The Nose–Hoover thermostat

D. J. Evans and B. L. Holian

Citation: The Journal of Chemical Physics 83, 4069 (1985); doi: 10.1063/1.449071
View online: http://dx.doi.org/10.1063/1.449071
View Table of Contents: http://scitation.aip.org/content/aip/journal/jcp/83/8?ver=pdfcov
Published by the AIP Publishing

Articles you may be interested in

The Nosé–Hoover looped chain thermostat for low temperature thawed Gaussian wave-packet dynamics
J. Chem. Phys. 140, 194106 (2014); 10.1063/1.4875517

Decomposition-order effects of time integrator on ensemble averages for the Nosé-Hoover thermostat
J. Chem. Phys. 139, 064103 (2013); 10.1063/1.4817194

Crooks equation for steered molecular dynamics using a Nosé-Hoover thermostat

J. Chem. Phys. 125, 164101 (2006); 10.1063/1.2360273

Solving the NoséHoover thermostat for Nuclear Pasta

AIP Conf. Proc. 841, 558 (2006); 10.1063/1.2218236

A configurational temperature Nosé-Hoover thermostat

J. Chem. Phys. 123, 134101 (2005); 10.1063/1.2013227

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
132.204.37.217 On: Mon, 01 Dec 2014 17:53:18
 The Nose-Hoover thermostat
Denis J. Evans and Brad Lee Holian 8 )
Research Schoolo/Chemistry, Australian National University, G.P. O. Box 4, Canbe"o, ACT. 2601, Australia
(Received 4 June 1985; accepted 10 July 1985)
We derive equilibrium fluctuation expressions for the linear response of many body systems
thermostated by the Nose-Hoover thermostat. We show that in the thermodynamic limit this
response is the same as that of the corresponding Gaussian isothermal system. Numerical
comparisons for shear flow show that the calculated shear viscosity coefficient is remarkably
independent of thermostatting method.

I. INTRODUCTION analyze theoretically, and they work equally well for equilib-
Recently there have been revolutionary developments rium and nonequilibrium systems. These methods are often
in molecular dynamics methodology. In its 30 year history called Gaussian methods since Gauss' Principle of Least
the molecular dynamics technique has remained almost un- Constraint can often be used to generate the equations of
changed. Newton's equations of motion are solved for a sys- motion.
tem of particles subject to periodic boundary conditions. Recently a number of alternative methods have been
Time averages for such a system represent fairly closely mi- proposed. The "momentum scaling" technique proposed by
crocanonical ensemble averages. The microcanonical en- Haile and Gupta cannot thermostat nonequilibrium systems
semble is however very inconvenient both for theoretical and even for equilibrium systems it contains a number of
analysis and also for comparisons of computed results with difficulties. s,6 The recent algorithm proposed by Berendsen7
experimental data. By comparison it was appreciated fairly is difficult to analyze theoretically since the Berendsen equa-
early that Monte Carlo simulations could be performed in a tions of motion are not time reversible.
wide variety of more useful ensembles. At the time it was In this paper we analyze some of the properties of a
generally thought that if one was interested in the equilibri- dynamics originally proposed by Nose. 8 ,9 In their original
um properties of a system where the pressure, enthalpy, or form these equations were very cumbersome to implement in
temperature (rather than the energy) were the preferred state actual molecular dynamics computer programs. Hoover 1o
variables, then one had no choice but to use Monte Carlo has recently cast Nose's equations in a form which is closely
techniques. related to the Gaussian schemes mentioned earlier.
Of course people had been cheating for some time. The principle differences between Nose-Hoover (NH)
Many people rescaled the second moment of the atomic ve- dynamics and the earlier Gaussian schemes is that NH dy-
locities to control the temperature of their molecular dynam- namics allows fluctuations in the state variables whereas the
ic simulations. 1,2 Moreover the results of these ad hoc rescal- Gaussian schemes rigidly clamp these variables. To date all
ing algorithms produced reliable results provided the analysis of the NH equations has been confined to their equi-
timesteps used were sufficiently short. In 1981 Anderson librium properties. In this paper we will consider the dyna-
developed a hybrid Monte Carlo molecular dynamics algo- mical implications of the NH equations.
rithm3 which was capable of generating both the canonical
and isothermal/isobaric ensembles. The stochastic nature of II. NOSE-HOOVER DYNAMICS
the Andersen algorithms made them difficult to analyze
We will present a short outline of the Nose equations of
theoretically-particularly for dynamical properties.
motion and their transformation to the more useful noncan-
In the meantime a number of workers had been using
onical Nose-Hoover form. We believe that some of the cru-
molecular dynamics to simulate nonequilibrium systems in-
cial points of this derivation have not been very explicit in
cluding nonequilibrium steady states. For these problems,
previous discussions.
standard molecular dynamics was more than merely incon-
Consider the Nose Hamiltonian H N ,
~en~ent, it was incapable of creating homogeneous nonequi-
bbnum steady states. The solution to this far more difficult HN = Ho{q,p/s) +gkTlns + (p~/2Q), (1)
problem was found simultaneously and independently by where Ho is the usual Hamiltonian for a classical many body
~oover and ~vans. 2,4 Their solution was to rewrite the equa- system except that everywhere one would normally expect to
tions of classical mechanics so that thermodynamic varia- see a momentum p appearing, it is replaced by pis. Thus
bles other than the energy, were constants of the motion.
These methods are deterministic, simple to implement and Ho(q,p/s) = L (p;ls)2/2m + <I>(q). (2)

The microcanonical ensemble for such a system is described

by the partition function

')Permanent addres~: MS B221, Los Alamos National Laboratory, Los

Alamos, New MeXIco 87545.
z= J dqdpdp. ds8(HN -E) (3)

J. Chern. Phys. 83 (8), 15 October 1985 0021-9606/85/204069-06$02.10 @ 1985 American Institute of Physics 4069

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
132.204.37.217 On: Mon, 01 Dec 2014 17:53:18
 4070 D. J. Evans and B. L. Holian: The Nose-Hoover thermostat

and the probability that the system will be observed at a

particular point in phase space dq, dp, dps' ds, will be uni-
iL'
-=-
1 a [p;
, a
_o-+(F;-sp;)o-
s s m aq; ap;
form provided that the Nose Hamiltonian takes the specified
valueE. a (p,2
+ss-+ I-' -gkT ) - a ] . (11)
The Nose equations of motion are as maps
it; =p;lms2, Using Eqs. (5), (6), and (7) and the fact that
p;=F;,
(4) i.)
as
=i.)
as
p' p
+p~o~)
ap; s'
I
(12)
s=PsIQ,
we find that
. = ~ p; _ gkT iL 'Is = iL. (13)
Ps ~mSJ s·
Substituting into the equation for the propagator [Eq. (10)]
Hoover has recently shown that these equations can be cast
we see that the NH p propagator is identical to the corre-
in a simpler, more easily understood form by transforming to
sponding Nose propagator
the scaled momentum
(14)
p; = p;fs (5)
Trajectories under NH and Nose dynamics follow the same
and to scaled time
paths but they traverse these paths at different (nonuniform)
rates. Scaled time trajectories viewed from the perspective of
t'= EdtiS. (6)
Nose dynamics, evolve under non-Galilaean time. This has
This same transformation had been used by Nose. However important consequences for evaluating the full NH partition
he did not take full advantage of the resulting simplifications function.
to the equations of motion. To evaluate the new partition function we will first
We should also point out that both Nose and Hoover transform to the scaled momentum p; leaving time in its
transformed Ps to p; = p.ls. However as we shall see this original Nose (Galilaean) form. The Nose partition function
additional transformation is an unnecessary complexity. for a three-dimensional system becomes
With this minor difference in mind we shall define a friction
(15)
constants,
s=P.lQ· (7)
We'will now consider the transformation to non-Galilaean
In scaled momentum p' and scaled time t ' the equations of
time. Although we are at this stage only concerned with time
motion become
independent, equilibrium distribution functions the effect of
introducing non-Galilaean transformations of time is most
dq;fdt' = p;lm , easily seen by considering its influence on an equilibrium
dp;ldt' = F; - sp; , (8) time average. If ( )' denotes an equilibrium t ' -time average
then as Nose pointed out, 8
ds
dt'
=J..
Q
[I p;2 -gkT]
m (A)'= ;, iT'dt'A(t'), (16)

with the subsidiary equation of motion for s (if needed) which Using Eq. (6) this becomes
is
( dtA (t')/s
dsldt' = ss. (9) (A )' = _J_oo--=_ __ (17)
Equation (9) is subsidiary to (8) since it is not needed to com- iT dt Is
pute the trajectories of the N interacting particles. We shall
refer to Eqs. (8) as the Nose-Hoover (NH) equations of mo-
tion. However since the propagators are identical [Eq. (14)], we
We must emphasize that the introduction of "scaled find that
time" t " has important and unusual implications for the sta- (A )' = (A Is)/(1/s) . (18)
tistical mechanical relationships of the Nose and NH sys- Equation (18) shows that the partition function in scaled
tems. The meaning of scaled time can be better understood time Z' must be
by considering the transformation of propagators in the two
dynamical systems. The propagator of phase variables in
NH dynamics is
Z'= f dqdp'dPsdsS3N-ltJ(HN-E). (19)

. .. =exPR ItiL' Integrating the irrelevant variable s using the fact that
elL t -dt, (10)
o s tJ[F(s)] = .5(s - so) (20)
where eXPR is the time ordered exponential with latest times F'(s) ,
to the right [see the discussion in Ref. lOla)]. From Eqs. (8) where F(so) = 0, we find that

J. Chern. Phys., Vol. 83, No.8, 15 October 1985

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
132.204.37.217 On: Mon, 01 Dec 2014 17:53:18
 D. J. Evans and B. L. Holian: The Nose-Hoover thermostat 4071

Z'=fdqdP'ds.!J..-exp 3N [E-Ho-~QS2]. (21) @: 0.40

gkT gkT ~ 0.35
From Eq. (21) we see that g = 3N and Q = 3NkTr where T i=
U 0.30 p'= 0.7
is some arbitrary time constant which governs the rate at Z

which the equations of motion damp out fluctuations in the

it0.25
1""=1.0
N=32
Z ,'= 0.09622
00.20
kinetic energy.
In our derivation we have considered the microcanoni- 15i= 0.15
cal ensemble rather than the molecular dynamics ensemble. a: 0.10
~

We have not included considerations of momentum conser- en MEDIAN

Ci
vation. It is easy to see that l',Pi is a constant of the motion
0.05

o -3 -2 -1
t 0 3
generated by the Nose Hamiltonian. That P; is not a constant
of the motion can be seen from the equivalent NH equations
of motion. FIG. 1. This figure shows a comparison of predicted and observed distribu-
In what follows we shall refer only to the NH form of the t.
tions for The system studied was an equilibrium soft disk fluid.
equations of motion. Henceforth, we will refer to P; as Pi'
The NH equations of motion can be written in momentum
One of the results which is central to the proof is the
conserving form as
following. Consider a set of phase variables {Ua } whose
qi = (Pi - po)lm , equilibrium averages vanish
Pi = Fi - S(Pi - Po), (22) (26)

t=~[I (Pi:PO)2 -3(N-l)kT],

We will be interested in sets oflocal, extensive variables
Ua = 0 (N). To be local and extensive we mean that firstly it
where Po = l',p;IN in which case the partition function is is possible to decompose Ua intoN formally identical inten-
easily seen to be sive terms Uai ,

(23) (27)

and secondly we require that

with A A
(Uai UfJj>c = 0 for all a,/3, ifrij >rc , (28)
Q = 3{N - l)kTr (24)
where rc is some cutoff distance such as the range of the
and intermolecular potential.
Ho =I (Pi - Po)2 + ~(q) . (25)
If properties (26), (27), and (28) are satisfied then
2m
In Fig. 1 we show computer simulation results for the
distribution function p(5). The system studied was 32 soft
spheres l,6(r) = 4e(alr)12 truncated at ria = 1.5. The state
point simulated was kT I € = 1.0, p£i3 = 0.7. The relaxation
time T [Eq. (24)] was taken to be 1"(€/(m~)1/2 = 0.09622. Examples of zero mean, local, extensive variables are the
This value was chosen because it is approximately equal to fluctuations in the internal energy or the fluctuations in
the Maxwell relaxation time (or "collision" time) Tc for the PV-the pressure tensor times the volume. Equation (29)
system. Figure 1 shows that within statistical uncertainties S shows that averages of products of such variables are exten-
is distributed in a slightly skewed Gaussian fashion with a sive. The result can be generalized to include mUltiple time,
variance which agrees with the predicted value [Eq. (23)]. correlation functions.
The thermodynamic properties of the NH simulation were
also found to agree with estimates obtained using standard (30)
Newtonian molecular dynamics or Gaussian isothermal
methods (see Sec. IV for further details).
The only difference in the proof being that the cutoff dis-
tance rc must be increased to rc + c Max {t a } where c is the
III. NONEQUILIBRIUM PROPERTIES OF NH DYNAMICS sound speed.
A. Equilibrium time correlation functions We will now use this result to compare equilibrium time
correlation functions computed under Nose-Hoover and
We will now show that in the absence of external fields Newtonian dynamics.
(Fe = 0) eqUilibrium time correlation functions computed by Let
using the NH propagator or the Newtonian propagator are
equivalent in the thermodynamic limit. Our method is essen- (31)
tially the same as that used by Evans and Morrissll,12 to denote a canonical ensemble time correlation function com-
show the thermOdynamic equivalence of Gaussian isother- puted under Newtonian dynamics. Let C (tNH) denote the
mal and Newtonian equilibrium time correlation functions. corresponding function computed under Nose-Hoover dy-

J. Chem. Phys., Vol. 83, No. 8,15 October 1985

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
132.204.37.217 On: Mon, 01 Dec 2014 17:53:18
 4072 D. J. Evans and B. L. Holian: The Nose-Hoover thermostat

namics [Fe = 0 in Eq. (8)]. We will drop the use of primes It is convenient to define the dissipation functional J
when describing NH momentum Pi and rescaled NH time t '. from the adiabatic derivative of the internal energy Ho:
For such correlation functions to be of interest the phase
H~ = J(r)Fe , (38)
variables A, B can be taken to be extensive but with zero
mean. In most problems of interest they will also be local
variables. Thus C(tN) and C(tNH ) are both extensive
J= L(: oD; -F; oc;).
[= o (N)]. We assume that at t = 0 a Nose-Hoover ensemble char-
The difference between the two correlation functions acterized by the N-particle distribution
IiC (t) can be computed using the Dyson decomposition 11,12 2
of propagators as 1= exp -P(Ho +! Qs ) (39)

IiC(t) = l' ds(A/L,.I.t-S)Ii/L"'B) + o (1i 2 ) , (32)

J Jdsexp-p(Ho+~Qs2)
dI'

is suddenly subject to a constant external field, Fe. We shall

where Ii, the difference between the respective Liouvilleans compute a formal expression for the N-particle response of
is from Eq. (8) seen to be the system. Now
a .a
Ii = - s L Pi 0 api + S as . (33) I(t) = e-il'lo, (40)
where il is the f-Liouvillean:
Normally phase variables of interest will not have any ex-
plicit dependence upon S and since the Newtonian Liouville ilA =(to~+t~)A +A (~ot+~t)
operator is also independent ofS' the last term in Eq. (33) can ar as ar as
be ignored. =iLA +AA. (41)
IfB is an analytic function of momenta Pi' it is clear that
L is called a p or phase Liouvillean and A is called the phase
l:Pi alapi will transform intensive variables into new in-
0
space compression factor since
tensive variables. Similarly it transforms extensive variables
into new extensive variables. Thus, (-1I/)dlldt=A.

IiC(t) = - l' ds (A (SN - tN)sB'(SN)c> (34)

The main trick that is required to convert Eq. (40) into a
more useful form is to use of the Dyson equation to express
the I propagator in terms of p propagators. Morriss and
where B ' is some new extensive variable. Evans recently proved the resu1t 12 that
From the equations of motion [Eq. (8)] and Eq. (24) for Q
we see that S is intensive. This means that IiC (t ) is 11N times
the average of a product of 3 zero mean
e - ilt = exp [1' ds A( - S)] e - iLt . (42)

(A) = (B') = (s) = 0), local extensive variables. Employ- Substituting this propagator transformation into Eq. (40)
ing our lemma (30) concerning such products we see that and using the fact that
IiC(t)/C(tN) = o (1IN) . (35) iL [Ho +! Qs 2 ] =JFe - 3NkTs (43)
we find
B.Responsetheory
We will now consider the response of an NH system to (44)
an external field Fe. For simplicity we will assume that the
external field is time independent. The equations of motion This is the Kawasaki form 13 of the nonlinear response of
are the N-particle distribution function. Equation (44) for the
Nose-Hoover response is the same formal expression as that
qi = pJm + Ci Fe , for both the adiabatic and the Gaussian isothermal response.
p=Fi +DiFe -SPi' (36) The only difference between the three cases is due to differ-
t= (KIKo-I)/r. ences in the propagator used to generate J ( - s). In the pres-
ent case it is the field dependent Nose-Hoover propagator.
This is the external field form of the Nose-Hoover equations
Linearizing Eq. (44) and taking the average of an arbi-
ofmotionforl:Pi = O. The equation of motion fors has been
trary zero mean phase variable B yields the usual Kubo
obtained by substituting Eq. (24) into Eq. (22). The preset
form l4 [Eq. (14)] for the linear response
value of K is Ko = It = 3NkT 12. The variables C,D are
phase variables which describe the coupling of the external
field to the system. As usual we assume the adiabatic incom-
(B (t) = l' ds X(s)F, (45)

pressibility of phase space or AIf'. Thus where

a +-·D·
L-·C. a =0. (37) x(t) =P (B(t)Jo) . (46)
aqi api I I
In Eq. (46) the propagation is derived from the field free NH
A sufficient but not necessary condition for AIr is the exis- equations of motion (8). Thus the susceptibility X(t) is given
tence of a Hamiltonian which could generate the adiabatic by an eqUilibrium time correlation function. The ensemble
(S = 0) equations of motion. average is to be taken over the NH equilibrium ensemble

J. Chern. Phys., Vol. 83, No.8, 15 October 1985

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
132.204.37.217 On: Mon, 01 Dec 2014 17:53:18
 D. J. Evans and B. L. Holian: The Nose-Hoover thermostat 4073

(23). This result is ergodically consistent because in an ergo- A practical difficulty with the Berendsen equations is
dic system time averaging of a single-phase trajectory will that they are very difficult to employ in nonequilibrium sim-
generate the entire equilibrium ensemble average. ulations at a specified state point. In the steady state the
Furthermore we know from Sec. III A and Eq. (35) in dissipation JFe [Eq. (38)] must balance the energy removed
particular, that in the thermodynamic limit there is no dif- by the thermostat 2KssSss, thus
ference between the susceptibility generated by NH dynam-
ics and the corresponding ensemble of Newtonian trajector-
ies. This means that in the thermodynamic limit the
(J)ssFe = (! Kss (~: - 1)) . (49)

adiabatic (Newtonian), isothermal (Gaussian isothermal) Thus only at eqUilibrium when Fe = 0 does To actually
and the Nose-Hoover linear responses are all equivalent. equal the temperature. In spite of these obvious difficulties
This is of course not true in the nonlinear regime. we thought that a comparison with the Berendsen scheme
might be useful in a practical demonstration of the essential
IV. NUMERICAL COMPARISONS OF NON-NEWTONIAN independence of the computed properties to the details of the
MOLECULAR DYNAMICS ALGORITHMS thermostating mechanism.
The system studied was the same as that employed in
In Secs. II and III we have seen that NH dynamics
our earlier analysis of the S distribution at equilibrium. The
yields correct predictions of the thermodynamic and linear
32 soft spheres were sheared using the now standard SHod
transport properties of many-body systems. We decided to
algorithm. IS We have recently proved that under adiabatic
employ the NH thermostat to nonequilibrium molecular dy-
conditions this algorithm correctly describes both the linear
namics simulations of planar Couette flow. This would test
and nonlinear properties of Couette flow arbitrarily far from
the usefulness of the method in calculating the nonlinear
equilibrium. IS
transport properties of systems.
If we assume that a linear velocity profile is stable then
The method was tested by comparing it with the Gaus-
thermostating is achieved in any of the various schemes by
sian isothermal, Gaussian isoenergetic and Berendsen
recognizing that
schemes. The Gaussian schemes are well known from pre-
vious work and will not be further described here. The Ber-
(50)
endsen method is very new and so we will give a very brief
description. provides a convenient definition of the temperature. In Eq.
It is basically a first order version of the NH scheme. (50) Y is the shear rate au)Jay. The incorporation of Gaus-
The equations of motion are sian thermostats into the Sllod equations for shear flow have
q=p;/m, been described many times before. 12
(47) Table I shows a comparison ofthe thermodynamic and
Pi =Fi -SPi'
transport properties for 32 soft spheres at a state point kT I
s=( K1Ko-l)/r. (48) € = 1.0, pcf3 = 0.7, yoiml€) I 12 = 1.0. It is important to rea-
The preset temperature To is obtained from Ko = 3NkT012. lize that this state point is well into the Non-Newtonian re-
The first thing to note about the Berendsen equations is that gime with a shear viscosity which is only 71 % of its Newto-
in contrast to all the Gaussian schemes and to NH dynamics, nian (y-+O) value. The hydrostatic pressure is approximately
it is irreversible. This leads to profound theoretical difficul- 5% greater than at equilibrium. Consequently the linear
ties when one tries to analyze either the equilibrium or non- analysis of Sec. III is not useful in the theoretical interpreta-
eqUilibrium properties of Berendsen dynamics. As far as is tion of these results.
known to the authors the form of its equilibrium distribution Table I shows results for NH simulations using three
function is unknown as is the form of its linear response different values for r ranging from approximately ten Max-
susceptibility. It is easily shown however that in contrast to well relaxation ("collision") times (re) down to rJ3. For
all the Gaussian schemes and NH dynamics, that the canoni- r < re the equations of motion become stiff, requiring very
cal distribution function is not a solution of the Berendsen short integration time steps. NH dynamics becomes very
equations of motion. inefficient for smallr. From the table we see that for the run

TABLE I. This Table shows the comparative abilities of a number of different thermostats in maintaining shearing steady states. All units are reduced by the
standard molecular parameters. The columns from left to right are: temperature, internal energy, shear viscosity, second moments of the internal energy and
kinetic energy, the run length in 1000's oftime steps and the integration time step. Note the greater efficiency of the two Gaussian methods. The reader can
find a relative comparison of thermostats as applied to equilibrium systems in the review by Evans and Hoover (Ref. 16).

Method kTIE E/E llcr(mE)I/2 (~E/E)2 (ME/E)2 kts ~(E/mif)

Gaussian isothermal 1.0 147.6 ± 0.1 1.524 ± 0.007 49.3 ± 10 0.0 50 0.003
Gaussian isoenergetic 0.996 ± .002 147.38 1.537 ± 0.008 0.0 27.6 ± 2 40 0.003
NH *lmif)I/2 =1 1.0001 + ± 0.0002 147.4 ± 0.2 1.55 ± 0.02 108.9 ± 2 60± 3 50 0.003
NH =0.09622 0.998 ± 0.001 147.2 ± 0.2 1.56 ± 0.02 110 ± 4 52 ± 1 40 0.003
=0.0333 0.998 ± 0.001 147.7 ± 0.4 1.56 ± 0.04 83 ± 3 43 ± 1 73 0.001
Berendsen Ts = 0.78 1.006 ± 0.003 148.4 ±0.05 1.56 ± 0.03 62±30 30± 10 60 0.003

J. Chern. Phys., Vol. 83, No.8, 15 October 1985

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
132.204.37.217 On: Mon, 01 Dec 2014 17:53:18
 4074 D. J. Evans and B. L. Holian: The Nose-Hoover thermostat

is approached-in particular for T ~ Tc (thermostating times

larger than the equilibrium collision time)-the kinetic
crossing rate is linear in the thermostating rate, with the
10 adiabatic limit extrapolated to be to/tx -6; to is the usual
time unit (mcrl€)1/2 (by comparison, to/2Tc -5). For ther-
y. - 1.0
p. - 0.7
mostating times shorter than the Maxwell time, equations of
T· - 1.0 motion get stiffer and the kinetic crossing rate increases in a
N - 32 nonlinear fashion. There is some indication that as T--o,
t x- I approaches a new linear regime with a higher slope of
1/2I1r. In the ideal gas limit, harmonic analysis of the NH
equations of motion [Eqs. (36)] yields t x-I = (V2hr)T-I. In
a sense, such a limit is ach)eved when T--o, since the mo-
FIG. 2. This figure shows the mean kinetic crossing rate ToITx as a function
of the NH thermostating rate T 01T. The system studied was the same as in
menta of particles change radically while their coordinates
Fig. 1 except that the fluid was subject to a reduced shear rate of unity. The jiggle about essentially frozen positions. Hence, the magni-
solid curve gives the observed results while the dashed curve gives the re- tude of <i> compared to k is relatively small most of the time.
sults of a harmonic analysis of the corresponding ideal gas. It is interesting to note, however, that the T--o limit of the
NH equations of motion does not produce the isokinetic, or
with T(€lmcr)I/2 = 0.0333, in spite of the fact that this run Gaussian thermostated results, since the NH kinetic energy
consumed the most computer time (73 000 time steps), it fluctuations remain canonical (i.e., the kinetic heat capacity
resulted in the poorest statistics for the computed shear vis- per particle is still ~ k ), rather than approaching the isokine-
cosity. NH dynamics will also become inefficient r is made tic value, zero. This behavior is entirely analogous to the
too large since this will slow down the canonical averaging. inability of classical statistical mechanics to give correct low
From the table one can see that the Gaussian methods temperature specific heats. In classical statistical mechanics
are more than twice as accurate as any of the other methods each degree of freedom will remain equipartitioned no mat-
which means that if all the methods were required to yield an ter how high the characteristic frequencies become.
accuracy of 112%, both of the Gaussian methods could In conclusion, we have shown that the Nose-Hoover
achieve this target in less than 114 of the computer time thermostat generates the canonical ensemble at equilibrium
required by any of the other methods. and gives correct linear nonequilibrium results which agree
All methods yield, within statistical uncertainties the with other methods of thermostating. It also appears that in
same value of the shear viscosity. At present we have no nonequilibrium steady states far from equilibrium, each of
theoretical understanding of why the properties of a nonlin- the thermostating algorithms discussed here yield essential-
ear system should be so robust with respect to thermostating ly identical numerical results. We have no theoretical expla-
mechanism. nation for this. It also seems that of the methods studied
Finally, we were interested in investigating the effect of here, the Gaussian isothermal method provides more effi-
the Nose relaxation time T, upon kinetic energy relaxation in cient calculations of transport coefficients than the other
nonequilibrium simulations such as shear flow. It is clear methods. The Nose-Hoover and Berendsen algorithms al-
from the foregoing discussion that Gaussian and NH ther- low the effects of thermostating to be extrapolated out of
mostating yield the same time averages and linear susceptibi- nonequilibrium molecular dynamics simulations so as to
lities for variables which are not constants of the motion. It is yield underlying adiabatic rates.
not so clear, however, what the effect of temperature regula-
tion will be upon kinetic energy relaxation. In Gaussian iso-
thermal dynamics the kinetic energy is of course a constant
of the motion. We might expect that in NH dynamics that
the kinetic energy relaxation time will be roughly propor-
tional to T. It also seems plausible that this relaxation time IL. V. Woodcock, Chern. Phys. Lett. 10,257 (1971).
20. J. Evans, J. Chern. Phys. 78, 3297 (1983).
will be shorter the further we move away from equilibrium. lH. C. Andersen, J. Chern. Phys. 72, 2384 (1980).
We decided to examine the kinetic relaxation rate 4W. G. Hoover, A. J. C. Ladd, and B. Moran, Phys. Rev. Lett. 48, 1818
Tk I(T- I ), as a function of NH thermostating rate T- I • In (1982).
particular can the kinetic relaxation rate be extrapolated sJ. M. Haile and S. Gupta, J. Chern. Phys. 79, 3067 (1983).
60. J. Evans and G. P. Morriss, J. Chern. Phys. 81, 3749 (1984).
back to its adiabatic value Tk 1(0), where the thermostating 7H. J. C. Berendsen, J. P. M. Postma, W. F. van Gunsteren, A. Oi Nola,
has been turned off? A more convenient statistic to gather in and J. R. Haak, J. Chern. Phys. 81, 3684 (1984).
the NH-thermostated steady-shear experiments is the kinet- 8S. Nose, J. Chern. Phys. 81,511 (1984).
9g. Nose, Mol. Phys. 52, 255 (1984).
ic mean crossing time tx defined as total trajectory time di- lOW. G. Hoover, Phys. Rev. A 31, 1695 (1985).
vided by the number of times that the kinetic energy K lOaB. L. Holian and O. J. Evans, J. Chern. Phys. (to be published).
crosses its mean, or preset value K = Ko; tx is related to the 110. J. Evans and G. P. Morris, Chern. Phys. 87,451 (1984).
correlation time by tx - 2TK' 12G. P. Morriss and O. J. Evans, Mol. Phys. 54, 629 (1985).
13T. Yamada and K. Kawasaki, Prog. Theor. Phys. 38, 1031 (1967).
In Fig. 2 we show the mean kinetic crossing rate t x- I, as 14R. Kubo, J. Phys. Soc. Jpn. 12,570 (1957).
a function of thermostating rate T- I for the NH-thermostat- ISO. J. Evans and G. P. Morriss, Phys. Rev. A 30, 1528 (1984).
16
ed shear flow experiments. Notice that as the adiabatic limit 0. J. Evans and W. G. Hoover, Annu. Rev. Fluid Mech. 18,243 (1985).

J. Chern. Phys., Vol. 83, No. 8,15 October 1985

This article is copyrighted as indicated in the article. Reuse of AIP content is subject to the terms at: http://scitation.aip.org/termsconditions. Downloaded to IP:
132.204.37.217 On: Mon, 01 Dec 2014 17:53:18