i
( 2) ( 23 ) [% ]
n=0
0.99347730595809
0.986994190946033
1.97581932955457
1.98488300460022
n=1
0.0176626552439774
0.0113692253334241
0.03532029252115
0.022743605435127
n=0
0.561992616901985
0.555509501889384
1.12235546332121
1.11251297949093
n=1
0.0077875115237713
0.0014940816137620
0.01557453448717
0.002988666325620
n=0
1.1469316184861
1.14044850347372
2.27900989536222
2.29547239448988
n=1
0.0215961385162
0.0153027086069013
0.04318461222319
0.030612394578089
n=0
0.582424287323269
0.575941172311012
1.16287900081569
1.15371115811809
n=1
0.0081408387380244
0.0018474088278264
0.01628110933248
0.003695412683857
n=0
1.63476681521783
1.62828370020502
3.23294408314287
3.29276043373661
n=1
0.0436595557790853
0.0373661258683988
0.08727704375430
0.074775452000452
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
_____________________30
the Frobenius norm is used (norm 4 in Eq. 27) to compute coefficient " " , with errors of less than
around 1% for n=1.
7.
CONCLUDING REMARKS
In this paper, a well-known property of the Neumann series was explored to derive lower and upper
bounds for expected value and variance of the response of stochastic systems. To increase the
efficiency of the proposed scheme, equivalence between norms of finite dimension operators were
considered.Problems of stochastic beam bending and axial reaction-diffusion where addressed herein
to illustrate theproposed methodology. Uncertain problem parameters (beam and foundation
stiffness, membrane and ?? stiffness) were represented as parameterized stochastic processes.
Solutions employing the proposed Monte Carlo Neumann scheme were computed for five
equivalent norms using one or two terms in the Neumann expansion (n=0 or n=1). Deviations of the
proposed Monte Carlo Neumann bounds were computed by comparing to a reference solution,
computed by direct Monte Carlo simulation. It was shown that accurate bounds can be obtained
using the proposed Monte Carlo Neumann scheme, using as few as one or two terms in the
Neumann expansion (n=0 or n=1). The most accurate bounds were obtained with norm 2
( ) , but
2
these were also shown to be computationally expensive to compute. Accurate results were obtained
by using the Frobeniusnorm
studied herein. It remains to be investigated weather this holds for other problems as well. The
proposed Monte Carlo Neumann bounding scheme proposed herein is a viable alternative or
complement to the solution of uncertainty propagation problems in mechanics.
ACKNOWLEDGEMENTS
The authors gratefully acknowledge theBrazilian National Research Council (CNPq) for sponsoring
this research.
REFERENCES
Babuska I; Tempone R, Zouraris GE. (2005) Solving elliptic boundary value problems with uncertain
coefficients by the finite element method: the stochastic formulation. Computer Methods in Applied
Mechanics and Engineering, v. 194, n. 12-16, p. 1251-1294.
Ghanem R. &Spanos P.D., 1991: Stochastic Finite Elements: A Spectral Approach, Dover, NY.
vila S Jr. CR, Beck AT, Mantovani G, Azikri H. (2010) Galerkin Solution of Stochastic Beam Bending on
Winkler Foundations. Computer Modeling in Engineering & Science; v. 1729, p. 1-31.
Avila & Beck: Efficient Bounds for the MC-Neumann Sol. of Stochastic Systems
_____________________31
Brenner SC, Scott LR. (1994), The Mathematical Theory of Finite Element Methods. Springer-Verlag.
Chakraborty S., DeyS.S. (1998) A stochastic finite element dynamic analysis of structures with uncertain
parametersInternational Journal of Mechanical Sciences, v. 40, n. 11,p. 1071-1087.
Chakraborty S., BhattacharyyaB. (2002) An efficient 3D stochastic finite element method, International
Journal of Solids and Structures, v. 39, n. 9, p. 2465-2475.
Lei Z., QiuC.(2000)Neumann dynamic stochastic finite element method of vibration for structures with
stochastic parameters to random excitationComputers & Structures, v. 77, n. 6, p. 651-657.
Hyuk-Chun Noh, Taehyo Park(2006) Monte Carlo simulation-compatible stochastic field for application to
expansion-based stochastic finite element methodComputers & Structures, v. 84, n. 3132,p. 2363-2372.
AdhikariS. A reduced spectral function approach for the stochastic finite element analysisComputer Methods
in Applied Mechanics and Engineering, v. 200, n. 2122, 2011, p. 1804-1821.
Chakraborty S., DeyS.S.Stochastic finite element simulation of random structure on uncertain foundation
under random loading,International Journal of Mechanical Sciences, v. 38, n. 11, 1996, p. 1209-1218.
Chakraborty S., DeyS.S. Stochastic finite element method for spatial distribution of material properties and
external loading, Computers & Structures, v. 55, n. 1, 1995, p. 41-45.
ArajoJ.M., AwruchA.M.,On stochastic finite elements for structural analysisComputers & Structures, v. 52,
n. 3, 1994, p. 461-469.
Bhattacharyya B., ChakrabortyS.,NE-MCS technique for stochastic structural response sensitivity, Computer
Methods in Applied Mechanics and Engineering, v. 191, n. 4950, 2002, p. 5631-5645.
D. Kincaid e W. Cheney (2002). Numerical Analysis: mathematics of scientific computing, 3a ed., AMS.
Li C.F., Feng Y.T., OwenD.R.J. (2006) Explicit solution to the stochastic system of linear algebraic equations,
Computer Methods in Applied Mechanics and Engineering, v. 195, n. 4447, p. 6560-6576.
Rahman S., XuH. A univariate dimension-reduction method for multi-dimensional integration in stochastic
mechanicsProbabilistic Engineering Mechanics, v. 19, n. 4, October 2004, p. 393-408
Chakraborty S., SarkarS. K.,Analysis of a curved beam on uncertain elastic foundationFinite Elements in
Analysis and Design, v. 36, n. 1, 2000, p. 73-82.
SchevenelsM., LombaertG., DegrandeG., ClouteauD.The wave propagation in a beam on a random elastic
foundation, Probabilistic Engineering Mechanics, v. 22, n. 2, 2007, p. 150-158.
Rao MM, Swift JR. (2010), Probability Theory with Applications. Springer; 2nd ed.
Olsson AMJ,SandbergGE.(2002) Latin Hypercube Sampling for Stochastic Finite Element
Analysis.Journal of Engineering Mechanics; v. 128, n. 1, p. 121-125.
Golub GH, Van LoanCF, (2012) Matrix Computations, Johns Hopkins Studies in the Mathematical Sciences.
Yoshida K. (1978) Functional analysisSpringer Berlin.