JOURNAL OF CHEMICAL PHYSICS

VOLUME 121, NUMBER 18

8 NOVEMBER 2004

Time-dependent exchange-correlation current density functionals with memory

Yair Kurzweil and Roi Baera)
Department of Physical Chemistry and the Lise Meitner Minerva-Center for Computational Quantum Chemistry, The Hebrew University of Jerusalem, Jerusalem 91904, Israel

Received 21 April 2004; accepted 11 August 2004 Most present applications of time-dependent density functional theory use adiabatic functionals, i.e., the effective potential at time t is determined solely by the density at the same time. This paper discusses a method that aims to go beyond this approximation, by incorporating memory effects: the derived exchange-correlation potential will depend not only on present densities but also on the past. In order to ensure the potentials are causal, we formulate the action on the Keldysh contour for electrons in electromagnetic elds, from which we derive suitable KohnSham equations. The exchange-correlation action is now a functional of the electron density and velocity eld. A specic action functional is constructed which is Galilean invariant and yields a causal exchange-correlation vector potential for the KohnSham equations incorporating memory effects. We show explicitly that the net exchange-correlation Lorentz force is zero. The potential is consistent with known dynamical properties of the homogeneous electron gas in the linear response limit . 2004 American Institute of Physics. DOI: 10.1063/1.1802793

I. INTRODUCTION

One of the important open theoretical challenges in molecular physics is the description of the dynamics of interacting electrons, for example, in a molecule or a cluster. In many typical cases the electrons start from their ground state and are subject to an external perturbation, such as a longitudal electric eld Eext (r,t) ext (r,t), where ext is an applied scalar potential imposed, for example, by the nuclei of the molecule within the BornOppenheimer approximation or by external sources. Runge and Gross1 proved that the time-dependent electron density n(r,t) determines the external potential up to a purely time-dependent additive function. Thus the density uniquely determines the external electric eld Eext (r,t). Ghosh and Dhara2 extended this result proving that the particle current-density j(r,t) determines the pair of external potentials ext and Aext up to an arbitrary gauge, i.e., the external electromagnetic elds, Bext Aext and Eext ext c 1 Aext where c is the speed of light , are uniquely determined by the particle current-density j. This forms a basis for a class of successful approaches to electron dynamics, known collectively as time-dependent density functional theory TDDFT Ref. 1 and time-dependent current-density functional theory TDCDFT Refs. 2 4 . Both analogous in many aspects to density functional theory DFT of Kohn and co-workers.5,6 In TDCDFT, the unique correspondence between the external elds and the particle current-density j(r,t) is used to set up a map between interacting and noninteracting electron systems: one looks for a system of noninteracting electrons having the same time-dependent TD current density and s controlled by a unique external electromagnetic eld (Eext or
a

Author to whom correspondence should be 972-2-6513742. Electronic mail: roi.baer@huji.ac.il

addressed.

Fax:

s Bext ). The class of current densities that can be generated in both an interacting and a noninteracting electron system is called noninteracting v -representable current densities. For this class, the effective electromagnetic elds are unique1,2,7 and therefore constitute a mapping of the interacting electron system upon a more tractable noninteracting one. In DFT, the issue of v -representability can be settled, based on the quantum mechanical minimum principles, for a wide class of cases see a discussion of this topic in Ref. 8 . An analogous result for TDDFT has not been established. One desirable way to formulate such a mapping, assuming it exists, for a given system is via an action principle. Because of the enormous complexity of the problem, only approximate mappings can be constructed in actual applications. The basic idea was outlined by Runge and Gross1 RG , who used the time-dependent quantum-mechanical action principle. The simplest approximations are adiabatic s in the sense that they give a prescription to build Eext and s Bext at time t based solely on the current density at that same time. Such an approach is expected to be appropriate for slowly varying external elds. In general, however, the effective elds at time t depend on the current density also at earlier times, i.e., they must have memory. Even in the linear response regime this is important, as excited states of double electron character cannot be accounted for using merely adiabatic functionals.9 Memory effects may be even more important for stronger fast varying elds, for example, the electronic correlation energy can be positive, an effect that cannot be accounted for using adiabatic functionals.10 It is the purpose of this paper to present a simple prescription for constructing a KohnSham mapping for TDCDFT which includes memory effects. To date, such a prescription is only partially available.3,4 A recently published method by Tokatly and Pankratov11 also attempts to achieve

0021-9606/2004/121(18)/8731/11/$22.00

8731

2004 American Institute of Physics

Downloaded 30 Oct 2004 to 132.64.1.37. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp

8732

J. Chem. Phys., Vol. 121, No. 18, 8 November 2004

Y. Kurzweil and R. Baer

this goal in a different way, based on the Landau Fermiliquid theory where the local Lorentz force is a divergence of a stress tensor. In this paper we expand and give in full detail the method which was recently proposed12 allowing construction of a functional having the following properties. a It is based on a TDCDFT action principle. It is therefore robust and, one hopes, applicable beyond linear response. The present formulations of memory prescriptions3,4,13 apply directly to potentials and usually cannot be derived from a three-dimensional 3D action principle. b It is causal, i.e., the external effective elds at time t depend only on the past behavior of the system. Here we formulate the RG theory on a Keldysh contour. The resulting theory is thus causal. While resulting in a different functional, our approach is heavily based on the ideas and results of van Leeuwen.14 c It obeys basic symmetry rules of quantum mechanics, i.e., it does not allow electrons to exert a net force upon themselves and it is gauge invariant. The rst of these conditions, termed Galilean invariance.15,16 Galilean invariance ensures compliance of TDDFT dynamics with the Harmonic potential theorem of Dobson.16,17 d It is consistent with known static and dynamical properties of the homogeneous electron gas HEG when subject to weak external electromagnetic elds. Here we follow parametrizations of such properties.18 20 The idea of assembling an approach that is Galilean invariant GI and is consistent with the linear response properties of the HEG goes back to the work of Dobson, Bunner, and Gross3 DBG . Our approach, however, is different in several aspects, most notably, we give here a fully 3D prescription, treating both the transverse as well as longitudal response properties. The mathematical intricacies of such a functional are rather involved. We nd that a relatively transparent, tractable, and formally appealing formulation is obtained when the functionals are made to depend on the velocity eld u j/n and density n instead of on the current-density j. The utility of using the velocity eld has been stressed by Vignale and co-workers,4,13 who use it to build a linear response TDCDFT approach. Some applications showing the utility of the memory effects have been recently published.21,22 With the velocity eld u(r,t), the concept of uid parcel trajectory is naturally dened.3,4,17 Thus the electronic density is viewed as a uid, and each uid parcel ows along a trajectory. The memory effects are easily described within a parcel, because there they are local. This is a Lagrangian description of the electron dynamics and should be contrasted with the Eulerian system, where a xed coordinate system is used to describe the density. Finally, since other parts of the theory are more naturally described in the Eulerian system, a Lagrangian-to-Eulerian transformation must be made after formulation of the relevant exchangecorrelation potentials. The structure of the paper is as follows. In Sec. II we formulate a causal theory of TDCDFT, based on a similar approach for TDDFT.14 Next we introduce a specic action functional including memory effects Sec. III . In Sec. IV we

take the appropriate functional derivatives to obtain the exchange-correlation potentials. The parametrization of the functionals kernel functions is derived in Sec. V. A discussion and summary is then given in Sec. VI.
II. TDCDFT ACTION

In TDDFT time-dependent Schrodinger equations are set up for one-particle orbitals from which the density and current density can be constructed. This is similar to DFT, where time-independent equations are solved for noninteracting electrons in an effective potential composed of the external potential, the Hartree potential, and the exchange correlation potential.6 In DFT the exchange-correlation XC energy functional E XC n 0 n 0 (r) being the ground state number density of the electronic system is a basic concept, determining the exchange-correlation potential through functional derivation, v XC(r) E XC n 0 / n 0 (r). In TDDFT the analogous concept, introduced by Runge and Gross,1 is the XC action functional S XC n , where n(r,t) is the TD number density. RG assumed that functional derivation of S XC yields the exchange-correlation potential as in DFT. One very successful case is TD adiabatic local density approximation TD-ALDA , where the DFT local density approximation LDA functional is used as a memory-less action. TD-ALDA is very successful for computing some dynamical properties of molecules.2327 Yet, it is inadequate in processes that involve motion of electrons over long distances22,28 or in cases of double excitations.29 Thus, it is important to develop methods beyond the adiabatic assumption and several attempts in this direction have been made.4,10,30,31 The RG procedure, as well as any attempt to derive the potentials from a time-dependent functional of the density was found to violate causality.14 This can be remedied by formulating the action on a Keldysh contour, which was used earlier in the context of TDDFT.14,32,33 The work of van Leeuwen14 has especially an elegant way of dealing with causality using the Keldysh contour. We now generalize this method for the magnetic eld case, obtaining a formulation of TDCDFT. Consider a system of N e electrons under an external electromagnetic eld. We specically focus on the dynamics in the time interval 0, f . The Keldysh contour is a parametrization t( ), using a parameter 0, f . This function, t( ), is taken from 0 to t f and then back to zero: t(0) t( f ) 0. We start from the Schrodinger equation of a system of N e electrons on a Keldysh contour: it
1

0.

2.1

The Hamiltonian here is of the general form, allowing external electromagnetic elds,
Ne

H
i 1

pi a ri , 2

v ext ri ,

1 2i

1 . rij 2.2

Despite the formal similarity to the Schrodinger equation, Eq. 2.1 is different since it is solved on the contour. To recover the physical equation, we must introduce physical potentials on the contour, i.e., v (r, ) v p r,t( ) and

Downloaded 30 Oct 2004 to 132.64.1.37. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp

J. Chem. Phys., Vol. 121, No. 18, 8 November 2004

Exchange-correlation current density

8733

a(r, ) ap r,t( ) . Only then will the solution be of the form ( ) p t( ) , where p (t) is the physical wave function. We introduce the following action: a A Re i ln
0

f ,0

2.3

continuity equation 2.10 . A better approach would be to dene a functional of n and j, both treated as independent variables, and add Eq. 2.10 as a constraint. Furthermore, we will need in the next sections a functional expressed in terms of n and the velocity eld u j/n, thus we dene S n,u as S n,u nu S n r, S u r, and S n r, t r, a
1

where the evolution operator is a solution of

Ne

it

,0
i 1

pi ri a 2

1 2i

1 U rij

w r, u r,

n r, 2.15
f

,0 , 2.4

. are

with U(0,0) being the identity operator and r, a a r,

0

The functional derivatives for 0 r, a w r,

v ext r,

2.5

n r,

2.16

is the vector potential in the xed gauge we are going to use. It is easy to show that in this gauge j r, with r, jH and
Ne

A r, a

Re

U f ,0 H r, j U f ,0 0

w r, w r, .

u r, 2.17

0 0

2.6

U 0, U j

,0 ,

2.7

0, f there are additional function terms. These disAt appear when physical densities are used, thus we do not consider them. We may analogously dene the exchange correlation action S XC n,u using Eqs. 2.13 and 2.15 by S n,u S 0 nu SH n S XC n,u . 2.18

r j
i 1

a pi ri

r ri

c.c.

2.8

is the particle current-density operator. One can also dene the density as n r, Re
0

The noninteracting action S 0 n,u in this case is tractable, since we may let the electrons evolve under the potential s a according to the Schrodinger equations k it
1 k

U
0

n H r, U f ,0
f ,0

0 0

2.9

r,

1 2

a p s r,

r,

2.19

It is straightforward to show that the continuity equation holds: t

1

From Eq. 2.12 , applied to the noninteracting electrons, we get S0 u r, S0 n r, t s r, a s r, a

1

n "j 0.

2.10

w s r, w s r, .

n r, u r,

, 2.20

Furthermore, when a physical vector potential is plugged in the expressions, j and n become the corresponding gaugeinvariant physical functions. One can now perform a Legendre transform j S A a d 3r
C

w s r,

d j r,

a r,

2.11

Equating the functional derivatives of Eq. 2.18 leads to a a n s u s a a w ws w ws XC , a

t

Functional differentiation gives S j r, r, a . 2.12

w ws

2.21 v H v XC ,

We then write this functional in the KohnSham KS way, dening the exchange-correlation action XC as S j S S0 j SH n S XC j . 2.13

with w s being the Lagrange multipliers for noninteracting electrons, aXC S XC / u, and v XC S/ n. This leads to s XC /n W, a a a W
0

2.22
1

S 0 is the action for a set of noninteracting electrons and SH n 1 2 d 3 rd 3 r

C

u"aXC /n v H v XC t

n r,

n r , r r

2.14

The functional in Eq. 2.13 will be awkward to handle S because of the functional dependence of n on j through the

where W w w s . The formalism is gauge invariant. One can apply any gauge transform to Eq. 2.19 and transform part of the longitudinal component of the vector potential into an external potential. This corresponds to a different choice of W in Eq. 2.22 . Indeed, for constructing denite

Downloaded 30 Oct 2004 to 132.64.1.37. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp

8734

J. Chem. Phys., Vol. 121, No. 18, 8 November 2004

Y. Kurzweil and R. Baer

KohnSham equations, a certain gauge must usually be chosen. Once this is done, the equations have the form i
t k

r,t

1 2

p as r,t

v s r,t

r,t

2.23

tials are used in the KohnSham scheme, Eq. 2.23 , where the effective potentials are, analogous to Eq. 2.24 with a 0,
v s r,t v p r,t v H r,t v ALDA r,t v GIXC r,t

along with as a aXC /n, 2.24

v s v ext v H v XC u"aXC /n.

u r,t aGIXC r,t /n r,t , as r,t aGIXC r,t /n r,t .

3.4

Equations 2.23 and 2.24 form a set of equations which must be solved self-consistently. The basic issue now is the construction of an approximation to the exchange-correlation action S XC n,u , which we do next.
III. ACTION WITH MEMORY

The potentials v GIXC and aGIXC should be derived from a GI action functional, the denition of which we discuss now. Consider two coordinate systems, r of the lab frame and r of a moving frame, where r x r 3.5

We now describe the basic principles of our approach for building an action principle for TDCDFT with memory. For simplicity, we assume the interacting system is not subject to a magnetic eld, thus a 0. We separate the functional into an instantaneous response part and a memory part: S XC S A S GIXC . 3.1

and x is an arbitrary accelerated trajectory with x 0 0 for simplicity in pseudotime accelerated frame. A functional S n,u is considered GI if n S ,u S n,u , 3.6

u where (r, ) and (r, ) are the density and velocity eld in n the accelerated frame, r, n r, u n r x u r x , , , 3.7 t
1

We assume S A is the ALDA, i.e., the functional that yields the following XC potential:
v XC,ALDA r,t

d dn

LDA

n n

n n r,t

3.2

where XC is the exchange-correlation energy per particle in the HEG at its ground state. The functional S GIXC will be expressed in terms of quantities that are zero at time zero, before any external timedependent perturbation is applied. It is a functional not only of the density, but also the electron uid velocity eld u(r,t) j(r,t)/n(r,t), where j(r,t) is the current density of the uid.13 The velocity is initially zero we assume no external magnetic elds at time zero . In a more general treatment, S A will also include the stationary velocity eld which exists in a static magnetic eld, but here we assume it is a functional of the density only. The functional must observe causality. This is done by introducing a pseudotime and a function t( ), which starts at t 0, changes to the nal time t f , and then is driven backwardsfrom t f to 0. The way we use the pseudotime Keldysh technique is discussed in Appendix A. Based on that technique we set out to construct a functional S GIXC n,u from which the XC potentials can be inferred by functional derivation the actual derivatives and potentials are given in Sec. IV :
v GIXC r,t

When a physical velocity eld u(r, ) up r,t( ) and trajectory x( ) xp t( ) are used in Eq. 3.7 the usual Galilean transformation is recovered: p r,t u up r x t ,t xp t . 3.8

According to Newtons third law, electrons should not produce a net force upon themselves. This requirement imposes a strict condition on the XC potentials, as discussed in Ref. 16, namely, that the net XC force on the electrons is zero. Now, what are the XC forces in the present the ory? Since the XC forces are of formal similarity to electromagnetic potentials, the force they produce per volume should be recovered from the Lorentz force of electromagnetic theory. The Lorentz force is discussed in detail in Ref. 34, where it is also shown that electromagnetic elds are related to their c 1 A and potentials by the following relations: E B A. In the case of a owing charge distribution in an electromagnetic eld, the Lorentz force per volume exerted on a charge distribution owing with velocity u is FL E uB . 3.9

S GIXC n,u n r, S GIXC n,u u r,

,
p

3.3 ,
p

In our case, the XC potentials are dened slightly differently, an analogy with electromagnetism is obtained by setting c 1, v , Aa, and n. The form for the XC potentials given in Eq. 3.4 leads to effective elds given by EGIXC v GIXC GIXC"u a BGIXC aGIXC , P aGIXC , 3.10

aGIXC r,t

where the subscript p denotes evaluation at the physical density n p and velocity eld up . A function f ( ) is called physical if there exists a function f p (t) such that f ( ) f p t( ) , see Appendix A for more details. These poten-

where naGIXC aGIXC . Plugging the elds of Eq. 3.10 into Eq. 3.9 yields the XC force density. In Appendix B a TD-

Downloaded 30 Oct 2004 to 132.64.1.37. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp

J. Chem. Phys., Vol. 121, No. 18, 8 November 2004

Exchange-correlation current density

8735

CDFT generalization of the results of Ref. 16 is given, showing Galilean invariance implies that the net XC force dened this way is zero. In order to impose Galilean invariance on the XC action functional we use a Lagrangian coordinate system.3,17 We introduce the trajectory function R r, , describing the position at pseudotime of an electron uid parcel, which at 0 was at r obeying the equations of motion R r, t u R r, , , R r,0 r, 3.11

potential derived from it is zero though because the functional derivative is done with respect to any not necessarily physical density and velocity elds . In order to obtain the GIXC scalar- and vectorpotentials, aGIXC(r, ) S GIXC / u(r, ), v GIXC(r, ) S GIXC / n(r, ), we need to rst compute the position and velocity Jacobians. The trajectory-position Jacobian matrix is calculated in Appendix C, resulting in Ii j r, R
ij

r,

R i r,

/ rj .

4.1

where u r, is the velocity eld. When a physical velocity eld is used, u(r, ) up r,t( ) , this denition is compatible with the physical trajectory dRp /dt up , where R(r, ) Rp r,t( ) . The Galilean transform of R is R r, r x note: R , R r, x , because r is the R r, position at time 0 when the two frames are identical . Following Refs. 3 and 17, we introduce the Lagrangian veu R r, , and density N(r, ) locity eld U r, n R(r, ), , noting that N r, N r, , U r, U r, x . 3.12

The trajectory-position Jacobian matrix I(r,t) tells us how a path is affected when its initial position is perturbed. In actual computations it can be determined directly from Eq. 4.1 . Furthermore, from Eqs. 4.1 and 3.11 , I(r, ) (r, ) ( ) u ( )U(r, ), from which I t t R(r, ), I(r, ) where u i j u i (r, )/ r j , thus the Jacobian matrix is given by I r, exp
0

u R r,

4.2

Since N N and U U, any functional of N and U is trivially GI i.e., obeys S N,U S N ,U ), we can write a general GIXC action as follows: S GIXC n,u N r, d 3r
0
f

d
0

t ,U r,

d , , , 3.13

where is the -ordering symbol earlier times to the right . The Jacobian determinant ensures particle conservation by correcting for particle density when physical function N(r, ) N p r,t( ) are used see Appendix C : I(r,t) N p (r,0)/N p (r,t). Next, we need also the trajectory-velocity Jacobian matrix Gi j r , ;r, Ri r , / u j r, . 4.3

,N r,

,U r,

where (N ,N ,U ,U , , ) is an appropriate kernel functional. This equation opens the door for a generalized gradient approximation in the time domain. Furthermore, this general XC action is not limited to the linear perturbation regime. To demonstrate the practicality of this GIXC action, we choose a simple nontrivial functional, which in the linear response regime coincides with known forms:35 S GIXC n,u d 3r
0
f

This function tells us how a trajectory originating at r changes at time as a result of a perturbation in the velocity eld applied at position r at time . Taking the derivative of Gi j , one arrives after some manipulawith respect to tions at t
1

Gi j

ij

R r , k ui R r , ,

r Gk j .

d
0

FL N r , t

,t "U r , t U r , 3.14

"U r , ,t .

This equation of motion for G can be readily solved using Eq. 4.2 : Gi j r , ;r, I r , I r ,
1 ij

FT N r , U r ,

R r ,

r .

4.4

The response kernels F L and F T , analogous to the XC energy per particle in LDA, carry the information of a generic physical system, such as that of the HEG. The functions F L,T can then be obtained from approximated form of the linear response LR functions, as discussed in the following section.
IV. EXCHANGE CORRELATION POTENTIALS

When physical density and velocity elds are plugged into Eq. 3.14 , it can be shown that S GIXC is identically zero use Eq. A4 for providing this . This does not mean that the

In this expression it is evident that the only trajectory affected by the perturbation is that which passes at the spaceis the Heavitime point of the velocity perturbation. side function enforcing causality: only the future trajectory is affected. The position Jacobians play the role of propagators. I(r , ) 1 propagates backward from perturbation forward from time zero to time to time zero and I r , present time . An important property of G is its spatial sparsity: it is strictly zero unless r and r refer to the same uid element. G is nonlocal in time, but in any application the memory functionals require only a limited time nonlocality. Thus G can be computed on the y. With the Jacobians,

Downloaded 30 Oct 2004 to 132.64.1.37. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp

8736

J. Chem. Phys., Vol. 121, No. 18, 8 November 2004

Y. Kurzweil and R. Baer

we can write the functional derivatives, determining how changes in the Eulerian elds affect the Lagrangian variables: N r , N r , U r , Ui r , / n r, / u r, / n r, / u j r, R r , N r , 0, t
1

r I r ,
1

/t , G r , ;r, ,

Gi j r ,

;r,

From these relations we compute the potentials, S GIXC n r, S GIXC u r,

f

V , r

0
f

N , 3 r d d , rt n r, U , r u r, N , r d 3 d , rt u r,

Memory is evident in Eq. 4.5 , since the potential is sensitive to the past behavior of the velocity eld. The nal potentials do not depend on the Keldysh contour, as required. The fact that the scalar potential is zero is simply a specic choice of gauge. The longitudal part of as aGIXC /n can be converted to a scalar GIXC potential by an appropriate gauge transform. It is interesting to note that Eq. 4.5 is consistent with the LR theory of the HEG. We note that in LR, we expand all quantities to rst order. The rst order change in density is n 1 . The rst order part of the velocity eld u1 is the leading order since we assume initial zero magnetic elds . This is also true for the vector potential. Therefore, I I o(u1 ), R(r,t) r o(u1 ), and Eq. 4.5 becomes similar to the time domain form of the LR result for a HEG:4 aXC,1 r,t
t 0

A , r

F L n 0 ,t t "u1 r,t
t 0

V , r

S GIXC / N(r, ) and A(r, ) S GIXC / where V(r, ) U(r, ). It is straightforward to verify, using Eqs. A3 and A5 , that when physical densities and velocities are plugged into these expressions and the causality property F L,T (N, 0) 0 enforced, the vector potential comes out causal, having both transverse and longitudal parts: aGIXC R r,t ,t I r,t
t 0 1

F T n 0 ,t t u1 r,t .

4.8

V. THE KERNEL FUNCTIONS

I r,t

dt Ap r,t I r,t ,

4.5

where Ap r,t
t 0

dt F L N r,t ,t t "U r,t

t 0

There are some exact results on the dynamical properties of the HEG in the LR regime. The relevance of these for TDDFT has been discussed in several references.4,13,18,20,36 38 These known and extrapolated propL,T erties are encapsulated in the functions f XC ( ), parametri18,20,38 zations of which have been discussed. The results we L,T present are based on the f XC ( ) of Qian and Vignale.20 In Appendix D we derive explicit expressions for the kernel functions, based on the known HEG response. The nite-memory kernel is derived, M L,T N,t 2n 2
0

f L,T N, i
2

dt F T N r,t ,t t U r,t .

sin td

4.6

F L,T N . 5.1

Furthermore, the GIXC potential is identically zero, i.e.,

v GIXC R r,t ,t

I r,t

V p r,t

0.

4.7

In terms of which the kernels to be used in Eq. 4.6 are given by F L N,t F T N,t where F L,T N
ad FT N

Notice that the left-hand side of Eq. 4.5 gives the vector potential at R(r,t). In actual applications, this will have to be transferred to the Eulerian systems coordinates r. Since R(r,t) is known, this should not present a problem. The determination of the vector potential is done in two stages. a First, the vector potential is determined by Eq. 4.6 in a way similar to linear response theory. However, the integral is not a convolution and cannot be performed in frequency space. One possible approximation is to use N(r,t 0 ) instead of N(r,t ) in Eq. 4.6 . t 0 is a representative time for example, t 0 can be equal to t . This may facilitate the computation since the vector potential Ap will now be a convolution. b Once Ap is obtained, a transformation to the Eulerian frame takes place, via Eq. 4.5 . This transformation involves that Jacobians, which act as translation operators along the uid parcel trajectory.

FL N FT N

M L N,t ,
ad FT N t

5.2 M T N,t , 5.3

N 2 f i L,T N,0 , N 2 f T N,0 . i 5.4

The memory functions M L,T (n) are shown in Fig. 1, where the density is described by the Wigner Seitz parameter r s ad (3/4 n) 1/3. The adiabatic constant F T is shown in Fig. 2. Summarizing, we only need the time derivative of the Lagrangian vector potential Ap in Eq. 4.5 , which is composed of three parts: Ap r,t Amem r,t p Ap r,t Aad r,t , p 5.5

where the memory part,

Downloaded 30 Oct 2004 to 132.64.1.37. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp

J. Chem. Phys., Vol. 121, No. 18, 8 November 2004

Exchange-correlation current density

8737

FIG. 1. The longitudinal left and transverse right memory function kernels for various densities values of r s ). Based on the LDA PW92 Ref. 42 functional and the Qian-Vignale f XC parametrization of Ref. 20.

Amem r,t p

t 0 t 0

dt M L N r,t ,t t "U r,t dt M T N r,t ,t t U r,t , 5.6

is evaluated by using a limited history of the uid velocity and density because M L,T (n,t) is short ranged in t. The innite response given by Ap r,t F L N r,t "U r,t F T N r,t U r,t , 5.7

is evaluated with no need to refer to history. And the adiabatic part Aad r,t p
t 0 ad dt F T N r,t

U r,t ,

5.8

which can be calculated on the y with no need for memory, since this integral can be incremented at each time step. We have shown explicit kernels for the memory functional and discussed their application stressing the fact that all calculations are doable, in the sense that they involve only a limited account of the history of the system.
VI. SUMMARY AND DISCUSSION

In this paper we formulated an action approach to TDCDFT for electronic systems in a TD electromagnetic eld, with the aim of constructing a memory action functional. Our formulation of the action is a generalization of

ad FIG. 2. The constant F T as a function of electron density parameter r s ).

the action devised for TDDFT.14 We derived simple memory functionals that are robust due to their inherent Galilean invariance and may be useful for taking into account memory effects in time-dependent calculations with strong elds. The use of a Lagrangian framework, as rst suggested in Refs. 17 and 3 allows a full formulation of the memory effects in a Galilean invariant way. The functionals are compatible with linear response properties, where the Lagrangian and Eulerian frames are identical. Comparing with the theory of DBG Ref. 3 , from which the present approach has been inspired, we nd results are different in several aspects. While DBG derive their theory from an elegant application of Newtons third law, it is not clear if their potential can be derived from an action principle. The same question can be raised when comparing to the theory of Tokatly and Pankratov,11 which assumes that Lorentz force is a divergence of a stress tensor. In our treatment, we use a more general assumption and impose GI using a transformation from the Lagrangian to the Eulerian frame. The importance of this method still needs to be examined by application to various benchmark systems. Furthermore, in the present approach the transverse part of the response is fully included. In principle, the present approach can encompass more elaborate ansatz than the one introduced in Eq. 3.14 . Future work will then address two issues related to this. A nonNewtonian-liquid approach should be attempted, where the derivatives of the Lagrangian velocity will be inserted in a nonlinear way. Another issue is spatial locality. The current formalism is based on the local density approximation. Thus, there is no account of nonlocal spatial effects. While simply using the adiabatic GGA as a memory-less action functional in place of S A in Eq. 3.1 is a possibility, a more rigorous attempt, which combines nonlocality in space and in time is still a challenge. The incorporation of these equations in a real-time timedependent scheme is more involved than the simple localdensity approximation. However, the inclusion of memory effects may be the only way to improve the approximations we currently have for the dynamics of electrons in molecules. While the present approach is far from addressing all the known properties of the true functional for example, the initial state dependence39 is not addressed here , we believe it is a step forward, supplying a formalism that can be applied to general electronic structure systems.

Downloaded 30 Oct 2004 to 132.64.1.37. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp

8738

J. Chem. Phys., Vol. 121, No. 18, 8 November 2004

Y. Kurzweil and R. Baer

ACKNOWLEDGMENT

The authors gratefully acknowledge the support of the German Israel Foundation.

APPENDIX A: CAUSALITY VIA THE KELDYSH CONTOUR

A special technique must be used to enforce causality in the action. Following van Leeuwen,14 the Keldysh time contour is used. To explain the basic idea, let us dene the Keldysh pseudotime. Suppose the relevant interval for the physical time t is 0,t f (t f can be innity . We dene a pseudotime 0, f for some f 0 and a parametrization t( ) which maps 0, f onto 0,t f with t(0) t( f ) 0. An important related concept is a physical time function: a function f ( ) f may depend on other variables as well for which there exists a function f p (t) such that f ( ) f p t( ) is a physical time function. Peuckert,32 Rajagopal,33 and van Leeuwen,14 describe an action, based on the Keldysh time contour, which by derivation produces causality-respecting potentials. Our approach to causality is also based on the Keldysh contour. Consider the following functional A of a pseudo-time-dependent function n( ): A n
0
f

The potential v p (t) is now causal in the sense that the functional derivative v p (t)/ n(t ) is evidently zero when t t. Another approach, more in the spirit of the Keldysh formalism40 is to consider the time ordered functions on the contour, dening F (1,2) for 1 2 and F (1,2) for 1 2 . Upon returning to physical quantities, only the combi(t 2 t 1 )(F F ), called the retarded kernel, nation F R survives. Such an approach was taken by van Leeuwen in his calculation of the response under TDDFT.14 Thus, our formulation is akin to the retarded functional in the Keldysh theory.

APPENDIX B: ZERO NET XC FORCE

In this appendix we show that the Galilean invariance of the XC action ensures zero net XC Lorentz force. The expression for the Lorentz force is given by Eq. 3.9 , where the elds of Eq. 3.10 are used. Let us assume that Eq. 3.6 holds and prove that the net XC Lorentz force is zero. We mount a frame of reference slightly perturbed at some time 0 see Eq. 3.7 . The density and velocity elds appear slightly distorted in the perturbed frame primed quantities : n u x n r u r x x xt
1

, ,

, t . S XC S XC n ,u S XC u
1

t
1

d n

F n d
1.

,t

B1

A1

The functional derivative is A n n

f

From S XC n,u d
C

Galilean invariance, 0, thus we have S XC n x n x"u t

1

F n

,t

d 3r

0. B2

n
0

F1 n

,t

1,

A2 where F 1 is the derivative of F by its rst argument. By plugging in a physical time function n( ) n p t( ) one obtains, after some manipulation,
vp t

Expressing this in indices, using the convention that repeated indices are summed, using Eq. 3.3 , and multiplying by 1, for later convenience, the equation is transformed into
v jn

ai jui t

aj

x j d 3 r 0,

B3

A n n t

0 p t

dt 2 F n t 2 ,t 2 t
t 0

where a aGIXC na and v v GIXC for brevity. Integrating by parts the rst term and using the fact that x j is arbitrary, we have A3 n j v na i j u i t
1

dt 1 n p t 1 F 1 n t ,t t 1 .

In deriving Eq. A3 , we used the following fact: for any physical time function f p (t) and pseudotimes 1 and 2 t
1

8 a na j j n d 3 r 0.

B4

Using the continuity equation, we have n Since i j u i a n

j jv

2 1

f t

0.

A4

18

na i j u i j a

nu i d 3 r 0.

B5

A crucial point: by choosing in Eq. A1 , F(n,t) to be causal, i.e., to be zero whenever t 0, the functional derivative of A4 is reduced to a single term,
vp t

j (a i u i ) t

u i j i , we have a
18

a v iu i

nu i j i j a a

nu i d 3 r 0. B6

A n n t

t p 0

n p t 1 F 1 n t ,t t 1 dt 1 .

A5 Integrating the last term by parts, we have

Downloaded 30 Oct 2004 to 132.64.1.37. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp

J. Chem. Phys., Vol. 121, No. 18, 8 November 2004

Exchange-correlation current density

8739

a v iu i

18

n u i j i u i i j d 3 r 0. a a B7

I r,

exp
0

u R r,

C6

Finally, we use the identity u and write n

j

u i j i u i i j a a

B8

a v iu i

18

ua

d 3 r 0.

B9

where is a time ordering operator earlier times appear to the right and u i j u i / r j . Consider a small volume element d 3 r. The number of particles in this element is N(r,0)d 3 r. At time t the element has moved to R(r,t), its shape and volume changed but the number of elements N(r,t) I(r,t) 1 d 3 r must still be the same, thus N r,t I r,t
1

The integrand in the curly brackets is the average XC force P a per particle. Because a and ( v "u) a are gauge invariant this force too is gauge invariant. For the usual TDDFT without vector potentials this expression reduces to the expression of Vignale,16 v nd 3 r 0.
APPENDIX C: LAGRANGIAN FRAME

N r,0 .

C7

This equation is a useful way to compute the Jacobian determinant. This result can be derived more rigorously from Eq. C6 . Consider the determinant of a small time slice t I is simply an ordered product of such slices, and the determinant of a product is the product of determinants . Because e A e tr A for any operator, we have e
u R r,t ,t t

In this appendix we review several properties of the Lagrangian quantities. Consider the Lagrangian density N(r, ) n R(r, ), . Taking the time derivative, in obvious notation, we have N r, n R r, , , u R r, , t . C1

e tr

u R r,t ,t

C8

However, tr u R r,t ,t "u R r,t ,t C9

and using Eq. C3 we nd tr u R r,t ,t t ln N r,t , C10

n R r,

Now, consider the physical density n(r, ) n p r,t( ) , velocity eld, etc., then, with omission of the subscript p N r,t t n t u"n R r,t ,t . C2

and plugging into Eq. C8 gives e

u R r,t ,t t

N r,t . N r,t t

C11

Using the continuity equation 2.10 evaluated at R(r,t), we obtain N r,t t N r,t "u R r,t ,t 0. C3

Taking the product of all time slices we observe that the Jacobian determinant equals N r,t t N r, t N r,t N r,t t N r,t 2 t N r,0 conrming Eq. C7 . N r,t , N r,0 C12

Next, we consider the Jacobian Ii j r, R i r, / rj C4

APPENDIX D: THE RESPONSE KERNEL

of the Eulerian-to-Lagrangian transformation rR . The Jacobian is needed because after formulation in the Lagrangian frame, we must transform back to the Eulerian frame, where the other functionals such as the adiabatic, Hartree, and external functionals are dened. Its determinant, which is used in various integrals, is also discussed. In actual applications the Jacobian is readily available from the function R(r,t). In order to study its properties, it is instructive to obtain an equation of motion for it, which is obtained by taking the derivative of Eq. C4 with respect to r: Ii j r, ui rk Ik j r,t t
R r, ,

C5

Here we used Eq. 3.11 and the convention that repeated indices are summed over. Equation C5 is a differential equation on I, which together with the initial condition that I r,0 I, can be solved formally as

In this appendix we discuss the construction of the memory kernel from the LR functions f L,T (n, ). To simplify the notation, we denote this function as f ( ), i.e., we drop the L, T super and subscripts and the explicit dependence on n. Let us recall the denition of the function f L ( ). It arises in the context of LR treatment of the homogeneous electron gas of density n. A weak perturbation by some external eld, starting at t 0 ensues a density response n 1 (r,t) n(r,t) n. The time-dependent density n(r,t) is only slightly different from n and n 1 (r,t) is proportional to the strength of the perturbation. The change in density n 1 (r,t) induces a change in the XC potential v 1,XC(r,t) v XC(r,t) v XC(r), depending linearly on n 1 :
t

v 1,XC r,t

XC

r r ,t t n 1 r ,t dt d 3 r . D1

Downloaded 30 Oct 2004 to 132.64.1.37. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp

8740

J. Chem. Phys., Vol. 121, No. 18, 8 November 2004

t

Y. Kurzweil and R. Baer

Assuming further that a local density approximation is appropriate, i.e., (r r ) XC n(r),t XC(r r ,t t ) t we obtain
t

a1,XC t Because

dt
0

t 0

t n 1 t dt .

D11

(t) is causal, we have

t 0

v 1,XC t

t t n 1 t dt ,

D2 a1,XC t where t
0

t t n 1 t dt ,

D12

where for notational clarity we drop the n(r) dependence and the XC subscript. Also, for the HEG, n is independent of r. The function (t) is known to be composed of two parts. One is a function that has a Fourier transform (t) and the other is a function, associated with the innite frequency response of the HEG: t t f t . D3

(t) is the causal function

t

t dt

f .

D13

Using the second relation in Eq. 3.4 , a1,GIXC na1,XC , we obtain an expression for the GIXC vector potential
t

Because of causality, let us assume explicitly that the function (t) is zero for t 0 thus it is possible to extend the upper limit of integration to innity:
v 1,XC t

a1,GIXC t

n
0

t t n 1 t dt .

D14

Let us now plug in the continuity equation 2.10 , which in linear response regime is
t

t t n 1 t dt

f n1 t .

D4 n1 t D5 We have then a1,GIXC t n

Fourier transforming this convolution, we have

v 1,XC

"u1 t dt .

D15

n1

where n 1 ( ) n 1 (t)e i t dt with an analogous expression for v 1,XC) and f f

0

n 2

t t dt
0

t 0

"u1 t dt .

D16

t e i t dt.

D6

Integrating by parts, we have a1,GIXC t where t n2

t

The fact that f ( ) f is the Fourier transform of a causal function poses a constraint on its analytical structure. Replacing by a complex frequency z, where Im(z) 0, on the right-hand side of Eq. D6 yields a converging integral and thus constitutes an analytical continuation of f ( ) into the upper complex plane. From this fact, it is possible to derive KramersKronig KK relations41 fr and fi 1 P fr f d , D8 f 1 P fi d D7

t 0

t t "u1 t dt ,

D17

t dt ,
0

D18

and taking the time derivative, using Eq. D13 , we have t n2

t

dt
0

D19

In other words we nd that in order for the kernel of Eq. 4.6 to be compatible with the LR properties of the HEG assumed to be known as (t), or f ( ) f ), the following must hold: t n2 t 0 D20

where f is decomposed into its real and imaginary parts f fr ifi . D9

with the initial conditions 0 0, n2 f . D21

The function f ( ) contains also the adiabatic LDA response. This is encapsulated in the real zero frequency coefcient18 f( 0) f 0 (n) d 2 /dn 2 ( XCn), where XC(n) is the XC energy per particle for the HEG in its ground state. Because is real, the following is valid for t 0: 2 t fi fr sin td

Taking the Fourier transform of Eq. D20 and using Eq. D6 , we have t e i t dt n 2 f
0

D22

And nally solving for (t) at t 0 see Eq. D10 , we obtain D10 t 2n 2
0

cos td .

fi

sin td .

D23

Using a gauge transformation in Eq. D2 , we can transform v 1,XC(t) into a vector potential

The solution is therefore

Downloaded 30 Oct 2004 to 132.64.1.37. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp

J. Chem. Phys., Vol. 121, No. 18, 8 November 2004

Exchange-correlation current density

8741

2n 2
0

fi
2

sin td 0

D24 0 and

F T t n 2 f T t n 2 f T 0 . 0 i So, writing F T (t) F T functional is MT t 2n 2

0 ad FT t

D32 M T (t), the nite memory

and the constants and (0) n 2 f , thus 0, 2n 2

0

are selected to ensure fi

fT i
2

sin td

FT ,

D33

n f .
0 f i(

D25 )/ d f0

where FT
1 2

From the KK relations Eq. D7 , (2/ ) f , thus we have t n 2 f 0t 2n

2 0

n2 f T 0 , i

ad FT n2 f T 0 . i

D34

fi
2

sin td .

D26

In order to check this result further, let us Fourier transform (t), obtaining ( ) (n 2 / 2 ) f ( ), and using it in na1,XC 1,GIXC( ) ( )"u1 ( ) the Fourier vera sion of Eq. D17 , we obtain a1,XC n
2

"u1

D27

an expression that directly compares to results of Ref. 4. The term n 2 f 0 t in Eq. D26 is linear in time and gives after gauging back to a potential, the adiabatic LDA potential, v ALAD f 0 n 1 . This is because18 f 0 (n) (d 2 /dn 2 ) ( XCn). Thus it should not be part of the kernel F in Eq. 4.6 , since the adiabatic potential is obtained from the functional derivative of S A in Eq. 3.1 . We nd that the kernel in Eq. 4.6 must be given by FL t 2n 2
0

fL i
2

sin td .

D28

The function F L (t) does not decay to zero, instead we have lim F L t
t

n2 f i 0

FL .

D29

This constant in time is treated separately, i.e., we dene the nite memory functional F L (t) M L (t) F L with ML t 2n 2
0

fL i
2

sin td

FL .

D30

It is this kernel which is used in Eq. 4.6 . The constant F L (n) is treated separately and because it is constant, its incorporation poses no problem. A similar treatment applies to F T , only here we do not take out the ALDA part, thus FT t n 2 f Tt 0 2n 2
0

fT i
2

sin td .

D31

The behavior at large t is thus

E. Runge and E. K. U. Gross, Phys. Rev. Lett. 52, 997 1984 . S. K. Ghosh and A. K. Dhara, Phys. Rev. A 38, 1149 1988 . 3 J. F. Dobson, M. J. Bunner, and E. K. U. Gross, Phys. Rev. Lett. 79, 1905 1997 . 4 G. Vignale and W. Kohn, Phys. Rev. Lett. 77, 2037 1996 . 5 P. Hohenberg and W. Kohn, Phys. Rev. 136, B864 1964 . 6 W. Kohn and L. J. Sham, Phys. Rev. 140, A1133 1965 . 7 G. Vignale and M. Rasolt, Phys. Rev. B 37, 10685 1988 . 8 R. M. Dreizler and E. K. U. Gross, Density Functional Theory: An Approach to the Quantum Many Body Problem Springer, Berlin, 1990 . 9 H. Appel, E. K. U. Gross, and K. Burke, Phys. Rev. Lett. 90, 043005 2003 . 10 P. Hessler, N. T. Maitra, and K. Burke, J. Chem. Phys. 117, 72 2002 . 11 I. V. Tokatly and O. Pankratov, Phys. Rev. B 67, 201103-1 2003 . 12 Y. Kurzweil and R. Baer unpublished . 13 G. Vignale, C. A. Ullrich, and S. Conti, Phys. Rev. Lett. 79, 4878 1997 . 14 R. van Leeuwen, Phys. Rev. Lett. 80, 1280 1998 . 15 R. van Leeuwen and E. J. Baerends, Phys. Rev. A 49, 2421 1994 . 16 G. Vignale, Phys. Rev. Lett. 74, 3233 1995 . 17 J. F. Dobson, Phys. Rev. Lett. 73, 2244 1994 . 18 E. K. U. Gross and W. Kohn, Phys. Rev. Lett. 55, 2850 1985 . 19 C. A. Ullrich and G. Vignale, Phys. Rev. B 65, 245102 2002 . 20 Z. X. Qian and G. Vignale, Phys. Rev. B 65, 235121 2002 . 21 P. L. de Boeij, F. Kootstra, J. A. Berger, R. van Leeuwen, and J. G. Snijders, J. Chem. Phys. 115, 1995 2001 . 22 M. van Faassen, P. L. de Boeij, R. van Leeuwen, J. A. Berger, and J. G. Snijders, Phys. Rev. Lett. 88, 186401 2002 . 23 M. E. Casida, C. Jamorski, K. C. Casida, and D. R. Salahub, J. Chem. Phys. 108, 4439 1998 . 24 S. Hirata, M. Head-Gordon, and R. J. Bartlett, J. Chem. Phys. 111, 10774 1999 . 25 R. Baer and R. Gould, J. Chem. Phys. 114, 3385 2001 . 26 R. Baer, Chem. Phys. Lett. 264, 75 2002 . 27 R. Baer, D. Neuhauser, P. Zdanska, and N. Moiseyev, Phys. Rev. A 68, 043406 2003 . 28 A. Dreuw, J. Wiesman, and M. Head-Gordon, J. Chem. Phys. 119, 2943 2003 . 29 F. Aryasetiawan and O. Gunnarsson, Phys. Rev. B 66, 165119 2002 . 30 N. T. Maitra, K. Burke, and C. Woodward, Phys. Rev. Lett. 89, 023002 2002 . 31 Z. Qian, A. Constantinescu, and G. Vignale, Phys. Rev. Lett. 90, 066402 2003 . 32 V. Peuckert, J. Phys. C 11, 4945 1978 . 33 A. K. Rajagopal, Phys. Rev. A 54, 3916 1996 . 34 J. D. Jackson, Classical Electrodynamics Wiley, New York, 1999 . 35 T. K. Ng, Phys. Rev. Lett. 62, 2417 1989 . 36 N. Iwamoto and E. K. U. Gross, Phys. Rev. B 35, 3003 1987 . 37 R. Nifosi, S. Conti, and M. P. Tosi, Phys. Rev. B 58, 12758 1998 . 38 N. Iwamoto and E. K. U. Gross, Phys. Rev. B 35, 3003 1987 . 39 N. T. Maitra and K. Burke, Phys. Rev. A 6304, 042501 2001 . 40 P. Danielewicz, Ann. Phys. Leipzig 152, 239 1984 . 41 R. Kubo, M. Toda, and N. Hashitsume, Statistical Physics II: Nonequilibrium Statistical Mechanics Springer, Berlin, 1995 . 42 J. P. Perdew and Y. Wang, Phys. Rev. B 45, 13244 1992 .

Downloaded 30 Oct 2004 to 132.64.1.37. Redistribution subject to AIP license or copyright, see http://jcp.aip.org/jcp/copyright.jsp