The Immersed Interface Method (IIM) is derived to solve the corresponding Fokker-Planck equation of Brownian motion with pure dry friction, which is one of the simplest models of piecewise-smooth stochastic systems. The IIM is capable of treating a discontinuity in the drift of Fokker-Planck equation and it is readily extended to the dry and viscous friction model. Analytic results of the considered model are used to confirm the effectiveness and design accuracy of the scheme.
In recent years, piecewise-smooth stochastic systems (governed by piecewise-smooth stochastic differential equations) are usually used to describe biological and physical systems. Although for some simple piecewise-linear stochastic differential equations, analytical solutions of the transition probability distribution can be obtained [
In this paper, we attempt to solve numerically a Fokker-Planck equation with discontinuous drift, which results from a so-called Brownian motion with pure dry friction [
v ′ ( t ) = − σ ( v ( t ) ) + ξ ( t ) . (1.1)
Here σ ( v ( t ) ) denotes the sign of the velocity v ( t ) , representing the dry friction force. ξ ( t ) is the Gaussian white noise with zero mean and delta correlation 〈 ξ ( t ) ξ ( t ′ ) 〉 = 2 D δ ( t − t ′ ) with D > 0 . The notation 〈 · 〉 stands for the average overall possible realizations of the noise, and δ is the Dirac delta function. The transition probability distribution p ( v , t | v 0 , 0 ) of (1.1) satisfies the following Fokker-Planck equation [
∂ p ( v , t | v 0 , 0 ) ∂ t = ∂ ( σ ( v ) p ( v , t | v 0 , 0 ) ) ∂ v + D ∂ 2 p ( v , t | v 0 , 0 ) ∂ v 2 . (1.2)
The corresponding initial condition is p ( v , 0 | v 0 , 0 ) = δ ( v − v 0 ) if v ( 0 ) = v 0 for (1.1).
Since Equation (1.2) has a discontinuous drift σ ( v ) , we must deal with it carefully. The IIM is a sharp interface method which can accurately capture discontinuities in the solution and the flux. This method has been used for many problems, such as elliptic interface problems [
The rest of this paper is organized as follows. In Section 2, we derive the IIM for the Fokker-Planck Equation (1.2). The numerical results are compared with the analytical solutions in Section 3. In addition, the accuracy of the scheme is also obtained. Finally, conclusions are made in Section 4.
We set D = 1 for convenience. At the discontinuous point v = 0 , we have the matching condition for the solution,
p + = p − (2.1)
where + and − stand for the limiting values from the right- and left-hand sides of v = 0 . Integrating (1.2) across the discontinuity, we find
p v + = p v − − 2 p − (2.2)
and then
p v − = p v + + 2 p + (2.3)
by replacing p − with p + in (2.2).
It follows from (2.1) that p t + = p t − , that is
( σ ( v ) p ) v + + p v v + = ( σ ( v ) p ) v − + p v v −
according to Equation (1.2). Then using the relations (2.1)-(2.3) we have
p v v + = p v v − − 2 p v − + 2 p − (2.4)
and
p v v − = p v v + + 2 p v + + 2 p + . (2.5)
For the numerical scheme, we have first to truncate the computational domain ( − ∞ , + ∞ ) to a finite domain. Without loss of generality, let us assume the finite domain to be [ − L v , L v ] , where L v is a positive constant. Then we assume the probability vanishes at the boundary, i.e.,
p ( ± L v , t | v 0 , 0 ) = 0 . (2.6)
A uniform grid with step h = 2 L v / M is chosen here, where M is a positive constant.
Therefore, the grid points can be expressed as v i = − L v + ( i − 1 ) h , i = 1 , 2 , ⋅ ⋅ ⋅ , M + 1 with the discontinuous point being between v j and v j + 1 , v j ≤ 0 < v j + 1 .
We hope to develop finite difference scheme of the form
p i n + 1 − p i n τ = γ i , 1 p i − 1 n + γ i , 2 p i n + γ i , 3 p i + 1 n + C i n , i = 2 , 3 , ⋅ ⋅ ⋅ , M , (2.7)
where τ is the time-step size. This means that we need to determine the coefficients γ and the correction term C i n so that
γ i , 1 p i − 1 n + γ i , 2 p i n + γ i , 3 p i + 1 n + C i n ≈ { ∂ ( σ ( v ) p ) ∂ v + ∂ 2 p ∂ v 2 } v = v i . (2.8)
At a regular grid point v i , i ≠ j , j + 1 , the coefficients γ in the explicit difference scheme (2.7) are obtained by the standard approximation as follows
γ i , 1 = 1 h 2 − σ ( v i − 1 / 2 ) 2 h , γ i , 2 = σ ( v i + 1 / 2 ) − σ ( v i − 1 / 2 ) 2 h − 2 h 2 , γ i , 3 = 1 h 2 + σ ( v i + 1 / 2 ) 2 h , (2.9)
and the correction term C i n = 0.
At the irregular grid point v j , we expand p j − 1 n , p j n and p j + 1 n in Taylor series at the discontinuous point v = 0 to obtain
p j − 1 n = p − + v j − 1 p v − + 1 2 v j − 1 2 p v v − + O ( h 3 ) , (2.10)
p j n = p − + v j p v − + 1 2 v j 2 p v v − + O ( h 3 ) , (2.11)
p j + 1 n = p + + v j + 1 p v + + 1 2 v j + 1 2 p v v + + O ( h 3 ) . (2.12)
For Equation (2.12), using (2.1), (2.2) and (2.4), we have
p j + 1 n = ( 1 − 2 v j + 1 + v j + 1 2 ) p − + ( v j + 1 − v j + 1 2 ) p v − + 1 2 v j + 1 2 p v v − + O ( h 3 ) . (2.13)
Furthermore, substituting (2.10), (2.11) and (2.13) into (2.8) we have
γ j , 1 ( p − + v j − 1 p v − + 1 2 v j − 1 2 p v v − ) + γ j , 2 ( p − + v j p v − + 1 2 v j 2 p v v − ) + γ j , 3 ( ( 1 − 2 v j + 1 + v j + 1 2 ) p − + ( v j + 1 − v j + 1 2 ) p v − + 1 2 v j + 1 2 p v v − ) + C j n + O ( h ) = { ∂ ( σ ( v ) p ) ∂ v + ∂ 2 p ∂ v 2 } v = 0 − + O ( h ) . (2.14)
Then by arranging terms we obtain
p − ( γ j , 1 + γ j , 2 + ( 1 − 2 v j + 1 + v j + 1 2 ) γ j , 3 ) + p v − ( v j − 1 γ j , 1 + v j γ j , 2 + ( v j + 1 − v j + 1 2 ) γ j , 3 ) + p v v − ( 1 2 v j − 1 2 γ j , 1 + 1 2 v j 2 γ j , 2 + 1 2 v j + 1 2 γ j , 3 ) + C j n + O ( h ) = − p v − + p v v − + O ( h ) . (2.15)
Comparing both sides of (2.15), one obtains three equations for γ j , 1 , γ j , 2 and γ j , 3 as follows
γ j , 1 + γ j , 2 + ( 1 − 2 v j + 1 + v j + 1 2 ) γ j , 3 = 0 , (2.16)
v j − 1 γ j , 1 + v j γ j , 2 + ( v j + 1 − v j + 1 2 ) γ j , 3 = − 1 , (2.17)
v j − 1 2 γ j , 1 + v j 2 γ j , 2 + v j + 1 2 γ j , 3 = 2 , (2.18)
and the correction term C j n = 0.
Therefore, one can solve (2.16) - (2.18) to attain the coefficients of Equation (2.7) for i = j .
In a similar way, we can compute the coefficients at the irregular grid point v j + 1 from the equations
( 1 + 2 v j + v j 2 ) γ j + 1 , 1 + γ j + 1 , 2 + γ j + 1 , 3 = 0 , (2.19)
( v j + v j 2 ) γ j + 1 , 1 + v j + 1 γ j + 1 , 2 + v j + 2 γ j + 1 , 3 = 1 , (2.20)
v j 2 γ j + 1 , 1 + v j + 1 2 γ j + 1 , 2 + v j + 2 2 γ j + 1 , 3 = 2 , (2.21)
and the correction term C j + 1 n = 0.
For the Fokker-Planck Equation (1.2), using spectral decomposition method, one can get the transition probability distribution in closed analytic form [
p ( v , t | v 0 , 0 ) = 1 D p ^ ( 1 D v , 1 D t | 1 D v 0 , 0 ) , (3.1)
where
p ^ ( x , τ | x ′ , 0 ) = e − τ / 4 2 π τ e − ( | x | − | x ′ | ) / 2 e − ( x − x ′ ) 2 / ( 4 τ ) + e − | x | 4 [ 1 + erf ( τ − ( | x | + | x ′ | ) 2 τ ) ] (3.2)
is the transition probability distribution in non-dimensional units and
erf ( z ) = 2 π ∫ 0 z e − t 2 d t (3.3)
is the error function. In addition, when t → ∞ the Fokker-Planck Equation (1.2) admits a steady stationary state
p ( v ) = 1 2 D e − | v | / D . (3.4)
Let D = 1 , v 0 = 2 and the computing interval be [ − 10 , 10 ] . We choose the space-step h = 0.01 and the time-step τ = 5 × 10 − 5 . For simplicity, we take the analytic distribution (3.1) at time t = 0.01 as the initial condition for computing.
To see the accuracy of the scheme numerically, we consider the L 2 and L ∞ errors between the numerical solutions and the exact solutions defined by
‖ E r r o r ‖ L 2 = h ∑ i = 1 M ( p i − p ¯ i ) 2 , (3.5)
‖ E r r o r ‖ L ∞ = max 1 ≤ i ≤ M | p i − p ¯ i | , (3.6)
where p i is the numerical solution and p ¯ i is the exact solution. Then we calculate the order of accuracy. A small time-step τ = 1 × 10 − 5 and v 0 = 2 are chosen and the problem is recalculated from time t = 0.01 to t = 1 . As illustrated in
h | L 2 error | Order | L ∞ error | Order |
---|---|---|---|---|
0.04 | 6 .9701 × 10 − 4 | 8 .9056 × 10 − 4 | ||
0.02 | 1 .7441 × 10 − 4 | 2.00 | 2 .2432 × 10 − 4 | 1.99 |
0.01 | 4 .2682 × 10 − 5 | 2.03 | 5 .5016 × 10 − 5 | 2.03 |
0.005 | 9 .7086 × 10 − 6 | 2.14 | 1 .2371 × 10 − 5 | 2.15 |
We have used the IIM to solve a Fokker-Planck equation with discontinuous drift in this paper. The numerical results show that the developed scheme is effective and has second order of accuracy. Moreover, the scheme can be readily extended to other dry friction models and the numerical results obtained are important references to see whether the dry friction effect exists in engineering applications.
This work was supported by the National Natural Science Foundation of China (Grant Nos. 11571366 and 11601517).
Zhang, B.Y., Chen, Y.M. and Song, S.H. (2017) Immersed Interface Method for Fokker-Planck Equation with Discontinuous Drift. Journal of Applied Mathematics and Physics, 5, 1613-1619. https://doi.org/10.4236/jamp.2017.59133