You are on page 1of 10

JOURNAL OF CHEMICAL PHYSICS

VOLUME 114, NUMBER 2

8 JANUARY 2001

A new Fourier path integral method, a more general scheme for extrapolation, and comparison of eight path integral methods for the quantum mechanical calculation of free energies
Steven L. Mielke
Department of Chemistry, The Johns Hopkins University, Baltimore, Maryland 21218

Donald G. Truhlar
Department of Chemistry, Chemical Physics Program, and Supercomputer Institute, University of Minnesota, Minneapolis, Minnesota 55455-0431

Received 23 June 2000; accepted 12 July 2000 Using an isomorphism of Coalson, we transform ve different discretized path integral DPI methods into Fourier path integral FPI schemes. This allows an even-handed comparison of these methods to the conventional and partially averaged FPI methods as well as a new FPI method. It also allows us to apply to DPI methods a simple and highly effective perturbative correction scheme previously presented for FPI methods to account for the error due to retaining only a nite number of terms in the numerical evaluation of the propagator. We nd that in all cases the perturbative corrections can be extrapolated to the convergence limit with high accuracy by using a correlated sequence of affordable calculations. The Monte Carlo sampling variances of all eight methods studied are very similar, but the variance of the perturbative corrections varies markedly with method. The efciencies of the new FPI method called rescaled uctuation FPI and one of Fourier analog methods compare favorably with that of the original FPI method. The rescaled uctuation method not only proves practically successful, but it also gives insight into the origin of the dominant error in the conventional FPI scheme. 2001 American Institute of Physics. DOI: 10.1063/1.1290476

I. INTRODUCTION

Path integral methods,13 especially when coupled with Monte Carlo integration, provide a powerful means of calculating accurate quantal partition functions and hence absolute free energies. Two different families of techniques have evolved that are distinguished by the method used to represent the paths. Discretized path integral DPI methods1,2,46 represent a given path using a nite number of discrete coordinate-space points that are equidistant in imaginary time and are usually referred to as beads. Fourier path integral FPI methods1,2,721 represent the deviations of the paths from free-particle paths by a Fourier expansion typically a Fourier sine expansion . The Fourier expansion may simply be truncated to a nite number of terms, which is called conventional FPI C-FPI , or one may use an approach called partial averaging10,11 FPI PA-FPI in which the potential is replaced by an effective potential that approximately includes the higher-order Fourier components that are not retained after truncation. Coalson9 has shown that the DPI formulations can be mapped isomorphically onto Fourier-like formulations that can be implemented with only slight differences from the conventional FPI method. In the present paper, we use this approach to transform ve DPI methods into analog FPI schemes, and we compare these methods to the C-FPI and PA-FPI methods as well as to a new method introduced below. The relative efciency of various DPI and FPI methods has already been widely studied9,14,2225 and debated! . By
0021-9606/2001/114(2)/621/10/$18.00 621

using the Fourier analogs of the DPI methods we can employ essentially the same Monte Carlo sampling scheme for all eight methods, and this permits a more even-handed comparison of relative efciencies than has been available previously. As we have recently shown,20 one major advantage of FPI calculations is that the paths Fourier expansion length can be truncated at a moderate number of terms, and the effect of additional terms can be considered as a perturbation. A perturbative correction for the contribution from a specic number of additional terms can be explicitly calculated in another Monte Carlo calculation that is substantially less expensive than the primary calculation. Several perturbative corrections, for various numbers of additional retained terms, can be calculated simultaneously and in a correlated fashion such that the innite expansion limit can be obtained via extrapolation without signicant distortion due to statistical sampling error.20 A major goal of the present article is to show that these techniques for perturbative corrections and extrapolation of correlated calculations can be adapted with minor modications for use with the Fourier analogs of DPI methods. Another goal is to examine the effectiveness of these techniques as a function of which path integral method is employed. We will also examine whether some of the Fourier analogs of the DPI methods can provide performance better than that of the original FPI method. All specic applications in the present paper are for vibrational-rotational partition functions of molecules. For 2001 American Institute of Physics

Downloaded 25 Jan 2001 to 160.94.96.172. Redistribution subject to AIP copyright, see http://ojps.aip.org/jcpo/jcpcpyrts.html.

622

J. Chem. Phys., Vol. 114, No. 2, 8 January 2001

S. L. Mielke and D. G. Truhlar

mal results about convergence rates are fairly general in the following discussion we will assume that the potential is bounded from below and possesses four continuous derivatives, although most of results hold for less restrictive conditions , but conclusions drawn from specic applications may need to be retested if the methods are applied to other problems because different algorithms may be favorable for different kinds of problems.
II. THEORY

where is Plancks constant divided by 2 , and x Dx(s) x denotes the summation over all paths parameterized by imaginary time s and beginning at x and ending at x . In the C-FPI method we set x equal to x and expand the resulting closed paths in Eq. 3 in a Fourier series,
K

j s x

xj

k 1

a jk sin

k s

where K is the length of the Fourier expansion. After some simplication one obtains the expression Q
K

We begin by outlining the eight FPI Monte Carlo methods that we will compare. We then discuss the methods for obtaining perturbative corrections and accurately extrapolating them.
II.A. Conventional and partially averaged FPI methods

J T
sym

N 1

K 1 k 1

N K 1 k 1

dx j

da jk

exp
j

a2 jk 2
2 k

S x,a

The internal i.e., vibrational-rotational partition function of a molecule in its ground electronic state may be obtained by calculating the trace of the canonical density operator, i.e., Q T 1
sym

where the
2 k

are the uctuation parameters given by


2

2 2,

dx x,x;

where sym is a symmetry number, x is an N-dimensional point in mass-scaled Jacobi coordinates N is 3N A 3, where N A is the number of atoms , is 1/k BT, where k B is Boltzmanns constant, x,x ; x exp H x 2

is the scaling mass of the mass-scaled Jacobi coordinates, S(x,a) is the contribution of the potential energy to the action integral for a given path and is calculated by S x,a
0

dsV s , x

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 x,x ;
x x

Dx s exp

1
0

dsH x s

V(x) is the potential energy, and J(T) is the Jacobian1 of the transformation from the integral over paths to the integral over Fourier coefcients. Note that the kinetic energys contribution to the action has been explicitly integrated. Equation 5 can be put into an expression more appropriate for Monte Carlo integration by multiplying and dividing by the free-particle partition function and restricting the conguration space to a nite domain D. We then obtain

K 1 k 1

Q fp T
sym

dx
D

da exp
j

a2 jk 2
N 2 k

exp
K

x,a , 8

dx
D

da exp
j 1 k 1

a2 jk 2
2 k

where the free particle partition function is given by


N/2

implemented exactly like the C-FPI method except that the potential is replaced by an effective potential dened by11 9
PA V eff s x

Q fp T

VD

2 2 s
N

N/2

dp V s x p , 10

exp and V D is the volume of the domain D. Partition functions calculated with the C-FPI method converge asymptotically as O(1/K). 26 Instead of simply ignoring contributions from Fourier components with k K, one can use the approach of partial averaging.10,11 The effects of k K are approximated by invoking Gibbs inequality.2 The PA-FPI method may then be
i 1

p 2 /2 2 s i

where 2 s
k K 1 2 k

sin2 k s/

11

or more conveniently for computation,

Downloaded 25 Jan 2001 to 160.94.96.172. Redistribution subject to AIP copyright, see http://ojps.aip.org/jcpo/jcpcpyrts.html.

J. Chem. Phys., Vol. 114, No. 2, 8 January 2001

A new Fourier path integral method

623

2 s

s
k 1

2 k

sin2 k s/

12

xi exp
fp

T V exp xi P P xi ,xi
1;

The PA-FPI method gives a rigorous lower bound on the partition function and converges asymptotically as O(1/K 2 ). 26 Analytic integration of the N-dimensional Gaussian transform of Eq. 10 is rarely possible for realistic molecular potential energy functions, so a gradient expansion is usually employed
PA x V eff s

/ P exp

V xi xi

/2 ,

19

or trapezoidal Trotter TT approximation, xi exp V T V exp exp x 2P P 2P i


fp 1

xi ,xi

1;

/ P exp

V xi

V xi

/2 ,

20

V s x

2 s 2

N i 1

V s x x2 i

13

Even more rapidly convergent procedures can be derived using cumulant methods11,14 but these have not been widely used for numerical work due to their increased complexity.
II.B. Discretized path integral methods

The DPI methods can be derived by starting with the identity x,x; x,x; / P
P

14

where P is the number of equidistant time slices. Using Eq. 14 in Eq. 1 and inserting the resolution of the identity P times gives Q T 1
sym P

dx1

where fp(xi ,xi 1 ; / P) is the free particle density, is commonly used to reduce Eq. 15 to a form that can be readily implemented. It is quite common to see claims made in the literature for calculations using the MT or TT approximations that have been proven only for the Trotter approximation. Formal proofs of asymptotic convergence rates with these more approximate schemes are apparently not available, but as we will discuss below, it is possible to show that partition functions calculated with the MT and TT schemes converge asymptotically as O(1/P) and O(1/P 2 ), respectively. Considerable attention has been given to the calculation of higher order corrections to the Trotter approximation.22,28,3135 One of the most widely used expressions,22,32,33 which we will simply refer to as the TakahashiImada TI approximation, can be implemented by replacing the potential in Eq. 17 with an effective potential
TI V eff x

dxP

i 1

xi exp

H xi

15 with the requirement that xP dard Trotter approximation27 exp H exp


1

V x

1 24

V V.

21

x1 . At this point the stan-

T exp

V ,

16

or symmetric Trotter approximation28 exp H exp exp V exp V , T 17

2P

Partition functions calculated with this expression converge as O(1/P 4 ). 28,32 One could use Eq. 21 in conjunction with either Eq. 19 or Eq. 20 , but we will only consider the latter option, which converts the trapezoidal Trotter approximation into the trapezoidal TakahashiImada TTI approximation. In the following discussion we will assume that the TTI approximation has the same asymptotic convergence rate as the TI approximation. Another approach to treating Eq. 15 is to expand the step propagator in a power series.3638 This yields xi exp H xi P
1 fp

2P

xi ,xi i

1;

/P
n 1

where T and V denote the kinetic and potential energy operators, is sometimes invoked. Partition functions calculated using either of these approximations can be shown28 to converge as O(1/P 2 ). Either approximation, together with Eq. 15 , yields rigorous upper bounds on the exact partition function; in particular, it can be shown29,30 that Q T; P 2
p1

exp

i
n 2

/P

Wn x

. 22

The rst two terms are given by38


1

Q T; P 2

p2

Q T; P

p2 p1 . 18 and

W2 x

d V xi

xi

xi

23

Equation 15 , together with one of the Trotter or Trotter-type approximations, is still not in a form that is easy to evaluate, and an additional approximation is required. Either the midpoint Trotter MT ,

W3 x

i 2

d
0

V x

xi

xi 1 xi

; 24

Downloaded 25 Jan 2001 to 160.94.96.172. Redistribution subject to AIP copyright, see http://ojps.aip.org/jcpo/jcpcpyrts.html.

624

J. Chem. Phys., Vol. 114, No. 2, 8 January 2001

S. L. Mielke and D. G. Truhlar

higher order terms are available but are prohibitively expensive to evaluate numerically. We follow the notation of Makri and Miller37,38 by referring to the expansion through the terms W 3 (x) as the rst order propagator FOP , and we will refer to the expansion through the term W 2 (x) as the zero order propagator ZOP . We note that these expansions are accurate to O(1/P 2 ) and O(1/P), respectively.36,38 The MT scheme can be considered a midpoint-rule integration approximation of the ZOP scheme; since this integration is accurate to O(1/P 2 ), the MT scheme converges at the same asymptotic rate as the ZOP schemeO(1/P).
II.C. Fourier analogs of DPI methods

where x xi / xi
1

xi .

29

It is worthwhile to explicitly note that the convergence order in K of the Fourier analog methods is the same as the convergence order in P of the DPI methods from which they are derived.
II.D. A new FPI method

Discretized path integral methods, including the ve MT-DPI, TT-DPI, TTI-DPI, ZOP-DPI, and FOP-DPI that we have detailed above, can be transformed into FPI methods using an isomorphism established by Coalson.9 Coalson also showed that a P-point TT-DPI scheme is equivalent to the C-FPI scheme with an innite number of Fourier coefcients but with the action integrals integrated using a P-point trapezoidal rule.9 One can see from this that the TT-DPI scheme converges at the same rate as trapezoidal-rule integration, i.e., O(1/P 2 ); thus, we see that partition functions calculated with Eq. 20 have the same asymptotic convergence rate as those obtained with Eq. 17 . Coalson further showed9 that if we change the uctuation parameters in the C-FPI method from those of Eq. 6 to
2 k;K

1 2 K 1 sin k / 2 K 1

2,

25

that a Fourier expansion of length K produces a path for which the P equidistant-time points with P K 1 are distributed as in an innite Fourier expansion. Using this relationship, each DPI method can be transformed into an FPI method; this leads to ve Fourier analog methods that we will label MT-FPI, TT-FPI, TTI-FPI, ZOP-FPI, and FOPFPI. Specically the MT-FPI method is implemented by replacing the action integral in Eq. 7 by a P-point midpoint rule, S
MT-FPI

x,a

1 Pi

P 1

V xi xi

/2 ,

26

and the TT-FPI method is implemented by replacing the action integral with a P-point trapezoidal rule integration, S
TT-FPI

x,a

1 2P i

P 1

V xi

V xi

27

where, as usual, xP 1 x1 . The TTI-FPI method is obtained by replacing the potential in Eq. 27 by the effective potential in Eq. 21 . The ZOP-FPI method involves accurate integration of V(x) over a path obtained by connecting the adjacent pairs of discretized points with straight line segments rather than using the actual Fourier path; the FOP-FPI method can be implemented by integrating an effective potential over the same path. The effective potential, as obtained from Eqs. 23 and 24 , is
FOP V eff

At rst glance only slight differences distinguish the C-FPI and TT-FPI methods, and Coalson argued9 that the two methods are essentially the same. He noted however that the TT-FPI scheme converged somewhat faster than the C-FPI scheme in some limited numerical tests while the MTFPI scheme converged somewhat more slowly than the C-FPI scheme. Apart from the initial studies,9,11 there have apparently been no calculations that utilize Fourier analogs of DPI schemes. As noted above, the C-FPI scheme converges as O(1/K) which is undesirably slow, especially at low temperatures where large values of K are often required to yield accurate results. The TT-FPI scheme converges as O(1/K 2 ) and thus should provide a considerable advantage over the C-FPI scheme, provided that the Monte Carlo sampling variance is not greatly different between the two methods. This difference in convergence rates must be predominately due to the different uctuation parameters; to emphasize this we consider yet another FPI method which we will refer to as rescaled uctuation FPI RF-FPI and which differs from the C-FPI method only by the use of Eq. 25 instead of Eq. 6 . The RF-FPI scheme reduces to the TT-FPI scheme if quadratures over the paths Eq. 7 are integrated with a P-point trapezoidal rule, but we do not restrict ourselves to equidistant-time quadrature nodes in the RF-FPI method as explained in Sec. III. The points on the Fourier path determined by the uctuation parameters given by Eq. 25 are ideally distributed only at the P equidistant-time points, so extension of the integration in the RF-FPI method to the entire Fourier path introduces a slight deviation from the quadrature result that would be obtained on a K Fourier path. Since the TT-FPI result is a trapezoidal-rule approximation of the RF-FPI result, this deviation can be seen to be of order O(1/P 2 ) and thus the RF-FPI method converges as O(1/K 2 ). The TTI-FPI method cannot protably be similarly generalized since the O(1/P 2 ) error from the nite expansion of the paths would reduce the asymptotic convergence rate to O(1/K 2 ) from O(1/K 4 ). The RF-FPI method is expected to be useful in situations where we desire a continuous specication of the path, or where we can use the extra exibility in quadrature choice to integrate some or all of the paths with fewer than P quadrature points. If we consider Eq. 25 in detail, we see that the parameters depend explicitly on the maximum expansion length K or equivalently P and are enlarged compared to the parameters of Eq. 6 . The high-k parameters differ the most, and the expression
K k 1

V x

V,

28

2 sin2 k s n / k

k 1

2 k

sin2 k s n /

30

Downloaded 25 Jan 2001 to 160.94.96.172. Redistribution subject to AIP copyright, see http://ojps.aip.org/jcpo/jcpcpyrts.html.

J. Chem. Phys., Vol. 114, No. 2, 8 January 2001

A new Fourier path integral method

625

holds exactly for the points sn n 1 K 1 , n 1,2,...,K 1. 31

a jk;K instead of a jk
jk

jk k;K

34

35

One can view the higher convergence rate of the TT-FPI scheme or the RF-FPI scheme as compared to the C-FPI scheme as resulting from this k-dependent magnication of the uctuation parameters to compensate for the truncation of the path expansion to nite K. This situation can be compared with that for the PA-FPI scheme, which also converges as O(1/K 2 ), and where the uctuation parameters for k K determine via Eq. 11 the spatial extent over which the potential energy is Gaussian averaged. A notable difference though is that the TT-FPI and RF-FPI schemes achieve the faster O(1/K 2 ) rate of convergence without requiring Laplacian evaluations.

II.E. Perturbative corrections and their extrapolation

The perturbative correction approach20 can be dened via the elementary identity Q where Q corr,K,K T Q
K K

Q corr,K,K T ,

32

T ,

33

and where K is less than K. We consider K small enough that the rst term on the right-hand side of Eq. 32 can be affordably calculated and well converged; the second term on the right hand side of Eq. 32 is expensive to calculate but small in magnitude for sufciently large K ). We have previously shown20 that in the C-FPI method, for a given number of Monte Carlo samples, we can calculate Q K (T) and Q corr,K,K (T) with a similar relative sampling variance. Thus to achieve a given absolute accuracy we need substantially fewer samples for the expensive Q corr,K,K (T) term than we need for the inexpensive Q K (T) term. In order to calculate Q corr,K,K (T), we perform simultaneous calculations for several values of K, the lowest of which is taken to equal the K of the previous paragraph. In the C-FPI method, for each Monte Carlo sample we form paths for each value of K from a single set of random Fourier coefcients. Each of these paths begins and ends at the same conguration space sample point, x, and the lower-order paths have Fourier expansions that are truncated versions of the highest-order path. We can then accumulate statistics on Q K (T), and the various Q K (T), and Q corr,K,K (T) terms in a single run. Except for statistical errors, the perturbative corrections are 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. This procedure is also used for the PA-FPI method, but we must modify the approach for the other six methods as these have uctuation parameters that vary with K. For these cases we calculate the Fourier expansion coefcients at each conguration space point using a single set of random numbers but letting the parameters vary with K, i.e., we use

in the NK-dimensional Monte Carlo average over the a space in Eq. 8 . This generates a family of paths that have Fourier expansions that are as similar as possible while still yielding sequences of partition functions that have the correct asymptotic convergence rates. As we will see in the example calculations that follow, the Q corr,K,K (T) terms for these methods have higher variances than we obtained for correction terms in the C-FPI scheme, but we can still achieve substantial savings by using these techniques. Another possible extrapolation strategy exists for the DPI schemes. Since each of the P discretization points is distributed as in an innite Fourier expansion we can select subsets of these points to calculate lower order partition function approximants. In the case of the TT scheme this amounts to extrapolation of different trapezoidal-rule integration approximations of the same path. An attractive feature of this approach is that we can obtain additional lower order results without the need to perform any additional potential evaluations if we use either the TT or the TTI scheme. If we choose the largest desired value of P as a power of two, this same-path extrapolation scheme saves nearly a factor of two in the cost of potential evaluations as compared to the scheme of Eq. 34 . Unfortunately, we found numerically that the same-path extrapolation approach yields perturbative corrections that have a much higher variance than those of the similar-Fourier-expansion approach of Eq. 34 and thus the latter approach is preferable. We have two calculations of Q K (T) from the procedure above, one using a large number of samples and a less accurate result obtained during the calculation of Q corr,K,K (T); we distinguish these two results 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 enhance the accuracy of the nal results via Q corr,K,K T Q Q
K K ,L ,S

T T

Q corr,K,K T .

36

We perform calculations of Q corr,K,K (T) at three or more values of K, and we extract Q corr, ,K (T) by tting to the functional form Q corr,K,K T Q corr,
,K

A Kn

n 1,

37

where n is the leading order of the asymptotic convergence rate i.e., n 1 for the C-FPI, MT-FPI, and ZOP-FPI methods, n 2 for the PA-FPI, TT-FPI, FOP-FPI, and RF-FPI methods, and n 4 for the TTI-FPI method . Partition functions calculated with the Trotter approximation can be shown39 to be even functions of P; thus, one might expect that an expansion in even powers of K might be better than the form of Eq. 37 for the TT scheme. It seems likely however that this result holds only for the Trotter approxi-

Downloaded 25 Jan 2001 to 160.94.96.172. Redistribution subject to AIP copyright, see http://ojps.aip.org/jcpo/jcpcpyrts.html.

626

J. Chem. Phys., Vol. 114, No. 2, 8 January 2001


N j 1

S. L. Mielke and D. G. Truhlar

mation proper rather than the TT approximation; ts of the form of Eq. 37 perform as well as or better than a form involving only even powers for all of the methods considered here.

x 2. j

39

III. COMPUTATIONAL DETAILS

In order to illustrate the various path integral methods and compare their relative efciency, we performed partition function calculations at a temperature of 300 K for HCl using a potential that has been used previously16,40 to illustrate methods for calculating quantal free energies. Monte Carlo sampling, as implemented in our algorithm,16,17,19,20 involves sampling in two distinct spacesthe conguration space x and the Fourier space a each of which is sampled in an uncorrelated fashion. Given a sampling of the Fourier space, we construct a relative Fourier path; when this is added to a conguration space sample we obtain an absolute Fourier path. A large amount of computational effort is required to actually form the relative Fourier path. If we sample the x and a space at the same rate then the formation of the relative paths dominates the computational cost for typical problems. Instead we choose to sample the x space much more frequently than the a space typically 101000 times as often ; thus, a given relative path is reused many times.19 In particular, for the present study we reuse relative paths 100 times. This means that the sequence of absolute paths has some short-term correlation, but numerical tests19 indicate that Monte Carlo variances for this type of sampling can be very accurately estimated using formulas appropriate for uncorrelated sampling. If we reuse the relative paths a large number of times, the computational expense is strongly dominated by the cost of the potential evaluations. Our algorithm15,19 for the C-FPI scheme uses Gaussian quadrature to evaluate the action integrals in an effort to minimize the number of potential evaluations required for accurate integration. In the present paper we also use Gaussian quadrature for the PA-FPI and RF-FPI calculations. Since Gaussian quadrature uses irregularly spaced quadrature nodes we must form the relative paths using matrix multiplication.19 For the FPI analogs of the ve DPI methods we only need to determine the path at the K 1 discretized points that are evenly spaced in imaginary time, i.e., the set of points given in Eq. 31 . Equation 4 then becomes
K

We subdivide D into several concentric hyperannulii and sample these via an adaptively optimized stratied sampling scheme15,19 AOSS . We also sometimes employ importance sampling in the conguration space using functions of the atomatom distances.19,20 In many cases, particularly for systems of high dimensionality and low temperature, a large fraction of absolute paths that we sample contribute negligibly to the partition function. We have implemented a number of geometric and energetic screening criteria that permit us to identify such cases early in the action integral evaluation phase, and we can then save substantial computational effort by early termination of the evaluation of contributions from these unimportant paths.19 Extensive details on the implementation of our algorithms have been presented previously, and we refer the interested reader to these sources for additional details.15,17,19,20 The calculations presented here used an adaptively optimized stratied sampling scheme with a sampling domain that is dened by 1 50 a 0 and u 150 a 0 where the scaling mass is equal to the mass of an electron and which is subdivided into 20 equal volume strata. In each case a total of 2 107 samples was calculated; 10% of these are distributed uniformly in an initial probe phase and the remainder are distributed in 20 AOSS phases as explained previously.19 Masses of 1.007 825 and 34.968 852 amu are used for H and Cl, respectively. In the present study the same number of samples (2 107 ) was used to calculate both the Q K (T) results and the perturbative correction results; this facilitates comparisons of the statistical errors. In actual applications we would rene the results by performing a calculation with a large number of samples at a single moderate value of K and apply the correction procedure outlined in Sec. II.E.
IV. RESULTS

Accurate variational calculations are available16 for HCl at 300 K for the same potential as used here, and they can be used as benchmarks for the present results. Table I lists Q K (T) and its associated statistical errors for various K for the eight methods studied. In particular we tabulate the 2 statistical error, which is calculated via
N strata i 1

j sn x

xj

k 1

a jk sin

n 1 k , K 1

38

var f i , Ni

40

and we implement this via a fast Fourier sine transform. This is substantially faster than matrix multiplication generating the entire path via matrix multiplication requires O(2NK 2 ) operations, whereas the FFT procedure only requires O(NK log K) operations , but the path generation phase of the algorithm still presents a computational bottleneck; therefore, we still reuse relative paths to increase efciency. We restrict our conguration space domain D to a hyperannulus dened by 1 u where the hyper-radius is given by

where var( f i ) denotes the variance of the Monte Carlo sampling of the integrand in strata i, N strata is the number of strata, and N i is the number of samples in strata i. We also tabulate the 2 relative statistical error given by 2 /Q K (T) which we express as a percentage . Table I also gives partition functions extrapolated to K obtained by tting the last several points to the functional form of Eq. 37 . Figure 1 displays the unsigned truncation error vs K for various unextrapolated calculations. Table II displays perturbative corrections and associated errors for various FPI methods and values of K and K.

Downloaded 25 Jan 2001 to 160.94.96.172. Redistribution subject to AIP copyright, see http://ojps.aip.org/jcpo/jcpcpyrts.html.

J. Chem. Phys., Vol. 114, No. 2, 8 January 2001

A new Fourier path integral method


2

627

TABLE I. Partition functions and 2 statistical errors for various methods. The variational result Ref. 16 is 1.651 10 K 1 2 4 8 16 24 32 64 96 128 MT-FPI 1.379 4.800 2.352 1.046 5.242 3.833 3.204 2.362 2.109 1.988 1.648 1.7 4.5 3.1 1.7 9.7 7.4 6.3 4.7 4.3 4.0 0.12 0.09 0.13 0.16 0.19 0.19 0.20 0.20 0.20 0.20 ZOP-FPI 5.959 2.870 1.319 6.414 3.713 2.947 2.593 2.097 1.942 1.866 1.647 4.8 3.2 1.9 1.1 7.1 5.8 5.2 4.3 4.0 3.8 0.08 0.11 0.14 0.17 0.19 0.20 0.20 0.20 0.20 0.20 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
4 4 4 4 5 5 5 5 5 5 1 1 1 2 2 2 2 2 2 2 2

. PA-FPI TTI-FPI
4 3 2 2 2 2 2 2 2 2 2

C-FPI 6.231 2.508 1.088 5.079 3.022 2.490 2.252 1.930 1.831 1.784 1.649 5.3 3.1 1.8 9.7 6.1 5.0 4.6 3.9 3.7 3.6 0.08 0.12 0.16 0.19 0.20 0.20 0.20 0.20 0.20 0.20 10 10 10 10 10 10 10 10 10 10 10
4 4 4 5 5 5 5 5 5 5 1 1 1 2 2 2 2 2 2 2 2

TT-FPI 3.732 1.637 6.347 2.978 2.009 1.813 1.742 1.671 1.657 1.653 1.647 2 3.6 2.2 1.2 6.2 4.3 3.9 3.7 3.5 3.5 3.5 Q K (T) 10 1 10 1 10 2 10 2 10 2 10 2 10 2 10 2 10 2 10 2 10 2

RF-FPI 5.715 2.041 7.760 3.351 2.104 1.857 1.768 1.679 1.662 1.656 1.649 10 10 10 10 10 10 10 10 10 10 10
4 4 4 5 5 5 5 5 5 5 1 1 2 2 2 2 2 2 2 2 2

FOP-FPI 2.986 1.548 5.344 1.052 1.413 1.524 1.572 1.625 1.637 1.641 1.647 2.6 1.8 7.9 1.8 2.7 3.0 3.1 3.3 3.3 3.3 0.09 0.12 0.15 0.18 0.19 0.20 0.20 0.20 0.20 0.20 10 10 10 10 10 10 10 10 10 10 10
7 6 6 5 5 5 5 5 5 5 4 3 3 2 2 2 2 2 2 2 2

10 10 10 10 10 10 10 10 10 10
3 4 4 4 5 5 5 5 5 5

1 1 1 2 2 2 2 2 2 2

9.369 4.550 1.139 1.531 1.628 1.641 1.645 1.647 1.647 1.648 1.648 8.0 5.7 1.9 2.9 3.3 3.3 3.4 3.4 3.4 3.4 0.09 0.12 0.16 0.19 0.20 0.20 0.20 0.20 0.20 0.20 10 10 10 10 10 10 10 10 10 10

10 10 10 10 10 10 10 10 10 10 10
7 6 5 5 5 5 5 5 5 5

7.269 3.654 2.173 1.744 1.659 1.650 1.648 1.647 1.646 1.647 1.646 1.7 1.0 6.4 4.5 3.7 3.6 3.5 3.4 3.4 3.4 0.23 0.28 0.29 0.26 0.23 0.22 0.21 0.21 0.21 0.21

10 10 10 10 10 10 10 10 10 10 10
4 4 5 5 5 5 5 5 5 5

2 2 2 2 2 2 2 2 2 2 2

1 2 4 8 16 24 32 64 96 128 1 2 4 8 16 24 32 64 96 128

10 10 10 10 10 10 10 10 10 10

10 10 10 10 10 10 10 10 10 10

statistical error 10 4 5.2 10 4 2.8 10 4 1.5 10 5 7.4 10 5 4.6 10 5 3.9 10 5 3.7 10 5 3.4 10 5 3.4 10 5 3.4

10 10 10 10 10 10 10 10 10 10

10 10 10 10 10 10 10 10 10 10

10 10 10 10 10 10 10 10 10 10

2 relative % error 0.10 0.09 0.14 0.14 0.18 0.19 0.21 0.22 0.21 0.22 0.21 0.21 0.21 0.21 0.21 0.20 0.21 0.20 0.21 0.20

Table III compares TT-FPI partition functions and perturbative corrections obtained using the path sequence of Eq. 34 to calculations using the same-path extrapolation approach.
V. DISCUSSION

Several factors must be considered in evaluating the efciency of the FPI methods considered here. These include the value of K that is required to get reasonable results, the variance of the calculation of Q K (T), the variance of the perturbative corrections, the quadrature costs of evaluating the action integrals, and the additional costs incurred in some of the methods for evaluating gradients or Laplacians. As indicated in Fig. 1, the MT-FPI and ZOP-FPI methods have the largest errors for a given value of K and the slowest convergence rates. The poor performance of these methods is in marked contrast to the seemingly similar TTFPI scheme. Makri and Miller37 concluded that trapezoidal Trotter calculations are superior to midpoint Trotter calculations because the former is a two-point approximation to the exact integration of the straight-line-segment path used in the propagator power series schemes while the later is only a one-point integration approximation. In contrast, in the

present study we nd that the TT scheme is not only more accurate than the MT scheme but also substantially more accurate than the ZOP method. To correctly explain these trends we must consider the way that the paths enter into the DPI schemes. Equation 15 is an exact expression for Q(T); for a specic set of P discrete points, we still must integrate over all possible paths that pass through these points. Thus the P distinguished points in the DPI schemes can be thought of as representing a set of paths which we will denote C(x1 ,x2 ,...,xP ). The trapezoidal Trotter approximation involves an operator approximation as given in Eq. 17 as well as a P-point trapezoidal rule integration scheme to approximate the contribution of each member of the set C(x1 ,x2 ,...,xP ). The MT scheme involves a similar operator approximation but differs from the TT scheme in that its calculation involves potential evaluations at points that do not lie on each member of C(x1 ,x2 ,...,xP ). In particular, the statistical distribution of a midpoint between two adjacent discretized points is narrower than the distribution from which the discretized points themselves are chosen, and thus the MT scheme yields partition functions that are biased toward higher values.

Downloaded 25 Jan 2001 to 160.94.96.172. Redistribution subject to AIP copyright, see http://ojps.aip.org/jcpo/jcpcpyrts.html.

628

J. Chem. Phys., Vol. 114, No. 2, 8 January 2001

S. L. Mielke and D. G. Truhlar TABLE II. Perturbative corrections and statistical errors for various methods, K and K . Method MT-FPI ZOP-FPI K 64 64 128 128 128 64 64 128 128 32 32 32 32 64 64 64 128 128 64 64 64 64 64 8 8 8 8 8 16 16 16 16 16 8 8 8 8 8 16 16 16 16 16 K 96 128 192 256 384 96 128 192 256 64 128 192 256 128 192 256 192 256 96 128 96 128 256 16 24 32 64 128 24 32 64 96 128 16 24 32 64 128 16 24 32 64 128 Q corr,K,K (T) 2.53 3.73 7.45 1.11 1.48 9.83 1.46 4.65 6.93 7.16 9.01 9.37 9.49 1.85 2.21 2.33 3.60 4.79 1.70 2.29 1.16 1.59 2.04 1.10 1.14 1.16 1.17 1.17 1.31 1.66 1.90 1.93 1.95 9.42 9.57 9.73 9.74 9.73 9.20 1.07 1.21 1.21 1.20 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
3 3 4 3 3 4 3 4 4 4 4 4 4 4 4 4 5 5 4 4 4 4 4 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 5 4 4 4 4

2 error 5.0 7.4 1.5 2.2 3.0 2.1 3.1 9.8 1.4 3.8 4.1 4.1 4.2 1.3 1.4 1.4 4.2 4.5 1.2 1.4 6.3 8.2 1.1 1.1 1.1 1.2 1.2 1.2 4.0 4.6 4.9 5.0 5.0 2.4 2.4 2.4 2.4 2.4 8.4 9.2 9.7 9.8 9.8 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10 10
6 6 6 6 6 6 6 7 6 6 6 6 6 6 6 6 7 7 6 6 7 7 6 5 5 5 5 5 6 6 6 6 6 5 5 5 5 5 6 6 6 6 6

2 rel. % error 0.20 0.20 0.20 0.20 0.20 0.22 0.21 0.21 0.21 0.53 0.45 0.44 0.44 0.70 0.62 0.59 1.15 0.95 0.72 0.59 0.54 0.52 0.52 1.02 1.01 1.00 1.00 1.00 3.10 2.75 2.59 2.56 2.54 2.51 2.49 2.48 2.48 2.48 9.17 8.59 8.02 8.04 8.17

C-FPI

TT-FPI

RF-FPI FOP-FPI

PA-FPI

FIG. 1. Unsigned truncation error, Q (T) Q (T) , as a function of K for various FPI methods.

The ZOP scheme can be thought of as an operator approximation together with a rule for replacing each member of C(x1 ,x2 ,...,xP ) with a single path obtained by connecting the P discretization points with straight line segments. The ZOP scheme is identical to the pure-bead limit i.e., pure discretized limit of the mixed bead-Fourier scheme of Vorontsov-Velyaminov et al.41 They argue that replacing the potential evaluations at the P points with integration over the straight lines between beads gives a better representation of the potential energy. This argument neglects the consideration that points on the straight-line-segment path have a narrower statistical distribution than the full set of properly weighted points on members of the set C(x1 ,x2 ,...,xP ), and that using only this path introduces a bias towards partition functions that are too large. The mixed bead-Fourier method of VorontsovVelyaminov et al.41 augments a discretized path representation with Fourier expansions between adjacent beads. Vorontsov-Velyaminov et al.41 state that neither the purebead nor the pure-Fourier limits of their mixed beadFourier scheme are optimal. This result must be understood as a consequence of their choice of implementation such that the pure-bead limit reduces to the ZOP scheme instead of the much more rapidly converging TT scheme; if their algorithm is suitably modied, a TT pure-bead limit would probably be optimal. A better mixed-discretized-Fourier approach could be devised using TT-DPI and TT-FPI schemes, and this mixed method would converge as the inverse square of the number of path variables for any partitioning of the

TTI-FPI

work. Such a mixed representation might be useful as it would permit both extrapolation in the Fourier space and importance sampling in conguration space for multiple points along the path. A mixed scheme such as this might also facilitate use of more advanced stratied sampling strategies than those we currently employ. The FOP-FPI scheme also involves accurate integration over a path where the P discretization points are connected by straight lines. This method is equivalent to the mixed bead-Fourier approach of Vorontsov-Velyaminov et al.41 in the pure-bead limit with gradient partial averaging over the trivial Fourier paths consisting of straight lines connecting adjacent beads. It tends to converge monotonically from below and yields good accuracy and an O(1/K 2 ) convergence rate. It performs somewhat better than the TT-FPI scheme at low K, but at higher K the accuracy of the TT-FPI and FOP-FPI schemes are very similar. As a numerical

Downloaded 25 Jan 2001 to 160.94.96.172. Redistribution subject to AIP copyright, see http://ojps.aip.org/jcpo/jcpcpyrts.html.

J. Chem. Phys., Vol. 114, No. 2, 8 January 2001

A new Fourier path integral method

629

TABLE III. Q K (T), Q corr,K,K (T), and statistical errors calculated using two different path sequence approaches and the TT-FPI scheme. P 128 64 32 128 64 32 Q
P 1

(T) 10 10 10 10 10 10
2 2 2

2 error 3.4 3.4 3.6 3.4 3.4 3.6

rel. % error

Q corr, P

1,31

(T)

2 error 1.0 9.2 10 10 10 10


5 6

rel. % error 1.08 1.21 0.45 0.52

1.656 1.676 1.751 1.656 1.675 1.751

Same-path 10 5 10 5 10 5 10 10 10
5 5 5

approach with K 127 path 0.20 9.50 10 4 0.20 7.55 10 4 0.21


4 4

2 2 2

Sequence of Eq. 34 0.20 9.48 10 0.20 7.56 10 0.21

4.3 3.9

6 6

method though, the FOP-FPI scheme is extremely expensive since it requires costly Laplacian evaluations and because accurate integration of the straight-line-segment paths requires additional functional evaluations as compared to the number required for the TT-FPI scheme. Three of the eight methods we have considered here involve accurate integration of the full Fourier path. The conventional FPI scheme displays poor accuracy and exhibits slow convergence. The RF-FPI scheme performs very similarly to the TT-FPI scheme. The PA-FPI scheme shows rapid O(1/K 2 ) convergence from below and yields superior accuracy. Unfortunately the PA-FPI scheme requires expensive Laplacian evaluations. In many situations the cost of these Laplacian calculations increases approximately linearly with the dimensionality of the system and thus we expect that the increased performance from partial averaging will not typically be sufciently compelling to offset the increased cost. The MT-FPI, TT-FPI, and TTI-FPI schemes all require P potential evaluations to integrate each path. The C-FPI, RF-FPI, and PA-FPI schemes can occasionally produce accurate results with fewer than P-point quadratures, but typically somewhat more than P points are required for accurate integration; for small K values, substantially more than P points can be required. For the present system the quadrature costs are rather modest, but there have been reports23 of systems which require about three times as many quadrature points as path variables to achieve accuracy of better than 1.5% in the PA-FPI scheme. Among the methods considered here, the TTI-FPI scheme yields the most accurate results for a given value of K and possesses the fastest asymptotic convergence rate O(1/K 4 ). Unfortunately the asymptotic rate in not realized until fairly high K, and thus the performance is not as dramatic as one might initially expect. Also the gradient calculations needed for the effective potential of Eq. 21 make the method quite expensive. Still, the methods cost is comparable to that of the PA-FPI scheme while yielding somewhat greater accuracy except at very small K. One of the most important aspects to consider in evaluating the efciency of a Monte Carlo method is the variance of the sampling. One must consider the magnitude of the variance both as a function of method and as a function of K. It is a common problem in DPI calculations for the statistical errors to increase as P increases, and a number of methods have been proposed to alleviate this problem.4244 Topper18

has stated that the sampling variance typically decreases rapidly as a function of K in FPI schemes. We observe neither trend in the results given in Table I; the sampling variance does decrease as a function of K, but only at the same often sluggish rate as the partition function. The 2 relative error is essentially independent of the path integral method employed, and rapidly approaches a value of about 0.20%0.21% as K is increased. Thus we may conclude that any reports of differences in sampling variances are likely consequences of the Monte Carlo sampling strategy employed rather than a property of the path integral methods themselves. The variance of the perturbative corrections varies signicantly depending on the path integral method used. The C-FPI scheme yields correction terms with two-standarddeviation relative errors that are comparable to those of the underlying Q K (T) calculations about 0.21% for widely varying values of K and K . The 2 relative errors for the perturbative corrections of the MT-FPI and ZOP-FPI scheme are also fairly small for the useful ranges of K and K . For the TT-FPI, FOP-FPI, and RF-FPI methods the 2 relative errors for the perturbative corrections are all signicantly larger than the sampling errors of the underlying Q K (T) calculations. The relative variance of the perturbative corrections increases as the magnitudes of the corrections decrease, as K increases, and as K decreases. The PA-FPI and TTIFPI schemes show the largest relative statistical errors for the perturbative corrections; for the TTI-FPI method with K 16, the relative statistical error is over 40 times larger than the relative statistical error in the Q K (T) calculations. The PA-FPI and TTI-FPI path integral methods also have relatively large absolute statistical errors, and this lessens their performance advantages as we must either use more samples in calculating the perturbative term or use a higher value of K . The TT-FPI, RF-FPI, and FOP-FPI methods have relatively small absolute statistical errors in their perturbative corrections, which is a critical advantage for practical calculations. The statistical error of the perturbative corrections also varies depending on the sequence of paths used to obtain the results. In Table III we see that the statistical error for perturbative corrections in the TT scheme is a factor of about 2.4 larger if we use the same path approach instead of the similar Fourier expansion sequence of Eq. 34 . The

Downloaded 25 Jan 2001 to 160.94.96.172. Redistribution subject to AIP copyright, see http://ojps.aip.org/jcpo/jcpcpyrts.html.

630

J. Chem. Phys., Vol. 114, No. 2, 8 January 2001

S. L. Mielke and D. G. Truhlar

same path approach would thus require nearly 6 times as many samples to achieve the same accuracy as the use of Eq. 34 , and this additional cost greatly outweighs the savings from reusing potential evaluations. The extrapolated values in Table I are all very similar because results for each of the methods were calculated with the same random number sequence. We also performed one calculation with the TT-FPI method, K 32, and 5 108 samples; this gave a result of (1.7477 0.0007) 10 2 , where again the quoted error is 2 . Using this result and a K perturbative correction derived from the data in Table I, we obtain a corrected partition function of (1.652 0.001) 10 2 which is in excellent agreement with the variational result16 of 1.651 10 2 . We have expressed no opinion on whether the trapezoidal Trotter scheme or any of the other DPI methods is better implemented in the discretized or Fourier formthe optimal choice may vary with the problem and may depend strongly on sampling strategies. Even if the discretized representation proves more efcient for a particular problem, one can still make use of the Fourier representation to affordably calculate a perturbative term to correct for the truncation to a nite number of beads. Furthermore, we have pointed out that there may be some advantages to a mixed discretized-Fourier method obtained by combining the TTDPI and TT-FPI schemes.
VI. CONCLUDING REMARKS

methods are only likely to be efcient choices for lowdimensional systems. On balance we expect the TT-FPI and RF-FPI schemes to be the most widely useful of the methods considered here.
ACKNOWLEDGMENT

Partial support for this work was provided by the National Science Foundation under Grant No. CHE97-25965.
R. P. Feynman and A. R. Hibbs, Quantum Mechanics and Path Integrals McGrawHill, New York, 1965 . 2 R. P. Feynman, Statistical Mechanics Benjamin, Reading, 1972 . 3 L. S. Schulman, Techniques and Applications of Path Integrals Wiley, New York, 1986 . 4 D. Chandler and P. G. Wolynes, J. Chem. Phys. 74, 4078 1981 . 5 K. S. Schweizer, R. M. Stratt, D. Chandler, and P. G. Wolynes, J. Chem. Phys. 75, 1347 1981 . 6 B. J. Berne and D. Thirumalai, Annu. Rev. Phys. Chem. 37, 401 1986 . 7 W. H. Miller, J. Chem. Phys. 63, 1166 1975 . 8 J. D. Doll and D. L. Freeman, J. Chem. Phys. 80, 2239 1980 . 9 R. D. Coalson, J. Chem. Phys. 85, 926 1986 . 10 J. D. Doll, R. D. Coalson, and D. L. Freeman, Phys. Rev. Lett. 55, 1 1986 . 11 R. D. Coalson, D. L. Freeman, and J. D. Doll, J. Chem. Phys. 85, 4567 1986 . 12 D. L. Freeman and J. D. Doll, Adv. Chem. Phys. 70B, 139 1988 . 13 J. D. Doll, D. L. Freeman, and T. Beck, Adv. Chem. Phys. 78, 61 1990 . 14 R. D. Coalson, D. L. Freeman, and J. D. Doll, J. Chem. Phys. 91, 4242 1989 . 15 R. Q. Topper and D. G. Truhlar, J. Chem. Phys. 97, 3647 1992 . 16 R. Q. Topper, G. J. Tawa, and D. G. Truhlar, J. Chem. Phys. 97, 3668 1992 ; 113, 3930 E 2000 . 17 R. Q. Topper, Q. Zhang, Y.-P. Liu, and D. G. Truhlar, J. Chem. Phys. 98, 4991 1993 . 18 R. Q. Topper, Adv. Chem. Phys. 105, 117 1999 . 19 J. Srinivasan, Y. L. Volobuev, S. L. Mielke, and D. G. Truhlar, Comput. Phys. Commun. 128, 446 2000 . 20 S. L. Mielke, J. Srinivasan, and D. G. Truhlar, J. Chem. Phys. 112, 8758 2000 . 21 J.-K. Hwang, Theor. Chem. Acc. 101, 359 1999 . 22 H. Kono, A. Takasaka, and S. H. Lin, J. Chem. Phys. 88, 6390 1988 . 23 C. Chakravarty, M. C. Gordillo, and D. M. Ceperley, J. Chem. Phys. 109, 2123 1998 . 24 J. D. Doll and D. L. Freeman, J. Chem. Phys. 111, 7685 1999 . 25 C. Chakravarty, M. C. Gordillo, and D. M. Ceperley, J. Chem. Phys. 111, 7687 1999 . 26 M. Eleftheriou, J. D. Doll, E. Curotto, and D. L. Freeman, J. Chem. Phys. 110, 6657 1999 . 27 H. F. Trotter, Proc. Am. Math. Soc. 10, 545 1959 . 28 H. De Raedt and B. De Raedt, Phys. Rev. A 28, 3575 1983 . 29 S. Golden, Phys. Rev. 137, B1127 1965 . 30 C. J. Thompson, J. Math. Phys. 6, 1812 1965 . 31 M. Suzuki, Commun. Math. Phys. 51, 183 1976 . 32 M. Takahashi and M. Imada, J. Phys. Soc. Jpn. 53, 3765 1984 . 33 X.-P. Li and J. Q. Broughton, J. Chem. Phys. 86, 5094 1987 . 34 W. Janke and T. Sauer, Phys. Lett. A 165, 199 1992 . 35 M. Suzuka, Phys. Lett. A 180, 232 1993 . 36 Y. Fujiwara, T. A. Osborn, and S. F. J. Wilk, Phys. Rev. A 25, 14 1982 . 37 N. Makri and W. H. Miller, Chem. Phys. Lett. 151, 1 1988 . 38 N. Makri and W. H. Miller, J. Chem. Phys. 90, 904 1989 . 39 M. Suzuki, Phys. Rev. B 31, 2957 1985 . 40 G. J. Hogenson and W. P. Reinhardt, J. Chem. Phys. 102, 4151 1995 . 41 P. N. Vorontsov-Velyaminov, M. O. Nesvit, and R. I. Gorbunov, Phys. Rev. E 55, 1979 1997 . 42 E. L. Pollock and D. M. Ceperley, Phys. Rev. B 30, 2555 1984 . 43 M. Sprik, M. L. Klein, and D. Chandler, Phys. Rev. B 31, 4234 1985 . 44 W. Janke and T. Sauer, Chem. Phys. Lett. 201, 499 1993 .
1

We have compared eight different Fourier path integral methods including the conventional and partial averaging versions, a new Fourier method based on rescaled uctuations, and ve discretized schemes that have been transformed into Fourier schemes by using the isomorphism of Coalson.9 This isomorphism allows us to apply Monte Carlo sampling in the Fourier space as well as to adapt our perturbative correction and correlated extrapolation schemes to DPI methods. The Monte Carlo relative sampling variance is observed to have little dependence on the path integral method or the value of K. The sampling variance of the perturbative corrections does vary strongly with method as well as K and K . The C-FPI, MT-FPI, and ZOP-FPI schemes are observed to yield poor accuracy and slow convergence as compared to the other methods. The FOP-FPI method performs reasonably well as a function of the number, K, of terms in the Fourier series, but is extremely expensive and thus is always less efcient than either the PA-FPI or TT-FPI schemes. The TT-FPI scheme is observed to be both efcient and accurate. The new RF-FPI method performs similarly to the TT-FPI scheme. Both methods converge well without requiring gradients or Laplacians, and the absolute variance of their perturbative corrections is relatively small compared to the other methods. The PA-FPI and TTI-FPI methods yield good results but at considerable expense due to the need to calculate Laplacians and gradients respectively; thus these two

Downloaded 25 Jan 2001 to 160.94.96.172. Redistribution subject to AIP copyright, see http://ojps.aip.org/jcpo/jcpcpyrts.html.

You might also like