You are on page 1of 25

prepared for publication in J. Chem. Phys.

22 February 2000
Extrapolation and perturbation schemes for accelerating the
convergence of quantum mechanical free energy calculations via the
Fourier path-integral Monte Carlo method
Steven L. Mielke

Environmental Molecular Sciences Laboratory, Pacific Northwest National Laboratory,


Mail Stop K8-91, Richland, Washington 99352
Jay Srinivasan and Donald G. Truhlar
Department of Chemistry, Chemical Physics Program, and Supercomputer Institute,
University of Minnesota, Minneapolis, Minnesota 55455-0431
We present two simple but effective techniques designed to improve the rate of
convergence of the Fourier path-integral Monte Carlo method for quantum partition
functions with respect to the Fourier space expansion length, K, especially at low
temperatures. The first method treats the high Fourier components as a perturbation, and
the second method involves an extrapolation of the partition function (or perturbative
correction to the partition function) with respect to the parameter K. We perform a
sequence of calculations at several values of K such that the statistical errors for the set of
results are correlated, and this permits extremely accurate extrapolations. We
demonstrate the high accuracy and efficiency of these new approaches by computing
partition functions for H
2
O from 296 K to 4000 K and comparing to the accurate results
of Partridge and Schwenke.
________________

Present address: Department of Chemistry, The Johns Hopkins University, Baltimore,


MD 21218
2
I. INTRODUCTION
Fourier path integral methods
1-14
, especially when coupled with Monte Carlo
integration, provide an efficient means of calculating accurate quantal partition
functionsand hence absolute free energieswithout diagonalization of molecular
Hamiltonians. Thus these methods become increasingly attractive as the system size
increases and diagonalization becomes prohibitively expensive. Unfortunately, Fourier
path integral methods suffer from slow convergence with respect to the maximum
number, K, of Fourier coefficients used in the expansion of the paths. The canonical FPI
method is known to converge as O(1/K).
15
By using the technique of partial averaging
6,7
the convergence rate may be increased to O(1/K
2
).
15
Unfortunately, efficient use of this
technique requires that the potential function have an analytic Gaussian
transformwhich most molecular potential energy functions do not possessor that the
potential be expanded in a Taylor series. The latter approach, known as gradient partial
averaging,
6,7
requires the calculation of the diagonal elements of the Hessian which can
be very expensive in many important cases, especially for systems of large
dimensionality. Additionally, considerable human labor may be involved in coding the
Hessian for a desired potential. In the present paper we introduce two simple, yet highly
effective, techniques for ameliorating the costliness of the slow convergence of FPI
methods with respect to K. Both of these techniques can also be applied in the partial
averaging approach.
We illustrate the new techniques by using them in calculations of the absolute free
energy of water vapor over a temperature range of a factor of 13.5.
3
II. THEORY
II. A. Algorithm overview
We begin by reviewing basic theory and the algorithmic implementation of the
FPIMC method as we have used it.
10-12,14
The partition function for a system with
spinless nuclei may be obtained by calculating the trace of the canonical density operator,
i.e.,
Q T d ( ) ( , ; )

x x x (1)
where is a symmetry number equal to N
i
i
!

where N
i
is the number of atoms with a
given atomic weight,
16
x is an N-dimensional point in mass-scaled Jacobi coordinates,
is 1/k
B
T, where k
B
is Boltzmann's constant,
( , ; ) exp( ) x x x x H (2)
is the coordinate representation of the density operator, and H is the Hamiltonian
operator. Elements of the density operator may be expressed as path integrals
1,2

( , ; ) ( )exp [ ( )] x x x x
x
x


D s dsH s
1
0
h
h
(3)
where h is Planck's constant divided by 2 and

Dx
x
x
( ) s

denotes the summation over all


paths parameterized by s and beginning at x and ending at x.
After substituting Eq. (3) with x = x into Eq. (1), expanding the paths in a Fourier
series,
4
x s x a
k s
j j jk
k
K
( ) sin +
j
(
,
\
,
(

h
(4)
where K is the length of the Fourier expansion, and simplifying, we obtain the expression
Q T
J T
dx da
a
S
j
j
N
jk
k
K
j
N
jk
jk
k
K
j
N
( )
( )
exp ( , )
j
(
,
,
\
,
(
(
j
(
,
,
\
,
(
(

j
(
,
,
\
,
(
(


L
1 1 1
2
2
1 1
2
x a . (5)
In this equation,

jk
k
2
2
2 2
2

h
, (6)
where is the scaling mass of the mass-scaled Jacobi coordinates, S( , ) x a is the action
integral for a given path and is calculated by
S dsV s ( , ) [ ( )] x a x

0
h
(7)
where V(x) is the potential energy, and J(T) is the Jacobian of the transformation from the
integral over paths to the integral over Fourier coefficients and is given by
J T
j
N
jk
k
K
( )

2
1
2
1 1
h
. (8)
We put Eq. (5) into a form appropriate for Monte Carlo integration by multiplying
and dividing by the free particle partition function and restricting the configuration space
to a finite domain D. This yields
5
Q T
Q T
d d
a
S
d d
a
K
jk
jk
k
K
j
N
D
jk
jk
k
K
j
N
D
[ ]
( )
( )
exp exp ( , )
exp

j
(
,
,
\
,
(
(
[ ]
j
(
,
,
\
,
(
(

j
(
,
,
\
,
(
(





fp

x a x a
x a
2
2
1 1
2
2
1 1
2
2
, (9)
where
Q T V
D
j
N
fp
( )
/

j
(
,
\
,
(


2
2
1 2
1
h
(10)
and V
D
is the volume of the domain D. We choose V
D
to be a hyperannulus defined by

l u
where the hyperradius is defined by

x
j
j
N
2
1
. (11)
To improve the efficiency of the calculation we subdivide V
D
into several concentric
hyperannulii and sample these via an adaptively optimized stratified sampling scheme.
We also employ importance sampling in the configuration space, and we have
implemented a number of screening schemes to eliminate the work associated with
evaluating contributions from paths that have a negligible contribution. The discussion
of the new methods we present below does not require a detailed understanding of these
numerical issues so we refer the reader to previously published papers
10-12,14
for further
details.
Partition functions calculated with Eq. (9) include contributions from dissociated
species. These contributions are formally divergent as they scale linearly with the
volume of the system. (This problem is not unique to the FPI method; it always occurs
when one calculates free energies of bound species by using configuration integrals
6
without making the harmonic oscillator approximation.
17,18
) At high temperatures or for
weakly bound systems, the treatment of dissociated species must be carefully defined. In
the present study we wish to compare our results to partition functions obtained from
bound-state eigenvalue calculations; thus we wish to remove the contributions from
species with energies greater than the dissociation energy. This can be done by
neglecting contributions from all paths for which the energy of the configuration space
sample point, x, is above the energy needed for dissociation. (Note that other points on
the path are allowed to have energies above the dissociation limit.) Since it is not
convenient to include the kinetic energy in this criterion, we instead neglect contributions
from paths where the potential energy at x is above the dissociation limit. This removes
nearly all of the effects from dissociated species without excessive cost.
II. B. Perturbative corrections
Partition functions calculated via the FPIMC method converge asymptotically as
O(1/K). This slow convergence rate means that well converged results require large
Fourier expansion lengths, K. It is thus useful to consider the high-k elements of the
Fourier expansion as being a perturbation of the low-K result. The identity
Q T Q T Q T
K K K K [ ] [ ] , ,
( ) ( ) ( ) +
corr
, (12)
where
Q T Q T Q T
K K K K corr, , [ ] [ ]
( ) ( ) ( )

(13)
with K < K, forms the basis for our approach. The first term on the right hand side of
Eq. (12) is large in magnitude but may be affordably calculated, the second term on the
right hand side of Eq. (12) is expensive to calculate but small in magnitude (for
7
sufficiently large K ). Equation (12) will be useful if we can devise a Monte Carlo
integration scheme such that we need to calculate substantially fewer samples of the
expensive Q
corr,K,K
(T) term to achieve a given level of accuracy in Q
[K]
(T) than we
would need if we calculated Q
[K]
(T) directly. If, for instance, Q
[K ]
(T) is converged to
within 10% with respect to the Fourier space size, only 1% of the total number of
samples is needed for Q
corr,K,K
(T) in order to obtain comparable statistical errors for the
two terms.
In order to calculate Q
corr,K,K
(T), we perform simultaneous calculations for K
and K . For each Monte Carlo sample we form two paths from a single set of random
Fourier coefficients. Thus both paths begin and end at the same configuration space
sample point, x, and the first K Fourier coefficients of the higher order path are identical
to those of the lower order path. We accumulate statistics on Q
[K]
(T), Q
[K]
(T), and
Q
corr,K,K
(T) in a single run. Except for statistical errors, the perturbative correction is
calculated exactly by this treatment. We then calculate Q
[K ]
(T) using a substantially
larger number of samples than we use for the Q
corr,K,K
(T) run. We can calculate the
variance of our final result via
var[ ( )] var[ ( )] var[ ( )]
[ ] [ ] ,
Q T Q T Q T
K K K
+
corr
. (14)
The standard deviation, , in any quantity is taken (as usual) as
var/ N (15)
where N is the number of samples used to evaluate the given quantity.
We actually have two calculations of Q
[K ]
(T) from the procedure above, one
using a large number of samples and a less accurate result obtained concomitantly with
the calculation of Q
corr,K,K
(T); when necessary, we will distinguish these two results
8
with superscripts of L and S respectively to denote "large" and "small" samples. The
statistical errors in Q
corr,K,K
(T) and Q
[K ],S
(T) are highly correlated so we can obtain a
correction term of significantly greater accuracy via

( )
( )
( )
( )
, ,
[ ],
[ ],
, ,
Q T
Q T
Q T
Q T
corr K K
K
K
K K

L
S
corr
. (16)
It is difficult to precisely estimate the degree of improvement we obtain by Eq. (16) so
when we quote statistical error bars, we will use the conservative estimate of Eq. (14).
II. C. Extrapolation with respect to K
In the absence of statistical error, the partition function tends to converge
smoothly and monotonically with respect to the size of the Fourier space. It is thus very
tempting to perform calculations for a few moderate values of K, and then extrapolate the
resultstaking advantage of the known asymptotic convergence rateto the K = limit.
The one difficulty with this approach is that the statistical errors for the various
calculations limit the accuracy of the extrapolation procedure. We instead seek to form a
sequence of results that differ in K while having statistical errors that have a high
correlation. We proceed as above by calculating Q
[K]
(T) for several values of K
simultaneously where, for each sample point, all of the paths are determined using a
single common set of random Fourier coefficients. When calculated in this manner, the
sequence of results displays very regular convergence properties with respect to K even
when the statistical errors are very large. Even for a single sample point, results have a
very strong tendency to decrease monotonically with increasing K.
9
Calculating results simultaneously for several values of K in the manner
suggested above is also somewhat less expensive than performing independent
calculations. In general the sets of paths need not intersect except at x, which is not a
quadrature node when we use Gauss-Legendre quadrature. Thus we need the same
number of potential evaluations when doing simultaneous calculations as we would for
independent runs. However, all of the work required to select configuration space
samples (with a distribution governed by an arbitrary importance function) needs to be
done only once. Additionally we need to calculate only a single set of Gaussian random
numbers to sample the Fourier space. It would also be possible (using partitioned matrix
techniques) to save work during the formation of the paths provided that some quadrature
nodes are chosen the same for paths constructed with different values of K. If we sample
the configuration space much more frequently than we sample the Fourier space (i.e., if
we reuse the relative paths a large number of times) these possible savings are modest
since the potential evaluations usually dominate the computational cost. Thus we have
not chosen to exploit this option.
We can apply the extrapolation scheme not only to Q
[K]
(T) itself but also to the
perturbative corrections to Q
[K ]
(T). The relative error of each of the perturbative
corrections is approximately constant so we calculate the K = error bars via
( , , ) ( , , )
, ,
, ,
corr corr
corr
corr

K
Q
Q
K K
K
K K
. (17)
We can also refine the extrapolated correction using Eq. (16).
We perform calculations of Q
corr,K,K
(T) at three or more values of K, and we
extract Q
corr,,K
(T) by fitting to the functional form
10
Q T Q T
A
K
B
K
K K K corr corr , , , ,
( ) ( )

+ +
2
. (18)
II.D. Absolute Free Energies
Absolute internal free energies are obtained from the final partition functions by
G RT Q T ln ( ) (19)
and these may (if desired) be adjusted to any desired standard state by the addition of an
appropriate analytic contribution from overall translation.
III. CALCULATIONS AND DISCUSSION
We will demonstrate the effectiveness of the two new computational schemes on
partition function calculations of water. Water has been widely studied as a prototype for
vibrations in a polyatomic molecule.
19
For our purposes the best set of partition
functions obtained by eigenvalue summation for a given potential energy surface are
those of Partridge and Schwenke.
20
Therefore we have chosen to use their potential
surface and compare to their data. (Another recent calculation of water partition
functions via direct diagonalization has been presented by Harris et al.
21
)
For consistency with prior calculations we use nuclear (rather than atomic)
masses of 1836.152697 and 29148.94642 for H and O respectively. Partridge and
Schwenke
20
tabulate partition functions with a zero of energy set at the ground state
energy of E
0
= 4638.39 cm
1
(whereas we use the minimum of the potential as the zero
of energy) and they include a factor of 4 for nuclear spin effects (which we do not
include). We scale their results by a factor of exp(E
0
)/4 to compare with ours.
11
We performed calculations at ten temperatures in the range 296 K to 4000 K. For
each temperature we perform a calculation using a large number of samples and a modest
value of K (denoted K) using the parameters given in Table I. We then performed
calculations of Q
corr,K,K
(T) for several larger values of K using the parameters given in
Table II and extrapolated the results to the K = limit using Eq. (18). In both cases we
employed an importance function given by
f
R R R R R R
( ) exp x

j
(
,
\
,
(


j
(
,
\
,
(


j
(
,
\
,
(
,

,
,
]
]
]
]

HH HH
e
HH
OH OH
e
OH
OH OH
e
OH

2 2 2
(20)
where the R
XY
e
are taken to be the equilibrium atom-atom distances. The parameters we
used for the importance function of Eq. (20) are given in Table I. The width parameters
were roughly optimized at 4000 K and used without re-optimization at lower
temperatures. The variance reduction due to importance sampling is about a factor of
three at high temperature, but less at lower temperatures; however, the savings in
computer time are less than the naive estimate (which would be the square of the variance
reduction factor) because the screening options
14
are more effective when importance
sampling is not used. A more quantitative discussion of the efficiency savings afforded
by importance sampling is provided elsewhere.
14
Table III presents the results of the two sets of calculations at K = K. Tables IV
and V give the Q
corr,K,K
(T) calculations together with two different extrapolations to the
K = limit. Table VI compares the final FPIMC partition functions to the calculations
of Partridge and Schwenke.
20
Our partition function calculations contain errors from several sources. Statistical
error is likely to be the largest source of error. Our sampling is not entirely uncorrelated
12
since we reuse the same Fourier space samples with many configuration space samples
(100500 for the present set of calculations), but this is a rather short term correlation and
numerical tests
14
indicate that uncorrelated error estimates are accurate. Thus our
statistical error bars are calculated using formulas appropriate for uncorrelated sampling.
Our statistical error bars are not adjusted to reflect the improvements we obtain by using
Eq. (16) so they are likely to be somewhat pessimistic. Errors in the extrapolation to the
K = limit can also be non-negligible. In Tables IV and V we list extrapolations using
two different ranges of K values. The agreement between these two sets of results is
quite good especially given that the largest K value differs by a factor of 1.5 to 2 for each
temperature studied, and the disagreement is comparable to the statistical errors in the
perturbative corrections. Since doubling the values of the largest values of K roughly
halves the contribution to the partition function that must be recovered via extrapolation,
we expect that the observed differences between the results obtained by the two different
extrapolations is comparable to the actual extrapolation error. Quadrature errors in the
numerical evaluation of the action integral and errors from restricting the sampling to lie
within a finite region of space are also present but are expected to be much smaller than
the other errors as we have used very conservative numerical parameters.
The agreement we obtain between the FPIMC partition functions and the results
of Partridge and Schwenke
20
is excellent, with the largest deviation being less than 1%
and seven of the ten deviations being less than 0.2%. At the two highest temperatures the
small deviations we observe are still much larger than the expected errors in the FPIMC
calculations. The calculations of Partridge and Schwenke
20
are likely slightly too low at
the highest temperatures since their partition function calculations were limited to
13
170,625 eigenvalues. At the other eight temperatures the observed deviations are all
comparable to the FPIMC statistical error estimates. (Note that this is strictly a test of the
two methods of calculating the partition function for a given potential; we make no
attempt here to sort out the accuracy of various potentials.)
For the calculations presented here, the computational cost is dominated by the
low-K calculations, i.e., the cost of calculating Q
[K]
(T). At high temperatures (T > 1500
K), extremely accurate results can be obtained at modest cost. At lower temperatures the
calculations become substantially more expensive, and we chose not to converge the
results as tightly. The statistical error at 296 K is a factor of 20 higher than the statistical
error at 4000 K even though we used 200 times as many samples at the lower
temperature. To achieve the same level of accuracy for the results at 296 K would
require about 80,000 times as many samples as were used at 4000 K. The quadrature
requirements also increase at low temperature in proportion with the increasingly large
values of K required to get qualitatively accurate results. Thus the calculations at 296 K
would require about 1 million times as many potential function evaluations to achieve the
same relative error in Q(T) as was achieved at 4000 K.
The extrapolated corrections, Q
corr,K
(T), calculated here range from 2 to 34% of
the magnitude of the accurate partition function. Given these large magnitudes, the high
accuracy we observe is remarkable. In order to achieve the same degree of convergence
without using the extrapolation and perturbative correction schemes presented here we
would require K values that are at least 20 to 100 times as large as the K values presently
used. (This is a very conservative estimate because we are comparing only the truncation
error of a large-K calculation to the total error [extrapolation, statistical, etc.] of the final
14
results of our full calculations.) Since higher K involves sampling more degrees of
freedom and since the quadrature requirements scale linearly with K, the final cost for a
straightforward calculation would increase by at least the same factors. Thus the methods
presented here provide a considerable reduction in the computational cost.
Table VII presents the final internal free energies in order to illustrate the size of
the uncertainty on the energy scale, which is the familiar scale for applications. Clearly
the statistical errors are now smaller than experimental error, and the accuracy of the
calculation is limited almost entirely by the accuracy of the potential energy functions.
The results presented here are especially encouraging because the effort in the
diagonalization method used in Ref. 19 grows as a high power of system size whereas
effort for the path integral approach increases slowly with system size. For triatomic
systems at very low temperatures the diagonalization method requires only a few low
eigenvalues and is clearly the method of choice, but for high temperatures the present
method is considerably less expensive.
IV. CONCLUDING REMARKS
Perturbation theory and extrapolation methods are widely used in many areas of
physics. With respect to Monte Carlo algorithms though, although perturbation methods
are widely used,
22
extrapolation methods are less common, probably because they
involve taking small differences between large numbers, and these differences are easily
overwhelmed by statistical errors. We have shown, however, that extrapolation can be
carried out very stably by using the same random number stream to calculate all the
results that are to be extrapolated. The present application involves a novel type of
15
perturbation and extrapolation in which the high-frequency components of the path
integrals are a perturbation on the results obtained by neglecting these high-frequency
components. Similar approaches might be useful for other forms
22-24
of thermodynamic
perturbation theory.
By using both perturbation and extrapolation techniques we have presented two
easily implemented schemes to address the high computational costs associated with the
slow convergence of Fourier path-integral Monte Carlo methods with respect to the size
of the Fourier space. These methods were used to calculate very accurate partition
functions and free energies for water over a temperature range of 2964000 K. Savings
compared to FPIMC calculations without these new techniques are about a factor of 20 to
100. These new techniques hold considerable promise for calculating accurate partition
functions of much larger systems than previously studied.
ACKNOWLEDGMENTS
The authors are grateful to David Schwenke, Jonathan Tennyson, and Robert
Topper for comments on the original manuscript. A portion of this work was supported
by the Division of Chemical Sciences, Office of Basis Energy Sciences, of the U.S.
Department of Energy and performed in the William R. Wiley Environmental Molecular
Sciences Laboratory, a national scientific user facility sponsored by the Department of
Energy's Office of Biological and Environmental Research and located at Pacific
Northwest National Laboratory. Pacific Northwest National Laboratory is operated for
the Department of Energy by Battelle. Additional support was provided by the National
Science Foundation under grant number CHE97-25965.
16
References
1
R. P. Feynman and A. R. Hibbs, Quantum mechanics and path integrals (McGraw-Hill,
New York, 1965).
2
R. P. Feynman, Statistical Mechanics (Benjamin, New York, 1972).
3
W. H. Miller, J. Chem. Phys. 63, 1166 (1975).
4
J. D. Doll and D. L. Freeman, J. Chem. Phys. 80, 2239 (1980).
5
R. D. Coalson, J. Chem. Phys. 85, 926 (1986).
6
J. D. Doll, R. D. Coalson, and D. L. Freeman, Phys. Rev. Lett. 55, 1 (1986).
7
R. D. Coalson, D. L. Freeman, and J. D. Doll, J. Chem. Phys. 85, 4576 (1986).
8
D. L. Freeman and J. D. Doll, Adv. Chem. Phys. 70B, 139 (1988).
9
J. D. Doll, D. L. Freeman, and T. Beck, Adv. Chem. Phys. 78, 61 (1990).
10
R. Q. Topper and D. G. Truhlar, J. Chem. Phys. 97, 3647 (1992).
11
R. Q. Topper, G. J. Tawa, and D. G. Truhlar, J. Chem. Phys. 97, 3668 (1992).
12
R. Q. Topper, Q. Zhang, Y.-P. Liu, and D. G. Truhlar, J. Chem. Phys. 98, 4991 (1993).
13
R. Q. Topper, Adv. Chem. Phys. 105, 117 (1999).
14
J. Srinivasan, Y. L. Volobuev, S. L. Mielke, and D. G. Truhlar, Comp. Phys. Comm.,
in press.
15
M. Eleftheriou, J. D. Doll, E. Curotto, and D. L. Freeman, J. Chem. Phys. 110, 6657
(1999).
16
T. L. Hill, An Introduction to Statistical Thermodynamics (Addison-Wesley, Reading,
MA, 1960) p.119.
17
W. Forst, Theory of Unimolecular Reactions (Academic Press, New York, 1973).
17
18
D. A. McQuarrie, Statistical Mechanics (Harper & Row, New York, 1973) chapters 5
and 9.
19
L. Halonen, Adv. Chem. Phys. 104, 41 (1998).
20
H. Partridge and D. W. Schwenke, J. Chem. Phys. 106, 4618 (1997).
21
G. J. Harris, S. Viti, H. Y. Mussa, and J. Tennyson, J. Chem. Phys. 109, 7197 (1998).
22
T. P. Straatsma and J. A. McCammon, Annu. Rev. Phys. Chem. 43, 407 (1992).
23
W. Witschel and J. Bohmann, J. Phys. A 13, 275 (1980).
24
W. Witschel, Phys. Lett. 77A, 107 (1980).
18
Table I. The parameters used for calculating Q
[K]
(T).
________________________________________________________________________
parameter note 296 K 500 K 750 K 1000 K 15004000 K
________________________________________________________________________
K a 128 64 32 32 8
N
samples b
110
10
410
8
210
8
110
8
510
7
N
initial
c 110
9
210
7
110
7
510
6
510
5
N
PU
d 500 100 100 100 100

l
e 80 80 80 80 80

u
f 140 140 140 140 160
N
QR
g 6 2 1 1 1
N
QL
g 64 64 64 64 32
N
strata
h 20 20 20 20 20
N
sweeps
i 20 20 20 20 20

OH
j 0.17 0.17 0.17 0.17 0.17

HH
k 0.35 0.35 0.35 0.35 0.35
S
Vmax
l 7 20 20 none none
S
Vave
m 0.7 1.2 1.2 none none
S
EVave
n 1.5 2 2 none none
S
Nmin
n 8 8 8 none none
________________________________________________________________________
a)
the number of Fourier coefficients used to expand the paths
b)
the total number of Monte Carlo samples
c)
the number of Monte Carlo samples used before stratified sampling
d
)the number of times the relative paths are reused
19
e)
the lower limit of the hyperannulus (in mass-scaled bohrs) that defines the allowed
configuration space. The scaling mass in the mass-scaled coordinates was set at = 1 m
e
where m
e
is the mass of an electron.
f)
the upper limit of the hyperannulus (in mass-scaled bohrs) that defines the allowed
configuration space
g)
The action integral is evaluated using N
QR
repetitions of N
QL
-point Gauss-Legendre
quadrature.
h)
the number of strata used in the stratified sampling
i)
the number of the times the optimal sampling estimates are recalculated in the
adaptively optimized stratified sampling scheme
j)
the OH gaussian width parameter (in bohrs, not mass scaled) in the importance function
k)
the HH gaussian width parameter (in bohrs, not mass scaled) in the importance function
l)
If the potential energy (in eV) is above this value for any quadrature point the evaluation
of the action integral is aborted.
m)
The evaluation of the action integral is aborted once the mean potential energy (in eV)
of the path is known to exceed this value.
n
)
The evaluation of the action integral is aborted when both the estimated mean potential
energy (in eV) of the path exceeds S
EVave
eV, and a minimum of S
Nmin
quadrature
points have been evaluated.
20
Table II. Parameters used in the perturbative corrections.
a
________________________________________________________________________
parameter 296 K 500 K 750 K 1000 K 15004000 K
________________________________________________________________________
K 128 64 32 32 8
K 128384 64256 32128 32128 864
N
samples
210
8
110
7
310
6
110
6
510
5
N
initial
210
7
110
6
310
5
110
5
510
4
N
QR
9 8 4 4 1
N
QL
64 64 64 64 128
________________________________________________________________________
a
Unlisted parameters are the same as in Table I.
21
Table III. Partition functions and 2 statistical errors for the two calculations with
Fourier expansion length of K.
___________________________________________________________________________________________________
T K Q T
K,
( )
S
Q T
K,
( )
L
___________________________________________________________________________________________________
296 128 (9.42 0.19)10
9
(9.540 0.029)10
9
500 64 (1.881 0.018)10
4
(1.8972 0.0030)10
4
750 32 (3.036 0.019)10
2
(3.0252 0.0025)10
2
1000 32 0.4306 0.0029 0.42628 0.00030
1500 8 9.340 0.053 9.3755 0.0059
2000 8 51.64 0.21 51.513 0.026
2500 8 173.70 0.56 173.880 0.075
3000 8 449.4 1.3 449.22 0.18
3500 8 979.0 2.5 982.24 0.37
4000 8 1914.4 5.8 1915.83 0.75
___________________________________________________________________________________________________
22
Table IV. High-k corrections, Q T
K K corr, ,
( )

, with 2 statistical errors for specific K and extrapolated to the K = limit for
T = 2961000 K.
____________________________________________________________________________________________________________
K T = 296 K T = 500 K T = 750 K T = 1000 K
___________________________________________________________________________________________________________
64 (2.603 0.017)10
3
(2.111 0.015)10
2
96 (3.427 0.022)10
3
(2.794 0.019)10
2
128 (1.829 0.018)10
5
(3.834 0.025)10
3
(3.135 0.022)10
2
192 (2.402 0.024)10
5
256 (1.289 0.026)10
9
(2.682 0.027)10
5
384 (1.680 0.034)10
9

a
(2.401 0.049)10
9
b
(3.495 0.035)10
5
(5.035 0.033)10
3
(4.161 0.029)10
2

c
(3.494 0.035)10
5
b
(5.012 0.033)10
3
(4.128 0.028)10
2
___________________________________________________________________________________________________________
a
extrapolated using the three highest values of K in the table (this is the final calculation)
b
the extrapolation at 296 K and the second extrapolation at 500 K include the point with K = K and Q T
K K corr, ,
( )

= 0.
c
extrapolated using the second, third, and fourth highest values of K in the table (this is a test for comparison to the results of the
previous line)
23
Table V. High-k corrections, Q T
K K corr, ,
( )

, with 2 statistical errors for specific K and extrapolated to the K = limit for
T = 15004000 K.
___________________________________________________________________________________________________________
K T = 1500 K T = 2000 K T = 2500 K T = 3000 K T = 3500 K T = 4000 K
___________________________________________________________________________________________________________
12 0.4889 0.0038 1.518 0.009 3.220 0.018 5.712 0.031 9.026 0.051 13.29 0.09
16 0.7325 0.0050 2.288 0.012 4.875 0.023 8.681 0.038 13.704 0.062 20.21 0.11
20 0.8778 0.0058 2.753 0.014 5.882 0.025 10.469 0.042 16.537 0.067 24.40 0.12
32 1.0961 0.0069 3.452 0.016 7.396 0.029 13.216 0.048 20.862 0.075 30.77 0.13
64 1.2759 0.0078 4.039 0.018 8.676 0.033 15.512 0.053 24.513 0.082 36.16 0.14

a
1.4538 0.0089 4.631 0.021 9.973 0.038 17.814 0.061 28.207 0.094 41.63 0.17

b
1.4607 0.0090 4.620 0.021 9.926 0.038 17.962 0.061 28.268 0.094 41.60 0.17
___________________________________________________________________________________________________________
a
extrapolated using the three highest values of K in the table (this is the final calculation)
b
extrapolated using the second, third, and fourth highest values of K in the table (this is a test for comparison to the results of the
previous line)
24
Table VI. Comparison of FPIMC partition functions to those of Partridge and Schwenke.
____________________________________________________________________________________________________________
T Q(T) 2 % error % difference
_________________________ __________ _______________
(K) FPIMC PS
a
FPIMC FPIMC vs. PS
a
_______________________________________________________________________________________
296 (7.107 0.057)10
9
7.05310
9
0.80 0.77
500 (1.5446 0.0046)10
4
1.542810
4
0.30 0.12
750 (2.5235 0.0041)10
2
2.524810
2
0.16 0.05
1000 0.38509 0.00041 0.38493 0.11 0.04
1500 7.9161 0.0098 7.9285 0.12 0.16
2000 46.893 0.033 46.893 0.071 0.00
2500 163.896 0.084 163.789 0.051 0.07
3000 431.41 0.19 431.10 0.043 0.07
3500 953.94 0.38 951.05 0.040 0.30
4000 1874.18 0.77 1855.88 0.041 0.99
_______________________________________________________________________________________
a
Ref. 20
25
Table VII. Absolute internal free energies (kcal/mol).
__________________________________________________
T(K) present Ref. 20
__________________________________________________
296 11.036 0.005 11.041
500 8.719 0.003 8.721
750 5.484 0.002 5.483
1000 1.896 0.002 1.896
1500 6.167 0.004 6.172
2000 15.293 0.003 15.293
2500 25.333 0.003 25.330
3000 36.169 0.003 36.165
3500 47.717 0.003 47.696
4000 59.901 0.003 59.823
__________________________________________________