PHYSICAL REVIE% A

VOLUME 3S, NUMBER 2

JANUARY 15, 1987

Application of transient correlation functions to shear flow far from equilibrium

Gary P. Morriss and Denis

J. Evans

Research School of Chemistry, Australian National University, Canberra ACT2601 Australia

(Received 25 March 1986; revised manuscript received 4 August 1986)

Morriss and Evans recently developed a generalization of the Green-Kubo relations which is valid
for nonequilibrium steady states far from equilibrium. This formalism relates the nonequilibrium
response to transient time correlation functions which connect the nonequilibrium steady state to the
equilibrium state. In the linear regime, the transient time correlation functions reduce to simple
equilibrium Green-Kubo relations. The transient time correlation function method thus provides a
long-sought-after fundamental relation between nonequilibrium molecular dynamics algorithms and
the Green-Kubo formalism which is only valid close to equilibrium. In this paper we demonstrate
the use of the transient time correlation function formalism for isothermal planar Couette flow.
The results show that the nonlinear steady-state response can be calculated by integrating the appropriate transient response time correlation function. In particular, the nonlinear shear stress and
pressure calculated in this way agree with the values calculated directly.

I. INTRODUCTION
The Green-Kubo relations' are fundamental
to our
understanding of linear transport processes close to equilibrium. For linear transport they play a role analogous to
that played by the partition function in equilibrium statistical mechanics. Like the partition function, G-reenKubo relations are highly nontrivial to evaluate. They do
however provide an exact starting point for approximate
theoretical treatments. They can also be evaluated directMore
ly in equilibrium molecular dynamics simulatioris.
recently they have been used to develop efficient nonequilibrium molecular dynamics algorithms for the calculation of linear transport coefficients. A major limitation
for nonequilibrium statistical mechanics has been the lack
of a corresponding theoretical basis for nonlinear transThis has not been
port processes far from equilibrium.
for the lack of trying. However, essentially all previous
attempts have either resulted in incorrect expressions for
nonlinear transport, or have led to expressions which although formally exact, are nonetheless very difficult to
analyze and interpret.
In 1979 Dufty and Lindenfeld developed a readily interpretable expression for the nonlinear response which
took the form of an integral of a time correlation function. Unlike the Green-Kubo expressions, this correlation
function involved correlating the fluctuations of microscopic quantities in the initial equilibrium (or local equilibrium) state with the corresponding values of these quantities during the establishment of the nonequilibrium state
(the transient response). This transient correlation function can be constructed from an ensemble of trajectories,
which at the initial time are equilibrium states, but are
then propagated forward in time with a field-dependent
tranthe Dufty-Lindenfeld
Unfortunately
propagator.
sient correlation functions can easily be seen to diverge.
This is because in the absence of a thermostat, there can
be no steady state. The thermodynamic properties of the
35

system continue to change in time.

Recently Evans, Morriss, and Holian have made major
of the thermostatted
improvements in our understanding
behavior of nonequilibrium
systems. These developments
have their origins in the formulation by Hoover and
Evans ' of the so-called Gaussian thermostat. This thermostat is implemented by a change in the equations of
motion of 1V-particle systems so that the heat produced irreversibly in nonequilibrium
systems is continuously and
instantaneously removed. Although the original formulation of the thermostat was ad hoc, its validity as a means
of studying nonequilibrium steady states has since been
established. Evans and Morriss derived the equilibrium
X-particle distribution function for an isolated system
evolving under Gaussian isokinetic equations of motion.
They proved that in the thermodynamic limit, equilibrium
time correlation functions computed under Gaussian
dynamics, are the same as the corresponding correlation
functions computed under Newton's equations of motion.
Later it was verified that if one computes the linear
response of Gaussian thermostatted systems to an external
is
field, one does indeed find that the susceptibility
governed by an equilibrium time correlation function of
Green-Kubo form but with time propagation generated by
the field-free Gaussian isokinetic propagator.
This, combined with the earlier result showing the equivalence of
time correlation
Gaussian and Newtonian equilibrium
functions, shows that to linear order in the external field
the adiabatic and thermostatted responses are identical. '
have been-derived for
A corresponding series of results
the Nose-Hoover thermostat.
In this paper we give a simplified derivation of the
transient correlation function expressions for the thermostatted nonlinear response of X-particle systems to planar
Couette flow. We also show how the transient time correlation formalism is related to the isothermal generalizafor the nonlinear
of Kawasaki's expression'
tion,
response. We test the validity of the transient correlation

"

792

1987

'

The American Physical Society

..

APPLICATION OF TRANSIENT CORRELATION FUNCTIONS.

35

theory

by

using

molecular dynamics
nonequilibrium
of planar Couette flow, far from

(NEMD) simulations
equilibrium.

equivalence of the Schrodinger and Heisenberg representations. ' For time in-dependent external fields, differentiating the Heisenberg form with respect to time we obtain

d(8 (t) )

II. TRANSIENT CORRELATION FUNCTIONS

793

M(t)

FOR PLANAR COUETTE FLOW

'

that the equations of motion for an

It is well known'
X-body system undergoing planar Couette flow can be
written as,

=p;/m+nyy;,
pi
&x'YES
pi =Fi
q;

'pi'

'Ypx~ pr

)'

direction, y is the
of particle i meathe streaming velothermostat. When

(2)

the peculiar kinetic energy is a constant of the motion. If

we take the peculiar kinetic energy per degree of freedom
as a measure of the temperature, then the Gaussian thermostat maintains the system at a constant temperature.
The transformation
in (1) of the boundary condition
which drives planar Couette flow, into the form of an
external mechanical perturbation, is achieved by writing
the equations of motion in terms of the peculiar momenta
These equations (1), are known as the Sllod equap;. '
tions of motion for shear flow. '
Isothermal response theory shows that the adiabatic
rate of change of the internal energy Ho
X(p;/2m +4'),
plays an essential role in determining the thermostatted
response. The adiabatic derivative of the internal energy
is the dissipaJI', where
dHO/dt is usually written as
tive flux and I', is the external field. Strictly speaking
this is only true when AII (Refs. 8 and 9) is satisfied.
AII is satisfied by all the commonly used NEMD algorithms including the Sllod equations of motion for
Couette flow. If AII is not satisfied, the form of the
For planar
response equations is modified somewhat.
is easily seen to be the shear stress P
Couette flow
times the system volume V:

'

" "' +y;F=

rn

= JF, = yg

yPV .

Consider a canonical ensemble (or isothermal ensemble ' ) of systems suddenly subject to a constant shearing
deformation at t =0. The 2V-particle distribution function at some later time r is (r, t), where I denotes the
6X-dimensional position of the system in phase space
I =(q, p)=(x, y, z, p,p, p, ). (Each of the components
x, y, z, p,p,p, is itself an N-dimensional vector. ) The
average value of a phase variable 8 (I ) at time t is,

(B(t))= drB(I )f(t)=

where

the

second

the result that

This is only true
since in that cirno explicit time
and hence commutes with its associated
Integrating (5) by parts we see that,

ps

Ho

Eq. (5) we have used

dB(t)/dt =iL exp(iLt)8=exp(iLt)iLB.
if the external field is time independent
cumstance the Liouvillean iLcontains
dependence,
propagator.

where n is the unit vector in the x

strain rate, and p; is the momentum
sured in a coordinate frame moving at
city yy;. The term exp; is the Gaussian
o, is chosen to be

g (Ft

In deriving

equality

f drf(0)8(r;t),

is a consequence

(4)

of the

dt

= drB(t)

BI

rf(o) .

(6)

The boundary term vanishes because the distribution

function
(0), approaches zero as the magnitudes of the
momenta become infinite, and
(0) can be taken to be
periodic with respect to the coordinates. We are explicitly
using the simulation periodic boundary conditions in the

derivation.
Integrating (6) with respect to time we see that the nonlinear nonequilibrium response can be written as,

f f sr 8 (s)

(B(t) ) = (8(0) ) 'as

rf (O) .

The dynamics implicit in 8 (s) is of course driven by the

full field-dependent,
thermostatted equations of motion.
For a system subject to the thermostatted shearing deformation, I is given by Eqs. (1) and (2). If f(0) is the
Gaussian isokinetic or the canonical distribution function
it is straightforward to show that (8/Bl ) [I (0)]
is PVPf(0). Thus (8(t)) becomes,

( 8 (t) ) = (8 (0) ) py

ds

(8 (s)P(0) ) .

(8)

value of a
relates the nonequilibrium
phase variable B at time t, to the integral of a transient
time correlation function (the correlation between P in
the equilibrium starting state, P(0), and 8 at time s
after the field is turned on). The time-zero value of the
transient correlation function is an equilibrium property
of the system. For example, if B =P~, then the timezero value is (P (0)). Under some, but by no means all
circumstances, the values of B(s) and P(0) will become
uncorrelated at large times. If this is the case, the system
is said to exhibit mixing. The transient correlation function will then approach (8 (r) ) (P~(0) ), which is zero be (P~(0)) =0. For systems that exhibit mixing, Eq.
cause
(8) can therefore be rewritten as,

This expression

(8(t)) =(8(0)) pyv

f 'ds(bB(s)P(0)),

(9)

where,

&8(s)=8(s) (8(s)) .
The adiabatic systems treated by Dufty and Lindenfeld
cannot exhibit mixing because in the absence of a thermostat d(8(t))ldt does not go to zero at large times.
Thus the integral of the associated transient correlation
function cannot converge. Other systems which are not
expected to exhibit mixing are turbulent systems or sys-

GARY P. MGRRISS AND DENIS J. EVANS

tems which execute quasiperiodic oscillations.
The derivation of our fundamental Eq. (8), is simpler
than that given in Ref. 8. In that paper equation (8) was
derived from the Kawasaki form of the nonequilibrium
distribution functjon. The derivation depended upon a
lemma which gives an exact relationship between distribution function propagators and their adjoints, phase variable propagators.
It is also interesting to note that the
present derivation is given in the Heisenberg picture rather than the customary Schrodinger picture. '
We can use recursive substitution
to derive the
Kawasaki form for the nonlinear response from our transient time correlation formula, Eq. (8). The fact that we
can derive the transient and Kawasaki forms of response
from each other shows
theory
the fundamental
equivalence of both representations.
The first step in the
derivation of the Kawasaki representation is to rewrite the
transient correlation relation using iI. to denote the phase
variable Liouvillean,
and i W to denote its nonHermitean adjoint, the distribution function Liouvillean.
Thus dB/dt =iLB and df/dt i Wf. Using this notation Eq. (8) can be written as,

(B(t))= drBf(0)
P
yV

f f dI f(0)e' 'Be
ds

f dI Bf(0)
f f dI [e
13yV

ds

'

'L'J

(10)

'f(0)]Be '~'J .

Equation (11) can be written more simply as,

(B(t) ) =

f dI Bf(0)
Pyv f f dI B(0)J(
ds

s)f(s
).

(12)

We can now successively substitute the transient correladistribution function expression for the nonequilibrium
tion function namely,

f(t)=f(0)
into the right-hand

f dI Bf(0) PyV f s, f dI B(0)J(s, )f(0)

+(PyV) f ds, f
f dI B(0)J(s, )J( )f(0)+. .
= f dI B(0)exp Py f J( s) f (0)

(B(t)) =

35

PyV

ds
J(

s)f(s)

side of (12). This gives,

ds

'

ds

(14)

This is precisely the isothermal Kawasaki form of the

nonlinear response.
Comparing the transient time
correlation expression for the nonlinear response with the
Kawasaki representation, we see that the difference simply amounts to a time shift. In the transient time correlation form it is the dissipative flux J, which is evaluated at
time zero, whereas in the Kawasaki form the response
variable B, is evaluated at time zero. For equilibrium or
steady-state time correlation functions the stationarity of
averages means that such time shifts are essentially trivial.
For transient response correlation functions there is of
course no such invariance principle, consequently the time
translation transformation is accordingly more complex.
The evaluation of the time-dependent response using
the Kawasaki form directly, Eq. (14), is numerically unstable. The inevitable errors associated with the inaccuracy of the trajectory, as well as those associated with the finite grid size in the calculation of the integral itself, combine and are magnified by the exponential. This exponential is then multiplied by the phase variable B(0), before
the ensemble average is performed. In contrast the calculation of the response using the transient correlation expression, Eq. (9), is, as we shall see, far more stable.
It is trivial to see that in the linear regime the transient
correlation function expression for the system response
reduces to the usual Green-Kubo expression. The GreenKubo expression for the linear response is identical in
form to the corresponding transient correlation relation
except that the time evolution in the second term on the
right-hand side of (8) is governed by the field-free ther-

mostatted propagator in the linear case and by the thermostatted field-dependent propagator in the nonlinear regime. The Green-Kubo correlation function is an equilibrium correlation function. We know ' that in the thermodynamic limit there is no difference between the thermostatted equilibrium time correlation function and the
corresponding Newtonian or adiabatic correlation function.
The coincidence at small fields of the Green-Kubo and
transient correlation formulas means that unlike direct
NEMD, the transient correlation method can be used at
small fields. This is impossible for direct NEMD because
in the small-field limit the signal-to-noise ratio goes to
zero. The signal-to-noise ratio for the transient correlation function becomes equal to that for the equilibrium
Green-Kubo method
a small, but nonzero number.
Thus at small fields the efficiency of the transient correlation method is comparable to that of Green-Kubo. The
transient correlation function method forms a bridge between the Green-Kubo method which can only be used at
equilibrium, and direct NEMD which is the most efficient
strong-field method.
It is also easy to see that at short times there is no
the linear and nonlinear
stress
difference between
response. It takes time for the nonlinearities to develop.
The way to see this is to expand the transient time correlation function in a power series in yt. The first term in
this series for the response of the shear stress is just
V(I'z ) /kT, the infinite frequency shear modulus, G
Since this is an equilibrium property its value is unaffect-

'

APPLICATION OF TRANSIENT CORRELATION FUNCTIONS. . .

35
'ed by the strain

rate and is thus the same in both the

linear and nonlinear cases. If we look at the response of a
quantity like the pressure whose linear response is zero,
the leading term in the short-time expansion is nonlinear
in the strain rate and in. time. The linear response of
course is the first to appear.

III. PHASE-SPACE MAPPINGS

In this paper we use nonequilibrium molecular dynamics of shear flow in simple fluids to illustrate the transient
correlation function formalism. The initial ensemble of
starting states was generated from a single field-free
Gaussian isokinetic trajectory. At equally spaced intervals along this single field-free trajectory, shearing simulations were started and followed for approximately 200
time steps. The response calculated from the integral of
the transient correlation functions was compared with the
direct response computed by averaging (B(t)). In practice the calculation proceeded by performing an equilibrium isokinetic molecular dynamics simulation for X, time
run of X time steps.
steps followed by a nonequilibrium
Each of these cycles gives one initial phase
for the
transient correlation function. The number Nshould be
greater than the characteristic time required for the system to relax to a steady state. X, is chosen to ensure that
the initial phases are sufficiently uncorrelated. Here we
used X, =X=200. This process can be made more efficient if we use the equilibrium starting state generated by
X, time steps, to provide more than one initial phase for
the nonequilibrium trajectories.
We note that given an initial
starting
phase
I (x, y, px, py) then new starting phases, which occur
within the equilibrium distribution with the same probability as the I ~, can be obtained using the following mappings:

There is a second, more important, advantage of this

procedure. If we examine the transient response formula
(8), we see that at long time the correlation function becomes (B(ao))(P (0)). The steady-state average of
8 is usually nonzero. This is not the case for equilibrium
time correlation functions. To minimize the statistical
uncertainties in calculating the transient correlation integral, it is convenient to evaluate (Py(0)) in a constrained ensemble in which Q, Pxy(I. ;) is identically zero.
The phase mapping procedure described above achieves
this. If the shear stress, computed from the original starting phase is Py(I &), then the shear stress of I 2 is also
equal to P y(I ~), but the shear stresses of both I 3 and I'4
are equal to Py(I &). This means that the sum of the
shear stresses of these four starting phases is exactly zero.
The statistical difficulties at long time, associated with a
small nonzero value the average of P (0), are eliminated.
These mappings can be trivially generalized to three dimensions, and as the z coordinate does not appear in I'~
we obtain the same four starting phases and precisely the
same advantage is gained.

I,

&

I 2 (x, y, p, py),
~3

x&y~

I 4=(y~px~

Px~py )

I.

IV. NUMERICAL RESULTS

To test the validity of the transient correlation function

formalism we carried out a series of molecular dynamics
simulations of shear flow in two and three dimensions.
We used a second-order Runge-Kutta method to integrate
the equations of motion with a time step of 0.005t in
two dimensions,
and 0.004 t * in three dimensions
[t*= t(o (m /e)' ) ']. The Runge-II utta method was
used rather than the more common Gear or leap-frog algorithms because unlike the latter, the Runge-Kutta
method is self-starting.
Since we are interested in the
transient response we need to be able to calculate the system trajectories from a specified initial phase. The Gear
and leap-frog algorithms only achieve accuracy after an

py)

These three new configurations, along with the original

one, give four different starting states and four different
shearing trajectories from the single equilibrium. run of
X, time steps. Each of the mappings consists of a pair of
reflections in a coordinate or momentum axis. The mapping which produces I"2 for example, is obtained by reflecting in both the x and y axes. The resulting particle
configurations
are then imaged by the usual periodic
boundary conditions. The overall effect of this mapping
is to rotate the simulation cell through 180. In total there
are 2 states that can be obtained using the reflections of a
two-dimensional
phase space; however, only 2 of these
states will result in at most a sign change in the shear
stress. Only 2 of the remaining mappings lead to dif'ferent shearing trajectories. The shearing trajectories obtained from
and are identical. The probability of
each of these states occurring within the equilibrium distribution of states, is identical because the Hamiltonian
Hp is invariant under these reflections and inversions of

I;

I;

795

2D Soft Disks
0

0
D
Y

-2

=0. 1

= 0.5

= I-0

GK

-D

xy

GK

FIG. 1. The direct (D) and transient time correlation func{T) results for the shear stress in two dimensions at
y*=0. 1, 0.5, and 1.0. A typical error bar is included on the

tion

time correlation function result at t*=1. For

comparison we include (GK) the, Green-Kubo equilibrium estimate for the response
PyV ds
(P(s)P(0)) evaluated at
y*=0.5 and 1.0.

y*=1 transient

GARY P. MORRISS AND DENIS

796

initial startup period and are therefore unsuitable for calculating transient responses.
In each simulation the direct average of the shear stress,
pressure, normal stress difference, and thermostatic multiplier a, were calculated along with their associated transient correlation functions using typically 2000
3000
starting states. In Fig. 1 we present the results obtained
for the shear stress P~~=P (o /e) in two dimensions.
The two-dimensional
simulations used a system of 896
soft disks [/=4'(o lr)' ], at a density of p* =pa
=0.9238, a temperature, T*=kT/@= 1, and a range of
values of the strain rate, y*=y(m/e)'~ o. The potential
was truncated at r*=(r/cr) =1.5. The values of the shear
stress calculated from the transient correlation function
expression agree within error bars, with those calculated
directly. The errors associated with the direct average are
less than the width of the lines on the graph, whereas the
error in the integral of the transient correlation function is
approximately 15%.
The curves labeled GK give the linear response, as calculated from the Green-Kubo relations. The GK curves
thus give the response of a hypothetical Newtonian fluid
whose thermophysical properties exactly match the thermodynamic and linear transport properties of this soft
disk fluid. The reduced zero shear rate shear viscosity,
g*=rio (me) '~, at this state point is 3.8. When the
time integration is truncated at t*=1, the Green-Kubo
estimate is 3.7+0.2. The statistical uncertainties
of
Green-Kubo and transient correlation calculations of
equivalent length are similar. Both are substantially less
accurate, at these strain rates, than direct NEMD calculations. At a reduced strain rate of one we see that the
steady-state shear stress is reduced to approximately two
thirds of the value it would have if the fluid was comThis is reflected in both the direct
pletely Newtonian.
NEMD results and in the response predicted by the transient time correlation functions. They are in good statistical agreement. It is also clear that this fluid exhibits
stress overshoot, a well-known property of non-Newtonian
fluids. Stress overshoot is apparent both in the direct
NEMD results and in the transient time correlation func-

J. EVANS

35

tion predictions for y* of 0.5 and 1.

At the lower shear rate, y*=0. 1, the response is closer
to the linear prediction given by the Green-Kubo relations
and there is no evidence of stress overshoot. Indeed the
linear response is indistinguishable
from the transient
correlation
function result.
The presence of stress
overshoot indicates that for y &
1, the transient time
correlation function, the derivative of the response, is negative at long times. This has never been observed in the
equilibrium stress autocorrelation function.
The shear-induced increase of pressure with increasing
strain rate is known as shear dilatancy. It is an intrinsically nonlinear effect and is not observed in Newtonian
fluids nor is it predicted by the linear Green-Kubo theory.
In Fig. 2 we present the direct and transient correlation
function values of the difference between the pressure
p*=p(cr /E) and its equilibrium value (bp"=p*
po).
The Green-Kubo predictions are of course zero because
the equilibrium correlation function, (, bp(t)P&(0)) is exactly zero at all times. The agreement between the direct
average, and the value obtained from the transient correlation function expression at y =0.5 and 1.0 is impressive.
At y' =0. 1, we are in the turnover regime' where shear
dilatancy has all but vanished. It is important to note
that the agreement between theory and simulation shown
in Fig. 2, is a test of the predictions. of the theory for an
entirely nonlinear effect. It is a more convincing check on
the validity of the transient correlation formalism than
are the results for the shear stress because there is no
underlying linear effect.
simulation was performed for
The three-dimensional
the Lennard- Jones fluid ( P = 4m[( o /r) '
(o /r ) ] ), at itstriple point (p* =0.8442, T' =0.722). The simulations
employed 108 atoms and a potential cutoff at i =2. 5o.
The results are presented in Fig. 3. The agreement between the direct average and the transient correlation
function expression is again very good for both the shear
stress and the pressure.
Although it is not easy to see in the figures, the initial

-0.

3D Lennard-Jones

1.5

2D Soft Disks
P

1.0

=1.0
0.8

0.5

= 0.5
= 0. 722

xy

FIG. 2, The direct (D)

and transient time correlation function ( T) results for the pressure in two dimensions at y* =0. 1,
0.5, and 1.0.

FICi. 3. The direct (D) and transient time correlation function (T) results for the shear stress and pressure in three dirnensions at y*=1.0. Note the positive y axis is the reduced pressure while the negative y axis is the reduced shear stress.

APPLICATION OF TRANSIENT CORRELATION FUNCTIONS. . .

35

slope of the pressure response curves is zero. This contrasts with the initial slope of the shear stress response
which is
. This is in agreement with the predictions
of the transient time correlation formalism made in Sec.
II. Figures 1 and 3 clearly show that at short time the
stress is controlled by linear response mechanisms.
It
takes time for the nonlinearities to develop.
Comparing the statistical uncertainties of the transient
correlation and direct NEMD results shows that, at the
fields studied in these calculations, direct NEMO is about
2 orders of magnitude more efficient than the transient
correlation function approach. On the other hand, the
transient correlation method has a computational efficiency which is similar to that of the equilibrium Green-Kubo
method. It appears that the efficiency of the transient
correlation function method decreases slightly with in
creasing field strength, although further calculations are
required to confirm this observation. This is to be contrasted with direct NEMD whose efficiency increases
with increasing field. However, direct NEMD has zero
efficiency at zero field.

V. CONCLUSION
We have outlined and tested a new formalism for
analyzing the nonlinear response of many-body systems.
In the past, many people have conjectured that nonlinear
transport coefficients, or at least their derivatives with
respect to the perturbing field, might be related to steadystate time correlation functions in much the same way
that linear transport coefficients are related to equilibrium
fluctuations. The present formalism does not satisfy this
implicit wish. Instead we have shown that the nonlinear
response is related in a very simple way to correlations between the initial equilibrium state and the transient which
establishes the steady state. The simulation studies support the correctness of the resulting expressions.

~D. A. McQuarrie, Statistical Mechanics (Harper and Row, New

York, 1976).

D. J. Evans

and G. P. Morriss, Comput. Phys. Rep. 1, 297

(1984).
J. W. Dufty and M. J. Lindenfeld, J. Stat. Phys. 20, 259 (1979).
4W. G. Hoover, A. J. C. Ladd, and B. Moran, Phys. Rev. Lett.
48, 1818 (1982).
5D. J. Evans, J. Chem. Phys. 78, 3297 (1983).
D. J. Evans and G. P. Morriss, Phys. Lett. 98A, 1695 (1985).
7D. J. Evans and G. P. Morriss, Chem. Phys. 87, 451 (1984).
G. P. Morriss and D. J. Evans, Mol. Phys. 54, 629 (1985).
9D. J. Evans, in Proceedings of the 1985 Enrico Fermi Interna

tional School of Physics (North-Holland, Amsterdam, 19861.

0D. J. Evans and B. L. Holian, J. Chem. Phys. 83, 4069 (1985).
S. Nose, Mol. Phys. 52, 255 (1984); J. Chem. Phys. 81, 511
(1984); W. G. Hoover, Phys. Rev. A 31, 1695 (1985).
T. Yamada and K. Kawasaki, Prog. Theor. Phys. 38, 1031
(1967).

797

In contrast with many previous approaches, the transient time correlation formalism does not require the use
of ad hoc convergence factors as in Zubarev's method, '
nor does it require the use of difficult to interpret or
analyze projection operators. ' lt is also the first of any
of these methods to be tested directly, either by comparison with experiment or with simulation data.
At high fields, y
1, the transient correlation function
method is clearly not as efficient as the direct NEMD. At
low fields, however, this situation will be reversed. This is
because at low fields the transient time correlation functions reduce to Green-Kubo correlation functions. Although these functions are difficult to calculate with high
accuracy, they can be calculated. On the other hand,
direct NEMD calculations cannot be performed at zero
field. There is therefore a critical-field value below which
the transient correlation function approach is more efficient and above which direct NEMD will be more efficient. The transient time correlation approach should
therefore prove useful in attempts to settle the controver'
sy, over the form of the low-field, asymptotic nonlinear
response. Previously the only method which was applicable to very low (but nonzero) applied fields was the differential trajectory method.
The Kawasaki representation for the response does not appear to be directly useful
in computer simulations.
The most important aspect of the transient time correlation formalism is likely to be theoretical rather than
computational, however. It is possible that the transient
time correlation form of the nonequilibrium
X-particle
distribution function can be used to form a nonequilibrium partition function. Indeed, some preliminary results
are already available. Evans ' has recently used the transient time correlation function expression for the internal
energy of the steady state to derive a theoretical expression for the specific heat of a nonequilibrium steady state.
Preliminary simulation results
support the validity of
the specific-heat expressions so derived.

3D. J. Evans and G. P. Morriss, Phys. Rev. A 30, 1528 (1984).

A. J. C. Ladd, Mol. Phys. 53, 459 (1984).
B. L. Holian and D. J. Evans, J. Chem. Phys. 83, 3560 (1985).
~6D. J. Evans and G. P. Morriss, Phys. Rev. Lett. 51, 1776
(1983).
D. N. Zubarev, Nonequilibrium Statistical Thermodynamics
(Consultants Bureau, New York, 1974).
' B. Robertson, in Proceedings of the Maximum Entropy Formalism Conference, Massachusetts Institute of Technology,
1978 (unpublished); also see references cited in this paper.
~9See Nonlinear Properties of Fluids, edited by H. J. M. Hanley,
Physica 118A (1984) (special issue).
0G. V. Paolini, G. Ciccotti, and C. Massobrio, Phys. Rev. A 34,

1355 (1986).
2'D. J. Evans, Proceedings of the Fourth Australian

National
So-

Congress on Rheology, edited by R. R. Huilgol (Australian

ciety of Rheology, 1986).
D. J. Evans and G. P. Morriss (unpublished).