Fokker Planck1

Nonlinear Dynamics and Stochastic Mechanics, W. Langford, W. Kliemann, and N. Sri Namachchivaya, editors, Fields Institute Communications, vol.
9, AmericanMathematicalSociety,Providence,1996, 23-37. Presented at the lnternational Symposium on Nonlinear Dynamics and Stochastic Mechanics, Fields Institute for Research in Mathematical Sciences, Waterloo, Ontario, Canada, Aug 28 -Sept1, 1993.
ROBUST NUMERICAL SOLUTION OF THE FOKKER-PLANCK EQUATION FOR SECOND ORDER DYNAMICAL SYSTEMS UNDER PARAMETRIC AND EXTERNAL WHITE NOISE EXCITATIONS
Lawrence A. Bergman1 , Billie F. Spencer, Jr.2, Steven F. Wojtkiewicz3, and Erik A. Johnson4
1;3;4Department of Aeronautical and Astronautical Engineering
University of Illinois at Urbana-Champaign Urbana, Illinois 61801 USA and Geological Sciences University of Notre Dame Notre Dame, Indiana 46556 USA
2Department of Civil Engineering
the Fokker-Planck equation for second order systems under additive white noise excitation. Animation of these solutions has provided insights into previously unstudied nonstationary behavior of the transition probability density functions of these systems. This work has now been extended to encompass systems subjected to both additive and multiplicative white noise excitations. In this paper we will examine a sequence of Du ng systems having a negative linear restoring force, and we will demonstrate through inspection of the probability density functions that adding parametric excitation to the linear part of the sti ness term results in stabilization at the origin. We will also compute second moments for each system and show the limited accuracy of several closure methods for this class of systems.
Abstract. The nite element method has been previously applied to the solution of
1. Introduction
The nite element method has been previously applied to the solution of the Fokker-Planck equation for second order systems under additive white noise excitation Langley 1985 , Langtangen 1991 , Bergman and Spencer 1992 , Spencer and Bergman 1993 . In the case of the former two, only stationary probability density functions have been obtained while, for the latter two, the evolution of the transition probability density function has been reported for a number of example systems. More recently, an algorithm for systems subjected to both additive and multiplicative white noise excitations has been demonstrated Bergman, Wojtkiewicz, and Spencer 1993 through solution of a sequence of benchmark problems for which
This
paper is dedicated to Prof. S. T. Ariaratnam on the occasion of his sixtieth birthday. Typeset by AMS-TEX 1
Nonlinear Dynamics and Stochastic Mechanics, W. Langford, W. Kliemann, and N. Sri Namachchivaya, editors, Fields Institute Communications, vol. 9, AmericanMathematicalSociety,Providence,1996, 23-37. Presented at the lnternational Symposium on Nonlinear Dynamics and Stochastic Mechanics, Fields Institute for Research in Mathematical Sciences, Waterloo, Ontario, Canada, Aug 28 -Sept1, 1993.
analytical solutions exist for the stationary probability density function. Furthermore, post-processing now permits the rapid computation of statistical moments of arbitrary order, upcrossing statistics, marginal density functions, and so forth. We will discuss in some detail the results obtained for one of these systems, the hardening Du ng oscillator. We will examine the case wherein the linear part of the restoring force is negative. In the absence of parametric excitation, the oscillator exhibits three xed points: two stable and symmetrically disposed about the origin on the displacement axis, and one unstable at the origin. The stationary probability density function is, thus, bimodal Caughey 1963,1971 , and the dynamics of the evolution to stationarity are particularly rich. As parametric excitation is added to the linear part of the restoring force, the system appears to become increasingly stable at the origin. It is this behavior that we wish to examine in some detail. Furthermore, when the second moments for these systems, as computed from the nite element solution, are compared with those obtained from various closure methods, the error associated with the latter can be quite large. These results are examined and discussed as well.
2. The System
The system to be examined is the hardening Du ng oscillator with a negative linear term in the restoring force. The oscillator is subjected to both multiplicative white noise w1t and additive white noise w2t, the former in the linear part of the restoring force. The governing equations are given by: _ 2 X + 2
!0X + !0 f + w1tg X + X 3 = w2 t 1 Here,
X 0 = 0 E wi t
_ X 0 = 0 2
= 0 E wit1wj t2 = 2Dij t2 , t1 i; j = 1; 2 Rewritten in state space form, the governing equations become _ X1 = X2
2 2 3 = w2t , 2
!0X2 , !0 X1 + w1t , !0 X1 X1 0 = 0 X2 0 = 0 The corresponding Fokker-Planck equation is given by
@p @t @ @ 2 = , @x x2 p + @x 2
!0 x2 + !0 x1 + x2 1 1 2

_ X2
3
@ + 1 @x2 2Dp 2
4 5 6
subject to the initial condition

t!0
lim px; tjx0 = x , x0
where
x=
x1 x 2 T
x0 =
x10 x20 T
4 D = D11 !0 x2 + D22 1
and where we have assumed w1t and w2 t to be uncorrelated. The solution to the Fokker-Planck equation is the transition probability density function, which provides the probability of being at x; x at time t given that the _ probability of being at x0 ; x0 at time 0 is unity. From a computational stand_ point, it is important to relax the initial condition to avoid numerical instability in the rst few time steps Bergman and Spencer 1992 . Thus, we prescribe an initial Gaussian distribution centered at x0; x0 with dispersion controlled by the _ parameters X10 and X20 . Then, through a term-by-term weighted integration of the Fokker-Planck equation, the e ect of the initial condition is integrated out, giving
@f @t @ @ = , @x x2 f + @x

@2 2 2
!0 x2 + !0 x1 + x2 f + 1 @x2 2Df 7 1 2
with initial condition

fX0 x0 = fX0 x10 ; x20 =
x10
, X10
X10
x20
, X20
X20
8 9
where
f x; t =
is a joint density function of response and represents a standard normal distribution.
,1 ,1
px; tjx0fX0 x0 dx0
3. Solution by the Finite Element Method
We employ a standard Bubnov-Galerkin nite element method utilizing element shape functions of class C 0 . Within an element, the probability density function is interpolated according to
f x; t =
np X r=1 e Nr xfr t
10
where the Nr x, r = 1; : : :; np , are the element trial functions, np is the number of nodes in a single element, and the fre are the nodal values of the density. This expression is substituted into the weak form of eq. 7 on the element domain e , leading to the expressions for the element matrices
krs
Z
e
D @Nr @Ns , Ns @x @x
2 2
crs
g1
@Nr @x1
+ g2 @Nr @x
2
dx1dx2
11 12 13
and where
g1 x = x2
Z
e
Nr xNs xdx1 dx2
g2x
2 2 = ,2
!0 x2 , !0 x1 , !0 x3 1
Transformation of each element matrix into global coordinates and summation over the number of elements in the problem speci c mesh, Ne , leads to the global matrix evolution equation Cf_ + Kf = 0 14 subject to the initial conditions
f 0 = fX0
15
This equation is further discretized in time using the Crank-Nicholson method, which is an implicit scheme exhibiting second order accuracy and unconditional stability, thus allowing reasonably large time steps to be used throughout.
4. Application to the Du ng System
The nominal parameters of the systems to be examined herein are = 0:20 = 0:10
D22 !0
= 1:0 16
= ,1:0
= 0:40 and the initial distribution is Gaussian with mean and variance
X10
X10
= X20 = 0:0
2 = X20 = 0:5
17
In all analyses reported herein, a uniform mesh of 10,201 nodes 10,000 four-node quadrilateral elements was utilized, along with a uniform time step, t, of 1000 seconds. A typical solution run, which records 30 seconds of the system's response, required approximately 30 minutes of processing time on the Cray Y-MP located at the National Center for Supercomputing Applications at the University of Illinois at Urbana-Champaign.
System 1.
This is the baseline system, subjected only to the additive excitation. Crosssections of the stationary distribution, showing a comparison with the exact solution Caughey 1971 , are shown in Figure 1. The stationary upcrossing rate is plotted against barrier level in Figure 2. In addition, Figure 2 also shows a comparison of the centerpoint probability density function evolution, f 0; 0; t, as computed by the nite element method and a 100,000-realization Monte Carlo simulation. The evolution of the response probability density is shown at several times in Figure 3. The second moments are plotted vs. time in Figure 4, where comparison is made with Gaussian and 4th order cumulant closure results.
4
Stationary PDF (x 2 =0)

FEM Solution Exact Solution
0.01 0.008 0.006 0.004
Stationary PDF (x 1 =0)

FEM Solution Exact Solution
0.1
0.05 0.002 0 -10 0 -10
-5
0 5 Displacement x 1
10
-5
0 Velocity x 2
10
Figure 1. Cross-sections of the stationary PDF for System 1.
Stationary Upcrossing Rate 0.12

FEM Solution
x (a )
+ 1
Centerpoint PDF Evolution 0.4

FEM Solution Monte Carlo
0.1 0.3 0.08 0.06 0.04 0.1 0.02 0 -10 -5 0 Barrier Level 5 a 10 0 0 0.2
10
20 Time [seconds]
30
40
Figure 2. Stationary upcrossing rate, centerpoint PDF for System 1.
10 time = 1.257 secs
5 Velocity x 2
0.2
0.1
-5
0 -10 0 -5 0 5 Displacement x 1 10 x1 10 -10
10 0 x2
-10 -10
10
time = 2.827 secs
5 Velocity x 2
0.1
0.05
-5
0 -10 0 -5 0 5 Displacement x 1 10 x1 10 -10
10 0 x2
-10 -10
10
time = 30.47 secs 0.2
5 Velocity x 2 0.1
-5
0 10 0 x2 0 -10 -10 x1
10
-10 -10
-5
10
Figure 3. Probability density function snapshots for System 1.
12 10 8 6 4 2 0 0 10
FEM 4th Order Cumulant Closure Gaussian Closure
Second Moment E[X 2 ] 1
20 Time [seconds]
30
40
3 2 1 0 -1 -2 0
Second Moment E[X 1 X 2 ]
10
20 Time [seconds]
30
40
Second Moment E[X 2 ] 2
1.5
1
0.5 0
10
20 Time [seconds]
30
40
Figure 4. Second Moments for System 1.
System 2.
Here, the spectral density of the parametric excitation, D11, is set at 0.08, which is 20 of the additive term. Cross-sections of the stationary distribution are shown in Figure 5. The stationary upcrossing rate is plotted against barrier level in Figure 6. In addition, Figure 6 also shows a comparison of the centerpoint probability density function evolution, f 0; 0; t, as computed by the nite element method and a 100,000-realization Monte Carlo simulation. The evolution of the response probability density is shown at several times in Figure 7. The second moments are plotted vs. time in Figure 8, where comparison is made with Gaussian and 4th order cumulant neglect closure results.
Stationary PDF (x 2 =0) 0.06

FEM Solution

FEM Solution
0.05 0.04 0.03 0.02 0.01 0 -10 -5 0 5 Displacement x 1 10
0.025 0.02 0.015 0.01 0.005 0 -10 -5 0 Velocity x 2 5 10
Stationary Upcrossing Rate 0.12

FEM Solution
x (a)
+ 1

0.1 0.3 0.08 0.06 0.04 0.1 0.02 0 -10 -5 0 5 Barrier Level a 10 0 0 0.2
10
20 Time [seconds]
30
40
Figure 6. Stationary upcrossing rate, centerpoint PDF evolution for
System 2.
10 time = 1.414 secs 0.1 5 Velocity x 2 0.05
-5
0 10 0 x2 0 -10 -10 x1
10
-10 -10
-5
10
10
time = 3.77 secs 0.04
5 Velocity x 2 0.02
-5
0 10 0 x2 0 -10 -10 x1
10
-10 -10
-5
10
10
time = 31.42 secs 0.04
5 Velocity x 2 0.02
-5
0 10 0 x2 0 -10 -10 x1
10
-10 -10
-5
10
Second Moment 10 8 6 4 2 0 0
E[ X 1 ]
10
20 Time [seconds]
30
40
Second Moment 3 2 1 0 -1 -2 0
E[ X 1 X 2 ]
10
20 Time [seconds]
30
40
Second Moment 4
E[ X 2 ]
0 0
10
20 Time [seconds]
30
40

10
System 3.
Here, the spectral density of the parametric excitation, D11, is set at 0.24, which is 60 of the additive term. Cross-sections of the stationary distribution are shown in Figure 9. The stationary upcrossing rate is plotted against barrier level in Figure 10. In addition, Figure 10 also shows a comparison of the centerpoint probability density function evolution, f 0; 0; t, as computed by the nite element method and a 100,000-realization Monte Carlo simulation. The evolution of the response probability density is shown at several times in Figure 11. The second moments are plotted vs. time in Figure 12, where comparison is made with Gaussian and 4th order cumulant neglect closure results.

FEM Solution

FEM Solution
0.025 0.02 0.015 0.01 0.005 0 -10 -5 0 5 Displacement x 1 10
0.025 0.02 0.015 0.01 0.005 0 -10 -5 0 Velocity x 2 5 10
Stationary Upcrossing Rate

FEM Solution
x (a)
+ 1

0.3 0.1 0.2 0.05 0.1
0 -10
-5
0 5 Barrier Level a
10
0 0
10
20 Time [seconds]
30
40
Figure 10. Stationary upcrossing rate, centerpoint PDF evolution for
System 3.
11
10 time = 2.985 secs 0.06 5 Velocity x 2 0.04 0 0.02 0 10 0 -10 -10 -5 0 5 Displacement x 1 10 x2 -10 -10 0 x1
-5
10
10
time = 3.77 secs 0.03
5 Velocity x 2 0.02 0 0.01 0 10 0 -10 -10 -5 0 5 Displacement x 1 10 x2 -10 -10 0 x1
-5
10
10
time = 31.42 secs 0.03
5 Velocity x 2 0.02 0 0.01 0 10 0 -10 -10 -5 0 5 Displacement x 1 10 x2 -10 -10 0 x1
-5
10
12
Second Moment 10 8 6 4 2 0 0
E[ X 1 ]
10
20 Time [seconds]
30
40
Second Moment 3 2 1 0 -1 -2 0
E[ X 1 X 2 ]
10
20 Time [seconds]
30
40
Second Moment 6 5 4 3 2 1 0 0
E[ X 2 ]
10
20 Time [seconds]
30
40

13
5. Discussion of Results
Figure 3 illustrates several stages of the evolution of the joint probability density function of the response in the phase plane to the expected bimodal stationary distribution. For the parameters chosen, the peaks of the stationary density are well separated, implying that a minimal number of trajectories pass in the vicinity of the origin. This is borne out by Figure 2, a plot of the evolution of the density function at the origin which is de ned at time t = 0 by the initial distribution but decays rapidly to the stationary value. The cross sections of the stationary distribution through the origin are compared with the existing exact solution Caughey 1971 in Figure 1, where the error associated with the computational result is seen to be + negligible. The stationary upcrossing rate, x1 a, is plotted versus threshold level in Figure 2, where it is seen to assume the bimodal form of the density function and is, thus, distinctly di erent from that associated with a unimodal distribution. The evolution of the second moments of the distribution are shown in Figure 4 to be distinctly di erent from those predicted from the moment equations with closure. In particular, the solution for mean square displacement obtained by closure is not only out-of-phase with the nite element solution but also converges to an incorrect stationary value, with the 4th order cumulant closure results appearing to be somewhat better than the standard Gaussian. The error in the stationary mean square displacement is signi cant, ranging from 42.5 for 4th order cumulant neglect to 52.5 for Gaussian closure, compared to the nite element error of only 0.1. Figures 7 and 11 depict probability density evolution plots for the Du ng system with increasing levels of parametric excitation. It is apparent that, as the level of parametric excitation increases with respect to the level of external excitation, the probability density function tends to evolve with more probability mass concentrated in the vicinity of the origin. In fact, when the spectral density ratio approaches 0.60, the density function becomes strongly peaked around the origin, implying stability at that point Namachchivaya 1990 . The evolution of the joint probability density function at the origin is shown in Figures 6 and 10. Here, the stabilizing behavior of the parametric excitation is clearly seen, as the magnitude of the density function increases at stationarity with increasing levels of excitation. In all cases, the results of the Monte Carlo simulation show agreement with those of the nite element analysis. Examination of the stationary upcrossing rates, Figures 6 and 10, reveals that, as the parametric excitation increases, the bimodality decreases, the distribution more closely resembling the unimodal results in shape. Then, previously derived asymptotic results for extremal and rst passage distributions become germane and can be used for approximation. Finally, analysis of the second moments reveals that signi cant errors occur in both mean square displacement and velocity in the parametrically excited systems. The a posteriori error associated with the nite element solution can be quanti ed for the case of additive excitation. The error associated with the nite element solution for the stationary probability distribution can be de ned as e = exact , FEM . The norms of this error are
14
kek1 = max j
kek1 = max i kekF robenius =
jeij j = 0:0078 jeij j = 0:0080

= 0:0056 18
tr eT e
We have demonstrated the ability to accurately solve the Fokker-Planck equation for the Du ng oscillator, subjected to both additive and multiplicative excitations. We have demonstrated a post-processing capability allowing us to compute various derived quantities such as upcrossing rates and moments of arbitrary order. We have studied the e ect upon the stability of the system of parametric excitation acting on the linear part of the restoring force and have shown that it stabilizes an initially unstable xed point at the origin. Finally, we have demonstrated the limitations of several low order closure schemes in predicting second moments of response of Du ng systems. An alternate method through which the evolution of the TPDF can be evaluated is direct Monte Carlo simulation. Here, the sample functions can be generated directly from the stochastic di erential equations using, for example, the weakly convergent numerical schemes of Kloeden and Platen 1992 . Estimation of the density function as it evolves through time, then, involves the simultaneous processing of a large number of these sample functions into histograms, the accuracy of which, in the tail regions of the distribution, is inversely proportional to the number of samples. Recent experiments with the application of massively parallel computers to MCS Johnson, Wojtkiewicz, and Bergman 1994 show the distinct advantages of FEM in problems of small dimension. Monte Carlo simulation, however, may be the only viable approach to systems of large dimension for the foreseeable future.
6. Conclusions
Acknowledgments
The authors are grateful to the National Science Foundation for its support through contracts ECS 9224828, CEE 92000 4N, and MSS-94000 1N, the latter two through the National Center for Supercomputing Applications at the University of Illinois at Urbana Champaign.
References
L. A. Bergman and B. F. Spencer, Jr., Numerical Solution of the Transient Fokker-Planck Equation: The Movie, Proceedings of the 6th ASCE Specialty Conference on Probabilistic Mechanics and Structural and Geotechnical Safety, Denver, Colorado, July 8 10, 1992, pp. 519 522. L. A. Bergman, S. F. Wojtkiewicz, and B. F. Spencer, Jr., Numerical Solution of the Transient Fokker-Planck Equation: Animations, Presented at the International Symposium on Nonlinear Dynamics and Stochastic Mechanics, Waterloo, Ontario, August 29 September 1, 1993. T. K. Caughey, Derivation and Application of the Fokker-Planck Equation to Discrete Nonlinear Dynamic Systems Subjected to White Noise Excitation, Journal of the Acoustical Society of America 35 1963, 1683 1692. 15
Nonlinear Dynamics and Stochastic Mechanics, W. Langford, W. Kliemann, and N. Sri Namachchivaya, editors, Fields Institute Communications, vol. 9, AmericanMathematicalSociety,Providence,1996, 23-37. Presented at the lnternational Symposium on Nonlinear Dynamics and Stochastic Mechanics, Fields Institute for Research in Mathematical Sciences, Waterloo, Ontario, Canada, Aug 28 -Sept1, 1993. T. K. Caughey, Nonlinear Theory of Random Vibrations, Advances in Applied Mechanics Vol. 11 Chia-Shun Yih, ed., 1971, pp. 209 253. E. A. Johnson, S. F. Wojtkiewicz, and L. A. Bergman, Some Experiments with Massively Parallel Computation for Monte Carlo Simulation of Stochastic Dynamical Systems, Proceedings of the Second InternationalConference on Computational StochasticMechanics, Athens, Greece, June 13 15, 1994, in press. P. E. Kloeden and E. Platen, Numerical Solution of Stochastic Di erential Equations, SpringerVerlag, Berlin, 1992. R. S. Langley, A Finite Element Method for the Statistics of Non-Linear Random Vibration, Journal of Sound and Vibration 101 1985, 41 54. H. P. Langtangen, A General Numerical Solution Method for Fokker-Planck Equations with Applications to Structural Reliability, Probabilistic Engineering Mechanics 6 1991, 33 48. N. Sri Namachchivaya, Stochastic Bifurcation, Applied Mathematics and Computation 38 1990, Elsevier Science Publishing Co., Inc., 101 159. B. F. Spencer, Jr. and L. A. Bergman, On the Numerical Solution of the Fokker-Planck Equation for Nonlinear Stochastic Systems, Nonlinear Dynamics 4 1993, Kluwer Academic Publishers., 357 372.
16

Fokker Planck1

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Fokker Planck1

Uploaded by

Copyright:

Available Formats

Nonlinear Dynamics and Stochastic Mechanics, W. Langford, W. Kliemann, and N. Sri Namachchivaya, editors, Fields Institute Communications, vol.

4 5 6

subject to the initial condition

lim px; tjx0 = x , x0 

with initial condition

is a joint density function of response and   represents a standard normal distribution.

px; tjx0fX0 x0 dx0

3. Solution by the Finite Element Method

11 12 13

Nr xNs xdx1 dx2

4. Application to the Du ng System

Stationary PDF (x 2 =0)

0.01 0.008 0.006 0.004

Stationary PDF (x 1 =0)

0.05 0.002 0 -10 0 -10

Figure 1. Cross-sections of the stationary PDF for System 1.

Stationary Upcrossing Rate 0.12

Centerpoint PDF Evolution 0.4

Figure 2. Stationary upcrossing rate, centerpoint PDF for System 1.

0 -10 0 -5 0 5 Displacement x 1 10 x1 10 -10

time = 2.827 secs

0 -10 0 -5 0 5 Displacement x 1 10 x1 10 -10

time = 30.47 secs 0.2

Figure 3. Probability density function snapshots for System 1.

Second Moment E[X 2 ] 1

Second Moment E[X 1 X 2 ]

FEM 4th Order Cumulant Closure Gaussian Closure

Second Moment E[X 2 ] 2

Figure 4. Second Moments for System 1.

Stationary PDF (x 2 =0) 0.06

Stationary PDF (x 1 =0) 0.03

0.05 0.04 0.03 0.02 0.01 0 -10 -5 0 5 Displacement x 1 10

0.025 0.02 0.015 0.01 0.005 0 -10 -5 0 Velocity x 2 5 10

Figure 5. Cross-sections of the stationary PDF for System 2.

Stationary Upcrossing Rate 0.12

Centerpoint PDF Evolution 0.4

Figure 6. Stationary upcrossing rate, centerpoint PDF evolution for

time = 3.77 secs 0.04

time = 31.42 secs 0.04

Figure 7. Probability density function snapshots for System 2.

FEM 4th Order Cumulant Closure Gaussian Closure

FEM 4th Order Cumulant Closure Gaussian Closure

Figure 8. Second Moments for System 2.

Stationary PDF (x 2 =0) 0.03

Stationary PDF (x 1 =0) 0.03

0.025 0.02 0.015 0.01 0.005 0 -10 -5 0 5 Displacement x 1 10

0.025 0.02 0.015 0.01 0.005 0 -10 -5 0 Velocity x 2 5 10

Figure 9. Cross-sections of the stationary PDF for System 3.

Stationary Upcrossing Rate

Centerpoint PDF Evolution 0.4

0.3 0.1 0.2 0.05 0.1

Figure 10. Stationary upcrossing rate, centerpoint PDF evolution for

time = 3.77 secs 0.03

5 Velocity x 2 0.02 0 0.01 0 10 0 -10 -10 -5 0 5 Displacement x 1 10 x2 -10 -10 0 x1

time = 31.42 secs 0.03

5 Velocity x 2 0.02 0 0.01 0 10 0 -10 -10 -5 0 5 Displacement x 1 10 x2 -10 -10 0 x1

Figure 11. Probability density function snapshots for System 3.

FEM 4th Order Cumulant Closure Gaussian Closure

FEM 4th Order Cumulant Closure Gaussian Closure

Figure 12. Second Moments for System 3.

kek1 = max i kekF robenius =

jeij j = 0:0078 jeij j = 0:0080

You might also like

4 5 6

lim px; tjx0 = x , x0

is a joint density function of response and represents a standard normal distribution.

px; tjx0fX0 x0 dx0

11 12 13

Nr xNs xdx1 dx2