Fractional Fokker-Planck Equations for Subdiffusion with Space-and-Time-Dependent

Forces.

B.I. Henry∗
Department of Applied Mathematics, School of Mathematics and Statistics,
University of New South Wales, Sydney NSW 2052, Australia.

T.A.M. Langlands†
arXiv:1004.4053v1 [cond-mat.stat-mech] 23 Apr 2010

Department of Mathematics and Computing, University of Southern Queensland, Toowoomba Queensland 4350, Australia.

P. Straka
School of Mathematics and Statistics, University of New South Wales, Sydney NSW 2052, Australia.
(Dated: April 26, 2010)
We have derived a fractional Fokker-Planck equation for subdiffusion in a general space-and-
time-dependent force field from power law waiting time continuous time random walks biased by
Boltzmann weights. The governing equation is derived from a generalized master equation and is
shown to be equivalent to a subordinated stochastic Langevin equation.

PACS numbers: 05.40.Fb,02.50.Ey,05.10.Gg

Over the past few decades there has been an enormous a time-dependent force field F (t). The modified frac-
growth in the numbers of papers devoted to experimental tional Fokker-Planck equation, Eq.(3), was also derived
and theoretical aspects of anomalous diffusion [1, 2]. The for subdiffusion in dichotomously alternating force fields
landmark review by Metzler and Klafter in 2000 [1] has [8, 9], F (x)ξ(t) with ξ(t) = ±1, but a fractional Fokker-
been particularly influential, promoting the description Planck equation to model subdiffusion in general space-
of anomalous diffusion within the framework of continu- and-time-dependent force fields F (x, t) has remained elu-
ous time random walks (CTRWs) and fractional calcu- sive [8–11]. On the other hand, a subordinated stochastic
lus. There are now numerous applications utilizing this Langevin equation has been proposed for modelling sub-
approach in physics, chemistry, biology and finance [2]. diffusion in space-and-time-dependent force fields [10].
A central theoretical result in this research was the More recently [12], in the case of time-dependent forces,
derivation [3, 4] of a fractional Fokker-Planck (Smolu- the moments of the stochastic process defined by this
chowksi) equation [5] stochastic Langevin equation were shown to coincide with
  the moments of the modified fractional Fokker-Planck
∂P 1−α ∂2 1 ∂ equation, Eq.(3).
= 0 Dt κα 2 − F (x) P (x, t) (1)
∂t ∂x ηα ∂x There have been numerous papers on fractional
Fokker-Planck equations in recent years relying on ad
to describe the evolution of the probability density
hoc or phenomenological models [8, 11, 13, 14].
function P (x, t) for subdiffusion in an external space-
In this letter we derive the fractional Fokker-Planck
dependent force field F (x). In this equation,
equation,
Z t  
1 ∂ Y (t′ )
D 1−α
Y (t) = dt′ (2) ∂P ∂2 1 ∂
0 t
Γ(α) ∂t 0 (t − t′ )1−α = κα 2 − F (x, t) 0 Dt1−α P (x, t), (4)
∂t ∂x ηα ∂x
is the Riemann-Liouville fractional derivative, κα is a from power law waiting time CTRWs, using a generalized
fractional diffusion coefficient, ηα = (βκα )−1 is a frac- master equation, for subdiffusion in a space-and-time-
tional friction coefficient, and β is the inverse tem- dependent force field F (x, t). This fractional Fokker-
perature kB T . The fractional Fokker-Planck equation, Planck equation is shown to be formally equivalent to
Eq.(1), was derived from the continuous time random the subordinated stochastic Langevin equation in [10] for
walk model of Montroll and Weiss [6], with power law space-and-time-dependent forces. We also show that the
waiting times [3, 4]. original fractional Fokker-Planck equation, Eq.(1), gen-
More recently a modified fractional Fokker-Planck eralized by replacing F (x) with F (x, t), can be recovered
equation, from power law waiting time CTRWs in an ad-hoc gen-
  eralization of the CTRW particle balance equation if the
∂P ∂2 1 ∂
= κα 2 − F (t) 0 Dt1−α P (x, t), (3) diffusing particles respond to the force field at the start
∂t ∂x ηα ∂x
of the waiting time prior to jumping. These derivations,
was derived from power law waiting time CTRWs, us- and further extensions to include reactions, are described
ing a generalized master equation [7], for subdiffusion in in greater detail, for chemotactic forcing, in a related
 2

publication [15]. In the CTRW model, subdiffusion originates from a

Our starting point is the generalized master equation heavy-tailed waiting-time density with long-time be-
approach developed in [7, 16]. This approach utilizes haviour [1]
two balance conditions. The balance equation for the  −1−α
concentration of particles, ni (t), at the site i and time t κ t
ψ(t) ∼ (13)
is τ τ
dni (t) where α is the anomalous exponent, τ is the characteristic
= Ji+ (t) − Ji− (t), (5)
dt waiting-time, and κ is a dimensionless constant. Using a
Tauberian (Abelian) theorem for small s [17]
where Ji± are the gain (+) and loss (-) fluxes at the site
i. The second balance equation is a conservation equa- b
ψ(s) s1−α
tion for the arriving flux of particles at the point i. In ∼ Aα α (14)
b
Φ(s) τ
general, to allow for biased CTRWs in a space-and-time-
dependent force field, we write where Aα = κα(1−α)α
. In the special case of a Mittag-
Ji+ (t) = pr (xi−1 , t)Ji−1
− −
(t) + pl (xi+1 , t)Ji+1 (t), (6) Leffler waiting time density [18] the ratio, Eq.(14), is ex-
act and Aα = 1. The loss flux can now be obtained
where pr (x, t) and pl (x, t) are the probabilities of jumping by using the ratio, Eq(14), in Eq.(11) and inverting the
from x to the adjacent grid point, to the right and left Laplace transform. Noting that the Laplace Transform
directions respectively. The two balance equations can of a Riemann-Liouville fractional derivative of order α,
be combined to yield where 0 < α ≤ 1, is given by [19]
 α   α−1 
dni (t) − −
d f (t) αb d f (t) 
= pr (xi−1 , t)Ji−1 (t) + pl (xi+1 , t)Ji+1 (t) − Ji− (t). L (s) = s f (s) − (15)
dt dtα dtα−1 t=0
(7)
For CTRWs with a waiting time probability density func- this yields
tion ψ(t) the loss flux at site i is from those particles that
were originally at i at t = 0 and wait until time t to leave, Aα d1−α ni (t)
Ji− (t) = , (16)
and those particles that arrived at an earlier time t and
′
τ α dt1−α
wait until time t to leave, hence [16] where we have assumed that the last term in Eq. (15) is
zero. Using this result in Eq.(7) yields,
Zt
Ji− (t) = ψ(t)ni (0) +
′ ′
ψ(t − t )Ji+ (t ) dt .
′
(8) 
dni (t) Aα d1−α ni−1 (t)
= α pr (xi−1 , t)
0 dt τ dt1−α

We can combine Eq. (5) and Eq. (8) to obtain d1−α ni+1 (t) d1−α ni (t)
+pl (xi+1 , t) − .
(17)
dt1−α dt1−α
Zt " ′
#
′ ′ dni (t ) ′
The jump probabilities are biased by the external
Ji− (t) = ψ(t)ni (0) + ψ(t − t ) Ji− (t ) + dt ,
dt space-and-time-dependent force. Here we consider (near
0
(9) thermodynamic equilibrium) Boltzmann weights with
and then exp(−βV (xi+1 , t))
h i pr (xi , t) = C , (18)
b
Jbi− (s) = ψ(s)n b b− exp(−βV (xi , t))
i (0) + ψ(s) Ji (s) + sb
ni (s) − ni (0) ,
(10) and
where the hat denotes a Laplace transform with respect
to time and s is the Laplace transform variable. This exp(−βV (xi−1 , t))
pl (xi , t) = C . (19)
simplifies further to exp(−βV (xi , t))

b
ψ(s) The jump probabilities are determined at the end of the
Jbi− (s) = n
bi (s), (11) waiting time, when the particle must jump, so that
b
Φ(s)
pr (xi , t) + pl (xi , t) = 1, (20)
b
where Φ(s) is the Laplace transform of the survival prob-
ability which defines C, and then
Z ∞
e−βV (xi−1 ,t) − e−βV (xi+1 ,t)
Φ(t) = ψ(t′ ) dt′ . (12) pl (xi , t) − pr (xi , t) = . (21)
t e−βV (xi−1 ,t) + e−βV (xi+1 ,t)
 3

The spatial continuum limit of Eq.(17) can now be It is asumed that Bt and Ut are independent stochastic
obtained by setting xi = x and xi±1 = x ± ∆x and processes and the initial condition is Y0 = Z0 = 0. The
carrying out Taylor series expansions in x. Retaining stochastic process representing subdiffusion in a space-
terms to order (∆x)2 and using Eq.(20) yields and-time-dependent force field is postulated to be given
 by [10] Xt = Y (St ) where for t ≥ 0, St is the random
∂n(x, t) Aα ∂   time the process Ut exceeds t.
= α ∆x (pl (x, t) − pr (x, t)) 0 Dt1−α n(x, t)
∂t τ ∂x The stochastic differential equation, Eq.(26), belongs

∆x2 ∂ 2 1−α to the general class of stochastic processes driven by Lévy
+ D n(x, t) . (22) noise [20, eq.(6.12)]. The infinitesimal generator for the
2 ∂x2 0 t
process Eq.(26) is then given by [20, eq.(6.42)]
The Taylor series expansion of Eq.(21) yields
F (y, z) ∂ ∂2
∂V (x, t) Af (y, z) = f (y, z) + κ 2 f (y, z)
pl (x, t) − pr (x, t) ≈ β∆x + O(∆x3 ), (23) η ∂y ∂y
∂x Z ∞
α
+ [f (y, z + z ′ ) − f (y, z)] z ′−1−α dz ′ .
and then 0 Γ(1 − α)
  (27)
∂n(x, t) Aα β∆x2 ∂ ∂V (x, t) 1−α The Fokker-Planck evolution equation for the probability
= D n(x, t)
∂t τα ∂x ∂x 0 t density qt (y, z) of the process (Yt , Zt ) is given by
Aα ∆x2 ∂ 2
+ D1−α n(x, t) + O(∆x4 ).(24) ∂
2τ α ∂x2 0 t qt (y, z) = A† qt (y, z) (28)
2
∂t
In the limit ∆ → 0 and τ → 0, with κα = Aα2τ∆x α , and
ηα = (2βκα )−1 we recover the fractional Fokker-Planck where
equation, Eq.(4), for subdiffusion in an external space-  
† ∂2 ∂ F (y, z)
and-time-dependent force field A f (y, z) = κ 2 f (y, z) − f (y, z)
∂y ∂y η (29)
∂V (x, t) − 0 Dzα f (y, z)
F (x, t) = − . (25)
∂x
defines the operator adjoint to A.
The fractional Fokker-Planck equation, Eq.(4), can
Now we relate the densities pt (x) and qt (y, z) of the
also be derived from the subordinated stochastic
stochastic processes Xt and (Yt , Zt ) respectively. We
Langevin equation motivated by physical arguments [10]
write ω for a particular (random) path of the lat-
to model subdiffusion in a space-and-time-dependent
ter process, and note that the coordinates at time t,
force field. This representation can be formulated as a
(Yt (ω), Zt (ω)) and Xt (ω), are functions of ω. For a fixed
system of stochastic equations,
interval I we define the indicator function
      (
dYt F (Yt , Zt )η −1 (2κ)1/2 dBt
= dt+ , (26) 1 if x ∈ I
dZt 0 dUt δI (x) = (30)
0 if x ∈
/I
where Bt is a one-dimensional Brownian motion and Ut
is a α-stable Lévy subordinator in [0, ∞), 0 < α < 1. together with the auxiliary function


′ δI (Yt′ (ω)) if Zt′ − (ω) ≤ t ≤ Zt′ − (ω) + ∆z
H(t , ω, ∆z) = (31)
0 otherwise.

With the above notation we can write (Y, Z) before it exits the set R × [0, t], we have
Z X
pt (x) dx = hδI (Xt (ω))i (32) δI (Xt (ω)) = H (t′ , ω, ∆Zt′ (ω)) , (33)
I t′ >0

where the angle brackets represent an ensemble average since all summands equal zero except for t′ = St , in which
over all paths ω. Given that Xt can be interpreted as the case Yt′ (ω) = Xt (ω). The jumps ∆z = ∆Zt′ (ω) are
Y coordinate of the last position of the Markov process a Poisson point process on (0, ∞) whose characteristic
 4

α
measure has the density Γ(1−α) ∆z −1−α . We can now where the jump probabilities are evaluated at the start
combine Eqs.(32), (33) and use the compensation formula of the waiting time prior to jumping. Note that this
in [21, XII (1.10)], to write equation can be derived from the Montroll-Weiss CTRW
Z Z ∞ Z ∞  formalism (see e.g., [2]) if and only if the jumping prob-
α∆z −1−α abilities are independent of time. Our inclusion of time
pt (x)dx = H(t′ , ω, ∆z) d∆z dt′ .
I 0 0 Γ(1 − α) dependence is an ad-hoc generalization for time depen-
(34) dent jumps. Using Laplace transform methods as above
After integrating over ∆z and using Eq.(31) the right then leads to the discrete space evolution equation
hand side simplifies further to
Z ∞  dni Aα d1−α
(t − Zt′ (ω))−α = α 1−α {−ni (t) + pr (xi−1 , t)ni−1 (t)
′
dt δI (Yt′ (ω))δ[0,t] (Zt′ (ω)) . (35) dt τ dt
0 Γ(1 − α) +pl (xi+1 , t)ni+1 (t)} . (42)

The ensemble average is evaluated using the probability It follows from Eqs. (16) and (17) that this evolution
density qt (y, z) so that equation is the evolution equation for the loss flux in
Z Z ∞ Z Z t
the generalized mater equation approach. After taking
′ (t − z)−α the spatial continuum limit of Eq.(42) with Boltzmann
pt (x)dx = dt dy dz qt′ (y, z) ,
I 0 I 0 Γ(1 − α) weighted jumping probabilities we have
and thus   
∂n ∂ 1−α ∂ 2 n(x, t) 1 ∂ ∂V (x, t)
Z = 1−α κα + n(x, t) .
∞ ∂t ∂t ∂x2 ηα α ∂x ∂x
pt (x) = dt′ 0 It1−α qt′ (x, t), (36) (43)
0
This provides an interpretation of the fractional Fokker-
where 0 It1−α denotes the Liouville fractional integral of Planck equation
order 1 − α acting on t. It also follows that  
∂P 1−α ∂2 1 ∂
Z ∞ = 0 Dt κα 2 − F (x, t) P (x, t) (44)
1−α ∂t ∂x ηα ∂x
0 Dt pt (x) = qt′ (x, t)dt′ , (37)
0
Z ∞ as an equation for the loss flux in subdiffusion in a space-
∂ α ′
pt (x) = 0 Dt qt′ (x, t)dt . (38) and-time-dependent force field
∂t 0 It is straightforward [15] to obtain numerical solutions
We now solve Eqs.(28), (29) for 0 Dtα qt′ (x, t) and subsi- of the discrete space evolution equations for subdiffusion
tute this into Eq.(38) to obtain in space-and-time-dependent force fields, Eqs.(17), (42)
using an implicit time stepping method with the frac-
Z ∞ tional derivatives approximated using the L1 scheme [22].
∂ ∂2 1 ∂
pt (x) = κ 2 qt′ (x, t) − (F (x, t)qt′ (x, t)) The CTRW models described here are also easy to sim-
∂t 0 ∂x η ∂x
 ulate using Monte Carlo methods [15].
∂
− ′ qt′ (x, t) dt′ (39) This work was supported by the Australian Research
∂t Council. We are grateful to Eli Barkai for his comments
and finally, using Eq.(37), on our draft manuscript.

∂ ∂2 ∂ F (x, t) 1−α
pt (x) = κ 2 0 Dt1−α pt (x) − D pt (x)
∂t ∂x ∂x η 0 t
(40) ∗
Electronic address: B.Henry@unsw.edu.au
where we have used q∞ (x, t) = 0 and q0 (x, t) = δ(0,0) (x, t) †
Electronic address: t.langlands@usq.edu.au
and assumed t > 0. Equation (40) recovers the fractional [1] R. Metzler and J. Klafter, Phys. Rep. 339, 1 (2000).
Fokker-Planck equation for space-and-time-dependent [2] B. I. Henry, T. A.M. Langlands, and P. Straka, in
forces, Eq.(4). Complex Physical, Biophysical and Econophysical Sys-
A different fractional Fokker-Planck equation can be tems: World Scientific Lecture Notes in Complex Sys-
obtained from the following generalization of the CTRW tems, edited by R. L. Dewar and F. Detering (World
Scientific, Singapore, 2010), vol. 9, pp. 37–90.
particle balance equation [3] R. Metzler, E. Barkai, and J. Klafter, Europhys Letts 46,
431 (1999).
Zt n
′ ′ [4] E. Barkai, R. Metzler, and J. Klafter, Phys. Rev. E 61,
ni (t) = ni (0)Φ(t) + pr (xi−1 , t )ni−1 (t ) 132 (2000).
0 [5] R. Metzler, E. Barkai, and J. Klafter, Phys. Rev. Letts.
o′ ′ ′ ′ 82, 3563 (1999).
+pl (xi+1 , t )ni+1 (t ) ψ(t − t ) dt , (41) [6] E. Montroll and G. Weiss, J. Math. Phys. 6, 167 (1965).
 5

[7] I. M. Sokolov and J. Klafter, Phys. Rev. Lett. 97, 140602 [17] G. Margolin, Physica A 334, 46 (2004).
(2006). [18] E. Scalas, R. Gorenflo, F. Mainardi, and M. Raberto,
[8] E. Heinsalu, M. Patriarca, I. Goychuk, and P. Hänggi, Fractals 11, 281 (2003).
Phys. Rev. Lett. 99, 120602 (2007). [19] I. Podlubny, Fractional differential equations, vol. 198
[9] E. Heinsalu, M. Patriarca, I. Goychuk, and P. Hänggi, of Mathematics in Science and Engineering (Academic
Phys. Rev. E 79, 041137 (2009). Press, New York and London, 1999).
[10] A. Weron, M. Magdziarz, and K. Weron, Phys. Rev. E [20] D. Applebaum, Levy Processes and Stochastic Calculus,
77, 036704 (2008). 2nd Edition, vol. 116 of Cambridge Studies in Advanced
[11] Y.-M. Kang and Y.-L . Jiang, J. Math. Phys. 51, 023301 Mathematics (Cambridge University Press, Cambridge,
(2010). 2009).
[12] Y.-M. Kang and Y.-L . Jiang, Phys. Rev. Letts. 101, [21] D. Revuz and M. Yor, Continuous Martingales and Brow-
210601 (2008). nian Motion, 3rd Edition, vol. 293 of A Series of Compre-
[13] Y. P. Kalmykov, W. T. Coffey, and S. V. Titov, Phys. hensive Studies in Mathematics (Springer, Berlin, 1999).
Rev. E 74, 011105 (2006). [22] K. Oldham and J. Spanier, The Fractional Calculus:
[14] M. M. Mousa and A. Kaltayev, Z. Naturforsch. 64a, 788 Theory and Applications of Differentiation and Integra-
(2009). tion to Arbitrary Order, vol. 111 of Mathematics in Sci-
[15] T. A.M. Langlands and B. I. Henry, ence and Engineering (Academic Press, New York and
arXiv:submit/0000822 pp. 1–25 (2010). London, 1974).
[16] A.V. Chechkin, R. Gorenflo, and I.M. Sokolov, J. Phys.
A: Math. Gen. 38, L679 (2005).