Pramg.na, Vol. 15, No. 6, June 1981, pp. 437-455. © Print~ in India.

Two-state random walk model of diffusion. 2. Oscillatory diffusion

V B A L A K R I S H N A N * and G V E N K A T A R A M A N
*Department of Physics, Indian Institute of Technology, Madras 600 036, India
Reactor Research Centre, Kalpakkam 603 102, Tamil Nadu, India

MS received 8 November 1980; revised 12 May 1981

Abstract. Continuing our study of interrupted diffusion, we consider the problem of

a particle executing a random walk interspersed with localized oscillations during its
halts (e.g., at lattice sites). Earlier approaches proceed via approximation schemes
for the solution of the Fokker-Planck equation for diffusion in a periodic potential. In
contrast, we visualize a two-state random walk in velocity space with the particle
alternating between a state of flight and one of localized oscillation. Usingsimple,
physically plausible inputs for the primary quantities chaxact.c.rising~e random walk,
we employ the powerful continuous-time random walk formahsm to derive convenient
and tractable closed-form expressions for all the objects of interest: the velocity auto-
correlation, generalized diffusion constant, dynamic mobility, mean square displace-
ment, dynamic structure factor (in the Gaussian approximation), etc. The interplay
of the three characteristic times in the problem (the mean residence and flight times,
and the period of the 'local mode') is elucidated. The emergence of a number of
striking features of oscillatory diffusion (e.g., the local mode peak in the dynamic
mobility and structure factor, and the transition between the oscillatory and diffusive
regimes) is demonstrated.

Keywords. Diffusion; continuous-time random walk; dynamic mobility; velocity

autocorrelation; dynamic structure factor.

1. Introduction

In paper I o f this series (Balakrishnan and Venkataraman 1981), we have developed

in detail a two-state random walk model for the diffusion of a particle on a lattice.
The formalism is based on continuous-time random walk (CTRW) theory, incorpo-
rating arbitrary holding-time distributions for the state of localized residence at the
lattice sites and for that of flight between sites. The physical motivation for consi-
dering this problem, its ramifications, and the approaches used in certain studies
have already been explained at length in I, and will not be repeated here.
The emphasis in I was on the positional probability or the self-correlation function,
and on the roles played by the two time scales w0 (the mean residence time at a site)
and w1 (the mean inter-site flight time) in the diffusion process. In the present paper,
the focus is on " oscillatory diffusion " : the diffusing particle executes localized
oscillations whenever it is in residence at a site. The ' local m o d e ' frequency thus
introduces a third characteristic time scale into the problem. Our objective is to find
the dynamic mobility (and from it the diffusion constant and the structure factor),
by analysing the random process representing the velocity of the diffus'mg particle.
Our model is a simple one, based on straightforward physical assumptions. It is
not restricted to one-dimension, unlike more formal treatments o f diffusion in a

437
p.wl
438 V Balakrishnan and G Venkataraman

periodic potential that are based either on the Langevin equation (Fulde et al 1975)
or on the Fokker-Planck equation for the conditional density P(x, v, t I xo, Vo) (Diete-
rich et al 1977; Risken and Vollmer 1978; see also Das 1979; Hammerberg 1980).
For clarity and ease of comparison, however, we shall also restrict ourselves to the
one-dimensional case in this paper. Our results will exhibit clearly the effects of
the oscillatory interruptions upon the diffusive motion and vice versa, and will serve
as an illustration of the power as well as the easy generalizability of the CTRW ana-
lysis, given the simplicity of the premises upon which the model rests.
The organization of the paper is as follows. While only the autocorrelation
(v(O)v(t)> is required to determine the dynamic mobility, it turns out that we can
quite conveniently compute the statistical average

00 (~, t) = <exp i~ [v(t) -- v(0)]> (l)

itself, which of course provides additional information. This is the central calcula-
tion. The average concerned is, by definition, given by

0(~, t) = f dvo ~ dvf(vo) P (v, t I Vo) exp [i~ (v - Vo)] (2)

where f(v) is the equilibrium distribution of the velocity, and P(v, t [ vo) is the condi-
tional probability density of this (stationary) random variable. In § 2, we indicate
how P(v, t I v0) is constructed in the framework of the basic two-state random
walk theory, now formulated in velocity space. The physical models for the
functions characterizing the CTRW are laid down and explained. The analytical
expressions obtained for @(~, t) and its Laplace transform ~(~:, s) are presented.
Using these, we obtain in § 3 answers for the frequency-dependent mobility,
its real part (the dynamic mobility), and the velocity autocorrelation function.
Closed-form results are presented for these, various special cases and limits are re-
covered from the general expressions, and graphs are plotted to illustrate the features
of interest. In § 4, the generalized diffusion constant and the mean square displace-
ment are deduced and discussed. The dynamic structure factor is also considered,
in the Gaussian approximation, to bring out the effects of the ' mixing' of oscilla-
tion and diffusion upon the quasielastic and local-mode peaks. Section 5 returns to
tI)(G t) and the additional information that can be extracted from this quantity.
We conclude with a brief summary of the main results of this paper in § 6.

2. Calculation of the function 0(¢, t)

2.1 Construction of the conditional probability density

The formal construction o f the conditional probability density P(v, t I Vo) in a two-
state random walk model proceeds along lines that are already familiar (Singwi and
Sj6lander 1960; see also I). The diffusing particle alternates between a state of
flight, with a holding-time distribution q(t), and a state of localized residence (about
lattice sites), with a holding-time distribution p(t). Further, if a flight step begins with
a velocity v0 at the instant t0, let h(v, t ] vo, to)dr be the probability that the velocity
 Two-state random walk model 439

evolves to a value between v and v + dv at time t, in the same state (of flight). Simi-
larly, let g(v, t [ re, to) denote the corresponding probability density in the oscillatory
state. By enumerating all possible event sequences in the interval (0, t), one can
now construct P(v, t I re) in terms of the primary quantities p, q, g and h. We have
first
oo

P(v, t I re) = X[
n=0
(w° Wo
+ Wl) G~ (v, t l Vo) + (We + Wl) H, (v, t[ Vo) ,] (3)

where G~ [H~] denotes the conditional probability density for the velocity to evolve
from the value vo at t = 0 in the state of residence [flight] to the value v at time t,
via n intermediate transitions of the state of the particle. Multiple integrals (over
the epochs and velocities of the intermediate states) can be written down for G, and
H,. For example, we have (for n >/1)
t tz

G2n(v't[Vo): f dt~n"" f dtl f dv$n.., f dVlP(t--t~n) g(v, tlv~n,t~n)

0 0
• (--q' (ts~ -- t2n-t)) h (V~n,t~n [ V~n-1, t~n-1) (--P' (t2n-1 -- t~n-2))
"'" (--P~ (tl)) g (vl' tl I Vo, 0), (4)

where a prime denotes differentiation, and P0 (t) is the first-waiting-time distribution

associated with the distribution p (t). We shall not write down here the expressions
for G~,+t, H~,, H~,+ 1, GOand H o in order to save space.

2.2 Physical inputs for the functions p, q, g, h

The simple picture of oscillatory diffusion that we adopt provides tractable inputs into
the foregoing machinery for the functions p, q, g and h characterizing the CI'RW,
based on physical grounds. First, if the successive trausitions of state are uncorre-
lated (see I), p and q are single exponentials, i.e.,

p (t) = exp (--tfi'o), q (t) = exp (--t/rl), (5)

so that wo = T0, wt = ~'1- Further, P0 (t) -----p (t) and qo (t) = q (t) in this case. For
ease of writing, we shall sometimes use the notation ~'o, Yl for (1/%) and (1/rt) res-
pectively. Next, we assume that the particle behaves like a classical simple harmonic
oscillator whenever it falls into a residence state: i.e., its velocity then evolves deter-
ministically according to

v (t) = A % sin (% t + ~Oo), (6)

A being the amplitude of the motion. Here 4o is the random initial phase into which
the particle falls at the commencement of the state concerned. Thus our model for g
is simply

g (v, t I Vo, to) = 8 (v--A % sin [% (t--to) + Po]), (7)

440 V Balakrishnan and G Venkataramatt

where A wo sin $o ----vo. The occurrence of the 8-functions simplifies the theory con-
siderably, by &coupling the multiple integrals over the intermediate velocities in
sequences such as (4). Further, the integration over each initial velocity vo with which
every oscillatory state commences can be converted into one over the corresponding
phase $0, with a constant weight factor (or a priori occupation probability) equal to
(1/2~r), and over the range 0 to 2rr. With the above as inputs, the summation in (3)
can be carried out after a Laplace transform with respect to t. After a great deal of
algebra, one obtains a dosed expression for the transform ~ (~, s) of the object re-
quired, i.e., t~ (~, t). This expression is still a functional ofh and of the equilibrium
velocity distribution f The dependence on h is actually quite marginal: by virtue
of the conservation of probability, this function gets integrated out to unity (regardless
of its actual form) in all the intermediate states in which it occurs, because of the
decoupling mentioned earlier. It survives explicitly only in places where it repre-
sents the terminal diffusive state in G~,+~ and/-/2,. This circumstance* drastically
reduces the 'error' introduced by any approximate form assumed for the function h,
provided the latter satisfies certain basic physical requirements. Given this, and the
physical fact that h describes flight over a single lattice distance at most, it is not
unreasonable to take the velocity to be a constant over the step (i.e. to assume free
flight); in other words, to set

h(v, tlvo, to) = ( v - v0). (8)

Indeed, it is this approximation, together with the simple exponential form for q (t),
that specializes an otherwise general theory to the case of diffusion in a lattice. With
more general forms for q and h, the theory could easily be formulated to describe
oscillatory diffusion in a liquid or a disordered solid. (In the latter case, a distribution
in w0 could for example be incorporated in the function g). In a liquid, for instance,
the individual flight steps are over variable distances. In a sufficiently long flight
step, frictional effects (thermalization) must also be taken into account. It is there-
fore appropriate in this case to specify h itself as the solution of a ' diffusion' equation
(Singwi and Sjdlander 1960)--in the case of the velocity variable, we should choose
the solution of the (potential-free) Fokker-Planck equation. Thus h (v, t I v0, to)
would be a Gaussian distribution in the variable Iv -- vo exp (--y (t -- to))], y being
the friction constant or the reciprocal of the correlation time of the velocity. When
t < ~,-i, this solution tends to the 8-ftmction of (8). The latter is therefore a plausible
choice for the single-lattice-distance jumps (with a mean flight time -r1 ~ 10-12 sec,
say) occurring in the current problem.

2.3 The equilibrium velocity distribution

Finally, we turn to the specification of the equilibrium distribution f(v). Since the
flight steps are assumed to leave the velocity unaltered (see (8)), f(v) is essentially
determined by the motion in the oscillatory state. Here the velocity has the functional
form v = A % sin 4', and the occupation probability density in the space of the phase

*For, of the (N + 1) (N + 2)/2individual states taken into account in 2~ Gn, any approximation
for h affects only [(N + 1)/2] states. One therefore expects the relative error to become negligible
when a completesummation is done, Le., when N ~ o~. A similar statement holds goodfor 27Hn.
 Two-state random walk model 441

variable ~ is a constant, namely, 1/(2z0. Hence the relation f ( v ) d v = d¢/(2rr)

trivially yields (with 0 denoting the step function)

f(v) = 77"-1 ( A s ~O2 - - u2) -1/2 0 ( A O)0 - - I V 1), (9)

It remains to specify the amplitude A of the oscillatory motion. Equating the mean
square velocity in equilibrium to kBT/m (where m is the mass of the diffusing
particle), i.e., using the equipartition theorem, we get

A = (2 k B T/mto~o)l/s. (lO)
The normalized equilibrium distributionf(v) that we use then reads

f(v) = zr-1 (as -- v2)-1Is O(cr-- I v ]) (ll)

where a ( = A%) = (2k B T/m) 1/~. (12)

Incidentally, it is clcar that (10)-(12) also ensure that there is no preferred' condensa-
tion' of the diffusing particles into one or the other of the two states (flight and resid-
ence).
Before the results of the calculations are presented, a few further remarks on the
distribution (11) are in order. These are best made in the form of a comparison
with the conventional Maxwellian distribution of velocities,

f M(v) = (~ra2)-1/2 exp (--vS/(r2). (13)

B o t h f a n d f M are symmetric, and their second moments are of course equal. For the
Maxwellian, ~v°-") = (~s/4)" (2n)t/n!. The corresponding moment forf(v) is smaller
by a factor (l/n!). Being of compact support, f(v) must clearly have a negative
excess of kurtosis; the actual value turns out to be --3/2. And lastly, while the
characteristic function corresponding to (13) is

f dVfM(v ) exp (i~v) = exp (--~a~/4), (14)

--00

which is again a Gaussian, that for the distribution at hand is

(exp (i~v))equi I = f dv f(v) exp (i~v)

--00
21r
= (2zr)-I f d~ exp (i~a sin ¢) = Jo(a~), (15)
0
where e is given by (12). It is worth noting, too, that
2rr
(2zr)-1 f de exp [i~a sin (%t + ~)1 = Jo(¢~), (16)
0
442 V Balakrishnan and G Venkataraman
2ff

and (2rr)-1 f d$ exp [i~(r {sin (%t + $) -- sin $}] = J0(2,~ sin { oJ0t). (17)
0

These integrals occur in the calculations.

2.4 Result for 0(~, t)

Using the foregoing as input, we may calculate the statistical average 0(~, t) defined
in (1) and (2). Suppressing the (lengthy) algebra emirely, we present the final result
for the Laplace transform ~)(~, s):
O0

~(~, s) ----f dt <exp i~ [v(t) -- v(O)]) exp (--st)

o
O0

= [-r~ + ('r° + ~'1 + S'ro~'l)j~ (~) + (¢o + ¢1 + SZo¢l) f dt'

S(1 + S'ro)
0
× Jo(2(r~sin½%t')exp{--t'( s + "rol)]']/[(~o + ¢,) (1 + s~,)],
(18)
where (to recall the notation) % is the mean residence time at a site, ~'1 is the mean
flight time between steps, % is the ' local mode' frequency, and v = (2kB T/m) 11~.
The Bessel functions appear as a consequence of the oscillatory motion. In the
next section, we shall determine the dynamic mobility with the help of (18).
Inversion of the Laplace transform (18) yields after simplification the following
expression for 0(~, t):

I~(~, t) ---- <exp i~[v(t) -- v(O)]>

_ 1 IT1 exp (-- t/¢l) W %Jo(2Cre sin 1 toot) exp (-- t[%)
(¢o'+ ~1)

J~ (o~) . { ~ (I -- exp (-- t/¢o)) -- ~ (1 -- exp (-- t/'rl) }

+ (~o - ~1)
O0
1 1
-q-exp ( - - t / ~ ' , ) f dt'' J0(2o, sin ~ % t ' ) e x p ( - - t ' (~o--~1)) ].
o
(19)

Before proceeding further, we must check the conservation of probability. It is

evident that 0(0, t) must be equal to unity by definition, as must ~)(~, 0). It is easily
verified that the final expression in (19) meets these requirements. This is tanta-
mount to a confirmation that all possible event sequences have been properly included
in the determination of the conditional probability P(o, t [ vo).
 Two-state random walk model 443

The second important check is the t-+ 0o limit. Under equilibrium conditions,
one must have P(v, t --> 0o ] vo) ~ f(v) (fluctuation dissipation). Hence, from (2) and
(15),
lira ¢(~:, t) = <exp (i~v)>equi1 <exp (i~Vo)>equil = J~(a~). (20).
t--~ oO

Once again, an inspection of (19) affirms that this limiting value is recovered.

3. The velocity autocorrelation function and the dynamic mobility.

3.1 The frequency-dependent mobility

The velocity response is described by a generalized susceptibility, namely, the fre-

quency-dependent mobility ~(to). As is well-known from linear response theory, this
is given by
oo

~(to) = fl f dt <v(t) v(O)> exp (itot), (21)

o

i.e., by the Fourier-Laplace transform of the velocity autocorrelation. This is found

easily by expanding ~(~, s) as given by (18) in powers of $, and isolating the term
proportional to $~. Analytic continuation of the result to s = - i to is trivial.
We obtain, after using the fact that <v~) = (kBT[m), the compact expression

1 ['r~ + %(1 --itoro) (-ro+rl--ito % ~-1)]

(to) -- m(zo+~'l) (1--ito~'l) (1 --ito~-o)~ + too~ ~'~o -j. (22)

The effect of the' local mode' is contained in the second term in the square brackets.
The solution in the absence of oscillatory behaviour can be obtained by letting
too + 0o (and simultaneously A -+ 0: recall that Az ~o~/2 has been set equal to kBT/m )
so that the particle is static in tile state of residence. The result is

(to) + ~-~/ [m (% + ~'1) (1 -- ito ~'1)]- (23)

This differs from the standard ' free diffusion' expression

~o (to) = 1/m (3"1-- ito), (24)

(where Yl = l/z1 is the friction constant) only by the factor rl/(% + ~a). The latter
is easily understood as the fraction of time the particle spends in the 'diffusive'
state. Other limits and special cases are discussed in what follows.
The structure of the exact result for ~ (co) for diffusion in a periodic potential is a
complicated one. The analysis is complicated, and various systematic approximation
schemes have been developed (Fulde et al 1975; Dieterich et al 1977; Risken and
Vollmer 1978). The frequency-dependent mobility can be obtained as an infinite
444 V Balakrishnan and G Venkataraman

continued fraction, for instance. While the concise expression (22) clearly does not
possess this structure, owing to our adoption of a simple model of oscillatory diffu-
sion, it does provide a reasonable facsimile of the actual physical situation in that it
demonstrates most of the non-trivial features one may expect the susceptibility con-
cerned to possess. It is also sufficiently general to encompass several known results
as special cases, as will be seen shortly.

3.2 The velocity autocorrelation function

The autocorrelation function (v (t) v (0)7 conveys much information about the
random process representing the velocity of the diffusing par'tide, in a physically
perspicuous manner. Using (22) for ~ (oJ) and inverting the transform in (21), we
obtain

= (kB T I 1
(v (t) v (0)> \--m-/('YO-}-?I) [(7o--71)2-]-to~o] [70 (78- 7o 71AI-to~)

exp (--71 t) + 71 (7~-- Yo ~"1+ to02)COS (too t) exp (--7o t)

q- 7o 71% sin (too t) exp (--70 t)], (25)

where 7o = 1/%, 71 = 1/~'1 as defined earlier. It is trivially verified, on setting t = 0,

that (v~) = (k B T/m). For ready comparison, it may be recalled that the Langevin
equation for the Brownian motion of a simple harmonic oscillator (which is of course
distinct from that for Brownian motion in a periodic potential) yields the result
(Uhlenbeck and Ornstein 1930, Chandrasekhar 1943)

(v (t) v (0)) = (k B T/m) (cos ~ t -- (7/2T~) sin ~ t) exp (--7t/2), (26)

where y is the friction coefficient and ~2 = (to~ _ 7~/4). Note that there is no diffu-
sion in this instance, as the integral of (26) over t from 0 to oo vanishes. Indeed, the
displacement x (t) is itself a stationary random variable in this case, with an expo-
nentially decaying autocorrelation function.
Consider now the special cases deducible from (25). First, if there is free diffusion
with no halts at sites, % -+ 0, i.e., 7o --> oo. Then

(v (t) v (0)7 -> (k B T/m) exp (--71 t). (27)

This is just the answer obtained from the ordinary Langevin equation, and quite
evidently is equivalent to (24) for ~0 (~o). On the other hand, instantaneous jump
diffusion implies that -r1 ----0, so that

(v (t) v (0)) --> (k B T/m) exp (--70 t) cos too t, (28)

which is again readily understood--the correlation function for an oscillator, modu-

lated by the holding-time distribution exp (--70 t).
 Two-state random walk model 445

3.3 The dynamic mobility

The quantity of direct experimental interest is the dynamic mobility*

oO
/~ (oJ) ~ Re ~z (~o) = fl f dt (v (t) v (0)> cos oJt. (29)
0

(The integral on the right is just 1[4 times the power spectral density of the velocity
variable.) We find, from (22)~

'YO 'Yl [(Y~O--~O y1+%~)

/~ ("') - m (To+'Y0 [(to--'Y0 ~ + %z] L (-~~4o-,"-"3
_1 (Y]--Yo Yl-F2°J~o4-W%) -F 1 (Y]--Yo Y14-2°~2o--°J%)]. (30)
2 7o~-F(co+%)~ 2 ~,o~4-(~--V~o)
2
The simultaneous occurrence of diffusive and oscillatory characteristics, and the
build-up of the' local mode' peak in/z (oJ), are illustrated in figure I.

1.4

1-0

3
zk
E
fM
0.6

0.2

0 4 8 lc_.2
tOT
Figure 1. Variation of the dynamic mobility p,(ta) with frequency (equation (30)),
illustrating the occurrence of both 'diffusive and 'oscillatory peaks, in the case Y0 =
yl (-- 1/~-). Curves (a)through (d)refer respectively to to0r = 2, 3, 5 and 10.

*For instance, in the modelling of superionic conductance via diffusion in a periodic potential
IL(~o) is directly related to the conductivity.
 446 V Balakrishnan and G Venkataraman

As in earlier works (e.g., Dieterich et al 1977), it is convenient to normalize/~ (co)

by dividing it by the static mobility/z 0 = ~o (0) = 1/(m ~'1) corresponding to free diffu-
sion. Figure 2 shows the enhancement of the local mode peak at the expense of the
diffusive peak in the quantity/z (o~)/tZo, as ~'1 decreases relative to %. These plots are
complemented by those in figure 3, in which the family of curves corresponds to a
fixed value of'r 1 and varying values of%. (It is of course evident on physical grounds
that t h e ' local m o d e ' peak will disappear whenever % T0<I.) Comparing the total
picture gathered from figures 1-3 with that deduced by other approaches (Fulde et al
1975, Dieterich et al 1977, Risken and Vollmer 1978), we conclude that the results of
our simple model display all the essential features expected of the exact velocity res-
ponse in oscillatory diffusion*, without necessitating any involved analysis or approxi-
mation scheme. Note, for instance, the minimum in/z (o,) at a finite value of o, in
curves (b) and (c) of figure 2; as pointed out in Dieterich et al (1977), this feature is
peculiar to oscillatory diffusion (or diffusion in a periodic potential). It cannot be
reproduced by simpler approaches (such as a generalized Langevin equation with an
exponential memory kernel) that lead to a two-pole approximation to the continued

15

o 1.0
=:L
3
:::L

0.5 I'~

l °1 I I
0 0.4 0.8 1.2 1.6 2.0
w/w o
Figure 2. Normalized dynamic mobility #(co)iV0 as a function of o/co0, for varying
values of the mean flight time ¢1. All the curves correspond to co0~-o= 5, with (a)
coocx = 20, (b) coo~'1= 5, (c) cao~'l = 2, (d) co0~'l = 1.

*Compare, for instance, curve (b) of figure 3 with the experimental curve of the conductivity cr (~)
for the superionic conductor AgI, reported in Brttesch et al (1975).
 Two-state random walk model 447

3"0

._~ 2"0

1"0

0
0.4 0.8 1.2 1.6 2.0
to/to o
Figure 3. Normalized dynamic mobility/~(to)/#oas a function of to/to0, for varying
values of the mean residence time ~'o. All the curves correspond to too~'l= 1, with
(a) too'to = 20, (b) too"ro = 5, (c) too"ro = 2, (d) too'to = 1.

fraction representation for ~ (to). We emphasize, therefore, that (30) is more than a
mere' two-pole approximation' of this sort.

4. T h e s t a t i c m o b i l i t y a n d the structure f a c t o r

4.1 The generalized diffusion constant

The static mobility is measured by the diffusion constant D, according to

D = k B r ~, (0) = k B T/~ (0). (31)

Our theory yields the expression

D -----\ - ~ / L ( 1 -~-.o 2 a "4- 1, (32)

~'~ ~'~) (To + ~'1)J
which, albeit simple in appearance, has a non-trivial structure built into it. Note
in particular that D is not simply given by (a 2 + (X~osc))/2 (% + ~'1) (where a is
448 V Balakrishnan and G Venkataraman

the lattice constant and (XZosc) is the mean square displacement in the oscillatory
state), as one may expect at first sight and as is sometimes written down in the
literature. However, the two terms in the square brackets in (32) do represent res-
pectively ' oscillatory' a n d ' diffusive ' contributions to D.
When % = 0, the well-known free diffusion result D O = (k B Tim ~'1) is retrieved
from (32). Further, if the ' local m o d e ' is absent, i.e., if during residence at a site
the particle is static, then (32) yields (on letting % -> oo) the physically understand-
able result

D = D O-r1 / (% -t- "q). (33)

This is precisely the effective diffusion constant used, for instance, in the description
of diffusion in the presence of traps (Schroeder 1976).
The extraction of D in the jump diffusion limit (r 1 -~ 0) from (32) reqmres a bit of
care. In the first term, one may replace k B Tim by oJ~o (Xo~sc). In the second term,
since k B Tim = (v~-), and r 1 is the mean flight time between neighbouring sites,
we may write 2 k B Tim = a2/~-~ Hence (32) may be re-written as

aS
D -- (X~°sc) ( % "r°)~ ~ -+- (34)
% 1 q- (% %)2) 2 (% -4- ra)"

Letting 71 --> 0 in this representation yields D for jump diffusion. When there is no
oscillatory motion, the first term on the right in (34) disappears. The familiar formula
D -----a~/2 % is recovered. Note also that when % % = 1, and only then, is the
diffusion constant for jump diffusion given by (a S + (Xo~sc))/2 %.
Given the fact that (32) encompasses all the special cases above, it is of interest to
see how the static mobility varies with the ' friction' ~'1 (---- 1/~'1)- This is shown in
figure 4, in which we plot m % / z (0) (or m % D/k B T) against vl/oJo for various fixed
values of % %. A similar plot is depicted in figure 2 of Dieterich et al (1977).
However, the latter is restricted to the rather narrow range 0.3 ~ ~'1/% <~ 1.6, and
moreover there is no explicit parameter corresponding to % in the work referred
to*. Within the above range of 71/o~o, there is broad agreement with the results of
the present work, specifically with the curve corresponding to % % ~ 20. Finally,
it is easy to show from (32) that /~ (0) ,~/~o according as %z >< ~'o vl, i.e,, %
(% ~1)-1/~.
4.2 The mean square displacement

Making use of the stationary of the velocity, the mean square displacement in a time
interval t is given by
oo
((x (t) -- x (0)) e) = 2 f at' (t - t') (v (t') v (0)). (35)
0
*There is of course the parameter Vo/kB T, where Vo is the depth of the well in the periodic
potential; if interpreted as the rate of escape over the barrier represented by a maximum of the
potential, ~'0 = 1/,% can be related to this parameter,
 Two-state random walk model 449

10 2

101

:t I
E

io-I

-2
10 I I I
16 2 io-' i 101
~I~o
Figure 4. Variation of the static mobility/z(O) with friction. Curves (a) through (d)
refer respectively to ¢o0¢o= 1, 5, 20 and 100.

Substituting (25) for the autocorrelation function and integrating, we find

<(x(t)--x(O))~> = 2 Ot -- (70 -~ 71) [(70-- )'1)2 "-{- °921

X {c 1 [1 -- exp (--71 t)] -t- c~ [1 -- exp (-- 7o t) cos % t]

-t- c3 exp (--70 t) sin % t}, (36)

where D is the diffusion constant (as required), and the constants c~ are defined by

q = y,, (~'~- ~'o ~'~ -t-- %~)/~,~,

% = 71 -(2 72oo9o + (7,~ - 7o y, + ,,,oh (7o~ - ,Oo2)}16,~ + %~)~; (37)

ca = 7o 71 o9o (2 7~ -- 2 7o 71 -- 7g -¢- 3 oJ~o)/(7~ + o9~)2.

Once again, on letting 7o ~ oo, we recover the free diffusion result

((x (t) -- x (0))0~> = (2 Do/Tx) (71 t -- I d- exp (-- 71 t)) (38)

450 V Balakrishnan and G Venkataraman

where D O= (kB Tim Yl) as already defined. Similarly, if there is no local mode but
only static residence (i.e., % -+ co but y0< oo), we find for the mean square displace-
ment exactly the same functional form as in (38), with however D o replaced by the
effective diffusion constant Dorl/('ro + zl). These results tally with those derived
in I by a totally different route, namely, from a consideration of the random walk
problem on the lattice in real space. Finally, in the case of jump diffusion and in
the absence of the oscillatory motion, all that survives on the right-hand side in (36)
is the term 2Dt where D now stands for a~/2-ro. This too is as it should be, for the
self-correlation function in this special ease obeys (the discrete or lattice analogue of)
the simple diffusion equation, and the solution is a Gaussian, with a variance pro-
portional to t.

4.3 The dynamic structure factor

We now consider in brief the dynamic structure factor S(k, (o), as this aspect of diffu-
sion is among those closest to experiment. Our purpose is to comment on the manner
in which the mutual interference of oscillation and free diffusion is manifested in the
shape of S(k, (o). For sufficiently small values of k, a satisfactory approximation
to S(/c, ,o) is
OD

S(k, co) ~_ (lfir) f dt cos oJt exp -[--½k ~ ((x(t) -- x(O))~)). (39)
o

The mean square displacement has the form displayed in (36). Using this in (39),
we may draw the following conclusions. The quasi-elastic peak in S(k, oJ) has an
FWHM that is approximately equal to 2Dk 2, where D is the generalized diffusion
eonstant found earlier. The interplay of oscillatory and diffusive characteristics in
D is already explicit in (34). For sufficiently large values of % r 0, there is a secondary
' local mode' peak in S(k, a,) near co = %. This peak has an FWHM approxi-
rnately equal to 2(Dk ~ + Yo), and is therefore considerably broadened, relative to
the quasi-elastic peak.
Figure 5 illustrates the sort of result obtained by numerical integration of (39) after
(36) is inserted for the mean square displacement. The values chosen for the various
parameters are approximately the same as those employed in Dieterich et al (1977)
(see § 6 and figure 5 of their paper), for the sake of comparison. Thus we set 2try1~%
(denoted by I" in the latter paper) equal to unity, and

(k2/a,~) (kBT/m) ~ 2(k~/k~) (kBT / Vo) : 2 × (0.2) 2 >((0.3) : 0.024, (40)

and finally Yo/% = 20. Then

o0

oS(k, ,o) = f cos (z,o/o,o) exp ( - F(z)),

o

where F(z) = 0.0369z -- 0.2193 (1 -- exp (-- O.I6z)) + 0.0182

(1 -- exp (-- O.05z) cos z) -- 0.0027 exp (-- O.05z) sin z. (41)
 Two-state random walk model 451

-I
t/)

1631
I0.1
1
0.3
I
0.5
I
0.7
I
0-9
I
1.1 1-3
co/~ o
Figure 5. Dynamic structure factor as a function of co/oJ0, in the Gaussian approxi-
mation (the ordinate is on a logarithmic scale). The numerical values of the para-
meters are specified in (40) ft. Curve (a) corresponds to the full integrand in (41).
For comparison, curve (b) is the Lorentzian obtained by retaining only the first term
(0.0369z) in the expression (41) for F(z).

Curve (a) in figure 5 is a plot of the structure factor given by (41) as a function of the
frequency, exhibiting the additional local mode peak at o, ~ %. If we use merely
the term 0.0369z for F(z) the outcome is the Lorentzian plotted in curve (b). This
amounts to approximating the mean square displacement in (39) by its asymptotic
value 2Dr, and helps give an idea of what the quasielastic peak would look like if the
other contlibutions to the mean square displacement were absent.

5. The velocity increment distribution; connection with ' interpolation ' models

We have computed, in the foregoing, the autocorrelation (v(t)v(O)) (and all the
other quantities related to it such as ~(~o), D, etc.) without explicitly obtaining
first the joint probability density P(v, t; vo, O) ~ f ( v o) P(v, t ] Vo). This is because
the convolution-structured CTRW formalism makes it natural and convenient to
work in terms of an appropriate Fourier transform with respect to the velocity vari-
able*. It is evident from the definition (1.2) that ~(~:, t) is the transform of P(v, t; v0, 0)
with respect to the difference variable (v -- Vo). Clearly, 0(~:, t) does not contain
as much information as, say, the characteristic function X(¢t, ~:~; t) of the distribution
P(v, t; vo, 0), being a special case: 0(~:, t) = X(~:, --~:; t). It does, however, carry
more information on the velocity variable than is incumbent in the autocorrelation
(v(t) v (0)), and it is this aspect to which we now turn, foz the sake of completeness.

*And, of course, a Laplace transform with respect to the time.

452 V Balakrishnan and G Venkataramaa

5.1 The probability density of the ve'locity increment

The explicit form of 0(~:, t) has already been written down in (19), and the limiting
values for t--> 0, t ~ ~ (representing respectively the conservation of probability
and the fluctuation-dissipation theorem) have been checked. One may first ask what
the inversion, of the Fourier transform with respect to the variable ~: yields. The
answer, which we denote by P(u, t), is easily seen to be

oO
P(u, t) = (l/2zr) I d~ 0(~, t) cxp (-- iu~)
uOO

= ~ dyef(Vo) P(vo + u, t l ~o). (42)

The physical meaning of P(u, t) is quite evident: it is the net probability density
associated with a velocity #wrement u in the time interval t. Performing the Fourier
inversion of the function ~(~:, t) given by (19), we find

7r(ro + rl) P(u, t)

= 7rr1 8(u) exp (-- t/'q) + r o exp (-- t/re) (4or°"sin2 ½ coot -- u~-)-alz

× Ir~ [1 -- exp (--t/re) ] -- r~ [1 - - e x p (--t/ra)]l 0(2~ - - [ u [ )

TO ~ T1

÷ ox,.-.I.,
f [_..(._i)]
o
x (40 °"sin ~ ½ %t' -- u~')q/'" 0(12or sin ½ toot' [ -- I u 1), (43)

where, to recall (12), cr = (2koT/m)l/"- , and P-l/2 stands for the Legendre function of
order -- ½. The restriction of [ u [ to values -~ 2~ is easily understood, for the velocity
itself is restricted to the range -- ~ ~< v ~ ~.

5.2 Conditional density for free diffusion

It is essentially the interleaving of two random processes, with correlation times %

and r 1 respectively, that has necessitated the use of the powerful CTRW formalism
in the present work. At various stages, we have paused to consider the limiting cases
of free diffusion (% = 0) and jump diffusion (z1 = 0). We shall now show that
P(v, t t %) can itself be obtained explicitly in these special cases, from an inspection
of the structure of 0( ~:, t). This exercise is instructive, as it provides some insight
into the make-up of the CTRW approach to the general problem. We consider first
the case of free diffusion, i.e. % = O.
 Two.state random walk model 453

In the above limit, (19) reduces to

q~ (~:, t) ----=exp (--Yt t) + J~ (tr ~) [1--exp (--Yt t)]. (44)

The autocorrelation function corresponding to (44) is of course the simple exponential

already given in (27), which matches the standard Langevin equation result. We
have also compared the equilibrium distribution f (v) (see (9)), the characteristic
function of which is J0 (o ~), with the MaxweUian fM (v) (see (13)). Continuing the
comparison, in the conventional picture the velocity is a Ganssian random variable,
so that

(~,t) = exp I -kB~mT~z [1--exp (--Ttt)] 1 • (45)

Further, P (v, t lvo) satisfies the Fokker-Planck equation, and the solution is
the Ornstein-Uhlenbeck distribution (Utdenbeek and Ornstein 1930). What is
P (v, t [ re) (for free diffusion) in the present case ?
The answer is quite simple, and may indeed be guessed at from (44) itself. It is

P (v, t I v0) = 8 (v -- vo) exp (--n t) +f(v) [1-- exp (--Yl t)]. (46)

In other words, when one of two correlation times (here, ~'0) vanishes, the CTRW
result simplifies to an interpolation model for the conditional density: the expression
in (46) interpolates between the initial distribution 8 ( 0 - 0o) and the equilibrium
distribution f(v). The Chapman-Kolmogorov equation is also obeyed by (46), so
that the Markovian nature of the velocity process is retained. The structure of inter-
polation models for a Markov process, comparison with the solution of the Fokker-
Planck equation, and generalization to the non-Markovian case have already been
presented elsewhere (Balakrishnan 1979).

5.3 Conditional density for jump diffusion

When ~'t = 0, so that the inter-site flights are instantaneous jumps, (19) becomes

¢ (~, t) = Jo (2 o f sin ½ % t) exp (--Yo t) q- JZo (a ~) [1 -- exp (--Yot)].

(47)
The autocorrelation function is the exponential-modulated cosine already written
down in (28). The non-trivial deterministic evolution of the velocity in the resident
state (see (6)) precludes the process from being Markovian. However, P (v, t 10o)
can again be shown to fall within the purview of interpolation models of a some-
what more generalized nature (Balakrishnan 1979). It can be verified that the solu-
tion that leads to (47) is
e(v, t I vo) = $ [v--o sin (% t + ~o)] exp (--Yo t) + f(v)
x [1 - exp ( - r e t)], (48)
where o -----A % = (2kB T/m)11~(equation (12)), f(v) is given by (9) as always, and
¢o is defined by o sin ~,0 = re.

P.--2
454 V Balakrishnan and G Venkataraman

As already stated, when both r o and ~-x are non-zero, the structure of the theory
becomes non-trivial.

6. Concluding remarks

The significance of the general problem of interrupted diffusion has already been
expounded in paper I. Physical applications of current interest include superionic
conductance, diffusion in the presence of traps, and the diffusion of hydrogen in
metals--specifically, in the transition regime between jump diffusion in a lattice and
fluid-like diffusion, when the mean residence time at a site and the mean inter-site
flight times are comparable. Oscillatory diffusion, as occurs for instance in the pre-
sence of a local mode, complicates the problem further; a third characteristic time is
introduced. Earlier approaches address themselves to the problem of diffusion in a
periodic potential (in one dimension), and treat it in terms of a Fokker-Planck equa-
tion for the conditional probability density. Rather complicated systematic approxi-
mation schemes are then developed to evaluate the relevant response functions (such
as the frequency-dependent mobility, etc.), and to demonstrate the simultaneous
occurrence of oscillatory and diffusive effects in these. In contrast, we have visua-
lized a two-state generalized random walk model of the diffusion process, in which
the particle alternates between a state of flight between lattice sites and one of loca-
lized oscillation about a site. With the help of simple physical inputs for the primary
quantities characterizing this random walk, we have, in this paper, used the very
effective CTRW technique in velocity space to derive convenient and physically
sensible closed-form expressions for all the quantities of interest in oscillatory diffu-
sion. The structure of these expressions is quite non-trivial, while remaining emi-
nently tractable, and a number of known results are recovered as special cases on
passing to the appropriate limits. The result for the generalized diffusion constant
D that is displayed in (32), and the subsequent discussion, should serve as a convinc-
ing illustration of these statements.
The scope of this paper has throughout been restricted to the following broad
objective: the derivation of a theoretical description encompassing certain qualitative
features expected of oscillatory diffusion on the basis of a variety of experimental
observations (e.g., the frequency-dependent conductivity of certain superionic con-
ductors, the structure factor probed by neutron scattering studies of hydrogen in
metals, etc.) The features referred to include the transition region from jump
diffusion to fluid-like diffusion, the peak at a non-zero value of the frequency
in the dynamic mobility, the secondary peak in the dynamic structure factor, and
so on. The attainment of such an objective is clearly an essential prelude to the
detailed analysis of specific experimental findings. The latter would in any case
involve a tailoring of the general theory in its details, depending upon the particular
case under study. Not the least of the advantages in favour of the approach
adopted here is its ready amenability to the generalizations or modifications that
may be required in this regard: for example, the important one of extension to
three dimensions; the incorporation of more involved holding-time distributions
controlling the diffusion process; and the use of different functional forms for the
basic probability distributions g and h to take into account the distinct physical
circumstances encountered in different problems.
 Two-state random walk model 455

Acknowledgements

We are very grateful to V Umadevi for her invaluable help in carrying out the nume-
rical computations leading to the graphs presented in the figures.

References

Balakrishnan V 1979 Pramana 13 337

Balakrishnan V and Venkataraman G 1981 Pramana 16 109 (referred to as I)
Brttesch P, Str[ssler S and Zeller H R 1975 Phys. Status Solidi A31 217
Chandrasekhar S 1943 Rev. Mod. Phys. 15 1
Das A K 1979 Physica A98 528
Dieterich W, Peschel I and Schneider W R 1977 Z. Phys. B27 177
Fulde P, Pietronero L, Schneider W R and Str~ssler S 1975 Phys. Rev. Lett. 35 1776
Hammerberg J 1980 Physica A100 119
Risken H and Vollmer H D 1978 Z. Phys. B31 209
Schroeder K 1976 Z. Phys. B25 91
Singwi K S and Sj61ander A 1960 Phys. Rev. 119 863
Uhlenbeck G E and Ornstein L S 1930 Phys. Rev. 36 823
Wang M C and Uhlenbeck G E 1945 Rev. Mod. Phys. 17 323