Accuracy of free-energy perturbation calculations in molecular simulation. I. Modeling

Transcription

1 JOURNAL OF CHEMICAL PHYSICS VOLUME 114, NUMBER 17 1 MAY 2001 ARTICLES Accuracy o ree-energy perturbation calculations in molecular simulation. I. Modeling Nandou Lu and David A. Koke a) Department o Chemical Engineering, State University o New York at Bualo, Bualo, New York Received 6 December 2000; accepted 5 February 2001 We examine issues involved in applying and interpreting ree-energy perturbation FEP calculations in molecular simulation. We ocus in particular on the accuracy o these calculations, and how the accuracy diers when the FEP is perormed in one or the other direction between two systems. We argue that the commonly applied heuristic, indicating a simple average o results taken or the two directions, is poorly conceived. Instead, we argue that the best way to proceed is to conduct the FEP calculation in one direction, namely that in which the entropy o the target is less than the entropy o the reerence. We analyze the behavior o FEP calculations in terms o the perturbation-energy distribution unctions, and present several routes to characterize the calculations in terms o these distributions. We also provide prescriptions or the selection o an appropriate multistage FEP scheme based on how the important phase-space regions o the target and reerence systems overlap one another American Institute o Physics. DOI: / I. INTRODUCTION Free-energy perturbation FEP is one o the most widely used methods to compute ree energies via molecular simulation. 1 4 A single-stage FEP calculation 5 yields the ree-energy dierence between two systems as an ensemble average perormed on only one o them. The working equation can usually be put in the orm e A 1 A 0 e U 1 U 0 0. The 0 and 1 subscripts indicate properties or the two systems o interest; A is the Helmholtz ree energy, U is the conigurational energy, and 1/kT with k the Boltzmann s constant and T the absolute temperature. The angle brackets indicate an ensemble average perormed on the 0 system, which we call the reerence; the 1 system we call the target. Either system o interest can be selected to be the reerence, and thus the FEP calculation or a given pair o systems can be perormed in either o two directions. I the systems dier in the presence o a single molecule, the FEP calculation yields the chemical potential, and in this case the two implementations are commonly known as the Widom insertion 6,7 and Widom deletion 8 methods. In what ollows we will present the idea that in all FEP calculations one o the directions can be considered uniquely as an insertion calculation while the other represents a deletion, so we will adopt this terminology here to reer to the two directions o an FEP calculation. In principle, insertion and deletion methods yield equivalent results or the ree-energy dierence, but in practice the calculated results dier systematically It is well known that the calculation in one direction typically overestimates the ree-energy dierence, while calculation in the other direction underestimates it. What is not well appreciated is the asymmetry in the magnitude o the over- and underestimations. Many researchers treat the errors in the two methods as being o the same magnitude and opposite sign Thereore, a heuristic has emerged that the best estimate o the ree-energy dierence should be taken as the average o the insertion and deletion orward and reverse ree energies. This practice is wrong. It is the aim o this work to highlight this misconception and present a means to analyze the magnitude asymmetry. The reliability o the FEP calculations has two acets, representing the precision and accuracy, respectively. Precision describes reproducibility o the measurements, and can be characterized by a variance statistic. In a previous study 27 we set up quantitative models to describe the precision o FEP calculations, and showed that precision is closely related to the entropy dierence between the reerence and target systems. We applied our analysis to develop a prescription or selecting optimal intermediates in staged FEP calculations. The accuracy o a calculation is distinct rom its precision. It indicates the correctness o the estimate, or how the measurement result deviates rom the true or exact value in repeated measurements. Arguably, the accuracy o a result is much more important than its precision, yet it usually gets little attention. One reason is that in many situations the inaccuracy o an ensemble average is much less than its imprecision. However, this circumstance oten does not hold or FEP calculations, even in cases where other ensemble averages are o good accuracy. Indeed, FEP results could be in good precision but terrible accuracy. A simple and exa Electronic mail: koke@eng.bualo.edu /2001/114(17)/7303/9/$ American Institute o Physics

2 7304 J. Chem. Phys., Vol. 114, No. 17, 1 May 2001 N. Lu and D. A. Koke treme example is the deletion calculation o the chemical potential or the hard-sphere HS system. According to this method the chemical potential is always zero whatever the system condition. Obviously, this is a wrong answer except or zero density, but it has perect reproducibility. Fortunately the result is so obviously wrong that no one would make the mistake o interpreting its good precision as good accuracy. But this situation is anomalous, and comes about because the hard-sphere system has only two possible energy levels zero and ininity. Inaccuracy is much better disguised in systems only slightly more realistic than the hardsphere model, even one having only three energy levels. 28 Poor understanding o the accuracy in realistic systems leads to misinterpretations o simulation results and misapplications o simulation methodologies. In contrast to precision, where the block-average variance provides a good measure, there is no simple way to assess the accuracy o an FEP calculation. In this study we address this problem, and examine the inaccuracy in the FEP calculations as a unction o the simulation length and basic characteristics o the target and reerence systems. We analyze the source o the simulation error and develop the methodology and models to characterize it using energy distribution unctions. We also urther study the asymmetry behavior o the FEP calculations and correct some related misconceptions. In the next paper o this series we apply this understanding to develop an easy-to-use heuristic that gauges the inaccuracy o insertion FEP calculations. We present in Sec. II the analysis and modeling o the inaccuracy. In Sec. III we describe simulations used to veriy the inaccuracy model and heuristic. Simulation results and discussions are given in Sec. IV, and concluding remarks in Sec. V. II. BACKGROUND AND THEORY A. Generalized insertion and deletion Our previous studies 13,14,27,28 show that the entropy dierence between the reerence and target systems is an important quantity in determining both the accuracy and precision o FEP calculations. Thereore, it is convenient to identiy the FEP systems according to their entropy. In our study, the higher-entropy system is denoted H and the lowerentropy one is L. We deine the dierence o a quantity between two systems as (L) (H). For example, the reeenergy, potential-energy, and entropy dierences are deined as A A L A H, U U L U H, and S S L S H, respectively; by deinition, S 0. In a Widom insertion calculation or the chemical potential, the reerence system can be viewed as one in which the inserted molecule moves around without intererence rom any o the other molecules in the system. The extra reedom available to this molecule accords the reerence with a higher entropy than the target. Correspondingly, we deine generalized insertion and deletion directions or any FEP calculation based on the sign o entropy dierence along the perturbation direction. We call an FEP calculation a generalized insertion i the H system is used as the reerence, where the entropy dierence along the perturbation direction is less than zero, S 1 S 0 0 where the 1 and 0 subscripts indicate the target and reerence systems, respectively, as discussed in the Introduction. Similarly, a generalized deletion calculation uses the L system as the reerence, and proceeds in a direction o increasing entropy: S 1 S 0 0. In a generalized insertion calculation we have expa exp u H, 2 where u is the dierence o conigurational potential energy U(r N ) between the L and H systems, i.e., u U L (r N ) U H (r N ), the energy change encountered in a trial sampling. For convenience, we simply call u potential energy or energy in the ollowing, but one should keep in mind that it really means the potential energy change during the perturbation. The momentum contributions to the Hamiltonian are ignored in this study, assuming they are the same in both systems. The subscripted angle brackets... H indicate the canonical-ensemble average sampled according to the Hamiltonian or the higher-entropy system. Similarly, the ree-energy dierence in a generalized deletion calculation is given by expa exp u L. 3 Note that Eqs. 2 and 3 are equivalent, as shown by Shing and Gubbins 8,29 in the context o the chemical potential calculation. B. Entropy and coniguration space The reerence and target systems in an FEP calculation share the same phase space, but the important regions o phase space those that contribute most to their respective partition unctions may dier between the two systems. The entropy o a system is closely related to the size o the important region o phase space. The important region could be deined, or example, by considering the energydistribution unction P(U) P U 1 Q U eu, where Q and are the canonical and microcanonical partition unctions, respectively. We could deine a threshold value P* or P(U), such that an integral over only those energies U that have P(U) above the threshold P(U) P* would yield some substantial raction e.g., 99% o the integral taken instead over all energies. Then, any coniguration r N having an energy U(r N ) such that P U(r N ) P* would be considered an important coniguration. Note that a coniguration may be important i it has a low large negative energy, or i it is one o many conigurations large having a nonprohibitive energy. In the ollowing analysis, we will consider FEP calculations between systems in which the set o important conigurations or the low-entropy system designated L orms a wholly contained subset o the important conigurations ( H ) o the high-entropy system: L H, as shown in Fig. 1. This is the case, or example, or systems o hard spheres in which one sphere is L system or is not H system interacting with the others. It is very likely that this situation i.e., L H holds or other molecule-insertion system pairs in- 4

3 J. Chem. Phys., Vol. 114, No. 17, 1 May 2001 Accuracy o ree energy perturbation. I FIG. 1. Schematic diagram o phase space and important conigurations in an FEP calculation. The shaded region represents the important conigurations or the high entropy system ( H ). The white region inside and part o H is the important conigurations or the low-entropy system, L.The signiicant eature o this diagram is L is completely inside H. In real systems the shapes o the H and L regions are vastly more complex than the simple two-dimensional representations presented here. volving more realistic and complex potentials. In the event that the important regions do not relate this way, it is unlikely that FEP can be made eective without setting up stages that break the calculation into pieces involving pairs o systems that do ollow the relation L H. We discuss this matter urther in the Discussion section below. The conigurations in H that do not lie in L all outside because their energy is greater than the energies o the conigurations in L. Thus, we expect that conigurations in H will exhibit energy dierences u that are typically positive and large. The primary exceptions are those conigurations that lie also in L, where the conigurations will tend to have a value o u that is zero or negative. Note that this characterization o the value o u in H and L is more o an expectation than a rule; deviations rom this description, perhaps signiicant ones, might be ound in particular systems. Much o what ollows does not rely on this characterization, but it can provide guidance to the intuition. C. Energy distributions We can rewrite the ensemble averages in Eqs. 2 and 3 in terms o a one-dimensional integral weighted by the distribution o energy changes encountered in a simulation, as proposed by Shing and Gubbins. 8 Following their notation, we deine (u) and g(u) as the distribution unctions o the potential energy change u encountered in a simulation using insertion method H system as the reerence and deletion method L system reerence, respectively. Then, Eq. 2 becomes expa du exp u u, 5 where the integral goes rom to. Likewise, the deletion ree-energy ormula is given by expa du exp u g u. 6 Note that both (u) and g(u) are normalized; also, physical considerations place a lower bound on the energy or which g(u) is nonzero. A useul relation between the and g distributions is 8,30 u exp Ag u expu. The dierence in the ensemble-averaged energies between the L and H systems is given as a derivative o the ree-energy dierence, U(A)/. With Eqs. 5 7 the result can be expressed in terms o the energy distributions and their temperature derivatives U du u ln ug u 7 ln g u du u u. 8 These temperature derivatives can in turn be interpreted via luctuation ormulas ln u ln g u U H H H U H H, 9 U L L L U L L, 10 where the angle brackets represent an ensemble average in the H or L system as indicated, and U is the energy in the corresponding system; also, is the Dirac delta unction applied to the energy change u, or which the ensemble average then gives the distribution unctions: (u) H and g(u) L. These results show that the energy change cannot be obtained rom knowledge o the distribution unctions at only one temperature. But, it is interesting to consider the degree o correlation between the energy change u and the conigurational energy U in the H and L systems. According to the ormulas above, correlations in the H system are relevant or ranges o u where g(u) is non-negligible, i.e., or conigurations in L. While these conigurations have a avorable u, it is not likely that the energy U H will dier markedly or systematically rom the average U H H in this region. For particle insertion, these are conigurations in which a hole has opened up to permit the insertion o a molecule. Except or very high-density systems, opening such a hole need not be associated with a large change in the energy U H. In contrast, the correlations between u and U L in the L system, over the region H where (u) is nonnegligible, will be signiicant. The energy-change u is large in these regions, which indicates that U L will also be systematically large. From these arguments we anticipate that the derivative o ln (u) will be small, while the derivative o ln g(u) will be non-negligible. Thus, rom Eq. 8 we can approximate the energy change as a simple g integral, but cannot express it in terms o the integral U ug u du. 11 Assuming this result is valid, then the entropy dierence can also be expressed in terms o the and g distributions; thus S g u ln u /g udu, 12 which, again, is by deinition less than zero.

4 7306 J. Chem. Phys., Vol. 114, No. 17, 1 May 2001 N. Lu and D. A. Koke FIG. 2. A plot o typical and g distributions as unctions o potential energy change u. Both distributions are normalized. The associated Boltzmann actor integrands or the and g distributions are also schematically shown in the plot, as the dashed curves. Figure 2 shows the typical shape o the (u) and g(u) distributions, together with the Boltzmann actor that is averaged to obtain the ree-energy dierence. By looking at the plot it is clear that the important region or the insertion calculation is the low-energy region, while the high-energy region is critical to the deletion calculation. These regions correspond to the tails o the or g distributions. They are prone to poor sampling in a inite length simulation since they correspond to conigurations to be sampled with low probability. So, in terms o the and g distribution unctions, the error in the ree-energy calculation comes rom the poor sampling o the distribution tails, more explicitly, the low energy tail or insertion calculation and the high energy tail or deletion calculation. D. Modeling inaccuracy Let us generalize the problem in the ollowing way. Suppose a quantity X is an average o unction F(u) according to weighting distribution p(u), i.e., X p u F u du. 13 In connecting Eq. 13 to FEP calculations, the average quantity X is either exp( A) or the insertion calculation or exp( A) or the deletion calculation; p(u) is the weight unction, either (u) or g(u); and F(u) is exp(u) or exp(u) or the insertion and deletion calculations, respectively. During the simulation we sample u according to p(u) and inally measure the average X as a sample average o F(u). It is clear that the inaccuracy or error in the average X depends on how well p(u) is sampled. For a long but inite length simulation, the high-probability region o p(u) will be well sampled, while the lowprobability regions, or the wings o the distribution, are generally not well sampled. This situation is shown schematically in Fig. 3 a, where the shaded regions indicate energies that are not well sampled. For any given sample, one can identiy energy values, u 1 and u 2, such that u is never sampled or the regions (,u 1 ) and (u 2,). Each such energy is reerred to as a limit energy in the ollowing. The exact X is given by the averaging over whole range o u FIG. 3. Depiction o the probability distribution representing or g, sampled region, and the relative inaccuracy o the FEP calculations. a Shaded regions the tails o distribution p(u) are those that have not been well sampled during a inite length simulation, while the region in between is perectly sampled; b the relative inaccuracy o the insertion calculation is given by the area with energy less than u under the g distribution, while the relative inaccuracy o the deletion g calculation is given by the area with energy larger than u g under the distribution. X exact u 1 u p u F u du 2p u F u du u1 p u F u du. 14 u2 As part o our model development, we now assume that in a simulation u is sampled perectly according to p(u) or u between u 1 and u 2. Then, the X measured in the simulation will be u 2p u F u du, 15 X sim u1 and the simulation error in X is given by the dierence between X sim and X exact X X sim X exact u 1 p u F u du p u F u du. 16 u2 Now we return speciically to the FEP calculations. For the insertion calculation, the high-energy term in Eq. 16 has an extremely small value and thereore can be dropped. Similarly, the low-energy term in Eq. 16 or a deletion calculation can be ignored. We then can write the inaccuracy o insertion and deletion calculations as ollows: and E ins expa sim ins expa exact u u exp u du 17

5 J. Chem. Phys., Vol. 114, No. 17, 1 May 2001 Accuracy o ree energy perturbation. I E del expa sim del expa exact g u exp u du, ug 18 where u is the lowest limit energy sampled in an insertion calculation and u g is the highest limit energy sampled in a deletion calculation. We can see in Eqs. 17 and 18 that the inaccuracy o FEP calculations is given by the contributions o the important conigurations that are never sampled. It is helpul to look at the relative error ( e) o the FEP calculations. Knowing Eq. 7, the relationship between and g distributions, we have e ins expa ins sim expa exact expa exact u g u du 19 and e del expa del sim expa exact expa exact u du. ug 20 Note that or small errors, the magnitude o the relative error in exp(a) is roughly equal to the magnitude o the absolute error in A itsel. Equations 19 and 20 tell us that the relative inaccuracy o the insertion calculation is given by the area with energy less than u under the g distribution, while the relative inaccuracy o the deletion g calculation is given by the area with energy larger than u g under the distribution. This relation is depicted in Fig. 3 b. One may note that the relative errors o ree-energy calculation, as shown by Eqs. 19 and 20, are always less than or equal to zero since the area under an or g curve is non-negative. This implies that A sim ins A exact sim and A del A exact, or in other words, or an insertion calculation the ree-energy dierence is overestimated, and or a deletion calculation it is underestimated. This analysis agrees with the common understanding and observation o FEP calculations. However, this common understanding oten includes a misconception in the magnitude o the accuracy o the FEP calculations. The prevalent view is that both calculations are equally wrong, and the best result is obtained by splitting the dierence Equations 19 and 20 do not explicitly provide inormation as to the magnitude o inaccuracy. To urther look at the asymmetry problem, we need to study some properties o the and g distribution unctions. E. Asymmetry o FEP calculations Asymmetry in the FEP calculation is easily understood in the context o the preceding analysis by considering the overlap o the and g distributions. I the upper limit energy o the g distribution lies well above the bulk o the distribution, then the error in deletion is small; likewise, i the lower limit energy o lies below the bulk o g, the insertion FIG. 4. Asymmetry in accuracy o FEP calculations. The inaccuracy o the insertion deletion calculation is given by the shadow area under the g( ) distribution. The size o the area is related to the length o simulation, the overlap level o the energy distributions, and the breadth o each distribution. This series o plots show how the inaccuracy o both calculations changes as increasing the simulation length increases the degree o overlap o distributions. The insertion calculation, which typically samples the broader energy distribution, will give less simulation error than the deletion calculation. error is small. In the extreme cases o no overlap Fig. 4 a, both calculations give poor results, while or complete overlap both give equally good results. It is in the intermediate case that the asymmetry o the methods emerges. For some intermediate degree o overlap, the better result is obtained by sampling the distribution that is broader. As the amount o sampling is increased, and the limit energies move urther into the wings o the distributions, the broader distribution will encompass the narrower distribution irst, meaning that sampling on the broader distribution will provide a result o greater accuracy Figs. 4 b and 4 c. This analysis raises an obvious question: Which distribution is wider, the insertion distribution or the g deletion distribution? The distribution describes sampling o the system having the larger entropy, so it is tempting to conclude that it is always the wider distribution. For the Widom calculation o the chemical potential, this is indeed the case. However, we are unable to prove this as a general result, so at present we leave it as a reasonable speculation indicating that insertion FEP calculations are inherently more accurate than corresponding deletion calculations.

6 7308 J. Chem. Phys., Vol. 114, No. 17, 1 May 2001 N. Lu and D. A. Koke F. Most-likely analysis The next step in our model development is the ormulation o a means or estimating the limit energies that govern the inaccuracy o the FEP calculations. What inaccuracy can be expected or a simulation with given length M? Assuming prior knowledge o the and g distributions, it is possible to have a general answer to this question. It is clear rom the oregoing inaccuracy model that the magnitude o the error depends on the value o limit energy u g or insertion calculation or u or deletion calculation. Once we have the limit energy, we know the inaccuracy o the calculation. So, the question becomes what is the limit energy or a given simulation? The limit energy u or u g can be ound easily in a simulation. O course, or a given simulation length M we would ind dierent values o u or u g in dierent independent simulation runs. However, a distribution o limit energies P(u) can be deined to describe the limit energies observed in repeated runs o the same length M. This distribution will exhibit some maximum corresponding to the most likely limit energy u* to be observed in any simulation. Then, we identiy the most likely ree energy to be that obtained by integrating Eq. 17 or 18 to this most likely limit energy. Accordingly, we deine the inaccuracy as the dierence between the most-likely ree-energy outcome and the true value. Now, the key quantities needed are the probability unction P(u) and, rom this, the most-likely limit energy, u*. First, let us look at these quantities or an insertion calculation. The probability density that u is the lowest energy sampled in a simulation with M insertion trials is given by P u u u du u M The right-hand side o Eq. 21 is the product o the probability density that u u is sampled exactly once, and the probability that or all other (M 1) samplings the energy value is between u and ininity u u is never sampled. The most-likely limit energy u* is obtained by maximizing P(u), and thus u* should satisy P(u) / u u u * 0, or more conveniently ln P(u) / u u u * 0. Ater straightorward derivation starting with Eq. 21, weget or large M ln u M u*. u u u* 22 Similarly, or the deletion calculation we have u M 1 P u g g u g g g u du. 23 Then the most-likely limit energy u g * is ln g u Mg u u g *. u u g * 24 Equations 22 and 24 indicate the most-likely limit energies or the insertion and deletion FEP calculations with simulation length M, knowing the energy distribution unctions and g. Insertion o these limit-energy values into Eqs. 17 or 19 and 18 or 20 yields the most-likely inaccuracy o the corresponding FEP calculations. III. VERIFICATION AND SIMULATION TESTS As described in Sec. II D, our inaccuracy model agrees with some common qualitative observations and understanding o the FEP error. Here, we conduct simulation tests to assess quantitatively the ability o the model to describe the inaccuracy o FEP calculations. We perormed Monte Carlo simulation o simple Lennard-Jones LJ luids in the NVT ensemble. We consider a reduced density o 0.9 all units are made dimensionless by the LJ size and energy parameters, and two reduced temperatures: 0.9 and 2.0. The perturbation systems are deined as ollows: system H contains 107 particles interacting according to the LJ potential i.e., it has 107 LJ particles, and one particle that interacts with the others according to the hard-sphere potential with diameter 0.8. The LJ potential is truncated at a separation o 2.5, and no long-range correction is included. System L has 108 LJ particles. The insertion calculation ( S 0) corresponds to the perturbation direction rom system H to L, i.e., the reerence is system H. The trial insertion process involves the identity change o the HS to a LJ particle and the growth o the particle size. The deletion calculation uses system L as the reerence and involves the change o a LJ particle to a hard sphere. Both insertion and deletion FEP calculations were perormed during the simulation. This introduces a small approximation to the calculation, in that the deletion calculation should be perormed or 108 LJ spheres, but separate tests show that the error introduced by this consideration is negligible. A test o the inaccuracy model requires as input the exact energy distribution unctions. These we obtained by making histograms o the perturbation energy encountered in a long simulation with 5M cycles where each cycle comprises 108 attempts o MC particle displacement ollowed by one trial o insertion or deletion measurement. The true reeenergy dierence is also needed to gauge the inaccuracy o the FEP calculations. We obtained this value using the relationship described in Eq. 7, or more explicitly, exp(a) exp( u)g(u)/(u), applied to the measured and g distributions over their region o overlap. Particularly important to the analysis are the low-energy tail o distribution and the high-energy tail o g distribution, which typically have poor statistics. To improve them, we correct the low-energy tail o (u) using the inormation o the true exp(a) just obtained, along with g(u) which has very good statistics in the low-energy range. Similarly, the high-energy tail o g(u) is corrected using the true ree-energy dierence and the distribution. We examined the eect o the simulation length M since it is a very important quantity aecting the error o the simulation. Simulations with M 1000, 2000, 5000, and cycles were perormed independently. Ater each simulation, we have ree-energy dierence by both insertion and deletion calculations. Simulations o each length were repeated 200 times independently. Ater all repeating simulations, the

7 J. Chem. Phys., Vol. 114, No. 17, 1 May 2001 Accuracy o ree energy perturbation. I probability histogram o ree-energy results was made and the ree-energy dierence corresponding to the peak o the histogram was obtained and used as the most-likely value. Based on the most-likely ree-energy dierence and the true value, the most-likely inaccuracy is computed. In accord with the development, we report our results as the ractional error in the exponential o A instead o A directly. The most-likely inaccuracy observed rom the simulation is given as ollows. For insertion calculation E s ins expa sim expa true, 25 or deletion calculation s E del expa del expa true, 26 where the superscript s indicates the most-likely value observed rom simulations. For the comparison we also computed the most-likely inaccuracy according to the model. We irst computed the most-likely limit energies with the and g distribution unctions obtained rom the long simulation. The most-likely lowest energy sampled in an insertion calculation, u*, and the highest energy sampled in a deletion calculation, u g *, were calculated using Eqs. 22 and 24, respectively. Then, using the discrete orm o our model Eqs. 17 and 18, the most-likely inaccuracies are E ins * u i exp u i u i u* 27 or insertion calculations and E del * g u i exp u i 28 u i u g * or deletion calculations. Note the discrete distribution unctions (u i ) and g(u i ) are normalized. We compare the inaccuracies observed rom the simulations Eqs. 25 and 26 and those rom our most-likely model Eqs. 27 and 28 to veriy i the inaccuracy analysis and model are valid. IV. RESULTS AND DISCUSSION The most-likely inaccuracies computed directly by the simulation data and by the inaccuracy model or both insertion and deletion calculations are presented in Fig. 5 as a unction o simulation length M. We report results or two system temperatures: Fig. 5 a or T 0.9 and Fig. 5 b or T 2.0. The state condition T 0.9 and 0.9 corresponds to a LJ luid located slightly above the liquid solid coexistence temperature. 31 From Figs. 5 a and 5 b, we can see that the model results agree with the simulation inaccuracies very well in both cases. This gives us the conidence that the error analysis conducted is appropriate. The diiculties o sampling the complex important coniguration space can be simply studied by investigating the one-dimensional energy distribution unctions. The major error o the FEP calculations is due to the poor sampling o the energy distribution wings. FIG. 5. Comparison o the inaccuracy in the ree-energy change, evaluated by repeated simulations and by the model. Both the inaccuracy o insertion square and deletion circle calculations are plotted as a unction o simulation length M, together with the model predictions or insertion and or deletion. T 0.9 or a and T 2.0 or b. In both cases the system density is 0.9. In Fig. 5 there are more points worth noting. As we expected, the ree-energy errors or the insertion and deletion calculations have opposite sign, and both decrease with the increasing simulation length M. A longer simulation is able to sample the coniguration space better than a shorter simulation, and thereore gives the estimation o the ree-energy dierence with better statistics. From the model point o view, in a longer simulation the limit energy is pushed urther away toward the end o the energy tail. For an ininite simulation, the whole range o the and g distributions will be well sampled and the result will be exact. However, the speed o error reduction over the simulation length M is not the same or the insertion and deletion calculations. From Fig. 5 we can see the error o the insertion calculation decreases aster than its deletion counterpart, with increasing simulation length. It is more eicient to reduce the simulation error by increasing the simulation length or insertion calculations than deletion ones. We observed similar behavior or the HS or simple three-state models 28 and believe it is common or many simulation situations. This behavior also reveals the asymmetric character o the insertion and deletion calculations. The magnitudes o the simulation errors by insertion and deletion calculations are not the same, a point

8 7310 J. Chem. Phys., Vol. 114, No. 17, 1 May 2001 N. Lu and D. A. Koke which is clearly shown in Fig. 5. We ind the insertion calculation results in less error, or more reliable ree-energy estimation. One should not assume these two FEP calculations are symmetric only because the errors have dierent sign. It is worthwhile to consider how the superiority o insertion impacts the design o multistage FEP methods. Multistage methods 13,21,32 calculate the ree energy between a reerence and target by breaking the dierence into smaller parts. Thus A 1 A 0 A 1 A W A W A 0, 29 where the W subscript deines a system that in some way is intermediate between the target and reerence system. Given the 0, 1, and W systems, there are our ways in which the overall FEP calculation can be completed, diering in the selection o the target and reerence systems or the two subcalculations involved in computing the total dierence. Two are commonly known as umbrella sampling 33 and Bennett s method, 34 while the other two we have called 13 staged insertion and staged deletion, although these names are appropriate only i S 1 S W S 0. In previous work 27 we showed that the intermediate should be selected such that the entropy dierences S 1 S W and S W S 0 are equal. This result anticipated the present work in that it assumed that both FEP calculations would be computed in the insertion direction, and that in particular S 1 S W S 0. Such a selection o the intermediate system is possible only i the important conigurations o the target and reerence systems are related as described in Fig. 6 a. In our limited experience, examining Widom-type FEP calculations involving the insertion or partial insertion o a single molecule, this has always been the case. For other perturbations this relationship may not hold, and it is important then to be especially careul in the ormulation o the intermediate system. In particular Bennett s method or umbrella sampling may be more appropriate staging schemes. The cases in which they are appropriate or use are depicted in Figs. 6 b and 6 c. Umbrella sampling is needed i there is virtually no overlap o the important conigurations o the target and reerence systems. Then, the W system must be selected to encompass both o them. Bennett s method can be applied i the systems have some overlap in their important regions o phase space. Bennett s method is special in that the W system is never used as a reerence in one o the substage FEP calculations. As shown by Bennett, this eature permits some optimization to be applied in selecting the intermediate. Interestingly, i the target and reerence are related as shown in Fig. 1, with the target orming a complete subset o the reerence, then at least or hard spheres the optimization procedure yields an intermediate that is identical to the target, i.e., the deletion stage o the calculation is eliminated and the method reduces to a single-stage FEP calculation. Staged insertion is then the method o choice. Higher-order staging algorithms could be developed to extend these ideas. FIG. 6. Appropriate staging strategy in dierent situations according to the overlap o the important phase space. A two-stage FEP calculation is used as an example, where three systems, the reerence 0, the target 1, and the intermediate W are involved. In the phase-space diagrams, these three systems are represented as the partial shaded, black, and white regions, respectively. For FEP calculations containing more stages, the same idea can be applied. The relationship o the important conigurations o three systems is schematically shown. Inormation about choosing an appropriate intermediate system is described in the text or each case. a The important conigurations o the target system sits inside that o the reerence system completely. A staged insertion is the best choice or an accurate result. b No overlap region o the important conigurations o the reerence and target. Umbrella sampling is the most suitable technique. c The important conigurations o the target and reerence are partially overlapped. Bennett s method is appropriate or FEP calculations. V. CONCLUSIONS We are strongly o the opinion that FEP calculations should be perormed only in the insertion direction, i.e., the simulation should sample the system o larger entropy, and compute the ree-energy dierence or a perturbation to a target system o smaller entropy. This outcome ollows as a consequence o a more important and more restrictive criterion or obtaining accurate measurements in FEP calculations. It is necessary that the conigurations important to the target system orm a subset o the conigurations important to the reerence. This situation is depicted in Fig. 1. For small, localized perturbations such as those employed in chemical-potential calculations, the more restrictive condition is usually satisied by the less restrictive condition that looks only at the entropy dierence. In more complex cases, where the target cannot be selected such that its important conigurations are entirely a subset o the important conigu-

9 J. Chem. Phys., Vol. 114, No. 17, 1 May 2001 Accuracy o ree energy perturbation. I rations o the reerence, it is necessary to apply a staging approach to the ree-energy calculation. However, the design o the staging method should still be guided by the criteria outlined here. ACKNOWLEDGMENT Acknowledgment is made to the Donors o the Petroleum Research Fund, administered by the American Chemical Society, or support o this research. 1 D. Frenkel and B. Smit, Understanding Molecular Simulation: From Algorithms to Applications Academic, New York, M. P. Allen and D. J. Tildesley, Computer Simulation o Liquids Clarendon Press, Oxord, K. E. Gubbins, in Foundations o Molecular Modeling and Simulation, AIChE Symposium Series, edited by P. Cummings and P. Westmoreland, Vol. 97, No P. Kollman, Chem. Rev. 32, R. W. Zwanzig, J. Chem. Phys. 22, J. L. Jackson and L. S. Klein, Phys. Fluids 7 2, B. Widom, J. Chem. Phys. 39, K. S. Shing and K. E. Gubbins, Mol. Phys. 46 5, B. Guillot and Y. Guissani, Mol. Phys. 54 2, J. G. Powles, W. A. B. Evans, and N. Quirke, Mol. Phys. 46 6, J. G. Powles, Chem. Phys. Lett. 86 4, K.-K. Han, J. H. Cushman, and D. J. Diestler, J. Chem. Phys. 93, D. A. Koke and P. T. Cummings, Mol. Phys. 92 6, D. A. Koke and P. T. Cummings, Fluid Phase Equilibria 150, N. G. Parsonage, Mol. Phys. 89 4, N. Parsonage, J. Chem. Soc., Faraday Trans. 92 7, R. H. Henchman and J. W. Essex, J. Comput. Chem. 20 5, W. L. Jorgensen and C. Ravimohan, J. Chem. Phys. 83, D. A. Pearlman and P. A. Kollman, J. Chem. Phys. 90, R. H. Wood, W. C. F. Mühlbauer, and P. T. Thompson, J. Phys. Chem. 95, R. J. Radmer and P. A. Kollman, J. Comput. Chem. 18, C. Chipot, L. G. Gorb, and J.-L. Rivail, J. Phys. Chem. 98, W. L. Jorgensen, J. K. Buckner, S. Boudon, and J. Tirado-Rives, J. Chem. Phys. 89, D. A. Pearlman, J. Phys. Chem. 98 5, R. H. Wood, J. Phys. Chem , C. Chipot, C. Millot, B. Maigret, and P. A. Kollman, J. Phys. Chem , N. Lu and D. A. Koke, J. Chem. Phys. 111, N. Lu and D. A. Koke, in Foundations o Molecular Modeling and Simulation, AIChE Symposium Series, edited by P. Cummings and P. Westmoreland, Vol. 97, No K. S. Shing and K. E. Gubbins, Mol. Phys. 43 3, M. P. Allen, in Proceedings o the Euroconerence on Computer Simulation in Condensed Matter Physics and Chemistry, edited by K. Binder and G. Ciccotti Como, Italy, 1996, Vol. 49, pp R. Agrawal and D. A. Koke, Mol. Phys. 85 1, J. P. Valleau and D. N. Card, J. Chem. Phys. 57, G. M. Torrie and J. P. Valleau, J. Comput. Chem. 23, C. H. Bennett, J. Comput. Chem. 22,