PHYSICAL REVIEW VOLUME 157, NUMBER 1 5 MAY 1967

On the Velocity Autocorrelation in a Classical Fluid*

K . S. SlNGWI AND M . P . TOSI
Argonne National Laboratory, Argonne, Illinois
(Received 22 September 1966)

The evaluation of the velocity autocorrelation function and of its frequency spectrum in a classical
liquid is discussed in the framework of the formalism developed by Tjon for the spin autocorrelation function.
This is based on an integrodifferential equation involving a memory function, and is shown to be equivalent
to the well-known approach based on the time expansion of the correlation function. The use of a Guas-
sian memory function leads to fair agreement with the results of Rahman's numerical experiments for
liquid argon.

I. INTRODUCTION where
$W = <v(0)-v(0)r/^V (2)
T H E calculation of the velocity autocorrelation
function of an atom in a liquid is a difficult and as
yet an unsolved problem. I n a real fluid this function is
The frequency spectrum /(co) is defined by the Fourier
transform
not a simple exponential but has a negative part as its
characteristic feature. Recent work of Nijboer and 2 r"
Rahman 1 has shown that the representation of this f(f $(t)cos(o)t)dt. (3)
Wo
function by means of its time expansion is far from
satisfactory because the expansion converges very The time expansion of the correlation function, there-
slowly. Thus the correlation function does not even fore, reads
become negative when one includes the ft term in the
time expansion. This term already involves the fourth
moment of the frequency spectrum of the velocity *(0=E(-i)" < » » , (4)
autocorrelation, a quantity which can in principle be n-=0 (2»)!
evaluated from a knowledge of the interatomic po-
where (co2n)av are the even moments of the frequency
tential but is in fact very poorly known.
spectrum.
In this paper we follow a method which was used by Equation (1) implies the following recurrence re-
Tjon 2 to calculate the spin autocorrelation function and lationship 2 between the even moments of /(co) and the
is equivalent to summing the time expansion to infinite even moments a(2k) of the Fourier transform of the
order, though approximately. The method consists in kernel a (t):
evaluating the velocity autocorrelation function from a
linear integrodifferential equation whose kernel has
<^> av = - E <^-*>>ava<2*-2>. (5)
the meaning of a memory function. The memory func-
tion is taken to be a simple Gaussian whose parameters
are related to the diffusion coefficient and to the second Conversely, starting from the time expansion (4) and
moment of the frequency spectrum. An explicit calcula- using the recurrence relationship (5) and the identity
tion of the velocity autocorrelation function and of its
frequency spectrum for liquid argon a t two tempera-
tures yields fair agreement with the numerical experi- %n(\ — x)mdx=nlml/ (n+m+1)1,
ments of Rahman. A detailed discussion of the integro-
differential equation, and applications based on an one can establish Eq. (1) in a straightforward manner.
exponential kernel, have been presented recently by Equations (1) and (4), therefore, provide two equiva-
Berne, Boon, and Rice. 3 lent formulations of the problem, which are connected
by the relationship (5). T h e advantage of using the
II. THEORETICAL DISCUSSION former approach is that one can guess a reasonable
The integrodifferential equation for the normalized functional form for the kernel a(t), which has the
velocity autocorrelation function <£(/) is2 physical meaning of a memory function.
l The frequency spectrum /(w) can also be evaluated
d${t) r
= / a{r)Ht-r)dr, (1) directly from the Laplace transform of the kernel a (t)
dt Jo for imaginary argument. Indeed, b y multiplying both
members of Eq. (1) by e±i(at and integrating over time,
* Based on work performed under the auspices of the U. S. one finds
Atomic Energy Commission.
1
B. R. A. Nijboer and A. Rahman, Physica 32, 415 (1966).
2
J. A. Tjon, Phys. Rev. 143, 259 (1966). 1
3
B. J. Berne, J. P. Boon, and S. A. Rice, J. Chem. Phys. 45, /(«) = - (6)
1086 (1966). irLd (too) + ico a ( — ioo) — io).
157 153
154 K. S. SINGWI AND M. P. TOSI 157

where i 1 "

i.o <

<2(±iw) = (t)e±io}tdt. (7) T - 94.4 °K

Jo

In particular, one has 0.5 -

\o

/(o)=- / a(t)dt . (8) *(t>)

o
0.1 :
One sees immediately from Eq. (1) that, when a(t) 0
Vk

^A
0

0
°%
^<3«£*-o o o o o o o o o o
-0.1 -
is a delta function, $(/) is an exponentially decaying
function corresponding to the motion of a particle
governed by the Langevin equation. On the other ._ J ... » 1
5 10
hand, if a(t) is a constant, <£(/) is an oscillatory function (k B T/-h) t
of time corresponding to the case of an Einstein oscilla-
tor. Thus the two limiting cases are contained in Eq. (1) FIG. 2. Velocity autocorrelation in liquid argon at 94.4°K.
Curve A: <co2)av = 55Xl0 24 sec^, D==2.43X10-5 cm2 sec"1. Curve
for appropriate choices of the memory function. A B : (w2)av=55Xl024 sec"2, £ = 2.19X10- 5 cm2 sec"1. The circles
simple and reasonable choice for the memory function are from A. Rahman, Ref. 4.
in a liquid, which leads to the two limiting cases for
appropriate choices of the parameters, is a Guassian if the parameters satisfy the condition
form,
0.057^^/^1^0.943. (ID
a(t) = -Ae (9)
The parameters A and B of the Gaussian kernel can
We note that the adoption of an exponentially decaying be fixed from experimental data by means of Eq. (5)
memory function is not consistent with the relationship for n= 1 and of Eq. (8). These give
(5), since only the zeroth moment of its Fourier trans-
form exists. However, this form of the kernel will become A = (w2)av (12)
correct at large time.3 and
B^/A^W'fiO). (13)
For a Gaussian kernel, the frequency spectrum f(co)
tends to zero at high frequencies as or2 exp(—u2/4B).
On the other hand, at small frequencies /(a?) has the
form H „ = ( l / 3 ¥ ) / < g(r)VV(r)dt (14)

f co2r 16/B l\ 2 -j) and

f(0) = 2MD/7rkBT. (15)
Here, g(r) is the pair distribution function, <p(f) is the
Therefore, f(co) will have1 a positive islope in this limit interatomic potential, M the atomic mass, and D the
1.0 <
diffusion coefficient.
T»85.5 - K
III. APPLICATION TO LIQUID ARGON
We have evaluated the velocity autocorrelation
0.5 - function and its frequency spectrum for liquid argon at
85.5 and 94.4°K, using the data of Rahman.1-4 These
* ( tt)
) are <a>2)av=50X1024 sec~2 [This value should be pre-
ferred to the other values reported in Ref. (1) (A.
0.1 ^^^c 8 - Rahman, private communication).]] and T>— 1.88X10""5
0 Mo o o o o o o^J-o-o- 0-0 cm2 sec-1 at 85.5°K, and <a>2)av=55X1024 sec~2 and
-0.1 - £>=2.43X10-6 cm2 sec"1 at 94.4°K. We have also
\ A //
investigated the effects of small changes (by ±10%)
i I - in these numerical values. The results are presented in
(k B T/f»)t Figs. 1 to 4, together with the results obtained in the
numerical experiments of Rahman.1,4,5
FIG. 1. Velocity autocorrelation in liquid argon at 85.5°K.
Curve A: <a>2>av = 50X1024 sec"2, Z?=1.88XH)- 6 cm2 sec"1. Curve 4
B : <co2)av = 45Xl0 24 sec"2, Z)=1.88X10- 5 cm2 sec"1. The circles A. Rahman, Phys. Rev. 136, A405 (1964).
5
are from B. R, A. Bijboer and A. Rahman ? Ref. 1. A, Rahman, J. Chem. Phys. 45, 2585 (1966),
 157 VELOCITY AUTOCORRELATION IN CLASSICAL FLUID 155

The present approach clearly reproduces semi-

quantitatively some of the features of Rahman's
results. Thus the velocity autocorrelation function
becomes negative at approximately the correct time and
has a well-defined minimum; it reaches, however, a
deeper negative value and undergoes some additional
oscillations of small amplitude. Small changes in the
parameters do not affect these features markedly. The
theoretical frequency spectrum has a peak in approxi-
mate correspondence to the shoulder in Rahman's
spectrum, and the high-frequency tail of the spectrum
is also approximately reproduced. The low-frequency
peak in Rahman's spectrum is missed entirely. Small
changes in the parameters affect markedly the height 0 0.5 1.0 1.5
of the peak and, to a lesser extent, its position and the fiw/(kBT)
high-frequency tail. A decrease in the diffusion coef-
FIG. 4. Frequency spectrum of velocity autocorrelation in liquid
ficient or in the second moment of the spectrum argon at 94.4°K. Curves A and B : data as in Fig. 2. The broken
heightens and narrows the peak in /(co) and strengthens curve is from A. Rahman, Ref. 4.
the oscillations in $(*); this behavior is consistent with
the analysis given in Sec. II. lator with a very small damping. For large values of
Our results at 94.4°K, for the values of the parameters B/Af(u>) decreases continuously from its value at zero
quoted above, can be compared with the results obtained frequency, but $(/) still undergoes some strongly
by Berne et al.,3 at the same temperature and for the damped oscillations. Substantially larger values of B/A
same values of the parameters, but with an exponential are necessary to completely eliminate the oscillations.
kernel. The curves for <£(/) and the broad features of
/(co) are essentially the same. The peak in /(co) obtained
IV. CONCLUDING REMARKS
with a Gaussian kernel is, however, lower and more
spread out. While the present approach clearly represents a
We have also evaluated $(t) and /(co) for values of substantial improvement in the description of the
B/A just outside the bounds (11). For small values of velocity autocorrelation over the use of the Langevin
this ratio /(co) has a very high and narrow peak, and picture or of a truncated time expansion, it is also clear
<£(/) undergoes many slightly damped oscillations; the that the use of a simple memory function oversimplifies
particle is clearly being described as an Einstein oscil- the problem. The motion of a particle in a liquid involves
a diffusive component as well as a damped vibratory
component; the former contributes mostly to the low-
frequency part of the spectrum while the latter is ex-
pected to lead to a peak in the region of the shoulder in
Rahman's spectrum. 5 ' 6 By fitting the parameters of the
kernel to the diffusion coefficient and to the second mo-
ment of the spectrum one is taking into account both
components in an approximate manner. Our comparison
with Rahman's results suggests that the present ap-
proach is actually giving more weight to the second
component, and may, therefore, be describing the
liquid as more "solidlike" than it really is. A better
description could be developed in this framework by
using a more complicated kernel, at the expense of
increasing the number of parameters in the model.

ACKNOWLEDGMENT
f>cu/(kBT)
We would like to thank A. H . Lent for programming
FIG. 3. Frequency spectrum of velocity autocorrelation in liquid the calculations.
argon at 85.5°K. Curves A and B : data as in Fig. 1. Curve C:
(co 2 ) av =50Xl0 24 sec"2, Z) = 2.07X10- 5 cm2 sec"1. The broken
6
curve is from A. Rahman, Ref. 5. K. S. Singwi, Physica 31, 1257 (1965).