Page 1
,
.. ;I
PHYSICAL REVIEW A
VOLUME 31, NUMBER 3 ? MARCH 1985
Canonical dynamics: Equilibrium phase-space distributions
William G. Hoover
Department 0/Applied Science, University o/California at Davis-Livermore, Livermore, California 94550
(Received 18 September 1984)
Nose has modified Newtonian dynamics so as to reproduce both the canonical and the
isothermal-isobaric probability densities in the phase space of an N-body system. He did this by
scaling time (with s) and distance (with Vl/D in D dimensions) through Lagrangian equations of
motion. The dynamical equations describe the evolution of these two scaling variables and their two
conjugate momenta Ps and Pv. Here we develop a slightly different set of equations, free of time
scaling. We find the dynamical steady-state probability density in an extended phase space with
variables x, Px, V, Ii, and S, where the x are reduced distances and the two variables Ii and Sact as
thermodynamic friction coefficients. We find that these friction coefficients have Gaussian distri
butions. From the distributions the extent of small-system non-Newtonian behavior can be estimat
ed. We illustrate the dynamical equations by considering their application to the simplest possible
case, a one-dimensional classical harmonic oscillator.
I. ?INTRODUCTION
where the p' are the scaled momenta p Is. Thus the Ham
iltonian (1) generates the canonical probability distribution
independent ofthe values chosen for HNose and Q.
During
the canonical-ensemble
described, the volume Vand temperature T are held fixed.
Nose demonstrated the usefulness of these ideas by carry
ing out several dense-fluid simulations using the Hamil
tonian HNose'
By allowing length to vary,1 as well as time, Nose gen
eralized this work to include the isothermal-isobaric en
semble. These methods and ideas forge a remarkable link
between the ensembles of statistical theory and atomistic
dynamics. They suggest promising approaches for the in
vestigation of nonequilibrium systems.
Here we exhibit steady-state (equilibrium) distributions
for the new variables which play the role of thermo
dynamic friction coefficients. Our equations of motion
are very much like Nose's, but differ in that scaling of the
time is not required. The new results for distributions
make it possible to estimate finite-size effects on dynami
cal averages. In Sec. II we review Nose's canonical equa
tions of motion and introduce a version of them free of
time scaling. In Sec. III we formulate the phase-space
evolution
of the many-body
fNVT(q,P,r,;,Q) and exhibit a steady-state solution. We in
dicate the straightforward extension to include the isobar
ic case. With some additional effort, it seems likely that a
stress-tensor version of this ensemble could be constructed
along the lines pioneered by Rahman and Parrinello.9 In
the final section we illustrate the equations of motion with
some representative trajectories for a single classical oscil
lator.
Classical "constant-temperature" calculations have been
pursued for over a decade. 1,2 In this sense, "temperature"
is a measure of the instantaneous kinetic energy in a sys
tem. Thus the corresponding dynamical equations include
non-Newtonian accelerations designed to keep the kinetic
energy ~ p 2 12m constant.
thermal accelerations are useful in dissipative systems in
volving viscous flow, or heat flow, far from equilibrium.
Such systems would heat rapidly in the absence of con
straints. By now, many3-6 distinct sets of differential
equations of motion have been devised to keep the kinetic
energy constant. ?
A somewhat different kind of constant-temperature cal
culation strives to reproduce the canonical phase-space
distribution, so that the kinetic energy can fluctuate, with
a distribution proportional to exp( - ~ p 2 12mkT). Ob
taining the canonical distribution is desirable, at least in
equilibrium work, in order to correlate ?the results of
many-body simulations with Gibbs's and Jaynes's statisti
cal mechanics. Andersen7 has used occasional discontinu
ous "stochastic" collisions to induce the canonical distri
bution in many-body simulations.
Nose achieved a major advance by showing that the
canonical distribution can be generated ? with smooth,
deterministic, and time-reversible trajectories. To do this
he introduced a time-scale variable s, its conjugate
momentum p.. and a parameter Q. Nose's augmented
Hamiltonian8
calculations just
The non-Newtonian iso
probability density
'
HNose=<P(q)+ ~ p 2 / 2 m s 2
+(X + 1)kTlns +p}12Q ,
(1)
II. CANONICAL DISTRIBUTION FROM
NON-NEWTONIAN DYNAMICS (REF. 10)
contains a nonlinear collective potential in which the
time-scale variable s oscillates. Thus the system, with X ?
degrees of freedom, is coupled to a heat bath (described by
the variables sand Ps)' Nose proved that the microcanon
ical distribution in the augmented set of variables is
equivalent to a canonical distribution of the variables q,p',
The equations of motion from Nose's Hamiltonian (1)
are ?
q=plms 2, p=F(q), s=PsIQ, ?
(2) ?
p s = ~ p 2 I m s 3 - ( X + l ) k T l s . ?
31 1695 ?
© 1985 The American Physical Society
Page 2
1696
WILLIAM G. HOOVER
31
These coupled first-order equations take a simpler form if
the time scale is reduced by s, so that dto1d:::=:S
of the rates given in (2) can then be expressed as deriva
tives with respect to tnew (for which we will estill use the
superior dot notation)
All
q=plms, p
(3)
The somewhat inconvenient variable s can then be elim
inated from the equations (3) by rewriting the coordinate
evolution equations in terms of q, q, and ii:
q=plms (plms)sls=Flm -qpsIQ:::=:F(q)/m -;q .
(4)
The thermodynamic friction coefficient
appears in the second-order equations (4) evolves in time
according to a first-order equation
IQ which
(5)
Nose showed that the phase-space distribution resulting
from the equations (2) is canonical in the variables q,p Is.
In the next section we show that the distribution resulting
from equations (4) and (5) can be made canonical too, and
in such a way as to avoid time scaling. To do this we
redefine p:::=:mq and replace Nose's X + 1 by X obtain
inglO
q=plm, p=F(q)
Berendsen6 has just suggested a close relative of (6) in
which ; rather
tlEkin:::=: ~ p 2 / 2 m -XkT12.
equations are not reversible in time. The equations (6) are
much less severely damped than Berendsen's. An extreme
opposite limiting case, in which tlEkin is identically zero
and time reversibility is retained, has been achieved by set
ting the friction coefficient equal to ( ~ F p 1m)1(»21m)
or, equivalently, by "velocity scaling.,,1-4
than ¢
is proportional to
Notice that Berendsen's
III. PHASE-SPACE EVOLUTION OF fNVT(q,P,t;)
Because the variables q, p, and; used in (6) are in
dependent, we can easily calculate the components of the
flow of probability density f( q,p,;) in (2X + 1)
dimensional space. The equations governing the motion
in this space are not Hamiltonian. Therefore the deriva
tives aq laq and ap lap do not generally sum to zero.
Thus the analog of Liouville's equation, expressing the
conservative flow of probability with time, including flow
in the; direction, is
aflat+qaflaq
a f l a p + ~ a f l a ;
+f[aqlaq+aplap+a¢/a;] O. (7)
Consider a density function hvVT proportional to the fol
lowing exponential:
fNVT 0:: exp [
j<I>(q) +~ p 2 / 2 m +Q;2!211kTI (8)
The non vanishing terms in (7) obtained from this density
function are as follows:
q aflaq (jI k T ) ~ F p l m ,
~
paflap (f I k T ) ~ ( -F+;p)plm,
(9)
¢
(jlkT) [{- ~ p 2 l m
]/Qj;Q,
faplap -XkT;) .
Inspection shows that these terms sum to zero, provided
that the coefficient of kT in the dynamical equation (6)
for the fliction coefficient is chosen equal to the number
of independent degrees of freedom in the set q,p. In the
usual molecular dynamics simulation, with periodic boun
daries, the center of mass and its velocity are fixed so that
this number of degrees of freedom is D
dimensional N-body system. Thus the canonical distribu
tion (8) is a steady equilibrium solution of the flow equa
tion (7) and satisfies the equations of motion (6).
In commenting on an earlier draft of this manuscript,
Brad Holian pointed out that the phase-space distribution
(8) can be used to derive the equation of motion for the
friction coefficient;. To see this, note that the canonical
distribution (8) satisfies (7) if, and only if, ; follows the re
laxation equation (6) of Nose. Thus Nose's canonical
equations of motion are unique. Other relaxation equa
Jions, such as Berendsen's, cannot lead to the canonical
distribution (8).
To extend these ideas to the isothermal-isobaric case is
straightforward. Reduced coordinates x:::=:q IVl/D are in
troduced, as is also a fixed "external pressure" Pex! and
relaxation time T. The equations of motion
1) for a D
x=plmV1/D,p F (E+;)p,;Q
~ p 2 I m - X k T ,
(10)
E= V IDV, i:=' (P
W/?kT,
have
0:: VN -1 exp( - 11'1kT), where
1I':::=:<I>(x VIID) +~ p 2 / 2 m +Q;2/2
the steady equilibrium
solution
f'VPT
2 ~ k T I 2 + P e x t V .
(11)
IV. CANONICAL HARMONIC OSCILLATOR
To illustrate the changes in viewpoint discovered by
Nose we consider a one-dimensional harmonic oscillator
with the mass, force constant, and initial values of
p an taken to be unity. We consider equations for which
the values of and p2 have averaged values of unity.
The microcanonical equations of motion
and
q=p, p -q
(12)
generate closed elliptical trajectories in the two
dimensional qp phase space. See Fig. 1(a). For this same
oscillator Nose's canonical equations [with X in (2) taken
to be zero and s initially unity] take the form
q =p Is 2, P -q, S=PsIQ, Ps
lis .
(13)
Page 3
31
CANONICAL DYNAMICS: EQUILIBRIUM PHASE-SPACE ...
1697
(e )
FIG. 1. (a) Elliptical orbit for an oscillator described by
(12). The abscissa is q, the ordinate is p. The major-to-minor
axis ratio of unity has been increased in plotting to fit the Tek
tronix hard-copy screen area symmetrically. This same illerease
applies to each figure. All data were obtained on the Digital
Equipment Corporation VAX 11/780 computer at the Physics
Department (Lausanne) using a fourth-order Runge-Kutta in
tegration in double precision with time steps in the range 0.01
down to 0.001. (b) Long-time qp trajectory for Eqs. (13) or (14j
with initial values q
1, p
1, s = 1, ps =,0, and Q
Same as (b) with Q =0.1. (d) Long-time qp trajectory for
(15) with initial values q
1, p = 1, ;=0, and Q 1. (e) Same
as (d) with Q=0.1.
1. (c)
For large Q these equations simply reproduce the micro
canonical behavior shown in Fig. Ha). In Figs. l(b) and
Hc) we show trajectories for Q = 1. 0 and 0.1 using the
same initial conditions. For the larger Q, the trajectories
in qp space gradually fill in a region between two limiting
curves. For the smaller Q the trajectories develop more
nearly singular turning points and the size of the filled re
gion diminishes. When a new time is introduced, with
and
dtold
lW. T. Ashurst and W. G.Hoover, Phys. Rev. Lett. 31, 206
(1973).
2L. V. Woodcoek, Chern. Phys. Lett. 10,257 (1971).
3D. J. Evalls, J. Chern. Phys. 78, 3297 (1983),
4\"\,. G. A. J. C. Ladd, and B. Moran, Phys. Rev. Lett.
48,1818 (1982).
5J. M. Haile and S. Gupta, J. Chem. Phys. 79, 3067 (1983),
6See H. J.. C. Berendsen, .J. P. M. Postma, W. F. van Gunsteren,
A. DiNola, and J. R. Haak, J. Chern. Phys. 81, 3684 (1984l.
q=pls, p= -qs, s=sPsIQ, Ps
(p
1 ,
(14)
exactly the same trajectories are produced, but at different
rates, This is a good check of the numerical integration,
Finally, if we abandon time scaling and redefine p =q
we have
q=p, p= -q-Sp, ~ = ( p 2 _ 1 ) I Q .
(15)
Solutions for these equations appear in Figs. l(d) and l(e).
The small-Q limit of (15) can be inferred from these fig
ures. The oscillator moves between widely-separated turn
ing points at velocity ± 1.
These examples illustrate that a single oscillator is not
sufficiently chaotic to reproduce the canonical distribu
tion from a single initial condition. The trajectories are,
however, stable and cover a relatively large part of the os
cillator phase space for reasonable values of the parameter
Q. For unreasonable values of Q (either very small or
very large) it is not at all clear that even large systems will
behave in a canonical (as opposed to microcanonical) way.
A study of the number dependence and Q dependence of
the phase-space density for a series of small systems
might help to clarify this point.
ACKNOWLEDGMENTS
It is a pleasure to thank Shliichi Nose and Carl Moser
for stimulating conversations at the 1984 Centre Europeen
de Calcul Atomique et Moleculaire (Orsay, France)
Workshop on Constrained Dynamics. Professor Nose
kindly made several comments, correcting and clarifying
a previous version of this manuscript. Professor Philippe
Choquard kindly provided local support and facilities for
this work at Laboratoire de Physique Theorique, Ecole
Poly technique Fecterale de Lausanne, Switzerland. The
Academy of Applied Science supported related work at
the University of California at Davis-Livermore as well
as the cost of travel between California and Europe. This
work was partially supported by the Lawrence Livermore
National Laboratory under the auspices of the U.S.
Department of Energy under Contract No. W-7405
ENG-48.
For an oscillator this "new" approach gives (Lord) Rayleigh's
equation [Philos. Mag. 15,229 (1883)].
iH. C. Andersen, J. Chern. Phys. 72, 2384 (1980) and references
quoted therein.
8S. Nose, Mol. Phys.
255 (1984).
9M. Parrinello, A. Rahman, and P. Vashishta, Phys. Rev. Lett.
SO, 1073 (1983) and referenees quoted therein.
lOS. Nose, J. Chern. Phys. 81, 511 (1984), See Sec. II B for equa
tions equivalent to (2)--(6).
Download full-text