Fluctuation-dissipation theorem, kinetic stochastic integral and efficient simulations Markus Hu tter and Hans Christian O ttinger ET H-Zu rich, Department of Materials, Institute of Polymers, CH-8092 Zu rich, Switzerland Di†usive systems respecting the Ñuctuation-dissipation theorem with multiplicative noise have been studied on the level of stochastic di†erential equations. We propose an efficient simulation scheme motivated by the direct deÐnition of the “ kinetic stochastic integral Ï, which di†ers from the better known Itoü and the Stratonovich integrals. This simulation scheme is based on introducing the identity matrix, expressed in terms of the di†usion tensor and its inverse, in front of the noise term, and evaluating these factors at di†erent times. related to the di†usion tensor by 1 Introduction Di†usion equations are studied whenever Ñuctuations, or irreversible behaviour in general, are involved ; examples being particle di†usion, heat conduction, viscous Ñow and the description of complex Ñuids. For a d-dimensional variable x, these equations can be written in the form C D d d d 1 d p(x, t) p(x, t) \ [ É [A(x, t)p(x, t)] ] É D (x, t) É dx dt dx 2 dx (1) where the term containing the d-dimensional vector A is called drift, and the other containing the positive semi-deÐnite d ? d-matrix D is called the di†usion term. For the following, it is essential to notice that the di†usion tensor D is placed between the two partial derivatives. In systems where the drift originates from a gradient of a potential that is mediated through the di†usion or mobility tensor, i.e. A is of the form C D d 1 /(x, t) A(x, t) \ [ D (x, t) É dx 2 (2) one then Ðnds that the “ Boltzmann distribution Ï p(x) B exp[[/(x)] is a stationary solution of the above di†usion equation [eqn. (1)]. A solution of the di†usion equation can often not be found in closed form if the di†usion tensor depends on the state of the system, i.e. for multiplicative noise, and one is led to solve the di†usion equation numerically with Ðnite element methods. However, for many complex systems, such as polymer solutions or colloidal suspensions, the large number of degrees of freedom is prohibitive for a numerical solution of the di†usion equation. Rather than studying the system by means of the distribution function p(x, t) for the variable x, one thus looks for a set of trajectories x which have the distribution p(x, t), meaning that one tries to solve stochastic differential equations.1,2 A trajectory, which is a sample for the di†usion equation [eqn. (1)], is given by the following stochastic di†erential equation1h3 (Itoü version, denoted by the symbol “ I Ï) : C dx \ A(x, t) ] A BD 1 d É D (x, t) 2 dx dt ] B (x, t) I dW (3) where the random increments are composed of d@-dimensional Wiener increments dW and a d ? d@-matrix B , the latter being B (x, t) É B (x, t)T \ D (x, t) (4) where superscript T denotes the “ transpose Ï of matrix B . Eqn. (4) relating random forces and di†usive/dissipative dynamics is the well known Ñuctuation-dissipation theorem of the second kind,1,4 which is a key result of statistical mechanics. Choosing the Itoü version of the stochastic di†erential eqn. (3) means that the random increments B (x, t) I dW have mean value zero. Notice the divergence term in eqn. (3) : putting the di†usion tensor in eqn. (1) between the two derivatives is the reason for the divergence term in eqn. (3). In other words, for systems given by eqn. (1) with a drift term given by eqn. (2), the Ñuctuation-dissipation theorem and the Boltzmann stationary solution lead to an ItoüÈstochastic di†erential equation with this additional contribution to the drift term. We here propose a numerical integration scheme for stochastic di†erential equations given by eqn. (3), that circumvents the calculation of the divergence term completely, but rather constructs it by using a two-step scheme with the main goal of substantially reducing the computational needs in simulations. In order to motivate this integration scheme we Ðrst deÐne a new stochastic integral, from which we then heuristically show how the divergence term is constructed. One should mention that a stochastic integral deÐnition, which naturally respects the Ñuctuation-dissipation theorem, has been given already for the one-dimensional case by Tsekov,5 but we give here a non-trivial generalization to arbitrary dimensions. The latter stochastic integral shall be called “ kinetic Ï, as proposed by Klimontovich.6,7 2 Multi-dimensional deÐnition of the “ kinetic stochastic integral Ï There are currently three di†erent deÐnitions for stochastic integrals, the most famous of which are the Itoü and Stratonovich versions,2,3 respectively. While the transformations among di†erent types of integrals are straightforward in principle, certain integrals may be much more convenient for developing efficient simulation algorithms. They di†er by the time point at which the integrand is evaluated when writing the integral as a sum over Ðnite time intervals : Itoü evaluates the integrand at the starting point of the time interval, whereas Stratonovich uses the midpoint values. A third stochastic integral deÐnition has recently been given by Tsekov,5 using the endpoint values of the integrand. He proved this integral to respect the Ñuctuation-dissipation theorem naturally in the one-dimensional case. However, it does not J. Chem. Soc., Faraday T rans., 1998, 94(10), 1403È1405 1403 possess this desirable property in the multi-dimensional case. We give here a deÐnition for the kinetic stochastic integral with the main property of naturally respecting the Ñuctuationdissipation theorem, i.e. of incorporating the divergence term in eqn. (3), also for an arbitrary multi-dimensional case. A deÐnition of this new integral, which shall be denoted by the symbol ), in terms of the Itoü integral by P P B (x) ) dW À B (x) I dW ] 1 2 PC D d É D (x) dt dx does not provide us with any more insight for the construction of the desired integration scheme. Much more useful is the key result of this paper, which is the deÐnition of this integral in terms of the mean-square limit of a time-discretized sum, P M B (x) ) dW À ms [ lim ; 1 [D (x ) É D (x ) ] 1] 2 ti`1 inv ti M?= i/1 (5) É B (x ) É [W [ W ] ti ti`1 ti with D À B É B T and the Wiener increments *W À W ti ti`1 [ W having mean value S*W T \ 0 and variance ti ti S*W *W T \ d *t1. The deÐnition of the inverse D has tk ik inv ti to account for the fact that the null-space of D might be nontrivial, i.e. that some linear combinations of the variables x are not a†ected by the random and di†usive motion. We therefore deÐne the matrix D to be the null-matrix on the null-space inv of D and the inverse of D on the remaining subspace, i.e. the inverse of the diagonalized matrix D is deÐned by taking the inverse of the eigenvalues with the deÐnition 1 À 0. (As a con0 sequence, the null-spaces of D and D are identical at a given inv time t ). It is essential that D is evaluated at the same time as inv i B for the image of B and the image of D to be the same subspace. Furthermore, this makes clear that the deÐnition 1 À 0 0 on the null-space does not lead to any complications and is not related to any physical singularities, but only stands for excluding the non-di†usive subspace from the deÐnition of the stochastic integral. For a strictly positive-deÐnite di†usion tensor, D is simply the inverse of D . inv One can show heuristically that the above deÐnition [eqn. (5)] for the kinetic stochastic integral is indeed the sum of the Itoü integral and the desired divergence term by making a Taylor expansion of the Ðrst di†usion tensor in eqn. (5) and expressing the end-point value x in terms of the startpoint ti`1 value x through the stochastic integral eqn. (3) to Ðrst order ti in *W . One Ðnds in the mean-square sense ti d 1 É D (x) dt B (x) ) dW \ B (x) I dW ] (6) dx 2 P P PC D illustrating the existence of the non-random contribution in the kinetic integral, as the expectation value SB (x) ) dWT \ S(d/dx) É D (x)Tdt is non-vanishing in general. We point out that the factorization in the integrand and the evaluation of the integrand factors at di†erent times are essential for producing the divergence term. The occurrence of the divergence term when going to non-vanishing time steps, i.e. Ðnite time resolution, was used by Fixman8 when constructing Langevin equations and the numerical integration scheme for the simulation of polymer dynamics, but it was not formulated in terms of stochastic integration. 3 Construction of a numerical integration scheme It has been shown in the previous section that the di†erences between the Itoü and the kinetic versions of the stochastic integral is the inclusion of the extra divergence term in the kinetic integral, formally changing the equations for dx but leaving the solution x unchanged, since 1404 J. Chem. Soc., Faraday T rans., 1998, V ol. 94 G A(x) ] A BH 1 d Æ D (x) 2 dx dt ] B (x) I dW \ A(x) dt ] B (x) ) dW The main advantage of this kinetic integral deÐnition lies, however, in the possibility of constructing efficient numerical integration schemes for stochastic processes as in eqn. (3), without ever calculating the divergence of the di†usion tensor. Such integration schemes have already been used in Brownian dynamics simulations of dense colloidal suspensions,9 being of Ðrst order in the time step *t. However, to the best of our knowledge, no direct mathematical deÐnition of a kinetic stochastic integral for arbitrary dimensions has ever been given. We think that a deÐnition by means of stochastic integrals, such as eqn. (5), is necessary to derive more efficient numerical integration schemes, especially to develop higher order integration schemes. In the following, we heuristically derive a weak2 Ðrst order numerical integration scheme from our kinetic integral deÐnition. Writing down the kinetic version of the stochastic di†erential equation, dx \ A(x ) dt ] B (x ) ) dW t t t t = A(x ) dt ] 1 [D (x ] dx ) É D (x ) ] 1] É B (x ) É dW (7) inv t t 2 t t t t suggests the following two-step scheme for the numerical integration. The predictor step is given by (8) *xp \ A(x )*t ] B (x ) É *W ti ti ti ti where *W \ W [ W are the Wiener increments with ti ti`1 i mean value S*W T \ 0 and variance S*W *W T \ d *t1. ti tk ik ti This predictor value is now used in the Ðrst di†usion tensor term on the second line of eqn. (7) for the corrector step : *xc \ 1 [A(x ] *xp) ] A(x )]*t ti ti ti ti 2 ] 1 [D (x ] *xp) É D (x ) ] 1] É B (x ) É *W (9) 2 ti ti ti ti inv ti It can be shown that the above two-step integration scheme is weakly convergent to Ðrst order in the time step *t by expanding D (x ] *xp) up to Ðrst order in *W . For a conti ti ti stant di†usion tensor, the above scheme is even weakly convergent to second order. The term involving D (x ] *xp) ti ti replaces the calculation of the divergence of the di†usion tensor in the Itoü stochastic di†erenctial eqn. (3). Thus, the kinetic version is to be favoured to the Itoü version, except in cases where a closed expression for (d/dx) É D (x) can be given. For cases where only D but not D itself is explicitly inv known (the connection between D and D being as described inv in Section 2), the determination of B and especially of (d/dx) Æ D (x) is computationally very time consuming. However, the deÐnition of the kinetic stochastic integral can be reformulated in a way to circumvent this problem : using D É D É B \ B , we Ðnd with B À D É B from the deÐnition inv inv inv of the kinetic stochastic integral P M B (x) ) dW \ ms [ lim ; 1 [D (x ) ] D (x )] 2 ti`1 ti M?= i/1 (10) É B (x ) É [W [ W ] tia1 ti inv ti with B É B T \ D É B É B T É D T \ D É D É D T \D T \ D . inv inv inv inv inv inv inv inv For such systems, the predictor step is then put into the form *xp \ A(x )*t ] D (x ) É B (x ) É *W ti ti ti inv ti ti and the corrector step is given by *xc \ 1 [A(x ] *xp) ] A(x )]*t ti ti ti ti 2 ] 1 [D (x ] *xp) ] D (x )] É B (x ) É *W ti ti ti ti inv ti 2 (11) (12) The contributions to *xp and *xc in eqn. (11) and (12), respecti ti tively, which involve the di†usion tensor D , can then be solved for by using the tensor D and a conjugate gradient method. inv In this sense, it is crucial to note that the integral in eqn. (10) is only well deÐned if the null-space of D at t is equal to, inv i`1 or a subset of, the null-space at t ; this was not a restriction to i eqn. (5). Again, the term involving D (x ] *xp) replaces the ti ti calculation of the divergence of the di†usion tensor. This latter integration scheme has been used by Ball and Melrose in Brownian dynamics simulations of dense colloidal suspensions,9 where the time invariance of the null-space of D is inv motivated by the Galilean invariance of the equations of motions (i.e. total particle momentum lies in the null-space of D for all times). inv It should be recalled that the integration scheme proposed in eqn. (8) and (9) applies to a larger class of problems than that in eqn. (11) and (12), as the former requires no restrictions on the null-space of D or D . The deÐnition of the kinetic inv integral in eqn. (5) [and not eqn. (10)] is thus to be considered as the starting point when developing numerical integration schemes. 4 Conclusions A deÐnition for a stochastic integral was given that naturally respects the Ñuctuation dissipation theorem by including the divergence of the di†usion tensor in the arbitrarily multidimensional case. It has been shown that this integral deÐnition can be used to construct numerical integration schemes for stochastic di†erential equations without ever calculating the divergence of the di†usion tensor and thereby substantially reducing the computational needs. This work was supported by grant no. 2100-046728.96/1 from the Swiss National Science Foundation. References 1 2 3 4 5 6 7 8 9 H. C. Ottinger, Stochastic Processes in Polymeric Fluids : T ools and Examples for Developing Simulation Algorithms, 1st edn., Springer, Berlin, Heidelberg, 1996. P. E. Kloeden and E. Platen, Numerical Solution of Stochastic Differential Equations, Springer, Berlin, Heidelberg, 1992. C. W. Gardiner, Handbook of Stochastic Methods for Physics, Chemistry and the Natural Sciences, 2nd edn., Springer, Berlin, Heidelberg, New York, Toronto, 1985. R. Kubo, Rep. Prog. Phys., 1966, 29, 255. R. Tsekov, J. Chem. Soc., Faraday T rans., 1997, 93, 1751. Yu. L. Klimontovich, Physica A, 1990, 163, 515. Yu. L. Klimontovich, Physica A, 1992, 182, 121. M. Fixman, J. Chem. Phys., 1978, 69, 1527. R. C. Ball and J. R. Melrose, Physica A, 1997, 247, 444. Paper 8/00422F ; Received 15th January, 1998 J. Chem. Soc., Faraday T rans., 1998, V ol. 94 1405